Popularity Bias Debiasing in RecSys

Trained a BPR matrix-factorization recommender on MovieLens 25M to tackle popularity bias, the tendency to over-recommend blockbusters while better niche items go unseen. Benchmarked two fixes against the biased baseline: reweighting the training loss by inverse item popularity, and re-ranking the model's top candidates with a popularity penalty. A lightweight re-ranker cut average recommended-item popularity by ~30% for under a 10% drop in ranking accuracy, no retraining required. A single Pareto chart maps the full tradeoff, showing which method wins at each fairness target.

Dataset

MovieLens 25M

Models

BPR matrix factorization, inverse-popularity reweighting, popularity-penalty re-ranker

Evaluation

Recommended-item popularity, ranking accuracy, Pareto tradeoff

Results & insight

Re-ranker cut item popularity ~30% for under a 10% accuracy drop, no retraining