Uma Abordagem em Etapa de Processamento para Redução do Viés de Popularidade

Recommendation systems are designed to provide personalized suggestions to each user to enhance user experience and satisfaction across various applications. However, despite their widespread adoption and benefits, such as increased user retention and pro-fits, certain challenges persist, particularly popularity bias, which impacts the quality of recommendations. This bias introduces in-consistencies among user groups, resulting in issues such as lack of calibration, unfairness, and filter bubbles. To address these challenges, several studies have proposed calibration strategies to improve the quality of recommendations and achieve consistency among user groups, focusing on mitigating popularity bias. However, integrating these approaches into a unified model remains a challenge. This study proposes an innovative approach combining popularity-based personalized calibration with the Bayesian Personalized Ranking (BPR) method in the processing step. Our approach aims to provide consistent and fair recommendations while leveraging the efficiency gains of the BPR method. Experimental results on different datasets demonstrate the effectiveness of our modified approach in achieving comparable or superior results to state-of-the-art methods in terms of ranking, popularity, and fairness metrics.