Abstract
Recommendation Systems (RS) are often plagued by popularity bias. Whentraining a recommendation model on a typically long-tailed dataset, the modeltends to not only inherit this bias but often exacerbate it, resulting inover-representation of popular items in the recommendation lists. This studyconducts comprehensive empirical and theoretical analyses to expose the rootcauses of this phenomenon, yielding two core insights: 1) Item popularity ismemorized in the principal spectrum of the score matrix predicted by therecommendation model; 2) The dimension collapse phenomenon amplifies therelative prominence of the principal spectrum, thereby intensifying thepopularity bias. Building on these insights, we propose a novel debiasingstrategy that leverages a spectral norm regularizer to penalize the magnitudeof the principal singular value. We have developed an efficient algorithm toexpedite the calculation of the spectral norm by exploiting the spectralproperty of the score matrix. Extensive experiments across seven real-worlddatasets and three testing paradigms have been conducted to validate thesuperiority of the proposed method.