Self-Explainable Multi-Label Graph Neural Network for Correlated Evidence Attribution
SEMGNN is an end-to-end self-explainable multi-label graph neural network that jointly learns node classification and edge-level explanations within a unified training framework Unlike post-hoc explainers, SEMGNN explicitly models label-dependent evidence sharing, handling both weakly and negatively associated label pairs The model uses a sparse edge-mask explainer that identifies edges significantly contributing to each predicted label per target node Label-label correlations are leveraged to s
Analysis
TL;DR
- SEMGNN is an end-to-end self-explainable multi-label graph neural network that jointly learns node classification and edge-level explanations within a unified training framework
- Unlike post-hoc explainers, SEMGNN explicitly models label-dependent evidence sharing, handling both weakly and negatively associated label pairs
- The model uses a sparse edge-mask explainer that identifies edges significantly contributing to each predicted label per target node
- Label-label correlations are leveraged to simultaneously improve multi-label classification accuracy and enhance the faithfulness of individual label explanations
- Experiments on synthetic and real-world datasets (social networking, entertainment, life sciences) demonstrate competitive or improved predictive performance with more faithful and compact explanations
Why It Matters
This work addresses a critical gap in multi-label graph learning by integrating interpretability directly into the training process rather than relying on post-hoc explanation methods that fail to capture label-dependent evidence structures. For AI practitioners building graph-based systems in domains like bioinformatics or social network analysis, SEMGNN offers a practical path toward models that are both accurate and transparent about how different labels relate to specific graph structures.
Technical Details
- Architecture: SEMGNN combines a multi-label node classifier with a sparse edge-mask explainer in a single unified framework, trained end-to-end with a joint objective function
- Core innovation: Explicit modeling of label-label correlations to guide both classification and explanation, enabling distinct yet coherent evidence attribution across labels
- Explanation mechanism: The sparse edge-mask explainer identifies and weights edges contributing to each predicted label, producing compact label-conditioned explanations rather than generic graph-level attributions
- Benchmarking: Evaluated on synthetic datasets and real-world multi-label networks spanning social networking, entertainment, and life sciences domains, with comparisons against existing post-hoc graph explainers
- Key advantage over post-hoc methods: Post-hoc explainers cannot explicitly model how evidence should be shared or separated across labels, particularly for weakly or negatively correlated label pairs—SEMGNN directly addresses this limitation through its joint training objective
Industry Insight
- Organizations deploying graph neural networks in regulated or high-stakes domains (e.g., healthcare, finance) should prioritize self-explainable architectures over post-hoc approaches, as the latter risk producing misleading or incomplete attributions when label correlations are complex
- The explicit handling of negative label associations in SEMGNN is particularly relevant for domains like drug discovery and recommendation systems, where the absence of certain relationships is as informative as their presence
- Future multi-label graph learning research should treat interpretability as a first-class design constraint rather than an afterthought, as unified training frameworks consistently outperform sequential explain-then-predict pipelines in both faithfulness and computational efficiency
Disclaimer: The above content is generated by AI and is for reference only.