Multi-hop Information-based Graph Convolutional Network for Clustering

Clustering is an essential and demanding undertaking in data analysis. The combination of traditional neural networks and graph convolutional networks (GCNs) has been extensively discussed in clustering tasks, in which the deep clustering methods learn useful content information and the graph convolutional networks mine the structured neighboring information in the graph data. However, the existing works equally consider the importance of different features to clustering and only focus on the nearest neighboring information in the structured features, ignoring the features of the long distant neighbors in the multi-hop information, resulting in inferior performance. We propose a novel multi-hop information-based graph convolutional network (MIGCN) for clustering to overcome these disadvantages. Specifically, we fuse content features and structured features with adaptive weights, which are dynamically adjusted according to the training results. Meanwhile, we utilize a multi-hop information module to extract structured features. Moreover, we design a multi-supervision mechanism and guide the training and updating of the whole model. Our method achieves better results than previous deep clustering approaches because it takes into account the diverse features embedded in the neural network and a dynamic fusion strategy for clustering. Our model has undergone rigorous testing on widely-accepted standard datasets, employing both objective and subjective evaluation measures. The results consistently demonstrate that our model consistently outperforms state-of-the-art techniques, validating its superior performance.