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

📝 Paper Summary

Cross-Domain Recommendation (CDR) Out-of-Distribution (OOD) Generalization Causal Recommendation

CICDOR tackles both cross-domain and single-domain distribution shifts by learning dual-level invariant causal structures and utilizing LLMs to extract and deconfound observed variables from user reviews.

Core Problem

Standard Cross-Domain Recommendation assumes target domain data is identically distributed, failing to account for the simultaneous presence of Cross-Domain Distribution Shift (CDDS) and Single-Domain Distribution Shift (SDDS).

Why it matters:

Real-world systems face complex OOD environments where user preferences shift across domains (e.g., movies vs. books) and within domains (e.g., regional differences), degrading generalization.
Existing methods typically address either CDDS or SDDS in isolation, leading to information loss and error accumulation when shifts co-exist.
Prior causal approaches often overlook observed confounders in unstructured text or rely only on unobserved environmental factors.

Concrete Example: Consider a system recommending books to users in Beijing based on Movie domain data. Users in Beijing prefer Chinese classics (SDDS: regional shift vs. Hong Kong users), while book readers generally prefer delayed gratification unlike movie watchers (CDDS). A model must transfer invariant preferences (e.g., 'heroic adventure') that hold true despite both shifts, which current isolated methods fail to capture effectively.

Key Novelty

Causal-Invariant Cross-Domain Out-of-distribution Recommendation (CICDOR)

Proposes a dual-level causal preference learning module that disentangles user preferences into domain-specific and domain-shared components, learning separate causal structures for each to handle both CDDS and SDDS.
Introduces an LLM-guided confounder discovery module that uses Large Language Models to extract candidate causal variables from reviews, refined by the FCI algorithm to identify observed confounders.

Architecture

Conceptual illustration of Cross-Domain Distribution Shift (CDDS) and Single-Domain Distribution Shift (SDDS) using a Movie-to-Book recommendation scenario.

Evaluation Highlights

Achieves an average increase of 6.28% in HR@10 compared to the best-performing state-of-the-art baseline across various OOD scenarios.
Achieves an average increase of 9.42% in NDCG@10 compared to the best-performing state-of-the-art baseline across various OOD scenarios.

Breakthrough Assessment

7/10

Novel integration of LLMs for explicit confounder discovery in causal recommendation. Addresses a realistic dual-shift problem (CDDS + SDDS) often treated separately. Quantitative gains are significant, though full experimental details are missing from the text.

⚙️ Technical Details

Problem Definition

Setting: Cross-Domain Out-of-Distribution Recommendation

Inputs: Source domain data (users, items, interactions, reviews) and Target domain training data

Outputs: Interaction predictions in Target domain testing data, where the data distribution differs from the training distribution due to CDDS and SDDS

Pipeline Flow

LLM-Guided Confounder Discovery: Extract variables from reviews -> Causal Discovery (FCI) -> Confounder Pool
Dual-Level Causal Learning: Disentangle preferences -> Learn Domain-Specific/Shared Causal Structures -> Inference

System Modules

User Preference Disentanglement

Extract and separate user preferences into domain-specific (unique to source/target) and domain-shared (invariant across domains) components.

Model or implementation: Not detailed in provided text

LLM-Guided Confounder Discovery

Identify observed confounders from unstructured text. Uses LLM to extract variables and FCI algorithm to verify causal links.

Model or implementation: Large Language Model (specific variant not reported in text)

Dual-Level Causal Structure Learning

Learn DAGs for both domain-specific and domain-shared levels to infer causal-invariant preferences.

Model or implementation: Causal Structure Learning algorithm (e.g., based on DAGs)

Novel Architectural Elements

Dual-level causal structure modeling that explicitly separates and models domain-shared vs. domain-specific causal graphs.
Integration of LLM-based variable extraction with formal statistical causal discovery (FCI algorithm) for confounder identification.

Modeling

Base Model: Not reported in the provided text

Training Method: Not reported in the provided text

Compute: Not reported in the provided text

Comparison to Prior Work

vs. DICE/DCCL: These treat OOD as a debiasing problem for specific factors (conformity) and only handle SDDS, whereas CICDOR handles both SDDS and CDDS via generalized causal structures.
vs. InvPref: Addresses unobserved confounders (environments) but overlooks observed confounders in text; CICDOR explicitly extracts observed confounders using LLMs.
vs. Conventional CDR (e.g., alignment methods): These assume IID within the target domain and fail to address SDDS.

Limitations

Relies on the quality of LLM extraction; inaccurate variable extraction may propagate errors to the causal graph.
Experimental details (datasets, baselines tables) are not present in the provided text snippet, preventing verification of specific numeric claims beyond summary statistics.
Requires user reviews to be available for confounder extraction.

Reproducibility

No replication artifacts mentioned in the paper. Code URL not provided in the text snippet.

📊 Experiments & Results

Evaluation Setup

Cross-domain recommendation under Out-of-Distribution (OOD) settings involving both CDDS and SDDS.

Benchmarks:

Two real-world datasets (Cross-Domain Recommendation)

Metrics:

HR@10 (Hit Ratio)
NDCG@10 (Normalized Discounted Cumulative Gain)
Statistical methodology: Not explicitly reported in the paper

Main Takeaways

CICDOR outperforms state-of-the-art baselines by a significant margin (avg +6.28% HR@10, +9.42% NDCG@10), validating the importance of handling both CDDS and SDDS.
The dual-level causal structure effectively captures invariant preferences that transfer across domains even when distributions shift within the target domain.
LLM-guided confounder discovery contributes to performance by identifying and controlling for observed confounding variables in user reviews.

📚 Prerequisite Knowledge

Prerequisites

Cross-Domain Recommendation (CDR)
Causal Inference (Confounders, Backdoor Adjustment)
Directed Acyclic Graphs (DAGs)

Key Terms

CDDS: Cross-Domain Distribution Shift—differences in data distribution between the source and target domains (e.g., user behavior differences between Movie and Book platforms).

SDDS: Single-Domain Distribution Shift—differences in data distribution within a single domain due to factors like time or region (e.g., users in Beijing vs. Hong Kong).

IID: Independent and Identically Distributed—a standard statistical assumption that training and testing data come from the same probability distribution, which CICDOR challenges.

OOD: Out-of-Distribution—scenarios where the testing data distribution differs from the training data distribution.

Confounder: A variable that influences both the treatment (user preference) and the outcome (interaction), causing spurious correlations if not controlled.

FCI algorithm: Fast Causal Inference—a constraint-based causal discovery algorithm used to learn causal structures (DAGs) from data, capable of handling latent variables.

Backdoor Adjustment: A causal inference technique to estimate the causal effect of a variable by adjusting for confounding factors.

DAG: Directed Acyclic Graph—a graphical representation of causal relationships where edges represent influence and no loops exist.