Performance of User Behavior-Based Similarity on Top-N Recommendation

User-based collaborative filtering is one of the techniques in the recommendation systems that generate suggested items using the similarity between users. In some commercial systems, the recommended items’ accuracy is more important than the accuracy in predicting the rating. One of the factors that affect the accuracy of the recommendations is the similarity model. Recently, the development of similarity models has combined rating-based similarity and behavior-based similarity. However, the performance of these similarity models is only measured using rating prediction metrics without measuring the accuracy of the resulting recommendations. Therefore, this study aims to measure the behavior-based similarity performance in generating recommended items and compare it to the implementation of rating-based similarity. The similarity models use UPCF, UPCSim, and Cosine, which were tested on the MovieLens 100k dataset. The experimental results showed that the behavior-based similarity models outperformed the traditional similarity model (Cosine) with increasing precision, recall, and F1 score by 0.36, 0.32, and 0.34, respectively (for UPCSim) and 0.31, 0.25, and 0.28, respectively (for UPCF). In addition, the UPCSim similarity model outperformed the UPCF with an improvement of precision, recall, and F1 score by 0.05, 0.07, and 0.06.