Adversarial Alignment and Disentanglement for Cross-Domain CTR Prediction with Domain-Encompassing Features

TL;DR

This paper tackles cross-domain recommendation by moving beyond sole reliance on domain-invariant features to also exploit valuable non-aligned signals. It introduces A^2DCDR, which combines inter-domain adversarial alignment with intra-domain mutual-information minimization and a feature reconstruction step, forming domain-encompassing representations that fuse both shared and domain-specific information. A novel Domain-Constrained MMD (DC-MMD) regularization aligns invariant features while enriching non-aligned signals, and Target-Aware Feature Combination (TAFC) adaptively fuses disentangled representations with contextual data for accurate predictions. Empirical results on four real-world cross-domain datasets and online A/B testing show consistent gains over state-of-the-art methods, demonstrating practical applicability in large-scale systems. The approach offers a principled framework for richer cross-domain information transfer and sets the stage for multi-domain extensions and causal robustness in future work.

Abstract

Cross-domain recommendation (CDR) has been increasingly explored to address data sparsity and cold-start issues. However, recent approaches typically disentangle domain-invariant features shared between source and target domains, as well as domain-specific features for each domain. However, they often rely solely on domain-invariant features combined with target domain-specific features, which can lead to suboptimal performance. To overcome the limitations, this paper presents the Adversarial Alignment and Disentanglement Cross-Domain Recommendation ($A^2DCDR$ ) model, an innovative approach designed to capture a comprehensive range of cross-domain information, including both domain-invariant and valuable non-aligned features. The $A^2DCDR$ model enhances cross-domain recommendation through three key components: refining MMD with adversarial training for better generalization, employing a feature disentangler and reconstruction mechanism for intra-domain disentanglement, and introducing a novel fused representation combining domain-invariant, non-aligned features with original contextual data. Experiments on real-world datasets and online A/B testing show that $A^2DCDR$ outperforms existing methods, confirming its effectiveness and practical applicability. The code is provided at https://github.com/youzi0925/A-2DCDR/tree/main.

Figures (3)

• Figure 1: Illustration of existing cross-domain recommendation (CDR) approaches: (a) Blended CDR methods, which blend all representations across domains including noise; (b) Disentangled CDR methods, which focus on separating domain-invariant and domain-specific information; and (c) The proposed $A^2DCDR$ framework, which captures both domain-invariant and valuable non-aligned features through adversarial training and feature reconstruction.
• Figure 2: The $A^2DCDR$ architecture has three main modules: Inter-Domain Adaptation, Intra-Domain Disentangler, and Fusion of Representations. In the Inter-Domain Adaptation module, the model aligns cross-domain invariant representations and uses adversarial learning to enhance source domain representations. Next, in the Intra-Domain Disentangler module, a mutual information minimizer is used to separate the domain-specific representation, while a feature reconstructor preserves information integrity during disentanglement. Finally, in the Fusion of Representations module, the model combines domain-invariant and valuable non-aligned representations with original contextual features to generate the final prediction.
• Figure 3: t-SNE visualization of disentangled representations in Sport-Cloth domain pair. Observations: (1) Clear separation between the green (source domain-encompassing features, $\mathbf{h}^A_t$) and yellow (target domain-specific features, $\mathbf{h}^B_s$) clusters; (2) Overlapping centroids of the red (target domain-invariant features, $\mathbf{h}^B_t$) and green distributions; (3) Red features nested within the green cluster boundaries.