Publication

Abstract : Recommendation systems (RS) aim at prediction of user preferences for a given set of items. Conventionally, RS uses collaborative filtering, content-based filtering, or hybrid of both the approaches for generating recommendation lists. Each of these approaches suffers from one or the other problems such as cold-start, sparsity, scalability, processing streaming data, and low latency. Furthermore, with high frequency of data updates, the user preferences change too, that demands for latest recommendation list for each user based on recent activity, capturing the concept shift and eliminating the stale item preferences. Hence, we propose a solution approach based on matrix factorization, that is robust, handles sparse and streaming data. The user preferences are represented in a matrix form and are decomposed into smaller matrices for ease of interpretation and information retrieval. Four matrix factorization techniques, namely NMF, NMFALS, CUR, and SVD, have been used for empirical analysis, and it is found that SVD outperforms NMF, NMFALS, and CUR in terms of time and recommendation accuracy.