Collaborative Topic Regression with Multiple Graphs Factorization for Recommendation in Social Media

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 233–244, Dublin, Ireland, August 23-29 2014. Collaborative Topic Regression with Multiple Graphs Factorization for Recommendation in Social Media Qing Zhang Key Laboratory of Computational Linguistics (Peking University) Ministry of Education, China zqicl@pku.edu.cn Houfeng Wang∗ Key Laboratory of Computational Linguistics (Peking University) Ministry of Education, China wanghf@pku.edu.cn Abstract With a large amount of complex network data available from multiple data sources, how to effec- tively combine these available data with existing auxiliary information such as item content into the same recommendation framework for more accurately modeling user preference is an inter- esting and significant research topic for various recommender systems. In this paper, we propose a novel hierarchical Bayesian model to integrate multiple social network structures and content information for item recommendation. The key idea is to formulate a joint optimization frame- work to learn latent user and item representations, with simultaneously learned social factors and latent topic variables. The main challenge is how to exploit the shared information among multiple social graphs in a probabilistic framework. To tackle this challenge, we incorporate multiple graphs probabilistic factorization with two alternatively designed combination strate- gies into collaborative topic regression (CTR). Experimental results on real dataset demonstrate the effectiveness of our approach. 1 Introduction Many real-life data have representations in the form of multiple views (Liu et al., 2013). For example, web pages usually consist of both text content and hyperlink information; images on the web have relevant tags associated with their content. In addition, it is also common that in real networks comprising multiple types of nodes are connected by multiple types of links, forming heterogeneous information networks (HIN) (Huang et al., 2012). For example, in scientific community, various types of links are formed for different types of objects, i.e., author writes paper, venue publishes paper, reader labels tag, and so on. Therefore, with a large amount of complex network data available from multiple data sources, how to effectively combine this kind of rich structure with other auxiliary information such as content information into the same recommendation framework is an interesting and significant research topic for various recommender systems. This paper aims to model multiple social graphs into a principled hierarchy Bayesian framework to improve recommending performance. The basic idea in this paper is inspired by multi-view learning approach (Liu et al., 2013), i.e., leverag- ing the redundancy and consistency among distinct views (Kumar et al., 2011) to strengthen the overall performance. We extend this idea (Liu et al., 2013) originally for clustering problem to deal with rat- ing scarcity problem when modeling user preference for recommendation. Just as in general multi-view learning, each view of objective function is assumed to be capable of correctly classifying labeled exam- ples separately. Then, they are smoothed with respect to similarity structures in all views. Similarly, in this paper, we assume that our individual views of multiple user social relations are similar and comple- mentary with a shared latent structure. However, …

Content truncated to 500 words