Research On Modeling Approaches To Recommender Systems By Exploiting Multi-information

With the developing of Internet, Internet of Thing, e-commence, artificial intelligence, cloud computing and mobile computing, huge amount of new information and new products come out every day. In current Big Data era, information overload has become one of the core issues in various areas. As a result, recommender systems are playing an important role to deal with this problem. In fact, generalized recommendation technology has been used in many areas, including webpage ranking, spam filtering, online dating, etc., therefore recommender systems have been core technology in the areas of information retrieval, data mining, social network analysis, etc. This paper conducts deep analysis on the deficiencies of current recommender systems, and proposes novel and efficient solutions. Our work tackles with these deficiencies in terms of constructing machine learning models by exploiting multi-information. As a result, more precise representation of user preferences and item features can be learned from these models. Moreover, this paper designs new type recommender systems to meet newly emerging requirement.In real-world recommender systems, the minority of users/items usually account for a large amount of data while the majority of users/items are only associated with few data, which follows long-tail distributions. In real world, most users/items may be quite heterogeneous, which may fail to be captured by conventional recommender systems, especially in the case of the long-tail distribution. To address this issue, this paper proposes a latent feature based Bayesian heteroskedastic choice model(BHCM) which automatically group users/items by Bayesian nonparametric priors so as to generate nonidentically distributed features to represent the heterogeneities. Furthermore, BHCM is based on weaken binary assumption so as to represent the negative choices in a more reasonable way. That is, we designs a Bayesian heteroskedastic model to assign each choice or not choice with different confidence.In reality, the number of long-tail items and users is large but they only account for very few data, which are associated with three major challenges for current recommender system, namely data skewness, cold-start and shilling attack. Hence, the recommendations on the tail items or for the tail users tend to be unreliable. In this paper, we propose a Coupled Heteroscedastic Matrix Factorization(Co HMF) approach which consists of two coupled components: C-HMF for enhancing credibility and R-HMF for enhancing rarity. In specific, C-HMF and R-HMF recurrently refine regularization in terms of the empirical priors induced from each other so as to enable the estimates of parameters to produce reliable prediction on the long-tail items and users.A user usually has enough experience in some focused domains but lack of experience in other domains. Therefore, it is more helpful to conduct recommendation for users in their inexperienced domains by leveraging their preferences from their focused domains. As a result, Cross-Domain Collaborative Filtering(CDCF) has become an emerging research topic in recent years. In this paper, we propose a tensor factorization model Cross-Domain Triadic Factorization(CDTF) over the irregular triadic relation user-item-domain, which relaxes the constraint that requires identical items across all domains. As well as CDTF, another novel CDCF model, Bilinear Multilevel Analysis(BLMA), is also proposed, which seamlessly introduces multilevel analysis theory to the latent factor models. Specifically, we employ BLMA to more efficiently address the determinants of ratings from a hierarchical view by jointly considering the multi-level effects of domains, communities, and users so as to overcome the deficiencies in current CDCF approaches.Due to the social nature of human beings, group activities have become an integral part of our daily life, thus motivating the study on group-based recommender systems(GRS). However, most existing GRS approaches make recommendations through aggregating individual ratings or individual predictive results rather than considering the collective features that govern user choices made within a group. As a result, such methods are largely sensitive to data; hence they often fail to learn group preferences when the data are slightly inconsistent with predefined aggregation assumptions. To this end, we devise a novel GRS approach which accommodates both individual choices and group decisions in a joint model. With such a deep model, we can use high-level features, which are induced from lower-level features, to represent group preference so as to relieve the vulnerability of data.For each model mentioned above, we conducted thorough empirical evaluations with real-world datasets, and comparing with state-of-the-art approaches. The results prove that the models proposed in this paper can more effectively deal with the challenges in current recommender systems and satisfy newly emerging requirements.