An Experimental Study on the Accuracy and Efficiency of Some Similarity Measures for Collaborative Filtering Recommender Systems

Abstract

Similarity measure is the fundamental component used in collaborative filtering recommendation technique to provide ratings prediction to users by employing either item-based or user-based recommender algorithms. The collaborative filtering has been widely implemented using various similarity measures but ignores to consider the time taken by the similarity measures to make accurate predictions in different application domains. This paper attempted to assist recommender systems developers to understand appropriate similarity measure depending on the application domain under consideration with less execution time and error rate. It also takes the effect of neighbrhood sizes (k) on the prediction accuracy and efficiency into consideration. The experimental evaluations were conducted on the four similarity measures with the same dataset using Python programming language implementation. The evaluation metrics considered during the experiments are Execution Time, Mean Absolute Error and Root Mean Square Error. The results of the evaluation demonstrated that, Manhattan Distance similarity measure had the best accuracy as well as efficiency of predictions in this study