A Scalable Ensemble Architecture for Collaborative Filtering in Recommender Systems

Matrix decomposition methods are extensively used for Collaborative Filtering in Recommender Systems. This research work investigates the effectiveness of various Matrix decomposition methods for Collaborative Filtering (CF) to predict recommendations. There is a tradeoff between the scalability and quality of predictions; Recommendations made by Singular Value Decomposition (SVD) based algorithms are of high quality but matrix decomposition based on SVD is associated with high computational cost and also requires large memory space. Low-rank approximation methods like CUR, CMD and Colibri are much faster than SVD but compromises on the quality of predictions. The authors propose ensemble approach to each of these matrix approximation methods combined with SVD++ to improve the accuracy while maintaining the scalability. Extensive experimentation is done on MovieLens datasets and the results are in support of the authors claim.