betaFeedback

WorksPhysics and Astronomy

Beyond node strengths: the topological content of weighted bipartite networks

Abstract

Weighted bipartite networks encode the topology and intensity of interactions between two disjoint sets of nodes. However, it is unclear whether the observed pattern of connections is already explained by node interaction volumes or contains additional structural information. Here we introduce a hierarchy of maximum-entropy models for continuous weighted bipartite networks that preserves node strengths alone, strengths and global density, or strengths and the complete degree sequence. Across economic, technological, ecological, social and lexical systems, we find that strength constraints alone are generally insufficient to reproduce empirical structures. Yet the missing topological information is often low-dimensional: a single constraint controlling network density frequently provides the most parsimonious description, whereas networks with stronger residual degree heterogeneity require node-specific constraints. Beyond downstream inference, we also show how the choice of topological resolution changes which interactions and projected relationships are identified as statistically significant. Our results indicate that topology is essential, but often compressible, in the statistical modeling of weighted bipartite networks, and provide a scalable computational framework for fitting, sampling, model selection and validation within a common workflow.

Introduction

Bipartite weighted networks describe systems in which interactions occur between two distinct classes of entities and vary in intensity [1, 2]. Countries export different volumes of products [3, 4], plants receive repeated visits from pollinators [5, 6, 7], users interact with varying intensity over online communities [8, 9], and words occur with different frequencies across documents [10, 11, 12, 13]. In each case, the data contain two related but distinct forms of information: the topology specifies which pairs interact, whereas the weights quantify the intensity of the observed interactions.

A central problem in the analysis of these systems is determining whether topology carries information that is not already encoded by node interaction volumes – namely, their strengths. On one hand, strength values alone do not specify how weights are divided among existing links: the same volume can be realized through many weak interactions or a few strong ones, producing radically different topologies. On the other hand, in the absence of additional structural mechanisms, nodes with larger (smaller) strengths naturally tend to connect to more (less) partners. Whether the observed pattern of zero and positive weights can be explained by weighted marginals is therefore a statistical question requiring an explicit statistical benchmark.

Maximum-entropy network models provide a principled way to construct such benchmarks [14, 15, 16, 17]. They define the least biased probability distribution over networks that reproduces selected quantities of an empirical system. The choice of these quantities (the constraints of the entropy maximization procedure) determines which information is treated as explanatory; deviations from the resulting ensemble identify empirical structures that cannot be attributed to that information. Hence, a strength-only constrained model can test whether node activity is sufficient to explain both presence and weights of individual interactions. Adding constraints on the number of edges, globally or node by node, instead treats topology as an additional source of information.

The distinction between weighted and topological information is known to be consequential in monopartite networks. Strength-constrained models [16, 18, 19, 20] often predict networks that are much denser than observed, because entropy maximization tends to spread node activity across all possible links, whereas jointly constraining degrees and strengths can greatly improve reconstruction [21, 22, 23]. Intermediate heuristic schemes further showed that strengths can be complemented by limited topological information, such as the link density of the network [24, 25] or an estimate of it [26]. More recently, researchers have distinguished integrated from conditional mechanisms for generating links and assigning weights [27, 28, 29] and developed scalable algorithms for parameter inference [30].

The corresponding problem is less developed for bipartite networks. While maximum-entropy methods have been studied extensively for binary bipartite data and their one-mode projections [31, 32, 33, 34], existing continuous-weight models are limited to strength-only formulations [35], possibly adding the cost of establishing connections [36]. Consequently, it is unclear not only how topology can be constrained, but also whether this is needed to model empirical networks and at what resolution: is global density sufficient, or must the connectivity of every node be retained?

Here we answer this question by developing a unified nested hierarchy of canonical maximum-entropy models for continuous weighted bipartite data. The Bipartite Weighted Configuration Model (BiWCM) constrains the strength sequences of the two sets of nodes and leaves topology unconstrained. The Partial Bipartite Enhanced Configuration Model (P-BiECM) additionally constrains the total number of links, introducing one parameter for global edge density. The Bipartite Enhanced Configuration Model (BiECM) jointly constrains the strength and degree sequences of both sets, preserving node-specific connectivity.1 This hierarchy allows us to determine the minimum topological information required for an adequate statistical description of a given empirical system.

Across networks from five domains, we find that strengths alone are generally insufficient. Remarkably, however, topology does not always need to be specified node by node: a single density constraint often captures most of the information missing from the strength-only ensemble and is favored after penalizing model complexity. Local degree constraints become necessary when empirical connectivity remains heterogeneous after conditioning on strengths and density. Synthetic benchmarks with controlled density, heterogeneity and mesoscale organization reproduce this transition. These results establish topology as an essential, but often compressible, component of weighted bipartite network reconstruction. They also show that selecting an inadequate structural resolution can change which interactions and projected relations are considered statistically significant.

Results

A hierarchy of hypotheses about topology

Let W=(wiα)W=(w_{i\alpha}) be a non-negative weighted biadjacency matrix connecting NTN_{\mathrm{T}} nodes {i}\{i\} in layer T to N⊥N_{\perp} nodes {α}\{\alpha\} in layer ⊥\perp. Its binary support is biα=1R+(wiα)b_{i\alpha}=\mathbb{1}_{\mathbb{R}_{+}}(w_{i\alpha}) (1\mathbb{1} is the indicator function), hence the number of links is L=∑i,αbiαL = \sum_{i,\alpha} b_{i\alpha}. Node strengths and degrees are

si=∑αwiα,σα=∑iwiα,ki=∑αbiα,hα=∑ibiα.(1)s_i = \sum_{\alpha} w_{i\alpha}, \qquad\sigma_\alpha= \sum_i w_{i\alpha}, \qquad k_i = \sum_{\alpha} b_{i\alpha}, \qquad h_\alpha= \sum_i b_{i\alpha}. \tag*{(1)}

The three maximum entropy models we consider retain the same weighted marginals but progressively enrich the description of topology (Table 1). They therefore encode competing hypotheses, rather than merely increasingly complicated approximations to the same object.

Model# links LLdegrees {ki}\{k_i\}, {hα}\{h_\alpha\}# parameters qq
BiWCM--N⊤+N⊥N_{\top}+N_{\perp}
P-BiECM✓-N⊤+N⊥+1N_{\top}+N_{\perp}+1
BiECMimplied✓2(N⊤+N⊥)2(N_{\top}+N_{\perp})

Table 1. Hierarchy of continuous weighted bipartite models. Check marks denote independently constrained quantities. All models preserve the strength sequences; they differ in whether topology is unconstrained, represented by global density, or specified through the two degree sequences.

Full model derivations are given in Methods and in Section A of the Appendix. In a nutshell, BiWCM associates an exponential weight distribution to every link, hence describes a fully connected system. The two enhanced models instead combine an exponential density for positive weights with an atom at zero: their link probability depends jointly on topological and weighted information.

Model selection

Models are fitted on a given empirical network by maximum likelihood (see Methods and Section A of the Appendix for more details and the resolution correction used to compare likelihoods). Comparison among models {m}\{m\} is then performed using such maximized likelihood Lm\mathcal{L}_m as well as the Bayesian Information Criterion (BIC):

BICm=qmlog⁡(N⊤N⊥)−2log⁡Lm,(2)\mathrm{BIC}_m = q_m \log(N_{\top}N_{\perp}) - 2\log\mathcal{L}_m, \tag*{(2)}

where qmq_m is the effective number of free parameters and N⊤N⊥N_{\top}N_{\perp} is the number of dyads. Lower BIC values indicate a better balance between goodness of fit and model complexity. Further details can be found in Methods and Section A of the Appendix.

We investigated the performance of the various models upon a large corpus of economic, technological, ecological, social, and lexical systems (see Methods and Section B of the Appendix for further details on the empirical networks). For clarity, the economic, social, and technology datasets are repeated observations (for instance, the bipartite representation of international trade over different years), not independent systems.

Figure 1 pictorially reports likelihood and BIC ratios among BiWCM, P-BiECM and BiECM for our real data. We first assess that strength sequences alone cannot provide an adequate statistical description of the considered systems: BiWCM is consistently disfavored relative to BiECM, even after accounting for its lower number of parameters. Once strengths-alone modeling is rejected, the relevant question is how much topology must be retained. The P-BiECM adds only one independent topological constraint, whereas BiECM adds the heterogeneous degree sequences of both layers. As our hierarchy is genuinely nested (P-BiECM is a limit case of BiECM), the raw likelihood necessarily favors the more constrained model, but the improvement is not always sufficient to offset its parameter cost.

Model comparison across empirical weighted bipartite networks

Figure 1. Model comparison across empirical weighted bipartite networks. Likelihood (a) and BIC (b) ratios of BiWCM vs BiECM (horizontal axis) and of P-BiECM vs BiECM (vertical axis); BiECM is used as the benchmark for the comparison, as it has the largest set of constraints and thus the highest likelihood, by definition. Each point represents one of the 46 empirical networks, coloured according to the class it belongs to; the outline marks the best model, i.e. the one with the highest likelihood (a) or the lowest BIC (b). Horizontal and vertical lines mark the values of 1, where models become equivalent. Because log-likelihoods are negative, the likelihood ratios are greater than 1, and the farther from 1, the worse the numerator model performs relative to BiECM. For BIC, ratios are greater than 1 when the model in the numerator performs worse than BiECM and smaller when it performs better (because the lower the BIC, the more effective the model). While both BiWCM and P-BiECM have lower likelihoods, BIC selects P-BiECM in 32 networks and BiECM in 14, while BiWCM is never selected.

In ecological, social and lexical networks, P-BiECM frequently provides the smallest BIC, meaning that the degree sequences information can be compressed, for model-selection purposes, into one density constraint. Economic and technological matrices instead more often favor BiECM, indicating that local connectivity contains information not captured by strengths and density alone.

The domain labels summarize the observed regimes but do not explain them. Several attempts were made to relate model preference and cross-model disagreement to elementary properties of the input network (see Section C of the Appendix), but no clear trend allowing a reliable a priori criterion was observed.

Network reconstruction

We now test the model’s ability to reproduce the empirical topology in terms of individual links. For each model, the link probability is jointly determined by the imposed constraints (see Methods). Figure 2 shows that the Positive Predictive Value PPV for the existence of links obtained with P-BiECM is systematically lower than the one obtained with BiECM. The difference is appreciable for some ecological and social networks, while trade and technological systems tend to display higher PPV under both models. Thus, preserving the global density captures only part of the information contained in the empirical support: node-specific degree constraints improve the ability of the model to identify where interactions actually occur. Note that BiWCM is not included in this comparison as it basically cannot predict non-existing links.

Figure 2. Reconstruction of the observed links. Comparison between P-BiECM and BiECM in the reconstruction of the observed links of the 46 empirical networks, measured through the positive predictive value PPV=⟨TP⟩/(⟨TP⟩+⟨FP⟩)\mathrm{PPV} = \langle\mathrm{TP}\rangle/(\langle\mathrm{TP}\rangle+ \langle\mathrm{FP}\rangle), the expected fraction of predicted links that are observed (Methods). The dashed line is the identity; the median difference between P-BiECM and BiECM is −0.146-0.146. The specificity, i.e. the fraction of empirical zeros predicted as zeros, is reported in Section D of the Appendix.

We next test models against properties not imposed as constraints, including nearest-neighbour degree and strength, clustering and nestedness (definitions in Methods). In the following, we present results for a representative of our set of ecological networks (Figure 3, panels a-d), and a representative of our set of trade networks (Figure 3, panels e-h): more results can be found in Section E of the Appendix.

Figure 3. Reconstruction of unconstrained structural properties for an ecological network (Lake of the Woods, a–d) and a trade network (export 2020, e–h). a,e, Mesoscopic reconstruction of the average nearest-neighbour degree knnk_{nn} as a function of the node degree kk, and c,g, of the average nearest-neighbour strength snns_{nn} as a function of the node strength ss (one point per node and sample). b,f, Macroscopic reconstruction of the binary structure, clustering CC against NODF, and d,h, of the weighted structure, weighted clustering CWC_{W} against WNODA, with the marginal distributions of the samples. Red crosses: observed values; 2×1032 \times10^{3} samples per model (definitions in Methods). The BiWCM samples are complete graphs, so their binary observables are not informative and are not shown in a, b, e and f. In the dense trade network the weighted clustering saturates (CW≃1C_{W} \simeq1) in the data and in all ensembles, so this observable does not discriminate between the models (h).

As shown in all panels of Figure 3, BiECM generally provides the most accurate reconstruction of purely topological observables, as expected given that it accounts for the full degree sequences. For observables combining topology and weight, however, the relative performance of the models is less consistent across systems, as illustrated by comparing panels c and d with panels g and h. In the country-product matrix, for instance, the observed nearest-neighbour strength (snns_{nn}) patterns lie close to the BiECM expectations (panel g), suggesting that the model captures the local connectivity heterogeneity of the system. A similar trend emerges for ecological networks, although the expected distribution of snns_{nn} is broader (panel c).

More pronounced differences emerge for other weighted observables. In ecological networks, both BiECM and P-BiECM reproduce weighted nestedness (WNODA) reasonably well, whereas only BiECM effectively reconstructs weighted clustering (CWC_W; panel d). In international trade, the pattern is almost reversed (panel h): all models effectively reconstruct CWC_W, but none reproduces WNODA. Moreover, the WNODA results are statistically significant for all models, with the corresponding deviations falling on the right-hand side of the reference distributions for BiECM and on the left-hand side for P-BiECM.

These results highlight system-specific structural features that remain unexplained by the selected low-order constraints. In international trade, reproducing the degree and strength sequences is insufficient to capture weighted nestedness, pointing to the need for additional information about the hierarchical organisation of export baskets. By contrast, the higher-order patterns measured by CWC_W are already well captured by the models. In ecological networks, the picture is different: weighted nestedness is reproduced comparably well by both BiECM and P-BiECM, whereas the constraints encoded in P-BiECM are insufficient to reproduce weighted clustering, which is effectively captured only by BiECM.

This comparison complements, but should not be conflated with, information-theoretic model selection. A model may improve a particular observable while being disfavoured after accounting for its additional parameters. Conversely, selecting P-BiECM by BIC means that its loss of likelihood is not sufficient to justify the full degree sequences; it does not imply that P-BiECM reproduces every topological statistic as accurately as BiECM.

Statistical validation

The effect of including the topological information in the model may have deep implications. Maximum-entropy ensembles are frequently used not only to statistically model a network but also to identify patterns that are unlikely under a specified null hypothesis: that the imposed constraints explain everything of the empirical system [17]. When this null is rejected, it means that the detected pattern carries genuinely additional information beyond that used to define the model. Therefore, the choice of the null model and the ingredients it should include are fundamental for the conclusions to be deduced from the statistical validation.

Weights validation

We start with the statistical validation of link weights. The procedure is pictorially represented in panel a of Figure 4. For each observed positive weight, we computed its p-value (upper-tail probability) under each ensemble. The p-value calculation is analytical in the Lagrangian multipliers and, thus, it does not need any sampling (see Methods). P-values are then validated, controlling the false discovery rate across dyads [38] (again, see Methods for further details). Panel a of Figure 5 shows the p-value profile of the various models for an ecological mutualistic network between plants and ants: the same empirical matrix produces different p-value distributions and validated interaction sets under the three null hypotheses. Qualitatively, the same conclusions can be drawn for other empirical networks in our corpus.

Statistical validation of bipartite weights and co-volumes. Panel a, Bipartite Weighted Validation; panel b, Co-Volume Validation. Graphical schematics of observed networks, null-model distributions, and validated networks.

Figure 4. Statistical validation of bipartite weights and co-volumes. a, Each observed weight is compared with its distribution under the null model (analytical upper-tail p-value), and the links with a significant p-value are retained. b, The one-mode projection gives the co-volume of each pair of nodes; its null distribution is obtained by sampling the bipartite ensemble, and the pairs with a significant co-volume form the validated projection.

Figure 5. Validation of weights and co-volumes in the plants–ants network (41 plants, 51 ant species). Sorted p-values under the BiWCM, the P-BiECM and the BiECM for a, the 285 observed links (inset: zoom on p-values between 10−310^{-3} and 1), and for the co-volumes of the projections on b, plants and c, ants (pairs with non-zero co-volume; 10610^{6} samples per model). Dashed line: p-value = 0.05; dotted line: Benjamini–Hochberg threshold (FDR 5%) of the BiWCM. With the Benjamini–Hochberg procedure both the P-BiECM and the BiECM validate no link. The dashed line is shown only as a reference.

The different profiles highlight how the resolution of topological constraints changes link-level statistical validation. The strength-only BiWCM generally labels more interactions as anomalous (i.e. smaller p-values) because it does not reproduce empirical sparsity. Adding global density makes the benchmark more selective (i.e. greater p-values). Furthermore, adding node-specific degrees produces a further reduction where local connectivity is heterogeneous, as the profile of the p-value curve is lower for BiECM than P-BiECM. Across the empirical networks, P-BiECM and BiECM agree more closely with one another than either does with BiWCM, but their remaining disagreement is systematic rather than negligible.

Co-volumes validation

We finally apply statistical validation to empirical co-volumes resulting from one-mode projections of the bipartite systems. The choice to validate projected relationships through a null model defined on the bipartite system, rather than one estimated directly on the projected co-volume matrix, follows from two considerations.

First, the constraints that define the null hypothesis are quantities with a direct empirical meaning only at the bipartite level: they describe the activity of the original entities before projection, whereas co-volumes are derived quantities with no corresponding directly-observed constraint. A null model fitted directly on the projection would therefore include information from a derived object (i.e. the projection), instead of focusing on principled proxies for any generative mechanism of the underlying system.

Second, such a model offers no guarantee of preserving the topological and weighted structure of the bipartite network the projection originates from: because projection is a many-to-one, non-linear transformation of the biadjacency matrix, bipartite networks with markedly different structure (such as different degree heterogeneity, or a different resolution of topological information, as discussed above) can yield projections with similar co-volume distributions. A null model calibrated only on the projection would be blind to this upstream information, so the resulting validated relationships need not reflect the mechanisms that actually generate the observed pattern in the real bipartite system[33, 34]. Deriving the null distribution of co-volumes from the bipartite ensemble instead allows us to ask a sharper question: are the observed projected relationships expected once the bipartite generative mechanism has been properly accounted for?

The procedure is pictorially represented in panel b of Figure 4. Unlike the validation of the bipartite weights, due to the intricate dependencies, we could not find a closed-form formula for the calculation of the p-values of co-volumes. Therefore, they were calculated by sampling from the various ensembles.

