A User-based Collaborative Filtering Method to deal with Sparsity in Recommendation Systems by an unsupervised learning of Users’ Hidden Preferences

Abstract

This paper focusses on sparsity in User-based Collaborative Filtering (UCF) type of Recommendation Systems. UCF mainly depends on Users’ Similarity Calculation (USC). The idea of the proposed method to overcome the sparsity problem of UCF is, overcoming the sparsity effect on USC through imputation of missing ratings. In the proposed method imputation is carried out by finding the hid-den preferences of users through an unsupervised learning using K-Means, basing on given ratings and item features. The proposed method aims to provide adequate information to the similarity metric used for USC, even with sparse rating data. The proposed method is compared with some other approaches in terms of Root Mean Square Error (RMSE), Mean Absolute Error (MAE) and Mean Squared Error (MSE) values with varying levels of Sparsity and the RMSE, MAE and MSE are found to be the least for the proposed method. Also the RMSE of the proposed method for all the levels is close to 1 with the rating range of the dataset used being [1], [5] and the fact that the error-rate being almost constant across the sparsity levels shows that the proposed method is not greatly affected by sparsity.