Counteracting Popularity Bias Amplification in Bundle Recommendations with Latent Factor Constraints

Bundle recommendation has become an important task in recommender systems, aiming to recommend sets of items to users. At the bundle level, bundle recommendation suffers from popularity bias similar to traditional item recommendation. However, a previously underexplored phenomenon is observed: at the item level, popular items within a bundle significantly increase the likelihood of the entire bundle being recommended, further amplifying popularity bias and leading to Matthew effect. Through empirical analysis, we find that the over-correlation between popularity signals and dominant latent dimensions leads to this bias amplification. To address this issue, we propose BALF (Bundle Amplification-aware Latent Factor constraints), a method that introduces spectral norm-based regularization terms at both bundle and item levels. By penalizing the magnitude of the principal singular value, BALF constraints the dominance of popularity-correlated directions within the latent space, thereby preventing the model from over-relying on and memorizing popularity signals during training. We evaluate BALF using four backbone models on four widely used bundle recommendation datasets under both in-distribution (IID) and out-of-distribution (OOD) settings. Experimental results show that BALF consistently achieves state-of-the-art performance on OOD, while effectively balancing the trade-off between debiasing and overall recommendation accuracy on IID. The datasets and code are available at WUT-IDEA/BALF .