Similarity Measure and Instance Selection for Collaborative Filtering

Abstract

Collaborative filtering is used in recommender systems and e-commerce both to help customers find items they want to buy and to help businesses prepare goods they will offer. The k-Nearest Neighbor (KNN) method plays a central role in collaborative filtering by finding the k-nearest neighbors for a given user and employing them to predict the user's interests. This method suffers from problems of sparsity and nonscalability due to the sparseness of the user-item data set and the entire scan of such a data set in search of neighbors. These problems are solved with a matrix conversion method for compressing items into classes and an instance-selection method to narrow the scope of neighbor searching. The matrix conversion method avoids the "cold start" problem and makes predictions more accurate. The instance-selection method improves the performance of prediction without sacrificing accuracy. Combining these two methods results in better accuracy and performance.