Hashing Graph Convolution for Node Classification

EasyChair Preprint 1056

Abstract

Convolution on graphs has aroused great interest in AI due to its potential applications to non-gridded data. To bypass the ordering influence of adjacent nodes, the summing/average diffusion/aggregation is often imposed on local receptive fields in most prior works. However, the collapsing into one point tends to cause signal entanglements of neighbor nodes, which would decrease the discriminability of nodes. To address this problem, in this paper, we propose a simple but effective Hashing Graph Convolution (HGC) method by using global-hashing and local-projection on node aggregation for the task of node classification. In contrast to the conventional aggregation with a full collision, the hash-projection can greatly reduce the collision probability during gathering neighbor nodes and further better preserve original discrepancies of local regions. Another incidental effect of hash-projection is that the receptive field of each node is normalized into a common-size bucket space, which not only staves off the trouble of different-size neighbors and their orders but also make a graph convolution run like the standard shape-girded convolution. Considering the small training samples, further, we introduce a prediction-consistent regularization term into HGC to constrain the score consistency of unlabeled nodes in the graph. The extensive experiments on several node classification datasets demonstrate that hash-projection can indeed promote the performance and our HGC can achieve new state-of-the-art results.

Keyphrases: Discriminative Information, graph convolution, hash transform, node classification

@booklet{EasyChair:1056,
  author    = {Wenting Zhao and Zhen Cui and Chunyan Xu and Chengzheng Li and Tong Zhang and Jian Yang},
  title     = {Hashing Graph Convolution for Node Classification},
  howpublished = {EasyChair Preprint 1056},
  year      = {EasyChair, 2019}}