As can be expected, the null distribution of the co-volume matrix and the resulting validated projection are strongly model-dependent. In panels b and c of Figure 5, the p-value profiles for the various models are presented for the projection on the plants and the ants layer, respectively (10610^{6} samples per model). Similarly to what was observed for the weight validation, ignoring the topology makes many observations surprising for BiWCM, as in the case of the weight validation above. Otherwise stated, the co-volume distribution predicted by the models including the topological information (i.e. P-BiECM and BiECM) is narrower and closer to the observed values. Differently for the weighted validation, nonetheless, the profiles of P-BiECM and BiECM are well resolved, stating that the difference in the resolution of the topology is relevant when studying the co-volume distribution. Similarities and differences in the co-volume validation can be found in Sections E, F, G of the Appendix.

Discussion

Across the systems studied here, strength-constrained ensembles fail to account for empirical support, showing that the allocation of positive and zero interactions carries additional information. The important distinction, however, is not simply between models with and without topology. It is between the resolutions at which topology becomes informative.

The hierarchy developed here turns this distinction into a model-selection problem. The P-BiECM asks whether one global sparsity constraint is sufficient beyond the two strength sequences; the BiECM asks whether the connectivity of each node must additionally be retained. In many empirical systems, global density is the most economical explanation of the departure from a strength-only baseline. In others, broad or poorly predicted degrees make local constraints indispensable for an adequate description. Thus, accounting for topology does not necessarily require node-specific constraints.

This conclusion refines earlier results for monopartite weighted networks, where degree information was shown to improve reconstruction beyond strengths[21], and density-assisted procedures were shown to recover much of an unknown topology from limited information[24, 39]. Our contribution is to formulate and compare these hypotheses in a common canonical framework for continuous weighted bipartite data. It complements the recent microcanonical algorithm that preserves bipartite degrees and strengths exactly[37]. Canonical and microcanonical approaches answer related but distinct questions: the former represents uncertainty around constraints satisfied on average, whereas the latter randomizes within an exactly constrained state space - the two approaches are not equivalent in the presence of local constraints[40, 41].

Our results also expose a practical consequence of model choice. Validation of link weights and co-volumes is conditional on the null ensemble, even when applied to the same empirical matrix. A model that does not reproduce sparsity tends to overstate the unexpectedness of observed quantities. Conversely, adding local degree constraints can remove relationships whose apparent significance follows from heterogeneous connectivity. Validated networks should therefore be reported together with, and justified by, an explicit selection of their underlying null model.

Several limitations remain. First, the empirical corpus contains different temporal realizations of the same dataset and uneven numbers of networks across domains. Domain-level conclusions should therefore be interpreted at the level of independent systems rather than raw matrix counts.

Second, the present hierarchy stops at degree constraints. Mesoscale organization, spatial costs, or covariates may contain additional information and can be incorporated as further constraints when justified by the application.

Overall, the proposed model hierarchy provides a quantitative notion of topological compressibility. If one density parameter leads to a performance comparable to that obtained using two complete degree sequences after accounting for complexity, most low-order topological information is reducible to global sparsity. If it does not, local connectivity is an unavoidable component of the statistical description.

Methods

Data

We analysed weighted bipartite matrices from economic, technological, ecological, social and lexical systems. Economic networks connecting countries to exported products are derived from UN-COMTRADE[42]. Technological networks connect countries with technological CPC classes from patent inventors’ geolocations obtained from the PATSTAT database[43]. Ecological networks encode observed mutualistic or trophic interactions, obtained from public interaction databases, including IWDB and ICON[44, 45]. Social networks connect users to online communities or information sources, built from Reddit and related online-interaction datasets[46, 47]. Lexical networks connect documents to processed word stems for UK companies[48]. Basic statistics, data sources, and all preprocessing operations are reported in Section B of the Appendix.

Maximum-entropy ensembles

General framework

Given a real network G∗G^{*} (in the following, all quantities denoted by an asterisk are measured on the empirical network), we first define G\mathcal{G}, the set of all possible graphs having the same number of nodes as in the empirical system, from the empty graph to the fully connected one. We equip each representative of this set with a probability obtained through a constrained maximization of the Shannon entropy, defined as

S=−∑G∈GP(G)ln⁡P(G).(3)S=-\sum_{G\in\mathcal{G}}P(G)\ln P(G). \tag*{(3)}

Let us call C⃗\vec{C} the vector of the constraints. The result of the constraint maximization of the Shannon entropy in Eq. (3) returns a probability for each representative of G\mathcal{G} that depends on the value of C⃗\vec{C} measured on GG:

P(G∣θ⃗)=exp⁡[−H(G,θ⃗)]Z(θ⃗),H(G,θ⃗)=θ⃗⋅C⃗(G)=∑mθmCm(G),P(G|\vec{\theta})=\frac{\exp[-\mathcal{H}(G,\vec{\theta})]}{Z(\vec{\theta})},\qquad\mathcal{H}(G,\vec{\theta})=\vec{\theta}\cdot\vec{C}(G)=\sum_m\theta_m C_m(G),

where ZZ is the partition function and θ⃗\vec{\theta} is the vector of Lagrange multipliers. Parameters are estimated by imposing that the average value over the ensemble of the constraints is equal to the value measured on the empirical network:

⟨C⃗⟩=C⃗(G∗).\langle\vec{C}\rangle=\vec{C}(G^{*}).

Remarkably, such a condition is equivalent to the maximization of the likelihood; such an equivalence is valid only for maximum entropy models[15].

As the constraints considered in the present paper are linear either in the binary biadjacency matrix B\mathbf{B} or in the weighted biadjacency matrix W\mathbf{W}, the resulting probability distributions per graph factorize over probability distributions per node pairs[14, 15].

Bipartite Weighted Configuration Model.

The Bipartite Weighted Configuration Model (BiWCM) constrains the strength sequences for each layer, i.e. {si}\{s_i\} and {σα}\{\sigma_\alpha\}. With corresponding multipliers {Θi}\{\Theta_i\} and {Hα}\{H_\alpha\}, the weight of link (i,α)(i,\alpha) follows the distribution

qiαBiWCM(w)=λiαe−λiαw,λiα=Θi+Hα,w≥0(4)q_{i\alpha}^{\mathrm{BiWCM}}(w)=\lambda_{i\alpha}e^{-\lambda_{i\alpha}w},\qquad\lambda_{i\alpha}=\Theta_i+H_\alpha,\qquad w\ge0 \tag*{(4)}

hence ⟨wiα⟩=1/λiα\langle w_{i\alpha}\rangle=1/\lambda_{i\alpha}. As the distribution is continuous, it has no probability mass at zero; therefore, the BiWCM controls weighted activity, but not the topology of the network.

Partial Bipartite Enhanced Configuration Model.

P-BiECM constrains {si}\{s_i\}, {σα}\{\sigma_\alpha\}, and LL. Let δ\delta be the multiplier associated with LL. The dyadic law is a hurdle distribution,

qiαP−BiECM(w)={1−piαw=0piαλiαe−λiαww>0(5)q_{i\alpha}^{\mathrm{P-BiECM}}(w)= \begin{cases} 1-p_{i\alpha} & w=0\\ p_{i\alpha}\lambda_{i\alpha}e^{-\lambda_{i\alpha}w} & w>0 \end{cases} \tag*{(5)}

where piαp_{i\alpha} is the probability of link existence,

piα=e−δλiα+e−δ,(6)p_{i\alpha}=\frac{e^{-\delta}}{\lambda_{i\alpha}+e^{-\delta}}, \tag*{(6)}

and λiα=Θi+Hα\lambda_{i\alpha}=\Theta_i+H_\alpha (while the relation between λiα\lambda_{i\alpha}, Θi\Theta_i and HαH_\alpha is analogous to the one in Eq. (4), the numerical values of the Lagrangian multipliers are different for the two models). Although topology has one independent multiplier, connection probabilities vary across dyads through the strength multipliers entering λiα\lambda_{i\alpha}.

Bipartite Enhanced Configuration Model.

BiECM constrains {si}\{s_i\}, {σα}\{\sigma_\alpha\}, {ki}\{k_i\}, and {hα}\{h_\alpha\}. Denoting by {θi}\{\theta_i\} and {ηα}\{\eta_\alpha\} the multipliers of the latter two constraints, the functional form of the weight distribution is analogous to the one in Eq. (5), i.e.

qiαBiECM(w)={1−piαw=0piαλiαe−λiαww>0(7)q_{i\alpha}^{\mathrm{BiECM}}(w)= \begin{cases} 1-p_{i\alpha} & w=0\\ p_{i\alpha}\lambda_{i\alpha}e^{-\lambda_{i\alpha}w} & w>0 \end{cases} \tag*{(7)}

but now the probability of link existence depends also on the multipliers for the degree sequences:

piα=e−(θi+ηα)λiα+e−(θi+ηα).(8)p_{i\alpha}=\frac{e^{-(\theta_i+\eta_\alpha)}}{\lambda_{i\alpha}+e^{-(\theta_i+\eta_\alpha)}}. \tag*{(8)}

Due to the functional form of Eq. (8), the topology and positive weights are inferred jointly because piαp_{i\alpha} depends on both topological and strength multipliers.

In fact, solving the parameters of the BiECM can be computationally demanding. Therefore, in our toolkit, we provide the quasi-Bipartite Enhanced Configuration Model (qBiECM), i.e. a method inspired by the BiECM, but introducing a simplification in the derivation that dramatically reduces the computation time of the parameters. More details can be found in Section H of the Appendix.

Weights validation

For each observed positive bipartite weight wiα∗w_{i\alpha}^{*}, we compute a right-tail p-value under the selected null ensemble. For the BiWCM this tail probability is

piαlink=exp⁡(−λiαwiα∗).(9)p_{i\alpha}^{\mathrm{link}}=\exp(-\lambda_{i\alpha}w_{i\alpha}^{*}). \tag*{(9)}

For the topology-aware weighted models, i.e. BiECM and P-BiECM, the probability also includes the chance that the edge exists,

piαlink=piαexp⁡(−λiαwiα∗).(10)p^{\mathrm{link}}_{i\alpha}=p_{i\alpha}\exp\left(-\lambda_{i\alpha}w^{*}_{i\alpha}\right). \tag*{(10)}

This distinction is central when testing sparse weighted matrices, because topology-aware models can assign probability to empirical zeros whereas the BiWCM cannot. Because we have a p-value for each pair (i,α)(i,\alpha), the p-values are corrected for multiple hypothesis testing. The package implements both the Bonferroni correction and False Discovery Rate control; all the results use the latter, with the Benjamini–Hochberg procedure at significance level αsig=0.05\alpha_{\mathrm{sig}}=0.05 (see the “Multiple-hypothesis testing” paragraph below).

Co-volumes validation

We used the same framework to validate one-mode projections. For each pair of nodes in the projected layer, we compute an empirical weighted co-volume Vij∗V^{*}_{ij} from their shared neighbours in the opposite layer. Since there is not a simple closed-form formula for the distribution of co-volumes, the null distribution of VijV_{ij} is obtained by sampling from the fitted bipartite ensemble. The projection p-value is

pijproj=P(Vijnull≥Vij∗).(11)p^{\mathrm{proj}}_{ij}=\mathbb{P}\left(V^{\mathrm{null}}_{ij}\geq V^{*}_{ij}\right). \tag*{(11)}

It is estimated as the fraction of the SS sampled networks whose co-volume is at least the observed one, p^ij=S−1∑s=1SI(Vij(s)≥Vij∗)\hat{p}_{ij}=S^{-1}\sum_{s=1}^{S}\mathbb{I}\left(V^{(s)}_{ij}\geq V^{*}_{ij}\right). As for the links, the p-values are corrected for multiple hypothesis testing with the Benjamini–Hochberg procedure at αsig=0.05\alpha_{\mathrm{sig}}=0.05. When the procedure validates no pair for at least one model, which happens for all the ecological networks, all the models of that network are compared with the same fixed threshold, p^ij<0.05\hat{p}_{ij}<0.05 (Section D of the Appendix).

Finite observation resolution and comparable likelihoods

In Section A of the Appendix, the interested reader can find the explicit form of the log-likelihood densities ℓmodel\ell_{\mathrm{model}}, calculated for each of our models. Here we remark a non-trivial detail necessary for the understanding of the model comparison.

When calculating BIC, see Eq. (2), the log-likelihood appears. In our models, we have probability densities with respect to a mixed continuous measure instead and cannot be compared directly with probabilities assigned to recorded observations. Thus, comparing all three models on a common footing requires converting the continuous density into an actual probability.

In fact, empirical weights are recorded with finite resolution Δw\Delta w. We therefore compare models using probabilities integrated over the observation bins. For instance, for BiWCM,

PBiWCM(wiα=w)≃qiαBiWCM(w)1−e−λiαΔwλiα.(12)\mathbb{P}^{\mathrm{BiWCM}}(w_{i\alpha}=w)\simeq q^{\mathrm{BiWCM}}_{i\alpha}(w)\frac{1-e^{-\lambda_{i\alpha}\Delta w}}{\lambda_{i\alpha}}. \tag*{(12)}

Thus, the discretized BiWCM is a geometric distribution over the recorded bins and assigns a non-zero probability to an observed zero through the interval [0,Δw)[0,\Delta w). For any of our topology-aware models, applying the same trick gives

P(wiα=0)=1−piαe−λiαΔw,P(wiα=w>0)=1−e−λiαΔwλiαpiαλiαe−λiαw.(13)\mathbb{P}(w_{i\alpha}=0)=1-p_{i\alpha}e^{-\lambda_{i\alpha}\Delta w}, \qquad\mathbb{P}(w_{i\alpha}=w>0)=\frac{1-e^{-\lambda_{i\alpha}\Delta w}}{\lambda_{i\alpha}}p_{i\alpha}\lambda_{i\alpha}e^{-\lambda_{i\alpha}w}. \tag*{(13)}

These corrections add a sum of terms to ℓ\ell, over every dyad for the purely continuous BiWCM, and only over the dyads where a link is actually observed (biα=1b_{i\alpha}=1) for the topology-aware

P-BiECM and BiECM:

ln⁡LBiWCM=ℓBiWCM+∑i,αln⁡(1−e−λiαΔwλiα)(14)\ln\mathcal{L}_{\mathrm{BiWCM}}=\ell_{\mathrm{BiWCM}}+\sum_{i,\alpha}\ln\left(\frac{1-e^{-\lambda_{i\alpha}\Delta w}}{\lambda_{i\alpha}}\right) \tag*{(14)}
ln⁡LP−BiECM=ℓP−BiECM+∑(i,α)∣biα=1ln⁡(1−e−λiαΔwλiα)+∑(i,α)∣biα=0ln⁡(1−piαe−λiαΔw1−piα)(15)\ln\mathcal{L}_{\mathrm{P-BiECM}}=\ell_{\mathrm{P-BiECM}}+\sum_{(i,\alpha)\mid b_{i\alpha}=1}\ln\left(\frac{1-e^{-\lambda_{i\alpha}\Delta w}}{\lambda_{i\alpha}}\right)+\sum_{(i,\alpha)\mid b_{i\alpha}=0}\ln\left(\frac{1-p_{i\alpha}e^{-\lambda_{i\alpha}\Delta w}}{1-p_{i\alpha}}\right) \tag*{(15)}
ln⁡LBiECM=ℓBiECM+∑(i,α)∣biα=1ln⁡(1−e−λiαΔwλiα)+∑(i,α)∣biα=0ln⁡(1−piαe−λiαΔw1−piα)(16)\ln\mathcal{L}_{\mathrm{BiECM}}=\ell_{\mathrm{BiECM}}+\sum_{(i,\alpha)\mid b_{i\alpha}=1}\ln\left(\frac{1-e^{-\lambda_{i\alpha}\Delta w}}{\lambda_{i\alpha}}\right)+\sum_{(i,\alpha)\mid b_{i\alpha}=0}\ln\left(\frac{1-p_{i\alpha}e^{-\lambda_{i\alpha}\Delta w}}{1-p_{i\alpha}}\right) \tag*{(16)}

The log-likelihoods and BIC values reported in the main text are computed with this exact, integrated correction throughout, so that likelihoods are always compared on the same footing across the hierarchy, including on the exact zeros where the shortcut would otherwise fail silently.

Multiple-hypothesis testing

When handling a great number of p-values, multiple-testing corrections are required to control false positives. Let p1,…,pDp_1,\ldots,p_D be the p-values of the DD tested hypotheses (piαlinkp_{i\alpha}^{\mathrm{link}} for the links, p^ij\hat{p}_{ij} for the projections) and αsig\alpha_{\mathrm{sig}} the significance level. Under Bonferroni correction [49], hypothesis rr is rejected if

pr≤αsigD.(17)p_r\leq\frac{\alpha_{\mathrm{sig}}}{D}. \tag*{(17)}

Under False Discovery Rate criterion [38], once the p-values are sorted in increasing order, p(1)≤p(2)≤⋯≤p(D)p_{(1)}\leq p_{(2)}\leq\cdots\leq p_{(D)}, one identifies the greatest rank r∗r^* satisfying

p(r∗)≤r∗αsigD(18)p_{(r^*)}\leq\frac{r^*\alpha_{\mathrm{sig}}}{D} \tag*{(18)}

and rejects all hypotheses with p-values smaller than or equal to tBH=r∗αsig/Dt_{\mathrm{BH}}=r^*\alpha_{\mathrm{sig}}/D.

Topology and structural reconstruction

Positive Predictive Value For the topology-aware models the connection probabilities piαp_{i\alpha} give the expected confusion matrix with respect to the observed support biα∗b_{i\alpha}^*: ⟨TP⟩=∑biα∗=1piα\langle\mathrm{TP}\rangle=\sum_{b_{i\alpha}^*=1}p_{i\alpha} and ⟨FP⟩=∑biα∗=0piα\langle\mathrm{FP}\rangle=\sum_{b_{i\alpha}^*=0}p_{i\alpha}. The positive predictive value PPV=⟨TP⟩/(⟨TP⟩+⟨FP⟩)\mathrm{PPV}=\langle\mathrm{TP}\rangle/(\langle\mathrm{TP}\rangle+\langle\mathrm{FP}\rangle) is the expected fraction of predicted links that are observed; it is an analytic expectation.

Assortativity measures The average nearest-neighbour degree and strength of node ii are kinn=ki−1∑αbiαhαk_i^{nn}=k_i^{-1}\sum_{\alpha}b_{i\alpha}h_{\alpha} and sinn=ki−1∑αbiασαs_i^{nn}=k_i^{-1}\sum_{\alpha}b_{i\alpha}\sigma_{\alpha}, with hαh_{\alpha} and σα\sigma_{\alpha} the degree and strength of node α\alpha of the other layer. Their ensemble mean ⟨xi⟩\langle x_i\rangle and standard deviation σxi\sigma_{x_i} are Monte Carlo estimates from networks sampled from each fitted model, and the agreement with the data is summarised by RMSEz=[N−1∑izi2]1/2\mathrm{RMSE}_z=[N^{-1}\sum_i z_i^2]^{1/2}, with zi=(xi∗−⟨xi⟩)/σxiz_i=(x_i^*-\langle x_i\rangle)/\sigma_{x_i}, which is close to one when the observable is reproduced within the model’s own fluctuations.

