Conditional Information Bottleneck Approach for Out-of-Distribution Sequential Recommendation

Binquan Wu, Wenxuan Shen, Wenzheng Xie, Junhao Zheng, Yu Cheng, Qianli Ma

Keywords: Sequential Recommendation, Out of Distribution, Robust, Conditional Information Bottleneck

TL;DR: We propose a Conditional Information Bottleneck-guided approach for diverse and reliable OOD exploration, enhancing the robustness of sequential recommendation.

Abstract:

Sequential recommendation (SR) aims to suggest items users are most likely to engage with next based on their past interactions. However, in practice, SR systems often face the out-of-distribution (OOD) problem due to dynamic environmental factors (e.g., seasonal changes), leading to significant performance degradation in the testing phase.
Some methods incorporate distributionally robust optimization (DRO) into SR to alleviate OOD, but the sparsity of SR data challenges this. Other approaches use random data augmentations to explore the OOD, potentially distorting important information, as user behavior is personalized rather than random. Additionally, they often overlook users' varying sensitivity to distribution shifts during the exploration, which is crucial for capturing the evolution of user preferences in OOD contexts.
In this work, inspired by information bottleneck theory (IB), we propose the Conditional Distribution Information Bottleneck (CDIB), a novel objective that creates diverse OOD distributions while preserving minimal sufficient information regarding the origin distribution conditioned on the user. Building on this, we introduce a framework with a learnable, personalized data augmentation method using a mask-then-generate paradigm to craft diverse and reliable OOD distributions optimized with CDIB. Experiments on four real-world datasets show our model consistently outperforms baselines. The code is available at https://anonymous.4open.science/r/CDIB-51C8.

Primary Area: other topics in machine learning (i.e., none of the above)

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 11571