A Significant Fallacy of Built-in Correlation Functions in Recommender Systems

Especially in shopping and entertainment applications, having personalized recommendation modules is essential for customer comfort. In the frameworks of applications, the user or item similarity analysis is expected to be performed throughout the algorithm, which directs the user to the right items, thereby making suitable suggestions. Developers tend to choose readily available functions by taking advantage of the development environment; however, in some cases, internal functions of the programming platforms may be missing the requirements of the applied scientific field. In this study, the fallacy originating from the built-in Pearson correlation function utilized in user-based collaborative filtering is analyzed. The way of processing the vectors related to some arguments in the formula, such as handling the user rating history, may differentiate from the expected processing. This study shows how much the utilization of the inline function alters the outcomes. With the aid of three different datasets and various performance metrics, a large amount of experimental monitoring underlines the importance of handcrafted correlation functions special to the applied field but not the built-in methods. The results confirm that this simple fallacy has considerable implications.