DyGSSM: Multi-view Dynamic Graph Embeddings with SSM Gradient Update

Bizhan Alipourpijani, Serdar Bozdag

Keywords: Dynamic graph representation, State Space Model, Graph Convolutional Networks, Random Walk

Abstract:

Dynamic graphs whose topology and nodes evolve over time are ubiquitous in multiple real world domains such as social networks, finance, and healthcare. Traditional graph learning methods fail to capture structural changes and temporal patterns in dynamic graphs. Recent advances in dynamic graph representation learning, such as meta-learning-based approaches, have addressed some of these
challenges. However, existing methods still face three key limitations. First, most approaches capture either local or global structures of the graphs, neglecting to model both simultaneously. Second, meta-learning models often depend on user-specific window size, which must be carefully tuned for each dataset. A short window size may miss trends, and a long window size may blur recent updates.
Third, most methods work on only discrete-time or continuous-time dynamic graphs, resulting in suboptimal performance across different temporal settings. To address these limitations in dynamic graph representation learning, we propose
a novel method called DyGSSM (Multi-view Dynamic Graph Embeddings with SSM Gradient Update). We extract local and global feature at each snapshot and fuse them using a lightweight attention mechanism for link prediction. To capture
long-term dependencies when updating model parameters, we incorporate HiPPO (High-order Polynomial Projection Operators) algorithm, which has gained attention for its ability to efficiently optimize and preserve sequence history in State
Space Models (SSMs). DyGSSM is designed to handle both discrete-time and continuous-time dynamic graphs. Parameter comparisons show that DyGSSM requires substantially fewer parameters than the other methods. Extensive experiments on 12 public datasets demonstrate that DyGSSM outperforms baselines and state-of-the-art methods in 32 out of 36 evaluation metrics. The source code
and datasets are available at https://anonymous.4open.science/r/DyGSSM

Submission Type: Full paper proceedings track submission (max 9 main pages).

Submission Number: 118