Popularity Bias Analysis of Recommendation Algorithm Based on ABM Simulation

Abstract

With the rapid development of the Internet, recommendation algorithms are increasingly influencing consumers’ decision-making. The issue of fairness in recommendation algorithms, especially popularity bias, is also becoming more and more heated. For the research on popularity bias, most existing methods cannot balance long-term interactivity and low cost. Using the ABM simulation method, various variables in the recommender system can be controlled, and the long-term impact of the continuous interaction between the recommender system and the user can be studied, especially the influence of popularity bias. In this paper, we construct and implement a recommendation algorithm simulation framework based on ABM, and two algorithms, Item-Based CF and SVD, are respectively deployed on it to count item popularity distribution and Gini coefficient under multiple rounds of recommendation. The indicators, combined with the visualization results of user interest offset, are used to explore the popularity bias problem of two classic recommendation algorithms under multiple rounds of interaction. We also summarize the existing problems and make an outlook for future improvements.