Learning Long Range Spatio-Temporal Representations over Continuous Time Dynamic Graphs with State Space Models

2026-06-06 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

The paper introduces CTDG-SSM, a parameter-efficient state-space modeling framework for continuous-time dynamic graphs (CTDGs). This model addresses the challenge of capturing both long-range temporal (LRT) and long-range spatial (LRS) dependencies in evolving relational data, a limitation of existing event-driven and sequence-based models. CTDG-SSM derives from a novel continuous-time Topology-Aware higher-order polynomial projection operator (CTT-HiPPO), which jointly encodes temporal dynamics and graph structure by projecting classical HiPPO solutions through a polynomial of the Laplacian matrix. A computationally efficient discrete formulation is achieved using the zero-order hold approach. Across benchmarks including dynamic link prediction, dynamic node classification, and sequence classification, CTDG-SSM achieves state-of-the-art performance, showing large gains on datasets requiring LRT and LRS reasoning (e.g., LastFM, MOOC, Enron). It also demonstrates high parameter efficiency, using about one-tenth fewer parameters than competing methods, and robustness to structural perturbations.

Key takeaway

For AI Architects designing systems for evolving relational data, CTDG-SSM offers a robust solution for simultaneously addressing long-range temporal and spatial dependencies. You should consider integrating this parameter-efficient state-space framework, especially for applications like financial fraud detection or social network analysis where fine-grained, multi-hop interactions are critical. Its proven state-of-the-art performance and computational efficiency on benchmarks like LastFM and Enron suggest it can improve accuracy while reducing resource overhead.

Key insights

CTDG-SSM unifies state-space modeling with graph topology to capture long-range spatio-temporal dependencies in continuous-time dynamic graphs.

Principles

• Jointly encode temporal dynamics and graph structure.
• Leverage polynomial graph filters for multi-hop spatial info.
• State-space models can efficiently compress historical information.

Method

CTDG-SSM derives CTT-HiPPO by minimizing discrepancy between observed node features and graph-filtered polynomial approximations. This continuous formulation is then discretized using zero-order hold for efficient implementation.

In practice

• Apply CTDG-SSM for dynamic link prediction.
• Use for dynamic node and sequence classification.
• Consider higher-order polynomial filters for LRS.

Topics

• Continuous-Time Dynamic Graphs
• State Space Models
• Graph Neural Networks
• Long-Range Dependencies
• Dynamic Link Prediction
• Node Classification

Code references

• adhocmp122/CTDG-SSM

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, AI Architect

Related on AIssential

• Graph networks Frozen in Time Are Useless. Temporal Graph Networks Aren’t. — Temporal Graph Networks (TGNs) enable dynamic graph analysis by giving nodes memory and encoding temporal interactions.
• A Topological Characterization of Graph Neural Networks via Stochastic Block Model Embeddings on the n-Sphere — A topological framework characterizes GNNs via SBM embeddings on an n-sphere, creating low-dimensional "fingerprints" for model comparison and transfer learning.
• Bi-Level Chaotic Fusion Based Graph Convolutional Network for Stock Market Prediction Interval — Prediction intervals, not just point forecasts, are crucial for risk-aware financial market decision-making.
• M-CaStLe: Uncovering Local Causal Structures in Multivariate Space-Time Gridded Data — M-CaStLe extends causal discovery for multivariate space-time gridded data by jointly modeling within- and cross-variable structures.
• Graph Hierarchical Recurrence for Long-Range Generalization — Graph Hierarchical Recurrence (GHR) improves long-range generalization and parameter efficiency in graph learning.
• SkillDAG: Self-Evolving Typed Skill Graphs for LLM Skill Selection at Scale — Dynamic, graph-based structural retrieval significantly enhances LLM agent skill selection and performance over static methods.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.