DiffuRec: A Diffusion Model for Sequential Recommendation

Zihao Li,Aixin Sun,Chenliang Li

2023-10-30

Abstract:Mainstream solutions to Sequential Recommendation (SR) represent items with fixed vectors. These vectors have limited capability in capturing items' latent aspects and users' diverse preferences. As a new generative paradigm, Diffusion models have achieved excellent performance in areas like computer vision and natural language processing. To our understanding, its unique merit in representation generation well fits the problem setting of sequential recommendation. In this paper, we make the very first attempt to adapt Diffusion model to SR and propose DiffuRec, for item representation construction and uncertainty injection. Rather than modeling item representations as fixed vectors, we represent them as distributions in DiffuRec, which reflect user's multiple interests and item's various aspects adaptively. In diffusion phase, DiffuRec corrupts the target item embedding into a Gaussian distribution via noise adding, which is further applied for sequential item distribution representation generation and uncertainty injection. Afterward, the item representation is fed into an Approximator for target item representation reconstruction. In reverse phase, based on user's historical interaction behaviors, we reverse a Gaussian noise into the target item representation, then apply a rounding operation for target item prediction. Experiments over four datasets show that DiffuRec outperforms strong baselines by a large margin.

Information Retrieval

What problem does this paper attempt to address?

### Problems the Paper Attempts to Solve

This paper aims to address several key issues in Sequential Recommendation (SR):

1. **Multiple Latent Aspects of Items**: In real-world scenarios, items may contain different latent aspects. For example, in the movie recommendation scenario, movies have different themes or categories, and different movies will contain different theme distributions. Encoding these complex latent aspects into a single vector is challenging.

2. **Multiple Interests of Users**: Users' interests and preferences are diverse and unstable, and they may dynamically adjust according to different scenarios or periods. Therefore, users' attention to various latent aspects of items also changes. Static item representations are insufficient to capture these changes.

3. **Uncertainty**: Due to the diversity and evolution of users' interests, there is a certain degree of uncertainty in users' current preferences and intentions. Recommendation systems are also expected to reflect the diversity, novelty, and surprise of recommended items. Therefore, modeling users' current interests as a distribution is more appropriate.

4. **Target Item Guidance**: Finally, the target item can assist in understanding the user's current intention and serve as an auxiliary signal to guide the subsequent process. However, interactions involving the target item may bring significant computational costs and deep coupling in network design, so existing methods are only applicable at the ranking stage.

To address these issues, the authors propose a diffusion model-based approach—DiffuRec—for sequential recommendation tasks. By gradually adding Gaussian noise to the target item embeddings and using these noisy representations to generate distributional representations of historical items, the method captures multiple latent aspects of items and multiple interests of users. This approach not only enhances the model's representation capability but also introduces uncertainty to improve the model's robustness. Experimental results show that DiffuRec significantly outperforms nine strong baseline methods on four real-world datasets.