AI summaryⓘ
The authors investigate how a common training method biased toward popular items changes how user profiles relate to each other in recommendation systems. They focus on a measure called the mean-centered user covariance, which shows how unique users' preferences are, and find it tends to shrink toward random noise when popularity bias is present. They provide a way to predict when this shrinkage happens based on training settings and confirm their findings with experiments on a movie dataset. However, they also find that the shrinkage effect is small with typical settings and does not harm recommendation quality in practice. Their method can help practitioners check if their system is at risk of this effect without complex simulations.
collaborative filteringpopularity biasBPR traininguser covariancemean-centeringitem-noise covarianceregularizationembedding geometryMovieLens-25Mrecommendation system
Abstract
We study how popularity-biased BPR training reshapes the between-user geometry of collaborative-filtering embeddings. We work with the mean-centered user covariance $C=\tfrac1n U^\top H U$, the object that measures how distinguishable users are from one another, as opposed to the uncentered second moment used in prior work. We prove that under popularity-biased feedback with stationary items, $C$ converges to a steady state proportional to the item-noise covariance $Q$. Thus between-user spread collapses toward a noise floor. We derive a closed-form, computable phase boundary in the training hyperparameters $(α,λ_{neg},γ,d)$ separating contraction from expansion, and validate both directional predictions on MovieLens-25M. We then examine the limits of the effect. At deployment-scale regularization the predicted contraction is real and policy-driven but small, and it is not reflected in any recommendation-level metric we measured. The $α$-driven anisotropic-collapse mechanism operates only at regularization strengths that degrade the recommender. A deployment-time restoration intervention derived from the theory does not improve recommendation quality. The boundary is computable from a trained model's embeddings, item interaction counts, and training hyperparameters, so a practitioner can check whether a deployed system sits in the strong-collapse regime without simulating the feedback loop. In our experiments the boundary places deployable settings far from that regime.