The paper introduces the Denoising Diffusion Recommender Model (DDRM), a novel approach to enhance the robustness of recommender systems against noisy implicit feedback. DDRM leverages the multi-step denoising process of diffusion models to improve user and item embeddings from any recommender model. The key components of DDRM include: 1. **Forward Process**: Gaussian noises are injected into user and item embeddings, enhancing noise diversity. 2. **Reverse Denoising Process**: Iterative removal of noises is performed using a learnable neural network, guided by collaborative information. 3. **Inference**: A refined forward-reverse process is applied to a personalized starting point (average embeddings of historically liked items) to generate ideal items. The paper highlights the effectiveness of DDRM through extensive experiments on three datasets, demonstrating superior performance compared to other baselines. Key contributions include: - **Model-Agnostic Denoising**: DDRM can be integrated with any recommender model without modification. - **Denoising Modules**: Specialized modules for user and item denoising, incorporating collaborative information. - **Comprehensive Experiments**: Comprehensive evaluation across different datasets and settings, showing DDRM's robustness and efficiency. The paper also discusses the impact of various components and design variations, providing insights into the optimal hyper-parameters and architectural choices. Future work could focus on enhancing the denoising module and expanding DDRM to broader recommendation tasks.The paper introduces the Denoising Diffusion Recommender Model (DDRM), a novel approach to enhance the robustness of recommender systems against noisy implicit feedback. DDRM leverages the multi-step denoising process of diffusion models to improve user and item embeddings from any recommender model. The key components of DDRM include: 1. **Forward Process**: Gaussian noises are injected into user and item embeddings, enhancing noise diversity. 2. **Reverse Denoising Process**: Iterative removal of noises is performed using a learnable neural network, guided by collaborative information. 3. **Inference**: A refined forward-reverse process is applied to a personalized starting point (average embeddings of historically liked items) to generate ideal items. The paper highlights the effectiveness of DDRM through extensive experiments on three datasets, demonstrating superior performance compared to other baselines. Key contributions include: - **Model-Agnostic Denoising**: DDRM can be integrated with any recommender model without modification. - **Denoising Modules**: Specialized modules for user and item denoising, incorporating collaborative information. - **Comprehensive Experiments**: Comprehensive evaluation across different datasets and settings, showing DDRM's robustness and efficiency. The paper also discusses the impact of various components and design variations, providing insights into the optimal hyper-parameters and architectural choices. Future work could focus on enhancing the denoising module and expanding DDRM to broader recommendation tasks.