Robust representation learning for heterogeneous attributed networks

Abstract

The aim of heterogeneous attributed network embedding is mapping network into low-dimensional representations while preserving topological structure and attributed content. However, when the content similarity of two closely related nodes is small in the heterogeneous attributed network, we find that the embedding vectors obtained by combining network structure information and content information are bad. To tackle this problem, we propose a new robust representation learning model for Heterogeneous Attributed Networks Embedding with Dual-Space (HANE-DS), whose idea is to embed heterogeneous attributed network into dual spaces. Firstly, a heterogeneous attributed network is converted to a heterogeneous structure network and a heterogeneous content network. Secondly, the network nodes of the heterogeneous structure network are embedded into the structure space by using our proposed the structure embedding model, while the network nodes of the heterogeneous content network are embedded into the content space by applying our proposed the content embedding model. Finally, by utilizing the learning embedding vectors for downstream tasks, our approach can capture more comprehensive, more prosperous, and more reasonable information. Up to present, this is the first paper that focuses on the heterogeneous attributed network embedding based on Dual-Space. Three real-world networks and five synthetic networks to evaluate the HANE-DS model, and experimental results show that the HANE-DS can outperform baselines. For classification and clustering, the HANE-DS is superior to the baseline methods. For link prediction, the HANE-DS is in the top three of the baseline methods. For recommendation, the HANE-DS is also ahead of the baseline methods in terms of HR@K. On the synthetic networks, it is verified that the baseline models are sensitive to the content similarity of two closely related nodes. In comparison, our model is insensitive to that situation. In summary, the HANE-DS is a robust model for the heterogeneous attributed network embedding.