Advancing diversity in recommender systems: a model for enhancing societal well-being

Abstract Recommender systems are central to digital platforms, where they personalize information flows and influence user engagement. While these mechanisms improve convenience and satisfaction, they increasingly raise concerns about unintended consequences, such as reinforcing echo chambers, amplifying polarized viewpoints, and fostering patterns of overconsumption that resemble addictive behaviors. These effects can contribute to societal challenges, including reduced creativity, weakened critical thinking, and entrenched algorithmic biases. This article introduces an AI-based recommender model that addresses these risks by embedding diversity directly into the recommendation process. The approach incorporates three complementary dimensions of heterogeneity: variation in emotional tone, coverage across distinct content categories, and the balancing of political or ideological perspectives. The recommendation score is recalibrated using a weighted similarity method in which each dimension is assigned explicit parameters, allowing for flexible adjustment between accuracy and exposure to diversity. Experimental evaluations demonstrate that the model increases the variety of recommendations without significantly lowering predictive accuracy or user satisfaction. Compared with baseline approaches, including collaborative filtering, content-based filtering, and diversity-augmented re-ranking methods, the proposed model achieves substantially higher diversity (ILD = 0.60 vs. 0.30–0.50) and coverage (0.80 vs. 0.45–0.70) while maintaining comparable accuracy (MAP = 0.73 vs. 0.72–0.75). By broadening the scope of suggested content, the system fosters discovery of new interests while maintaining user engagement. Beyond performance outcomes, the method aligns with ethical design principles by promoting fairness, offering transparency through interpretable weighting schemes, and supporting autonomy by enabling users to adjust their diversity preferences. The results suggest that recommender systems can simultaneously serve personal relevance and societal well-being when diversity is treated as an integral design objective rather than a secondary constraint.