MLGIB: Multi-Label Graph Information Bottleneck for Robust GNNs

To tackle the issue of over-squashing in Graph Neural Networks (GNNs) for multi-label graphs, researchers have introduced the Multi-Label Graph Information Bottleneck (MLGIB). Over-squashing happens when information from rapidly expanding neighborhoods is condensed into fixed-dimensional formats. In multi-label graphs, adjacent nodes frequently have limited shared labels, while differing in numerous irrelevant ones, which weakens predictive signals. MLGIB approaches multi-label message passing as a constrained information transfer amidst irrelevant label noise, striking a balance between expressiveness and robustness by maintaining predictive signals and minimizing noise. It establishes a Markovian dependence space and formulates manageable variational bounds: the lower bound enhances mutual information with target labels, while the upper bound reduces irrelevant information. This research is available on arXiv (2605.13126).