Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Causality Enhancement for Cross-Domain Recommendation

Zhibo Wu, Yunfan Wu, Lin Jiang, Ping Yang, Yao Hu

TL;DR

This paper tackles data sparsity in recommender systems by introducing CE-CDR, a causality-enhanced cross-domain framework. It reformulates cross-domain recommendation as a causal graph and builds a causality-aware dataset via a Causality Labeling Module, then learns cross-domain causal representations with a Partially Labeled Causal Loss under a Direct Causality Modeling Module. The learned causal representations are adaptively fused into the target-domain model through a Causality Enhancement Module, enabling model-agnostic improvements. CE-CDR demonstrates strong offline gains across multiple real-world datasets and achieves production-level deployment with low latency, highlighting its practical impact for scalable, causality-aware cross-domain recommendations.

Abstract

Cross-domain recommendation forms a crucial component in recommendation systems. It leverages auxiliary information through source domain tasks or features to enhance target domain recommendations. However, incorporating inconsistent source domain tasks may result in insufficient cross-domain modeling or negative transfer. While incorporating source domain features without considering the underlying causal relationships may limit their contribution to final predictions. Thus, a natural idea is to directly train a cross-domain representation on a causality-labeled dataset from the source to target domain. Yet this direction has been rarely explored, as identifying unbiased real causal labels is highly challenging in real-world scenarios. In this work, we attempt to take a first step in this direction by proposing a causality-enhanced framework, named CE-CDR. Specifically, we first reformulate the cross-domain recommendation as a causal graph for principled guidance. We then construct a causality-aware dataset heuristically. Subsequently, we derive a theoretically unbiased Partial Label Causal Loss to generalize beyond the biased causality-aware dataset to unseen cross-domain patterns, yielding an enriched cross-domain representation, which is then fed into the target model to enhance target-domain recommendations. Theoretical and empirical analyses, as well as extensive experiments, demonstrate the rationality and effectiveness of CE-CDR and its general applicability as a model-agnostic plugin. Moreover, it has been deployed in production since April 2025, showing its practical value in real-world applications.

Causality Enhancement for Cross-Domain Recommendation

TL;DR

This paper tackles data sparsity in recommender systems by introducing CE-CDR, a causality-enhanced cross-domain framework. It reformulates cross-domain recommendation as a causal graph and builds a causality-aware dataset via a Causality Labeling Module, then learns cross-domain causal representations with a Partially Labeled Causal Loss under a Direct Causality Modeling Module. The learned causal representations are adaptively fused into the target-domain model through a Causality Enhancement Module, enabling model-agnostic improvements. CE-CDR demonstrates strong offline gains across multiple real-world datasets and achieves production-level deployment with low latency, highlighting its practical impact for scalable, causality-aware cross-domain recommendations.

Abstract

Cross-domain recommendation forms a crucial component in recommendation systems. It leverages auxiliary information through source domain tasks or features to enhance target domain recommendations. However, incorporating inconsistent source domain tasks may result in insufficient cross-domain modeling or negative transfer. While incorporating source domain features without considering the underlying causal relationships may limit their contribution to final predictions. Thus, a natural idea is to directly train a cross-domain representation on a causality-labeled dataset from the source to target domain. Yet this direction has been rarely explored, as identifying unbiased real causal labels is highly challenging in real-world scenarios. In this work, we attempt to take a first step in this direction by proposing a causality-enhanced framework, named CE-CDR. Specifically, we first reformulate the cross-domain recommendation as a causal graph for principled guidance. We then construct a causality-aware dataset heuristically. Subsequently, we derive a theoretically unbiased Partial Label Causal Loss to generalize beyond the biased causality-aware dataset to unseen cross-domain patterns, yielding an enriched cross-domain representation, which is then fed into the target model to enhance target-domain recommendations. Theoretical and empirical analyses, as well as extensive experiments, demonstrate the rationality and effectiveness of CE-CDR and its general applicability as a model-agnostic plugin. Moreover, it has been deployed in production since April 2025, showing its practical value in real-world applications.
Paper Structure (34 sections, 2 theorems, 17 equations, 7 figures, 4 tables)

This paper contains 34 sections, 2 theorems, 17 equations, 7 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Cross-Domain Recommendation
  4. Causal Recommendation
  5. Methodology
  6. Overview
  7. Problem Formulation
  8. Overall Architecture
  9. Causality Labeling Module
  10. Direct Causality Modeling Module
  11. Backbone model
  12. Partial Label Causal Loss
  13. Causality Enhancement Module
  14. Cross-Domain Self-Attention
  15. Cross-Domain Gated Attention
  16. ...and 19 more sections

Key Result

Proposition 1

(Partial Label Causal Loss). Given $s \perp\!\!\!\perp y \mid x$ and true propensity score $e(x)$, the causal model $f(x)$ is optimized by the following loss to learn a theoretically unbiased estimation of the true causal probability $p(y=1\mid x)$, with a partially labeled dataset $\mathcal{D}$: where $h(x) = \text{sg}\left[s+(1-s)\frac{f(x)(1-e(x))}{1-f(x)e(x)}\right]$, $\delta_1^f(x)=-\log(f(x

Figures (7)

  • Figure 1: Cross-domain causal illustration and causal graph.
  • Figure 2: Overall architecture of CE-CDR. (a) CLM: constructing causality-aware source-target behavior pairs; (b) DCMM: learning how source domain behaviors influence target domain preferences; (c) CEM: enhancing the target model for comprehensive user preference modeling.
  • Figure 3: Illustration of similarity calibration and fusion.
  • Figure 4: Analyzing how the causality score varies with used and excluded similarities during dataset construction.
  • Figure 5: Activation Difference of Cross-Domain Features.
  • ...and 2 more figures

Theorems & Definitions (2)

  • Proposition 1
  • Proposition 2