From Noise to Order: Learning to Rank via Denoising Diffusion

The article introduces DiffusionRank, a novel learning-to-rank (LTR) method utilizing a denoising diffusion-based generative approach, which models the full joint distribution over features and relevance labels. This model builds on TabDiff, enhancing traditional discriminative LTR objectives with generative alternatives, and demonstrates improved performance on four standard LTR datasets. This development is significant for practitioners as it opens new avenues for leveraging deep generative models in ranking tasks, potentially leading to more accurate relevance estimations.