Nestedness Nestedness is measured by NODF [50], the average over pairs of nodes ordered by decreasing degree of the overlap ∣Γi∩Γj∣/kj|\Gamma_i\cap\Gamma_j|/k_j between their neighbourhoods, and by its weighted counterpart WNODA [51], in which nodes are ordered by strength and the overlap counts the entries of jj that are smaller than those of ii.

Clustering Binary clustering is the Robins–Alexander coefficient C=4C4/L3C = 4C_4/L_3, the fraction of paths of length three closed into 4-cycles, computed with NetworkX[52]; weighted clustering is the two-mode coefficient of Opsahl[53], the weighted fraction of closed 4-paths, with each path weighted by the geometric mean of its weights.

For each of these observables, the observed value is compared with its sampled distribution through the z-score and the percentile of the observed value. Exact definitions, numbers of samples, and results are given in Section D of the Appendix.

Pseudocode and numerical implementation

The complete workflow, summarized in Algorithm 1, is implemented in an open-source Python package built on PyTorch, exposing a common interface for the three null models: fitting, sampling, and validation share the same routines across BiWCM, P-BiECM, and BiECM. Because the empirical analysis above requires fitting several competing models on each of many networks and drawing large sampling ensembles (SS from 10310^3 to 10610^6 networks per model, see Section D of the Appendix) for both reconstruction and projection validation, computational efficiency is a design requirement, not an afterthought, in the spirit of recent scalable likelihood-maximization solvers for constrained network ensembles[30]. All solvers are fully vectorized over dyads, can be chunked row-wise to bound memory use on large matrices, and are moved to GPU when explicitly requested via a device argument.

Parameter fitting solves the moment-matching equations underlying ⟨C⃗⟩=C⃗(G∗)\langle\vec{C} \rangle= \vec{C}(G^*). The default solver is a fixed-point iteration with Polyak-style momentum (step size γ=1\gamma= 1, momentum β=0.9\beta= 0.9 by default), used whenever degree and strength constraints are solved jointly (P-BiECM, BiECM); momentum can optionally be annealed toward β=0.999\beta= 0.999 over a warm-up window to stabilize early iterations without slowing late-stage convergence. The single-constraint BiWCM converges without momentum. A diagonal quasi-Newton update, with a model-specific step size tuned for stability, and an exact bisection scheme (60 steps in log-space, ∼10−18\sim10^{-18} precision) are provided as alternatives, the latter offering guaranteed monotonic convergence when the fixed-point iteration is unstable. Convergence is assessed on the maximum relative error between empirical and expected constraints: in all the analyses of this work a solution is accepted when this error is below ϵ=10−3\epsilon= 10^{-3} (the package default, ϵ=10−8\epsilon= 10^{-8} with at most 10410^4 iterations, is stricter and can be set by the user). Further details can be found in § 23 of the Appendix.

Algorithm 1: WBNM fitting–sampling–validation pipeline

Input: WW; model M∈{BiWCM,P-BiECM,BiECM,qBiECM}\mathcal{M} \in\{\mathrm{BiWCM},\mathrm{P\text{-}BiECM},\mathrm{BiECM},\mathrm{qBiECM}\} (Table 1); tolerance ε\varepsilon; number of samples SS; significance level αsig\alpha_{\mathrm{sig}}

Output: Fitted multipliers, BIC, validated bipartite links, validated projection

// 1. Multiplier resolution

C∗←C^{*} \leftarrow empirical constraints of WW (Eq. (1))

initialize {Θ,H,θ,η}(0)\{\Theta,H,\theta,\eta\}^{(0)}; v←0v \leftarrow0; t←0t \leftarrow0

repeat

Δ(t)←⟨C⟩(t)/C∗−1\Delta^{(t)} \leftarrow\langle C\rangle^{(t)}/C^{*}-1 // vectorized, chunked dyads (CPU/GPU)

v←βtv+(1−βt)Δ(t)v \leftarrow\beta_t v+(1-\beta_t)\Delta^{(t)}

{Θ,H,θ,η}(t+1)←{Θ,H,θ,η}(t)⊙(1+γv)\{\Theta,H,\theta,\eta\}^{(t+1)} \leftarrow\{\Theta,H,\theta,\eta\}^{(t)}\odot(1+\gamma v) // or quasi-Newton / bisection

t←t+1t \leftarrow t+1

until max⁡∣Δ(t−1)∣<ε\max|\Delta^{(t-1)}|<\varepsilon or t=tmax⁡t=t_{\max}

// 2. Model selection and link validation

L,BIC←Likelihood⁡(W,{Θ,H,θ,η})\mathcal{L},\mathrm{BIC}\leftarrow\operatorname{Likelihood}(W,\{\Theta,H,\theta,\eta\}) // Eq. (2)

piαlink←PValue⁡(wiα∗,{Θ,H,θ,η})p_{i\alpha}^{\mathrm{link}}\leftarrow\operatorname{PValue}(w_{i\alpha}^{*},\{\Theta,H,\theta,\eta\}) for every observed link // Eqs. (9), (10)

// 3. Sampling: reconstruction and projection validation

for s=1s=1 to SS do

W(s)←Sample⁡(M,{Θ,H,θ,η}); V(s)←V(W(s))W^{(s)}\leftarrow\operatorname{Sample}(\mathcal{M},\{\Theta,H,\theta,\eta\});\ V^{(s)}\leftarrow V(W^{(s)})

end

p^ij←S−1∑sI(Vij(s)≥Vij∗)\hat{p}_{ij}\leftarrow S^{-1}\sum_s\mathbb{I}(V_{ij}^{(s)}\ge V_{ij}^{*}) for every pair // estimate of Eq. (11)

// 4. Multiple-testing correction

return links and pairs selected by Benjamini–Hochberg at αsig\alpha_{\mathrm{sig}} (fixed threshold 0.05 when BH selects none) // Eq. (18)

Data availability

All datasets are publicly available from the repositories and original sources listed in Table S1 of the Appendix.

Code availability

The WBNM package implementing fitting, sampling, model selection and validation is available at wbnm.

Acknowledgements

Author contributions

Daniele Cirulli: Software, Formal analysis, Investigation, Data Curation, Validation, Visualization, Writing - Original Draft; Lorenzo Buffa: Software, Formal analysis, Investigation, Data Curation, Validation, Visualization, Writing - Original Draft; Giulio Cimini: Conceptualization, Supervision, Writing - Review & Editing; Dario Mazzilli: Conceptualization, Data Curation, Supervision, Writing - Review & Editing; Aurelio Patelli: Conceptualization, Methodology, Data Curation, Supervision, Writing - Review & Editing; Fabio Saracco: Conceptualization, Methodology, Data Curation, Supervision, Writing - Review & Editing, Project administration.

Competing interests

The authors declare no competing interests.

Model derivations and likelihoods

Dyadic distributions

All models are maximum-entropy ensembles whose constraints are sums of dyadic quantities, so that the probability of a biadjacency matrix factorises over the dyads (i,α)(i,\alpha). With λiα=Θi+Hα\lambda_{i\alpha}=\Theta_i+H_\alpha the multipliers of the strengths, the weighted models have

qiαBiWCM(w)=λiαe−λiαw,w≥0,(19)q_{i\alpha}^{\mathrm{BiWCM}}(w)=\lambda_{i\alpha}e^{-\lambda_{i\alpha}w},\qquad w\geq0, \tag*{(19)}
qiαhurdle(w)={1−piαw=0piαλiαe−λiαww>0(20)q_{i\alpha}^{\mathrm{hurdle}}(w)= \begin{cases} 1-p_{i\alpha} & w=0\\ p_{i\alpha}\lambda_{i\alpha}e^{-\lambda_{i\alpha}w} & w>0 \end{cases} \tag*{(20)}

where the connection probability of the hurdle (topology-aware) models is

piαP-BiECM=e−δλiα+e−δ,piαBiECM=e−(θi+ηα)λiα+e−(θi+ηα),piαqBiECM=e−(θi+ηα)1+e−(θi+ηα).(21)p_{i\alpha}^{\mathrm{P\text{-}BiECM}}=\frac{e^{-\delta}}{\lambda_{i\alpha}+e^{-\delta}},\qquad p_{i\alpha}^{\mathrm{BiECM}}=\frac{e^{-(\theta_i+\eta_\alpha)}}{\lambda_{i\alpha}+e^{-(\theta_i+\eta_\alpha)}},\qquad p_{i\alpha}^{\mathrm{qBiECM}}=\frac{e^{-(\theta_i+\eta_\alpha)}}{1+e^{-(\theta_i+\eta_\alpha)}}. \tag*{(21)}

Here δ\delta is the multiplier of the total number of links LL, and θi,ηα\theta_i,\eta_\alpha those of the degrees ki,hαk_i,h_\alpha. The BiCM is the purely binary model with piαBiCM=piαqBiECMp_{i\alpha}^{\mathrm{BiCM}}=p_{i\alpha}^{\mathrm{qBiECM}}. The qBiECM is discussed in Section H.

Density log-likelihoods

Summing ln⁡qiα(wiα∗)\ln q_{i\alpha}(w_{i\alpha}^{*}) over all dyads at the maximum-likelihood solution gives, for the BiCM,

ln⁡LBiCM=−∑iθiki∗−∑αηαhα∗−∑i,αln⁡(1+e−(θi+ηα)),(22)\ln\mathcal{L}_{\mathrm{BiCM}}=-\sum_i\theta_i k_i^{*}-\sum_\alpha\eta_\alpha h_\alpha^{*}-\sum_{i,\alpha}\ln\left(1+e^{-(\theta_i+\eta_\alpha)}\right), \tag*{(22)}

which is already the log-probability of the observed binary matrix. For the weighted models the same sum is a log-density, which we denote by ℓ\ell:

ℓBiWCM=−∑iΘisi∗−∑αHασα∗+∑i,αln⁡λiα,(23)\ell_{\mathrm{BiWCM}}=-\sum_i\Theta_i s_i^{*}-\sum_\alpha H_\alpha\sigma_\alpha^{*}+\sum_{i,\alpha}\ln\lambda_{i\alpha}, \tag*{(23)}
ℓP-BiECM=−δL∗−∑iΘisi∗−∑αHασα∗−∑i,αln⁡(1+e−δλiα),(24)\ell_{\mathrm{P\text{-}BiECM}}=-\delta L^{*}-\sum_i\Theta_i s_i^{*}-\sum_\alpha H_\alpha\sigma_\alpha^{*}-\sum_{i,\alpha}\ln\left(1+\frac{e^{-\delta}}{\lambda_{i\alpha}}\right), \tag*{(24)}
ℓBiECM=−∑iθiki∗−∑αηαhα∗−∑iΘisi∗−∑αHασα∗−∑i,αln⁡(1+e−(θi+ηα)λiα).(25)\ell_{\mathrm{BiECM}}=-\sum_i\theta_i k_i^{*}-\sum_\alpha\eta_\alpha h_\alpha^{*}-\sum_i\Theta_i s_i^{*}-\sum_\alpha H_\alpha\sigma_\alpha^{*}-\sum_{i,\alpha}\ln\left(1+\frac{e^{-(\theta_i+\eta_\alpha)}}{\lambda_{i\alpha}}\right). \tag*{(25)}

The BiCM depends only on the degree multipliers, the BiWCM only on the strength multipliers, the P-BiECM combines the strength term with the single global multiplier δ\delta, and the BiECM combines strength and degree terms in full.

Likelihood of the recorded data

The densities above are dimensional quantities and cannot be compared with probabilities. We therefore compare models through the probability of the recorded values, obtained by integrating each continuous component over the observation bin [w∗,w∗+Δw)[w^{*},w^{*}+\Delta w):

P{w∗≤W<w∗+Δw}=e−λw∗(1−e−λΔw)=q(w∗)1−e−λΔwλ.(26)\mathbb{P}\{w^{*}\leq W<w^{*}+\Delta w\}=e^{-\lambda w^{*}}\left(1-e^{-\lambda\Delta w}\right)=q(w^{*})\frac{1-e^{-\lambda\Delta w}}{\lambda}. \tag*{(26)}

The approximation q(w∗)Δwq(w^{*})\Delta w of Ref.[54] is recovered for λΔw≪1\lambda\Delta w\ll1, but it is not a probability in general: for an empirical zero of the BiWCM it equals λΔw\lambda\Delta w, which exceeds one whenever the expected weight 1/λ1/\lambda is smaller than Δw\Delta w, a frequent case at Δw=1\Delta w = 1 in sparse networks. With Eq. (12), the log-likelihoods used in all tables and figures are

ln⁡LBiWCM=ℓBiWCM+∑i,αln⁡1−e−λiαΔwλiα(27)\ln\mathcal{L}_{\mathrm{BiWCM}} = \ell_{\mathrm{BiWCM}} + \sum_{i,\alpha} \ln\frac{1-e^{-\lambda_{i\alpha}\Delta w}}{\lambda_{i\alpha}} \tag*{(27)}
ln⁡LP−BiECM=ℓP−BiECM+∑(i,α):biα∗=1ln⁡1−e−λiαΔwλiα+∑(i,α):biα∗=0ln⁡1−piαe−λiαΔw1−piα(28)\ln\mathcal{L}_{\mathrm{P-BiECM}} = \ell_{\mathrm{P-BiECM}} + \sum_{(i,\alpha):b^*_{i\alpha}=1} \ln\frac{1-e^{-\lambda_{i\alpha}\Delta w}}{\lambda_{i\alpha}} + \sum_{(i,\alpha):b^*_{i\alpha}=0} \ln\frac{1-p_{i\alpha}e^{-\lambda_{i\alpha}\Delta w}}{1-p_{i\alpha}} \tag*{(28)}
ln⁡LBiECM=ℓBiECM+∑(i,α):biα∗=1ln⁡1−e−λiαΔwλiα+∑(i,α):biα∗=0ln⁡1−piαe−λiαΔw1−piα(29)\ln\mathcal{L}_{\mathrm{BiECM}} = \ell_{\mathrm{BiECM}} + \sum_{(i,\alpha):b^*_{i\alpha}=1} \ln\frac{1-e^{-\lambda_{i\alpha}\Delta w}}{\lambda_{i\alpha}} + \sum_{(i,\alpha):b^*_{i\alpha}=0} \ln\frac{1-p_{i\alpha}e^{-\lambda_{i\alpha}\Delta w}}{1-p_{i\alpha}} \tag*{(29)}

For the BiWCM the result is a geometric distribution over the bins, and an empirical zero has probability 1−e−λiαΔw1-e^{-\lambda_{i\alpha}\Delta w}. For the hurdle models a recorded zero is either a structural zero or a link with weight below Δw\Delta w, so that it has probability (1−piα)+piα(1−e−λiαΔw)=1−piαe−λiαΔw(1-p_{i\alpha})+p_{i\alpha}(1-e^{-\lambda_{i\alpha}\Delta w})=1-p_{i\alpha}e^{-\lambda_{i\alpha}\Delta w}, while a recorded link of weight w∗w^* has probability piαe−λiαw∗(1−e−λiαΔw)p_{i\alpha}e^{-\lambda_{i\alpha}w^*}(1-e^{-\lambda_{i\alpha}\Delta w}). Since ℓ\ell already contains ln⁡(1−piα)\ln(1-p_{i\alpha}) for each empty dyad, the last sum replaces it with ln⁡(1−piαe−λiαΔw)\ln(1-p_{i\alpha}e^{-\lambda_{i\alpha}\Delta w}). With these corrections the probabilities of all recorded values sum to one in every model, and a recorded zero is treated in the same way, as the bin [0,Δw)[0,\Delta w), across the hierarchy.

The multipliers are the maximum-likelihood estimates of the continuous model, and the likelihood of the recorded data is evaluated at that solution. When λiαΔw≪1\lambda_{i\alpha}\Delta w \ll1 the two likelihoods differ by a constant and share the same maximum; otherwise the values reported are lower bounds of the maximised likelihoods of the recorded data, for every model alike. The smallest BIC gap between the best and the second-best model is 88 (Table E.1), so this does not affect model selection.

Information criteria

With n=N1N2n=N_1N_2 dyads and kk free parameters, BIC =kln⁡n−2ln⁡L=k\ln n-2\ln\mathcal{L} and AIC =2k−2ln⁡L=2k-2\ln\mathcal{L}, with k=N1+N2k=N_1+N_2 for the BiWCM (and the BiCM), N1+N2+1N_1+N_2+1 for the P-BiECM and 2(N1+N2)2(N_1+N_2) for the BiECM and the qBiECM. Gauge freedoms of the multipliers (one for each pair of multipliers entering only through a sum) are not subtracted; they shift kk by one or two units, which is negligible compared with the BIC gaps in Table E.1.

Dataset sources and preprocessing

We analyse 46 weighted bipartite matrices from seven systems (Table B.1): international trade, Reddit, a lexical corpus, Twitter, two families of technological matrices and ecological interaction networks. Repeated temporal snapshots are grouped under their parent system in every table. Basic statistics of each matrix are reported in Table B.2. Rows and columns with zero strength are removed before any analysis.

Trade. Country–product export matrices from UN Comtrade for the years 2012–2020 (biennial observations) [42]. The data cover 220 countries and 5204 products classified at the 6-digit level of the Harmonized System 2012 (HS12). Each weight represents the export value of a country for a given product, expressed in constant 2015 US dollars.

Reddit. Yearly matrices for 2013–2017, built from the Pushshift dumps [46, 47]. Rows are subreddits, which are the nodes of the validated projections, and columns are news domains: a subreddit and a domain are linked if the domain was shared in at least one post of that subreddit. The weight of a link is the number of times the domain was shared in the subreddit.

Lexical corpus. The lexical corpus was downloaded from Ref. [48]; it includes 599 sustainability or non-financial reports published from 2015 to 2023 by the 10 UK companies with the highest Company Market Cap listed on the London Stock Exchange. The preprocessing and the extraction of a weighted bipartite network between tokens and documents, in which weights encode the frequency of the various tokens in the different documents, follows a recipe similar to the one in Ref. [55]. All texts under analysis were processed with the Natural Language Processing tools made available by the nltk Python library [56]. As a preprocessing step, individual tokens were extracted by stripping punctuation and purely numeric entries, while alphanumeric tokens (i.e. tokens containing both numbers and words) were retained; stop words were removed as well. Stop-word filtering relied on the English list provided in the nltk.corpus submodule [56], which covers pronouns, demonstrative adjectives, prepositions, auxiliary verbs, and comparable function words. The resulting tokens were then reduced to their root form using the Snowball stemmer [56]. Stemming strips prefixes and suffixes from a word, collapsing its inflected or derived forms onto a common stem, so that occurrences of the same underlying concept can be counted consistently. Finally, the number of occurrences of each stem in each document was counted, providing a weighted biadjacency matrix in which rows are documents and columns are stems. Five yearly snapshots (2015, 2018, 2020, 2021, 2023) are analysed.

