Asymmetrically Weighted Cosine Similarity Measure for Recommendation Systems

Abstract

AbstractA recommendation system (RS) provides suggestions about a variety of items to the users according to their interests. RS tries to predict the preference or rating that a user is expected to give to an item. Many techniques/algorithms, such as collaborative filtering-based, content-based, matrix factorization, machine learning, and many more, are used to recommend items to the users. One of the widely used and simple techniques/algorithms is collaborative filtering-based in which similarity between the users (or the items) is determined to make the recommendations. Some of the well-known similarity measures are cosine similarity (CS) and Pearson correlation (PC). However, these measures consider that the similarity between the two users/items is the same in either way. Therefore, in this paper, we propose a new similarity measure, called the asymmetrically weighted CS (AWCS) measure for RSs, which is inspired by the CS measure. We compare its performance with the CS and PC using mean absolute error (MAE), root mean square error (RMSE), and F1 score. In addition to that, we examine the effect of the number of users and multiple thresholds chosen to compute predicted ratings on the performance of the RS.KeywordsRecommendation systemsCollaborative filteringSimilarity functionCosine similarity measurePearson correlation measureMean absolute errorRoot mean square error