Abstract: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.

What problem does this paper attempt to address?

This paper attempts to address the issue of implicit feedback noise in recommender systems. Specifically, recommender systems often encounter implicit feedback data (such as clicks, purchases, etc.) that contain noise, which can mislead the recommendation model's understanding of user preferences, thereby affecting recommendation performance. To tackle this challenge, the paper proposes a new method called the Denoising Diffusion Recommender Model (DDRM). ### Main Issues 1. **Noise in Implicit Feedback**: - Implicit feedback data (such as clicks, purchases) easily contain noise, which may come from various interfering factors (e.g., unconscious clicks by users). - Noisy data can mislead the recommendation model, resulting in poor recommendation performance. 2. **Limitations of Existing Methods**: - **Data Cleaning Methods**: These methods reduce the impact of noise through resampling or reweighting, but they rely on heuristic assumptions and have limited adaptability. - **Denoising Methods from a Model Perspective**: These methods enhance the model's noise resistance by injecting random noise into user-item interactions, but they require high representational capacity from the model. ### Solution The paper proposes a denoising method based on a diffusion model—Denoising Diffusion Recommender Model (DDRM), which aims to enhance the embedding representations of users and items through a multi-step denoising process, making them more robust to noisy feedback. ### Key Techniques 1. **Forward Process**: - Actively inject Gaussian noise into the embedding representations of users and items to generate noisy embeddings. - Gradually increase the noise through adjustable noise scales and steps. 2. **Reverse Denoising Process**: - Gradually remove the noise through a learnable neural network to restore clean embeddings. - Design specialized user denoising modules and item denoising modules to handle different types of noise distributions. 3. **Inference Stage**: - Use the average embedding of items liked by the user historically as the starting point, rather than pure noise, to add personalized information. - Guide the denoising process through collaborative information, making the generated ideal items more aligned with user preferences. ### Experimental Validation The paper conducts extensive experiments on three public datasets to validate the effectiveness of DDRM. The experimental results show that DDRM significantly outperforms other baseline methods on multiple metrics, especially under natural noise settings. ### Main Contributions 1. Propose a model-agnostic denoising diffusion recommender model that enhances the robustness of user and item representations to noisy feedback for any recommendation model. 2. Design user and item denoising modules that guide the denoising process through collaborative information and integrate personalized information as the starting point during the inference stage. 3. Conduct extensive experiments on three backend models to validate the effectiveness of DDRM on multiple public datasets. Through these techniques, DDRM can effectively address the noise issue in implicit feedback in recommender systems, improving the performance of recommender systems.

Denoising Diffusion Recommender Model

Federated Recommender System Based on Diffusion Augmentation and Guided Denoising

Diffusion Recommender Model

RecDiff: Diffusion Model for Social Recommendation

Data Collaborative Contrastive Recommendation Model with Self-Adaptive Noise

Multi-Step Denoising Scheduled Sampling: Towards Alleviating Exposure Bias for Diffusion Models

Learning Robust Recommender from Noisy Implicit Feedback

Denoising Implicit Feedback for Recommendation

Self-Guided Learning to Denoise for Robust Recommendation

Probabilistic and Variational Recommendation Denoising

Learning Robust Recommenders through Cross-Model Agreement

SDDRS: Stacked Discriminative Denoising Auto-Encoder Based Recommender System.

Conditional Denoising Diffusion for Sequential Recommendation

Residual Denoising Diffusion Models

Discrete Conditional Diffusion for Reranking in Recommendation

Denoising User-aware Memory Network for Recommendation

Recommendation Via Collaborative Diffusion Generative Model

EDiffuRec: An Enhanced Diffusion Model for Sequential Recommendation

Dual-domain Collaborative Denoising for Social Recommendation

Learning to Denoise Unreliable Interactions for Graph Collaborative Filtering

Denoising Multi-modal Sequential Recommenders with Contrastive Learning