Prediction of Preferences through Optimizing Users and Reducing Dimension in Collaborative Filtering System

摘要

Collaborative filtering systems employ statistical techniques to find a set of users known as neighbors, who have a history of agreeing with the target user. However, the problems associated with high dimensionality in the recommender systems have been discussed in several studies. In addition, the degree of correlation is computed between only two users. Although the preference correlation of two users may not be very high, their preferences can serve as useful data for preference prediction. The preference information of the two users, however, cannot be used to give a recommendation because the degree of their mutual correlation is low. The users preferences in collaborative filtering systems are not necessarily accurate information. Carelessly entered information stored in the collaborative filtering database must be excluded. In this paper, Entropy is used for optimizing users. Bayesian classification shall be used to classify items to lower the dimension of the user-item matrix in this paper. To extract the features of items this paper uses association word mining. Since this method is representing document as a set of association words, it prevents users from being confused by word sense disambiguation. Also, the users can be put into clusters by using the genetic algorithm on the classified items to solve the problems of previous clustering algorithm such as the need to predefine the number of clusters, the high sensitivity to noise in data, and the possibility their resulting data would converge with the region's optimal solution. Finally, it predicts typical preferences by using entropy to solve the problem depending on the correlation match between only two users. The typical preferences mean pseudo preferences that represent preferences of items within the group. A dynamic recommendation for an item is made using the typical preference. To give dynamic recommendations to users, the process of classifying users receiving recommendations into the most suitable group takes precedence.