Poisson kernel: Avoiding self-smoothing in graph convolutional networks

Graph convolutional network is now an effective tool to deal with non-Euclidean data, such as social behavior analysis, molecular structure and skeleton-based action recognition. kernel one of the most significant factors in graph networks extract nodes’ feature, some variants it have achieved highly satisfactory performance theoretically experimentally. However, there was limited research about how exactly different structures influence these kernels. Some existing methods used adaptive a given structure, which still not explore internal reasons. In this paper, we start from theoretical analysis spectral study properties kernels, revealing self-smoothing phenomenon its effect specific structured graphs. After that, propose Poisson that can avoid without training any kernel. Experimental results demonstrate our only works well on benchmark datasets where state-of-the-art work fine, but also evidently superior them synthetic datasets.