A Refined SVD Algorithm for Collaborative Filtering
This paper offers an incremental improvement to the initialization step of the SVD algorithm for collaborative filtering, which is relevant for recommender systems.
This paper addresses the problem of predicting user ratings for items in collaborative filtering. It proposes a refined initialization technique for the Singular Value Decomposition (SVD) algorithm, leveraging K-means clustering to improve upon an existing SVD initialization method. The authors demonstrate that their combined technique outperforms both SVD and K-means when used separately.
Collaborative filtering tries to predict the ratings of a user over some items based on opinions of other users with similar taste. The ratings are usually given in the form of a sparse matrix, the goal being to find the missing entries (i.e. ratings). Various approaches to collaborative filtering exist, some of the most popular ones being the Singular Value Decomposition (SVD) and K-means clustering. One of the challenges in the SVD approach is finding a good initialization of the unknown ratings. A possible initialization is suggested by [1]. In this paper we explain how K-means approach can be used to achieve the further refinement of this initialization for SVD. We show that our technique outperforms both initialization techniques used separately.