Analysis of the ALS-MF (Alternating Least Square) Algorithm Matrix Factorization) with SVD (Singular Value Decomposition) in the Collaborative Filtering Method

Abstract

The development of the gaming industry, reaching $175.8 billion in 2021, triggers the need for precise recommendation systems on various platforms such as consoles, mobile devices, and emerging trends like e-sports and VR. Collaborative filtering, a commonly used method, has weaknesses in handling new or unrated items. Previous research shows that the Alternating Least Squares (ALS) algorithm can improve the accuracy of recommendation systems on the Goodreads dataset, achieving an RMSE value between 0.86 and 0.88. Additionally, ALS demonstrated superiority with an RMSE value of 0.641 when compared to other recommendation systems. Based on this, our research uses the collaborative filtering method with the ALS algorithm to overcome its weaknesses and increase the accuracy of game recommendations. The ALS algorithm was also compared with a system using SVD to provide algorithm comparisons. The final results showed that the optimal ALS model was obtained with a hyperparameter value of rank 1, regularization of 0.1, and 5 iterations, using the sklearn split method for data sharing, and percentile scores for the rating conversion method. Furthermore, ALS proved superior to SVD, achieving a Root Mean Squared Error (RMSE) of 1.1142 for test data and 0.8873 for training data.