Deep linear graph attention model for attributed graph clustering

Abstract

Attributed graph clustering is an important task for grouping the nodes in a graph. In recent years, the algorithms based on graph convolutional networks (GCN) have achieved promising performance. However, almost all existing methods ignore that the nonlinearity between two consecutive GCN layers is unnecessary for improving the performance of attributed graph clustering, and may even harm the efficiency of the model. In this paper, we propose a novel deep linear graph attention model for attributed graph clustering (DLGAMC), which consists of an attention-based aggregation module and a similarity preserve module. Specifically, we simply exploit cosine similarity to construct the attention for aggregation, which does not need to learn extra attention parameters. It is worth noting that the attention we designed not only explicitly considers the similarity of attribute information, but also implicitly takes into account the local graph structure. To select the proper order of aggregation, we propose an adaptive strategy to evaluate the smoothness of node representations, where intra-cluster distance and inter-cluster distance are the key indicators in this process. To learn node representations for clustering, we design a similarity preserve module to preserve local similarity and global dissimilarity of the smooth features obtained by multiple aggregations, which is different from the ideas in reconstruction-based methods. Finally, k-means is performed on the learned representations to obtain the cluster partition. Experiments on several datasets show that our algorithm achieves great performance in attributed graph clustering tasks.