Similarity Measure Based on Low-Rank Approximation for Highly Scalable Recommender Systems

Abstract

Recommender systems are mostly used to make the appropriate personalized recommendation for different customers. Collaborative filtering recommendation is one of the most popular methods among E-commerce systems, but it has some shortcomings, such as cold starts, in which the system fails to consider items which no one in the community has rated previously, and sparse data, which is caused by a low number of rankings by users which results in a sparse similarity matrix. Most of the existing approaches have shortcomings of sparsity and scalability. In this paper we propose a method that approximates the matrix of users similarities with Nyström low-rank approximations and is based on Collaborative Filtering (CF). The proposed method avoids the high computation cost of Singular Value Decomposition (SVD) and also enables us to use the low-rank approximation of the similarity matrix to handle huge datasets with low computation costs. The experimental results show that the proposed approach can solve the problem of sparsity, while increasing the efficiency and scalability of the system.