Search papers, labs, and topics across Lattice.
The paper introduces A2G-DiffRec, a diffusion-based recommender system that mitigates popularity bias and enhances item-side fairness by adaptively guiding the main diffusion model with a less-trained "weak" version of itself. This adaptive autoguidance dynamically weighs the outputs of the main and weak models during training, using a popularity regularization term to encourage balanced item exposure. Experiments on three datasets demonstrate that A2G-DiffRec improves item-side fairness with only a small reduction in recommendation accuracy compared to existing methods.
Autoguidance—using a model to guide itself—can effectively reduce popularity bias in diffusion recommenders, leading to fairer item exposure without significantly sacrificing accuracy.
Diffusion recommender systems achieve strong recommendation accuracy but often suffer from popularity bias, resulting in unequal item exposure. To address this shortcoming, we introduce A2G-DiffRec, a diffusion recommender that incorporates adaptive autoguidance, where the main model is guided by a less-trained version of itself. Instead of using a fixed guidance weight, A2G-DiffRec learns to adaptively weigh the outputs of the main and weak models during training, supervised by a popularity regularization that promotes balanced exposure across items with different popularity levels. Experimental results on the MovieLens-1M, Foursquare-Tokyo, and Music4All-Onion datasets show that A2G-DiffRec is effective in enhancing item-side fairness at a marginal cost of accuracy reduction compared to existing guided diffusion recommenders and other non-diffusion baselines.
