Recommendation Models Based On Competitive Relationship

With the rapid development of the Internet, information overload is brought to users as the explosive growth of information available on the Web. Recommender systems are proposed as a solution to deal with the issue. Recommender systems provide personalized recommendations for users, which mine user preferences and have been widely researched and applied.Existing studies mostly assume that the ratings of different items from the same user are separately given, and that the ratings of the same item from different users are also separately given. However, the user rates not just considering the items themselves. More importantly, a rate will be given by considering the comparison with the history ratings of the other items or the comparison with the other ratings of the item. In other words, the rating is by large the result of comparison, and thus is relative. Basing on the real phenomenon, this thesis assumes that there are competitive relationships among the ratings of different items from the same user and among the ratings of the same item from different users. Four novel models are proposed for mining user preferences to provide personalized recommendations based on the competitive relationship, which is modeled by the probabilistic competition model. The main contributions of the thesis are presented as follows.A recommendation model based on the competitive relationship of contents is proposed. This model mines user preference based on the competitive relationship of contents of items. Firstly, it is assumed that a competitive relationship exists among items in the perspective of attracting user preferences, and thus the same user’s ratings of different items are results of item competitions. Secondly, the competitive relationship between items is assumed to be represented by the competitive relationship between contents of items, the competitive relationship between contents can be further represented by the competitive relationship between independent features lied in content. Basing on the above assumptions, the competitive relationship between content features is modeled by the competition model, and each user preference value of a feature is acquired to predict the user preference of a new item. Experiments on the two public datasets show that the model achieves high accuracy and fast recommendation speed and outperforms conventional recommendation approaches. Further, the recommendation performance of the model can improve as increasing competitive scales.A recommendation model based on the competitive relationship of contents with bigram features is proposed. On the basis of the first model, this model introduces bigram features based on co-occurrence to represent contents of items. The introduction of bigram features makes richer and more sophisticated competitive relationship of contents be modeled, and makes the model describe and distinguish user preferences on a finer granularity. Experiments on the two public datasets show that the model introducing bigram features as the content representation has better recommendation performance both on the two metrics than the content competitive relationship recommendation model, which is based on independent features. On the Netflix dataset, the model improves MAE by about0.9%and RMSE by about1.6%, comparing with the recommendation model based on the competitive relationship of contents.A collaborative filtering recommendation model based on competitive relationship is proposed. The model directly builds the competitive relationship of user rating data of items, and introduces competitive relationship into collaborative filtering. First of all, the model considers that a user’s rating of an item is usually compared with the other users’ratings of the item, and gets the competitive relationship among user preferences. The user preference values of the item are thereby acquired. Then, the user is represented by the user preference vector, which is built by the user preference values of different items. The similarities between users are computed, and recommendations are provided. Experiments on the two public datasets show that the model has higher accuracy than user-based collaborative filtering. Furthermore, the model is good at modeling the case that the same item gets diversified user ratings.A hybrid recommendation model based on competitive relationship is proposed. The model models competitive relationships from the above-mentioned two perspectives:the competitive relationships between items and between user preferences. The model generates a unified recommendation result, which combines the result of the recommendation model based on the competitive relationship of contents with bigram features and the one of the collaborative filtering recommendation model based on competitive relationship. By integrating the two complementary models, the model could deal with the over-specialization problem, the cold start problem and the sparsity problem existing in the two previous models. Experiments on the two public datasets show that the model has higher recommendation accuracy than the two single models. On the Netflix dataset, the model improves RMSE by about13.8%and9.1%, compared to the recommendation model based on the competitive relationship of contents with bigram features and the collaborative filtering recommendation model based on competitive relationship.