Interpreting and Understanding Graph Convolutional Neural Network using Gradient-based Attribution Methods
share
In order to solve the problem that convolutional neural networks (CNN) are difficult to process non-image type relational data, Kipf et al. proposed a graph convolutional neural network (GCN). The core idea is to perform two-fold information fusion for each node in a given graph during each iteration: the fusion of graph structure information and the fusion of node feature dimensions. Although GCN has been widely used in the fields of scene semantic relationship analysis, natural language processing, and few-shot learning because of its ability to combine generalization, owing to its two-information fusion involves mathematical irreversible calculations, it is hard for GCN to explain that the predicting reason for each node classification (i.e. attribution analysis). However, the existing attribution analysis methods cannot be directly applied to the GCN because compared with the independence among CNN input data, there is correlation between GCN input data. This leads to the existing attribution method only to obtain the partial contribution of the final decision of the GCN from target node feature, the complete contribution and the contribution from neighbor nodes features cannot be obtained. To this end, we propose a gradient attribution analysis method for GCN, NAM (Node Attribution Method), can get the contribution of the target node and its neighbor nodes to the GCN output. We also propose the NIV (Node Importance Visualization) method to visualize the target node of the GCN and its neighbor nodes based on the value of the contribution value. We use the perturbation analysis method to verify the effect of NAM based on the citation network dataset. The experimental results show that NAM can well learn the contribution of each node to the node classification prediction.
READ FULL TEXT