Source-linked AI summary
Denoising Diffusion Recommender Model
Jujia Zhao, Wenjie Wang, Yiyan Xu, Teng Sun, Fuli Feng, Tat-Seng Chua
TL;DR
Noisy implicit feedback can mislead models’ interpretation of user preferences. DDRM applies guided, multi-step diffusion to user and item embeddings, using personalized initialization at inference. Across three public datasets, it mostly outperforms backend models and other model-agnostic denoising methods.
Problem
Noisy implicit feedback can mislead recommender models’ interpretation of user preferences and hamper recommendation performance.
Source: §1 · p. 1Method
DDRM adds Gaussian noise to user and item embeddings, iteratively denoises them with collaborative guidance, and initializes inference from users’ historically liked items’ average embeddings.
Source: §1 · p. 2Results
DDRM mostly outperforms backend models and other model-agnostic denoising methods across all three datasets.
Source: §4.2 · p. 7Takeaways & Limitations
The reported cross-dataset results support DDRM’s effectiveness for robust representation learning in recommendation.
Source: §1 · p. 2, §4.2 · p. 7
Abstract
from arXiv · showhide
Recommender systems often grapple with noisy implicit feedback. Most studies alleviate the noise issues from data cleaning perspective such as data resampling and reweighting, but they are constrained by heuristic assumptions. Another denoising avenue is from model perspective, which proactively injects noises into user-item interactions and enhances the intrinsic denoising ability of models. However, this kind of denoising process poses significant challenges to the recommender model's representation capacity to capture noise patterns. To address this issue, we propose Denoising Diffusion Recommender Model (DDRM), which leverages multi-step denoising process of diffusion models to robustify user and item embeddings from any recommender models. DDRM injects controlled Gaussian noises in the forward process and iteratively removes noises in the reverse denoising process, thereby improving embedding robustness against noisy feedback. To achieve this target, the key lies in offering appropriate guidance to steer the reverse denoising process and providing a proper starting point to start the forward-reverse process during inference. In particular, we propose a dedicated denoising module that encodes collaborative information as denoising guidance. Besides, in the inference stage, DDRM utilizes the average embeddings of users' historically liked items as the starting point rather than using pure noise since pure noise lacks personalization, which increases the difficulty of the denoising process. Extensive experiments on three datasets with three representative backend recommender models demonstrate the effectiveness of DDRM.
1 INTRODUCTION
Noisy implicit feedback can mislead preference learning, while existing denoising approaches face limits from heuristic assumptions or the demands of direct noise-pattern modeling. DDRM uses guided, multi-step diffusion to robustify user and item embeddings, with personalized initialization at inference.
1 INTRODUCTION
- Noisy implicit feedback can mislead recommender models’ interpretation of user preferences and hamper recommendation performance.Implicit signals such as clicks may not reflect users’ actual satisfaction.Source: §1 · p. 1
- Data-cleaning methods such as resampling and reweighting rely on heuristic assumptions and require tuning to adapt to different backends and datasets.Examples include assuming large-loss interactions are noisy or relying on cross-model agreement.Source: §1 · pp. 1–2
- Direct model-perspective denoising can demand substantial representation capacity to capture noise patterns while converting noisy data into clean data.These methods may add random interaction noise or drop positive interactions during training.Source: §1 · p. 2
- DDRM is a model-agnostic plug-in that adds adjustable Gaussian noise to user and item embeddings, then iteratively removes it through a learnable reverse process.It is designed for embeddings from existing recommender models.Source: §1 · p. 2
- DDRM encodes collaborative information as reverse-process guidance and initializes inference with the average embeddings of a user’s historically liked items rather than pure noise.It also rounds the generated item embedding to an existing candidate using embedding similarity.Source: §1 · p. 2
- The authors instantiate DDRM on three backend models and report experiments across three public datasets.Source: §1 · p. 2
2 PRELIMINARY
Diffusion models use a forward process to add Gaussian noise and a reverse process to iteratively denoise data. Training learns the reverse distribution by optimizing a variational lower bound.
2 PRELIMINARY
- The forward process adds Gaussian noise to a data sample over T steps, with β_t setting the noise scale at step t.The text describes x_t as sampled from a Gaussian distribution and notes that it can be obtained directly from x_0.Source: §2 · p. 3
- The reverse process iteratively maps noisy data from x_T back toward x_0, using a neural network to predict the reverse Gaussian distribution.The text states this formulation under Gaussian forward transitions and sufficiently small β_t.Source: §2 · p. 3
- Training optimizes a variational lower bound for the predictive denoising distribution, aligning reverse-step distributions with tractable posteriors.The forward-process term L_T can be disregarded because it has no learnable parameters.Source: §2 · p. 3
3 METHOD
DDRM denoises pretrained user and item embeddings through a forward-reverse diffusion process, using collaborative guidance and a personalized inference starting point. Its training combines embedding reconstruction with recommendation loss, and inference ranks candidate items using generated embeddings.
