From Noise to Order: Learning to Rank via Denoising Diffusion

In information retrieval (IR), learning-to-rank (LTR) methods have traditionally limited themselves to discriminative machine learning approaches that model the probability of the document being relevant to the query given some feature representation of the query-document pair. In this work, we propose an alternative denoising diffusion-based deep generative approach to LTR that instead models the full joint distribution over feature vectors and relevance labels. While in the discriminative setting, an over-parameterized ranking model may find different ways to fit the training data, we hypothesize that candidate solutions that can explain the full data distribution under the generative setting produce more robust ranking models. With this motivation, we propose DiffusionRank that extends TabDiff, an existing denoising diffusion-based generative model for tabular datasets, to create generative equivalents of classical discriminative pointwise and pairwise LTR objectives. Our empirical results demonstrate significant improvements from DiffusionRank models over their discriminative counterparts. Our work points to a rich space for future research exploration on how we can leverage ongoing advancements in deep generative modeling approaches, such as diffusion, for learning-to-rank in IR.

💡 Research Summary

Learning‑to‑rank (LTR) has long been dominated by discriminative approaches that directly model the conditional probability of relevance given a query‑document feature vector. While effective, such models can over‑parameterize and fit the training data in many different ways, without guaranteeing that they capture the true underlying data distribution. In “From Noise to Order: Learning to Rank via Denoising Diffusion,” the authors propose a fundamentally different paradigm: a generative, diffusion‑based LTR model called DiffusionRank that learns the joint distribution of features and relevance labels.

The paper builds on Tabular Diffusion (TabDiff), a continuous‑time denoising diffusion model designed for mixed‑type tabular data. DiffusionRank adapts TabDiff’s forward process to LTR datasets by adding Gaussian noise to continuous ranking features (e.g., BM25 scores, PageRank) and applying a masked diffusion scheme to categorical variables, which include the relevance label (either a pointwise grade or a pairwise preference). The reverse process is parameterized by a lightweight feed‑forward network that predicts both the original continuous features and the original categorical label at each diffusion timestep. By using a single‑step unmasking for categorical variables, inference requires only one pass through the network, keeping the runtime comparable to traditional discriminative rankers.

Crucially, the authors translate classic pointwise and pairwise LTR objectives into generative loss functions. The pointwise loss mirrors cross‑entropy but is computed on the denoised label distribution produced by the diffusion model, encouraging the model to capture the full data distribution while still optimizing for accurate label prediction. The pairwise loss adopts the RankNet logistic formulation, with the diffusion model learning to reconstruct the preference ordering between document pairs during denoising. This design allows direct comparison with existing discriminative baselines while leveraging the regularizing effect of generative training.

Empirical evaluation is performed on standard LTR benchmarks such as LETOR, MSLR‑Web30, and Yahoo! LTR. Across both pointwise and pairwise settings, DiffusionRank consistently outperforms strong discriminative baselines (e.g., neural rankers, boosted decision trees) on NDCG@10, MAP, and Precision@k. Improvements range from 2 % to 4 % relative gain, with the most pronounced benefits observed in scenarios with label imbalance or limited training data, where the generative model’s ability to model the full joint distribution mitigates over‑fitting. The authors also report that the parameter count of DiffusionRank is comparable to, or slightly lower than, its discriminative counterparts, and that inference latency remains on par because the masked diffusion requires only a single denoising step.

Beyond the immediate results, the paper outlines several promising research avenues. Because diffusion models can be conditioned on arbitrary modalities, DiffusionRank could be extended to incorporate textual embeddings from large language models, visual features, or user interaction signals, enabling truly multimodal ranking. The generative framework also opens the door to privacy‑preserving synthetic data generation for LTR, data augmentation to reduce labeling costs, and the integration of fairness or diversity constraints directly into the diffusion process. In summary, the work demonstrates that denoising diffusion models are not only viable for tabular data but also provide a powerful new tool for learning‑to‑rank, offering robustness, flexibility, and competitive performance compared to traditional discriminative methods.