Attributed Graph Clustering via Deep Adaptive Graph Maximization

Due to the increasing popularity of the social networks, the detection and discovery of the hidden building blocks and their community structures are considered as the primary tasks on the graph (network) based data structures. Graph clustering is considered as a challenging task as it requires contribution of input graph’s topological and content data jointly. Graph Convolutional Neural Networks (GCNs) have demonstrated remarkable power in the domain of graph representation learning by merging both structural and content information of networks. While GCN based clustering methods are being used as the state-of-the-art alternative solution for graph clustering, these methods fail to capture global structural information of networks, considering a local neighborhood of each node. Here we propose an integrated novel graph convolutional clustering approach that enables us to extract the local and global structures of the graph based data along with the nodes content. Experimental studies on three real-world benchmark information networks approve our approach and confirm that our proposed method outperforms baseline methods significantly in graph clustering and link prediction tasks.