Optimal Hyperparameters for Collaborative Filtering: A Comparative Study

In recommendation systems, collaborative filtering algorithms play a vital role in offering personalized recommendations based on user-item interactions. This paper conducts a detailed comparative study of three collaborative filtering algorithms, including Alternating Least Squares (ALS), Singular Value Decomposition (SVD), and Neural Collaborative Filtering (NCF). The paper's objective is to provide the optimal hyperparameter configurations for the three algorithms, which would boost their predictive power, using the Movie Lens 25M dataset. Initially the algorithms are evaluated with default parameters whose results shows that SVD performs better in Mean Absolute Error (MAE) while ALS performs better in RMSE. However, after the hyperparameter tuning with multiple parameters, such as epoch, rank, maxIter, learning rate, and batch size, the most optimal hyperparameter configuration for the algorithms was defined. The results tells that each of the algorithm have the capacity to handle sparsity. According to the RMSE and MAE, ALS and SVD have a slight advantage over NCF. These results are essential for choosing the appropriate hyperparameters enhancing the accuracy of the algorithms and the recommendation system as a whole.