| With the rapid growth of information in the Internet, personalized recommendation system plays an important role on industry and academia. However, the traditional recommendation system faces a number of challenges. On one hand, there are known issues, such as data sparse, cold start, and extension problems among traditional recommendation algorithms. On the other hand, the traditional recommender systems cannot satisfy’ the demand of users with the rapid development of data. In fact, there is too much additional and useful information e.g. social relationships between users, the detailed content information of items, ratings of specific items and even users, all these can be used to improve recommendation systems. How to integrate the useful complementary information into the recommender systems is one of the big challenges in practice. The personalized recommendation system aims to enhance accuracy and ease the problems in the existing recommenders.By using the probabilistic matrix factorization model, this paper attempts to cope with the above issues, especially with multiple attributes inherent in the recommender systems explicitly or implicitly. The traditional recommendation algorithms usually ignored the social relationships between users. Actually, social relationships between users are able to ease the data sparse, cold start problems concerning traditional recommendation system. What’s more, relationships between users are diverse and implicit which has the potential to depict the complex relationships involved in recommender systems. In addition, as Micro-blogging, Wechat, labeling systems and other content-sharing systems emerging rapidly, more and more content information can be integrated into the recommender systems to provide users with more accurate recommendation and good experience. Other information such as time information is very important as well in terms of modeling the evolving of users’ behaviours. Incorporating time series into the recommendation algorithm will greatly improve the performance of recommender. Therefore, this paper aims to combine the complementary information to analyze and solve problems faced by the existing recommendation systems.This paper proposes a model combining probability matrix factorization model and the topic model LDA, called SC-PMF model, to recommend for users as for the relationships between users and content information of items. The experiments conduct on two data sets:Bibsonomy and CiteUlike show that the more information (the relationships between users and content information as well) shared by users, the better performance the model exhibits. It is apparent that the social relationships between users and content information of items can effectively improve the performance of recommendation algorithm.Encouraged by the above work, we further propose the PMFST model by adding rating matrix, social relationship matrix and time information based on probability matrix factorization model for real-time recommendation. This study constructs a hierarchical framework based on the categories of items. More specifically, we first divide the items into several classes by their categories, then, recommend the items’ class for the users. Take this a step further, the PMFST model can recommend a specific item within the given category. The hierarchical framework not only narrows the scale of data, but also reduces the calculating. Due to the characteristics of the hierarchical framework, it ideally fits into the situation of Internet. To testify the availability of proposed model PMFST and hierarchical framework, Epinions and Ciao two data sets from real-world are used. Experimental results show that PMFST model can effectively ease the cold start, data sparse problems faced by the traditional systems. Furthermore, it can improve the performance of the recommendation system in terms of evaluation metrics. The framework can be extended and applied to the large-scale data scenarios. |
PDF Full Text Request