Clustering Sparse Matrices Using Dimensional Reduction in Recommender Systems

The most advanced method for predicting ratings in recommender systems (RS) is collaborative filtering, which teaches us about similarities with other users or items. However, there is a drawback: the rating matrix we use for training is sparse. If a user does not rate an item, we cannot assume that the rating is zero in RS because this assertion implies that the user does not like the item. This is an error in RS's zero value assumption due to a lack of rating. The majority of machine learning algorithms must be reprogrammed to account for the lack of ratings. This problem can lead to other issues such as the inability to use public frameworks or open scripts and the lack of gain from parallel computing. In order to reduce sparsity in RS and enhance prediction quality and cluster identification, we concentrate on dimensional reduction strategies in this work. We investigate how applying alternative strategies operating in low dimensionality, coupled with similarity measures, improves cluster quality, predictions, and speeds up the learning process of models. We provide results over the publicly available MovieLens 1M dataset, which is the approach that the current RS has primarily adopted.