Deep Attention Fusion Network for Attributed Graph Clustering

The field of graph clustering has gained prominence in data mining and machine learning due to the surge in big data and social networks. Deep clustering networks, which are commonly used for graph clustering, extract information from both attributes and structures to perform clustering tasks. However, the heterogeneity between the structure and attribute features of nodes can limit the model’s clustering performance. To address this, a deep attention fusion network is proposed for attribute graph clustering. The network comprises an autoencoder and a graph autoencoder, which independently extract the attribute and structural features from the attribute graph. An attention-based information fusion module is also introduced. This module links the autoencoder and the graph autoencoder layer by layer, enhancing the fusion of the two types of heterogeneous information between layers. Additionally, the information enhancement module, which has the capability to further enhance the final embedding vector, contributes to the overall effectiveness of the model. The model is trained using self-supervised learning on the attribute graph dataset. The results of the experiment indicated that this method can significantly enhance the clustering performance.