Brownian Bridge Diffusion for Sequential Recommendation
Fuente:
arXiv
Gespeichert in:
| Hauptverfasser: | , , , , , , , |
|---|---|
| Format: | Preprint |
| Veröffentlicht: |
2025
|
| Schlagworte: | |
| Online-Zugang: | |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| _version_ | 1866915999411863552 |
|---|---|
| author | Bai, Yimeng Zhang, Yang Ding, Sihao Ruan, Shaohui Yao, Han Guan, Danhui Feng, Fuli Chua, Tat-Seng |
| author_facet | Bai, Yimeng Zhang, Yang Ding, Sihao Ruan, Shaohui Yao, Han Guan, Danhui Feng, Fuli Chua, Tat-Seng |
| contents | Diffusion models, known for their strong generative capability derived from iterative noising and denoising processes, have recently emerged as a promising paradigm for sequential recommendation. To incorporate user history for personalization, existing methods typically follow a history-guided denoising paradigm inspired by text-guided image generation, where target item representations are reconstructed from Gaussian noise conditioned on user historical interactions. However, this design remains fundamentally anchored to an "item $\leftrightarrow$ noise" formulation, introducing an additional noise-reconstruction burden that may distract the model from capturing user-specific preference structures. Motivated by this limitation, we revisit diffusion-based sequential recommendation from a preference-centric perspective and adopt a preference bridging design that enables a direct "item $\leftrightarrow$ history" transition instead of relying on Gaussian noise. Based on this idea, we propose Brownian Bridge Diffusion Recommendation (BBDRec), which leverages the Brownian bridge process to construct a structured diffusion trajectory between target items and user historical representations, thereby better aligning diffusion modeling with the intrinsic nature of recommendation. Extensive experiments on multiple public datasets show that BBDRec consistently outperforms representative sequential and diffusion-based recommendation baselines. The implementation code is publicly available at https://github.com/baiyimeng/BBDRec. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2507_06121 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Brownian Bridge Diffusion for Sequential Recommendation Bai, Yimeng Zhang, Yang Ding, Sihao Ruan, Shaohui Yao, Han Guan, Danhui Feng, Fuli Chua, Tat-Seng Information Retrieval Diffusion models, known for their strong generative capability derived from iterative noising and denoising processes, have recently emerged as a promising paradigm for sequential recommendation. To incorporate user history for personalization, existing methods typically follow a history-guided denoising paradigm inspired by text-guided image generation, where target item representations are reconstructed from Gaussian noise conditioned on user historical interactions. However, this design remains fundamentally anchored to an "item $\leftrightarrow$ noise" formulation, introducing an additional noise-reconstruction burden that may distract the model from capturing user-specific preference structures. Motivated by this limitation, we revisit diffusion-based sequential recommendation from a preference-centric perspective and adopt a preference bridging design that enables a direct "item $\leftrightarrow$ history" transition instead of relying on Gaussian noise. Based on this idea, we propose Brownian Bridge Diffusion Recommendation (BBDRec), which leverages the Brownian bridge process to construct a structured diffusion trajectory between target items and user historical representations, thereby better aligning diffusion modeling with the intrinsic nature of recommendation. Extensive experiments on multiple public datasets show that BBDRec consistently outperforms representative sequential and diffusion-based recommendation baselines. The implementation code is publicly available at https://github.com/baiyimeng/BBDRec. |
| title | Brownian Bridge Diffusion for Sequential Recommendation |
| topic | Information Retrieval |
| url | https://arxiv.org/abs/2507.06121 |