betaFeedback

WorksMathematics

Sparse Directed Network Recovery From Multivariate Count Time Series

Abstract

Count time series arise naturally in many scientific applications. Examples include species abundances collected over time, and gene expression counts over pseudotime. Network inference from such data can yield important insight into underlying biological mechanisms, but is challenging because count observations are discrete, overdispersed, and temporally dependent. We propose a multivariate Poisson log-normal model in which observed counts arise from a Poisson distribution with rates determined by a latent vector autoregressive (VAR) process. The latent VAR transition matrix induces directed temporal dependencies among components and serves as the target network parameter. We establish identifiability of the model under suitable conditions and develop two procedures for sparse estimation of the VAR transition matrix. Our first estimator is a sparse method-of-moments estimator, for which we establish consistent network recovery and parameter estimation in high-dimensional settings. Our second estimator is an L1-penalized variational inference estimator that is motivated by recent approaches for count data. We evaluate the proposed estimators through simulations under a range of sampling regimes. We further demonstrate the method on a macroinvertebrate abundance time series collected from European freshwater ecosystems, obtaining an interpretable ecological network.

Introduction

Multivariate count time series arise naturally in many biological applications, such as species abundances measured over time and longitudinal gene expression counts from measurements over time. A central goal is to recover an underlying network structure that captures interactions among system components (species in an environment, genes in a cell). However, unlike continuous-valued time series data, which have a well-developed modeling literature [21], multivariate count time series have a less unified modeling framework. Their analysis is often complicated by over-dispersion in addition to the dependence among components and the serial dependence over time inherent to multivariate time series. Related work includes network inference for categorical time series [33] and for Poisson and Bernoulli autoregressive models [16]. Existing work provides several approaches to the problem, reviewed in Davis et al. [11]. Early work developed integer-valued autoregressive models for univariate counts [23, 2], later extended to the multivariate setting by Latour [19]. However, these models do not admit an intuitive interpretation for network inference and are limited in their ability to capture general autocorrelation structures. Recent work addressed some of these limitations. For example, Yu and Zheng (2024) allows for flexible correlation in the noise component, but still does not allow one component to depend on past values of another. Fokianos et al. [14] developed multivariate count autoregressive models in which past observed counts directly influence future conditional means, with contemporaneous dependence captured through a copula. We instead treat counts as noisy observations of an underlying process whose dynamics define the target network. This separates temporal interactions in the latent log-rates from Poisson sampling variation.

To address the limitations of existing approaches, we adopt a multivariate Poisson log-normal (PLN) model [1], in which observed counts follow a Poisson distribution with rates driven by a latent Gaussian vector autoregressive (VAR) process. This model naturally accommodates count-valued data, over-dispersion, and dependence both within and across time points. Moreover, the transition matrix of the latent VAR process encodes directed temporal dependencies among components corresponding to Granger causal interactions [29], so its support can be interpreted as the underlying network structure.

Our proposed model is most closely related to Zhang et al. [40], who modeled multivariate Poisson counts using latent Gaussian VAR dynamics and estimated the parameters by Monte Carlo EM, with numerical studies primarily in low-dimensional settings. We focus on sparse recovery of the directed network encoded by the latent transition matrix, developing computationally scalable estimators with support-recovery guarantees. Our framework also accommodates covariates, multiple independent trajectories of unequal length, and ℓ1\ell_1-penalized estimation for sparse recovery. For network inference, our approach is closer to Chiquet et al. [8], who used a PLN model to infer an undirected network via sparsity in the latent precision matrix. Recent extensions of the PLN framework include classification, clustering, and dimension reduction (see Chiquet et al. [7] for an overview), as well as tree-structured PLN models for hierarchical count data [6]. Saha et al. [27] introduced the Poisson log-normal process, which extends the PLN framework by placing a Gaussian process prior on the Poisson log-rates for count prediction.

