Self-Explainable GNN Improves Multi-Label Classification and Attribution

This paper introduces the Self-Explainable Multi-Label Graph Neural Network (SEMGNN), an innovative approach that addresses the complexities of multi-label graph learning. In real-world scenarios, entities often belong to multiple categories or possess multiple attributes, and existing graph learning methods typically lack integrated interpretability. While post-hoc explainers exist, they often fail to adequately model how evidence should be shared or separated across correlated labels. SEMGNN overcomes these limitations by jointly learning a predictor and a sparse edge-mask explainer within a single, unified framework. This end-to-end design allows the model to not only classify multi-labeled nodes but also to identify the specific edges that significantly contribute to each predicted label. By explicitly incorporating label-label correlations, SEMGNN enhances both the accuracy of multi-label node classification and the fidelity of individual label explanations, ensuring that distinct yet coherent structural evidence supports each label. Experiments on synthetic and real-world datasets across various domains demonstrate SEMGNN's competitive predictive performance and its ability to provide more faithful and compact label-conditioned explanations.