Recommender Systems Using Limited Dataset Sizes

Abstract: In order to create personalized recommendations for users on services such as e-commerce websites and streaming platforms, recommender systems often utilize various machine learning techniques. A common technique used in recommender systems is collaborative filtering which creates rating predictions based on similar users’ interests. In this report we have compared two different approaches of implementing collaborative filtering by measuring their performance on movie rating datasets of varying sizes. The algorithms were singular value decomposition and imputation boosted collaborative filtering which were also compared to a baseline measurement of using the average rating as a prediction. The results show that utilizing singular value decomposition for a model-based approach is faster than using a memory-based approach with imputed data, but the memory-based approach gives more accurate predictions given a dataset size of more than around 25 users. This implies that the imputation boosted approach is mostly suitable for smaller dataset sizes of users, which might be found in less accessed services or services with infrequent recommendations. Singular value decomposition can instead be used for much larger dataset sizes albeit with a lower prediction accuracy, indicating that it can be used for services that handle more users and more frequent recommendations.