Pruning for Generalization: A Transfer-Oriented Spatiotemporal Graph Framework

Multivariate time series forecasting in graph-structured domains is critical for real-world applications, yet existing spatiotemporal models often suffer from performance degradation under data scarcity and cross-domain shifts. We address these challenges through the lens of structure-aware context selection. We propose TL-GPSTGN, a transfer-oriented spatiotemporal framework that enhances sample efficiency and out-of-distribution generalization by selectively pruning non-optimized graph context. Specifically, our method employs information-theoretic and correlation-based criteria to extract structurally informative subgraphs and features, resulting in a compact, semantically grounded representation. This optimized context is subsequently integrated into a spatiotemporal convolutional architecture to capture complex multivariate dynamics. Evaluations on large-scale traffic benchmarks demonstrate that TL-GPSTGN consistently outperforms baselines in low-data transfer scenarios. Our findings suggest that explicit context pruning serves as a powerful inductive bias for improving the robustness of graph-based forecasting models.

💡 Research Summary

The paper tackles a fundamental problem in graph‑based spatiotemporal forecasting: models such as Spatiotemporal Graph Convolutional Networks (STGCNs) perform well when abundant historical data are available, but their accuracy collapses under data scarcity or when the source and target domains have different graph topologies (e.g., different cities with distinct road networks). Existing transfer‑learning approaches focus on parameter sharing, domain adaptation, or knowledge distillation, yet they largely ignore the quality of the input graph itself. The authors argue that noisy or irrelevant graph context—especially boundary sensors that are heavily influenced by external regions not represented in the modeled graph—injects noise into message passing, inflates representation variance, and hampers cross‑domain generalization.

To address this, they propose TL‑GPSTGN, a transfer‑oriented spatiotemporal framework that explicitly prunes the graph before any learning takes place. The core component is the Graph Pruning Processor (GPP), which consists of three stages: (1) Information Entropy Analyzer (IEA), (2) dual‑criteria edge scoring, and (3) outer‑layer node removal.

• Entropy analysis discretizes each node’s traffic time series into B bins, builds an empirical distribution, and computes Shannon entropy Hᵢ. Higher entropy indicates richer, less trivial dynamics.
• Correlation scoring computes the absolute Pearson correlation rᵢⱼ between every pair of connected nodes.
• Edge importance combines these two signals: sᵢⱼ = Aᵢⱼ·rᵢⱼ·(Hᵢ+Hⱼ)/2. This score favors edges that link high‑entropy nodes with strong temporal alignment, while suppressing weak or noisy connections.

Edges with scores below a threshold τ (or a top‑k per node) are dropped. The remaining adjacency matrix is used to compute a pruned degree for each node; nodes whose degree falls below a minimum d_min are deemed “outer‑layer” or boundary nodes and are removed together with their incident edges. This pruning can be iterated L times to peel multiple outer rings, yielding a compact subgraph that retains the most informative intra‑region connectivity while discarding peripheral noise.

After pruning, each node’s time series is normalized (e.g., z‑score) before feeding into a standard STGCN backbone. The STGCN follows the familiar “TCL–GCL–TCL” pattern (Temporal Convolution → Graph Convolution → Temporal Convolution) and finally passes through a prediction head and a “Reductor” that inverts the normalization to output traffic values in the original scale.

Training proceeds in two stages: (i) Source pre‑training on a data‑rich city (e.g., METR‑LA) where the full pipeline—including GPP—is applied to learn transferable spatiotemporal filters; (ii) Target fine‑tuning where the same GPP is applied to the target city, and the pretrained parameters are fine‑tuned using only a small labeled subset (5‑20 % of the target data). Because the same pruning logic is used in both domains, the source and target graphs become more comparable, reducing domain shift at the input level.

Experiments use three large traffic datasets—METR‑LA, PEMS‑BAY, and PEMSD7—with 15‑minute and 30‑minute forecasting horizons. Baselines include naive methods (HA, ARIMA), classic neural models (FNN, FC‑LSTM), and the original STGCN. Evaluation metrics are MAE, RMSE, and MAPE.

Key findings:

• In single‑domain (in‑domain) forecasting, TL‑GPSTGN matches or slightly trails STGCN on some datasets, confirming that pruning does not hurt core predictive power while reducing graph complexity.
• In transfer scenarios, TL‑GPSTGN consistently outperforms STGCN across all target datasets and horizons, especially when the target data budget is low (5‑10 %). The performance gap widens as the source‑to‑target data ratio decreases, demonstrating superior sample efficiency.
• Qualitative plots show that TL‑GPSTGN tracks sudden traffic spikes and drops more faithfully than STGCN, indicating that the pruned subgraph preserves the most informative dynamics while filtering out noisy boundary influences.

Limitations and future work: The pruning thresholds (τ, d_min) are hyper‑parameters that may need domain‑specific tuning; the entropy calculation requires discretization, potentially discarding fine‑grained information; and the current pruning pipeline is non‑differentiable, preventing end‑to‑end learning. The authors suggest meta‑learning for automatic hyper‑parameter selection and differentiable graph sparsification techniques as promising extensions.

In summary, TL‑GPSTGN introduces a novel “structure‑aware context selection” paradigm for transfer learning in graph‑based spatiotemporal forecasting. By pruning away non‑optimized graph context before model training, it provides a powerful inductive bias that improves robustness, generalization, and data efficiency when moving across heterogeneous graph domains.