A Comparative Study of Different Similarity Metrics in Highly Sparse Rating Dataset

Abstract

Recommender system has been popularly used for recommending products and services to the online buyers and users. Collaborative Filtering (CF) is one of the most popular filtering approaches used to find the preferences of users for the recommendation. CF works on the ratings given by the users for a particular item. It predicts the rating that is not explicitly given for any item and build the recommendation list for a particular user. Different similarity metrics and prediction approaches are used for this purpose. But these metrics and approaches have some issues in dealing with highly sparse datasets. In this paper, we sought to find the most accurate combinations of similarity metrics and prediction approaches for both user and item similarity based CF. In this comparative study, we deliberately instill sparsity of different magnitudes (10, 20, 30 and 40%) by deleting given ratings in an existing dataset. We then predict the deleted ratings using different combinations of similarity metrics and prediction approach. We assessed the accuracy of the prediction with the help of two evaluation metrics (MAE and RMSE).