Improving Collaborative Filtering With Psychology Model

Because the amount of data on the web grows tremendously, it becomes difficulty for users searching useful information. Because of the diversity of users'background, they need more specific information. Under this scenario, how to provide the users with the exact information they need, becomes a most focused problem. One typical solution to this problem is the personalized recommendation systems. Among the existing recommendation systems, collaborative filtering system is the most widely used one. Recommendation systems predict users'interests, and provide the users with items they may like. Many studies show that E-commerce sites have made great profit from recommender systems.Although E-commerce has benefited from recommendation systems, the performance of existing recommendation systems are still unsatisfactory, and need improvement. In order to provide high quality recommendations to the users, we propose to study user's psychology to understand their thinking and behavior pattern. This thesis applied psychology theory to improve collaborative filtering from three aspects, including rating collection, similarity measure, and rating prediction.First, this thesis proposed a purchase prediction model based on the attitude-behavior relationship theory. The proposed purchase prediction model can analyze the web user's interests, figure out user's attitude to the objects through his browsing behavior, and identify the user's most wanted item from a large number of items. Experiments on E-commerce data set showed that the proposed purchase prediction model can effectively predict the user's buying behavior. Then we implemented the proposed purchase prediction model to improve the collaborative filtering algorithm. We regard the probability of purchase as implicitly collected ratings of the users, i.e., use the purchase prediction model as rating collection model. We adopted the predicted ratings to improve both user-based and item-based collaborative filtering algorithms. We compared the improved algorithms with the traditional collaborative filtering algorithms which only use user click data, and experimental results showed that improved methods perform better than traditional ones.Tversky's Contrast Model pointed out, when people define whether two objects are similar, they do not only judge by the common characteristics, but also take different characteristics into consideration. Based on Tversky's theory, this thesis proposed a new user similarity measure by taking both common and different feature of user's ratings into consideration. We applied a basic fractional function and an exponential function to calculate the similarity in collaborative filtering algorithms. We compared our proposed similarity measure with cosine similarity, pearson's correlation and spearman's correlation. Experiments were conducted on two data sets, Movie Lens and Book-Crossing data sets. Experimental results showed that our basic fractional function slightly outperformed other measures, while exponential function significantly outperformed the other similarity measures.At last, we improved rating prediction model in collaborative filtering algorithm. We introduced the number of neighbors involved in the voting into the model as a weighting factor. Experiments on Movie Lens and Book-Crossing data sets showed that, the weighting factor improved the recommendation performance. The results proved that the number of neighbors involved in the voting did have an effect on the rating prediction.This thesis applied psychology theory to improve collaborative filtering from three aspects, including information collection, similarity measure, and rating prediction. The experimental results proved that the proposed models improved collaborative filtering algorithms.