Enhanced singular values decomposition technique with ensemble learning for collaborative filtering recommender system

Abstract

The matrix factorization technique is arguably one of the most widely employed collaborative filtering recommender systems. The recommendation results are derived from the user-item relationships discovered within some lower-dimensional latent space. Unfortunately, the optimal size of latent space dimension is generally data-dependent and is often tuned utilizing the time-consuming cross-validation scheme. Thus, this research proposes to leverage the power of the ensemble learning method to facilitate hyper-parameter selection and improve the recommendation performance. The recommender system results were obtained by combining the prediction results of multiple Singular Value Decomposition models generated from learning processes with different latent space dimensions. Experimental results based on MovieLen100K, MovieLen1M, BookCrossing, and Filmtrust datasets demonstrated that the proposed technique outperformed a tuned Singular Value Decomposition technique in terms of RMSE and MAE values. Additionally, the recommendation model creation time was reduced because no other brute-force process for figuring out the optimal size of latent space dimension was required. Furthermore, the research results also indicate that ensemble learning that pays more attention to lower-dimensional latent spaces tends to generalize better.