1 Title

        Diffusion Recommender Model（Wenjie Wang、Yiyan Xu、Fuli Feng、Xinyu Lin、Xiangnan He、Tat-Seng Chua）【SIGIR '23】

2 Conclusion

        . In light of the impressive advantages of Diffusion Models (DMs) over traditional generative models in image synthesis, this paper proposes a novel Diffusion Recommender Model (named DiffRec) to learn the generative process in a denoising manner. In addition, it extend traditional
DMs to tackle the unique challenges in recommendation: high resource costs for large-scale item prediction and temporal shifts of user preference.

3 Good Sentences

        1、Nevertheless, we cannot directly graft the forward process in the image domain due
to the necessity of generating personalized recommendations. To retain personalized information in users’ corrupted interactions,we should avoid corrupting users’ interaction histories into pure
noises like in image synthesis.（The differences between diffusion model and Diffusion Recommender Model）
        2、Generative models, such as MultiVAE and DiffRec, predict the interaction probabilities 𝒙ˆ0 over all items simultaneously, requiring extensive resources and limiting large-scale item prediction in industry. To reduce the costs, we offer L-DiffRec, which clusters items for dimension compression via multiple VAEs and conducts diffusion processes in the latent space.（The shortcomings of previous work and improvement of this study）
        3、Since user preference might shift over time, it is crucial to capture temporal information during DiffRec learning. Assuming that more recent interactions can better represent users’ current preferences, we propose a time-aware reweighting strategy to assign larger weights to users’ later interactions.（The driving force behind this improvement）

---

本文提出了一种扩散推荐模型，以及它的两个扩展，用latent space的L-DiffuRec和引入了时间戳的T-DiffuRec。

DiffRec 主要由两部分组成：对于给定的用户历史交互
(1) 前向过程加入高斯噪声逐步破坏交互信息

值得注意的是与原始diffusion不同，本文设计了一个新的noise schedule，不过与原始相同，前向过程是没有参数的。

(2) 反向过程中模型逐步去噪并恢复原始信息。

        与图像生成任务不同，为保证用户的个性化信息，DRM在训练时并没有将用户交互破坏为纯噪声，并且在训练和推断时均减少了前向过程中添加的噪声。

利用VAE学习latent embedding，在latent embedding上进行diffusion