Interpreting and Understanding Graph Convolutional Neural Network using Gradient-based Attribution Method

To solve the problem that convolutional neural networks (CNNs) are difficult to process non-grid type relational data like graphs, Kipf et al. proposed a graph convolutional neural network (GCN). The core idea of the GCN is to perform two-fold informational fusion for each node in a given graph during each iteration: the fusion of graph structure information and the fusion of node feature dimensions. Because of the characteristic of the combinatorial generalizations, GCN has been widely used in the fields of scene semantic relationship analysis, natural language processing and few-shot learning etc. However, due to its two-fold informational fusion involves mathematical irreversible calculations, it is hard to explain the decision reason for the prediction of the each node classification. Unfortunately, most of the existing attribution analysis methods concentrate on the models like CNNs, which are utilized to process grid-like data. It is difficult to apply those analysis methods to the GCN directly. It is because compared with the independence among CNNs input data, there is correlation between the GCN input data. This resulting in the existing attribution analysis methods can only obtain the partial model contribution from the central node features to the final decision of the GCN, but ignores the other model contribution from central node features and its neighbor nodes features to that decision. To this end, we propose a gradient attribution analysis method for the GCN called Node Attribution Method (NAM), which can get the model contribution from not only the central node but also its neighbor nodes to the GCN output. We also propose the Node Importance Visualization (NIV) method to visualize the central node and its neighbor nodes based on the value of the contribution...