Twitter/X. We used tweets concerning the three most prominent political events debated on Italian Twitter/X in 2022, the same events analysed in Ref. [57]. Since, at the time of the download of the data, the platform name was still Twitter, this will be the name of the dataset in the following. For each event, a set of pertinent keywords was defined and used to retrieve tweets via Twitter’s Academic API. The three datasets comprise:

  • the debate around the election of the President of the Republic (ended with the re-election of Sergio Mattarella on January 29, 2022).

  • the debate around the political crisis that led to the fall of the Draghi government on July 21, 2022.

  • the debate around the Italian general elections held on September 25, 2022, which led to Giorgia Meloni becoming Prime Minister.

We emphasize that the data were collected prior to the introduction of user verification as part of X’s Premium subscription (https://en.wikipedia.org/wiki/Twitter_verification).

For each dataset, we extracted from the collected tweets, following Refs. [57, 58], a bipartite weighted network, whose two layers are the verified users and the standard (i.e., non-verified) users. The weight of the link between a verified user vv and a standard user ss is a non-negative integer: the number of times ss reposted messages originally authored by vv. Such an analysis can effectively extract users’ political leanings; see Ref. [57, 58, 59].

Technology.

  • Country–technology patent matrices (CPC) for five biennial snapshots from 2010 to 2018. The data cover 47 major patent-producing countries and approximately 640 technology classes defined by the 4-digit Cooperative Patent Classification (CPC), restricted to sections A–H (excluding Y). Not all classes are active in every year. Weights represent fractional patent counts.

  • Country-level scientific citation matrices (CIT) for five biennial snapshots from 2010 to 2018, taken from [60]. The data cover approximately 200 countries, with minor variations across years due to low publication activity in smaller countries. Countries are identified through authors’ affiliations, and weights represent citations received by scientific papers, fractionally

SystemLayer 1 (rows, projected)Layer 2 (columns)WeightΔw\Delta wSnapshots
TradeCountriesProductsExport value10−1210^{-12}2012–2020 (5)
RedditSubredditsNews domainsTimes the domain is shared12013–2017 (5)
LexicalDocumentsWord stemsStem occurrences12015–2023 (5)
TwitterVerified usersNon-verified usersRetweets (counts)13 events
Tech CITCountriesTechnology classesCitations (counts)10−1210^{-12}2010–2018 (5)
Tech CPCCountriesCPC classesFractional patent counts10−1210^{-12}2010–2018 (5)
EcologicalPlants / hostsAnimals / parasitesVisits, counts, intensity or prevalence1 or 0.121 matrices

Table B.1. Empirical systems. Meaning of the two layers and of the weights, observation resolution Δw\Delta w used in the likelihoods (the largest power of ten such that every recorded weight is an integer multiple of it, capped at 10−1210^{-12} for monetary and fractional data), and number of snapshots. Projections are always computed on layer 1. Sources and preprocessing are described in the text.

adjusted for the number of authors. The scientific classification accounts for 312 distinct scientific domains. The dataset is based on the Microsoft Academic Graph (MAG), whose legacy is now maintained within the OpenAlex project.

Ecological networks. Mutualistic and host–parasite interaction matrices from IWDB and ICON[44, 45]. Weights are counts of visits or interactions (integers, Δw=1\Delta w = 1) or, for the host–parasite matrices of the Canadian lakes, parasite intensity and prevalence recorded with one decimal (Δw=0.1\Delta w = 0.1). Some matrices of the source collection coincide: Beehler 1 and Beehler 2 are identical, Snow 1 and Seed disp. Snow are the same matrix up to a permutation of rows and columns, and the intensity and prevalence matrices of McGregor lake are identical. For the other lakes, intensity and prevalence share the same support but have different weights. We keep a single copy of each duplicated matrix (Beehler, Seed disp. Snow and the McGregor intensity matrix), so that the corpus contains 46 distinct matrices.

Observation resolution. All likelihoods are evaluated on the recorded values (Section A). For each matrix, the resolution Δw\Delta w is the largest power of ten such that every recorded weight is an integer multiple of it. Count data have Δw=1\Delta w = 1; ecological intensities and prevalences have Δw=0.1\Delta w = 0.1; trade values and fractional technology counts show no finite decimal resolution and are assigned Δw=10−12\Delta w = 10^{-12}.

NetworkSystemN1N_1N2N_2LLdd⟨k⟩\langle k\rangle⟨w⟩\langle w\rangleρks\rho_{ks}YYΔw\Delta w
Trade 2012Trade22052046.14×1056.14 \times10^50.5372275.46×1075.46 \times10^70.6350.18710−1210^{-12}
Trade 2014Trade22052045.26×1055.26 \times10^50.461943.52×1073.52 \times10^70.5950.18910−1210^{-12}
Trade 2016Trade22052045.49×1055.49 \times10^50.4792023.16×1073.16 \times10^70.6140.19110−1210^{-12}
Trade 2018Trade22052045.61×1055.61 \times10^50.492073.76×1073.76 \times10^70.5970.18910−1210^{-12}
Trade 2020Trade22052045.46×1055.46 \times10^50.4762013.54×1073.54 \times10^70.6220.1910−1210^{-12}
Reddit 2013Reddit2166.88×1046.88 \times10^41.59×1051.59 \times10^50.01074.618.980.8610.781
Reddit 2014Reddit2484.56×1044.56 \times10^41.32×1051.32 \times10^50.01165.748.620.880.7581
Reddit 2015Reddit2844.54×1044.54 \times10^41.48×1051.48 \times10^50.01146.469.270.8860.7351
Reddit 2016Reddit3996.78×1046.78 \times10^42.31×1052.31 \times10^58.52×10−38.52 \times10^{-3}6.7616.60.8680.7631
Reddit 2017Reddit4616.59×1046.59 \times10^42.28×1052.28 \times10^57.52×10−37.52 \times10^{-3}6.8831.30.8540.781
Lexical 2015Lexical401.65×1041.65 \times10^47.21×1047.21 \times10^40.1098.735.90.8660.751
Lexical 2018Lexical562.21×1042.21 \times10^41.06×1051.06 \times10^50.08589.596.160.8620.7681
Lexical 2020Lexical692.74×1042.74 \times10^41.43×1051.43 \times10^50.075510.46.850.8530.7781
Lexical 2021Lexical762.99×1042.99 \times10^41.52×1051.52 \times10^50.066910.16.820.8450.7881
Lexical 2023Lexical893.54×1043.54 \times10^41.83×1051.83 \times10^50.058210.37.420.8390.7961
Twitter Q22Twitter13616.56×1046.56 \times10^41.45×1051.45 \times10^51.62×10−31.62 \times10^{-3}4.321.440.9420.7671
Twitter electionTwitter29761.32×1051.32 \times10^54.60×1054.60 \times10^51.17×10−31.17 \times10^{-3}6.821.990.9480.6881
Twitter crisisTwitter66673.22×1053.22 \times10^58.85×1058.85 \times10^54.13×10−44.13 \times10^{-4}5.391.470.9440.7451
Tech CIT 2010Tech CIT2003123.87×1043.87 \times10^40.6215117950.6250.11110−1210^{-12}
Tech CIT 2012Tech CIT1993124.00×1044.00 \times10^40.64515714410.6120.10410−1210^{-12}
Tech CIT 2014Tech CIT2003124.03×1044.03 \times10^40.64615810410.6270.098110−1210^{-12}
Tech CIT 2016Tech CIT2043124.00×1044.00 \times10^40.6291555390.6390.098210−1210^{-12}
Tech CIT 2018Tech CIT1903123.29×1043.29 \times10^40.55513167.60.7010.1110−1210^{-12}
Tech CPC 2010Tech CPC476501.23×1041.23 \times10^40.40235.310.80.8820.20610−1210^{-12}
Tech CPC 2012Tech CPC476481.30×1041.30 \times10^40.42637.310.70.8870.20210−1210^{-12}
Tech CPC 2014Tech CPC476481.33×1041.33 \times10^40.43638.210.90.890.19710−1210^{-12}
Tech CPC 2016Tech CPC476531.35×1041.35 \times10^40.43938.510.50.8950.19310−1210^{-12}
Tech CPC 2018Tech CPC4662072160.25321.74.540.9030.27610−1210^{-12}
BeehlerEcological1118380.1922.623.530.7550.6781
Seed disp. BeehlerEcological9311190.4275.959.990.9020.4931
Seed disp. SnowEcological19292110.3838.7994.50.830.3971
Plant–pollinator BHEcological102121670.1362.933.290.8290.7861
Plants–antsEcological41512850.1366.22.260.9750.4241
Aishihik int.Ecological729780.3844.3343.60.740.670.1
Aishihik prev.Ecological729780.3844.33460.8370.5920.1
CanadaEcological1038870.2293.6218.90.7530.6731
Cold Lake int.Ecological1040910.2283.6418.40.7440.6780.1
Cold Lake prev.Ecological1040910.2283.6424.90.7630.7190.1
Huron prev.Ecological33973160.09874.8625.50.790.5850.1
Lake of the WoodsEcological311464460.09855.0429.70.8360.6150.1
Lake of the Woods prev.Ecological311443840.0864.3931.60.8240.6270.1
McGregora^{a}Ecological14511140.163.5111.50.7740.7010.1
Parsnip int.Ecological17531580.1754.5112.70.7760.5780.1
Parsnip prev.Ecological17531580.1754.519.840.7810.5860.1
Smallwood int.Ecological625530.3533.4217.80.6970.6871
Smallwood prev.Ecological625530.3533.4224.40.6830.6391

Table B.2. Basic statistics of the empirical bipartite networks. Number of nodes in the two layers (N1,N2N_1,N_2), number of links LL, density d=L/(N1N2)d=L/(N_1N_2), average degree ⟨k⟩=2L/(N1+N2)\langle k\rangle=2L/(N_1+N_2), average positive weight ⟨w⟩\langle w\rangle, Spearman correlation between strengths and degrees ρks\rho_{ks}, weight disparity YY (Methods) and observation resolution Δw\Delta w. a^{a} Intensity matrix; the prevalence matrix in the source data is identical to it and is not included.

Structural drivers: which network properties favour the global-density model?

The P-BiECM and the BiECM differ only in the topological constraint: one global density against the full degree sequences. To identify the network properties that make the local constraints necessary, we compare the two models on the real networks and on an Erdős–Rényi (ER) family in which the degrees carry no information beyond the density (Algorithm G.1, Section G).

Figure C.1 shows the ratios ℓP-BiECM/ℓBiECM\ell_{\mathrm{P\text{-}BiECM}}/\ell_{\mathrm{BiECM}} of the log-likelihoods (both negative, so a ratio larger than one means that the P-BiECM fits worse) and BICP-BiECM/BICBiECM\mathrm{BIC}_{\mathrm{P\text{-}BiECM}}/\mathrm{BIC}_{\mathrm{BiECM}} (smaller than one when the P-BiECM is selected).

In the ER family, the likelihood ratio is always above one, since the BiECM has more parameters, but it approaches one as the density grows and the weight disparity decreases (ρs=−1.00\rho_s=-1.00 and +1.00+1.00), and the BIC ratio stays below one for every realisation: BIC recovers the generating global-density structure across the whole range of density.

Among the real networks, the same monotonic trends hold for density (ρs=−0.89\rho_s=-0.89 for the likelihood ratio and +0.83+0.83 for the BIC ratio) and weight disparity (+0.90+0.90 and −0.85-0.85), while the strength–degree correlation is a weaker predictor (+0.49+0.49 and −0.57-0.57) and the degree variance is not informative (+0.05+0.05 and +0.18+0.18). Nonetheless, the ER family reproduces the sign of most of the trends observed in real networks (panels a, b, d, e, f, h), but not their magnitude: ER ensembles span likelihood and BIC ratios within about 10−210^{-2}, whereas empirical networks reach up to 0.4, and occupy a region of the observables’ space that the ER family does not cover. The only observables showing a sizeable monotonic association with model preference in the empirical data are the density and the weight disparity (Spearman ∣ρs∣≈0.8|\rho_s| \approx0.8–0.90.9). However, the scatter around these trends is large (e.g. panel h, where at high disparity the BIC ratio spans roughly 0.77–1.0), so none of them yields a reliable a priori criterion. BIC-based selection therefore remains necessary.

Figure C.1. Global versus local topological constraints in synthetic and real networks. a–d, Ratio between the log-likelihoods of the P-BiECM and of the BiECM and e–h, ratio between their BIC values, against the strength–degree Spearman correlation (a,e), the density (b,f), the degree variance (c,g) and the weight disparity (d,h). Blue: ER networks at fixed total weight (Algorithm G.1; 30 densities, 3 realisations each). Grey: the empirical networks of Table B.2. The boxes report the Spearman correlation for the ER family and, in parentheses, for the real networks. The vertical axis of a–d is broken to show both groups.

Reconstruction and validation definitions

Mesoscale observables

With biα=I(wiα>0)b_{i\alpha} = \mathbb{I}(w_{i\alpha} > 0), degrees ki=∑αbiαk_i = \sum_{\alpha} b_{i\alpha}, hα=∑ibiαh_\alpha= \sum_i b_{i\alpha} and strengths si=∑αwiαs_i = \sum_\alpha w_{i\alpha}, σα=∑iwiα\sigma_\alpha= \sum_i w_{i\alpha}, the average nearest-neighbour degree and strength of a node of layer 1 are

kinn=1ki∑αbiαhα,sinn=1ki∑αbiασα,(30)k_i^{nn} = \frac{1}{k_i}\sum_\alpha b_{i\alpha}h_\alpha,\qquad s_i^{nn} = \frac{1}{k_i}\sum_\alpha b_{i\alpha}\sigma_\alpha, \tag*{(30)}

set to zero when ki=0k_i = 0, and analogously for layer 2. Note that sinns_i^{nn} is the average strength of the neighbours of ii, not the weight-averaged nearest-neighbour degree.

Reconstruction errors (Table E.2). For each network and model we sample S=300S = 300 networks from the fitted ensemble and compute xi(s)x_i^{(s)}, x∈{knn,snn}x \in\{k^{nn},s^{nn}\}, for every node ii of a layer and every sample ss. The ensemble mean and standard deviation of node ii are

⟨xi⟩=1S∑s=1Sxi(s),σxi=[1S∑s=1S(xi(s)−⟨xi⟩)2]1/2,(31)\langle x_i\rangle= \frac{1}{S}\sum_{s=1}^{S}x_i^{(s)},\qquad\sigma_{x_i} = \left[\frac{1}{S}\sum_{s=1}^{S}\left(x_i^{(s)}-\langle x_i\rangle\right)^2\right]^{1/2}, \tag*{(31)}

and the deviation of the observed value is measured by the node z-score zi=(xi∗−⟨xi⟩)/σxiz_i = (x_i^*-\langle x_i\rangle)/\sigma_{x_i}, with σxi\sigma_{x_i} floored at 10−1210^{-12}. The table reports the root mean square of the z-scores over the NN nodes of the layer,

RMSEz=[1N∑i=1Nzi2]1/2.(32)\mathrm{RMSE}_z = \left[\frac{1}{N}\sum_{i=1}^{N} z_i^2\right]^{1/2}. \tag*{(32)}

The root mean square is used instead of the average of ziz_i because deviations of opposite sign would otherwise cancel. If the model is correct, each ziz_i has zero mean and unit variance and RMSEz≃1\mathrm{RMSE}_z \simeq1; values much larger than one indicate deviations that the fluctuations of the model cannot account for. The unscaled error RMSE=[N−1∑i(xi∗−⟨xi⟩)2]1/2\mathrm{RMSE} = [N^{-1}\sum_i(x_i^*-\langle x_i\rangle)^2]^{1/2} is computed as well and released with the parameter files; it is not comparable across networks because it carries the units of xx. For the P-BiECM and the BiECM on the Twitter matrices the 300 samples are the first 300 samples of the projection Monte Carlo (Section D, validation of projections). Under the BiWCM every dyad is positive with probability one, so kinnk_i^{nn} is the same in every sample and its RMSEz\mathrm{RMSE}_z is not defined; sinn=T/N2s_i^{nn}=T/N_2 depends only on the total weight TT of the sample, and on Twitter TT is drawn from a normal distribution with the exact mean and variance of a sum of N1N2N_1N_2 independent exponentials instead of sampling the dense matrices.

Macroscale observables

Nestedness. The binary NODF50{}^{50} is computed on each layer after ordering the nodes by decreasing degree: for each ordered pair i<ji<j the paired overlap is Nij=∣Γi∩Γj∣/kjN_{ij}=|\Gamma_i\cap\Gamma_j|/k_j, the fraction of the neighbours of jj that are also neighbours of ii. NijN_{ij} is averaged over the pairs of each layer and NODF =50(N‾rows+N‾cols)=50(\overline{N}_{\mathrm{rows}}+\overline{N}_{\mathrm{cols}}), in percent. Unlike [50], pairs with equal degree are not set to zero and the two layers have equal weight; the same definition is applied to the data and to every sample, so the comparison is unaffected. Its weighted counterpart is WNODA51{}^{51}: nodes are ordered by decreasing strength, a pair with si>sjs_i>s_j contributes the fraction of the positive entries of jj that are strictly smaller than the corresponding entries of ii (pairs with si=sjs_i=s_j contribute zero), and the contributions are averaged over the pairs of both layers, in percent.

Clustering. The binary bipartite clustering is the Robins–Alexander coefficient 61{}^{61} as implemented in NetworkX52{}^{52}, C=4C4/L3C=4C_4/L_3, with C4C_4 the number of 4-cycles and L3L_3 the number of paths of length three. The weighted clustering CWC_{\mathrm{W}} is the Opsahl coefficient 53{}^{53}: the fraction of the total value of 4-paths ii–α\alpha–jj–β\beta–ll that are closed (i.e. ii and ll share a neighbour other than α,β\alpha,\beta), with the value of a path given by the geometric mean of its four weights.

Representative networks (main-text reconstruction figure). For Trade 2020 and Lake of the Woods we sample S=2000S=2000 networks from each model. At the mesoscale, besides RMSEz\mathrm{RMSE}_z, we report the median relative error med⁡i[∣xi∗−⟨xi⟩∣/xi∗]\operatorname{med}_i[|x_i^*-\langle x_i\rangle|/x_i^*] and the fraction of nodes whose observed value lies within the central 95% interval (2.5th to 97.5th percentile) of their sampled values; a correct model gives a fraction close to 0.95. For a macroscale observable y∈{NODF,C,WNODA,CW}y\in\{\mathrm{NODF},C,\mathrm{WNODA},C_{\mathrm{W}}\}, computed once per sample, we report the zz-score z=(y∗−⟨y⟩)/σyz=(y^*-\langle y\rangle)/\sigma_y and the percentile of y∗y^* in the sampled distribution, which is 0 or 100 when the observed value lies outside all samples. Results are in Table D.1.

