A Weighted Meta-graph Based Approach for Mobile Application Recommendation on Heterogeneous Information Networks

• Published in: Service-Oriented Computing Vol. 11236; pp. 404 - 420
• Main Authors: Xie, Fenfang, Chen, Liang, Ye, Yongjian, Liu, Yang, Zheng, Zibin, Lin, Xiaola
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer International Publishing AG 2018; Springer International Publishing
• Series: Lecture Notes in Computer Science
• Subjects: Deep neural network; Factorization; Heterogeneous information network; Meta-graph; Mobile app recommendation
• ISBN: 9783030035952; 3030035956
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-030-03596-9_29

Explosive growth in the number of mobile applications (apps) makes it difficult for users to find relevant apps. Therefore, it is an urgent task to recommend desired apps for users. Traditional approaches focus on exploiting the context information, user’s interest, privacy, security and other features for app recommendation. Most of them do not consider heterogeneous information network (HIN) in the scenario of mobile app recommendation. HIN contains rich structure and semantic information, and it can satisfy various requirements of users and generate better recommendation results. In this paper, we propose a Weighted Meta-Graph based approach for app Recommendation, called WMGRec, on HIN. Specifically, we firstly introduce the concept of weighted meta-graph, which not only distinguishes different rating scores to depict the subtle semantics but also utilizes meta-graph to capture complex semantics. And then, we apply weighted meta-graph to measure the semantic similarity between users and apps. Furthermore, we leverage non-negative matrix factorization on user-app similarity matrix to obtain user latent features and app latent features. Finally, the concatenated user and app latent features are fed into the factorization machine & deep neural network model to learn the higher-order interactions and get the final prediction score. Extensive experiments conducted on two real-world datasets validate the effectiveness of the proposed approach compared to state-of-the-art recommendation algorithms.