Many existing PLN methods rely on variational inference (VI), estimating model parameters by maximizing the evidence lower bound (ELBO) [5]. More recent work has proposed likelihood-based alternatives, including maximum marginal likelihood estimation [18], motivated by composite likelihood ideas [10, 35]. Relatedly, Stoehr and Robin [31] combine composite likelihood with importance sampling. Full likelihood-based estimation requires integration over the latent trajectories, making computation challenging as the number of components and time points increases. Marginal likelihood approaches, such as Huang et al. [18], reduce this burden by using low-dimensional marginal distributions, but still require numerical integration and optimization.

We propose a method-of-moments (MoM) estimator as well as its l1l_{1}-penalized version to encourage sparsity in high dimensions. Our penalized MoM estimator is related to Düker et al. [13], who studied sparse latent VAR estimation for count time series. Their model generates counts by thresholding latent Gaussian variables, whereas ours uses a Poisson observation model conditional on the latent Gaussian states. Our proposed estimator scales well in high-dimensional settings and has theoretical guarantees on recovering the support of the network of interest under usual conditions for sparse estimation. We also develop a VI estimator and penalized VI estimator to compare our method to the prevailing estimation strategy used by PLN models in the literature.

The main methodological contributions of this work are twofold. First, we formulate network recovery for multivariate count time series under a dynamic PLN model in which the support of the latent VAR transition matrix defines the target directed network. Second, we develop an estimation procedure for sparse network recovery along with theoretical guarantees. We illustrate the practical utility of our method using macroinvertebrate species abundance data from Danish freshwater environments in the TREAM (Time seRies of European freshwAter Macroinvertebrates) dataset [37].

Methodology

Model and network target

We consider nn independent realizations of a multivariate count time series, where ii indexes the realization and the timepoints for each realization are indexed along a grid of equally spaced timepoints 1,2,…,Ti1, 2, \ldots, T_i, where TiT_i is the latest timepoint at which the realization is observed. At time tt, let Yt(i)∈Z≥0dY_t^{(i)} \in\mathbb{Z}_{\geq0}^d, denote the vector of counts for dd components, Xt(i)∈RpX_t^{(i)} \in\mathbb{R}^p, the vector of observed covariates (excluding an intercept term), and Ot(i)∈RO_t^{(i)} \in\mathbb{R} an offset. Let Zt(i)∈RdZ_t^{(i)} \in\mathbb{R}^d denote the latent state for sample ii at time tt and let ut(i)u_t^{(i)} denote the corresponding noise term. Further, let Σ∈Rd×d\boldsymbol{\Sigma} \in\mathbb{R}^{d \times d} be a positive definite covariance matrix, A∈Rd×dA \in\mathbb{R}^{d \times d} the latent transition matrix, β0∈Rd\beta_0 \in\mathbb{R}^d the intercept vector, and β∈Rp×d\beta\in\mathbb{R}^{p \times d} the matrix of regression coefficients. Since higher-order VAR processes can be written in “companion form” as VAR(1) processes, for simplicity, we focus on a latent VAR(1) model [21]. Specifically, we assume

Zt(i)=AZt−1(i)+ut(i),ut(i)∼i.i.d.N(0,Σ)(1)Z_t^{(i)} = A Z_{t-1}^{(i)} + u_t^{(i)}, \qquad u_t^{(i)} \mathrel{\mathop{\sim}^{\mathrm{i.i.d.}}} N(0,\boldsymbol{\Sigma}) \tag*{(1)}
Ytj(i)∣Ztj(i)∼Pois.⁡{exp⁡(Ot(i)+β0j+Xt(i)⊤βj+Ztj(i))}(2)Y_{tj}^{(i)} \mid Z_{tj}^{(i)} \sim\operatorname{Pois.}\left\{\exp\left(O_t^{(i)} + \beta_{0j} + X_t^{(i)\top}\beta_j + Z_{tj}^{(i)}\right)\right\} \tag*{(2)}

We additionally assume that, conditional on the latent trajectory, covariates, and offsets, the counts are independent across time points and components. Suppressing the conditioning on covariates and offsets, the joint distribution of the latent states and observed counts factorizes as

p(y1:T,z1:T)=p(z1)∏t=2Tp(zt∣zt−1)∏t=1T∏j=1dp(ytj∣ztj).p(y_{1:T}, z_{1:T}) = p(z_1) \prod_{t=2}^{T} p(z_t \mid z_{t-1}) \prod_{t=1}^{T} \prod_{j=1}^{d} p(y_{tj} \mid z_{tj}).

