Research On Dynamics Of Collaborative Filtering Algorithms

The rapid growth of the Internet and Electronic commerce brings many conveniences for people’s life; at the same time it is easy for people to feel trampled by the Massive information. People have to spend a lot of time to obtain useful information. Search engine can help people to retrieve information, but in many cases people’s demand is not clear, they don’t know how to search. Recommender System (RS) solve this problem successfully, it use people’s profile and rating data to establish interest model. RS can push useful information for people, effectively reduce the cost of searching information and improve the user experience.Collaborative filtering (CF) is the most successful algorithm for RS. It can find neighbor users and items through user’s rating history, and then predict user rating with the neighbors. Despite the great success of CF, it is still facing many problems, such as cold start, sparse problem, and interest changing. The paper tries to solve the interest changing problem in probabilistic latent semantic (PLSA) algorithm.PLSA is an algorithm based on semantic analysis, it map user’s interest and item’s attribute into a latent variable z, then get the average score of variable z for each item and the probability of a user belong to z, at last it can predict the rating score for every user. The classic PLSA algorithm can’t process the interest changing problem, the paper save it with a rating window, separate long-term interests and short-term interests with two models. The long-term interest can be obtained with PLSA algorithm, the short-term interest is computed through similar items in rating window, then mix the two interests and get the final score. The result of experiment indicates than the Dynamic PLSA algorithm can successfully handle interest changing problem and obtain higher recommendation accuracy.