Prediction of the empirical zeros

For a topology-aware model with connection probabilities piαp_{i\alpha}, the expected confusion matrix with respect to the observed support biα∗b_{i\alpha}^* is

⟨TP⟩=∑biα∗=1piα,⟨FP⟩=∑biα∗=0piα,⟨FN⟩=∑biα∗=1(1−piα),⟨TN⟩=∑biα∗=0(1−piα),(33)\langle\mathrm{TP}\rangle=\sum_{b_{i\alpha}^*=1}p_{i\alpha},\qquad \langle\mathrm{FP}\rangle=\sum_{b_{i\alpha}^*=0}p_{i\alpha},\qquad \langle\mathrm{FN}\rangle=\sum_{b_{i\alpha}^*=1}(1-p_{i\alpha}),\qquad \langle\mathrm{TN}\rangle=\sum_{b_{i\alpha}^*=0}(1-p_{i\alpha}), \tag*{(33)}

from which PPV =⟨TP⟩/(⟨TP⟩+⟨FP⟩)=\langle\mathrm{TP}\rangle/(\langle\mathrm{TP}\rangle+\langle\mathrm{FP}\rangle), the fraction of predicted links that are observed, and SPC =⟨TN⟩/(N1N2−L∗)=\langle\mathrm{TN}\rangle/(N_1N_2-L^*), the fraction of empirical zeros that are predicted as zeros. These are analytic expectations, not Monte Carlo estimates. The main text reports the PPV; Supplementary Fig. D.1 reports both for the P-BiECM and the BiECM.

NetworkModelknnk^{nn} RMSEz_zknnk^{nn} med. rel. err.knnk^{nn} in 95%snns^{nn} RMSEz_zsnns^{nn} med. rel. err.snns^{nn} in 95%
Trade 2020BiWCM–––48.480.400.20
P-BiECM57.990.174.5×10−34.5 \times10^{-3}10.500.310.13
BiECM4.377.7×10−37.7 \times10^{-3}0.490.870.030.96
Lake of the WoodsBiWCM–––70.500.690.00
P-BiECM10.150.580.032.300.410.58
BiECM0.990.100.870.870.080.90
NetworkModelNODF zzNODF perc.CC zzCC perc.WNODA zzWNODA perc.
Trade 2020BiWCM––––−161.3-161.30.0
P-BiECM367.6100.0172.6100.0209.9100.0
BiECM−15.1-15.10.06.4100.0−53.7-53.70.0
Lake of the WoodsBiWCM––––−183.7-183.70.0
P-BiECM22.2100.016.6100.03.7100.0
BiECM−1.1-1.114.4−0.4-0.435.21.896.5

Table D.1. Reconstruction of the two networks of the main-text figure. 2000 networks sampled from each fitted model. Top: nearest-neighbour degree and strength of the row nodes, with RMSEz_z, the median relative error ∣xi∗−⟨xi⟩∣/xi∗\lvert x_i^*-\langle x_i\rangle\rvert/x_i^* and the fraction of nodes whose observed value falls in the central 95% interval of the sampled values. Bottom: zz-score (y∗−⟨y⟩)/σy(y^*-\langle y\rangle)/\sigma_y and percentile of the observed value in the sampled distribution for NODF, Robins–Alexander clustering CC and WNODA. BiWCM networks are complete, so knnk^{nn}, NODF and CC are not random under that model (–). The weighted clustering CWC_W is not shown because it equals one in Trade 2020 under every model; in Lake of the Woods its zz-score is 5.0 for the P-BiECM and −0.39-0.39 for the BiECM.

Validation of bipartite weights

For an observed positive weight, the upper-tail probability under a hurdle model is piαlink=piαexp⁡(−λiαwiα∗)p_{i\alpha}^{\mathrm{link}}=p_{i\alpha}\exp(-\lambda_{i\alpha}w_{i\alpha}^*), with piα=1p_{i\alpha}=1 for the BiWCM.

Validation of projections

For two nodes i,ji,j of layer 1 the empirical co-volume is Vij∗=∑αwiα∗wjα∗V_{ij}^*=\sum_\alpha w_{i\alpha}^*w_{j\alpha}^*. Its null distribution is estimated from SS networks sampled from the fitted ensemble, and the p-value pijproj=P(Vijnull≥Vij∗)p_{ij}^{\mathrm{proj}}=\mathbb{P}(V_{ij}^{\mathrm{null}}\ge V_{ij}^*) is estimated as

p^ij=1S∑s=1SI(Vij(s)≥Vij∗).(34)\widehat{p}_{ij}=\frac{1}{S}\sum_{s=1}^{S}\mathbb{I}(V_{ij}^{(s)}\ge V_{ij}^*). \tag*{(34)}

The real networks use S=103S=10^3, except plants–ants (S=106S=10^6) and the Twitter matrices, whose Benjamini–Hochberg thresholds are below 10−210^{-2} and are not resolved with 10310^3 samples: for the P-BiECM and the BiECM we use S=3×104S=3\times10^4 (Q22) and S=105S=10^5 (election, crisis), accumulated over independent runs with distinct seeds. A threshold tBHt_{\mathrm{BH}} is resolved when tBHS≥10t_{\mathrm{BH}}S\ge10 exceedances are expected; this holds for every network and model except the BiECM on election and crisis (tBHS≃2t_{\mathrm{BH}}S\simeq2), whose validated sets may still shrink with more samples. The number of samples of every network is given in Table E.3. Pairs are validated with the Benjamini–Hochberg procedure at αsig=0.05\alpha_{\mathrm{sig}}=0.05; when it does not reject any hypothesis for at least one model, all models of that network are validated with the common threshold p^ij<0.05\widehat{p}_{ij}<0.05, so that the sets being compared are obtained with the same rule (column Thr. of Table E.3). This happens for all ecological matrices, whose small size makes the FDR procedure too conservative.

For the Twitter matrices (crisis: 6667×3218246667\times321824) a single BiWCM sample is complete and too expensive to generate in the numbers needed to resolve their thresholds. Since the BiWCM weights are independent exponentials with means μiα=1/λiα\mu_{i\alpha} = 1/\lambda_{i\alpha}, the cumulants of VijV_{ij} are exact,

Figure D.1. Prediction of the observed support by the topology-aware models. a, Positive predictive value and b, specificity of the P-BiECM against the BiECM for the empirical networks of Table B.2. The dashed line is the identity; the boxes report the median difference P-BiECM −BiECM.

κ1=∑αμiαμjα,κ2=3∑αμiα2μjα2,κ3=26∑αμiα3μjα3,κ4=426∑αμiα4μjα4,(35)\kappa_1 = \sum_{\alpha} \mu_{i\alpha}\mu_{j\alpha}, \quad\kappa_2 = 3\sum_{\alpha}\mu_{i\alpha}^{2}\mu_{j\alpha}^{2}, \quad\kappa_3 = 26\sum_{\alpha}\mu_{i\alpha}^{3}\mu_{j\alpha}^{3}, \quad\kappa_4 = 426\sum_{\alpha}\mu_{i\alpha}^{4}\mu_{j\alpha}^{4}, \tag*{(35)}

and the BiWCM p-values are obtained from the Edgeworth expansion of the tail to fourth order. On the election matrix, where the Monte Carlo estimate is available, the two procedures give the same FDR threshold and an F1 overlap of 0.9998 between the validated sets.

Monte Carlo resolution. With SS samples the smallest non-zero p-value is 1/S1/S, and the standard error of p^ij\hat{p}_{ij} is [pijproj(1−pijproj)/S]1/2[p_{ij}^{\mathrm{proj}}(1-p_{ij}^{\mathrm{proj}})/S]^{1/2} (0.007 at pijproj=0.05p_{ij}^{\mathrm{proj}} = 0.05 for S=103S = 10^{3}), so the fixed threshold 0.05 is well resolved with S=103S = 10^{3}. The Benjamini–Hochberg thresholds are much smaller on the largest matrices, and the number of samples is increased until they are resolved (tBHS≥10t_{\mathrm{BH}}S \ge10), with the two exceptions given above.

Complete empirical results

Tables E.1, E.2 and E.3 report, for every network of the corpus, the model comparison, the reconstruction errors and the comparison of the validated projections. Model names are the same in every table. BIC selects the P-BiECM in 32 networks and the BiECM in 14 (all trade and Tech CIT matrices and the Tech CPC matrices up to 2016); AIC, which penalises parameters less, selects the BiECM in 40 networks. The BiWCM is never selected.

NetworkBiWCMP-BiECMBiECMBiWCMP-BiECMBiECMΔBIC2\Delta\mathrm{BIC}_2BICAIC
Trade 2012−4.86×107-4.86 \times10^7−2.78×107-2.78 \times10^7−2.75×107-2.75 \times10^79.72×1079.72 \times10^75.57×1075.57 \times10^75.52×1075.52 \times10^74.41×1054.41 \times10^5BiECMBiECM
Trade 2014−4.79×107-4.79 \times10^7−2.38×107-2.38 \times10^7−2.35×107-2.35 \times10^79.58×1079.58 \times10^74.76×1074.76 \times10^74.71×1074.71 \times10^74.60×1054.60 \times10^5BiECMBiECM
Trade 2016−4.78×107-4.78 \times10^7−2.47×107-2.47 \times10^7−2.44×107-2.44 \times10^79.57×1079.57 \times10^74.94×1074.94 \times10^74.90×1074.90 \times10^74.33×1054.33 \times10^5BiECMBiECM
Trade 2018−4.80×107-4.80 \times10^7−2.53×107-2.53 \times10^7−2.50×107-2.50 \times10^79.61×1079.61 \times10^75.06×1075.06 \times10^75.02×1075.02 \times10^74.52×1054.52 \times10^5BiECMBiECM
Trade 2020−4.79×107-4.79 \times10^7−2.46×107-2.46 \times10^7−2.43×107-2.43 \times10^79.59×1079.59 \times10^74.92×1074.92 \times10^74.87×1074.87 \times10^74.50×1054.50 \times10^5BiECMBiECM
Reddit 2013−1.50×107-1.50 \times10^7−9.50×105-9.50 \times10^5−6.72×105-6.72 \times10^53.11×1073.11 \times10^73.04×1063.04 \times10^63.62×1063.62 \times10^65.84×1055.84 \times10^5P-BiECMBiECM
Reddit 2014−1.14×107-1.14 \times10^7−7.89×105-7.89 \times10^5−6.32×105-6.32 \times10^52.36×1072.36 \times10^72.32×1062.32 \times10^62.75×1062.75 \times10^64.30×1054.30 \times10^5P-BiECMBiECM
Reddit 2015−1.30×107-1.30 \times10^7−8.97×105-8.97 \times10^5−7.42×105-7.42 \times10^52.68×1072.68 \times10^72.54×1062.54 \times10^62.98×1062.98 \times10^64.37×1054.37 \times10^5P-BiECMBiECM
Reddit 2016−2.73×107-2.73 \times10^7−1.48×106-1.48 \times10^6−1.20×106-1.20 \times10^65.58×1075.58 \times10^74.13×1064.13 \times10^64.73×1064.73 \times10^65.95×1055.95 \times10^5P-BiECMBiECM
Reddit 2017−3.07×107-3.07 \times10^7−1.53×106-1.53 \times10^6−1.22×106-1.22 \times10^66.25×1076.25 \times10^74.19×1064.19 \times10^64.72×1064.72 \times10^65.29×1055.29 \times10^5P-BiECMBiECM
Lexical 2015−7.49×105-7.49 \times10^5−3.11×105-3.11 \times10^5−2.81×105-2.81 \times10^51.72×1061.72 \times10^68.42×1058.42 \times10^51.00×1061.00 \times10^61.63×1051.63 \times10^5P-BiECMBiECM
Lexical 2018−1.37×106-1.37 \times10^6−4.69×105-4.69 \times10^5−4.22×105-4.22 \times10^53.05×1063.05 \times10^61.25×1061.25 \times10^61.47×1061.47 \times10^62.17×1052.17 \times10^5P-BiECMBiECM
Lexical 2020−2.08×106-2.08 \times10^6−6.45×105-6.45 \times10^5−5.78×105-5.78 \times10^54.56×1064.56 \times10^61.69×1061.69 \times10^61.95×1061.95 \times10^62.62×1052.62 \times10^5P-BiECMBiECM
Lexical 2021−2.47×106-2.47 \times10^6−6.97×105-6.97 \times10^5−6.20×105-6.20 \times10^55.37×1065.37 \times10^61.83×1061.83 \times10^62.12×1062.12 \times10^62.85×1052.85 \times10^5P-BiECMBiECM
Lexical 2023−3.39×106-3.39 \times10^6−8.65×105-8.65 \times10^5−7.66×105-7.66 \times10^57.31×1067.31 \times10^62.26×1062.26 \times10^62.59×1062.59 \times10^63.32×1053.32 \times10^5P-BiECMBiECM
Twitter Q22−8.93×107-8.93 \times10^7−9.87×105-9.87 \times10^5−8.42×105-8.42 \times10^51.80×1081.80 \times10^83.20×1063.20 \times10^64.14×1064.14 \times10^69.36×1059.36 \times10^5P-BiECMBiECM
Twitter election−3.92×108-3.92 \times10^8−3.05×106-3.05 \times10^6−2.53×106-2.53 \times10^67.87×1087.87 \times10^88.76×1068.76 \times10^61.04×1071.04 \times10^71.64×1061.64 \times10^6P-BiECMBiECM
Twitter crisis−2.15×109-2.15 \times10^9−6.97×106-6.97 \times10^6−5.87×106-5.87 \times10^64.30×1094.30 \times10^92.10×1072.10 \times10^72.59×1072.59 \times10^74.87×1064.87 \times10^6P-BiECMBiECM
Tech CIT 2010−1.94×106-1.94 \times10^6−1.30×106-1.30 \times10^6−1.29×106-1.29 \times10^63.88×1063.88 \times10^62.61×1062.61 \times10^62.60×1062.60 \times10^61.06×1041.06 \times10^4BiECMBiECM
Tech CIT 2012−1.92×106-1.92 \times10^6−1.34×106-1.34 \times10^6−1.33×106-1.33 \times10^63.84×1063.84 \times10^62.68×1062.68 \times10^62.67×1062.67 \times10^61.07×1041.07 \times10^4BiECMBiECM
Tech CIT 2014−1.91×106-1.91 \times10^6−1.34×106-1.34 \times10^6−1.33×106-1.33 \times10^63.83×1063.83 \times10^62.68×1062.68 \times10^62.67×1062.67 \times10^61.04×1041.04 \times10^4BiECMBiECM
Tech CIT 2016−1.91×106-1.91 \times10^6−1.30×106-1.30 \times10^6−1.29×106-1.29 \times10^63.82×1063.82 \times10^62.61×1062.61 \times10^62.60×1062.60 \times10^61.14×1041.14 \times10^4BiECMBiECM
Tech CIT 2018−1.66×106-1.66 \times10^6−1.02×106-1.02 \times10^6−1.01×106-1.01 \times10^63.33×1063.33 \times10^62.05×1062.05 \times10^62.04×1062.04 \times10^61.10×1041.10 \times10^4BiECMBiECM
Tech CPC 2010−8.39×105-8.39 \times10^5−3.80×105-3.80 \times10^5−3.76×105-3.76 \times10^51.69×1061.69 \times10^67.66×1057.66 \times10^57.66×1057.66 \times10^5591BiECMBiECM
Tech CPC 2012−8.38×105-8.38 \times10^5−3.99×105-3.99 \times10^5−3.95×105-3.95 \times10^51.68×1061.68 \times10^68.06×1058.06 \times10^58.05×1058.05 \times10^5674BiECMBiECM
Tech CPC 2014−8.40×105-8.40 \times10^5−4.08×105-4.08 \times10^5−4.04×105-4.04 \times10^51.69×1061.69 \times10^68.23×1058.23 \times10^58.23×1058.23 \times10^5786BiECMBiECM
Tech CPC 2016−8.47×105-8.47 \times10^5−4.14×105-4.14 \times10^5−4.10×105-4.10 \times10^51.70×1061.70 \times10^68.36×1058.36 \times10^58.35×1058.35 \times10^5854BiECMBiECM
Tech CPC 2018−7.41×105-7.41 \times10^5−2.21×105-2.21 \times10^5−2.19×105-2.19 \times10^51.49×1061.49 \times10^64.49×1054.49 \times10^54.51×1054.51 \times10^51373P-BiECMBiECM
Beehler−236-236−150-150−137-137626458581123P-BiECMP-BiECM
Seed disp. Beehler−501-501−450-450−420-42012271131129195.9P-BiECMP-BiECM
Seed disp. Snow−1464-1464−1199-1199−1112-1112323027062830123P-BiECMBiECM
Plant-pollinator BH−1351-1351−675-675−557-557351321672734568P-BiECMBiECM
Plants–ants−2256-2256−1039-1039−928-928521627883262474P-BiECMBiECM
Aishihik int.−802-802−559-559−517-517179513151416101P-BiECMBiECM
Aishihik prev.−1136-1136−665-665−623-623246315261629104P-BiECMBiECM
Canada−602-602−433-433−398-398148911581366209P-BiECMP-BiECM
Cold Lake int.−1279-1279−667-667−624-624285816391846208P-BiECMP-BiECM
Cold Lake prev.−1674-1674−761-761−712-712364818282024195P-BiECMBiECM
Huron prev.−1.03×104-1.03 \times10^4−2903-2903−2689-26892.17×1042.17 \times10^468647476613P-BiECMBiECM
Lake of the Woods−1.67×104-1.67 \times10^4−4252-4252−3912-39124.59×1044.59 \times10^41.00×1041.00 \times10^41.08×1041.08 \times10^4802P-BiECMBiECM
Lake of the Woods prev.−1.62×104-1.62 \times10^4−3745-3745−3522-35223.39×1043.39 \times10^4896899861018P-BiECMBiECM
McGregor^a−1926-1926−864-864−801-801427921632456293P-BiECMP-BiECM
Parsnip int.−3089-3089−1273-1273−1184-1184665530293320290P-BiECMBiECM
Parsnip prev.−2602-2602−1190-1190−1119-1119568028643191327P-BiECMBiECM
Smallwood int.−339-339−269-269−238-23883469978788.3P-BiECMBiECM
Smallwood prev.−414-414−301-301−272-27298376285593.7P-BiECMP-BiECM

Table E.1. Table E.1: Likelihood-based model comparison. Log-likelihood of the recorded data, evaluated at the fitted parameters, with each empty cell given the probability 1−piαe−λiαΔw1 - p_{i\alpha}e^{-\lambda_{i\alpha}\Delta w} of a value below Δw\Delta w (main text, Methods) (probabilities integrated over the observation bins of width Δw\Delta w, Table B.2) and BIC =kln⁡(N1N2)−2ln⁡L= k\ln(N_1N_2) - 2\ln\mathcal{L}, with k=N1+N2k = N_1 + N_2 (BiWCM), N1+N2+1N_1 + N_2 + 1 (P-BiECM) and 2(N1+N2)2(N_1 + N_2) (BiECM). The largest likelihood and the smallest BIC are in bold. ΔBIC2\Delta\mathrm{BIC}_2 is the gap between the best and the second-best model: since its smallest value over the corpus is 88.3, the BIC weight of the selected model, [1+∑m≠beste−ΔBICm/2]−1[1 + \sum_{m \ne\mathrm{best}} e^{-\Delta\mathrm{BIC}_m/2}]^{-1}, exceeds 1−10−81 - 10^{-8} in every network. The last two columns give the model selected by BIC and by AIC =2k−2ln⁡L= 2k - 2\ln\mathcal{L}. a^a Intensity matrix; the prevalence matrix in the source data is identical to it and is not included.

