Rich heterogeneous information preserving network representation learning. (December 2020)

• Record Type: Journal Article
• Authors: Yu, Bin; Hu, Jinzhi; Xie, Yu; Zhang, Chen; Tang, Zhouhua
• Is Part Of: Pattern recognition. Volume 108(2020:Dec.)
• Publication Date: 2020-12
• Subjects: Network representation learning -- Heterogeneous information -- Autoencoder; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2020.107564
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: A novel and unified heterogeneous information network representation learning framework is proposed, which integrates node proximities and semantic information. A meta-path based random walk strategy and the autoencoders are employed to preserve semantic and structural information of heterogeneous information networks. The comprehensive evaluations validate the superiority of our model against the state-of-the-arts. Abstract: Network representation learning has attracted increasing attention recently due to its applicability in network analysis. However, most existing network representation learning models only focus on preserving fragmentary aspects of network information, either node proximities or fixed semantic information. In this paper, we propose a novel algorithm named Rich Heterogeneous Information Preserving Network Representation Learning (HIRL), which integrates the high-order proximity among nodes and semantic information into a generic framework by exploiting a flexible autoencoder network. Based on the proposed algorithm, we can explore the hidden information in heterogeneous information networks through any custom form of path schema, and represents different types of nodes in a continuous and common vector space. Moreover, the proposed HIRL is applicable to homogeneous information networks. Extensive experimental results demonstrate that our approach can effectively preserve the information in networks under various path schemas, and performs better on real-world applications such as network reconstruction, link prediction, and node classification compared with the state-of-the-art methods.