Research On Maximum Entropy Recommendation Algorithm Combining User Attributes And Interests

E-commerce has provided great convenience for people, while the massivegrowing information makes people do not know how to choose his needing and result in"information overload". Therefore, personalized recommendation technology obtainswidely attention in the e-commerce activities, it can provide the goods information topeople that they may be interested quickly and accurately on the network, and also itenhances the purchase probability and benefits for businesses. Collaborative filteringhas been concerned widely by researchers and achieved great success. There is verypractical significance for using collaborative filtering technology to provide efficientrecommendation service and good experience for users when they are shopping on line.This thesis has analyzed multiple aspects of the collaborative recommendationtechnology and emphatically studies the application of the maximum entropy model forrecommendation system, establishes a collaborative filtering model based on mixedtechnologies. There has done following special study work in this thesis.First, this thesis studies and compares several major recommendation technologiesdeeply, including the basic structure of recommendation system, the realization ofseveral mainstream technologies usually applied and the comparison of their superiorityand insufficient in the practical application. Especially for the collaborative filtering,this thesis gives detailed description of the common algorithms, and analyses a varietyof problems the collaborative filtering technology facing and gives some of theimproving methods that solve those problems.Second, this thesis studies the theory of maximum entropy model and appliesmaximum entropy model to the personalized recommendation, achieves a new ratingpredict method differ from the traditional neighbor based method. In the new model,how to predict the rating classification of item is relying on the information of userinterests and attributes, the classification probability of model has the best balanceunder meet all feature constrains.Third, this thesis integrates the basic attributes information of user into maximumentropy model, establishes the feature constraints relying on the ratings of predicteditem and user attributes, this can provides stability predict results even when the ratinginformation is not enough.Fourthly, this thesis integrates the user interests information into maximum entropy model, the user interests is user-item rating in the thesis, selects conditional projectsbased on the method of maximum rating relevance, and establishes the featureconstraints relying on user-item rating, the maximum relevance metrics computes basedon the mutual information of information theory.Finally, this thesis conducts the test on the proposed method in several aspectsusing the MovieLens datasets, and compares our method with the traditional neighborbased collaborative filtering algorithm, the experimental results show that the proposedmethod has higher accuracy in predicting unknown user ratings.