Treffer: Dynamic graph representation learning with disentangled information bottleneck.

Dynamic graph representation learning recently garnered enormous research attention. Despite the notable successes of existing methods, they usually characterize dynamic graphs as a perceptual whole and learn dynamic graph representations within an entangled feature space, which overlook different temporal dependencies inherent in the data. Specifically, the evolution of dynamic graphs is usually decided by a dichotomy in properties: time-invariant properties and time-varying properties. Existing holistic works fail to distinguish these temporal properties and may suffer suboptimal performance in downstream tasks. To tackle this problem, we propose to learn macro-disentangled dynamic graph representations based on the Information Bottleneck theory, leading to a novel dynamic graph representation learning method, Disentangled Dynamic Graph Information Bottleneck (DDGIB). Our DDGIB explicitly embeds the dynamic graphs into a time-invariant representation space and a time-varying representation space. The time-invariant representation space encapsulates stable properties across the temporal span of dynamic graphs, whereas the time-varying representation space encapsulates time-fluctuating properties. The macro disentanglement on the temporal dependencies facilitates the representations' performance on downstream tasks. Furthermore, we theoretically prove the sufficiency and macro disentanglement of DDGIB. The sufficiency demonstrates that DDGIB can achieve sufficient representations for any possible downstream tasks, while the macro disentanglement certifies that DDGIB can embed the different temporal properties into their corresponding temporal representation space. Extensive experimental results on various datasets and downstream tasks demonstrate the superiority of our method.
(Copyright © 2025. Published by Elsevier Ltd.)

Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.