3 METHOD
- DDRM applies forward Gaussian-noise injection and iterative reverse denoising to pretrained user and item embeddings, with specialized modules guiding denoising and inference.The framework applies the refined process to a selected starting point to generate an item aligned with user preferences.Source: §3 · p. 3
3.1 DDRM Framework
- A linear variance schedule regulates forward noise, while separate user and item modules iteratively denoise embeddings using collaborative and step information to guide reconstruction.The modules use reconstruction MLPs to predict original embeddings; step information indicates the current noise level.Source: §3.1 · pp. 3–4
3.2 Optimization
- Training combines embedding-reconstruction loss with BPR recommendation loss; DDRM denoises the user and positive item, and optional reweighting lowers weights for lower positive-score instances.The loss balance factor adjusts the relative weights of reconstruction and BPR losses.Source: §3.2 · pp. 4–5
3.3 Inference
- Inference starts from the average embeddings of a user's historically interacted items, applies forward-reverse denoising conditioned on the user embedding, and ranks candidates by inner-product similarity.The procedure selects the top-k candidate items by similarity score.Source: §3.3 · p. 5
4 EXPERIMENTS
Experiments compare DDRM with baselines across settings, test component contributions, and examine design variations.
4.1 Experimental Settings
- Experiments use three public datasets, chronological 7:1:2 positive-interaction splits, noisy training and validation data, clean-positive testing, and Recall/NDCG@10/20.Natural-noise settings add false positives; random-noise settings add sampled unobserved interactions.Source: §4.1.1 · pp. 5–6, §4.1.2 · p. 6
4.2 Overall Performance (RQ1)
- Across three datasets, DDRM mostly outperforms backend models and model-agnostic denoisers under natural noise.It also brings suboptimal backends to performance comparable with or above DiffRec.Source: Table 2 · p. 6, §4.2 · p. 7
4.3 In-depth Analysis
- With noise ratios from 0% to 60%, performance declines overall, while DDRM consistently beats the backend and R-CE across noise levels.Similar trends were reported for Amazon-book and ML-1M, though their figures were omitted.Source: Figure 3 · p. 7, §4.3.1 · p. 7
- Reconstruction loss consistently beats the backend, whereas reweighted loss varies across datasets and backends and generally contributes less.In some cases, reweighted loss falls below backend performance.Source: Figure 4 · p. 7, §4.3.2 · p. 8
- Item-only denoising still outperforms the backend but trails full DDRM, supporting the contribution of denoising both user and item representations.User-only denoising was not feasible because inference requires denoising item embeddings to generate recommendations.Source: Figure 5 · p. 7, §4.3.2 · p. 8
- Optimal inference performance occurs below 100 diffusion steps; excessive steps or noise scale can impair personalization, so key hyperparameters need careful selection.Fewer steps also reduce computation and time costs.Source: Figure 6 · p. 8, §4.3.4 · p. 9, §4.3.5 · p. 9
- The Transformer architecture underperforms the MLP, and replacing personalized initialization with pure noise substantially reduces performance.The authors attribute the Transformer result to an unsuitable inductive bias and the pure-noise result to lost personalization.Source: Table 3 · p. 8, §4.3.3 · p. 8
- The linear variance schedule outperforms linear, cosine, and binomial alternatives by injecting noise more gradually while preserving personalization.The authors report improved training stability relative to cosine and binomial schedules.Source: §4.3.3 · p. 8
5 RELATED WORK
Prior denoising work addresses noisy implicit feedback through data cleaning or model-side augmentation, while diffusion-based recommendation has focused on other recommendation goals. DDRM applies diffusion to denoise naturally noisy implicit feedback.
5 RELATED WORK
- Data-cleaning methods use assumptions to remove or downweight noisy interactions, limiting adaptability across backend models and datasets.Model-perspective approaches augment interactions or corrupt them for reconstruction, but may require neural networks to capture complex real-world noise distributions.Source: §5 · p. 9
- Diffusion recommenders have modeled interaction probabilities, generated content, or addressed data sparsity, with several embedding-level methods focused on sequential recommendation.The cited methods include CODIGEM and DiffRec for interaction probabilities, DiffuRec and CDDRec for sequential recommendation, and DiffuASR for item sequences.Source: §5 · p. 9
- DDRM differs by using diffusion to denoise implicit feedback and support more robust representation learning.The related work contrasts this focus with diffusion methods that overlook natural false-positive interactions.Source: §5 · p. 9
6 CONCLUSION AND FUTURE WORK
DDRM is a plug-in model that applies a forward-reverse diffusion process to user and item embeddings, using collaborative guidance and personalized inference initialization. Experiments report superiority over competitive baselines, while future work may expand its denoising module and task coverage.
6 CONCLUSION AND FUTURE WORK
- DDRM injects Gaussian noise into existing recommender embeddings and iteratively removes it to bolster robust representation learning amid noisy feedback.A denoising module encodes collaborative information as guidance; inference starts from average embeddings of historically liked items and grounds the generated embedding to item candidates.Source: §6 · p. 9
- Experiments report DDRM's superiority over other competitive baselines.Source: §6 · p. 9
- Future work may enrich the denoising module with neural networks and adapt DDRM to sequential or bundle recommendations.Source: §6 · p. 9