Overlapping Clusters in Cluster Convolutional Networks

A popular research topic in Graph Convolutional Net-works (GCNs) is to speedup the training time of the network. The main bottleneck in training GCN is the exponentially growing of computations. In Cluster-GCN based on this fact that each node and its neighbors are usually grouped in the same cluster, considers the clus-tering structure of the graph, and expand each node’s neighborhood within each cluster when training GCN. The main assumption of Cluster-GCN is the weak rela-tion between clusters; which is not correct at all graphs. Here we extend their approach by overlapped clustering, instead of crisp clustering which is used in Cluster-GCN. This is achieved by allowing the marginal nodes to con-tribute to training in more than one cluster. The evalua-tion of the proposed method is investigated through the experiments on several benchmark datasets. The exper-imental results show that the proposed method is more efficient than Cluster-GCN, in average.