Interpreting Deep Graph Convolutional Networks with Spectrum Perspective

Graph convolutional network (GCN) architecture is the basis of many neural networks and has been widely used in processing graph-structured data. When dealing with large sparse data, deeper GCN models are often required. However, suffer from performance degradation as number layers increases. The mainstream attribution current research over-smoothing, there also gradient vanishing, training difficulties, etc., so a consensus cannot be reached. In this paper, we theoretically analyze problem by adopting spectral graph theory to globally consider propagation transformation components architecture, conclude that over-smoothing caused matrices not key factor for degradation. Afterwards, addition using conventional experimental methods, proposed an analysis strategy under guidance random matrix singular value distribution model weight matrix. We concluded leading component. context lack on degradation, paper proposes systematic strategy, well theoretical empirical evidence.