Abstract
Recommender systems include a broad scope of applications and are associated with subjective preferences, indicating variations in recommendations.As a field of data science and machine learning, recommender systems require both statistical perspectives and sufficient performance monitoring.In this paper, we propose diversified similarity measurements by observing recommendation performance using generic metrics.Considering user-based collaborative filtering, the probability of an item being preferred by any user is measured.Having examined the best neighbor counts, we verified the test item bias phenomenon for similarity equations.Because of the statistical parameters used for computing in a global scope, there is implicit information in the literature, whether those parameters comprise the focal point user data statically.Regarding each dynamic prediction, user-wise parameters are expected to be generated at runtime by excluding the item of interest.This yields reliable results and is more compatible with real-time systems.Furthermore, we underline the effect of significance weighting by examining the similarities between a user of interest and its neighbors.Overall, this study uniquely combines significance weighting and test-item bias mitigation by inspecting the fine-tuned neighborhood.Consequently, the results reveal adequate similarity weight and performance metric combinations.