Performance Evaluation of Collaborative Filtering Recommender Algorithms

Recommender systems are used with frequency so that content/items are offered in a personalized way for each user, and it is important that these algorithms can accurately recommend content/items to these users (good predictive performance), as well as have a satisfactory computational performance - so that it is not necessary to use too many computational resources. Thus, this paper aims to evaluate some recommendation algorithms that use the memory-based Collaborative Filtering (CF) technique and to evaluate the influence of similarity metrics on the performance of these algorithms. Both algorithms and metrics are available in the scikit-surprise library. Two public databases were used: MovieLens 100k and MovieLens 1M. After the experiments, it was observed that the choice of similarity metric might influence the predictive performance and the prediction and training time of the algorithms. The MSD metric was the one that stood out in influencing, in a positive way, these results. It was also noticed that the database could influence both the predictive performance of the algorithms, as well as the RAM consumption.