We assume that the nn trajectories are independent and follow this model. Further, we assume all eigenvalues of AA have modulus less than 1 and that the latent process is stationary. For each trajectory, the latent process is independent of the covariates and offsets. Letting ΣZ∈Rd×d\boldsymbol{\Sigma}_Z \in\mathbb{R}^{d \times d} denote the stationary covariance matrix of ZtZ_t, we have ΣZ=AΣZA⊤+Σ\boldsymbol{\Sigma}_Z = A\boldsymbol{\Sigma}_Z A^\top+ \boldsymbol{\Sigma} [21].

Our target of inference is the network of Granger-causal interactions among the components of the latent process ZtZ_t, as defined in Shojaie and Fox [29]. In the proposed model, AjkA_{jk} represents the lagged effect of latent component kk on latent component jj, and Ajk≠0A_{jk} \ne0 corresponds to a directed edge from component kk to component jj.

Identifiability

Under conditions on the design matrix of observed covariates and the covariance of the latent VAR process, the proposed model is identifiable.

Lemma 1. Let Ntot:=∑i=1nTiN_{\mathrm{tot}} := \sum_{i=1}^{n} T_i, and let X∈RNtot×p\mathbf{X} \in\mathbb{R}^{N_{\mathrm{tot}} \times p} denote the matrix of observed covariates obtained by stacking Xt(i)\mathbf{X}_{t}^{(i)} over all samples and time points. If the augmented design X‾=[1,X]\overline{\mathbf{X}} = [\mathbf{1}, \mathbf{X}] has full column rank, and ΣZ\boldsymbol{\Sigma}_{Z} is invertible, then the model is identifiable.

Proof. The proof is given in Appendix A.

We next describe the two proposed estimation procedures for the model parameters. We start with the MoM estimator and then discuss the alternative VI estimator.

Method of Moments Estimator

We develop a method-of-moments estimator for the model parameters θ:={A,β0,β,Σ}\theta:= \{A, \beta_{0}, \beta, \Sigma\}, with primary interest in AA, whose support determines the directed network.

We first define some notation. Let ∥⋅∥1\lVert\cdot\rVert_{1}, ∥⋅∥2\lVert\cdot\rVert_{2}, and ∥⋅∥∞\lVert\cdot\rVert_{\infty} denote the usual ℓ1\ell_{1}, Euclidean, and maximum norms of a vector. For a matrix MM,

∥M∥max⁡=max⁡j,k∣Mjk∣,∥M∥1=max⁡k∑j∣Mjk∣,∥M∥∞=max⁡j∑k∣Mjk∣.\lVert M \rVert_{\max} = \max_{j,k} |M_{jk}|,\qquad\lVert M \rVert_{1} = \max_{k} \sum_{j} |M_{jk}|,\qquad\lVert M \rVert_{\infty} = \max_{j} \sum_{k} |M_{jk}|.

We write ∥M∥2\lVert M \rVert_{2} for its spectral norm, and ∥M∥F\lVert M \rVert_{F} for its Frobenius norm.

