An Evaluation of Backpropagation Interpretability for Graph Classification with Deep Learning
Kenneth Teo Tian Shun, Eko Edita Limanta, Arijit Khan
Abstract
The end-to-end learning in convolutional neural networks (CNNs) and their ability to extract localized deep features, have turned them into powerful tools for learning from a large corpus of data including graphs. Deep neural networks such as CNNs are "black-box", therefore various interpretability methods have been developed to understand which aspects of the input data drive the decisions of the network. However, interpretability for graph convolutional neural networks (GC-NNs) is an open area of research. To this end, we investigate three backpropagation-based interpretability methods: saliency map with contrastive gradients (CG), gradient-weighted class activation mapping (Grad-CAM), and deep learning important features (DeepLIFT) in conjunction with three state-of-the-art GCNNs: GCNN+GAP, DGCNN, and DIFFPOOL, as well as with their variants, for the graph classification problem. We discuss novel challenges and our solutions towards integrating these deep learning frameworks, measuring their efficiency and performance both qualitatively and quantitatively. With our extensive empirical analysis over five real-world graph datasets from different categories, we report their quantitative, visualization, and active subgraph based performance, compare them with the classic significant subgraph mining results, summarize their trade-offs and surprising findings. We conclude by discussing our recommendations on the road ahead.