A Bi-Channel Co-Clustering Algorithm for Heterogeneous Information Networks

Network clustering has received increasing attention for its ubiquitous real-world applications. Heterogeneous information network (HIN) clustering improves traditional homogeneous network clustering, as HIN reserves heterogeneity of nodes and relations to enhance clustering. However, existing HIN clustering studies based on graph neural networks (GNNs) ignore different weights of node features and topology structures on clustering. Moreover, these methods only cluster target nodes of a single type, while do not consider the auxiliary of nodes of other types in HINs, which significantly degrades their performance. To this end, we propose a bi-channel co-clustering algorithm for heterogeneous information networks, abbreviated B3C, which is capable of merging node features and topology structures, as well as capturing the hidden correlations between heterogeneous nodes, in order to achieve effective HIN clustering. Specifically, we first design a simple yet effective bi-channel encoder to aggregate neighborhood information w.r.t. topology structure and a similarity matrix. Then, self-training based clustering is performed to jointly optimize the cluster assignments while learning HIN representations. Next, the co-clustering mechanism is used to cluster nodes of different types simultaneously. Finally, we adopt the triplet-center loss to obtain discriminative node embeddings, so that similar nodes are condensed and dissimilar nodes are separated. Extensive experiments on public datasets demonstrate that the designed bi-channel encoder shows significant improvements over widely used GNN encoder and B3C outperforms the state-of-the-art learning-based HIN clustering competitors.