An integrated topic modeling and auto-encoder for semantic-rich network embedding and news recommendation

In recent years, network representation learning is considered as a crucial research direction which explicitly supports multiple problems in information analysis and mining (INAM) domain. Among downstream tasks of INAM, news recommendation is considered as an important task, especially in semantic-rich/heterogeneous networks. Most of previous news recommendation models are mostly relied on the collaborative filtering (CF) approach. The CF-based techniques support to analyze the historical user–item interacting relationships. These relationships are used to extract latent factors and characterize the user’s preferences which are later utilize to facilitate the different recommendation tasks. Recent attempts also focus on the integrations of news recommendation with complex representation learning techniques to leverage the accuracy performance. Even multiple models have gained remarkable performance recently, they still encounter challenges. These challenges are related to the sparsity of user–item interaction data and thorough topic-driven user’s preference characterization. Moreover, these recent deep neural embedding-based recommendation models also suffer several limitations which are related to the capability of multi-viewed data embedding. Specifically, in the context of semantic-rich/network heterogeneity, they might be unable to fully incorporate the global structural representations of user–item interactions as well as associated data sources. Mainly motivated by remaining challenges, in this paper we propose novel topic-driven heterogeneous information network embedding-based technique which is aimed to effectively deal with the news recommendation, called as THIN4Rec. Our proposed THIN4Rec model enables to jointly capture the rich-semantic latent features of textual data and global structural representations of user–item interactions in the context of heterogeneous networks. The rich-semantic and structural representations of users and items are then used to improve the accuracy performance of news recommendation task. Extensive experiments in benchmark datasets prove the effectiveness of our works in comparison with recent state-of-the-art baselines.