Explainable link prediction based on multi-granularity relation-embedded representation

Jianxing Zheng

doi:10.1016/j.knosys.2021.107402

Return

Explainable link prediction based on multi-granularity relation-embedded representation

2021-10-01

PRE

DOI：10.1016/j.knosys.2021.107402

Original

Original request for help

Save

Abstract

En 中文

Existing link prediction methods focus on mining relations of nodes in terms of network structure, ignoring rich attributes of nodes. In the micro-blog social networks, text contents describe users' diverse behaviors, which depicts ones' multi-dimensional and multi-granularity preferences. Thus, based on users' multi-granularity interest, we can predict and explain social relations among users. In this work, we develop an explainable social relation extraction method based on hierarchical semantic specificity matrices of user interest. First, according to users' micro-blog contents, we model user interest subjects and design their semantic specificity matrices. Then, on the basis of interest subjects, we learn multi-dimensional and multi-granularity semantic specificity matrices with translation mechanism, which reflects user pairs' follow relation causes from multiple aspects. Finally, by leveraging multi-granularity semantic specificity matrices, we predict and explain users' social follow relation. To justify our proposal, we conduct extensive experiments on Sina Weibo and Tencent Weibo datasets. The experimental results demonstrate that the proposed model outperforms state-of-the-art methods in Hits, MeanRank and MRR metrics. The proposal can effectively predict users' follow relation from the aspects of multi-granularity interest subjects, and improve the accuracy of link prediction in social networks. (c) 2021 Elsevier B.V. All rights reserved.

Keywords:

Social relation extraction

User interest profile

Multi-granularity relation-embedded

representation

Explanation

AI Summary

Key information extracted from the uploaded paper, including a brief overview, abstract, background, key highlights, visual analysis, and future outlook.