Cross-Domain Recommendation

Updated 2 May 2026

Cross-Domain Recommendation is a strategy that transfers user-item interactions from a data-rich source to a sparse target domain, mitigating cold-start challenges.
Key methodologies include embedding mapping, deep/meta-learning, diffusion models, and graph-based integration to effectively bridge domain gaps.
Empirical deployments demonstrate performance gains of 8–15% in metrics such as HR@10 and NDCG@10, while addressing issues like negative transfer and distribution shifts.

Cross-Domain Recommendation (CDR) addresses the fundamental challenge of data sparsity and cold-start in recommender systems by leveraging auxiliary information from related domains. In CDR, user–item interaction signals from one or more rich source domains are transferred to bolster performance in a sparse or emerging target domain. Rigorous methodologies have emerged for bridging domains at the embedding, graph, signal, motif, causal, and language level, providing both practical pipelines for cold-start scenarios and a unified lens to study transferability and generalization in recommendation. This article surveys formal definitions, methodological foundations, algorithmic paradigms, advanced architectures, unified frameworks, and contemporary research directions central to CDR.

1. Formal Problem Definition and CDR Scenarios

Let $(\mathcal{U}^s, \mathcal{V}^s, R^s)$ be the source domain with user set $\mathcal{U}^s$ , item set $\mathcal{V}^s$ , and interactions $R^s$ ; $(\mathcal{U}^t, \mathcal{V}^t, R^t)$ analogously denote the target domain. Overlapping users (or in some settings, overlapping items) enable a bridge: $\mathcal{U}^o = \mathcal{U}^s \cap \mathcal{U}^t$ . The CDR objective is to learn a function $f: \mathcal{U}^{?} \times \mathcal{V}^t \to \hat{y}$ (prediction or ranking) for users with sparse or no $R^t$ , leveraging signals from $R^s$ (Zhu et al., 2021).

CDR is taxonomized into several scenarios:

Single-Target CDR: Enhances recommendations in a specific sparse target using a rich source (Zhu et al., 2021).
Dual-Target and Multi-Target CDR: Performs simultaneous transfer and enhancement across two or more domains (Zhu et al., 2021, Zhu et al., 2023, Cao et al., 2024).
Inter-Domain vs. Intra-Domain Transfer: Inter-domain transfer concerns cross-domain user (or item) cold-start recommendations; intra-domain leverages transfer within the same domain to improve performance for sparse users/items (Lee et al., 2024).
Cross-System Recommendation (CSR): Transfers across systems built on different resources (items/apps/users) (Zhu et al., 2021, Zhu et al., 2020).

Challenges include severe sparsity in $R^t$ , limited overlap ( $\mathcal{U}^s$ 0), domain heterogeneity, and the risk of negative transfer under distribution shifts (Zhu et al., 22 May 2025).

2. Algorithmic Paradigms and Foundations

2.1 Embedding and Mapping (EMCDR)

Canonical CDR methods first learn independent latent embeddings (via MF, BPR, or neural recommenders) in each domain. They then fit an explicit mapping $\mathcal{U}^s$ 1 using overlapping users: $\mathcal{U}^s$ 2 This paradigm, while simple and modular, generalizes poorly with small $\mathcal{U}^s$ 3, resulting in overfitting and sharp loss minima (Zeng et al., 2024).

2.2 Deep and Meta-Learning Approaches

Deeper mapping functions (MLPs, DNNs) better capture nonlinear dependencies between domains. Meta-learning (TMCDR) simulates many cold-start adaptation tasks for overlapping users, learning mapping networks $\mathcal{U}^s$ 4 optimized for rapid adaptation to new domains or users with limited data (Zhu et al., 2021). These approaches present notable gains over standard EMCDR, especially in low-overlap regimes (Zhu et al., 2021, Li et al., 20 Jan 2025).

2.3 Diffusion Models and Explicit Preference Integration

Diffusion models (DMCDR) introduce generative, stepwise information injection. Source-domain user preference signals, encoded via transformer layers, are conditionally injected into diffusion steps to reconstruct target-domain user representations for cold-start users. Objective terms include both standard recommendation and diffusion-elbo denoising losses, highlighting the benefit of explicit, stepwise transfer (Li et al., 20 Jan 2025).

2.4 Graph Neural, Heterogeneous, and Contrastive Models

Heterogeneous graphs capture rich cross-domain structures: user–item, taxonomy, content, and social edges. GNN-based CDR (GA, CCDR, DIDA-CDR, UniCDR⁺, MOP) leverage node/edge heterogeneity, augmented neighbor sampling, and motif-based or triplet structures to encode transferable and robust user/item representations (Zhu et al., 2021, Xie et al., 2021, Zhu et al., 2023, Cao et al., 2024, Hao et al., 2023). Both intra- and inter-domain contrastive learning objectives are deployed at various granularities (user, taxonomy, neighbor, motif) to better align domain-shared and domain-specific features.

2.5 Signal-Processing and Nonparametric Frameworks

Graph signal processing (CGSP) frames CDR as a signal reconstruction task on cross-domain similarity graphs. Personalized user signals are filtered across the whole graph, which linearly combines target-only and source-bridged similarities using an alpha-mixing parameter, yielding robust and efficient recommendations even as overlap ratios diminish (Lee et al., 2024).