We begin by estimating β\beta. Let γ∈R(p+1)×d\gamma\in\mathbb{R}^{(p+1)\times d} be defined columnwise by γj:=[β0j+12ΣZ,jj,β1j,…,βpj]⊤∈Rp+1\gamma_{j} := [\beta_{0j} + \frac{1}{2}\Sigma_{Z,jj}, \beta_{1j}, \ldots, \beta_{pj}]^{\top} \in\mathbb{R}^{p+1}, and index the rows of γ\gamma from 0 to pp. Since γ1:p=β\gamma_{1:p} = \beta, estimating γ\gamma yields an estimator of β\beta. Moreover, E[Ytj(i)∣Xt(i),Ot(i)]=exp⁡{Ot(i)+γ0j+Xt(i)⊤γ1:p,j}\mathbb{E}[Y_{tj}^{(i)} \mid X_{t}^{(i)}, O_{t}^{(i)}] = \exp\{O_{t}^{(i)} + \gamma_{0j} + X_{t}^{(i)\top}\gamma_{1:p,j}\}, so γ\gamma can be estimated from the first moment by fitting a Poisson regression. Concretely, we define the objective Lj(g)L_j(g) for g=[g0 g1:p]⊤∈Rp+1g = [g_0\ g_{1:p}]^{\top} \in\mathbb{R}^{p+1} as

Lj(g)=−1Ntot∑i=1n∑t=1Ti{Ytj(i)(Ot(i)+g0+Xt(i)⊤g1:p)−exp⁡(Ot(i)+g0+Xt(i)⊤g1:p)},L_j(g) = -\frac{1}{N_{\mathrm{tot}}}\sum_{i=1}^{n}\sum_{t=1}^{T_i}\left\{Y_{tj}^{(i)}\left(O_t^{(i)} + g_0 + X_t^{(i)\top}g_{1:p}\right) - \exp\left(O_t^{(i)} + g_0 + X_t^{(i)\top}g_{1:p}\right)\right\},

and obtain estimates for γj\gamma_j (j=1,…,dj = 1, \ldots, d) as γ^j=arg⁡min⁡gLj(g)\hat{\gamma}_j = \arg\min_g L_j(g).

