IMPACT OF IMPUTATION ON CLUSTER-BASED COLLABORATIVE FILTERING APPROACH FOR RECOMMENDATION SYSTEM

Abstract

The Collaborative Filtering (CF) widely used in Recommendation System commonly suffers the sparsity issue since the unobserved rating entries usually over dominance the observed ones. A clustering technique is an alternative solution that can solve the problem. However, no in-depth work has investigated how the missing entries should be mitigated and how the cluster-based approach can be implemented. In this study, we show how the imputed cluster-based approach deals with the missing entries, improving the recommendation quality. The framework of our method consists of four main stages: rating imputation to replace the missing entries, K-means clustering to group users or items based on the imputed rating data, CF-based prediction model, and generating the list of top-N recommendation. This paper uses three variations of imputation techniques, i.e., null, mean, and mode. The cluster-based approach is employed by using the K-Means as the clustering technique, and either the user-based or the items-based model as the CF approach. Experiment results show that the null imputation technique gives the best results when dealing with the missing entries. This finding indicates that the implementation of the clustering technique is sufficient for solving the sparsity issue such that imputing the missing entries is not necessary. We also show that our imputed cluster-based CF methods always outperform the traditional CF methods in terms of the F1-Score metric.

Keywords

• clustering, collaborative filtering, imputation, sparsity