2.6 LLMs for CDR

Emergent research demonstrates that LLMs, when precisely prompted with user histories and domain-specific guidance, can outperform neural CDR baselines in both rating prediction and ranking tasks (Liu et al., 10 Mar 2025, Vajjala et al., 2024). Prompt-based approaches remove the need for architecture re-engineering, leveraging pretrained models' capabilities to cross-map semantic features across domains, though performance degrades as domain gap widens.

3. Unified and Scenario-Invariant Architectures

While many early CDR models are scenario-expert (vertical, single scenario), recent advances push for scenario-invariant, universal frameworks (Cao et al., 2024, Hao et al., 2023). UniCDR⁺, for example, introduces domain-specific and domain-shared embedding layers with adaptive aggregation, disentanglement, and both hard/soft contrastive losses, unified across static, sequential, multi-action, intra/inter, and multi-domain settings. Motif-based prompt models (MOP) encode shared topological patterns across domains for transfer learning, employing both pre-training and prompt-tuning strategies for universal compatibility (Hao et al., 2023).

These frameworks demonstrate deployment viability at industrial scale, e.g., in Kuaishou's Living Room RecSys (Cao et al., 2024) and WeChat Top Stories (Xie et al., 2021), affirming their scalability and generalizability.

4. Causal and Robustness-Oriented CDR

Distribution shift has been a persistent barrier for CDR methods assuming i.i.d. samples. Causal-invariant CDR (CICDOR) formalizes both cross-domain and within-domain OOD shifts using dual-level structural causal models: $\mathcal{U}^s$ 5 where nodes encode both user attributes and shared/specific preferences. Causal invariance is enforced via acyclicity, path, and root constraints on learned adjacency matrices. LLM-guided confounder discovery modules bootstrap causal identification from natural-language reviews, integrating LLM output with classical CI and FCI tests for robust deconfounding (Zhu et al., 22 May 2025). Empirically, CICDOR outperforms strong baselines under both user-degree and regional distribution shifts.

Sharpness-aware methods (SCDR) directly regularize the loss geometry during mapping function training, penalizing sharp minima via adversarial perturbations and PAC-Bayes motivated objectives: $\mathcal{U}^s$ 6 demonstrating empirically and theoretically improved generalization and adversarial robustness on cold-start tasks (Zeng et al., 2024).

5. Practical Applications, Empirical Insights, and Deployment

Contemporary CDR systems deliver tangible impact in both academia and industry. Benchmark evaluations across Amazon, Douban, MovieLens, and WeChat datasets evidence substantial performance gains for modern CDR frameworks over single-domain baselines, especially when target domains are extremely sparse or overlap ratios are low (Vajjala et al., 2024, Li et al., 20 Jan 2025, Chen et al., 2023). Best-in-class methods demonstrate:

HR@10/NDCG@10 improvements of 8–15% over previous SOTA, especially in dual-target and multi-domain settings (Zhu et al., 2023, Hao et al., 2023).
Robustness to overlap ratios, hyperparameters, sharpness, and adversarial perturbations (Zeng et al., 2024, Chen et al., 2023).
Online and offline deployment success, including boosts in CTR, GAUC, ECPM, and recommendation diversity (Xie et al., 2021, Xie et al., 2021, Cao et al., 2024).

A prominent insight is that careful disentanglement of domain-shared and domain-specific components, along with structured contrastive or causal-invariant regularization, is critical to avoid negative transfer and oversmoothing. Plug-and-play adapters (Chen et al., 2023) mitigate catastrophic forgetting and reduce engineering overhead, while graph signal and motif-based approaches are particularly effective for efficiency and OOD robustness in real-world, highly sparse scenarios (Lee et al., 2024, Hao et al., 2023).

6. Open Challenges and Research Trajectories

Several directions remain central for advancing CDR:

Heterogeneous and Multimodal CDR: Improving the ability to align and transfer across domains with divergent side-information via graph or attention-based bridges (Zhu et al., 2021, Zhu et al., 2021).
Sequential and Streaming Extensions: Incorporating temporal cross-domain dynamics, e.g., using RNN or Transformer architectures to transfer preference sequences (Ma et al., 2023, Cao et al., 2024).
Causal Discovery at Scale: Integrating observed and latent confounder extraction from multimodal data (reviews, images, features) using LLMs and causal inference (Zhu et al., 22 May 2025).
Privacy-Preserving Transfer: Leveraging federated and secure computation to learn cross-domain mappings without exposing sensitive data (Zhu et al., 2021).
Parameter and Scenario Efficiency: Developing architectures that unify across multiple domains, overlaps (user, item, both), actions, and interaction types with few additional parameters (Cao et al., 2024, Hao et al., 2023).
LLMs and Prompt Learning: Scaling prompt-based CDR while controlling for context window limitations, semantic gap, and dynamic prompt design (Liu et al., 10 Mar 2025, Vajjala et al., 2024).

Negative transfer, model oversmoothing, and distributional robustness remain critical theoretical and practical concerns, calling for continued research in causal modeling, invariance principles, dynamic contrastive losses, and adaptive regularization.

References: