ECTI TRANSACTIONS ON COMPUTER INFORMATION TECHNOLOGYVolume 16, No. 04, Month DECEMBER, Year 2022, Pages 469 - 478
Multiple latent space ensemble for matrix factorization based collaborative filtering
Vasin Jinopong, Dussadee Praserttitipong, Jakramate Bootkrajang
Abstract Download PDFMatrix factorization is arguably one of the most widely employed collab- orative ltering techniques for recommender systems. The recommendation is obtained based on the user-item relationships discovered within some lower-dimensional latent space. The optimal latent space is generally data-dependent and often needs to be selected using the time-consuming cross-validation scheme. In this paper, we propose to leverage the power of the ensemble method not only to facilitate the hyper-parameter selection but also to improve the predictive performance of the system. Specically, we studied ways to combine predictions from multiple Singular Value Decomposition models, each operates in its own latent space. Experimental results based on MovieLen100K, MovieLen1M, Bookcrossing and Filmtrust datasets demonstrated that the ensembles outperformed a tuned single model in terms of RMSE and MAE while requiring no additional model selection step. Ensemble sizes experiment have shown that the 21 submodel of the ensemble models produce better results than the 14, 8 and standalone model. However, it takes longer to complete. We also found that an ensemble that pays more attention to lower-dimensional latent spaces tends to generalize better.
Collaborative Fil- tering, Ensemble, Matrix Fac- torization, Singular Values De- composition