# Denoising Diffusion Recommender Model

**Authors:** Jujia Zhao, Wenjie Wang, Yiyan Xu, Teng Sun, Fuli Feng, Tat-Seng Chua  
**arXiv:** [2401.06982](https://arxiv.org/abs/2401.06982)  
**Date:** 2024-01-13  
**Summary page:** https://www.summarizepaper.com/en/paper/2401.06982v4/

## Key points

- Recommender systems face challenges with noisy implicit feedback
- Data cleaning techniques such as resampling and reweighting are commonly used to address these challenges, but they have limitations due to heuristic assumptions
- The Denoising Diffusion Recommender Model (DDRM) is proposed to overcome these limitations
- DDRM utilizes diffusion models to strengthen user and item embeddings from any recommender model
- Controlled Gaussian noises are introduced in the forward process and iteratively removed in the reverse denoising process to enhance embedding robustness against noisy feedback
- Guidance for the reverse denoising process is provided by a dedicated denoising module that encodes collaborative information
- In the inference stage, DDRM uses average embeddings of users' historically liked items as a personalized starting point instead of pure noise
- Extensive experiments on three datasets with three backend recommender models demonstrate the effectiveness of DDRM in improving recommendation performance
- Various design variations were explored to validate DDRM's performance across different settings

## Summary

Recommender systems face challenges with noisy implicit feedback, which are often addressed through data cleaning techniques such as resampling and reweighting. However, these methods have limitations due to their heuristic assumptions. To overcome this issue, the Denoising Diffusion Recommender Model (DDRM) is proposed. <br>
 of diffusion models is utilized by DDRM to strengthen user and item embeddings from any recommender model. By introducing controlled Gaussian noises in the forward process and iteratively removing them in the reverse denoising process, DDRM enhances embedding robustness against noisy feedback. <br>
Key to this approach is providing guidance for the reverse denoising process and selecting an appropriate starting point for inference. A dedicated denoising module encodes collaborative information as guidance while in the inference stage, DDRM uses average embeddings of users' historically liked items as a personalized starting point instead of pure noise. <br>
Extensive experiments on three datasets with three backend recommender models demonstrate the effectiveness of DDRM in improving recommendation performance. Additionally, various design variations were explored to validate DDRM's performance across different settings as depicted in Figure 6.