Return
Effective and Efficient Temporal Graph Neural Networks via Polynomial Spectral Sparsification
Z
X
L
T
DOI:10.1109/tii.2026.3686195.png)
Abstract
En 中文
Temporal graph neural networks (T-GNNs) have emerged as an effective paradigm for learning dynamic node representations by modeling the temporal evolution of graph structures. Among them, personalized PageRank (PPR)-based T-GNNs have recently gained increasing attention due to their strong empirical performance and theoretical grounding. Nevertheless, existing PPR-based T-GNN methods suffer from two fundamental limitations. First, their aggregation mechanisms are typically restricted to lower order neighborhoods, which hinders the capture of higher order temporal dependences and leads to degraded performance. Second, feature propagation and feature transformation are often tightly coupled, resulting in limited scalability on large temporal graphs. To overcome these challenges, we propose temporal personalized PageRank with spectral sparsification (TPPSS), a novel framework that seamlessly incorporates temporal information into the PPR formulation while decoupling propagation from transformation through a polynomial spectral sparsification. A key technical contribution of <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">TPPSS</i> lies in a unified matrix-based formulation that replaces the conventional rowwise computation of PPR-like scores, enabling efficient and scalable. Moreover, by leveraging spectral sparsification, <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">TPPSS</i> constructs a sparse computational graph that effectively suppresses noise and redundancy. Extensive experiments on six datasets showcase the effectiveness, efficiency, and scalability of our solutions compared to eight competitors.
Keywords:
Graph neural networks (GNNs)
personalized PageRank (PPR)
temporal graphs
Journal
IF:
9.9
Papers:
8.3K
Citations:
6.0W
