How Do Recommendation Models Amplify Popularity Bias? An Analysis from the Spectral Perspective

2024-04-18

Web Search and Data Mining

AI Summary

Recommendation systems often exacerbate popularity bias in long-tail scenarios, leading to overexposure of popular items. This paper is the first to reveal—via spectral analysis—that popularity information concentrates predominantly in the principal singular spectrum of the prediction score matrix, and that model low-rank approximation amplifies this bias through dimensional collapse. To address this, we propose a provably debiasing regularization method based on the principal spectral norm, accompanied by an efficient spectral norm approximation algorithm scalable to large datasets. Evaluated across seven real-world datasets and three evaluation paradigms, our approach significantly improves exposure fairness for long-tail items while maintaining or even enhancing recommendation accuracy. This work provides both a novel theoretical framework for bias attribution and a practical, controllable debiasing solution grounded in spectral analysis.

Technology Category

• Machine Learning: Learning Preferences or Rankings
• Data Mining & Knowledge Management: Recommender Systems
• Computer Vision: Bias, Fairness & Privacy

Application Category

• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Social Networks and Social Media: Fairness and bias in social network and social media analysis
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

Recommendation Systems (RS) are often plagued by popularity bias. When training a recommendation model on a typically long-tailed dataset, the model tends to not only inherit this bias but often exacerbate it, resulting in over-representation of popular items in the recommendation lists. This study conducts comprehensive empirical and theoretical analyses to expose the root causes of this phenomenon, yielding two core insights: 1) Item popularity is memorized in the principal spectrum of the score matrix predicted by the recommendation model; 2) The dimension collapse phenomenon amplifies the relative prominence of the principal spectrum, thereby intensifying the popularity bias. Building on these insights, we propose a novel debiasing strategy that leverages a spectral norm regularizer to penalize the magnitude of the principal singular value. We have developed an efficient algorithm to expedite the calculation of the spectral norm by exploiting the spectral property of the score matrix. Extensive experiments across seven real-world datasets and three testing paradigms have been conducted to validate the superiority of the proposed method.

Problem

• Analyzes how recommendation models amplify popularity bias.
• Identifies item popularity memorized in principal spectrum.
• Proposes spectral norm regularizer to reduce popularity bias.

Innovation

• Spectral norm regularizer reduces popularity bias.
• Efficient algorithm speeds up spectral norm calculation.
• Empirical and theoretical analyses reveal bias mechanisms.