Dynamic Graph Information Bottleneck

Information bottleneck method
DOI: 10.48550/arxiv.2402.06716 Publication Date: 2024-02-09
ABSTRACT
Dynamic Graphs widely exist in the real world, which carry complicated spatial and temporal feature patterns, challenging their representation learning. Graph Neural Networks (DGNNs) have shown impressive predictive abilities by exploiting intrinsic dynamics. However, DGNNs exhibit limited robustness, prone to adversarial attacks. This paper presents novel Information Bottleneck (DGIB) framework learn robust discriminative representations. Leveraged (IB) principle, we first propose expected optimal representations should satisfy Minimal-Sufficient-Consensual (MSC) Condition. To compress redundant as well conserve meritorious information into latent representation, DGIB iteratively directs refines structural flow passing through graph snapshots. meet MSC Condition, decompose overall IB objectives DGIB$_{MS}$ DGIB$_C$, channel aims minimal sufficient representations, with guarantees consensus. Extensive experiments on real-world synthetic dynamic datasets demonstrate superior robustness of against attacks compared state-of-the-art baselines link prediction task. best our knowledge, is work graphs grounded information-theoretic principle.
SUPPLEMENTAL MATERIAL
Coming soon ....
REFERENCES ()
CITATIONS ()
EXTERNAL LINKS
PlumX Metrics
RECOMMENDATIONS
FAIR ASSESSMENT
Coming soon ....
JUPYTER LAB
Coming soon ....