Graph Convolutional Network Using a Reliability-Based Feature Aggregation Mechanism

TL;DR: This work presents a Graph Convolutional Network using a Reliability-based Feature Aggregation Mechanism called GraphRFA, where the neighbors for each node are sample according to different kinds of link reliability and further aggregate feature information from different reliability-specific neighborhoods by a dual feature aggregation scheme.

Abstract: Graph convolutional networks (GCNs) have been proven extremely effective in a variety of prediction tasks. The general idea is to update the embedding of a node by recursively aggregating features from the node’s neighborhood. To improve the training efficiency, modern GCNs usually sample a fixed-size set of neighbors uniformly or sample according to nodes’ importance, instead of using the full neighborhood. However, both the sampling strategies ignore the reliability of a link between the target node and its neighbor, which can be implied by the graph structure and may seriously impact the performance of GCNs. To deal with this problem, we present a Graph Convolutional Network using a Reliability-based Feature Aggregation Mechanism called GraphRFA, where we sample the neighbors for each node according to different kinds of link reliability and further aggregate feature information from different reliability-specific neighborhoods by a dual feature aggregation scheme. We also theoretically prove that our aggregation scheme is permutation invariant for the graph data, and provide two simple but effective instantiations satisfying such scheme. Experimental results demonstrate the effectiveness of GraphRFA on different datasets.