Networkknnk^{nn}, layer 1 BiWCMknnk^{nn}, layer 1 P-BiECMknnk^{nn}, layer 1 BiECMknnk^{nn}, layer 2 BiWCMknnk^{nn}, layer 2 P-BiECMknnk^{nn}, layer 2 BiECMsnns^{nn}, layer 1 BiWCMsnns^{nn}, layer 1 P-BiECMsnns^{nn}, layer 1 BiECMsnns^{nn}, layer 2 BiWCMsnns^{nn}, layer 2 P-BiECMsnns^{nn}, layer 2 BiECM
Trade 2012–63.53.61–9.261.3825.57.830.74321.94.610.336
Trade 2014–59.83.92–10.81.5132.58.10.86127.94.530.398
Trade 2016–61.94.12–10.21.4446.6100.88639.84.930.433
Trade 2018–61.34.41–9.981.4637.88.270.8634.24.910.41
Trade 2020–58.54.38–10.81.4446.910.40.94240.85.130.458
Reddit 2013–2746.07–52.60.85962962390.68682111.30.84
Reddit 2014–1875.21–30.10.99142171151.064466.911
Reddit 2015–1885.32–200.955405095.20.9541913.561.03
Reddit 2016–2325.8–39.11.277621951.1152112.61.21
Reddit 2017–2395.38–49.81.1864791401.034769.071.22
Lexical 2015–70.73.69–3.280.89552718.80.90255.81.630.991
Lexical 2018–82.63.95–30.8591061251.26741.630.987
Lexical 2020–90.43.94–3.340.854113623.41.0772.91.740.997
Lexical 2021–98.83.66–3.480.913145229.61.11072.071.09
Lexical 2023–1064–4.260.897143829.31.051132.351.11
Twitter Q22–1345.07–16.91.31384431.83.1756284.431.27
Twitter election–64.33.91–16.61.25419926.73.061.71×1041.71 \times10^{4}4.571.23
Twitter crisis–69.34.53–30.11.42340317.33.524.44×1044.44 \times10^{4}10.31.4
Tech CIT 2010–19.11.18–5.340.82213.73.590.56513.32.70.0672
Tech CIT 2012–18.71.13–5.260.79914.73.640.55512.22.640.105
Tech CIT 2014–18.71.22–5.190.831153.740.569132.740.0935
Tech CIT 2016–18.31.13–5.370.88214.74.010.52915.32.970.0915
Tech CIT 2018–18.11.24–5.470.88618.24.150.57624.53.660.16
Tech CPC 2010–151.55–4.570.716387.310.62137.52.610.421
Tech CPC 2012–14.11.43–4.440.75831.66.20.51838.22.530.413
Tech CPC 2014–14.51.44–4.550.75226.66.130.34530.72.490.376
Tech CPC 2016–14.71.38–4.590.7333.26.920.377362.60.398
Tech CPC 2018–11.31.48–4.470.75851.66.560.80364.72.390.65
Beehler–1.50.921–1.130.6255.280.6660.4815.860.7160.558
Seed disp. Beehler–1.851.01–1.420.7555.760.9880.5054.811.140.676
Seed disp. Snow–1.470.701–2.360.6275.311.340.8117.070.7980.494
Plant–pollinator BH–4.240.931–3.491.6112.41.610.83330.43.480.617
Plants–ants–4.630.952–3.630.84632.31.920.71827.12.10.844
Aishihik int.–0.6891.83–0.9720.5254.850.8270.2652.460.9560.955
Aishihik prev.–1.232.06–1.20.5115.770.5120.5562.630.7490.437
Canada–0.7221.39–1.140.6459.641.140.2291.670.5510.548
Cold Lake int.–0.7791.39–1.250.65211.61.130.2091.950.5610.574
Cold Lake prev.–0.7621.35–1.470.69711.61.040.4632.090.6670.592
Huron prev.–5.441.07–2.570.89593.53.170.8258.571.010.721
Lake of the Woods–9.970.993–1.560.85968.12.30.84512.60.8980.735
Lake of the Woods prev.–5.371.24–1.60.87956.61.750.87413.30.8580.742
McGregorᵃ–1.741.23–1.750.64418.81.320.553.930.8020.65
Parsnip int.–2.861.23–1.050.60224.62.031.025.480.7130.615
Parsnip prev.–2.141.27–1.070.64215.70.9060.6976.320.8750.578
Smallwood int.–2.641.56–1.490.6052.140.8490.4562.610.7750.651
Smallwood prev.–2.431.58–1.420.6172.860.9350.5882.390.6240.423

Table E.2. Table E.2: Reconstruction of the nearest-neighbour degree and strength. Standardised root mean square error RMSEz=[1N∑i(xi∗−⟨xi⟩)2/σxi2]1/2\mathrm{RMSE}_z = \left[\frac{1}{N}\sum_i (x_i^* - \langle x_i\rangle)^2/\sigma_{x_i}^2\right]^{1/2} of knnk^{nn} and snns^{nn} on the two layers, with ensemble averages and standard deviations estimated from sampled networks (Supplementary Note S3). The smallest error is in bold. BiWCM networks are complete with probability one, so kinnk_i^{nn} is the same in every sample and its RMSEz\mathrm{RMSE}_z is undefined (–).ᵃ Intensity matrix; the prevalence matrix in the source data is identical to it and is not included.

NetworkThr.SS (W/P/E)BiWCMP-BiECMBiECMW–PW–EP–EW–PW–EP–E
Trade 2012FDR103/103/10310^3/10^3/10^30.8670.8450.7740.9870.9430.9560.8940.7710.875
Trade 2014FDR103/103/10310^3/10^3/10^30.8210.7880.7040.9790.9220.9420.8910.780.889
Trade 2016FDR103/103/10310^3/10^3/10^30.8280.8010.7170.9830.9280.9440.9020.7820.883
Trade 2018FDR103/103/10310^3/10^3/10^30.8260.8030.7230.9850.9330.9470.9120.7930.885
Trade 2020FDR103/103/10310^3/10^3/10^30.8050.7770.6870.9820.920.9380.9140.7940.886
Reddit 2013FDR103/103/10310^3/10^3/10^30.8930.8910.5420.9990.7550.7560.9480.5510.573
Reddit 2014FDR103/103/10310^3/10^3/10^30.8650.8570.4160.9950.6490.6530.8810.5870.645
Reddit 2015FDR103/103/10310^3/10^3/10^30.9260.9150.5120.9940.7110.7150.7820.4420.527
Reddit 2016FDR103/103/10310^3/10^3/10^30.9190.9110.5970.9960.7870.790.8410.4810.551
Reddit 2017FDR103/103/10310^3/10^3/10^30.9170.9050.6130.9930.8010.8070.810.4960.599
Lexical 2015FDR103/103/10310^3/10^3/10^30.8260.7490.6960.9510.9010.9330.8580.7630.867
Lexical 2018FDR103/103/10310^3/10^3/10^30.8180.7440.6660.9530.8890.9080.8520.7130.742
Lexical 2020FDR103/103/10310^3/10^3/10^30.8050.7380.7460.9570.9490.9560.8670.8070.863
Lexical 2021FDR103/103/10310^3/10^3/10^30.8670.8170.7820.970.9430.9490.8190.7150.723
Lexical 2023FDR103/103/10310^3/10^3/10^30.8830.8470.8570.9790.9690.9650.8360.7610.775
Twitter Q22FDRE/3⋅1043\cdot10^4/3⋅1043\cdot10^40.07480.03217.74×10−37.74\times10^{-3}0.6010.1870.3880.9730.9780.994
Twitter electionFDRE/10510^5/10510^50.02958.49×10−38.49\times10^{-3}4.45×10−44.45\times10^{-4}0.4470.02960.08920.9650.9710.988
Twitter crisisFDRE/10510^5/10510^50.01493.84×10−33.84\times10^{-3}2.82×10−42.82\times10^{-4}0.4090.0370.08340.970.9810.989
Tech CIT 2010FDR103/103/10310^3/10^3/10^30.9510.8940.8050.9690.9170.9470.5840.5120.861
Tech CIT 2012FDR103/103/10310^3/10^3/10^30.9470.8970.8220.9730.930.9560.6130.5330.86
Tech CIT 2014FDR103/103/10310^3/10^3/10^30.9660.9250.8270.9780.9220.9440.5260.4380.839
Tech CIT 2016FDR103/103/10310^3/10^3/10^30.950.8980.8090.9720.920.9470.610.5240.846
Tech CIT 2018FDR103/103/10310^3/10^3/10^30.9450.8730.7330.960.8740.9130.5540.4790.872
Tech CPC 2010FDR103/103/10310^3/10^3/10^30.9750.950.7440.9870.8650.8760.50.2750.69
Tech CPC 2012FDR103/103/10310^3/10^3/10^30.9540.920.7310.9820.8650.8820.6220.3960.686
Tech CPC 2014FDR103/103/10310^3/10^3/10^30.9540.9320.7750.9890.8970.9080.6250.3670.665
Tech CPC 2016FDR103/103/10310^3/10^3/10^30.9780.9540.7680.9880.880.8910.5470.2820.61
Tech CPC 2018FDR103/103/10310^3/10^3/10^30.8410.5880.2050.8240.3920.5160.6990.6020.908
Beehler0.05103/103/10310^3/10^3/10^30.2730.036400.235000.9920.9710.978
Seed disp. Beehler0.05103/103/10310^3/10^3/10^30.1940.083300.6000.9790.960.974
Seed disp. Snow0.05103/103/10310^3/10^3/10^30.2870.09360.04680.4920.2460.50.9540.9480.979
Plant–pollinator BH0.05103/103/10310^3/10^3/10^30.3760.1895.82×10−45.82\times10^{-4}0.6712.06×10−32.06\times10^{-3}2.04×10−32.04\times10^{-3}0.9890.9740.975
Plants–ants0.05106/106/10610^6/10^6/10^60.4540.1021.22×10−31.22\times10^{-3}0.3685.36×10−35.36\times10^{-3}0.02350.9910.9340.952
Aishihik int.0.05103/103/10310^3/10^3/10^30.190.1430.04760.8570.40.50.9570.8950.849
Aishihik prev.0.05103/103/10310^3/10^3/10^30000000.9650.8350.829
Canada0.05103/103/10310^3/10^3/10^30.0222000000.9750.530.501
Cold Lake int.0.05103/103/10310^3/10^3/10^30.0222000000.9740.5810.549
Cold Lake prev.0.05103/103/10310^3/10^3/10^30.066700.0222000.50.9810.8440.822
Huron prev.0.05103/103/10310^3/10^3/10^30.3010.1160.01330.5550.06020.08820.9840.8840.906
Lake of the Woods0.05103/103/10310^3/10^3/10^30.3330.03010.01080.1660.050.3160.9730.8070.812
Lake of the Woods prev.0.05103/103/10310^3/10^3/10^30.3050.03688.60×10−38.60\times10^{-3}0.2140.04110.190.9740.9170.939
McGregora^a0.05103/103/10310^3/10^3/10^30.1320.065900.667000.9060.9330.93
Parsnip int.0.05103/103/10310^3/10^3/10^30.140.04410.02940.480.3480.60.9440.9310.971
Parsnip prev.0.05103/103/10310^3/10^3/10^30.1910.06620.02210.5140.2070.50.9720.9530.968
Smallwood int.0.05103/103/10310^3/10^3/10^30000000.9340.9340.996
Smallwood prev.0.05103/103/10310^3/10^3/10^30000000.930.9090.996

Table E.3. Table E.3: Validation of the weighted projections on layer 1. Fraction of validated pairs for each model, F1 overlap between the sets of validated pairs and Spearman correlation between the pp-values of all pairs (W = BiWCM, P = P-BiECM, E = BiECM). Column SS: number of networks sampled from each fitted ensemble to estimate the pp-values (W/P/E); E denotes the analytic pp-values of the BiWCM on the Twitter matrices (Edgeworth expansion of the exact cumulants, Section D). Column Thr.: pairs are validated with the Benjamini–Hochberg threshold at αsig=0.05\alpha_{\mathrm{sig}}=0.05 (FDR) when it exists for all three models, otherwise with the common threshold p^ij<0.05\widehat{p}_{ij}<0.05 (0.05). All thresholds correspond to at least ten expected exceedances at the given SS, except for the BiECM on Twitter election and Twitter crisis, where the Benjamini–Hochberg threshold (2.2×10−52.2\times10^{-5} and 1.4×10−51.4\times10^{-5}) corresponds to about two exceedances at S=105S=10^5, and the validated sets of that model may still shrink with more samples. The pp-value correlations do not depend on the threshold. a^a Intensity matrix; the prevalence matrix in the source data is identical to it and is not included.

Network properties and projection disagreement

This analysis is exploratory. Its purpose is to see which properties of the input networks appear to accompany the differences between the validated projections of the BiWCM, the P-BiECM and the BiECM, and to guide the controlled experiments of Section G; it is not meant to establish causal effects. For comparability, all pairs are validated here with the common threshold p^<0.05\hat{p}<0.05 for every network and model, so the validated fractions and F1 overlaps of the non-ecological networks differ from those of Table E.3, which use the Benjamini–Hochberg procedure.

For each of the 46 networks we consider nine outcomes: the Spearman correlation between the p-values of each pair of models, which does not depend on any threshold, the F1 overlap between the validated sets of each pair and the validated fraction of each model. The F1 overlap is not defined when neither model validates any pair (three ecological networks for the pairs including the BiWCM, five for the P-BiECM–BiECM pair), and these networks are excluded from the corresponding tests. We use twelve descriptors: the five used for the synthetic families (density dd and mean degree ⟨k⟩\langle k\rangle, as in Table B.2, the variance of the degrees of both layers pooled, and the mean and variance of the non-zero weights) and seven further ones (the layer sizes N1N_1 and N2N_2, the mean binary co-occurrence Vˉ=∑i≠j∑αbiαbjα/[N1(N1−1)]\bar{V}=\sum_{i\ne j}\sum_{\alpha}b_{i\alpha}b_{j\alpha}/[N_1(N_1-1)], the coefficients of variation of the degrees and of the weights, the strength–degree Spearman correlation and the weight disparity). Mean weight and weight variance depend on the units of each system and are compared across systems only indicatively.

We use only non-parametric tests: Spearman correlations across networks with Benjamini–Hochberg correction over the 108 descriptor–outcome pairs (adjusted p-values pBHp_{\mathrm{BH}}), and, for the five descriptors of the synthetic families, Kruskal–Wallis tests across terciles with effect size ηH2=(H−k+1)/(n−k)\eta_H^2=(H-k+1)/(n-k). Since snapshots of the same system are not independent, we also report a within-system permutation test, in which the outcome is permuted among networks of the same system (10410^4 permutations; adjusted p-values pBHwithinp_{\mathrm{BH}}^{\mathrm{within}}), and a 95% bootstrap confidence interval (CI) over the seven systems (5×1035\times10^3 resamples).

Many descriptors are correlated with the outcomes across the corpus (66 of 108 pairs with pBH<0.05p_{\mathrm{BH}}<0.05, Fig. F.1a), but only a few of these associations persist within systems (eight), so most of them reflect differences between systems. We therefore read the results only as indications.

  • Degree variance. Among the five descriptors of the synthetic families, it is the one that appears most associated with the outcomes (mean ηH2=0.48\eta_H^2=0.48, against 0.43 for the mean degree, 0.36 for density, 0.20 for the weight variance and 0.06 for the mean weight; Fig. F.1c). Networks with larger degree variance tend to have higher F1 overlaps (ρs\rho_s between 0.68 and 0.82) and lower correlations between p-values (ρs\rho_s between −0.47 and −0.28; Fig. F.1d,e). A possible reading is that, when degrees are heterogeneous, the models validate similar sets of pairs but rank them differently. In the real networks the degree variance also grows with size (Fig. F.1b), so its role cannot be separated here; this is one motivation for the degree-heterogeneity family of Section G.

  • Density. The associations are weak (∣ρs∣≤0.35|\rho_s|\le0.35) and not monotonic: the middle tercile, mostly ecological networks, has the lowest validated fractions (median 0.14 for the BiWCM against 0.83 and 0.95 in the other terciles; Fig. F.1f,g). In the real networks density does not appear to be a good indicator of projection disagreement.

  • Co-occurrence. The mean co-occurrence Vˉ\bar{V} shows the clearest association, the only one that persists both within systems and in the bootstrap over systems: it is correlated with the F1 overlap of the BiWCM–BiECM (ρs=0.87\rho_s=0.87, 95% CI [0.54, 0.95]) and P-BiECM–BiECM pairs (0.84, [0.39, 0.94]). Networks with more co-occurrences per pair seem to provide more evidence for every model, so that more pairs are validated by all of them.

  • Weight descriptors. Mean weight, weight variance, strength–degree correlation and weight disparity show no clear pattern.

Figure F.1. Input-network properties and projection disagreement in the 46 empirical networks (exploratory). All pairs are validated with p^<0.05\hat{p} < 0.05. a, Spearman correlation between twelve input descriptors (rows; above the line, the five descriptors of the synthetic families) and nine projection outcomes: the Spearman correlation between the pp-values and the F1 overlap of the validated sets for each pair of models (W: BiWCM, P: P-BiECM, E: BiECM), and the validated fraction of each model. Dots: pBH<0.05p_{\mathrm{BH}} < 0.05 over the 108 tests; asterisks: also pBHwithin<0.05p_{\mathrm{BH}}^{\mathrm{within}} < 0.05 in the within-system permutation test. b, Spearman correlation between the descriptors; the five descriptors of the synthetic families are framed. c, Kruskal–Wallis effect size ηH2\eta_{\mathrm{H}}^2 across the terciles of these five descriptors; dots: pBH<0.05p_{\mathrm{BH}} < 0.05. d–g, F1 overlap (d,f) and p-value correlation (e,g) between the BiWCM and the BiECM against the degree variance (d,e) and the density (f,g), coloured by system. Above each panel: Spearman correlation ρs\rho_s, 95% CI and pBHwithinp_{\mathrm{BH}}^{\mathrm{within}} of the within-system permutation test.