References

  1. [1]Aitchison, J. and Ho, C. H. (1989). The multivariate poisson-log normal distribution. Biometrika 76, 643–653.
  2. [2]Al-Osh, M. A. and Alzaid, A. A. (1987). First-order integer-valued autoregressive (inar(1)) process. Journal of Time Series Analysis 8, 261–275.
  3. [3]Beck, A. and Teboulle, M. (2009). A fast iterative shrinkage-thresholding algorithm for linear inverse problems. SIAM Journal on Imaging Sciences 2, 183–202.DOI
  4. [4]Belloni, A. and Chernozhukov, V. (2013). Least squares after model selection in high-dimensional sparse models. Bernoulli 19, 521–547.
  5. [5]Blei, D. M., Kucukelbir, A., and McAuliffe, J. D. (2017). Variational inference: A review for statisticians. Journal of the American Statistical Association 112, 859–877.
  6. [6]Chaussard, A., Bonnet, A., Gassiat, E., and Corff, S. L. (2025). Tree-based variational inference for poisson log-normal models.arxiv.org/abs/2406.17361
  7. [7]Chiquet, J., Mariadassou, M., and Robin, S. (2021). The poisson-lognormal model as a versatile framework for the joint analysis of species abundances. Frontiers in Ecology and Evolution Volume 9 - 2021,.DOI
  8. [8]Chiquet, J., Robin, S., and Mariadassou, M. (2019). Variational inference for sparse network reconstruction from count data. In Chaudhuri, K. and Salakhutdinov, R., editors, Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pages 1162–1171. PMLR.
  9. [9]Cortés-Guzmán, D., Bowler, D. E., Forio, M. A. E., Goethals, P., Karaouzas, I., Moulinec, A., Sinclair, J. S., Vannevel, R., Haase, P., and Welti, E. A. R. (2026). Recovering european river invertebrate communities homogenize or differentiate depending on anthropogenic stress. Global Change Biology 32, e70716. e70716 3092046.DOI
  10. [10]Cox, D. R. and Reid, N. (2004). A note on pseudolikelihood constructed from marginal densities. Biometrika 91, 729–737.DOI
  11. [11]Davis, R. A., Fokianos, K., Holan, S. H., Joe, H., Livsey, J., Lund, R., Pipiras, V., and Ravishanker, N. (2021). Count time series: A methodological review. Journal of the American Statistical Association 116, 1533–1547.DOI
  12. [12]Dempster, A. P., Laird, N. M., and Rubin, D. B. (1977). Maximum likelihood from incomplete data via the em algorithm. Journal of the Royal Statistical Society: Series B (Methodological) 39, 1–22.DOI
  13. [13]Düker, M.-C., Lund, R., and Pipiras, V. (2024). High-dimensional latent Gaussian count time series: Concentration results for autocovariances and applications. Electronic Journal of Statistics 18, 5484 – 5562.arxiv.org/abs/2301.00491
  14. [14]Fokianos, K., STØVE, B., TJØSTHEIM, D., and DOUKHAN, P. (2020). Multivariate count autoregression. Bernoulli 26, pp. 471–499.
  15. [15]Haase, P., Bowler, D., Baker, N., Bonada, N., Domisch, S., García Márquez, J., Heino, J., Hering, D., Jähnig, S., Schmidt-Kloiber, A., Stubbington, R., Altermatt, F., Álvarez Cabria, M., Amatulli, G., Angeler, D., Archambau​d, G., Jorrín, I., Aspin, T., Azpiroz, I., and Welti, E. (2023). The recovery of european freshwater biodiversity has come to a halt. Nature 620, 1–7.
  16. [16]Hall, E. C., Raskutti, G., and Willett, R. (2016). Inference of high-dimensional autoregressive generalized linear models. arXiv preprint arXiv:1605.02693 .arxiv.org/abs/1605.02693
  17. [17]Holt, E. and Miller, S. (2011). Bioindicators: Using organisms to measure. Nature 3, 8–13.
  18. [18]Huang, Y., Wang, C., Tang, J., Wu, W., and Xi, R. (2025). A generic family of graphical models: Diversity, efficiency, and heterogeneity. In Singh, A., Fazel, M., Hsu, D., Lacoste-Julien, S., Berkenkamp, F., Maharaj, T., Wagstaff, K., and Zhu, J., editors, Proceedings of the 42nd International Conference on Machine Learning, volume 267 of Proceedings of Machine Learning Research, pages 25876–25907. PMLR.
  19. [19]Latour, A. (1997). The multivariate ginar(p) process. Advances in Applied Probability 29, 228–248.
  20. [21]Lütkepohl, H. (2005). New Introduction to Multiple Time Series Analysis. Spring Berlin, Heidelberg.
  21. [22]Matthews, B. (1975). Comparison of the predicted and observed secondary structure of t4 phage lysozyme. Biochimica et Biophysica Acta (BBA) - Protein Structure 405, 442–451.pubmed.ncbi.nlm.nih.gov/1180967
  22. [23]McKenzie, E. (1985). Some simple models for discrete variate time series. JAWRA Journal of the American Water Resources Association 21, 645–650.
