Analyzing Trade-Offs Between Accuracy and Popularity Bias in Diffusion Recommender Systems via Post-Hoc Reranking
Author
Frydensbjerg, Christoffer Nørtoft
Term
4. term
Education
Publication year
2026
Submitted on
2026-06-04
Pages
7
Abstract
Recommender systems (for movies, music, products, etc.) often suffer from popularity bias: very popular items keep getting recommended, while less popular “long-tail” content stays hidden. Even advanced diffusion-based recommenders such as DiffRec and D3Rec perform well but still show this skew. Existing fixes like MACR act during training, yet they require retraining and altering the model, which is inconvenient when using pretrained systems. This thesis studies a lightweight alternative: post-processing reranking, which adjusts the order of suggested items to curb the dominance of popularity. We evaluate three strategies—Log Popularity Penalty (LogPop), Maximal Marginal Relevance (MMR), and Inverse Popularity Weighting (IPW)—on DiffRec and D3Rec, and compare them to MACR. Across multiple datasets, we assess both accuracy and diversity using common metrics (Recall, NDCG), as well as catalog coverage, entropy, and popularity measures. We also vary a trade-off parameter, λ, that balances relevance and diversity. Our results show that reranking consistently reduces popularity bias, but with different trade-offs between accuracy and diversity. LogPop behaves stably and predictably; IPW achieves strong bias reduction but is sensitive to λ; and MMR offers the most balanced overall performance. Overall, reranking is an effective and flexible alternative to training-time debiasing, though outcomes depend strongly on the chosen method and parameter settings.
Anbefalingssystemer (fx til film, musik og produkter) har ofte en indbygget popularitetsbias: meget populære emner får endnu flere anbefalinger, mens mindre kendt “long-tail”-indhold sjældent bliver vist. Selv avancerede, diffusionsbaserede anbefalingsmodeller som DiffRec og D3Rec leverer høj ydeevne, men lider stadig af denne skævhed. Kendte modforanstaltninger som MACR virker under træningen af modellen, men kræver gentræning og ændringer af modellen, hvilket er upraktisk, når man bruger fortrænede systemer. I denne afhandling undersøger vi et letvægtsalternativ: efterbehandling med omrangering, hvor man justerer rækkefølgen af de foreslåede elementer for at dæmpe popularitetens dominans. Vi tester tre strategier—Log Popularity Penalty (LogPop), Maximal Marginal Relevance (MMR) og Inverse Popularity Weighting (IPW)—på DiffRec og D3Rec og sammenligner dem med MACR. På tværs af flere datasæt måler vi både nøjagtighed og mangfoldighed med almindelige mål som Recall og NDCG samt katalogdækning, entropi og popularitetsmål. Vi varierer en afvejningsparameter, λ, der styrer balancen mellem relevans og diversitet. Resultaterne viser, at omrangering konsekvent mindsker popularitetsbias, men med forskellige afvejninger mellem nøjagtighed og mangfoldighed. LogPop opfører sig stabilt og forudsigeligt; IPW reducerer bias kraftigt, men er følsom over for λ; og MMR giver den mest afbalancerede samlede præstation. Samlet set er omrangering en effektiv og fleksibel erstatning for debiasering under træning, men effekten afhænger af metodevalg og indstilling af parametre.
[This apstract has been rewritten with the help of AI based on the project's original abstract]