HANE-SHC: Heterogeneous attributed network embedding with structural homophily contents

Yue Fu, Shuliang Zhao, Yongliang Wu

Abstract: Heterogeneous attributed networks (HANs) are ubiquitous in real-world, spanning domains from the academic network to the movie network. Heterogeneous attributed network embedding (HANE), a promising analytical technique in recent years, aims to generate low-dimensional dense vectors containing network structure and attribute content. Compared to the embedding approaches that only consider topology structure, researchers embed the network by leveraging both network topology and content to improve the quality of embedding vectors. However, when two nodes with a connected edge may have a small content similarity value, while two nodes without a connected edge may have a large content similarity value, so, in this situation, it is not appropriate to utilize the network topology and nodes’ content to generate embedding vectors. To overcome this challenge, we propose a novel approach of Heterogeneous Attributed Networks Embedding based on Structural Homophily Contents, abbreviated as HANE-SHC, which is based on convolutional neural networks to leverage both the network structure and the content of nodes connected with the embedded node to generate the representation vectors of the embedded node. To our knowledge, this is the first study that proposes the Structural Homophily Contents to enhance the heterogeneous attributed network embedding. Through comprehensive evaluations on downstream tasks including clustering, classification, link prediction, and visualization, the experimental results demonstrate that the HANE-SHC outperforms the baselines. On the artificial synthetic networks with various network structure-content dissimilarity scores, the HANE-SHC model consistently produces high-quality embedding vectors regardless of the network dissimilarity scores between structure and content, significantly outperforming existing approaches.

External IDs:dblp:journals/apin/FuZW25