Effects of Binary Similarity Metrics in Recommender Systems for Jester Jokes Dataset

Recommendation systems have become an indispensable element for online vendors that continiuosly increase their e-commerce activity. On the other hand, customers ask the system to get recommendations for the target product. The quality of the prediction becomes important from accuracy and efficiency point of view. Accuracy and efficiency of prediction may vary depending on different factors. In this study, we implemented k-modes clustering algorithm on a dense dataset called Jester Jokes. We studied effects of eleven binary similarity metrics on the quality of prediction while implementing collaborative filtering approach. The metrics used are: Anderberg (A), Dice (D), Gower2 (G), Hamann (H), Jaccard (J), Kulczynski (K), Ochiai (O), Pearson (P), Simple Matching (SM), Sokal and Sneath (SS), and Yule (Y). Different numbers of test and train users as well as different number of clusters were tested. To check the quality of predictive performance, the evaluation criteria f-measure and classification accuracy were used. Online durations were also recorded to check the computational efficiency of the system. As a result, G performed the worst from the f-measure point of view. A, H, and Y were also among the worst performers among the metrics. On the other hand, J was giving promising results for most of the experiments. Other similarity metrics had mixed results. In terms of the classification accuracy, all binary similarity metrics had mixed results. From the efficiency standpoint, only SS performed good in all experiments. H, P, SM, and Y were among the better performers in most experiments. On the other hand, D, G, J, K, and O were among the worst performers. A similarity metric provided mixed performance. Our study is one of the first studies comparing numerous binary similarity metrics for recommender systems based on both predictive and computation performance.