Clustering with Adaptive Unsupervised Graph Convolution Network

Graph clustering has become one of the most challenging problems in Deep Learning in recent years. There are a number of methods for classifying nodes, including Graph Convolution Network (GCN), a deep semi-supervised learning method. In this chapter, based on GCN architecture, we propose a deep unsupervised learning scheme. The main contributions are as follows. First, the whole architecture is trained with two unsupervised learning losses based on kernelized features and spectral smoothness. Second, spectral smoothing uses an adaptive and additional graph matrix associated with the predicted soft cluster assignments (node representations) and adaptively integrates additional structure information during the learning phase. The adaptive fused graph used for loss of spectral smoothness takes into account structural information coming from both data features and node deep representations. With the proposed objective function, we are able to develop a powerful graph-based deep clustering. Experiments on four benchmark datasets show that our proposed unsupervised GCN provides better clustering performance than other Graph Neural Network-based techniques for most datasets.