  23. [24]Meinshausen, N. and Bühlmann, P. (2010). Stability selection. Journal of the Royal Statistical Society Series B: Statistical Methodology 72, 417–473.
  24. [25]Nagaev, S. V. (1979). Large deviations of sums of independent random variables. The Annals of Probability pages 745–789.DOI
  25. [26]Neykov, M., Ning, Y., Liu, J. S., and Liu, H. (2018). A unified theory of confidence regions and testing for high-dimensional estimating equations. Statistical Science 33, 427–443.arxiv.org/abs/1510.08986
  26. [27]Saha, A., Gandrakota, A., and Morozov, A. V. (2026). Poisson log-normal process for count data prediction.arxiv.org/abs/2602.05926
  27. [28]Sexton, A., Beisel, J.-N., Staentzel, C., Wolter, C., Tales, E., Jérôme, B., Buijse, A. D., Martínez Fernández, V., Wantzen, K., Jähnig, S., Garcia de Leaniz, C., Schmidt-Kloiber, A., Haase, P., Forio, M., Archambaud, G., Fruget, J.-F., Dohet, A., Evtimova, V., Csabai, Z., and Jeliazkov, A. (2024). Inland navigation and land use interact to impact european freshwater biodiversity. Nature Ecology & Evolution 8, 1–11.DOI
  28. [29]Shojaie, A. and Fox, E. B. (2022). Granger causality: A review and recent advances. Annual review of statistics and its application 9, 289–319.arxiv.org/abs/2105.02675
  29. [30]Sinclair, J., Welti, E., Altermatt, F., Álvarez Cabría, M., Arovita, J., Baker, N., Barešová, L., Barquín, P., Bonacina, L., Cañedo-Argüelles, M., Csabai, Z., Eyto, E., Dohet, A., Dörflinger, G., Eriksen, T., Evtimova, V., Feio, M., Ferréol, M., and Haase, P. (2024). Multi-decadal improvements in the ecological quality of european rivers are not consistently reflected in biodiversity metrics. Nature Ecology & Evolution 8, 1–12.DOI
  30. [31]Stoehr, J. and Robin, S. (2026). Composite likelihood inference for the poisson log-normal model: Composite likelihood inference for the poisson log-normal model. Statistics and Computing 36,.DOI
  31. [32]Tank, A., Fox, E. B., and Shojaie, A. (2019). Identifiability and estimation of structural vector autoregressive models for subsampled and mixed-frequency time series. Biometrika 106, 433–452.DOI
  32. [33]Tank, A., Li, X., Fox, E. B., and Shojaie, A. (2021). The convex mixture distribution: Granger causality for categorical time series. SIAM journal on mathematics of data science 3, 83–112.DOI
  33. [34]Toman, M. J. and Dall, P. C. (1997). The diet of erpobdella octoculata (hirudinea: Erpobdellidae) in two danish lowland streams. Archiv für Hydrobiologie 140, 549–563.DOI
  34. [35]Varin, C. (2008). On composite marginal likelihoods. AStA Advances in Statistical Analysis 92, 1–28.DOI
  35. [36]Wainwright, M. J. (2009). Sharp thresholds for high-dimensional and noisy sparsity recovery using ℓ₁-constrained quadratic programming (lasso). IEEE Trans. Info. Theory 55, 2183–2202.
  36. [37]Welti, E. A., Bowler, D. E., Sinclair, J. S., Altermatt, F., Álvarez Cabría, M., Amatulli, G., Angeler, D. G., Archambaud, G., Arrate Jorrín, I., Aspin, T., and et al. (2024). Time series of freshwater macroinvertebrate abundances and site characteristics of european streams and rivers. Scientific Data 11,.DOI
  37. [38]Xia, L. and Shojaie, A. (2026). Statistical inference for high-dimensional generalized estimating equations. Biostatistics 27, kxag013.DOI
  38. [40]Zhang, C., Chen, N., and Li, Z. (2017). State space modeling of autocorrelated multivariate poisson counts. IISE Transactions 49, 518–531.DOI
  39. [41]Zieritz, A., Brian, J. I., Sousa, R., Aldridge, D. C., Atkinson, C. L., Douda, K., Vaughn, C., Bespalaya, Y., Richmond, T., Čmiel, A. M., Crisp, A., Dobler, A. H., Ercoli, F., Esteves, E., Ferreira-Rodríguez, N., Geist, J., González, I. S., Halabowski, D., Hoos, P., Hopper, G. W., Hyvärinen, H., Ilarri, M., Lewin, I., Lipińska, A. M., Mageroy, J. H., Nizzoli, D., Ollard, I., Österling, M., Riccardi, N., Rock, S. L., Sjönberg, T., Taskinen, J., Urbanič, G., Urbańska, M., Yu, Q., and Vaz, A. S. (2025). A global meta-analysis of ecological functions and regulating ecosystem services of freshwater bivalves. Limnology and Oceanography 70, 2982–2997.DOI

Paper details

vlnc:W_9k44p2g3
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 1

downloads · 0

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

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents