In this paper
AbstractIntroductionBackground and Related WorkHydra: A Multi-task, Transfer Learning ModelProblem DefinitionHydra: Shared Trunk with Task-Specific HeadsSpectral Path for Global ConnectivitySpatial Path for Local ConnectivityPrediction HeadEnd-to-End TrainingExperimental EvaluationTraining Setup and AlgorithmClassification and Regression ResultsHydra Generalization and EfficiencyImpact of Multi-Task TrainingConclusionAcknowledgmentsAppendixLimitationsBroader ImpactComputational Complexity of HydraDatasetBaselinesTask FormalizationsClassification TasksRegression TasksImportance of Temporal Property PredictionEvaluation MetricsAblation StudiesTransfer to Social Interaction NetworksFurther Analysis of Scaling and Model CapacityScaling Behavior in HydraExtended Network and Task ResultsReferencesWorksComputer Science
Hydra: Towards Transferable Multi-Task Learning on Temporal Graphs
Abstract
Real-world evolving networks are naturally modeled as temporal graphs (TGs), where capturing temporal dynamics is essential for predicting future graph properties that support downstream decision-making. Existing temporal graph methods have been developed primarily for single-task prediction, and little is known about their generalization across tasks or transfer to unseen networks. This leaves the challenge of multi-task graph property prediction in TGs largely open. We address this challenge by introducing Hydra, a novel architecture that integrates local connectivity features from temporal GNNs with a spectral learning module that captures global connectivity patterns. This design enables joint learning of local and global information under a multi-task objective. In multi-task classification, Hydra achieves an 8.9% relative gain in AUC over the strongest competitor. In multi-task regression, Hydra achieves competitive results in all three tasks, while obtaining the best results in two tasks with a 8.2% relative gain in MAE compared to the strongest baseline. Moreover, Hydra delivers these gains with a 22% reduction in training time compared to temporal transfer models. These results provide the first systematic evidence that multi-task transferable learning on temporal graphs is effective. By delivering consistent top-ranked performance, Hydra highlights multi-task training on temporal graphs as a promising direction toward adaptable foundation models for temporal graphs.
Introduction
Temporal graph prediction addresses the challenge of forecasting the future state and the trajectory of networks that evolve over time and possess an underlying relational structure. Unlike traditional graph analysis that assumes a static topology, temporal graphs are characterized by continuously changing nodes, edges, and features, representing dynamic real-world phenomena such as social network interactions [49, 16], traffic flow [38, 29], and biological processes [12]. The core difficulty lies in capturing both the complex spatial dependencies encoded in the graph structure and the temporal dynamics that govern its evolution [19]. Effective solutions must therefore move beyond static graph representation learning to models that can jointly capture spatio-temporal patterns, enabling accurate prediction of future interactions, node properties, and system-wide events. This capability underpins applications such as disease outbreak detection [40], fraud prevention [11], recommendation systems [45], and transportation network planning [6, 46]. Despite progress in temporal graph learning, two fundamental challenges remain open.
Multi-task Temporal Graph Prediction. The first challenge is multi-task temporal graph property prediction. Existing temporal graph methods commonly focus on one prediction objective at a time, while many real applications require forecasting multiple graph-level properties from the same evolving network. For example, one may need to predict both community evolution and connectivity growth. By learning these tasks jointly, models can exploit the shared spatio-temporal dynamics that shape them, yielding more robust generalization and a deeper understanding of system-wide behavior. Multi-task temporal graph prediction requires representations that transfer across heterogeneous graphs, remain stable under temporal drift, and still capture task-specific signals, which existing temporal GNNs and graph-level baselines cannot provide. Our framework overcomes these limits by using a shared trunk that fuses global structural dynamics with local temporal information to support robust multi-graph, multi-task prediction.
Transferability Across Networks. The second challenge is transferability across networks. In real-world dynamic settings, new networks constantly emerge, data may be sparse or incomplete, and training from scratch is costly. A practical solution is to pre-train temporal graph models on multiple observed networks and transfer them to unseen ones without retraining [24, 37, 48, 33]. This requires learning representations that capture generalizable temporal patterns while scaling effectively with data size and model capacity. Unlike static graphs, temporal graphs add the complexity of evolving structures and features, which makes transfer and scaling more difficult. Yet, no prior work examines how to build a representation trunk that captures cross-network temporal invariances while supporting multiple tasks through flexible, lightweight heads. The tasks we study match real system needs: forecasting structural stability, connectivity growth, and shrinkage, which are standard aggregate indicators used in risk analysis and network monitoring.
In this work, we address both challenges by introducing HYDRA, a novel temporal multi-task graph neural network (GNN) designed for graph-level multi-property predictions with strong transferability across diverse networks. To capture global structural information, HYDRA leverages graph Laplacian descriptors, while a Temporal Graph Neural Network (TGNN) models local node interactions and temporal dynamics with a self-attention pooling layer.
This trunk–head design separates a shared spectral–spatio–temporal backbone that captures cross-network, task-invariant dynamics from lightweight task-specific heads that flexibly support heterogeneous objectives. The resulting spectral and spatio-temporal representations are fused and passed through task-specific prediction heads (Figure 1). Trained end-to-end in a multi-task setting, HYDRA achieves state-of-the-art performance on six benchmark tasks spanning classification and regression.
Figure 1. HYDRA overview. HYDRA is trained once on multiple networks and, without additional training, can be directly applied to unseen networks to predict multiple properties simultaneously.
Significantly, it generalizes across unseen networks, consistently outperforming models trained from scratch on those networks. We also analyze its scaling behavior, showing that performance can improve as more training networks are incorporated.
Our contributions are summarized as follows:
First study on multi-task learning for temporal graphs. We introduce the first model that jointly handles multiple temporal graph prediction tasks without retraining, achieving state-of-the-art performance while addressing task interference and cross-network invariance.
Novel integration of TGNNs with spectral methods. We incorporate Laplacian descriptors into TGNNs for graph-level prediction, enabling cross-network transfer by capturing structural invariances absent from purely temporal or purely spectral models.
Comprehensive study of transfer and scaling. Hydra achieves an 8.9% AUC gain in classification and 8.2% MAE gain in two regression tasks over the strongest baselines while remaining competitive on the third task without any task-specific training. Moreover, transferability to unseen networks can improve as the number of training datasets increases.
Background and Related Work
Temporal Graph Property Prediction. Graph property prediction has been widely studied in the static setting, where various methods [26, 54, 53] have shown strong performance in tasks such as molecular property prediction [44, 51] and social network analysis [4]. Building upon these advances, temporal graph property prediction has recently received increasing attention. One pioneering approach to extend static models to temporal settings is GraphPulse [41], which combines recurrent architectures with topological data analysis tools to capture temporal patterns. However, existing work has been limited to training on a single network [15, 13], except for MiNT [42], which trains across multiple networks and demonstrates transferability to unseen networks during inference. However, MiNT is limited to a single classification task, and its computational cost is high because of its hyperbolic backbone. In this work, we go beyond this limitation by addressing multiple classification and regression tasks in a scalable manner.
Multi-Task Learning. Multi-task learning has been extensively studied as a means to improve generalization by jointly optimizing related tasks and leveraging shared information [3, 36, 57]. In graph learning, several approaches integrate learning strategies to enhance generalization, achieving strong performance across diverse datasets and downstream tasks [17, 18, 22, 28]. However, while these advances have driven substantial progress in static graph learning, applying multi-task learning to temporal graphs remains unexplored.
Temporal Graph Transferability. Recent advances in temporal graph learning have produced models tailored for tasks such as link prediction [34, 14, 8, 43]. While these methods achieve strong results on specific tasks and datasets, their applicability is often narrow: models trained for one task or on one network may fail to generalize to new settings. Unlike static graphs, temporal graphs introduce additional transfer challenges, including evolving topologies, distribution shifts, and non-stationary dynamics. These factors make it difficult for a model to retain predictive power outside its training environment. Addressing transferability is important for real-world use cases where networks are constantly changing, data collection is costly, and retraining for every task or dataset is impractical [47, 15, 32]. A central open problem is to design models that capture generalizable temporal patterns so that knowledge gained from past networks or tasks can be reused in unseen ones [30, 25]. Developing such models would enable scalable, efficient, and continuous prediction in dynamic environments. Hydra uses supervised multi-task pre-training with direct zero-shot prediction, rather than self-supervised pre-training followed by target-specific fine-tuning. We view it as a step toward temporal graph foundation models within the evaluated tasks and domains.
Hydra: A Multi-task, Transfer Learning Model
Problem Definition
We define a discrete-time temporal graph as a sequence of snapshots , where and denote the node and edge sets at time , and represents associated node or edge features. In this work, we consider a collection of temporal graphs , each evolving independently over time. For each temporal graph we associate task-specific labels for prediction tasks.
Definition 1 (Multi-Task Temporal Graph Property Prediction). The objective is to learn a model
that ingests all snapshots of all graphs in up to time and jointly predicts graph-level properties at time for each graph. The parameters are optimized jointly across , and the trained model is required to generalize to temporal graphs not included in .
Hydra is evaluated on two categories of prediction tasks across multiple temporal graphs. In the classification setting, for , the objective is to predict binary outcomes indicating the directional change of graph-level properties between consecutive snapshots. In the regression setting, for , the model predicts numerical values that quantify the magnitude of graph-level properties at the next snapshot. In principle, this division allows Hydra to handle both discrete and continuous objectives within a single unified architecture.
Multi-task learning provides the foundation for this setting [57, 35]. Instead of training a separate model for each task, it enables simultaneous prediction, allowing shared temporal representations to capture interdependencies between tasks [9]. HYDRA builds on this foundation by unifying diverse temporal prediction tasks under a single architecture. Specifically, we focus on a shared trunk architecture with task-specific heads, as it has proven effective in balancing information sharing and mitigating negative transfer [9]. Joint task learning lets the model exploit shared temporal dynamics rather than treating each task independently. Figure 2 shows the HYDRA framework.
Figure 2. HYDRA framework. The architecture integrates spatial and spectral paths with a unified multi-task design, enabling HYDRA to learn transferable representations across diverse temporal graph tasks.
Hydra: Shared Trunk with Task-Specific Heads
The central design choice of HYDRA is to couple a shared representation trunk with lightweight task-specific heads. We adopt this structure because temporal graph property prediction requires representations that are both transferable across heterogeneous graphs and adaptable to multiple tasks. The name HYDRA, from the water serpent in Greek and Roman mythology with multiple heads, emphasizes this shared representation architecture with multiple heads. The trunk captures common temporal–structural patterns once, while task-specific heads provide minimal but sufficient specialization.
HYDRA instantiates the predictor (Definition 1) with the shared trunk that produces embeddings for , followed by task-specific heads that map to predictions .
The trunk jointly leverages (i) spectral descriptors that summarize global connectivity patterns and (ii) spatial GNN-based encoders that model local temporal interactions. These components provide complementary global and local information while avoiding redundant representation learning.
As shown in Figure 2, the trunk consists of a spectral path, which encodes global descriptors, and a spatial path, which models local temporal dynamics. Their outputs are concatenated into the shared embedding , which is then passed to task-specific decoder heads, , each producing a prediction .
Spectral Path for Global Connectivity
Laplacian eigenvalues effectively encode global graph structural properties [27]. For example, in star, path, and fully-connected graphs, the distribution of the Laplacian eigenvalues is distinct and captures their unique topologies. However, computing all eigenvalues of the graph Laplacian matrix is highly expensive and only feasible for small graphs with thousands of nodes. Thus, it is important to efficiently approximate the spectrum of the Laplacian.
Network Density of States. To capture information from the Laplacian spectrum, we use the network density of states (DOS), the distribution of the eigenvalues of the graph Laplacian matrix [10], as follows:
Definition 2 (Density of States). Given a temporal graph snapshot with adjacency matrix and degree matrix , let the (unnormalized) Laplacian be . We use the normalized Laplacian with eigenvalues . The density of states is then defined as a discrete distribution supported on the spectrum of :
Intuitively, DOS summarizes the distribution of Laplacian eigenvalues and thus global connectivity. The Kernel Polynomial Method [50] approximates this distribution without explicitly computing eigenvalues or eigenvectors. We discretize the reconstructed density into 20 equal-width bins over , yielding a fixed-dimensional descriptor irrespective of graph size; implementation details are given in Appendix § 11.
Spectral Memory. To stabilize spectral signals and capture short-term trends, we maintain a fixed-length FIFO memory of recent DOS descriptors. At snapshot , let denote the spectral feature (e.g., a DOS vector). For prediction at time , the memory contains the last descriptors up to : . After prediction at , the new descriptor is computed and rolled into the memory: , . The FIFO memory provides a recent structural trajectory rather than a potentially noisy single snapshot, while discarding outdated descriptors.
Attention over Spectral Memory. Rather than uniform averaging, we apply attention to emphasize historically similar or predictive patterns. With learnable projections , we compute , , . Attention weights and the attentive spectral embedding are , . Unlike uniform averaging, attention can emphasize relevant history and distinguish transient fluctuations from persistent structural changes. The DOS descriptors of each snapshot are fixed, while their temporal aggregation is learned through .
Spatial Path for Local Connectivity
We integrate spatial and temporal dynamics by combining a Graph Convolutional Network (GCN) [21] with Gated Recurrent Units (GRUs) [7]. Similar approaches have proven effective in temporal settings [58]. At snapshot , a GCN encodes structural features , and a GRU updates node states with temporal memory . The representation is thus spatially grounded and temporally enriched.
Pooling and Graph Readout. We apply Self-Attention Graph Pooling (SAGPool) [23] to emphasize structurally important nodes. Node scores are computed as , where projects node embeddings to scalar scores. The top- nodes are retained to form the pooled graph with embeddings . We then compute a graph-level vector using mean pooling over the selected nodes: , . SAGPool reduces noise from low-activity nodes and highlights structurally central actors, which is particularly relevant in sparse networks where activity is concentrated in a few key nodes [41].
Prediction Head
We fuse spatial and spectral signals to produce a joint temporal state. At each step, the shared hidden state of the prediction head , where denotes concatenation of spatial and spectral features. The recurrent update captures cross-modal dependencies over time. Task-specific MLPs then map the shared state to predictions: , . This balances efficiency through a shared backbone with flexibility through task-specific heads.
End-to-End Training
Hydra supports both classification and regression objectives. For classification, each head is optimized with Binary Cross-Entropy loss; for regression, with Mean Squared Error. The total loss in each setting is the sum over task-specific losses within that group. Following prior work on loss balancing [56, 39], we train the two groups separately, as naïvely combining heterogeneous objectives degraded optimization stability. This preserves Hydra’s multi-task benefits within each group while maintaining transferability across unseen networks.
Experimental Evaluation
Training Setup and Algorithm
Datasets. We evaluate HYDRA on the MiNT benchmark of 84 ERC-20 temporal transaction networks [42], randomly selecting 64 networks for training/validation and 20 held-out networks for zero-shot transfer evaluation. Within each network, snapshots are split 70%–15%–15% into training, validation, and test sets. Appendix § 12 provides detailed dataset statistics. This setting lets us evaluate whether HYDRA generalizes across network topologies and temporal patterns. During each epoch, we present networks in random order, and reset historical embeddings at the start of each network to avoid cross-network leakage. The shared trunk is optimized jointly, while all task-specific heads are updated under a multi-task loss with early stopping on validation performance. We train models for 100 to 250 epochs with an initial learning rate of . We apply early stopping based on validation AUC for classification and MAE for regression, with a patience of 20 epochs and a tolerance of . We use binary cross-entropy for classification tasks and mean squared error for regression tasks. We optimize using Adam [20]. We evaluate each task on 20 unseen test networks across three independent runs, yielding 60 result points per task. We report the average results from three different runs for each task separately, using ROC-AUC for classification objectives and Mean Absolute Error for regression objectives. We pre-train HYDRA with 8, 16, and 32 networks to study how scale affects transferability. The primary scaling study uses up to 32 networks; the 64-network and capacity analyses are in Appendix § 18. Unless otherwise specified, all reported results refer to HYDRA-32, as this configuration offers strong overall performance.
Tasks. We train HYDRA in two multi-task settings: a classification model with three classification heads and a regression model with three regression heads. The classification tasks jointly predict whether the Largest Connected Component (LCC), the number of nodes, or the number of edges will grow or shrink. The regression tasks estimate the number of influential nodes (i.e., nodes with degree greater than five), new nodes, and edges. Appendix § 14 provides full task definitions.
Models.
We compare HYDRA against HTGN [52], GC-LSTM [5], EvolveGCN [31], GraphPulse [41], ROLAND [55], TGCN [59], and WinGNN [61], and additionally MiNT [42] for classification (MiNT is the only prior framework explicitly designed for transferability across temporal graphs for classification). Hydra and MiNT are trained only on source networks and evaluated on the same 20 held-out networks without target-network training or fine-tuning. Single-network baselines are trained separately on the training split of each target network. Appendix § 13 provides baseline details.
Compute Resources. Experiments were run on a dedicated cluster compute node equipped with an NVIDIA H100 GPU (80 GB), dual 64-core AMD EPYC CPUs, and 512 GB RAM.
Classification and Regression Results
Table 1 presents the aggregate classification results across three tasks: LCC Growth/Shrinkage, Node Growth/Shrinkage, and Edge Growth/Shrinkage. For each task, we report three metrics: () the number of datasets where the method achieved first place, () the average ranking across datasets, and () the average area under the ROC curve (metric formulations are provided in appendix § 15). HYDRA consistently achieves the strongest performance across all metrics, outperforming both state-of-the-art specifically trained models and the transferable baseline. Per-network means and standard deviations over three runs are reported in Appendix § 20.
| LCC-G/S | Node-G/S | Edge-G/S | |||||||
| 1st | Rank | AUC | 1st | Rank | AUC | 1st | Rank | AUC | |
| Single Models | |||||||||
| HTGN | 2 | 4.40 | 0.648 | 0 | 3.65 | 0.615 | 1 | 4.85 | 0.684 |
| GC-LSTM | 0 | 6.00 | 0.572 | 0 | 6.65 | 0.489 | 1 | 5.80 | 0.609 |
| EvolveGCN | 0 | 6.30 | 0.585 | 0 | 7.00 | 0.487 | 0 | 6.10 | 0.626 |
| GraphPulse | 3 | 3.50 | 0.686 | 2 | 2.95 | 0.652 | 6 | 3.80 | 0.712 |
| ROLAND | 0 | 6.95 | 0.556 | 0 | 6.70 | 0.480 | 0 | 6.30 | 0.542 |
| TGCN | 1 | 6.10 | 0.583 | 0 | 6.15 | 0.530 | 0 | 5.60 | 0.633 |
| WinGNN | 0 | 6.10 | 0.585 | 0 | 6.20 | 0.528 | 0 | 5.55 | 0.642 |
| Transfer Models | |||||||||
| MiNT | 1 | 3.95 | 0.672 | 3 | 3.55 | 0.616 | 4 | 3.30 | 0.727 |
| HYDRA (ours) | 13 | 1.70 | 0.759 | 15 | 1.95 | 0.734 | 8 | 2.80 | 0.753 |
Table 1. Table 1: Classification task results. Across LCC-GS, Node-GS, and Edge-G/S classification, HYDRA consistently outperforms both specifically trained and transferable baselines.
Table 2 summarizes the regression results across New Node Count, Influential Node Count, and Edge Count Regression. For each task, we report three metrics: () the number of datasets where the method achieved first place, () the average ranking across datasets, and () the average mean absolute error. HYDRA is the best model on two of the three tasks and remains very close to the best baseline on Influential Node Count, demonstrating strong overall transferability for regression Arrows indicate directionality: higher is better, lower is better.
| New Node Count | Influential Node Count | Edge Count | |||||||
| 1st | Rank | MAE | 1st | Rank | MAE | 1st | Rank | MAE | |
| Single Models | |||||||||
| HTGN | 4 | 3.45 | 0.039 | 4 | 3.52 | 0.050 | 2 | 3.95 | 0.049 |
| TGCN | 1 | 4.10 | 0.042 | 1 | 4.85 | 0.059 | 1 | 4.60 | 0.054 |
| GCLSTM | 2 | 4.00 | 0.043 | 5 | 3.80 | 0.053 | 4 | 4.58 | 0.057 |
| ROLAND | 3 | 4.35 | 0.052 | 2 | 4.55 | 0.057 | 3 | 4.58 | 0.050 |
| EGCN | 0 | 3.90 | 0.040 | 0 | 4.20 | 0.052 | 1 | 5.30 | 0.054 |
| GraphPulse | 1 | 5.75 | 0.071 | 1 | 5.90 | 0.078 | 2 | 5.72 | 0.073 |
| WinGNN | 4 | 4.05 | 0.054 | 4 | 4.28 | 0.059 | 5 | 3.93 | 0.057 |
| Transfer Model | |||||||||
| HYDRA (ours) | 8 | 2.50 | 0.035 | 5 | 4.90 | 0.056 | 6 | 3.35 | 0.046 |
Table 2. Table 2: Regression task results. Across New Node Count, Influential Node Count, and Edge Count regression, HYDRA achieves the best performance on two tasks and is competitive on the third, all without additional training.
Arrows indicate directionality: higher is better, lower is better.
objectives. HYDRA produces the most top-ranked results in both classification and regression settings, outperforming all baselines in five out of six tasks. The reported gains average task-wise relative improvements over all three classification tasks (8.9%) and the two winning regression tasks (8.2%); Appendix § 15 gives the calculation. Even in the one regression task where HYDRA isn’t the leader, it stays competitive with the strongest baseline. Notably, HYDRA is always tested on unseen networks, while single models specifically trained for a task are evaluated on the test portion of the same networks they were trained on. Despite this tougher setting, HYDRA consistently delivers better or comparable results, making it a reliable and adaptable framework for temporal graph learning.
Ablation Studies. We assess the impact of HYDRA’s key components by disabling the DOS module and the pooling layer. We provide a comprehensive analysis in Appendix § 16, including an additional edge-count regression DOS ablation. As summarized in Table 3, DOS improves performance from an average AUC of 0.697 to 0.719, raising the number of best-performing datasets from 6 to 14 and lowering the mean rank from 1.70 to 1.30. Pooling yields an even larger effect, increasing average AUC from 0.719 to 0.753 and improving the mean rank from 1.85 to 1.15, with gains on 17 out of 20 datasets. These results confirm that DOS provides complementary global spectral information, while pooling enhances the spatial path by emphasizing structurally important nodes, together strengthening HYDRA’s generalization ability.
| Ablation | Avg. AUC w/o | Avg. AUC w/ | #1st-Place w/o | #1st-Place w/ |
| Density of States | 0.697 | 0.719 | 6 | 14 |
| Attention-based Pooling | 0.719 | 0.753 | 3 | 17 |
Table 3. Table 3: Meta ablation summary for Edge G/S classification. Average AUC and #1st-Place counts are reported with and without the corresponding component. Per-network results are given in Appendix Tables 6 and 7.
Hydra Generalization and Efficiency
In this section, we analyze how the number of networks used during HYDRA pre-training affects its performance in predicting the characteristics of unseen networks at inference time. Figure 3 presents the results. For classification tasks, we observe that increasing the number of training networks generally leads to a higher #1^{st}-Place. This metric is a strong indicator of performance, as it reflects the number of unseen networks on which the model achieves the best results at inference time. In contrast, the correlation between the number of pre-training networks and metrics such as Avg. Rank or Avg. AUC is weaker, with only negligible differences across settings. Notably, HYDRA-32 consistently achieves the best or on-par performance across all classification benchmarks. For regression tasks, we observe a clearer scaling trend: performance steadily improves from HYDRA-8 to HYDRA-32 across all indicators. Specifically, #1^{st}-Place increases with scale, while both Avg. Rank and Avg. MAE decrease, indicating better predictive performance. Given its overall performance, we adopt HYDRA-32 as our primary model and focus our subsequent evaluation on it. Table 11 in Appendix § 19 also reports the performance of the 8-, 16-, and 32-network variants.

Figure 3. Scaling results for HYDRA-8, HYDRA-16, and HYDRA-32, pre-trained on 8, 16, and 32 networks, respectively, and evaluated zero-shot on the same 20 held-out networks, across classification and regression tasks. HYDRA-32 achieves competitive classification results and the lowest average regression errors among these settings. Evaluation measures are #1^{st}-Place (↑), Avg. Rank (↓), Avg. AUC (↑), and Avg. MAE (↓).
Table 11: Performance of HYDRA across six tasks as the number of training networks increases from 8 to 32.
| Dataset | HYDRA-8 | HYDRA-16 | HYDRA-32 |
| MIR | 0.769 (±0.003) | 0.769 (±0.007) | 0.764 (±0.001) |
| DOGE2.0 | 0.613 (±0.083) | 0.573 (±0.133) | 0.633 (±0.064) |
| MUTE | 0.755 (±0.016) | 0.763 (±0.003) | 0.748 (±0.025) |
| EVERMOON | 0.624 (±0.005) | 0.582 (±0.031) | 0.655 (±0.040) |
| DERC | 0.734 (±0.010) | 0.725 (±0.006) | 0.742 (±0.004) |
| ADX | 0.753 (±0.027) | 0.767 (±0.004) | 0.718 (±0.051) |
| HOICHI | 0.582 (±0.029) | 0.552 (±0.050) | 0.558 (±0.047) |
| SDEX | 0.762 (±0.008) | 0.748 (±0.068) | 0.743 (±0.026) |
| BAG | 0.963 (±0.010) | 0.952 (±0.021) | 0.969 (±0.009) |
| XCN | 0.862 (±0.017) | 0.821 (±0.070) | 0.878 (±0.009) |
| ETH2x-FLI | 0.717 (±0.030) | 0.710 (±0.009) | 0.678 (±0.031) |
| stkAAVE | 0.776 (±0.009) | 0.776 (±0.003) | 0.779 (±0.007) |
| GLM | 0.750 (±0.011) | 0.746 (±0.010) | 0.763 (±0.005) |
| QOM | 0.703 (±0.012) | 0.707 (±0.013) | 0.719 (±0.012) |
| WOJAK | 0.502 (±0.103) | 0.352 (±0.157) | 0.412 (±0.060) |
| DINO | 0.903 (±0.029) | 0.905 (±0.021) | 0.910 (±0.013) |
| Metis | 0.685 (±0.002) | 0.632 (±0.091) | 0.693 (±0.007) |
| REPv2 | 0.728 (±0.022) | 0.721 (±0.023) | 0.689 (±0.013) |
| TRAC | 0.756 (±0.005) | 0.745 (±0.014) | 0.765 (±0.003) |
| BEPRO | 0.858 (±0.023) | 0.806 (±0.094) | 0.865 (±0.010) |
| 1st-Place Count↑ | 6 | 3 | 12 |
| Avg. Rank↓ | 1.85 | 2.45 | 1.70 |
| Avg. AUC↑ | 0.740 | 0.718 | 0.734 |
Table 11. (a) Classification: Node G/S
Computation and Time Efficiency. The per-snapshot complexity of HYDRA reduces to a scalable , where is the number of edges, is the number of nodes, and is the number of Chebyshev moments used in the DOS module (see Appendix § 11 for details). Because MiNT must be retrained separately for each task, training three tasks requires about three full runs on 64 networks. In contrast, HYDRA trains all tasks jointly in a single run on 32 networks and achieves a 22× speedup, completing all three tasks in about 16,700 seconds compared to more than 370,000 seconds for MiNT (Figure 4a). This efficiency makes HYDRA far more practical and scalable for multi-task training and inference.
Figure 4. (a) Total training cost for three classification tasks. HYDRA trains once for all tasks jointly, while MiNT must be retrained separately for each task. The dashed line indicates that MiNT is evaluated at 64 networks, while HYDRA’s best result occurs at 32 networks. (b) Inference cost on TRAC dataset. HYDRA performs a single zero-shot classification pass for all three tasks jointly, whereas single-network baselines require full training before they can be used for inference, even on a single task (LCC–G/S here); reported times reflect the cost of inference on an unseen network using HYDRA versus single models.
Figure 4: Training and inference efficiency of HYDRA. (a) HYDRA reduces training cost by avoiding retraining. (b) HYDRA drastically lowers inference time by requiring only one forward pass.
HYDRA inference time also improves significantly. HYDRA requires only a single zero-shot inference pass to predict all three tasks on an unseen network, completing end-to-end evaluation on the largest test network (TRAC) in about 56 seconds (see Figure 4b).
Single-network baselines, in contrast, require retraining from scratch for every dataset and task before they can be used for inference, with training times ranging from several thousand seconds to over 9,000 seconds even for a single task. Extending these models to multiple tasks or multiple networks multiplies the cost linearly, making them infeasible at scale. HYDRA’s design eliminates the need for repeated training and heavy preprocessing, providing unique efficiency in both training and inference.
Transfer beyond ERC-20. We additionally train Hydra-Social on six social networks and Hydra-Mix on those networks plus 32 ERC-20 networks, holding out MathOverflow and RedditB. On RedditB, Edge G/S AUC rises from 0.604 (Social) to 0.769 (Mix); on MathOverflow, Hydra-Mix reaches 0.630, below MiNT-Mix’s 0.782. These results support mixed-domain transfer but show that gains are not uniform (Appendix § 17).
Impact of Multi-Task Training
To assess how multi-task learning shapes the shared temporal representation, we examine edge count prediction as a reference task and train three model variants: single-task (edge count only), two-task (edge count and new node count), and the full three-task HYDRA configuration. Table 4 shows the results.
| Dataset | HYDRA (One Task) | HYDRA (Two Tasks) | HYDRA (Three Tasks) |
| MIR | 0.057 (± 0.004) | 0.034 (± 0.002) | 0.016 (± 0.005) |
| DOGE2.0 | 0.169 (± 0.016) | 0.150 (± 0.015) | 0.187 (± 0.008) |
| MUTE | 0.053 (± 0.020) | 0.067 (± 0.011) | 0.021 (± 0.022) |
| EVERMOON | 0.029 (± 0.021) | 0.042 (± 0.010) | 0.017 (± 0.000) |
| DERC | 0.024 (± 0.016) | 0.035 (± 0.010) | 0.021 (± 0.003) |
| ADX | 0.023 (± 0.010) | 0.027 (± 0.006) | 0.025 (± 0.008) |
| HOICHI | 0.071 (± 0.006) | 0.068 (± 0.011) | 0.027 (± 0.020) |
| SDEX | 0.060 (± 0.023) | 0.046 (± 0.007) | 0.095 (± 0.043) |
| BAG | 0.061 (± 0.018) | 0.067 (± 0.020) | 0.041 (± 0.020) |
| XCN | 0.037 (± 0.017) | 0.026 (± 0.010) | 0.062 (± 0.015) |
| ETH2x-FLI | 0.030 (± 0.013) | 0.038 (± 0.007) | 0.023 (± 0.001) |
| stkAAVE | 0.026 (± 0.008) | 0.021 (± 0.003) | 0.051 (± 0.022) |
| GLM | 0.092 (± 0.009) | 0.083 (± 0.017) | 0.087 (± 0.008) |
| QOM | 0.025 (± 0.003) | 0.023 (± 0.005) | 0.036 (± 0.013) |
| WOJAK | 0.025 (± 0.023) | 0.039 (± 0.017) | 0.010 (± 0.004) |
| DINO | 0.031 (± 0.018) | 0.041 (± 0.010) | 0.017 (± 0.007) |
| Metis | 0.047 (± 0.014) | 0.054 (± 0.016) | 0.034 (± 0.010) |
| REPv2 | 0.103 (± 0.002) | 0.102 (± 0.003) | 0.111 (± 0.008) |
| TRAC | 0.029 (± 0.016) | 0.039 (± 0.012) | 0.021 (± 0.001) |
| BEPRO | 0.031 (± 0.025) | 0.046 (± 0.012) | 0.011 (± 0.008) |
| Average | 0.051 | 0.052 | 0.046 |
| Best Count | 1 | 7 | 12 |
Table 4. MAE of edge count prediction under one, two, and three task HYDRA configurations, reported with standard deviations.
The trend is non-monotonic: average MAE is 0.051, 0.052, and 0.046 for one, two, and three tasks, respectively. The three-task model achieves the best results on 12 out of 20 datasets and the lowest overall average MAE, compared to 7 and 1 for the two-task and single-task variants respectively. These gains are consistent with additional tasks providing complementary supervisory signals that encourage the shared trunk to learn richer temporal and structural patterns, suggesting benefits from multi-task supervision overall.
Conclusion
We have introduced HYDRA, a novel temporal graph-level multi-task and multi-network model designed to address the complexities of dynamic network analysis. HYDRA combines a spatial path, which captures local connectivity through temporal GNNs, along with a spectral path that attends to global structural patterns, forming a new architectural design for temporal graph learning. This unified framework enables HYDRA to handle diverse prediction tasks simultaneously while transferring effectively to unseen networks. Empirically, HYDRA consistently outperforms state-of-the-art baselines across benchmarks, achieving strong performance without requiring additional training on target networks. These results highlight HYDRA as the first architecture to bring together spectral and spatial pathways for transferable, multi-task learning on temporal graphs. Looking ahead, this work opens up a promising research direction toward more generalizable temporal models that can support a wide range of tasks.
Acknowledgments
This work was partially supported by the Canadian NSERC Discovery Grant RGPIN-2020-05665: Data Science on Blockchains, the National Science Foundation under grant DMS-2229417, and the Simons Foundation under grant SFM-21269. This research is also supported by the Canadian Institute for Advanced Research (CIFAR AI chair program) and the EPSRC Turing AI World-Leading Research Fellowship No.EP/X040062/1 and EPSRC AI Hub No. EP/Y028872/1. The authors acknowledge the UCF Advanced Research Computing Center and Texas Advanced Computing Center (TACC) at UT Austin.
Appendix
Limitations
In this work, we propose HYDRA, a novel architecture for multi-task temporal graph property prediction. Our evaluation covers ERC-20 transaction networks [42] and two unseen social networks. Transfer to traffic, citation, and biological temporal graphs remains untested; broader domain coverage is needed.
Here, we also focus on the graph property prediction task, as we expect the model to learn the evolution of the global graph structure and transfer that learning across networks. One limitation is that we explore fewer other tasks, such as link prediction and node classification. Training networks have different sequence lengths, but we do not independently vary temporal duration in a controlled experiment. This is also partly due to the nature of the dataset, where many nodes are transient and only appear once in a network; it is more difficult to generalize and predict the behavior of these one-time nodes.
DOS is a fixed, binned spectral summary and may discard structural detail; learned attention does not make the descriptor itself learnable. Our ablations establish the value of the spectral path, not superiority over learned spectral alternatives. We do not include matched spectral-transformer or other recent graph-pretraining baselines, which require adaptation to this multi-network graph-level setting. Increasing pre-training diversity may also introduce task-dependent negative transfer or capacity constraints (Appendix § 18). Future directions include broader multi-domain pre-training, self-supervised objectives, and prompt- or head-based adaptation.
Broader Impact
This work advances temporal graph learning by enabling a single model to handle multiple tasks and transfer across networks without retraining. This capability reduces computational cost and improves the transferability of TGNNs, with potential impact in areas such as network security, financial fraud detection, and real-time social network analysis, where evolving systems require continuously adaptive, multi-objective modeling.
Moreover, the principles of multi-network and multi-task pre-training introduced here may inspire similar advances in other temporal data domains, including healthcare, transportation, and climate science. More broadly, this work opens a new direction toward generalizable temporal graph models that can be trained across distinct networks, taking an important step toward Temporal Graph Foundation Models.
Computational Complexity of Hydra
DOS computation. We rescale the normalized Laplacian as , mapping its spectrum from to . KPM estimates 50 Chebyshev moments using 50 random probing vectors and stochastic trace estimation. The recurrence requires only sparse matrix–vector products. Jackson damping suppresses the Gibbs phenomenon, which we map back to and discretize into 20 equal-width bins. The complexity below uses for the moment count and treats the probe count as fixed.
For a snapshot with nodes and edges, the per-snapshot complexity of HYDRA is
where is the hidden dimension, is the number of task-specific heads, and is the number of Chebyshev moments used to approximate the DOS with a kernel polynomial method. The term arises from spatial message passing, since each edge propagates a -dimensional embedding. The term comes from recurrent or attention-based updates. Pooling with top- selection contributes , reflecting attention-score computation for all nodes followed by partial sorting. The spectral DOS module adds , as each moment requires one sparse matrix–vector multiplication. Finally, the cost of task-specific heads is . Over networks with snapshots each, the overall training cost scales as
In our implementation, the spatial hidden dimension is fixed to 16, the DOS descriptor dimension is 20, and we train with task-specific heads. Substituting these values, the per-snapshot complexity of HYDRA becomes
where and . The terms correspond to spatial message passing (), recurrent/attention updates (), pooling with top- selection (), DOS spectral approximation with Chebyshev moments (), and three task heads on a 36-dimensional joint embedding (108). For each snapshot, the overall training complexity scales as
In practice, with hidden size and Chebyshev moment parameters fixed, the per-snapshot complexity of HYDRA reduces to , dominated by sparse matrix–vector multiplications in the DOS module.
Dataset
We evaluate HYDRA on temporal transaction networks from the MiNT benchmark [42], the first large-scale dataset designed for training and evaluating temporal graph models across multiple heterogeneous networks. MiNT contains 84 ERC-20 token transaction graphs collected from the Ethereum blockchain between 2017 and 2023, spanning more than six years of activity. The biggest MiNT token network contains 128,159 unique addresses and 554,705 transactions, while the smallest token network has 1,454 nodes. Each network is represented as a sequence of weekly snapshots, where nodes correspond to wallet addresses and edges represent token transfers. Across their full duration, networks range from tens of thousands to over 100K unique nodes and up to several million edges, reflecting the scale of real-world token ecosystems.
A distinctive feature of MiNT is its strong inductiveness: new addresses and transactions appear continuously, making prediction inherently open-world. This reflects the real behavior of blockchain networks, where adoption cycles, liquidity shocks, and market events constantly introduce novel participants and structural patterns. The long temporal duration further amplifies this effect, as networks exhibit bursts of rapid growth, periods of stagnation, and sudden fragmentation or collapse. Such novelty and surprise factors make MiNT particularly challenging, as models must generalize across highly dynamic trajectories rather than stationary or repetitive patterns.
We summarize detailed statistics of each token network in the MiNT datasets in Table 5. Most networks have more than 10k nodes and over 100k edges. The lifespan of MiNT networks varies from 107 days to 6 years, and each day has at least one transaction. As the table shows, the token networks have quite high surprise values with an average of 0.82.
| Token | #Node | #Transaction | #Snapshots (days) | Growth rate | Novelty | Surprise | Token | #Node | #Transaction | #Snapshots (days) | Growth rate | Novelty | Surprise |
| ARC | 11325 | 70968 | 606 | 0.43 | 0.32 | 0.88 | Metis | 52586 | 343141 | 907 | 0.44 | 0.48 | 0.89 |
| CELR | 65350 | 235807 | 1691 | 0.49 | 0.56 | 0.96 | cDAI | 52753 | 358050 | 1437 | 0.45 | 0.46 | 0.9 |
| CMT | 86895 | 205961 | 309 | 0.45 | 0.72 | 0.92 | BITCOIN | 34051 | 347054 | 178 | 0.48 | 0.39 | 0.63 |
| DRGN | 113453 | 341849 | 2164 | 0.44 | 0.57 | 0.97 | INJ | 60472 | 312822 | 1113 | 0.46 | 0.52 | 0.98 |
| GHST | 35156 | 180955 | 1146 | 0.43 | 0.51 | 0.93 | MIM | 23038 | 269366 | 885 | 0.44 | 0.4 | 0.89 |
| INU | 8556 | 66315 | 154 | 0.27 | 0.41 | 0.59 | GLM | 53385 | 234912 | 1080 | 0.5 | 0.53 | 0.96 |
| IOTX | 63079 | 288469 | 1993 | 0.45 | 0.56 | 0.99 | Mog | 14590 | 240680 | 107 | 0.37 | 0.38 | 0.55 |
| QSP | 117977 | 299671 | 2178 | 0.45 | 0.67 | 0.99 | DPI | 40627 | 234246 | 1150 | 0.49 | 0.5 | 0.86 |
| REP | 83282 | 224843 | 346 | 0.46 | 0.69 | 0.96 | LINA | 45342 | 227147 | 1144 | 0.45 | 0.46 | 0.95 |
| RFD | 23208 | 173695 | 169 | 0.3 | 0.39 | 0.6 | Yf-DAI | 22466 | 226875 | 1158 | 0.42 | 0.31 | 0.87 |
| TNT | 88247 | 316352 | 1216 | 0.43 | 0.55 | 0.93 | BOB | 42806 | 212099 | 199 | 0.35 | 0.48 | 0.73 |
| TRAC | 71667 | 299181 | 2110 | 0.46 | 0.54 | 0.97 | RGT | 35277 | 211932 | 1110 | 0.44 | 0.46 | 0.98 |
| RLB | 28033 | 240291 | 129 | 0.43 | 0.49 | 0.76 | TVK | 42539 | 208082 | 1062 | 0.41 | 0.48 | 0.93 |
| steCRV | 19079 | 211538 | 1033 | 0.45 | 0.53 | 0.9 | RSR | 50645 | 205906 | 659 | 0.47 | 0.62 | 0.91 |
| ALBT | 63042 | 434881 | 1152 | 0.43 | 0.44 | 0.89 | WOJAK | 34341 | 198653 | 201 | 0.37 | 0.48 | 0.73 |
| POLS | 128159 | 554705 | 1132 | 0.45 | 0.61 | 0.94 | ANT | 36517 | 200262 | 1107 | 0.47 | 0.46 | 0.93 |
| SWAP | 69230 | 509769 | 1213 | 0.46 | 0.45 | 0.79 | LADYS | 37486 | 192176 | 181 | 0.37 | 0.52 | 0.79 |
| SUPER | 83299 | 502030 | 986 | 0.47 | 0.46 | 0.85 | ETH2x-FLI | 11008 | 199088 | 965 | 0.47 | 0.28 | 0.84 |
| RARI | 87186 | 502960 | 1207 | 0.43 | 0.47 | 0.91 | TURBO | 38638 | 189048 | 189 | 0.33 | 0.48 | 0.72 |
| KP3R | 39323 | 493258 | 1102 | 0.43 | 0.33 | 0.88 | REPv2 | 39061 | 191367 | 1194 | 0.48 | 0.5 | 0.97 |
| MIR | 79984 | 444998 | 1066 | 0.45 | 0.43 | 0.92 | NOIA | 29798 | 185528 | 1133 | 0.46 | 0.37 | 0.7 |
| aUSDC | 23742 | 475680 | 1067 | 0.46 | 0.4 | 0.73 | 0x0 | 21531 | 182430 | 283 | 0.51 | 0.46 | 0.81 |
| LUSD | 25852 | 430473 | 943 | 0.48 | 0.36 | 0.87 | PSYOP | 25450 | 168896 | 169 | 0.32 | 0.39 | 0.59 |
| PICKLE | 28498 | 430262 | 1149 | 0.48 | 0.34 | 0.69 | ShibDoge | 40023 | 134697 | 680 | 0.43 | 0.53 | 0.8 |
| DODO | 47046 | 390443 | 1131 | 0.47 | 0.45 | 0.91 | ADX | 14567 | 123755 | 1188 | 0.44 | 0.4 | 0.91 |
| YFII | 43964 | 391984 | 1196 | 0.44 | 0.44 | 0.96 | BAG | 11860 | 122634 | 298 | 0.31 | 0.44 | 0.87 |
| STARL | 71590 | 369913 | 856 | 0.46 | 0.48 | 0.86 | QOM | 21757 | 118292 | 598 | 0.46 | 0.41 | 0.81 |
| LQTY | 34687 | 374230 | 943 | 0.45 | 0.34 | 0.91 | BEPRO | 26521 | 120261 | 1132 | 0.46 | 0.48 | 0.87 |
| FEG | 118294 | 367584 | 1007 | 0.4 | 0.62 | 0.92 | AIOZ | 29231 | 119926 | 947 | 0.43 | 0.49 | 0.89 |
| AUDIO | 91218 | 362685 | 1108 | 0.45 | 0.58 | 0.95 | PRE | 40476 | 118625 | 1113 | 0.5 | 0.55 | 0.86 |
| OHM | 45728 | 377068 | 690 | 0.43 | 0.46 | 0.88 | CRU | 19990 | 117712 | 1144 | 0.5 | 0.43 | 0.9 |
| WOOL | 18684 | 351178 | 716 | 0.41 | 0.18 | 0.41 | POOH | 27245 | 111641 | 193 | 0.26 | 0.49 | 0.69 |
| DEFI | 24777 | 111205 | 824 | 0.45 | 0.49 | 0.83 | DAI25 | 13648 | 187050 | 1068 | 0.45 | 0.46 | 0.82 |
| stKAAVE | 37355 | 110924 | 1128 | 0.42 | 0.57 | 0.71 | ORN | 44010 | 239451 | 1134 | 0.46 | 0.47 | 0.87 |
| BTFRLY | 8450 | 85371 | 453 | 0.48 | 0.34 | 0.74 | DOGE2.0 | 7664 | 79047 | 123 | 0.45 | 0.38 | 0.66 |
| SDEX | 9127 | 104869 | 240 | 0.41 | 0.44 | 0.75 | HOICHI | 5075 | 77361 | 436 | 0.36 | 0.32 | 0.71 |
| XCN | 20085 | 104185 | 607 | 0.46 | 0.42 | 0.84 | EVERMOON | 7552 | 79868 | 163 | 0.24 | 0.35 | 0.52 |
| HOP | 37004 | 102650 | 514 | 0.41 | 0.46 | 0.88 | MUTE | 12426 | 82345 | 977 | 0.43 | 0.46 | 0.95 |
| MAHA | 18401 | 96180 | 749 | 0.43 | 0.47 | 0.91 | cvUSDD | 2950 | 88647 | 174 | 0.61 | 0.37 | 0.73 |
| DINO | 15837 | 94140 | 358 | 0.44 | 0.44 | 0.74 | SLP | 6675 | 95368 | 1151 | 0.43 | 0.36 | 0.91 |
| bendWETH | 1454 | 96898 | 593 | 0.51 | 0.21 | 0.51 | sILV2 | 12838 | 92905 | 611 | 0.4 | 0.34 | 0.48 |
| PUSH | 14501 | 93103 | 936 | 0.46 | 0.38 | 0.83 | SPONGE | 25852 | 90468 | 184 | 0.31 | 0.66 | 0.81 |
Table 5. All token networks’ statistics.
For our experiments, we follow the MiNT protocol and split the dataset into 64 financial networks for training and 20 additional networks for unseen testing. This setup lets HYDRA learn transferable temporal representations from diverse source networks and evaluate zero-shot generalization on new target networks available only at inference time.
Baselines
In this section, we provide further details on the temporal graph learning models we used as baselines for our work.
TGCN [59] is a combination of GCN and GRU. In particular, GCN is used to learn complex topological structures, while GRU is used to model embeddings dynamically to capture temporal dependence.
HTGN [52] leverages the power of hyperbolic geometry, which is well-suited for capturing hierarchical structures and complex relationships in temporal networks. HTGN maps the temporal graph into hyperbolic space and utilizes hyperbolic graph neural networks and hyperbolic gated recurrent neural networks to model the evolving dynamics. It incorporates two key modules that are hyperbolic temporal contextual self-attention (HTA) and hyperbolic temporal consistency (HTC)-to ensure that temporal dependencies are effectively captured and that the model is both stable and generalizable across various tasks.
GraphPulse [41] addresses the challenge of learning from nodes and edges with different timestamps, which many existing models struggle with. It combines two key techniques: the Mapper method from topological data analysis to extract clustering information from graph nodes and Recurrent Neural Networks (RNNs) for temporal reasoning. This principle approach helps capture both the structure and dynamics of evolving graphs.
GCLSTM [5] combines a Graph Convolutional Network (GCN) and Long Short-Term Memory (LSTM) units to handle both the structural and temporal aspects of evolving networks. The GCN is used to capture the local structural properties of the network at each snapshot, while the LSTM learns the temporal evolution of these snapshots over time.
EvolveGCN [31] is designed to capture the temporal dynamics of graph-structured data. Instead of relying on static node embeddings, EvolveGCN evolves the parameters of a graph convolutional network (GCN) over time. By using a recurrent neural network (RNN) to adapt the GCN parameters, this model is capable of dynamically adjusting during both training and testing, allowing it to handle evolving graphs, even when node sets vary significantly across different time steps.
ROLAND [55] is a dynamic graph learning framework that models node representations as hierarchical states, updated recurrently to capture temporal dependencies in evolving graphs. It supports scalable training using techniques like truncated backpropagation through time and meta-learning. In our DTDG setting, we use ROLAND to benchmark its performance and adaptability across diverse temporal networks.
WinGNN [61] uses a simple GNN to model topological information from the graph as other models existing in the literature. However, to model temporal dependencies, WinGNN proposes a novel mechanism of random gradient aggregation and a meta-learning strategy. In particular, WinGNN computes the frame-wise loss of the current snapshot and passes the loss gradient to the next to model graph dynamics without using RNN-based modules. Then it introduces the randomized sliding window to acquire the window-aware gradient on consecutive snapshots, and the two types of gradients are aggregated to update the GNN modules.
Task Formalizations
In this section, we provide detailed definitions for all classification and regression tasks considered in HYDRA. Each temporal snapshot corresponds to a 7-day interval, and the property prediction setup follows GraphPulse [41]. We also note that some of these tasks and the corresponding labels were processed specifically for this work, ensuring consistency and comparability across networks.
Setting , , and days, we establish a practical graph property with a 7-day prediction window. This choice is particularly relevant in financial contexts, such as Ethereum asset networks, where it can guide investment decisions [1].
Classification Tasks
Node Growth/Shrinkage (Node G/S). Let denote the set of unique nodes active between times and . The task predicts whether the number of nodes increases in the prediction interval :
Importance. Node growth measures adoption, reflecting the entry of new addresses, while shrinkage signals attrition. In token ecosystems, this corresponds to market expansion or decline.
Edge Growth/Shrinkage (Edge G/S). Let denote the set of transactions in . The model predicts whether transaction activity grows in the next interval:
Importance. Edge growth captures changes in liquidity and market engagement, with direct implications for trading activity and token valuation.
Largest Connected Component Growth/Shrinkage (LCC G/S). Let denote the size of the largest connected component during . The task is to predict whether connectivity expands:
Importance. LCC growth reflects stronger structural integration, improving liquidity and market stability in blockchain networks.
Regression Tasks
New Node Count. Let denote the set of nodes that appear for the first time in the prediction interval. The regression target is:
Importance. New node prediction quantifies adoption and user acquisition, a key indicator of ecosystem growth.
Edge Count. The task is to predict the number of transactions in the future interval:
Importance. Transaction forecasts provide fine-grained estimates of liquidity and demand surges in decentralized markets.
Influential Node Count. Define influential nodes as those with a degree greater than 5 in the prediction interval. Empirically, degree distributions of ERC-20 token graphs are heavy-tailed [42]: most nodes appear only once or twice, and a small fraction appear hundreds or thousands of times. When we plot the cumulative distribution of node degrees, there is usually a sharp drop in frequency after degree 1–2, followed by a long but thinner tail. Setting the threshold at 5 sits just beyond this long tail cutoff, ensuring that only the top 10–20% of nodes in activity are retained as influential. In practice, this excludes wallets that perform only a handful of transfers while capturing the repeat participants who actually shape liquidity and flow.
Let . The regression target is:
Importance. Influential nodes represent hubs such as active traders, liquidity providers, or contracts, which shape the stability and price dynamics of token ecosystems.
Importance of Temporal Property Prediction
Monitoring global structural dynamics in temporal graphs is essential in domains where the evolution of the entire network, rather than individual edges, drives operational decisions. In blockchain transaction networks, most transactions occur only once and do not repeat, so link-level prediction is not informative. Graph properties instead track global indicators such as changes in connectivity, activity, and influential participants, which often serve as early signals of liquidity risk, instability, or ecosystem decline [1, 2]. Since thousands of Ethereum-based tokens evolve with their own user bases and activity patterns [60], training a separate temporal model for each network is not practical. A single model that learns from many heterogeneous networks and jointly predicts several graph properties offers a more scalable and informative solution. Similar considerations arise in communication, payment, and social systems, where forecasting network-level behaviors helps detect anomalies, anticipate demand, and understand system health [19]. A multi-task approach is particularly valuable because these properties move together and provide complementary signals, leading to more reliable and actionable predictions than solving each task independently.
Evaluation Metrics
We evaluate models using both performance scores and ranking-based statistics. Below we provide formal definitions of each metric used in the main paper.
First-Place Count. For a given task, the first-place count measures how often a method achieves the top performance across datasets. Let denote the set of datasets and the set of methods. For dataset , let , where is the task-specific score (AUC for classification, MAE for regression). Then the first-place count for method is .
Average Rank. Each method is ranked per dataset according to performance. Let be the rank of method on dataset (lower is better). The average rank of method is .
Average AUC (Classification). For binary classification, the Area Under the ROC Curve (AUC) for dataset is , where are predicted scores and are ground-truth labels. The average AUC of method is .
Average MAE (Regression). For regression, the Mean Absolute Error (MAE) on dataset is , where are predictions and ground truth labels. The average MAE of method is .
Relative Gain Across Tasks. Relative gain quantifies Hydra’s improvement over the strongest baseline by comparing the average metric values across all 20 test datasets for each task. For classification (where higher AUC is better), the gain for task is defined as
where denotes the average AUC of method on task across all test datasets.
For regression (where lower MAE is better), the gain is
where denotes the average MAE of method on task across all test datasets.
Finally, the overall gain across a group of tasks (e.g., all classification tasks) is computed as the mean of the per-task gains:
For LCC, node, and edge growth/shrinkage, the relative AUC gains are 10.64%, 12.58%, and 3.58%, respectively (mean 8.9%). For new-node and edge-count regression, the relative MAE reductions are 10.26% and 6.12% (mean 8.2%); this regression summary excludes influential-node count, where Hydra is not the best model, although it remains competitive without any specific training on tested networks.
Ablation Studies
To better understand HYDRA components, we conduct an ablation study on the edge growth/shrinkage task. The ablations reflect different stages in the model’s incremental development. First, we compare the initial version of HYDRA without pooling, evaluating the impact of adding Density of States (DOS) features in the spectral path. Next, we consider models that already include DOS and assess the effect of introducing attention-based pooling in the spatial path. Results are reported in Appendix Tables 6 and 7. Together, these comparisons show that DOS enhances the capture of global spectral structure, while pooling improves the selection of informative subgraphs, confirming their complementary roles in HYDRA’s design and performance.
Impact of DOS. To assess the contribution of the DOS module, we compare HYDRA with and without DOS features (Table 6). This setup isolates the effect of spectral information on predictive performance, holding all other components fixed. The results show that incorporating DOS substantially improves performance on most datasets: HYDRA with DOS achieves the best results in 14 out of 20 networks, compared to 6 without DOS. On average, DOS raises AUC from 0.697 to 0.719 and improves the mean rank from 1.70 to 1.30. Gains are particularly large on challenging datasets such as EVERMOON, DOGE2.0, and DINO, where DOS more effectively captures global spectral structure. These findings highlight that DOS provides complementary information to the spatial path, leading to stronger generalization across temporal networks.
| Dataset | HYDRA w/o DOS | HYDRA w DOS |
| MIR | 0.791 (±0.014) | 0.755 (±0.013) |
| DOGE2.0 | 0.500 (±0.102) | 0.732 (±0.044) |
| MUTE | 0.675 (±0.030) | 0.699 (±0.011) |
| EVERMOON | 0.426 (±0.129) | 0.796 (±0.219) |
| DERC | 0.748 (±0.029) | 0.761 (±0.030) |
| ADX | 0.678 (±0.019) | 0.703 (±0.006) |
| HOICHI | 0.792 (±0.052) | 0.572 (±0.016) |
| SDEX | 0.599 (±0.134) | 0.261 (±0.129) |
| BAG | 0.759 (±0.294) | 0.874 (±0.017) |
| XCN | 0.822 (±0.073) | 0.825 (±0.038) |
| ETH2x-FLI | 0.716 (±0.020) | 0.736 (±0.026) |
| stkAAVE | 0.708 (±0.021) | 0.719 (±0.006) |
| GLM | 0.755 (±0.194) | 0.810 (±0.015) |
| QOM | 0.639 (±0.015) | 0.698 (±0.024) |
| WOJAK | 0.548 (±0.081) | 0.633 (±0.081) |
| DINO | 0.672 (±0.044) | 0.843 (±0.031) |
| Metis | 0.785 (±0.050) | 0.713 (±0.027) |
| REPv2 | 0.761 (±0.039) | 0.736 (±0.028) |
| TRAC | 0.794 (±0.041) | 0.732 (±0.007) |
| BEPRO | 0.775 (±0.022) | 0.786 (±0.009) |
| 1st-Place Count↑ | 6 | 14 |
| Avg. Rank ↓ | 1.70 | 1.30 |
| Avg. AUC ↑ | 0.697 | 0.719 |
Table 6. DOS ablation study on task Edge G/S. AUC values are reported. Best results per dataset are in bold.
Impact of Attention-based Pooling. The ablation on the edge growth/shrinkage task (Table 7) shows the contribution of the pooling mechanism within HYDRA’s spatial path. When pooling is enabled, the model achieves higher AUC on 17 out of 20 datasets, with the average score improving from 0.719 to 0.753 and the mean rank from 1.85 to 1.15. The improvement is particularly noticeable when pooling helps the spatial path focus on structurally important nodes, resulting in stronger graph-level representations. These findings indicate that pooling is an essential component for producing compact, informative graph-level representations and strengthen HYDRA’s ability to capture temporal dynamics.
| Dataset | HYDRA w/o Pooling | HYDRA w Pooling |
| MIR | 0.755 (±0.013) | 0.793 (±0.002) |
| DOGE2.0 | 0.732 (±0.044) | 0.897 (±0.089) |
| MUTE | 0.699 (±0.011) | 0.701 (±0.098) |
| EVERMOON | 0.796 (±0.219) | 0.818 (±0.046) |
| DERC | 0.761 (±0.030) | 0.839 (±0.008) |
| ADX | 0.703 (±0.006) | 0.722 (±0.087) |
| HOICHI | 0.572 (±0.016) | 0.591 (±0.103) |
| SDEX | 0.261 (±0.129) | 0.348 (±0.066) |
| BAG | 0.874 (±0.017) | 0.969 (±0.008) |
| XCN | 0.825 (±0.038) | 0.844 (±0.012) |
| ETH2x-FLI | 0.736 (±0.026) | 0.712 (±0.045) |
| stkAAVE | 0.719 (±0.006) | 0.732 (±0.011) |
| GLM | 0.810 (±0.015) | 0.850 (±0.009) |
| QOM | 0.698 (±0.024) | 0.745 (±0.003) |
| WOJAK | 0.633 (±0.081) | 0.585 (±0.067) |
| DINO | 0.843 (±0.031) | 0.895 (±0.003) |
| Metis | 0.713 (±0.027) | 0.733 (±0.004) |
| REPv2 | 0.736 (±0.028) | 0.772 (±0.016) |
| TRAC | 0.732 (±0.007) | 0.722 (±0.001) |
| BEPRO | 0.786 (±0.009) | 0.800 (±0.002) |
| 1st-Place Count↑ | 3 | 17 |
| Avg. Rank ↓ | 1.85 | 1.15 |
| Avg. AUC ↑ | 0.719 | 0.753 |
Table 7. SAG pooling ablation study on task Edge G/S. AUC values are reported. Best results per dataset are in bold.
Regression DOS ablation. The additional edge-count ablation reduces reported average MAE from 0.044 to 0.034 with DOS. DOS wins on 11 networks, the variant without DOS on six, and three are tied at the reported precision (Table 8). This complements the classification ablation, while improvements remain dataset-dependent.
| Dataset | HYDRA w/o DOS | HYDRA w DOS |
| MIR | ||
| DOGE2.0 | ||
| MUTE | ||
| EVERMOON | ||
| DERC | ||
| ADX | ||
| HOICHI | ||
| SDEX | ||
| BAG | ||
| XCN | ||
| ETH2x-FLI | ||
| stkAAVE | ||
| GLM | ||
| QOM | ||
| WOJAK | ||
| DINO | ||
| Metis | ||
| REPv2 | ||
| TRAC | ||
| BEPRO | ||
| Strict wins ↑ | 6 | 11 |
| Reported avg. ↓ | 0.044 | 0.034 |
Table 8. DOS ablation for edge-count regression. Values are means with standard deviations in parentheses over three runs. Lower is better; best means per dataset are in bold.
Transfer to Social Interaction Networks
To assess Hydra’s applicability beyond transaction networks, we adopt a cross-domain evaluation setting similar to MiNT, using the same social datasets and training/test split. Hydra-Social is trained on LastFM, SuperUser, Email-Eu, AskUbuntu, CollegeMsg, and StackOverflow; MathOverflow and RedditB are held out for zero-shot testing. Hydra-Mix combines social and transaction networks, using the same six social training networks together with 32 ERC-20 training networks. Both Hydra variants are evaluated without target-network training or fine-tuning, whereas HTGN is trained separately on each target network.
As shown in Table 9, Hydra-Mix improves over Hydra-Social on both held-out social networks. On RedditB, it achieves a mean AUC of 0.769 compared with 0.656 for target-trained HTGN, demonstrating effective transfer to an unseen social network without target-specific training. On MathOverflow, mixed-domain training improves Hydra’s mean AUC from 0.584 to 0.630, although it remains below HTGN’s 0.788. These results support Hydra’s applicability beyond transaction networks and the potential benefit of mixed-domain pre-training, while showing that transfer performance remains dependent on the target network.
| Model | MathOverflow | RedditB |
| HTGN | ||
| Hydra-Social | ||
| Hydra-Mix |
Table 9. Edge G/S on unseen social networks: mean AUC (± standard deviation) over three runs. Hydra variants transfer zero-shot; HTGN is target-trained. Best mean AUC per network is bold.
Further Analysis of Scaling and Model Capacity
To explore future scaling directions, we extend pre-training beyond the primary Hydra-32 configuration to 64 networks and evaluate on the same 20 unseen networks. With the base architecture, Hydra-64 maintains a similar mean Edge G/S AUC (0.7546 versus 0.7534), while mean edge-count MAE increases from 0.0457 to 0.0680. This task-dependent response motivates examining model capacity alongside pre-training scale. Per-network classification results, including standard deviations, are reported in Table 10.
| Dataset | Hydra-32 | Hydra-64 |
| MIR | 0.793 (0.002) | 0.796 (0.006) |
| DOGE2.0 | 0.897 (0.089) | 0.841 (0.009) |
| MUTE | 0.701 (0.098) | 0.764 (0.008) |
| EVERMOON | 0.818 (0.046) | 0.784 (0.102) |
| DERC | 0.839 (0.008) | 0.825 (0.011) |
| ADX | 0.722 (0.087) | 0.746 (0.031) |
| HOICHI | 0.591 (0.103) | 0.643 (0.050) |
| SDEX | 0.348 (0.066) | 0.343 (0.013) |
| BAG | 0.969 (0.008) | 0.960 (0.019) |
| XCN | 0.844 (0.012) | 0.836 (0.016) |
| ETH2x-FLI | 0.712 (0.045) | 0.732 (0.012) |
| stkAAVE | 0.732 (0.011) | 0.755 (0.017) |
| GLM | 0.850 (0.009) | 0.840 (0.003) |
| QOM | 0.745 (0.003) | 0.746 (0.014) |
| WOJAK | 0.585 (0.067) | 0.608 (0.026) |
| DINO | 0.895 (0.003) | 0.881 (0.018) |
| Metis | 0.733 (0.004) | 0.711 (0.028) |
| REPv2 | 0.772 (0.016) | 0.774 (0.015) |
| TRAC | 0.722 (0.001) | 0.719 (0.005) |
| BEPRO | 0.800 (0.002) | 0.787 (0.012) |
Table 10. Edge G/S on the 20 held-out networks: mean AUC ( standard deviation) over three runs. Best mean AUC per network is bold.
We increase the number of learnable layers alone or jointly increase the number of layers and the embedding dimension by factors of two and three. As shown in Figure 5, doubling both reduces Hydra-64’s mean edge-count MAE from 0.0680 to 0.0527, a 22.5% reduction, partially closing the gap to Hydra-32. Increasing layers alone yields smaller gains, and tripling both is less effective than doubling them. Hydra-32 therefore remains the stronger regression configuration in this comparison.
Figure 5. Effect of model capacity on Hydra-64 edge-count prediction. Values are means over three runs, averaged across 20 held-out networks; variability is not shown. The dashed line indicates Hydra-32’s base configuration. Lower MAE is better.
These findings highlight the importance of model capacity when expanding pre-training: jointly increasing depth and embedding dimension partially recovers regression performance, while increasing depth alone yields smaller gains. This provides a concrete basis for investigating how architectural choices interact with pre-training scale. A valuable future direction is to jointly study model capacity and training-network composition to identify when broader pre-training translates into improved transfer across tasks.
Scaling Behavior in Hydra
We conducted scaling-trend experiments to systematically evaluate HYDRA under different training pack sizes. Three model variations were trained using 8, 16, and 32 networks, enabling us to observe how increasing the number of networks in the training loop affects performance. This setup provides a consistent way to study HYDRA’s behavior under varying amounts of training data across multiple tasks. For each task, we present results for all three HYDRA variants. These results demonstrate how HYDRA scales with the number of training networks and are shown in the following tables. Each table provides a detailed breakdown by task and model variation, offering a comprehensive view of performance under various scaling setups. Table 11 reports the detailed results for each HYDRA variation trained with different datapacks across all six tasks. Subsections (a)–(c) present classification tasks (Node G/S, Edge G/S, LCC G/S), while (d)–(f) correspond to regression tasks (New Node Count, Influential Node Count, Edge Count).
Extended Network and Task Results
This section reports the full per-network results for all classification and regression tasks, using the same evaluation protocol as in the main text. These tables complement the summary figures by showing the complete distribution of baseline and HYDRA performance across datasets.
Classification. Extended classification results are provided for Edge-G/S, LLC-G/S, and Node-G/S in Table 12, Table 13, and Table 14, respectively.
| Dataset | HTGN | GC-LSTM | EvolveGCN | GraphPulse | ROLAND | TGCN | WinGNN | MiNT | HYDRA (Ours) |
| MIR | 0.750 ±0.005 | 0.768 ±0.026 | 0.745 ±0.015 | 0.689 ±0.097 | 0.228 ±0.060 | 0.749 ±0.026 | 0.742 ±0.015 | 0.836 ±0.016 | <u>0.793 ±0.002</u> |
| DOGE2.0 | 0.590 ±0.059 | 0.538 ±0.000 | 0.551 ±0.022 | 0.384 ±0.180 | 0.513 ±0.022 | 0.487 ±0.044 | 0.577 ±0.038 | 0.538 ±0.038 | 0.897 ±0.089 |
| MUTE | 0.649 ±0.015 | 0.593 ±0.030 | 0.617 ±0.010 | 0.779 ±0.004 | 0.289 ±0.042 | 0.557 ±0.068 | 0.593 ±0.054 | 0.673 ±0.013 | 0.701 ±0.098 |
| EVERMOON | 0.512 ±0.023 | 0.562 ±0.179 | 0.451 ±0.046 | 0.519 ±0.130 | 0.349 ±0.119 | 0.463 ±0.149 | 0.525 ±0.114 | 0.517 ±0.039 | 0.818 ±0.046 |
| DERC | 0.683 ±0.013 | 0.703 ±0.022 | 0.669 ±0.009 | 0.769 ±0.040 | 0.405 ±0.357 | 0.743 ±0.077 | 0.674 ±0.044 | 0.798 ±0.027 | 0.839 ±0.008 |
| ADX | 0.769 ±0.018 | 0.723 ±0.002 | 0.718 ±0.004 | 0.784 ±0.002 | 0.761 ±0.011 | 0.674 ±0.034 | 0.733 ±0.023 | 0.679 ±0.024 | 0.722 ±0.087 |
| HOICHI | 0.807 ±0.047 | 0.857 ±0.000 | 0.856 ±0.001 | 0.714 ±0.010 | 0.815 ±0.036 | 0.836 ±0.034 | 0.769 ±0.101 | 0.765 ±0.018 | 0.591 ±0.103 |
| SDEX | <u>0.762 ±0.034</u> | 0.720 ±0.002 | 0.733 ±0.028 | 0.436 ±0.030 | 0.483 ±0.254 | 0.759 ±0.039 | 0.726 ±0.000 | 0.614 ±0.020 | 0.348 ±0.066 |
| BAG | 0.673 ±0.227 | 0.196 ±0.179 | 0.329 ±0.040 | 0.934 ±0.020 | 0.418 ±0.016 | 0.334 ±0.171 | 0.485 ±0.105 | 0.931 ±0.028 | 0.969 ±0.008 |
| XCN | 0.668 ±0.099 | 0.306 ±0.092 | 0.512 ±0.067 | 0.821 ±0.004 | 0.765 ±0.015 | 0.703 ±0.037 | 0.586 ±0.029 | 0.851 ±0.043 | 0.844 ±0.012 |
| ETH2x-FLI | 0.610 ±0.059 | 0.670 ±0.009 | 0.688 ±0.010 | 0.666 ±0.047 | 0.621 ±0.023 | 0.647 ±0.020 | 0.617 ±0.056 | 0.729 ±0.015 | 0.712 ±0.045 |
| stkAAVE | 0.702 ±0.042 | 0.368 ±0.011 | 0.397 ±0.022 | 0.743 ±0.006 | 0.591 ±0.122 | 0.577 ±0.129 | 0.572 ±0.018 | 0.709 ±0.022 | 0.732 ±0.011 |
| GLM | 0.830 ±0.029 | 0.451 ±0.003 | 0.501 ±0.033 | 0.769 ±0.018 | 0.559 ±0.357 | 0.531 ±0.008 | 0.530 ±0.004 | 0.831 ±0.024 | 0.850 ±0.009 |
| QOM | 0.633 ±0.017 | 0.612 ±0.001 | 0.618 ±0.002 | 0.775 ±0.011 | 0.641 ±0.003 | 0.647 ±0.032 | 0.645 ±0.099 | 0.647 ±0.019 | 0.745 ±0.003 |
| WOJAK | 0.479 ±0.005 | 0.484 ±0.000 | 0.505 ±0.023 | 0.467 ±0.030 | 0.529 ±0.005 | 0.516 ±0.021 | 0.511 ±0.026 | 0.524 ±0.027 | 0.585 ±0.067 |
| DINO | 0.730 ±0.195 | 0.874 ±0.028 | 0.868 ±0.029 | 0.801 ±0.020 | 0.497 ±0.092 | 0.544 ±0.314 | 0.628 ±0.251 | 0.779 ±0.113 | 0.895 ±0.003 |
| Metis | 0.715 ±0.122 | 0.646 ±0.023 | 0.688 ±0.027 | 0.812 ±0.011 | 0.696 ±0.108 | 0.709 ±0.033 | 0.690 ±0.039 | 0.760 ±0.025 | 0.733 ±0.004 |
| REPv2 | 0.760 ±0.012 | 0.725 ±0.014 | 0.709 ±0.002 | 0.830 ±0.001 | 0.751 ±0.003 | 0.696 ±0.035 | 0.744 ±0.026 | 0.789 ±0.020 | 0.772 ±0.016 |
| TRAC | 0.712 ±0.071 | 0.748 ±0.000 | 0.748 ±0.000 | 0.767 ±0.001 | 0.495 ±0.223 | 0.741 ±0.012 | 0.752 ±0.007 | 0.785 ±0.008 | 0.722 ±0.001 |
| BEPRO | 0.655 ±0.038 | 0.632 ±0.019 | 0.610 ±0.012 | 0.783 ±0.003 | 0.439 ±0.125 | 0.744 ±0.074 | 0.736 ±0.018 | 0.782 ±0.003 | 0.800 ±0.002 |
| 1st-Place Count↑ | 1 | 1 | 0 | 6 | 0 | 0 | 0 | 4 | 8 |
| Avg. Rank↓ | 4.85 | 5.80 | 6.10 | 3.80 | 6.30 | 5.60 | 5.55 | 3.30 | 2.80 |
| Avg. AUC↑ | 0.684 | 0.609 | 0.626 | 0.712 | 0.542 | 0.633 | 0.642 | 0.727 | 0.753 |
Table 12. AUC results for the Edge Growth/Shrinkage prediction task (classification). Best results are in bold, second best are underlined.
| Dataset | HTGN | GC-LSTM | EvolveGCN | GraphPulse | ROLAND | TGCN | WinGNN | MiNT | HYDRA (Ours) |
| MIR | 0.745 ±0.023 | 0.585 ±0.128 | 0.575 ±0.146 | 0.800 ±0.008 | 0.536 ±0.275 | 0.585 ±0.055 | 0.749 ±0.020 | 0.845 ±0.035 | 0.815 ±0.007 |
| DOGE2.0 | 0.446 ±0.164 | 0.387 ±0.294 | 0.583 ±0.115 | 0.333 ±0.042 | 0.411 ±0.232 | 0.464 ±0.182 | 0.595 ±0.176 | 0.661 ±0.047 | 0.702 ±0.160 |
| MUTE | 0.574 ±0.022 | 0.579 ±0.022 | 0.578 ±0.033 | 0.647 ±0.014 | 0.624 ±0.037 | 0.567 ±0.007 | 0.641 ±0.061 | 0.582 ±0.078 | 0.704 ±0.085 |
| EVERMOON | 0.494 ±0.127 | 0.512 ±0.112 | 0.548 ±0.152 | 0.463 ±0.034 | 0.491 ±0.157 | 0.624 ±0.004 | 0.603 ±0.041 | 0.527 ±0.118 | 0.667 ±0.045 |
| DERC | 0.717 ±0.035 | 0.591 ±0.010 | 0.553 ±0.044 | 0.727 ±0.009 | 0.481 ±0.131 | 0.523 ±0.103 | 0.582 ±0.043 | 0.689 ±0.096 | 0.808 ±0.022 |
| ADX | 0.753 ±0.013 | 0.599 ±0.012 | 0.604 ±0.030 | 0.661 ±0.006 | 0.606 ±0.189 | 0.621 ±0.017 | 0.611 ±0.062 | 0.587 ±0.014 | 0.727 ±0.093 |
| HOICHI | 0.746 ±0.010 | 0.749 ±0.001 | 0.745 ±0.003 | 0.730 ±0.017 | 0.360 ±0.121 | 0.750 ±0.002 | 0.635 ±0.183 | 0.722 ±0.034 | 0.638 ±0.075 |
| SDEX | 0.911 ±0.104 | 0.721 ±0.138 | 0.601 ±0.105 | 0.808 ±0.050 | 0.425 ±0.047 | 0.770 ±0.231 | 0.555 ±0.282 | 0.382 ±0.280 | 0.816 ±0.025 |
| BAG | 0.493 ±0.043 | 0.291 ±0.180 | 0.480 ±0.052 | 0.900 ±0.010 | 0.463 ±0.019 | 0.463 ±0.141 | 0.490 ±0.080 | 0.893 ±0.074 | 0.976 ±0.006 |
| XCN | 0.566 ±0.199 | 0.481 ±0.160 | 0.533 ±0.257 | 0.681 ±0.005 | 0.569 ±0.204 | 0.638 ±0.045 | 0.549 ±0.133 | 0.827 ±0.025 | 0.887 ±0.014 |
| ETH2x-FLI | 0.561 ±0.037 | 0.529 ±0.017 | 0.547 ±0.009 | 0.653 ±0.047 | 0.499 ±0.135 | 0.549 ±0.019 | 0.505 ±0.090 | 0.618 ±0.025 | 0.687 ±0.032 |
| stkAAVE | 0.623 ±0.077 | 0.581 ±0.085 | 0.551 ±0.102 | 0.662 ±0.004 | 0.532 ±0.140 | 0.543 ±0.102 | 0.489 ±0.105 | 0.688 ±0.019 | 0.748 ±0.012 |
| GLM | 0.761 ±0.031 | 0.481 ±0.073 | 0.636 ±0.123 | 0.749 ±0.014 | 0.802 ±0.037 | 0.425 ±0.005 | 0.489 ±0.079 | 0.818 ±0.074 | 0.848 ±0.012 |
| QOM | 0.658 ±0.150 | 0.509 ±0.100 | 0.562 ±0.022 | 0.747 ±0.006 | 0.627 ±0.134 | 0.419 ±0.044 | 0.546 ±0.152 | 0.645 ±0.109 | 0.729 ±0.010 |
| WOJAK | 0.378 ±0.028 | 0.489 ±0.133 | 0.394 ±0.079 | 0.550 ±0.036 | 0.360 ±0.005 | 0.481 ±0.092 | 0.414 ±0.172 | 0.492 ±0.107 | 0.500 ±0.083 |
| DINO | 0.706 ±0.120 | 0.796 ±0.023 | 0.710 ±0.034 | 0.661 ±0.026 | 0.523 ±0.238 | 0.773 ±0.043 | 0.731 ±0.037 | 0.561 ±0.006 | 0.818 ±0.049 |
| Metis | 0.679 ±0.039 | 0.687 ±0.018 | 0.672 ±0.016 | 0.783 ±0.007 | 0.672 ±0.103 | 0.657 ±0.014 | 0.634 ±0.042 | 0.780 ±0.041 | 0.731 ±0.001 |
| REPv2 | 0.730 ±0.007 | 0.653 ±0.015 | 0.644 ±0.027 | 0.752 ±0.001 | 0.658 ±0.103 | 0.646 ±0.025 | 0.683 ±0.014 | 0.742 ±0.041 | 0.764 ±0.020 |
| TRAC | 0.733 ±0.009 | 0.629 ±0.005 | 0.623 ±0.004 | 0.686 ±0.001 | 0.606 ±0.171 | 0.620 ±0.005 | 0.599 ±0.026 | 0.762 ±0.028 | 0.781 ±0.004 |
| BEPRO | 0.694 ±0.009 | 0.595 ±0.008 | 0.557 ±0.058 | 0.725 ±0.004 | 0.482 ±0.146 | 0.536 ±0.031 | 0.582 ±0.063 | 0.628 ±0.017 | 0.830 ±0.006 |
| 1st-Place Count↑ | 2 | 0 | 0 | 3 | 0 | 1 | 0 | 1 | 8 |
| Avg. Rank↓ | 4.40 | 6.00 | 6.30 | 3.50 | 6.95 | 6.10 | 6.10 | 3.95 | 1.70 |
| Avg. AUC↑ | 0.648 | 0.572 | 0.585 | 0.686 | 0.556 | 0.583 | 0.585 | 0.672 | 0.759 |
Table 13. AUC results for the LCC Growth/Shrinkage prediction task (classification). Best results are in bold, second best are underlined.
| Dataset | HTGN | GC-LSTM | EvolveGCN | GraphPulse | ROLAND | TGCN | WinGNN | MiNT | HYDRA (Ours) |
| MIR | 0.545 ±0.030 | 0.537 ±0.033 | 0.528 ±0.081 | 0.633 ±0.066 | 0.472 ±0.040 | 0.532 ±0.021 | 0.525 ±0.043 | 0.622 ±0.030 | 0.764 ±0.001 |
| DOGE2.0 | 0.427 ±0.065 | 0.633 ±0.014 | 0.400 ±0.061 | 0.403 ±0.035 | 0.260 ±0.000 | 0.627 ±0.034 | 0.693 ±0.044 | 0.750 ±0.014 | 0.763 ±0.064 |
| MUTE | 0.518 ±0.023 | 0.448 ±0.008 | 0.475 ±0.051 | 0.677 ±0.008 | 0.411 ±0.093 | 0.458 ±0.009 | 0.563 ±0.015 | 0.606 ±0.056 | 0.748 ±0.025 |
| EVERMOON | 0.585 ±0.059 | 0.606 ±0.021 | 0.488 ±0.088 | 0.463 ±0.002 | 0.427 ±0.137 | 0.567 ±0.030 | 0.548 ±0.116 | 0.614 ±0.071 | 0.655 ±0.040 |
| DERC | 0.662 ±0.051 | 0.492 ±0.069 | 0.503 ±0.084 | 0.611 ±0.049 | 0.551 ±0.113 | 0.447 ±0.003 | 0.517 ±0.044 | 0.569 ±0.004 | 0.742 ±0.004 |
| ADX | 0.678 ±0.017 | 0.505 ±0.043 | 0.509 ±0.022 | 0.701 ±0.003 | 0.557 ±0.082 | 0.484 ±0.048 | 0.504 ±0.018 | 0.507 ±0.037 | 0.718 ±0.051 |
| HOICHI | 0.687 ±0.004 | 0.718 ±0.007 | 0.685 ±0.020 | 0.745 ±0.006 | 0.347 ±0.084 | 0.718 ±0.002 | 0.526 ±0.188 | 0.492 ±0.120 | 0.558 ±0.047 |
| SDEX | 0.824 ±0.106 | 0.364 ±0.148 | 0.817 ±0.032 | 0.865 ±0.011 | 0.779 ±0.018 | 0.755 ±0.202 | 0.757 ±0.072 | 0.861 ±0.025 | 0.743 ±0.026 |
| BAG | 0.735 ±0.075 | 0.337 ±0.089 | 0.166 ±0.066 | 0.897 ±0.016 | 0.390 ±0.088 | 0.391 ±0.219 | 0.351 ±0.098 | 0.685 ±0.038 | 0.969 ±0.009 |
| XCN | 0.476 ±0.012 | 0.466 ±0.012 | 0.407 ±0.176 | 0.671 ±0.020 | 0.430 ±0.144 | 0.483 ±0.036 | 0.355 ±0.017 | 0.505 ±0.006 | 0.878 ±0.009 |
| ETH2x-FLI | 0.628 ±0.022 | 0.548 ±0.001 | 0.548 ±0.002 | 0.615 ±0.020 | 0.488 ±0.063 | 0.553 ±0.036 | 0.586 ±0.098 | 0.411 ±0.066 | 0.678 ±0.031 |
| stkAAVE | 0.517 ±0.039 | 0.543 ±0.043 | 0.456 ±0.069 | 0.643 ±0.005 | 0.661 ±0.037 | 0.425 ±0.029 | 0.466 ±0.056 | 0.561 ±0.007 | 0.779 ±0.007 |
| GLM | 0.706 ±0.014 | 0.566 ±0.014 | 0.516 ±0.105 | 0.595 ±0.003 | 0.493 ±0.149 | 0.575 ±0.019 | 0.610 ±0.024 | 0.720 ±0.045 | 0.763 ±0.005 |
| QOM | 0.647 ±0.094 | 0.492 ±0.003 | 0.485 ±0.001 | 0.705 ±0.002 | 0.592 ±0.080 | 0.495 ±0.004 | 0.409 ±0.051 | 0.572 ±0.017 | 0.719 ±0.012 |
| WOJAK | 0.417 ±0.143 | 0.338 ±0.068 | 0.357 ±0.104 | 0.500 ±0.000 | 0.202 ±0.018 | 0.488 ±0.080 | 0.314 ±0.029 | 0.618 ±0.035 | 0.412 ±0.060 |
| DINO | 0.845 ±0.015 | 0.323 ±0.148 | 0.444 ±0.052 | 0.686 ±0.007 | 0.330 ±0.115 | 0.615 ±0.070 | 0.600 ±0.292 | 0.735 ±0.005 | 0.910 ±0.013 |
| Metis | 0.589 ±0.049 | 0.483 ±0.052 | 0.566 ±0.012 | 0.652 ±0.029 | 0.574 ±0.040 | 0.549 ±0.011 | 0.510 ±0.006 | 0.616 ±0.012 | 0.693 ±0.007 |
| REPv2 | 0.650 ±0.004 | 0.519 ±0.023 | 0.515 ±0.019 | 0.662 ±0.008 | 0.597 ±0.028 | 0.534 ±0.029 | 0.626 ±0.058 | 0.710 ±0.129 | 0.689 ±0.013 |
| TRAC | 0.670 ±0.031 | 0.527 ±0.016 | 0.524 ±0.003 | 0.610 ±0.055 | 0.546 ±0.024 | 0.584 ±0.011 | 0.522 ±0.057 | 0.600 ±0.027 | 0.675 ±0.003 |
| BEPRO | 0.500 ±0.055 | 0.332 ±0.010 | 0.356 ±0.015 | 0.707 ±0.005 | 0.490 ±0.016 | 0.372 ±0.100 | 0.420 ±0.018 | 0.561 ±0.022 | 0.865 ±0.010 |
| 1st-Place Count↑ | 0 | 0 | 0 | 2 | 0 | 0 | 0 | 3 | 15 |
| Avg. Rank↓ | 3.65 | 6.65 | 7.00 | 2.95 | 6.70 | 6.15 | 6.20 | 3.55 | 1.95 |
| Avg. AUC↑ | 0.615 | 0.489 | 0.487 | 0.652 | 0.480 | 0.530 | 0.528 | 0.616 | 0.734 |
Table 14. AUC results for the Node Growth/Shrinkage prediction task (classification). Best results are in bold, second best are underlined.
Regression. Extended regression results are provided for Edge Count, New Node Count, and Influential Node Count in Table 15, Table 16, and Table 17, respectively.
| Dataset | HTGN | TGCN | GCLSTM | ROLAND | EGCN | GraphPulse | WinGNN | HYDRA (Ours) |
| MIR | 0.059 (±0.007) | 0.044 (±0.009) | 0.047 (±0.003) | 0.039 (±0.000) | 0.057 (±0.015) | 0.059 (±0.001) | 0.046 (±0.010) | 0.016 (±0.005) |
| DOGE2.0 | 0.101 (±0.035) | 0.063 (±0.017) | 0.106 (±0.029) | 0.052 (±0.003) | 0.092 (±0.031) | 0.046 (±0.000) | 0.045 (±0.003) | 0.187 (±0.008) |
| MUTE | 0.025 (±0.006) | 0.038 (±0.004) | 0.017 (±0.001) | 0.040 (±0.007) | 0.049 (±0.005) | 0.025 (±0.002) | 0.027 (±0.003) | 0.021 (±0.022) |
| EVERMOON | 0.010 (±0.001) | 0.021 (±0.013) | 0.025 (±0.007) | 0.016 (±0.016) | 0.030 (±0.010) | 0.235 (±0.005) | 0.025 (±0.004) | 0.017 (±0.000) |
| DERC | 0.038 (±0.015) | 0.059 (±0.011) | 0.016 (±0.005) | 0.060 (±0.008) | 0.023 (±0.007) | 0.023 (±0.003) | 0.032 (±0.001) | 0.021 (±0.003) |
| ADX | 0.017 (±0.001) | 0.018 (±0.003) | 0.016 (±0.001) | 0.017 (±0.002) | 0.021 (±0.002) | 0.019 (±0.001) | 0.016 (±0.000) | 0.025 (±0.008) |
| HOICHI | 0.034 (±0.010) | 0.020 (±0.001) | 0.046 (±0.013) | 0.020 (±0.003) | 0.034 (±0.013) | 0.044 (±0.002) | 0.028 (±0.005) | 0.027 (±0.020) |
| SDEX | 0.080 (±0.029) | 0.128 (±0.046) | 0.058 (±0.007) | 0.121 (±0.002) | 0.085 (±0.025) | 0.106 (±0.005) | 0.128 (±0.008) | 0.095 (±0.043) |
| BAG | 0.022 (±0.003) | 0.023 (±0.017) | 0.025 (±0.002) | 0.030 (±0.016) | 0.027 (±0.005) | 0.063 (±0.001) | 0.260 (±0.064) | 0.041 (±0.020) |
| XCN | 0.074 (±0.006) | 0.112 (±0.042) | 0.107 (±0.024) | 0.120 (±0.035) | 0.121 (±0.011) | 0.118 (±0.000) | 0.072 (±0.018) | 0.062 (±0.015) |
| ETH2x-FLI | 0.055 (±0.015) | 0.079 (±0.019) | 0.177 (±0.057) | 0.040 (±0.026) | 0.066 (±0.016) | 0.144 (±0.009) | 0.030 (±0.011) | 0.023 (±0.001) |
| stkAAVE | 0.083 (±0.008) | 0.087 (±0.005) | 0.092 (±0.017) | 0.104 (±0.012) | 0.079 (±0.010) | 0.096 (±0.006) | 0.100 (±0.004) | 0.051 (±0.022) |
| GLM | 0.072 (±0.010) | 0.058 (±0.005) | 0.063 (±0.002) | 0.058 (±0.003) | 0.060 (±0.001) | 0.076 (±0.001) | 0.054 (±0.003) | 0.087 (±0.008) |
| QOM | 0.042 (±0.005) | 0.085 (±0.020) | 0.053 (±0.030) | 0.057 (±0.030) | 0.069 (±0.013) | 0.055 (±0.001) | 0.046 (±0.006) | 0.036 (±0.013) |
| WOJAK | 0.009 (±0.002) | 0.012 (±0.003) | 0.013 (±0.004) | 0.016 (±0.008) | 0.013 (±0.007) | 0.057 (±0.008) | 0.006 (±0.002) | 0.010 (±0.004) |
| DINO | 0.069 (±0.020) | 0.025 (±0.009) | 0.040 (±0.019) | 0.014 (±0.004) | 0.039 (±0.011) | 0.087 (±0.002) | 0.021 (±0.008) | 0.017 (±0.007) |
| Metis | 0.038 (±0.002) | 0.054 (±0.001) | 0.047 (±0.003) | 0.057 (±0.008) | 0.053 (±0.004) | 0.066 (±0.006) | 0.043 (±0.001) | 0.034 (±0.010) |
| REPv2 | 0.117 (±0.013) | 0.108 (±0.004) | 0.115 (±0.004) | 0.106 (±0.001) | 0.128 (±0.036) | 0.119 (±0.001) | 0.118 (±0.001) | 0.111 (±0.008) |
| TRAC | 0.026 (±0.004) | 0.036 (±0.010) | 0.061 (±0.006) | 0.023 (±0.004) | 0.036 (±0.003) | 0.017 (±0.000) | 0.040 (±0.014) | 0.021 (±0.001) |
| BEPRO | 0.009 (±0.001) | 0.009 (±0.003) | 0.009 (±0.002) | 0.015 (±0.017) | 0.007 (±0.001) | 0.007 (±0.000) | 0.007 (±0.002) | 0.011 (±0.008) |
| 1st-Place Count↑ | 2 | 1 | 4 | 3 | 1 | 2 | 5 | 6 |
| Avg. Rank↓ | 3.95 | 4.60 | 4.58 | 4.58 | 5.30 | 5.72 | 3.93 | 3.35 |
| Avg. MAE↓ | 0.049 | 0.054 | 0.057 | 0.050 | 0.054 | 0.073 | 0.057 | 0.046 |
Table 15. MAE results for the Edge Count prediction task (regression). Best results are in bold, second best are underlined.
| Dataset | HTGN | TGCN | GCLSTM | ROLAND | EGCN | GraphPulse | WinGNN | HYDRA (ours) |
| MIR | 0.031 (±0.002) | 0.028 (±0.001) | 0.039 (±0.005) | 0.025 (±0.018) | 0.030 (±0.001) | 0.025 (±0.001) | 0.037 (±0.011) | 0.013 (±0.004) |
| DOGE2.0 | 0.046 (±0.009) | 0.064 (±0.021) | 0.044 (±0.014) | 0.073 (±0.019) | 0.051 (±0.031) | 0.157 (±0.010) | 0.098 (±0.009) | 0.092 (±0.008) |
| MUTE | 0.021 (±0.003) | 0.041 (±0.005) | 0.030 (±0.003) | 0.048 (±0.002) | 0.053 (±0.003) | 0.031 (±0.001) | 0.051 (±0.015) | 0.025 (±0.005) |
| EVERMOON | 0.010 (±0.004) | 0.017 (±0.008) | 0.026 (±0.009) | 0.029 (±0.019) | 0.028 (±0.017) | 0.022 (±0.005) | 0.022 (±0.002) | 0.012 (±0.005) |
| DERC | 0.028 (±0.001) | 0.034 (±0.018) | 0.016 (±0.024) | 0.043 (±0.024) | 0.019 (±0.004) | 0.043 (±0.011) | 0.027 (±0.003) | 0.015 (±0.004) |
| ADX | 0.014 (±0.001) | 0.010 (±0.001) | 0.011 (±0.001) | 0.024 (±0.029) | 0.012 (±0.000) | 0.024 (±0.002) | 0.011 (±0.004) | 0.016 (±0.005) |
| HOICHI | 0.044 (±0.004) | 0.035 (±0.025) | 0.053 (±0.028) | 0.042 (±0.028) | 0.039 (±0.023) | 0.066 (±0.001) | 0.027 (±0.004) | 0.029 (±0.008) |
| SDEX | 0.075 (±0.002) | 0.093 (±0.006) | 0.069 (±0.002) | 0.088 (±0.002) | 0.077 (±0.008) | 0.087 (±0.002) | 0.090 (±0.025) | 0.063 (±0.006) |
| BAG | 0.023 (±0.007) | 0.031 (±0.008) | 0.019 (±0.007) | 0.053 (±0.006) | 0.030 (±0.004) | 0.054 (±0.000) | 0.230 (±0.062) | 0.052 (±0.002) |
| XCN | 0.014 (±0.007) | 0.015 (±0.006) | 0.017 (±0.007) | 0.012 (±0.012) | 0.017 (±0.007) | 0.040 (±0.001) | 0.015 (±0.002) | 0.009 (±0.002) |
| ETH2x-FLI | 0.031 (±0.010) | 0.041 (±0.001) | 0.069 (±0.010) | 0.020 (±0.004) | 0.030 (±0.002) | 0.092 (±0.001) | 0.028 (±0.003) | 0.030 (±0.004) |
| stkAAVE | 0.128 (±0.018) | 0.136 (±0.008) | 0.128 (±0.018) | 0.154 (±0.005) | 0.124 (±0.005) | 0.151 (±0.000) | 0.147 (±0.008) | 0.078 (±0.017) |
| GLM | 0.066 (±0.000) | 0.068 (±0.001) | 0.068 (±0.000) | 0.068 (±0.002) | 0.067 (±0.002) | 0.092 (±0.000) | 0.066 (±0.000) | 0.094 (±0.007) |
| QOM | 0.035 (±0.006) | 0.038 (±0.010) | 0.032 (±0.020) | 0.018 (±0.010) | 0.035 (±0.020) | 0.033 (±0.004) | 0.029 (±0.007) | 0.012 (±0.001) |
| WOJAK | 0.008 (±0.001) | 0.009 (±0.001) | 0.015 (±0.003) | 0.029 (±0.017) | 0.014 (±0.003) | 0.067 (±0.005) | 0.007 (±0.001) | 0.009 (±0.000) |
| DINO | 0.061 (±0.017) | 0.024 (±0.002) | 0.028 (±0.019) | 0.013 (±0.005) | 0.030 (±0.005) | 0.085 (±0.001) | 0.051 (±0.029) | 0.018 (±0.007) |
| Metis | 0.034 (±0.010) | 0.045 (±0.006) | 0.041 (±0.006) | 0.043 (±0.005) | 0.042 (±0.004) | 0.054 (±0.001) | 0.034 (±0.002) | 0.034 (±0.010) |
| REPv2 | 0.061 (±0.003) | 0.061 (±0.004) | 0.075 (±0.003) | 0.055 (±0.002) | 0.063 (±0.004) | 0.068 (±0.000) | 0.063 (±0.000) | 0.066 (±0.004) |
| TRAC | 0.043 (±0.021) | 0.030 (±0.003) | 0.071 (±0.021) | 0.021 (±0.006) | 0.025 (±0.009) | 0.018 (±0.000) | 0.044 (±0.010) | 0.022 (±0.005) |
| BEPRO | 0.012 (±0.002) | 0.010 (±0.002) | 0.011 (±0.017) | 0.011 (±0.011) | 0.009 (±0.000) | 0.012 (±0.000) | 0.006 (±0.001) | 0.004 (±0.001) |
| 1st-Place Count↑ | 4 | 1 | 1 | 0 | 0 | 1 | 4 | 8 |
| Avg. Rank↓ | 3.45 | 4.10 | 4.00 | 4.35 | 3.90 | 5.75 | 4.05 | 2.50 |
| Avg. MAE↓ | 0.039 | 0.042 | 0.043 | 0.052 | 0.040 | 0.071 | 0.054 | 0.035 |
Table 16. MAE results for the New Node Count prediction task (regression). Best results are in bold, second best are underlined.
Table 17: MAE results for the Influential Node Count prediction task (regression). Best results are in bold, second best are <u>underlined</u>.
| Dataset | HTGN | TGCN | GCLSTM | ROLAND | EGCN | GraphPulse | WinGNN | HYDRA (ours) |
| MIR | 0.114 (±0.003) | 0.119 (±0.005) | 0.105 (±0.020) | 0.115 (±0.020) | 0.114 (±0.017) | 0.127 (±0.002) | <u>0.082 (±0.007)</u> | 0.039 (±0.005) |
| DOGE2.0 | 0.064 (±0.031) | 0.090 (±0.023) | 0.053 (±0.021) | 0.070 (±0.015) | 0.059 (±0.021) | 0.087 (±0.004) | 0.091 (±0.009) | 0.126 (±0.032) |
| MUTE | 0.028 (±0.004) | 0.042 (±0.002) | 0.018 (±0.002) | 0.051 (±0.006) | 0.042 (±0.012) | <u>0.021 (±0.001)</u> | 0.045 (±0.033) | 0.045 (±0.027) |
| EVERMOON | 0.011 (±0.002) | <u>0.014 (±0.006)</u> | 0.018 (±0.008) | 0.014 (±0.004) | 0.032 (±0.015) | 0.235 (±0.004) | 0.026 (±0.006) | 0.038 (±0.024) |
| DERC | 0.069 (±0.007) | 0.084 (±0.001) | 0.053 (±0.002) | 0.104 (±0.011) | 0.058 (±0.010) | <u>0.048 (±0.001)</u> | 0.077 (±0.003) | 0.033 (±0.017) |
| ADX | 0.016 (±0.003) | <u>0.015 (±0.000)</u> | 0.012 (±0.001) | 0.022 (±0.009) | <u>0.015 (±0.001)</u> | 0.020 (±0.001) | <u>0.015 (±0.004)</u> | 0.030 (±0.001) |
| HOICHI | 0.039 (±0.011) | 0.020 (±0.007) | 0.046 (±0.017) | 0.030 (±0.015) | 0.034 (±0.016) | 0.047 (±0.009) | <u>0.024 (±0.016)</u> | 0.055 (±0.028) |
| SDEX | <u>0.037 (±0.011)</u> | 0.049 (±0.022) | 0.031 (±0.012) | 0.058 (±0.017) | <u>0.037 (±0.005)</u> | 0.067 (±0.009) | 0.066 (±0.025) | 0.042 (±0.019) |
| BAG | 0.009 (±0.000) | 0.017 (±0.002) | 0.030 (±0.008) | 0.019 (±0.010) | 0.030 (±0.009) | 0.064 (±0.001) | 0.074 (±0.094) | 0.075 (±0.027) |
| XCN | 0.061 (±0.008) | 0.155 (±0.009) | <u>0.055 (±0.003)</u> | 0.124 (±0.005) | 0.066 (±0.008) | 0.080 (±0.004) | 0.159 (±0.029) | 0.126 (±0.033) |
| ETH2x-FLI | 0.050 (±0.007) | 0.053 (±0.035) | 0.100 (±0.014) | 0.023 (±0.005) | 0.044 (±0.010) | 0.108 (±0.003) | <u>0.033 (±0.010)</u> | <u>0.033 (±0.020)</u> |
| stkAAVE | 0.062 (±0.001) | 0.080 (±0.009) | 0.062 (±0.001) | <u>0.057 (±0.005)</u> | 0.061 (±0.006) | 0.108 (±0.009) | 0.066 (±0.003) | 0.042 (±0.013) |
| GLM | 0.066 (±0.002) | 0.057 (±0.001) | 0.067 (±0.004) | <u>0.057 (±0.005)</u> | 0.054 (±0.002) | 0.082 (±0.000) | 0.053 (±0.002) | 0.075 (±0.017) |
| QOM | 0.060 (±0.004) | 0.088 (±0.040) | 0.061 (±0.032) | 0.074 (±0.018) | 0.075 (±0.015) | 0.076 (±0.004) | <u>0.062 (±0.012)</u> | 0.039 (±0.020) |
| WOJAK | 0.007 (±0.001) | <u>0.008 (±0.003)</u> | 0.012 (±0.003) | 0.027 (±0.010) | 0.011 (±0.004) | 0.066 (±0.019) | <u>0.007 (±0.001)</u> | 0.036 (±0.027) |
| DINO | 0.078 (±0.036) | 0.024 (±0.004) | 0.068 (±0.001) | <u>0.024 (±0.003)</u> | 0.036 (±0.007) | 0.087 (±0.003) | 0.022 (±0.000) | 0.034 (±0.002) |
| Metis | 0.056 (±0.003) | 0.078 (±0.004) | 0.068 (±0.005) | 0.085 (±0.030) | 0.073 (±0.004) | 0.071 (±0.010) | <u>0.063 (±0.001)</u> | 0.046 (±0.008) |
| REPv2 | 0.115 (±0.000) | 0.118 (±0.006) | 0.119 (±0.002) | <u>0.115 (±0.001)</u> | 0.138 (±0.027) | 0.127 (±0.000) | 0.126 (±0.001) | 0.129 (±0.022) |
| TRAC | 0.045 (±0.005) | <u>0.058 (±0.002)</u> | 0.079 (±0.002) | 0.055 (±0.009) | 0.056 (±0.005) | 0.033 (±0.005) | 0.074 (±0.009) | 0.043 (±0.023) |
| BEPRO | 0.016 (±0.006) | 0.014 (±0.003) | 0.012 (±0.002) | 0.013 (±0.001) | 0.012 (±0.002) | 0.012 (±0.000) | 0.010 (±0.001) | 0.033 (±0.018) |
| 1<sup>st</sup>-Place Count↑ | 4 | 1 | 5 | 2 | 0 | 1 | 4 | 5 |
| Avg. Rank↓ | 3.52 | 4.85 | 3.80 | 4.55 | 4.20 | 5.90 | 4.28 | 4.90 |
| Avg. MAE↓ | 0.050 | 0.059 | 0.053 | 0.057 | 0.052 | 0.078 | 0.059 | 0.056 |
Table 17.
For classification, HYDRA achieves the highest AUC on the majority of datasets across Edge-G/S, LLC-G/S, and Node-G/S. For regression HYDRA consistently delivers the lowest or near-lowest MAE, remaining competitive even on tasks where another baseline occasionally leads. These results confirm that HYDRA’s advantages are not limited to averages: the model transfers robustly across heterogeneous temporal networks and maintains strong performance in the zero-shot setting.
| Dataset | HYDRA-8 | HYDRA-16 | HYDRA-32 |
| MIR | 0.800 (±0.007) | 0.796 (±0.001) | 0.793 (±0.002) |
| DOGE2.0 | 0.859 (±0.022) | 0.897 (±0.089) | 0.897 (±0.089) |
| MUTE | 0.757 (±0.004) | 0.764 (±0.004) | 0.701 (±0.098) |
| EVERMOON | 0.698 (±0.180) | 0.750 (±0.024) | 0.818 (±0.046) |
| DERC | 0.823 (±0.018) | 0.826 (±0.002) | 0.839 (±0.008) |
| ADX | 0.760 (±0.004) | 0.768 (±0.015) | 0.722 (±0.087) |
| HOICHI | 0.606 (±0.015) | 0.617 (±0.026) | 0.591 (±0.103) |
| SDEX | 0.288 (±0.023) | 0.331 (±0.049) | 0.348 (±0.066) |
| BAG | 0.969 (±0.009) | 0.955 (±0.015) | 0.969 (±0.008) |
| XCN | 0.847 (±0.006) | 0.845 (±0.010) | 0.844 (±0.012) |
| ETH2x-FLI | 0.735 (±0.006) | 0.738 (±0.005) | 0.712 (±0.045) |
| stkAAVE | 0.744 (±0.008) | 0.743 (±0.007) | 0.732 (±0.011) |
| GLM | 0.849 (±0.003) | 0.841 (±0.015) | 0.850 (±0.009) |
| QOM | 0.755 (±0.015) | 0.762 (±0.008) | 0.745 (±0.003) |
| WOJAK | 0.627 (±0.014) | 0.561 (±0.101) | 0.585 (±0.067) |
| DINO | 0.875 (±0.030) | 0.889 (±0.009) | 0.895 (±0.003) |
| Metis | 0.735 (±0.004) | 0.697 (±0.062) | 0.733 (±0.004) |
| REPv2 | 0.778 (±0.001) | 0.784 (±0.009) | 0.772 (±0.016) |
| TRAC | 0.713 (±0.010) | 0.711 (±0.011) | 0.722 (±0.001) |
| BEPRO | 0.783 (±0.009) | 0.743 (±0.085) | 0.800 (±0.002) |
| 1st-Place Count↑ | 6 | 7 | 9 |
| Avg. Rank↓ | 1.95 | 1.95 | 2 |
| Avg. AUC↑ | 0.75 | 0.751 | 0.753 |
Table 11b. (b) Classification: Edge G/S
| Dataset | HYDRA-8 | HYDRA-16 | HYDRA-32 |
| MIR | 0.819 (±0.004) | 0.817 (±0.001) | 0.815 (±0.007) |
| DOGE2.0 | 0.762 (±0.010) | 0.667 (±0.175) | 0.702 (±0.160) |
| MUTE | 0.714 (±0.049) | 0.756 (±0.006) | 0.704 (±0.085) |
| EVERMOON | 0.600 (±0.057) | 0.639 (±0.029) | 0.667 (±0.045) |
| DERC | 0.819 (±0.003) | 0.813 (±0.005) | 0.808 (±0.022) |
| ADX | 0.691 (±0.066) | 0.770 (±0.021) | 0.727 (±0.093) |
| HOICHI | 0.610 (±0.044) | 0.640 (±0.041) | 0.638 (±0.075) |
| SDEX | 0.784 (±0.017) | 0.810 (±0.048) | 0.816 (±0.025) |
| BAG | 0.968 (±0.013) | 0.957 (±0.021) | 0.976 (±0.006) |
| XCN | 0.863 (±0.011) | 0.863 (±0.029) | 0.887 (±0.014) |
| ETH2x-FLI | 0.708 (±0.011) | 0.703 (±0.001) | 0.687 (±0.032) |
| stkAAVE | 0.753 (±0.011) | 0.757 (±0.003) | 0.748 (±0.012) |
| GLM | 0.853 (±0.006) | 0.844 (±0.020) | 0.848 (±0.012) |
| QOM | 0.725 (±0.009) | 0.730 (±0.002) | 0.729 (±0.010) |
| WOJAK | 0.558 (±0.032) | 0.458 (±0.142) | 0.500 (±0.083) |
| DINO | 0.878 (±0.021) | 0.883 (±0.032) | 0.818 (±0.049) |
| Metis | 0.727 (±0.006) | 0.669 (±0.100) | 0.731 (±0.001) |
| REPv2 | 0.781 (±0.004) | 0.785 (±0.009) | 0.764 (±0.020) |
| TRAC | 0.772 (±0.006) | 0.768 (±0.008) | 0.781 (±0.004) |
| BEPRO | 0.826 (±0.015) | 0.800 (±0.058) | 0.830 (±0.006) |
| 1st-Place Count↑ | 6 | 7 | 7 |
| Avg. Rank↓ | 1.95 | 2 | 2 |
| Avg. AUC↑ | 0.761 | 0.756 | 0.759 |
Table 11c. (c) Classification: LCC G/S
| Dataset | HYDRA-8 | HYDRA-16 | HYDRA-32 |
| MIR | 0.056 (±0.057) | 0.040 (±0.042) | 0.013 (±0.004) |
| DOGE2.0 | 0.068 (±0.023) | 0.060 (±0.027) | 0.092 (±0.008) |
| MUTE | 0.073 (±0.036) | 0.036 (±0.025) | 0.025 (±0.005) |
| EVERMOON | 0.070 (±0.053) | 0.038 (±0.045) | 0.012 (±0.005) |
| DERC | 0.071 (±0.049) | 0.038 (±0.040) | 0.015 (±0.004) |
| ADX | 0.069 (±0.047) | 0.029 (±0.035) | 0.019 (±0.005) |
| HOICHI | 0.064 (±0.038) | 0.048 (±0.030) | 0.029 (±0.008) |
| SDEX | 0.074 (±0.021) | 0.072 (±0.011) | 0.063 (±0.006) |
| BAG | 0.072 (±0.023) | 0.056 (±0.009) | 0.052 (±0.002) |
| XCN | 0.067 (±0.056) | 0.036 (±0.044) | 0.009 (±0.002) |
| ETH2x-FLI | 0.053 (±0.034) | 0.032 (±0.023) | 0.030 (±0.004) |
| stkAAVE | 0.085 (±0.037) | 0.057 (±0.024) | 0.078 (±0.017) |
| GLM | 0.097 (±0.013) | 0.103 (±0.008) | 0.094 (±0.007) |
| QOM | 0.073 (±0.054) | 0.034 (±0.040) | 0.012 (±0.001) |
| WOJAK | 0.076 (±0.059) | 0.034 (±0.042) | 0.009 (±0.000) |
| DINO | 0.065 (±0.053) | 0.046 (±0.042) | 0.018 (±0.007) |
| Metis | 0.072 (±0.030) | 0.037 (±0.016) | 0.034 (±0.010) |
| REPv2 | 0.083 (±0.024) | 0.074 (±0.015) | 0.066 (±0.004) |
| TRAC | 0.074 (±0.042) | 0.027 (±0.015) | 0.026 (±0.004) |
| BEPRO | 0.081 (±0.067) | 0.034 (±0.046) | 0.004 (±0.001) |
| 1st-Place Count↑ | 0 | 2 | 18 |
| Avg. Rank↓ | 2.90 | 1.95 | 1.15 |
| Avg. MAE↓ | 0.072 | 0.047 | 0.035 |
Table 11d. (d) Regression: New Node Count
| Dataset | HYDRA-8 | HYDRA-16 | HYDRA-32 |
| MIR | 0.080 (±0.059) | 0.059 (±0.004) | 0.039 (±0.005) |
| DOGE2.0 | 0.065 (±0.034) | 0.094 (±0.003) | 0.126 (±0.032) |
| MUTE | 0.103 (±0.070) | 0.078 (±0.034) | 0.045 (±0.027) |
| EVERMOON | 0.107 (±0.067) | 0.072 (±0.018) | 0.038 (±0.024) |
| DERC | 0.063 (±0.065) | 0.056 (±0.012) | 0.033 (±0.017) |
| ADX | 0.084 (±0.070) | 0.054 (±0.009) | 0.030 (±0.001) |
| HOICHI | 0.115 (±0.066) | 0.088 (±0.045) | 0.055 (±0.028) |
| SDEX | 0.087 (±0.027) | 0.083 (±0.068) | 0.042 (±0.019) |
| BAG | 0.141 (±0.064) | 0.109 (±0.045) | 0.075 (±0.027) |
| XCN | 0.071 (±0.033) | 0.096 (±0.008) | 0.126 (±0.033) |
| ETH2x-FLI | 0.066 (±0.059) | 0.052 (±0.007) | 0.033 (±0.020) |
| stkAAVE | 0.067 (±0.049) | 0.063 (±0.008) | 0.042 (±0.013) |
| GLM | 0.108 (±0.045) | 0.088 (±0.014) | 0.075 (±0.017) |
| QOM | 0.066 (±0.055) | 0.055 (±0.004) | 0.039 (±0.020) |
| WOJAK | 0.116 (±0.068) | 0.074 (±0.006) | 0.036 (±0.027) |
| DINO | 0.085 (±0.066) | 0.059 (±0.009) | 0.034 (±0.002) |
| Metis | 0.082 (±0.056) | 0.072 (±0.022) | 0.046 (±0.008) |
| REPv2 | 0.127 (±0.015) | 0.130 (±0.004) | 0.129 (±0.022) |
| TRAC | 0.069 (±0.053) | 0.059 (±0.003) | 0.043 (±0.023) |
| BEPRO | 0.105 (±0.068) | 0.064 (±0.005) | 0.033 (±0.018) |
| 1st-Place Count↑ | 3 | 0 | 17 |
| Avg. Rank↓ | 2.70 | 2.05 | 1.25 |
| Avg. MAE↓ | 0.090 | 0.075 | 0.056 |
Table 11e. (e) Regression: Influential Node Count
| Dataset | HYDRA-8 | HYDRA-16 | HYDRA-32 |
| MIR | 0.087 (±0.057) | 0.035 (±0.027) | 0.016 (±0.005) |
| DOGE2.0 | 0.115 (±0.023) | 0.135 (±0.032) | 0.187 (±0.008) |
| MUTE | 0.092 (±0.036) | 0.048 (±0.017) | 0.021 (±0.022) |
| EVERMOON | 0.085 (±0.053) | 0.041 (±0.029) | 0.017 (±0.000) |
| DERC | 0.082 (±0.049) | 0.033 (±0.024) | 0.021 (±0.003) |
| ADX | 0.082 (±0.047) | 0.029 (±0.018) | 0.025 (±0.008) |
| HOICHI | 0.090 (±0.038) | 0.062 (±0.006) | 0.027 (±0.020) |
| SDEX | 0.080 (±0.021) | 0.077 (±0.018) | 0.095 (±0.043) |
| BAG | 0.099 (±0.023) | 0.065 (±0.005) | 0.104 (±0.020) |
| XCN | 0.066 (±0.056) | 0.027 (±0.009) | 0.062 (±0.015) |
| ETH2x-FLI | 0.084 (±0.034) | 0.035 (±0.019) | 0.023 (±0.001) |
| stkAAVE | 0.065 (±0.037) | 0.025 (±0.004) | 0.051 (±0.022) |
| GLM | 0.106 (±0.013) | 0.101 (±0.007) | 0.087 (±0.008) |
| QOM | 0.074 (±0.054) | 0.032 (±0.019) | 0.036 (±0.013) |
| WOJAK | 0.090 (±0.059) | 0.034 (±0.035) | 0.010 (±0.004) |
| DINO | 0.085 (±0.053) | 0.040 (±0.028) | 0.017 (±0.007) |
| Metis | 0.087 (±0.030) | 0.041 (±0.005) | 0.034 (±0.010) |
| REPv2 | 0.127 (±0.024) | 0.105 (±0.003) | 0.111 (±0.008) |
| TRAC | 0.086 (±0.042) | 0.029 (±0.020) | 0.021 (±0.001) |
| BEPRO | 0.094 (±0.067) | 0.037 (±0.036) | 0.011 (±0.008) |
| 1st-Place Count↑ | 1 | 5 | 14 |
| Avg. Rank↓ | 2.85 | 1.75 | 1.40 |
| Avg. MAE↓ | 0.089 | 0.052 | 0.046 |
Table 11f. (f) Regression: Edge Count
References
- [1]N. C. Abay, C. G. Akcora, Y. R. Gel, M. Kantarcioglu, U. D. Islambekov, Y. Tian, and B. Thuraisingham. Chainnet: Learning on blockchain graphs with topological features. In J. Wang, K. Shim, and X. Wu, editors, 2019 IEEE International Conference on Data Mining, ICDM 2019, Beijing, China, November 8-11, 2019, pages 946–951. IEEE, 2019. doi: 10.1109/ICDM.2019.00105. URL https://doi.org/10.1109/ICDM.2019.00105.
- [2]C. G. Akcora, Y. R. Gel, and M. Kantarcioglu. Blockchain networks: Data structures of bitcoin, monero, zcash, ethereum, ripple, and iota. WIREs Data Mining Knowl. Discov., 12(1), 2022. doi: 10.1002/WIDM.1436. URL https://doi.org/10.1002/widm.1436.
- [3]R. Caruana. Multitask learning. Machine learning, 28(1):41–75, 1997.
- [4]H. Chen, H. Yin, X. Sun, T. Chen, B. Gabrys, and K. Musial. Multi-level graph convolutional networks for cross-platform anchor link prediction. In R. Gupta, Y. Liu, J. Tang, and B. A. Prakash, editors, KDD ’20: The 26th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Virtual Event, CA, USA, August 23-27, 2020, pages 1503–1511. ACM, 2020. doi: 10.1145/3394486.3403201. URL https://doi.org/10.1145/3394486.3403201.
- [5]J. Chen, X. Wang, and X. Xu. GC-LSTM: graph convolution embedded LSTM for dynamic network link prediction. Appl. Intell., 52(7):7513–7528, 2022. doi: 10.1007/S10489-021-02518-9. URL https://doi.org/10.1007/s10489-021-02518-9.
- [6]Y. Chen, T. Shu, X. Zhou, X. Zheng, A. Kawai, K. Fueda, Z. Yan, W. Liang, and K. I. Wang. Graph attention network with spatial-temporal clustering for traffic flow forecasting in intelligent transportation system. IEEE Trans. Intell. Transp. Syst., 24(8):8727–8737, 2023. doi: 10.1109/TITS.2022.3208952. URL https://doi.org/10.1109/TITS.2022.3208952.
- [7]K. Cho, B. van Merrienboer, D. Bahdanau, and Y. Bengio. On the properties of neural machine translation: Encoder-decoder approaches. In D. Wu, M. Carpuat, X. Carreras, and E. M. Vecchi, editors, Proceedings of SSST at EMNLP 2014, Eighth Workshop on Syntax, Semantics and Structure in Statistical Translation, Doha, Qatar, 25 October 2014, pages 103–111. Association for Computational Linguistics, 2014. doi: 10.3115/V1/W14-4012. URL https://aclanthology.org/W14-4012/.DOI
- [8]F. Cornell, O. Smirnov, G. Z. Gandler, and L. Cao. On the power of heuristics in temporal graphs. In A. B. Blaas, P. D’Costa, F. Feng, A. Kriegler, I. Mason, Z. Pan, T. Uelwer, J. Williams, Y. Xie, and R. Yang, editors, Proceedings on "I Can’t Believe It’s Not Better: Challenges in Applied Deep Learning" at ICLR 2025 Workshops, 28 April 2025, Singapore, volume 296 of Proceedings of Machine Learning Research, pages 37–46. PMLR, 2025. URL https://proceedings.mlr.press/v296/cornell25a.html.arxiv.org/abs/2502.04910
- [9]M. Crawshaw. Multi-task learning with deep neural networks: A survey. CoRR, abs/2009.09796, 2020. doi: 10.48550/ARXIV.2009.09796. URL https://doi.org/10.48550/arXiv.2009.09796.
- [10]K. Dong, A. R. Benson, and D. Bindel. Network density of states. In A. Teredesai, V. Kumar, Y. Li, R. Rosales, E. Terzi, and G. Karypis, editors, Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, KDD 2019, Anchorage, AK, USA, August 4-8, 2019, pages 1152–1161. ACM, 2019. doi: 10.1145/3292500.3330891. URL https://doi.org/10.1145/3292500.3330891.
- [11]Y. Duan, G. Zhang, S. Wang, X. Peng, Z. Wang, J. Mao, H. Wu, X. Jiang, and K. Wang. Cat-gnn: Enhancing credit card fraud detection via causal temporal graph neural networks. CoRR, abs/2402.14708, 2024. doi: 10.48550/ARXIV.2402.14708. URL https://doi.org/10.48550/arXiv.2402.14708.
- [12]M. M. Hosseinzadeh, M. Cannataro, P. H. Guzzi, and R. Dondi. Temporal networks in biology and medicine: a survey on models, algorithms, and tools. Netw. Model. Anal. Health Informatics Bioinform., 12(1):10, 2023. doi: 10.1007/S13721-022-00406-X. URL https://doi.org/10.1007/s13721-022-00406-x.
- [13]Q. Huang, X. Wang, S. X. Rao, Z. Han, Z. Zhang, Y. He, Q. Xu, Y. Zhao, Z. Zheng, and J. Jiang. Bencht emp: A general benchmark for evaluating temporal graph neural networks. In 40th IEEE International Conference on Data Engineering, ICDE 2024, Utrecht, The Netherlands, May 13-16, 2024, pages 4044–4057. IEEE, 2024. doi: 10.1109/ICDE60146.2024.00310. URL https://doi.org/10.1109/ICDE60146.2024.00310.
- [14]S. Huang, F. Poursafaei, J. Danovitch, M. Fey, W. Hu, E. Rossi, J. Leskovec, M. M. Bronstein, G. Rabusseau, and R. Rabbany. Temporal graph benchmark for machine learning on temporal graphs. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, editors, Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. URL https://papers.nips.cc/paper_files/paper/2023/hash/066b98e63313162f6562b359f62671288-Abstract-Datasets_and_Benchmarks.html.arxiv.org/abs/2307.01026
- [15]S. Huang, F. Poursafaei, J. Danovitch, M. Fey, W. Hu, E. Rossi, J. Leskovec, M. M. Bronstein, G. Rabusseau, and R. Rabbany. Temporal graph benchmark for machine learning on temporal graphs. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, editors, NeurIPS 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023.arxiv.org/abs/2307.01026
- [16]V. Jalili. The temporal graph of bitcoin transactions. In D. Belgrave, C. Zhang, L. N. Montoya, H. Lin, R. Pascanu, P. Koniusz, M. Ghassemi, N. Chen, I. V. M. Rúžič, and A. Loaiza-Bonilla, editors, Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2025, NeurIPS 2025, San Diego, CA, USA, December 2-7, 2025 / Mexico City, Mexico, November 30 - December 5, 2025, 2025. URL http://papers.nips.cc/paper_files/paper/2025/hash/2c894fcb96e23035f0fcfaf09e3135a2-Abstract-Datasets_and_Benchmarks_Track.html.
- [17]W. Jin, X. Liu, X. Zhao, Y. Ma, N. Shah, and J. Tang. Automated self-supervised learning for graphs. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022. URL https://openreview.net/forum?id=rFbR4Fv-D6-.arxiv.org/abs/2106.05470
- [18]M. Ju, T. Zhao, Q. Wen, W. Yu, N. Shah, Y. Ye, and C. Zhang. Multi-task self-supervised graph neural networks enable stronger task generalization. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. URL https://openreview.net/forum?id=1tHAZRqtfM.arxiv.org/abs/2210.02016
- [19]S. M. Kazemi, R. Goel, K. Jain, I. Kobyzev, A. Sethi, P. Forsyth, and P. Poupart. Representation learning for dynamic graphs: A survey. J. Mach. Learn. Res., 21:70:1–70:73, 2020. URL https://jmlr.org/papers/v21/19-447.html.arxiv.org/abs/1905.11485
- [20]D. P. Kingma and J. Ba. Adam: A method for stochastic optimization. In Y. Bengio and Y. LeCun, editors, 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. doi: 10.48550/ARXIV.1412.6980. URL https://doi.org/10.48550/arXiv.1412.6980.
- [21]T. N. Kipf and M. Welling. Semi-supervised classification with graph convolutional networks. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net, 2017. URL https://openreview.net/forum?id=SJU4ayYgl.arxiv.org/abs/1609.02907
- [22]K. Kläser, B. Banaszewski, S. Maddrell-Mander, C. McLean, L. Müller, A. Parviz, S. Huang, and A. W. Fitzgibbon. Minimol: A parameter-efficient foundation model for molecular learning. CoRR, abs/2404.14986, 2024. doi: 10.48550/ARXIV.2404.14986. URL https://doi.org/10.48550/arXiv.2404.14986.
- [23]J. Lee, I. Lee, and J. Kang. Self-attention graph pooling. In K. Chaudhuri and R. Salakhutdinov, editors, Proceedings of the 36th International Conference on Machine Learning, ICML 2019, 9-15 June 2019, Long Beach, California, USA, volume 97 of Proceedings of Machine Learning Research, pages 3734–3743. PMLR, 2019. URL http://proceedings.mlr.press/v97/lee19c.html.arxiv.org/abs/1904.08082
- [24]R. Levie, W. Huang, L. Bucci, M. M. Bronstein, and G. Kutyniok. Transferability of spectral graph convolutional neural networks. J. Mach. Learn. Res., 22:272:1–272:59, 2021. URL https://jmlr.org/papers/v22/20-213.html.
- [25]Y. Liang, H. Wen, Y. Xia, M. Jin, B. Yang, F. Salim, Q. Wen, S. Pan, and G. Cong. Foundation models for spatio-temporal data science: A tutorial and survey. In L. Antonie, J. Pei, X. Yu, F. Chierichetti, H. W. Lauw, Y. Sun, and S. Parthasarathy, editors, Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining, V.2, KDD 2025, Toronto ON, Canada, August 3-7, 2025, pages 6063–6069. ACM, 2025. doi: 10.1145/3711896.3736552. URL https://doi.org/10.1145/3711896.3736552.
- [26]Y. Luo, L. Shi, and X. Wu. Can classic gnns be strong baselines for graph-level tasks? simple architectures meet excellence. In A. Singh, M. Fazel, D. Hsu, S. Lacoste-Julien, F. Berkenkamp, T. Maharaj, K. Wagstaff, and J. Zhu, editors, Forty-second International Conference on Machine Learning, ICML 2025, Vancouver, BC, Canada, July 13-19, 2025, volume 267 of Proceedings of Machine Learning Research. PMLR / OpenReview.net, 2025. URL https://proceedings.mlr.press/v267/luo25h.html.arxiv.org/abs/2502.09263
- [27]B. Mohar, Y. Alavi, G. Chartrand, and O. Oellermann. The laplacian spectrum of graphs. Graph theory, combinatorics, and applications, 2(871-898):12, 1991.
- [28]R. Nassif, S. Vlaski, C. Richard, J. Chen, and A. H. Sayed. Multitask learning over graphs: An approach for distributed, streaming machine learning. IEEE Signal Process. Mag., 37(3):14–25, 2020. doi: 10.1109/MSP.2020.2966273. URL https://doi.org/10.1109/MSP.2020.2966273.
- [29]S. Ockerman, A. Gueroudji, T. Mallick, Y. He, L. Pouchard, R. B. Ross, and S. Venkataraman. PGT-I: scaling spatiotemporal gnns with memory-efficient distributed training. In Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis, SC 2025, St. Louis, MO, USA, November 16-21, 2025, pages 217–236. ACM, 2025. doi: 10.1145/3712285.3759880. URL https://doi.org/10.1145/3712285.3759880.
- [30]J. Pan, M. Nayyeri, O. Mohammed, D. Hernández, R. Zhang, C. Cheng, and S. Staab. Towards foundation model on temporal knowledge graph reasoning. CoRR, abs/2506.06367, 2025. doi: 10.48550/ARXIV.2506.06367. URL https://doi.org/10.48550/arXiv.2506.06367.
- [31]A. Pareja, G. Domeniconi, J. Chen, T. Ma, T. Suzumura, H. Kanezashi, T. Kaler, T. B. Schardl, and C. E. Leiserson. Evolvegcn: Evolving graph convolutional networks for dynamic graphs. In The Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI 2020, The Thirty-Second Innovative Applications of Artificial Intelligence Conference, IAAI 2020, The Tenth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2020, New York, NY, USA, February 7-12, 2020, pages 5363–5370. AAAI Press, 2020. doi: 10.1609/AAAI.V34I04.5984. URL https://doi.org/10.1609/aaai.v34i04.5984.
- [32]Y. Qi, J. Wu, H. Xu, and M. Guizani. Blockchain data mining with graph learning: A survey. IEEE Trans. Pattern Anal. Mach. Intell., 46(2):729–748, 2024. doi: 10.1109/TPAMI.2023.3327404. URL https://doi.org/10.1109/TPAMI.2023.3327404.
- [33]R. Ranjan, V. Hudovernik, M. ZNidar, C. I. Kanatsoulis, R. R. Upendra, M. Mohammadi, J. Meyer, T. Palczewski, C. Guestrin, and J. Leskovec. Relational transformer: Toward zero-shot foundation models for relational data. In C. Vondrick, B. Hariharan, C. Raffel, L. Pinto, D. Yang, and A. Faust, editors, The Fourteenth International Conference on Learning Representations, ICLR 2026, Rio de Janeiro, Brazil, April 23-27, 2026. proceedings.iclr.cc, 2026. URL https://proceedings.iclr.cc/paper_files/paper/2026/hash/7107d4d2e837bde2171c6b71b5bde954-Abstract-Conference.html.arxiv.org/abs/2510.06377
- [34]E. Rossi, B. Chamberlain, F. Frasca, D. Eynard, F. Monti, and M. M. Bronstein. Temporal graph networks for deep learning on dynamic graphs. CoRR, abs/2006.10637, 2020. doi: 10.48550/ARXIV.2006.10637. URL https://doi.org/10.48550/arXiv.2006.10637.
- [35]S. Ruder. An overview of multi-task learning in deep neural networks. arXiv preprint arXiv:1706.05098, 2017.arxiv.org/abs/1706.05098
- [36]S. Ruder. An overview of multi-task learning in deep neural networks. CoRR, abs/1706.05098, 2017. doi: 10.48550/ARXIV.1706.05098. URL https://doi.org/10.48550/arXiv.1706.05098.
- [37]L. Ruiz, L. F. O. Chamon, and A. Ribeiro. Transferability properties of graph neural networks. IEEE Trans. Signal Process., 71:3474–3489, 2023. doi: 10.1109/TSP.2023.3297848. URL https://doi.org/10.1109/TSP.2023.3297848.
- [38]Z. A. Sahili and M. Awad. Spatio-temporal graph neural networks: A survey. CoRR, abs/2301.10569, 2023. doi: 10.48550/ARXIV.2301.10569. URL https://doi.org/10.48550/arXiv.2301.10569.
- [39]O. Sener and V. Koltun. Multi-task learning as multi-objective optimization. In S. Bengio, H. M. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, editors, Advances in Neural Information Processing Systems 31: Annual Conference on Neural Information Processing Systems 2018, NeurIPS 2018, December 3-8, 2018, Montréal, Canada, pages 525–536, 2018. URL https://proceedings.neurips.cc/paper/2018/hash/432aca31e3e43593f35a30c8f65edce-Abstract.html.arxiv.org/abs/1810.04650
- [40]A. Senthilkumar, M. Gupte, and S. Shridevi. Dynamic spatial-temporal graph model for disease prediction. International Journal of Advanced Computer Science and Applications, 13(6), 2022.DOI
- [41]K. Shamsi, F. Poursafaei, S. Huang, T. G. B. Ngo, B. Coskunuzer, and C. G. Akcora. Graph-pulse: Topological representations for temporal graph property prediction. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. URL https://openreview.net/forum?id=DZqiC2sPTY.
- [42]K. Shamsi, T. G. B. Ngo, R. Shirzadkhani, S. Huang, F. Poursafaei, P. Azad, R. Rabbany, B. Coskunuzer, G. Rabusseau, and C. G. Akkora. Mint: Multi-network transfer benchmark for temporal graph learning. In D. Belgrave, C. Zhang, L. N. Montoya, H. Lin, R. Pascanu, P. Koniusz, M. Ghassemi, N. Chen, I. V. M. Ruíz, and A. Loaiza-Bonilla, editors, Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2025, NeurIPS 2025, San Diego, CA, USA, December 2-7, 2025 / Mexico City, Mexico, November 30 - December 5, 2025, 2025. URL http://papers.nips.cc/paper_files/paper/2025/hash/c5548168cb7324f714365a971dfe76d1-Abstract-Datasets_and_Benchmarks_Track.html.DOI
- [43]W. Shi, Z. Wen, Q. Zhang, C. Zhang, and Z. Zhang. Temporal graph thumbnail: Robust representation learning with global evolutionary skeleton. In C. Vondrick, B. Hariharan, C. Raffel, L. Pinto, D. Yang, and A. Faust, editors, The Fourteenth International Conference on Learning Representations, ICLR 2026, Rio de Janeiro, Brazil, April 23-27, 2026. proceedings.iclr.cc, 2026. URL https://proceedings.iclr.cc/paper_files/paper/2026/hash/df29d63af05cb91d705cf06ba5945b9d-Abstract-Conference.html.
- [44]M. Sypetkowski, F. Wenkel, F. Poursafaei, N. Dickson, K. Suri, P. Fradkin, and D. Beaini. On the scalability of gns for molecular graphs. In A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. M. Tomczak, and C. Zhang, editors, Advances in Neural Information Processing Systems 37: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, December 10 - 15, 2024, 2024. URL http://papers.nips.cc/paper_files/paper/2024/hash/2345275663a15ee92a06bc957be54a2c-Abstract-Conference.html.arxiv.org/abs/2404.11568
- [45]H. Tang, S. Wu, X. Sun, J. Zeng, G. Xu, and Q. Li. TCGC: temporal collaboration-aware graph co-evolution learning for dynamic recommendation. ACM Trans. Inf. Syst., 43(1):5:1–5:27, 2025. doi: 10.1145/3687470. URL https://doi.org/10.1145/3687470.
- [46]M. J. Uddin, S. Changani, and B. Coskunuzer. T3former: Temporal graph classification with topological machine learning. In S. Koenig, C. Jenkins, and M. E. Taylor, editors, Fourtieth AAAI Conference on Artificial Intelligence, Thirty-Eighth Conference on Innovative Application of Artificial Intelligence, Sixteenth Symposium on Educational Advances in Artificial Intelligence, AAAI 2026, Singapore, January 20-27, 2026, pages 26028–26036. AAAI Press, 2026. doi: 10.1609/AAAI.V40i31.39804. URL https://doi.org/10.1609/aaai.v40i31.39804.
- [47]S. Wang, J. Cao, and P. S. Yu. Deep learning for spatio-temporal data mining: A survey. IEEE Trans. Knowl. Data Eng., 34(8):3681–3700, 2022. doi: 10.1109/TKDE.2020.3025580. URL https://doi.org/10.1109/TKDE.2020.3025580.
- [48]Y. Wang, W. Fan, S. Wang, and Y. Ma. Towards graph foundation models: A transferability perspective. CoRR, abs/2503.09363, 2025. doi: 10.48550/ARXIV.2503.09363. URL https://doi.org/10.48550/arXiv.2503.09363.
- [49]Z. Wang, X. Wang, and H. Wang. Temporal graph neural networks for money laundering detection in cross-border transactions. Academia Nexus Journal, 3(2), 2024.
- [50]A. Weiße, G. Wellein, A. Alvermann, and H. Fehske. The kernel polynomial method. Reviews of modern physics, 78(1):275–306, 2006.
- [51]O. Wieder, S. Kohlbacher, M. Kuenemann, A. Garon, P. Ducrot, T. Seidel, and T. Langer. A compact review of molecular property prediction with graph neural networks. Drug Discovery Today: Technologies, 37:1–12, 2020.DOI
- [52]M. Yang, M. Zhou, M. Kalander, Z. Huang, and I. King. Discrete-time temporal network embedding via implicit hierarchical learning in hyperbolic space. In F. Zhu, B. C. Ooi, and C. Miao, editors, KDD ’21: The 27th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Virtual Event, Singapore, August 14-18, 2021, pages 1975–1985. ACM, 2021. doi: 10.1145/3447548.3467422. URL https://doi.org/10.1145/3447548.3467422.
- [53]M. Yang, W. Feng, Y. Shen, and B. Hooi. Towards better graph representation learning with parameterized decomposition & filtering. In A. Krause, E. Brunskill, K. Cho, B. Engelhardt, S. Sabato, and J. Scarlett, editors, International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA, volume 202 of Proceedings of Machine Learning Research, pages 39234–39251. PMLR, 2023. URL https://proceedings.mlr.press/v202/yang23c.html.
- [54]C. Ying, T. Cai, S. Luo, S. Zheng, G. Ke, D. He, Y. Shen, and T. Liu. Do transformers really perform badly for graph representation? In M. Ranzato, A. Beygelzimer, Y. N. Dauphin, P. Liang, and J. W. Vaughan, editors, Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual, pages 28877–28888, 2021. URL https://proceedings.neurips.cc/paper/2021/hash/f1c1592588411002af340cbaedd6fc33-Abstract.html.
- [55]J. You, T. Du, and J. Leskovec. ROLAND: graph learning framework for dynamic graphs. In A. Zhang and H. Rangwala, editors, KDD ’22: The 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Washington, DC, USA, August 14 - 18, 2022, pages 2358–2366. ACM, 2022. doi: 10.1145/3534678.3539300. URL https://doi.org/10.1145/3534678.3539300.
- [56]T. Yu, S. Kumar, A. Gupta, S. Levine, K. Hausman, and C. Finn. Gradient surgery for multi-task learning. In H. Larochelle, M. Ranzato, R. Hadsell, M. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. URL https://proceedings.neurips.cc/paper/2020/hash/3fe78a8acf5f1da99de95303940a2420c-Abstract.html.arxiv.org/abs/2001.06782
- [57]Y. Zhang and Q. Yang. A survey on multi-task learning. IEEE Trans. Knowl. Data Eng., 34(12): 5586–5609, 2022. doi: 10.1109/TKDE.2021.3070203. URL https://doi.org/10.1109/TKDE.2021.3070203.
- [58]L. Zhao, Y. Song, C. Zhang, Y. Liu, P. Wang, T. Lin, M. Deng, and H. Li. T-GCN: A temporal graph convolutional network for traffic prediction. IEEE Trans. Intell. Transp. Syst., 21(9): 3848–3858, 2020. doi: 10.1109/TITS.2019.2935152. URL https://doi.org/10.1109/TITS.2019.2935152.
- [59]L. Zhao, Y. Song, C. Zhang, Y. Liu, P. Wang, T. Lin, M. Deng, and H. Li. T-GCN: A temporal graph convolutional network for traffic prediction. IEEE Trans. Intell. Transp. Syst., 21(9): 3848–3858, 2020. doi: 10.1109/TITS.2019.2935152. URL https://doi.org/10.1109/TITS.2019.2935152.
- [60]J. Zhu, A. Khan, and C. G. Akcora. Data depth and core-based trend detection on blockchain transaction networks. Frontiers Blockchain, 7, 2024. doi: 10.3389/FBLOC.2024.1342956. URL https://doi.org/10.3389/fbloc.2024.1342956.
- [61]Y. Zhu, F. Cong, D. Zhang, W. Gong, Q. Lin, W. Feng, Y. Dong, and J. Tang. Wingnn: Dynamic graph neural networks with random gradient aggregation window. In A. K. Singh, Y. Sun, L. Akoglu, D. Gunopulos, X. Yan, R. Kumar, F. Ozcan, and J. Ye, editors, Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD 2023, Long Beach, CA, USA, August 6-10, 2023, pages 3650–3662. ACM, 2023. doi: 10.1145/3580305.3599551. URL https://doi.org/10.1145/3580305.3599551.



