Research On Probabilistic Classification Model Based Recommendation Algorithm

How does user retrieve information from massive Internet data efficiently orinformation provider provide information to user accurately, it’s a difficult task in suchan information overload era. Personalized recommendation system can track the user’sinterest automatically and recommend products, web pages to user which may ofinterest to them; it’s an effective tool to handle the information overload problem.Traditional User-based and Item-based algorithm facing data sparsity and cold startproblem, they may be failed to give recommendation to user, because of there have notenough information about user’s interest in data sparsity circumstance.In order to alleviate the data sparsity and cold start problem, we treat personalizerecommendation as a probability classification problem in this article, using user’shistorical ratting information as an important gist for classification, and gather togetherwith user demographic data, item property and context information to make user’sinterest learning more accurate in data sparsity system.In this article, use classic generalized linear classification method such as logisticclassification method and maximum entropy classification method to learn ourprobabilistic model. We introduced the mathematical principia of logistic and maximumentropy method, and how to model our problem with them. At last, we get ourprobabilistic classification model based on logistic method (LC) and maximum entropymethod (MaxEnt).Finally, we take experiments on actual transaction dataset; tune the parameters ofour model; then compare the accuracy of recommendation with the traditionalcollaborative filtering algorithm. The recommendation accuracy of model LC andMaxEnt improved by4.85%and5.42%than traditional algorithm respectively. LC andMaxEnt has more stable performance than traditional method on sparse data sets.