A Re-evaluation of Deep Learning Methods for Attributed Graph Clustering

Attributed graph clustering aims to partition the nodes in a graph into groups such that the nodes in the same group are close in terms of graph proximity and also have similar attribute values. Recently, deep learning methods have achieved state-of-the-art clustering performance. However, the effectiveness of existing methods remains unclear due to two reasons. First, the datasets used for evaluation do not support fully the goal of attributed graph clustering. The category labels of nodes are only relevant to node attributes, and nodes with the same category label are often distant in the graph. Second, existing methods for the attributed graph clustering are complex and consist of several components. There is lack of comparisons of methods composed of different components from existing methods. This study proposes six benchmark datasets that support better the goal of attributed graph clustering and reports the performance of existing representative methods. Given that existing methods leave room for improvement on the proposed benchmark datasets, we systematically analyze five aspects of existing methods: encoded information, training networks, fusion mechanisms, loss functions, and clustering result generation. Based on these aspects, we decompose existing methods into modules and evaluate the performance of reconfigured methods based on these modules. According to the experimental results on the proposed benchmark datasets, we identify two promising configurations: (i) taking the attribute matrix as input to a graph convolutional network and (ii) layer-wise linear fusing deep neural network and graph attention network. And we also find that complex loss function fails to improve the clustering performance.