A Study Of Personalized Recommender Systems Based On Object-oriented Thought And Typical User Group

With the development of the Internet and the progress of information tech-nology, Recommender Systems have been a powerful tool, which can help humans obtain valuable knowledge from massive information, and avoid the information explosion. However, there are two key issues in the practice of Recommender Systems, one is how to use rich context information correctly to improve rec-ommendation performance, and another is how to efficiently handle amass data from vast users and items. These problems in practical application has given a great challenge to existing Recommender Systems. To that end, in this thesis, we make a focused study of dealing with these problems with contextual informa-tion processing, feature selection and typical user finding, so that the efficiency and accuracy of Recommender Systems can be both improved. Our contributions could be summarized as follows:Firstly, we propose a method to represent Objects in Recommender Systems based on feature-value pairs, and design a framework named Object-oriented Rec-ommender System (ORS). Though some models have been developed with some specific contextual information, they lack capacity to systematically and dynami-cally incorporate multiple types of additional contextual information. Therefore, we discuss how to introduce more additional contextual information for recom-mendation based on object-oriented thought. Specifically, the context information are uniformly represented as feature-value pairs, and the Objects are the collec-tions of the feature-value pairs. Then computing the similarity of the Objects by the relationships of feature-value pairs, the recommendation list is generated by the collaborative filtering method. We design the ORS framework and develop a similarity model named Objected-oriented Bayesian Network (OBN). The exper-imental results on the real-world data set of tourism domain show that the ORS framework leads to better recommendation performances.Secondly, we propose Objected-oriented Topic Model(OTM), and design the feature selection method using feature information entropy. In the practical ap-plication of Recommendation Systems, there is a realistic problem that how to use rich contextual information correctly to improve recommended results. In this thesis, the study found that different types of context information have d-ifferent contributions to recommendation. Based on the findings, in this paper, we propose the OTM model which can extract users’implicit interests from the hybrid contextual information, calculate the feature information entropies of dif-ferent context information by the OTM model, then select the features with more contribute and lower entropy. The experimental result on the real-world data set shows that the OTM model has better recommendation effect with explainable topic distribution, also can be used for selecting effective features.Lastly, we provide a concept of Typical User Group (TUG) as a represen-tative user subset in Recommender Systems, and develop a modified TUG-based Collaborative Filtering (TUG-CF) algorithm. For being convenient to recommen-dation study, researchers are used to select a small user subset, but rarely consider the representation of the subset. So that we defined the TUG as a representative user subset with higher item cover rate and rating accuracy. We also developed TUG-CF, which can discover the nearest neighbors in TUG with both lower com-putational cost and higher representation. Experimental results on the real-world data set show that TUG is better than other user subsets on the comparison of item cover rate or rating accuracy, and TUG-CF has better recommendation results than traditional collaborative filtering methods.