Enhancing Concept Completeness for Graph Neural Networks via Side-Channel
Yicong Li, Kuanjiu Zhou, Mingyu Fan, Shaozhao Zhai, Muhammad Usman Arshad · 2024
In the realm of interpreting Graph Neural Networks (GNNs), concept-based interpretability stands out for its ability to recognize diverse subgraph structures as semantic concepts, elucidating GNN predictions. While existing methods in this domain have shown promise in enhancing human understanding, they often fall short in ensuring the completeness of the concept set. These interpreters may lack essential concepts, impacting interpretation accuracy. The root of this issue lies in the absence of the Markov assumption between the set of concepts and predictions, resulting in non-zero mutual information between input features outside the concept set and GNN predictions. To address this, a novel concept self-expansion method, GENII, is introduced. GENII employs a side-channel to learn potential concepts and known concepts crucial for accurate predictions, bypassing the Markov assumptions and enhancing conceptual completeness.