A comparison of matrix factorization algorithms for a movie recommender system
Michel Tabari, Rawand Sultani · KTH Publication Database DiVA (KTH Royal Institute of Technology) · 2018
Recommendation systems is a growing technique for providing a better user experience for discovering new content on a platform. It can be implemented in many contexts such as Netflix for recommending movies. There are many ways to implement recommendation systems. This paper investigated two of these methods - Weighted Alternating Least Squares and Stochastic Gradient Descent - which fall into the category of matrix factorization and measured their performance in regards to time taken for training, error convergence and prediction quality. To our help we have used TensorFlow, a machine learning framework developed by Google which have been providing us with algorithms, models for training, and testing. The results showed that the Weighted Alternating Least Squares model proved to be better in terms of prediction quality: We also found that the quality of our predictions relied heavily on the model's parameters, since optimal predictions for a model can be found through the correct tuning. We concluded that the choice of model depends heavily on the data set investigated, and that optimal parameters for one model cannot simply be transferred to another model.