Reproduce “State Space Models on Temporal Graphs: A First-Principles Study” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Jintang Li, Ruofan Wu, Xinzhou Jin, Boqun Ma, Liang Chen, Zibin Zheng

This research introduces GRAPHSSM, a framework for modeling temporal graphs that extends State Space Models (SSMs) through a novel GHIPPO abstraction incorporating Laplacian regularization. The framework demonstrates state-of-the-art performance in node classification on various temporal graph benchmarks, offering enhanced scalability and efficiency, particularly for large-scale and long-sequence graphs.

Key insights

  • Integrating graph topology via Laplacian regularization into the online function approximation objective (GHIPPO) leads to a theoretically sound, piecewise LTI state-space representation for temporal graphs.
  • The challenge of unobserved graph mutations in discrete-time temporal graphs can be effectively addressed by a mixed discretization strategy that combines GNN-based inter-node mixing and feature/representation-level intra-node mixing.
  • The 'hippo' initialization for the state matrix A, coupled with applying sophisticated mixing mechanisms primarily at the first SSM layer, optimizes performance for temporal graph modeling.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.