Causal-Invariant Cross-Domain Out-of-Distribution Recommendation

Jiajie Zhu, Yan Wang, Feng Zhu, Pengfei Ding, Hongyang Liu, Zhu Sun

Abstract

Cross-Domain Recommendation (CDR) aims to leverage knowledge from a relatively data-richer source domain to address the data sparsity problem in a relatively data-sparser target domain. While CDR methods need to address the distribution shifts between different domains (CDDS), they typically assume independent and identical distribution (IID) between training and testing data within the target domain. However, this IID assumption rarely holds in real-world scenarios due to single-domain distribution shift (SDDS). These co-existing distribution shifts lead to out-of-distribution (OOD) environments that hinder effective knowledge transfer and generalization. To address this, we propose CICDOR, a framework that learns dual-level causal structures to infer domain-specific and domain-shared causal-invariant user preferences. By integrating Large Language Models (LLMs) with conventional causal discovery, CICDOR extracts observed confounders for effective deconfounding, enabling accurate causal-invariant preference inference. Extensive experiments demonstrate the superior recommendation accuracy of CICDOR over state-of-the-art methods across various OOD scenarios.

Introduction: The Double-Shift Problem

Cross-Domain Recommendation (CDR) is a common strategy to help "data-poor" domains by borrowing knowledge from "data-rich" domains. However, most CDR models make a dangerous assumption: that the data distribution stays the same within the target domain (the IID assumption). In reality, two types of shifts often happen simultaneously:

  1. Cross-Domain Distribution Shift (CDDS): Fundamental differences in user behavior between domains (e.g., a user might seek "instant gratification" from a movie blockbuster but "delayed gratification" from a classic novel).
  2. Single-Domain Distribution Shift (SDDS): Shifts within a single domain due to temporal or regional factors (e.g., a user in Hong Kong might prefer international bestsellers, while a user in Beijing prefers Chinese classics).
Visualizing co-existing distribution shifts (CDDS and SDDS) in cross-domain recommendation.

When both shifts occur, recommendation accuracy plummets because models learn "spurious correlations" that fail in new, Out-of-Distribution (OOD) environments.

Methodology: The CICDOR Framework

The proposed CICDOR framework tackles these shifts by identifying stable, "causal-invariant" user preferences that remain accurate even when environments change. It utilizes a three-pronged approach:

  • Dual-Level Causal Preference Learning: The system learns two sets of causal structures (Directed Acyclic Graphs). One captures domain-shared preferences (universal traits) and the other captures domain-specific traits, allowing the model to transfer universal knowledge while respecting domain uniqueness.
  • LLM-Guided Confounder Discovery: CICDOR uses an LLM to propose potential causal factors from unstructured text (e.g., "price sensitivity"). To ensure statistical rigor, these factors are filtered through a formal Fast Causal Inference (FCI) algorithm and Markov Blanket extraction.
  • Causal Deconfounding: Once confounders are identified, the model uses backdoor adjustment to remove their influence, isolating the pure, invariant user preference.
The CICDOR architecture, integrating LLM-guided discovery with causal preference learning.

Experimental Results

CICDOR was evaluated on massive real-world datasets from Douban and Amazon. To test robustness, the authors created specific OOD scenarios involving User Degree Shift (testing on power users) and Region Shift (testing on specific city populations).

Key findings include:

  • Accuracy Boost: CICDOR achieved an average increase of 6.28% in Hit Ratio (HR@10) and 9.42% in NDCG@10 over state-of-the-art baselines.
  • Stability: As the intensity of the distribution shift increased, CICDOR’s performance degraded much more slowly than other models, proving its superior generalization.
  • Ablation Success: Removing either the dual-level learning or the LLM-confounder module led to significant performance drops, validating the necessity of both causal math and semantic understanding.

Conclusion

CICDOR demonstrates that while distributions change, the underlying causal mechanisms of user preference are often stable. By leveraging LLMs as "causal proxies" backed by statistical discovery, the framework can effectively model the "why" behind user choices. Future work will focus on addressing unobserved confounders and optimizing the computational complexity of the iterative LLM-FCI loop.

Summary updated

Abstract

Cross-Domain Recommendation (CDR) aims to leverage knowledge from a relatively data-richer source domain to address the data sparsity problem in a relatively data-sparser target domain. While CDR methods need to address the distribution shifts between different domains, i.e., cross-domain distribution shifts (CDDS), they typically assume independent and identical distribution (IID) between training and testing data within the target domain. However, this IID assumption rarely holds in real-world scenarios due to single-domain distribution shift (SDDS). The above two co-existing distribution shifts lead to out-of-distribution (OOD) environments that hinder effective knowledge transfer and generalization, ultimately degrading recommendation performance in CDR. To address these co-existing distribution shifts, we propose a novel Causal-Invariant Cross-Domain Out-of-distribution Recommendation framework, called CICDOR. In CICDOR, we first learn dual-level causal structures to infer domain-specific and domain-shared causal-invariant user preferences for tackling both CDDS and SDDS under OOD environments in CDR. Then, we propose an LLM-guided confounder discovery module that seamlessly integrates LLMs with a conventional causal discovery method to extract observed confounders for effective deconfounding, thereby enabling accurate causal-invariant preference inference. Extensive experiments on two real-world datasets demonstrate the superior recommendation accuracy of CICDOR over state-of-the-art methods across various OOD scenarios.
BibTeX
@misc{zhu2025causal,
  author = {Jiajie Zhu and Yan Wang and Feng Zhu and Pengfei Ding and Hongyang Liu and Zhu Sun},
  title = {{Causal-Invariant Cross-Domain Out-of-Distribution Recommendation}},
  year = {2025},
  eprint = {2505.16532},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2505.16532},
}
Figure