Research On Recommendation Algorithms Via Incorporating Side Information

With the rapid development of internet technology, it is difficult for general users to find valuable information from the huge volume of data, i.e., general users are con-fronted with the serious information overload problem. Recommender systems have drawn lots of attention from industry and academia since they alleviate the informa-tion overload problem, by providing users with personalized information, products or services to satisfy their tastes and preferences.In practice, there are several challenges for recommender systems, such as data sparsity, scalability, cold start, accuracy and interpretability etc. Various of recom-mendation methods have been proposed to overcome one or several such problems. However, only utilizing past activities of users to make recommendations can not ef-fectively solve the inherent problems existed in recommender systems. Recently, var-ious of side information becomes available, such as item attribute information, social network information, geographical information and review text of users etc. Rich side information is a beneficial complement to users' past activities. As a result, the avail-ability of rich side information brings opportunities for us to alleviate the problem of lack of information in recommender systems. Meanwhile, how to incorporate side information into recommender systems to boost the quality of recommendation has become an important research problem in the field of recommender systems. In this thesis, we focus on the issues of data sparsity, scalability, cold start, accuracy etc., and study side information enhanced recommendation algorithms based on existing works in the field of collaborative filtering, social-network-based recommender systems and point-of-interest recommender systems etc.The main contributions and innovations are summarized as follows:1. We propose item attribute coupled based matrix factorization for item recom-mendation.Currently, most of exiting matrix factorization based recommendation methods only focus on dealing with the cold start user problem but ignore the cold start item problem. In addition, there are few suitable similarity measures for these content en-hanced matrix factorization approaches to compute the similarity between categorical items. In this work, we propose attributes coupling based matrix factorization method by incorporating item-attribute information to overcome the above problems. Specifi-cally, item-attribute information is used to constrain the process of matrix factorization via an item relationship regularization term. The item relationship regularization ter-m makes two item-specific latent feature vectors as similar as possible if those two items have similar attribute contents. Moreover, in order to capture the relationship be-tween items, Coupled Object Similarity (COS) is adopted to measure the interactions or couplings between items. Experimental results on two real data sets show that our proposed method outperforms the state-of-the-art recommendation methods, and can effectively cope with the cold start item problem.2. We propose an item recommendation method by integrating user social status and matrix factorization.With the popularity of social network applications, more and more recommender systems utilize trust relationships of users to improve the performance of recommenda-tion algorithms. However, most of existing social-network-based recommendation al-gorithms ignore the following problems:(1) in different domains, users tend to trust d-ifferent friends; (2) the degree of influence that a user is affected by their trusted friends is different in different domains since the user has different social status in differen-t domains. In this work, we first infer domain-specific social trust relation networks based on original users' rating information and social network information, and then compute each user's social status by leveraging PageRank algorithm for each specific domain. Finally, we propose a novel recommendation algorithm by integrating users' social status with matrix factorization model. Experimental results on real-world da-ta sets show that our proposed approach outperforms traditional social-network-based recommendation algorithms.3. We propose a point-of-interest recommendation method by exploiting location significance and user authority.The increasing popularity of smart devices and the growing development of GPS and Web2.0 promote the emergence of location-based social networks (LBSNs). Point-of-interest recommender systems have been indispensable in LBSNs since they provide users with new places which user may be interested in, by mining users'preferences from various of data sources. Although some approaches are proposed by taking POI as items in traditional recommender systems, traditional POI recommendation algorithms have the following problems:1) most of existing POI recommendation algorithms sim-plify users'check-in frequencies at a location, i.e.. regardless how many times a user checks-in a location, they only use binary values to indicate whether a user has visited a POI; 2) matrix factorization based POI recommendation algorithms totally treat users' check-in frequencies as ratings in traditional recommender systems and model user-s'check-in behaviors using the Gaussian distribution; 3) little methods systematically consider the effects of location significance and user authority on users'final check-in decision processes. In this work, we integrate probabilistic factor model and location significance to model users'check-in behaviors, and propose a location significance and user authority enhanced probabilistic factor model. Specifically, a hybrid model of HITS and PageRank is adapted to compute user authority and location significance. Moreover, user authorities are treated as personalized factors to weight users'implicit feedback. Experimental results on two real world data sets show that our proposed approach outperforms the state-of-the-art POI recommendation algorithms.4. We propose a ranking based Poisson matrix factorization model for point-of-interest recommendation.In addition to simplifying users'check-in frequencies at a location as well as total-ly treating users'check-in frequencies as ratings in traditional recommender systems, most of traditional POI recommendation methods ignore that users'check-in feedback is implicit and only positive examples are observed in POI recommendation. These methods only fit observed check-ins via point-wise regression and do not consider the partial orders of POIs. However, for POI recommendation, we care about more the preference order than the number of check-ins. In this work, we propose a ranking based Poisson matrix factorization model for POI recommendation. Specifically, we first utilize the Poisson distribution instead of the Gaussian distribution to model users' check-in behaviors. Then, we use the Bayesian Personalized Ranking metric to op-timize the loss objective function of Poisson matrix factorization and fit the partial orders of POIs. Finally, we leverage a regularized term encoding geographical influ- ence to constrain the process of Poisson matrix factorization. Experimental results on real-world data sets show that our proposed approach outperforms traditional POI recommendation algorithms.