Synthetic network experiments

We use three synthetic families, all with N1=N2=500N_1=N_2=500. The ER family at fixed total weight (Algorithm G.1) tests model selection between global and local topological constraints (Section C). The transition family (Algorithm G.2) and the degree-heterogeneity family (Algorithm G.3) test how the information encoded in topology and weights changes the validated projections. All networks of a family are fitted with the BiCM, the BiWCM, the P-BiECM and the BiECM. The transition family moves the block structure from the links to the weights, testing the hierarchy of the null models (topology only, strengths only, or both[21]); the degree-heterogeneity family varies the degree variance, which appears associated with projection disagreement in the real networks (Section F). Along the transition density, weights and degrees change together, while the degree-heterogeneity family keeps density and weights fixed; in both, only the direction of the effects is compared with the real networks.

Algorithm G.1: Erdős–Rényi family at fixed total weight.

Input: N1=N2=500N_1=N_2=500; grid p∈[10−2,0.5]p\in[10^{-2},0.5] (30 log-spaced values); total weight Wtot=107W_{\mathrm{tot}}=10^7; 3 independent realisations

draw uiα∼U(0,1)u_{i\alpha}\sim\mathcal{U}(0,1) once per realisation

for each pp do

biα←I[uiα<p]b_{i\alpha}\leftarrow\mathbb{I}[u_{i\alpha}<p], L←∑i,αbiαL\leftarrow\sum_{i,\alpha}b_{i\alpha}

draw shares gℓ∼Gamma⁡(1,1)g_\ell\sim\operatorname{Gamma}(1,1) for the LL links, (mℓ)∼Multinomial⁡(Wtot−L;gℓ/∑ℓgℓ)(m_\ell)\sim\operatorname{Multinomial}(W_{\mathrm{tot}}-L;g_\ell/\sum_\ell g_\ell)

wℓ←1+mℓw_\ell\leftarrow1+m_\ell on each link, so that ∑i,αwiα=Wtot\sum_{i,\alpha}w_{i\alpha}=W_{\mathrm{tot}} with integer weights (Δw=1\Delta w=1)

Algorithm G.2: Transition family: the block structure moves from the topology to the weights.

Input: N1=N2=500N_1=N_2=500, two equal blocks; grid δ∈[10−3,1]\delta\in[10^{-3},1] (50 log-spaced values); one realisation per value

for each δ\delta do

pin←0.30+0.70δp_{\mathrm{in}}\leftarrow0.30+0.70\delta, pout←0.01+0.99δp_{\mathrm{out}}\leftarrow0.01+0.99\delta

win←102+(104−102)δw_{\mathrm{in}}\leftarrow10^2+(10^4-10^2)\delta, wout←102−(102−1)δw_{\mathrm{out}}\leftarrow10^2-(10^2-1)\delta

for each pair (i,α)(i,\alpha) do

(p,w)←(pin,win)(p,w)\leftarrow(p_{\mathrm{in}},w_{\mathrm{in}}) if ii and α\alpha are in the same block, else (pout,wout)(p_{\mathrm{out}},w_{\mathrm{out}})

biα∼Bernoulli⁡(p)b_{i\alpha}\sim\operatorname{Bernoulli}(p), wiα←biαuiαw_{i\alpha}\leftarrow b_{i\alpha}u_{i\alpha}, uiα∼U(0.9w,1.1w)u_{i\alpha}\sim\mathcal{U}(0.9w,1.1w)

Algorithm G.3: Degree-heterogeneity family at fixed density and block contrast.

Input: N1=N2=500N_1=N_2=500, two equal blocks; target densities pin=0.70p_{\mathrm{in}}=0.70, pout=0.20p_{\mathrm{out}}=0.20; grid δ∈{0,0.025,…,0.45}\delta\in\{0,0.025,\ldots,0.45\} (19 values); 3 independent realisations

draw zi,zα∼N(0,1)z_i,z_\alpha\sim\mathcal{N}(0,1) once per realisation

for each δ\delta do

target degrees κi∝eδzi\kappa_i\propto e^{\delta z_i}, κα∝eδzα\kappa_\alpha\propto e^{\delta z_\alpha}, rescaled to the mean degree implied by pin,poutp_{\mathrm{in}},p_{\mathrm{out}}

piα←xiyαciα1+xiyαciαp_{i\alpha}\leftarrow\frac{x_i y_\alpha c_{i\alpha}}{1+x_i y_\alpha c_{i\alpha}}, ciα=1c_{i\alpha}=1 within blocks, ciα=coutc_{i\alpha}=c_{\mathrm{out}} across blocks

solve for xi,yα,coutx_i,y_\alpha,c_{\mathrm{out}} such that ∑αpiα=κi\sum_\alpha p_{i\alpha}=\kappa_i, ∑ipiα=κα\sum_i p_{i\alpha}=\kappa_\alpha and the expected density across blocks equals poutp_{\mathrm{out}}

biα∼Bernoulli⁡(piα)b_{i\alpha}\sim\operatorname{Bernoulli}(p_{i\alpha}), wiα←biαuiαw_{i\alpha}\leftarrow b_{i\alpha}u_{i\alpha}, uiα∼U(0.5⋅105,1.5⋅105)u_{i\alpha}\sim\mathcal{U}(0.5\cdot10^5,1.5\cdot10^5)

ER family. The topology is homogeneous at every density, so the degrees carry no information beyond LL and the global-density model is the correct topological description. Fixing the total weight while increasing the density spreads the same volume over more links, so density and weight disparity vary together. The family therefore checks whether BiC recovers the global-density model (it does in all 90 networks, Fig. C.1) and how close to the decision boundary the likelihood ratio gets as the network densifies.

Transition family. At small δ\delta the two blocks are encoded in the topology (pin=0.30p_{\mathrm{in}} = 0.30, pout=0.01p_{\mathrm{out}} = 0.01) with nearly uniform weights; at δ=1\delta= 1 the network is complete and the blocks are encoded only in the weights (10410^{4} against 1). The family is therefore an interpolation between a block structure carried by the binary links and one carried only by the weights.

Degree-heterogeneity family. Degree heterogeneity increases with δ\delta while the expected density within and across blocks is kept fixed: the connection probabilities have the BiCM form with a block factor, so they remain below one for every δ\delta and neither the density nor the block contrast changes with heterogeneity. Weights are nearly uniform, so the strengths follow the degrees. In the figures we use δ≤0.325\delta\le0.325: beyond this value the F1 overlap between the BiCM and the BiWCM no longer changes beyond the variability between realisations. A previous version of this family, with multiplicative lognormal propensities piα=min⁡(1,pblockαqiβα)p_{i\alpha} = \min(1, p_{\mathrm{block}\alpha}q_i\beta_\alpha), was discarded because the truncation at one made density and block contrast decrease together with δ\delta.

Validation and sample size. For each network and model, the projection p-values are estimated from S=103S = 10^{3} sampled networks, and pairs are validated with the fixed threshold p^<0.05\hat{p} < 0.05, the same for all models, so that the overlap between validated sets reflects the null models and not different thresholds. The threshold is fifty times the p-value resolution 1/S1/S, and the Monte Carlo standard error at pijproj=0.05p_{ij}^{\mathrm{proj}} = 0.05 is 0.007; with S=103S = 10^{3} the validated sets are therefore stable, while the Benjamini–Hochberg thresholds can fall below 1/S1/S and would not be resolved. Each figure reports, besides the F1 overlap, the Pearson correlation between the p-values of all pairs, which does not depend on any threshold.

Results. In the transition family (Fig. G.1) the agreement between the BiCM and the BiWCM falls from F1 ≃0.93\simeq0.93 to zero as δ\delta grows (ρs=−1.00\rho_s = -1.00), while the agreement between the BiWCM and the BiECM rises from 0.38 to one (ρs=+0.99\rho_s = +0.99): as the structure moves into the weights, the topology-only model loses the signal and the topology-aware weighted model converges to the strength-only one, because the network becomes complete. The same pattern holds for the p-value correlations and for all pairs involving the P-BiECM (panel c).

In the degree-heterogeneity family (Fig. G.2) the two agreement measures behave differently, unlike in the transition family where both increase together. The F1 overlap between the BiCM and the BiWCM decreases from 1.00 to about 0.91 as the degree variance grows (ρs=−0.98\rho_s = -0.98), and the Pearson correlation between their p-values decreases accordingly (from 1.000 to 0.97 at δ=0.325\delta= 0.325). The F1 overlap between the BiWCM and the BiECM instead increases, from about 0.45 to 0.68 (ρs=+0.84\rho_s = +0.84), while the correlation between their p-values decreases, from 0.996 to 0.978 (to 0.917 at δ=0.45\delta= 0.45). The two observations are consistent: the pairs validated by the BiECM are almost all (95–100%) also validated by the BiWCM, and as heterogeneity grows the BiWCM validates fewer pairs (from 50% to 41% of all pairs) and the BiECM more (from 15% to 21%). For nested sets the F1 overlap is 2b/(a+b)2b/(a+b), with aa and bb the two validated fractions, so it increases when the two sizes become closer, even if the p-values of the two models become less correlated. The increase of the F1 overlap therefore does not indicate a convergence of the two null models: the threshold-independent p-value correlation shows that agreement decreases with degree heterogeneity, for both pairs (panels f and g) and for all pairs of models (panel c). The same decrease of the p-value correlation is found in the sparse variant of the family (pin=0.20p_{\mathrm{in}} = 0.20, pout=0.05p_{\mathrm{out}} = 0.05, from 0.89 to 0.49) and in the previous version of the family. Density, mean degree and mean weight are uncorrelated with all the agreement measures (panel c), confirming that the controlled variable is heterogeneity alone. In this dense family the Spearman correlation between the p-values stays close to 0.92 for both pairs, so heterogeneity changes the values of the p-values more than their ranking; in the sparse variant the Spearman correlation decreases as well, from 0.96 to 0.48.

Figure G.1. Null-model agreement across a topology-to-weight interpolation (Algorithm G.2). a,b, Weighted biadjacency matrices at low and high δ\delta (white: no link). c, Spearman correlation between input-network properties and two agreement measures between the projections of each pair of models: the F1 overlap between the validated sets and the Pearson correlation between the p-values. d,e, F1 overlap between the BiCM and the BiWCM and between the BiWCM and the BiECM against δ\delta. f,g, Pearson correlation between the p-values of all pairs for the same two pairs of models. Colour encodes δ\delta, and the circled markers are the networks of a and b. Pairs are validated with p^<0.05\hat{p}<0.05 for all models, with 10310^3 samples per network and model.

Figure G.2. Null-model agreement under increasing degree heterogeneity at fixed density (Algorithm G.3). a,b, Weighted biadjacency matrices at low and high δ\delta. c, Spearman correlation between input-network properties and the agreement measures between pairs of models, as in Fig. G.1. d,e, F1 overlap between the BiCM and the BiWCM and between the BiWCM and the BiECM against the degree variance, for three realisations per δ\delta. f,g, Pearson correlation between the p-values of all pairs for the same two pairs of models: unlike the F1 overlap in e, it decreases for both pairs. Grey crosses are means and standard deviations over realisations at fixed δ\delta, and the circled markers are the networks of a and b. Pairs are validated with p^<0.05\hat{p} < 0.05 for all models, with 10310^{3} samples per network and model.

Approximate enhanced inference: qBiECM

Solving the parameters of the BiECM can be computationally demanding. Therefore, in our toolkit, we also provide the quasi-Bipartite Enhanced Configuration Model (qBiECM), a method inspired by the BiECM that introduces a simplification in the derivation and reduces the computation time of the parameters2.

The maximisation of the entropy is split into two steps. First, a BiCM[31] is fitted to the empirical degrees, which gives the connection probabilities piαqBiECM=e−(θi+ηα)/[1+e−(θi+ηα)]p_{i\alpha}^{\mathrm{qBiECM}} = e^{-(\theta_i+\eta_\alpha)}/[1+e^{-(\theta_i+\eta_\alpha)}], independent of the strength multipliers. Second, the exponential weight kernel is fitted to the strengths conditionally on these probabilities, i.e. by imposing ∑αpiα/λiα=si∗\sum_{\alpha}p_{i\alpha}/\lambda_{i\alpha}=s_i^* and ∑ipiα/λiα=σα∗\sum_i p_{i\alpha}/\lambda_{i\alpha}=\sigma_\alpha^*. The resulting dyadic law has the same hurdle form as the BiECM. The approximation neglects the dependence of the topology on the weights and is not the joint maximum-entropy solution; for this reason the qBiECM is not a member of the hierarchy used for model selection. Its log-likelihood, evaluated with the same expressions as the BiECM, can exceed that of the BiECM, because the two models are not optimised over the same parametrised family.

The qBiECM converges on all networks with the coordinate solver, at a cost similar to or lower than that of the BiECM (§ 23).

Numerical implementation and convergence

The complete workflow is implemented in an open-source Python package built on PyTorch, exposing a common interface for the null models (BiCM, BiWCM, P-BiECM, BiECM and qBiECM) and complementing existing maximum-entropy solvers such as NEMtropy[30]. All solvers are vectorised over dyads, can be chunked row-wise to bound memory use on large matrices, and run on GPU when requested.

Parameter fitting solves the moment-matching equations ⟨C⟩=C∗\langle C\rangle=C^*. Three solvers are available.

  • Fixed point. The multipliers are updated multiplicatively with the relative constraint error, x(t+1)=x(t)⊙(1+γv(t))x^{(t+1)}=x^{(t)}\odot(1+\gamma v^{(t)}), where vv is a Polyak-style momentum average of ⟨C⟩/C∗−1\langle C\rangle/C^*-1 (step γ=1\gamma=1, momentum β=0.9\beta=0.9, optionally annealed to 0.999 during a warm-up window). The BiWCM converges without momentum.

  • Quasi-Newton. A diagonal approximation of the Jacobian of the constraints is used as preconditioner, with a model-specific step size.

  • Coordinate. Each multiplier is updated in turn by solving its own equation with the others fixed (bisection in log-space, 60 steps), which guarantees monotonic convergence when the fixed point is unstable.

Convergence is assessed on the maximum relative error between empirical and expected constraints, max⁡∣⟨C⟩/C∗−1∣\max|\langle C\rangle/C^*-1|: in all the analyses of this work a solution is accepted when this error is below 10−310^{-3}.

Benchmark. Figure Figure I.Figure 1 shows convergence time and number of iterations against network size for a selection of empirical networks spanning the whole range of sizes, and Table I.Table 1 summarises convergence rate, time and iterations of the three solvers for each model on 45 networks (Twitter crisis was not benchmarked), from ∼102\sim10^2 to ∼4×108\sim4 \times10^8 dyads. In the benchmark a run stops when the maximum relative error is below 10−310^{-3} or after 5000 iterations. The coordinate solver converges for every model and network; fixed point is the fastest when it converges, while quasi-Newton is the least reliable for the P-BiECM and the qBiECM (75% and 84% of the networks, respectively). Wall-clock time grows roughly linearly with the number of dyads for all solvers, and the BiECM and the qBiECM are one to two orders of magnitude slower than the BiWCM at fixed size.

Figure I.1. Computational performance of the fitting solvers. a–c, Convergence time against the number of nodes N1+N2N_1 + N_2 for the BiWCM, the BiECM, the P-BiECM and the BiCM, with the coordinate (a), fixed-point (b) and quasi-Newton (c) solvers. d,e, Number of iterations (d) and convergence time (e) of the BiECM for the three solvers. In each panel the networks are selected among those for which all the plotted model–solver pairs are available, taking up to two networks per logarithmic interval of N1+N2N_1 + N_2 (10 intervals). Statistics over all networks are in Table I.Table 1.

Numerical uncertainty of the likelihood. For a canonical model with constraints CC, parameters θ\theta and log-likelihood L(θ)=−θTC∗−ln⁡Z(θ)\mathcal{L}(\theta) = -\theta^{\mathsf{T}} C^* - \ln Z(\theta), the score at an approximate solution is ∇L=−[C∗−Eθ(C)]\nabla\mathcal{L} = -[C^* - \mathbb{E}_{\theta}(C)]. A parameter perturbation δθ\delta\theta therefore changes the likelihood to first order by δL≃−[C∗−Eθ(C)]Tδθ\delta\mathcal{L} \simeq-[C^* - \mathbb{E}_{\theta}(C)]^{\mathsf{T}}\delta\theta. Close to the optimum the Hessian is minus the covariance matrix of the constraints, so the remaining likelihood gap can be estimated from the constraint residuals and the inverse covariance (with a pseudoinverse along the gauge directions). This estimate sets the scale below which likelihood or BIC differences should not be interpreted.

ModelSolverConverged (%)Time (s) medianTime (s) maxIterations medianIterations max
BiCMFixed point1000.0720.69491316
Quasi-Newton1008.4×10−38.4 \times10^{-3}22.72557
Coordinate1000.11229614
BiWCMFixed point980.0198.541599
Quasi-Newton876.1×10−36.1 \times10^{-3}92.24390
Coordinate1000.2882592487
P-BiECMFixed point960.032741891797
Quasi-Newton730.0248050157
Coordinate1000.475572524225
BiECMFixed point930.1047.95941790
Quasi-Newton1000.31132924483746
Coordinate1001.14115526188
qBiECMFixed point890.0620010843365
Quasi-Newton820.0227791021393
Coordinate1000.34247395221511

Table I.1. Convergence and computational cost of the solvers. Each model is fitted with the three solvers on the 45 empirical networks of Fig. I.Figure 1 (Twitter crisis excluded). A run is converged when the maximum relative error on the constraints falls below 10−310^{-3} within 5000 iterations. Times and iterations refer to converged runs (single CPU run per network and solver).

