An improved collaborative filtering algorithm combining content-based algorithm and user activity

Abstract

Collaborative filtering, which plays an important role in making personalized recommendation, is one of the most traditional and effective recommendation algorithms. However, this algorithm suffers the sparse user rating record matrix problem which would result in poor recommendation precision. A usual approach to alleviate this problem is filling empty values with user average rating value. This approach solve the sparse matrix problem to some degree, but the inaccuracy of the filling values seriously impact the veracity of recommendation. To further enhance the recommending precision, this paper propose a new method of user-based collaborative filtering based on predictive value padding. This algorithm would predict the empty values in user-item matrix by integrating content-based recommendation algorithm and user activity level before calculating user similarity. It considers both the role of user and the item attributes in order to make a more accurate prediction. Experimental results on movie-lens dataset has shown that our new algorithm improves recommendation accuracy significantly compared with traditional user-based collaborative filtering algorithm and has an obvious advantage over the recommendation result after padding with average rating value as well.