ScoreCAM GNN : a generalization of an optimal local post-hoc explaining method to any geometric deep learning models
Adrien Raison, Bourdon, Pascal, Helbert, David · HAL (Le Centre pour la Communication Scientifique Directe) · 2022
Graph neural networks have shown impressive results in many daily applications. As many others deep approaches, they inherently lack of interpretability regarding the decisonal procedure designed while being optimized. Under the post-hoc local model-based paradigm, explaining methods suited for graph often try to find relevant subgraphs but the inherent geometric strucuture of graph turn this task in an untractable combinatorial optimization problem. Many relevant methods have been designed but often lack to provide meaningful results. Under the geometric deep learning framework, convolutional neural networks are particular case of graph neural networks. In this study we extend and generalize an explaining methods suited for convolutional neural network to graph neural network thank to theorical ground of the geometric deep learning framework. To show the relevance of our generalization, we lead a theoritical study regarding geometric priors those two models share and the computational cost impact it induces for explaining such models. We also lead a qualitative study on real-world dataset often used in the literature and compare our method to state-of-the-art methods. Finally, we lead a quantitative study between our method and benchmarked methods with respect to objectives metrics widely-used in the literature. We show that our method achieve stronger results regarding those three settings.