References

  1. [1]Stephen P. Borgatti and Martin G. Everett. Network analysis of 2-mode data. Social Networks, 19(3):243–269, August 1997. ISSN 0378-8733. doi: 10.1016/S0378-8733(96)00301-2. URL http://dx.doi.org/10.1016/S0378-8733(96)00301-2.
  2. [2]Matthieu Latapy, Clémence Magnien, and Nathalie Del Vecchio. Basic notions for the analysis of large two-mode networks. Social Networks, 30(1):31–48, January 2008. ISSN 0378-8733. doi: 10.1016/j.socnet.2007.04.006. URL http://dx.doi.org/10.1016/j.socnet.2007.04.006.
  3. [3]César A. Hidalgo and Ricardo Hausmann. The building blocks of economic complexity. Proceedings of the National Academy of Sciences, 106(26):10570–10575, 2009. doi: 10.1073/pnas.0900943106. URL https://www.pnas.org/doi/abs/10.1073/pnas.0900943106.DOI
  4. [4]Andrea Tacchella, Matthieu Cristelli, Guido Caldarelli, Andrea Gabrielli, and Luciano Pietronero. A new metrics for countries’ fitness and products’ complexity. Scientific reports, 2:723, 10 2012. doi: 10.1038/srep00723.DOI
  5. [5]Enrico L. Rezende, Jessica E. Lavabre, Paulo R. Guimarães, Pedro Jordano, and Jordi Bascompte. Non-random coextinctions in phylogenetically structured mutualistic networks. Nature, 448(7156):925–928, August 2007. ISSN 1476-4687. doi: 10.1038/nature05956. URL http://dx.doi.org/10.1038/nature05956.
  6. [6]Samir Suweis, Filippo Simini, Jayanth R. Banavar, and Amos Maritan. Emergence of structural and dynamical properties of ecological mutualistic networks. Nature, 500(7463): 449–452, August 2013. ISSN 1476-4687. doi: 10.1038/nature12438. URL http://dx.doi.org/10.1038/nature12438.
  7. [7]Carlos Gracia-Lázaro, Laura Hernández, Javier Borge-Holthoefer, and Yamir Moreno. The joint influence of competition and mutualism on the biodiversity of mutualistic ecosystems. Sci. Rep., 8(1), dec 2018. ISSN 20452322. doi: 10.1038/s41598-018-27498-8.
  8. [8]Demival Vasques Filho and Dion R. J. O’Neale. Transitivity and degree assortativity explained: The bipartite structure of social networks. Phys. Rev. E, 101:052305, May 2020. doi: 10.1103/PhysRevE.101.052305. URL https://link.aps.org/doi/10.1103/PhysRevE.101.052305.
  9. [9]James Flamino, Alessandro Galeazzi, Stuart Feldman, Michael W. Macy, Brendan Cross, Zhenkun Zhou, Matteo Serafino, Alexandre Bovet, Hernán A. Makse, and Boleslaw K. Szymanski. Political polarization of news media and influencers on twitter in the 2016 and 2020 us presidential elections. Nature Human Behaviour, March 2023. ISSN 2397-3374. doi: 10.1038/s41562-023-01550-8. URL http://dx.doi.org/10.1038/s41562-023-01550-8.
  10. [10]Inderjit S. Dhillon. Co-clustering documents and words using bipartite spectral graph partitioning. In Proceedings of the Seventh ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’01, pages 269–274, New York, NY, USA, 2001. Association for Computing Machinery. ISBN 158113391X. doi: 10.1145/502512.502550. URL https://doi.org/10.1145/502512.502550.
  11. [11]Christopher Andrew Bail. Combining natural language processing and network analysis to examine how advocacy organizations stimulate conversation on social media. Proceedings of the National Academy of Sciences, 113(42):11823–11828, 2016. doi: 10.1073/pnas.1607151113. URL https://www.pnas.org/doi/abs/10.1073/pnas.1607151113.DOI
  12. [12]Martin Gerlach, Tiago P. Peixoto, and Eduardo G. Altmann. A network approach to topic models. Science Advances, 4(7):eaaq1360, 2018. doi: 10.1126/sciadv.aaq1360. URL https://www.science.org/doi/abs/10.1126/sciadv.aaq1360.DOI
  13. [13]Andrea Simonetti, Alessandro Albano, Michele Tumminello, and T. Di Matteo. Statistically validated network for analysing textual data. Applied Network Science, 10(1), February 2025. ISSN 2364-8228. doi: 10.1007/s41109-025-00693-z. URL http://dx.doi.org/10.1007/s41109-025-00693-z.
  14. [14]Juyong Park and M. E J Newman. Statistical mechanics of networks. Physical Review E - Statistical, Nonlinear, and Soft Matter Physics, 70, 2004. ISSN 15393755. doi: 10.1103/PhysRevE.70.066117.DOI
  15. [15]Diego Garlaschelli and Maria I. Loffredo. Maximum likelihood: Extracting unbiased information from complex networks. Phys. Rev. E, 78:015101, Jul 2008. doi: 10.1103/PhysRevE.78.015101. URL https://link.aps.org/doi/10.1103/PhysRevE.78.015101.DOI
  16. [16]Tiziano Squartini and Diego Garlaschelli. Analytical maximum-likelihood method to detect patterns in real networks. New Journal of Physics, 13(8):083001, August 2011. ISSN 1367-2630. doi: 10.1088/1367-2630/13/8/083001. URL http://dx.doi.org/10.1088/1367-2630/13/8/083001.
  17. [17]Giulio Cimini, Tiziano Squartini, Fabio Saracco, Diego Garlaschelli, Andrea Gabrielli, and Guido Caldarelli. The statistical physics of real-world networks. Nature Reviews Physics, 1(1):58–71, Jan 2019. ISSN 2522-5820. doi: 10.1038/s42254-018-0002-6. URL https://doi.org/10.1038/s42254-018-0002-6.
  18. [18]M. Ángeles Serrano, Marián Boguñá, and Romualdo Pastor-Satorras. Correlations in weighted networks. Phys. Rev. E, 74:055101, Nov 2006. doi: 10.1103/PhysRevE.74.055101. URL https://link.aps.org/doi/10.1103/PhysRevE.74.055101.DOI
  19. [19]Pavel N Krivitsky. Exponential-family random graph models for valued networks. Electron J Stat, 6:1100–1128, 2012. ISSN 1935-7524 (Print); 1935-7524 (Electronic); 1935-7524 (Linking). doi: 10.1214/12-EJS696.DOI
  20. [20]O. Sagarra, C. J. Pérez Vicente, and A. Díaz-Guilera. Statistical mechanics of multiedge networks. Physical Review E, 88(6):062806–, 12 2013. doi: 10.1103/PhysRevE.88.062806. URL https://link.aps.org/doi/10.1103/PhysRevE.88.062806.DOI
  21. [21]Rossana Mastrandrea, Tiziano Squartini, Giorgio Fagiolo, and Diego Garlaschelli. Enhanced reconstruction of weighted networks from strengths and degrees. New Journal of Physics, 16 (4):043022, apr 2014. doi: 10.1088/1367-2630/16/4/043022. URL https://dx.doi.org/10.1088/1367-2630/16/4/043022.
  22. [22]Tiziano Squartini, Rossana Mastrandrea, and Diego Garlaschelli. Unbiased sampling of network ensembles. New Journal of Physics, 17(2):023052, February 2015. ISSN 1367-2630. doi: 10.1088/1367-2630/17/2/023052. URL http://dx.doi.org/10.1088/1367-2630/17/2/023052.
  23. [23]Domenico Di Gangi, Fabrizio Lillo, and Davide Pirino. Assessing systemic risk due to fire sales spillover through maximum entropy network reconstruction. Journal of Economic Dynamics and Control, 94:117–141, 2018. ISSN 0165-1889. doi: 10.1016/j.jedc.2018.07.001. URL http://dx.doi.org/10.1016/j.jedc.2018.07.001.
  24. [24]Giulio Cimini, Tiziano Squartini, Diego Garlaschelli, and Andrea Gabrielli. Systemic risk analysis on reconstructed economic and financial networks. Scientific Reports, 5(1):15758, 2015. doi: 10.1038/srep15758. URL https://doi.org/10.1038/srep15758.
  25. [25]Giulio Cimini, Tiziano Squartini, Andrea Gabrielli, and Diego Garlaschelli. Estimating topological properties of weighted networks from limited information. Physical Review E, 92 (4):040802–, 10 2015. doi: 10.1103/PhysRevE.92.040802. URL https://link.aps.org/doi/10.1103/PhysRevE.92.040802.DOI
  26. [26]Tiziano Squartini, Giulio Cimini, Andrea Gabrielli, and Diego Garlaschelli. Network reconstruction via density sampling. Applied Network Science, 2(1):3, 2017. doi: 10.1007/s41109-017-0021-8. URL https://doi.org/10.1007/s41109-017-0021-8.
  27. [27]Andrea Gabrielli, Rossana Mastrandrea, Guido Caldarelli, and Giulio Cimini. Grand canonical ensemble of weighted networks. Phys. Rev. E, 99:030301, Mar 2019. doi: 10.1103/PhysRevE.99.030301. URL https://link.aps.org/doi/10.1103/PhysRevE.99.030301.DOI
  28. [28]Federica Parisi, Tiziano Squartini, and Diego Garlaschelli. A faster horse on a safer trail: generalized inference for the efficient reconstruction of weighted networks. New Journal of Physics, 22(5):053053, May 2020. ISSN 1367-2630. doi: 10.1088/1367-2630/ab74a7. URL http://dx.doi.org/10.1088/1367-2630/ab74a7.
  29. [29]Marzio Di Vece, Diego Garlaschelli, and Tiziano Squartini. Deterministic, quenched, and annealed parameter estimation for heterogeneous network models. Physical Review E, 108 (5), November 2023. ISSN 2470-0053. doi: 10.1103/physreve.108.054301. URL http://dx.doi.org/10.1103/PhysRevE.108.054301.
  30. [30]Nicolò Vallarano, Matteo Bruno, Emiliano Marchese, Giuseppe Trapani, Fabio Saracco, Giulio Cimini, Mario Zanon, and Tiziano Squartini. Fast and scalable likelihood maximization for exponential random graph models with local constraints. Scientific Reports, 11(1):15227, 2021. ISSN 2045-2322. doi: 10.1038/s41598-021-93830-4. URL https://doi.org/10.1038/s41598-021-93830-4.
  31. [31]Fabio Saracco, Riccardo Di Clemente, Andrea Gabrielli, and Tiziano Squartini. Randomizing bipartite networks: the case of the world trade web. Scientific Reports, 5(1):10595, Jun 2015. ISSN 2045-2322. doi: 10.1038/srep10595. URL https://doi.org/10.1038/srep10595.
  32. [32]Stanislao Gualdi, Giulio Cimini, Kevin Primicerio, Riccardo Di Clemente, and Damien Challet. Statistically validated network of portfolio overlaps and systemic risk. Scientific Reports, 6(1):39467, 2016. doi: 10.1038/srep39467. URL https://doi.org/10.1038/srep39467.
  33. [33]Fabio Saracco, Mika J Straka, Riccardo Di Clemente, Andrea Gabrielli, Guido Caldarelli, and Tiziano Squartini. Inferring monopartite projections of bipartite networks: an entropy-based approach. New Journal of Physics, 19(5):053022, may 2017. doi: 10.1088/1367-2630/aa6b38. URL https://dx.doi.org/10.1088/1367-2630/aa6b38.
  34. [34]Giulio Cimini, Alessandro Carra, Luca Ddomenicantonio, and Andrea Zaccaria. Meta-validation of bipartite network projections. Communications Physics, 5(1):76, Apr 2022. ISSN 2399-3650. doi: 10.1038/s42005-022-00856-9. URL https://doi.org/10.1038/s42005-022-00856-9.
  35. [35]Matteo Bruno, Dario Mazzilli, Aurelio Patelli, Tiziano Squartini, and Fabio Saracco. Inferring comparative advantage via entropy maximization. Journal of Physics: Complexity, 4(4):045011, December 2023. ISSN 2632-072X. doi: 10.1088/2632-072X/ad1411. URL https://dx.doi.org/10.1088/2632-072X/ad1411.
  36. [36]Lorenzo Buffa, Dario Mazzilli, Riccardo Piombo, Fabio Saracco, Giulio Cimini, and Aurelio Patelli. Maximum entropy modelling of sub-optimal transport. Communications Physics, 9(1):39, 2025. doi: 10.1038/s42005-025-02468-5. URL https://doi.org/10.1038/s42005-025-02468-5.
  37. [37]G. Glaviano and S. Miccichè. Unbiased randomization of weighted bipartite networks with exact degree and strength sequences. Phys. Rev. E, 114:014312, Jul 2026. doi: 10.1103/rn1f-zsgb. URL https://link.aps.org/doi/10.1103/rn1f-zsgb.DOI
  38. [38]Yoav Benjamini and Yosef Hochberg. Controlling the false discovery rate: A practical and powerful approach to multiple testing. Journal of the Royal Statistical Society: Series B (Methodological), 57(1):289–300, 1995. doi: https://doi.org/10.1111/j.2517-6161.1995.tb02031.x. URL https://rss.onlinelibrary.wiley.com/doi/abs/10.1111/j.2517-6161.1995.tb02031.x.
  39. [39]Tiziano Squartini, Assaf Almog, Guido Caldarelli, Iman van Lelyveld, Diego Garlaschelli, and Giulio Cimini. Enhanced capital-asset pricing model for the reconstruction of bipartite financial networks. Physical Review E, 96(3):032315–, 09 2017. doi: 10.1103/PhysRevE.96.032315. URL https://link.aps.org/doi/10.1103/PhysRevE.96.032315.DOI
  40. [40]Tiziano Squartini, Joey de Mol, Frank den Hollander, and Diego Garlaschelli. Breaking of ensemble equivalence in networks. Phys. Rev. Lett., 115:268701, Dec 2015. doi: 10.1103/PhysRevLett.115.268701. URL https://link.aps.org/doi/10.1103/PhysRevLett.115.268701.DOI
  41. [41]Diego Garlaschelli, Frank den Hollander, and Andrea Roccaverde. Ensemble nonequivalence in random graphs with modular structure. Journal of Physics A: Mathematical and Theoretical, 50(1):015001, November 2016. ISSN 1751-8121. doi: 10.1088/1751-8113/50/1/015001. URL http://dx.doi.org/10.1088/1751-8113/50/1/015001.
  42. [42]Patelli A, Mazzilli D, Sbardella A, Tacchella A, Pietronero L, and Marschinski R. Economic complexity analysis of export prices. Scientific analysis or review KJ-09-24-051-EN-N (online), Luxembourg (Luxembourg), 2024.
  43. [43]Aurelio Patelli, Giulio Cimmini, Emanuele Pugliese, and Andrea Gabrielli. The scientific influence of nations on global scientific and technological development. Journal of Informetrics, 11(4):1229–1237, 2017. ISSN 1751-1577. doi: https://doi.org/10.1016/j.joi.2017.10.005. URL https://www.sciencedirect.com/science/article/pii/S1751157717301335.
  44. [44]University of São Paulo. Interaction web database (IWDB). urlhttp://www.ecologia.ib.usp.br/iwdb/, 2024. Accessed: 2024-06-18.
  45. [45]University of Colorado. Index of complex networks (ICON). urlhttps://icon.colorado.edu, 2024. Accessed: 2024-06-18.
  46. [46]Jason Baumgartner, Savvas Zannettou, Brian Keegan, Megan Squire, and Jeremy Blackburn. The pushshift reddit dataset. https://arxiv.org/abs/2001.08435, 2020.
  47. [47]Daniele Cirulli, Antonio Desiderio, Giulio Cimini, and Fabio Saracco. Polarization and echo chambers in reddit’s political discourse, 2025. URL https://arxiv.org/abs/2510.27467.
  48. [48]Jonida Carungu and Alessia Pat uelli. 599 sustainability reports by UK-listed firms (2015–2023), 8 2026. URL https://figshare.com/articles/dataset/599_sustainability_reports_by_UK-listed_firms_2015_2023_/33339564.DOI
  49. [49]C.E. Bonferroni. Teoria statistica delle classi e calcolo delle probabilità. Pubblicazioni del R. Istituto superiore di scienze economiche e commerciali di Firenze. Seeber, 1936. URL https://books.google.it/books?id=3CY-HQAACAAJ.
  50. [50]Mário Almeida-Neto, Paulo Guimarães, Jr P R Guimaraes, Rafael D. Loyola, and Werner Ulrich. A consistent metric for nestedness analysis in ecological systems: Reconciling concept and measurement. Oikos, 117(8):1227–1239, 2008. ISSN 00301299. doi: 10.1111/j.0030-1299.2008.16644.x.
  51. [51]Rafael B. P. Pinheiro, Gabriel M. F. Felix, Carsten F. Dormann, and Marco A. R. Mello. A new model explaining the origin of different topologies in interaction networks. Ecology, 100 (9):e02796, 2019.
  52. [52]Aric A. Hagberg, Daniel A. Schult, and Pieter J. Swart. Exploring network structure, dynamics, and function using NetworkX. In Proceedings of the 7th Python in Science Conference (SciPy 2008), pages 11–15, 2008.
  53. [53]Tore Opsahl. Triadic closure in two-mode networks: Redefining the global and local clustering coefficients. Social Networks, 35(2):159–167, 2013.
  54. [54]Thomas M. Cover and Joy A. Thomas. Elements of Information Theory. Wiley-Interscience, Hoboken, NJ, 2nd edition, 2006. ISBN 978-0-471-24195-9.
  55. [55]Riccardo Dainelli and Fabio Saracco. Bibliometric and social network analysis on the use of satellite imagery in agriculture: An entropy-based approach. Agronomy, 13(2), 2023. ISSN 2073-4395. doi: 10.3390/agronomy13020576. URL https://www.mdpi.com/2073-4395/13/2/576.
  56. [56]Edward Loper and Steven Bird. Nltk: The natural language toolkit, 2002. URL https://arxiv.org/abs/cs/0205028.
  57. [57]Stefano Guarino, Ayoub Mounim, Guido Caldarelli, and Fabio Saracco. Leveraging content producer networks and user perception to detect online discursive communities. Scientific Reports, 16(1), March 2026. ISSN 2045-2322. doi: 10.1038/s41598-026-39477-5. URL http://dx.doi.org/10.1038/s41598-026-39477-5.
  58. [58]Carolina Beccati, Guido Caldarelli, Renaud Lambiotte, and Fabio Saracco. Extracting significant signal of news consumption from social networks: the case of twitter in italian political elections. Palgrave Communications, 5:1–16, 12 2019. ISSN 20551045. doi: 10.1057/s41599-019-0300-3.
  59. [59]Guido Caldarelli, Oriol Artime, Giulia Fischetti, Stefano Guarino, Andrzej Nowak, Fabio Saracco, Petter Holme, and Manlio De Domenico. The physics of news, rumors, and opinions. Physics Reports, 1186:1–75, August 2026. ISSN 0370-1573. doi: 10.1016/j.physrep.2026.05.002. URL http://dx.doi.org/10.1016/j.physrep.2026.05.002.
  60. [60]Aurelio Patelli, Lorenzo Napolitano, Giulio Cimini, and Andrea Gabrielli. Geography of science: Competitiveness and inequality. Journal of Informetrics, 17(1):101357, 2023. ISSN 1751-1577. doi: https://doi.org/10.1016/j.joi.2022.101357. URL https://www.sciencedirect.com/science/article/pii/S1751157722001109.

Paper details

vlnc:W_uxzjuzes
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents