Multiview learning of homogeneous neighborhood of nodes for the node representation of heterogeneous graph

Abstract

Multiview learning has caught the interest of many graph researchers because it can learn richer information about graphs from different views. Recently, multiview learning, as a novel paradigm in learning, has been widely applied to learn nodes representation of heterogeneous graphs, such as MVSE, HeMI, etc., they only utilize the local homogeneous neighborhood information of nodes, which degrades the quality of nodes representation. We are aware that the heterogeneous graph representation aims to drive the representation of a node to be near the homogeneous neighbors that are similar to it in the heterogeneous graph and far wary from heterogeneous neighbors. Besides, in the heterogeneous graph, linked nodes are more likely to be dissimilar, but remote nodes may have some similarities. Therefore, we can move the locality of a node to discover more homogenous neighbors’ information to improve the quality of node representation. In this work, we propose an unsupervised heterogeneous graph embedding technique that is simple yet efficient; and devise a systematic way to learn node embeddings from the local and global views of the homogeneous neighborhood of nodes by introducing a regularization framework that minimizes the disagreements among the local and global node embeddings under the specific meta-path. Inspired by Personal PageRank graph diffusion, we expand an infinite meta path-based restart random walk to obtain global homogenous neighbors of nodes and construct a meta path-based diffusion matrix to represent the relation between global homogenous neighbors and nodes. Finally, we employ mini-batch gradient descent to train our model to reduce computational consumption. Experimental findings demonstrate that our approach outperforms a wide variety of baselines on different datasets when it comes to node classification and node clustering tasks, with a particularly impressive 7.22% improvement over the best baseline on the ACM dataset.