Explanations for Graph Neural Networks via Layer Analysis
Qinfeng Li, Xinrui Kang, Wenyuan Li, Dong Liang · Atlantis Highlights in Intelligent Systems/Atlantis highlights in intelligent systems · 2023
Like many deep learning models, graph neural networks (GNNs) are regarded as black boxes and lack interpretability.Therefore, it is difficult for GNNs to be fully trusted by humans to be applied to various life scenarios.Based on this problem, we propose a new interpretability method called LAExplainer, which is used to explain GNNs hierarchically at the model level.In particular, LAExplainer not only focuses on the overall interpretation of the model, but also analyzes the interpretation problems between layers.Our approach interprets the middle-level process of the model through layer-by-layer analysis, and uses it as a basis to guide the construction of sub-graphs to reduce the size of the sub-graph set, which effectively explain the overall model.In addition, the approach will analyze the importance of model features and produce an adjustable principal component selection mechanism.In terms of evaluation indicators, we propose to set hyperparameters so that the two results of Fidelity and Sparsity can be changed simultaneously by adjusting the hyperparameters during the interpretation of GNNs.Experimental results show that our proposed method is effective in synthetic data sets and real data sets, and the results of the visualized sub-graphs are more in line with human understanding.