Data-Isolated Cross-Domain Sequential Recommendation

Updated 12 January 2026

The paper proposes model merging and adapter techniques that transfer cross-domain sequential knowledge without sharing raw data, enhancing recommendation performance.
It employs contrastive, meta-learning, and variational methods to mitigate cold-start issues and alleviate data sparsity in isolated domains.
Empirical results show significant gains in ranking metrics such as Recall and NDCG across datasets while upholding strict privacy guarantees.

Data-isolated cross-domain sequential recommendation addresses the challenge of leveraging knowledge across multiple recommendation domains in settings where raw user interaction data cannot be exchanged between domains. This scenario arises due to privacy, legal, or organizational boundaries that prohibit direct data sharing, leading to disjoint user and item spaces across domains. Approaches in this area focus on enhancing generalization and recommendation quality in each isolated domain, especially under cold-start and data sparsity constraints, by transferring knowledge without violating data isolation.

1. Formal Problem Setting

In data-isolated cross-domain sequential recommendation, each domain $D_k$ possesses its own disjoint user set $U_k$ and item set $I_k$ , satisfying $U_k \cap U_{k'} = \emptyset$ and $I_k \cap I_{k'} = \emptyset$ for all $k \ne k'$ (Kim et al., 5 Jan 2026, Hou et al., 30 Oct 2025). User interactions are available only within their respective domains, and raw interaction sequences $\{u: u \in U_k\}$ are not shared across domains. The research objective is to build a universal recommendation model $f_{\text{merge}}(\cdot)$ that, without direct access to out-of-domain data, can generalize to new domains and accurately predict the next-item for any user $u$ in any domain $k$ , effectively leveraging cross-domain knowledge in a privacy-compliant manner.

Formally, the recommendation model aims to maximize the probability of the next interaction for each user, i.e.,

$\max_\Theta\ \mathbb{P}\bigl(v_{|s_u|+1}=v\mid s_u\,,\,\mathcal{K}(\mathcal{D})\bigr)\,,\quad \forall\,u\in\mathcal{U}_k.$

Here $\mathcal{K}(\mathcal{D})$ denotes transferred cross-domain knowledge, and $\Theta$ includes backbone and adapter parameters. The evaluation is usually based on ranking metrics (NDCG@ $k$ , Recall@ $k$ ) in the target domain.

2. Core Methodologies

Several methodological families target data-isolated cross-domain sequential recommendation, unified by their avoidance of raw data sharing:

a) Model Merging

Model merging constructs universal recommenders by aggregating fine-tuned, domain-private models at the parameter level (Hou et al., 30 Oct 2025, Kim et al., 5 Jan 2026). The process involves:

Independently fine-tuning models $f_k(\cdot\,|\theta_k)$ in each domain using local data.
Merging parameter deltas relative to a shared base model: $\theta_{\rm merge} = \theta_{\rm base} + \sum_k w_k (\theta_k - \theta_{\rm base})$ , where $w_k$ are learned or uniform weights.
Optimization leverages synthetic pseudo-users (e.g., using each item as a single-step pseudo-user), with collaborative losses combining distillation (KL divergence of predicted distributions) and top-1 recommendation losses.
Notably, MergeRec (Kim et al., 5 Jan 2026) introduces pseudo-user data and collaborative optimization to compute merging weights, achieving up to 17.21% improvement in Recall@10 over prior merging baselines, and enabling privacy-compliant, scalable transfer.

b) Adapter and LoRA-based Approaches

Adapter-based and LoRA (Low-Rank Adaptation) frameworks such as WeaveRec (Hou et al., 30 Oct 2025) attach lightweight, domain-specific LoRA modules to a frozen LLM backbone. Each branch is trained on either the target domain or a hybrid mixture of the target plus one source. At inference, modules are merged via convex averaging, yielding a single adapter that benefits from cross-domain knowledge while incurring no extra inference cost. WeaveRec supplies a theoretical guarantee that the hybrid LoRA model reduces the expected target error compared to source-only training.

c) Meta-Learning, Contrastive, and Denoising Techniques

Methods like MACD (Xu et al., 2023) utilize plug-and-play contrastive denoising atop standard sequential backbones. Multi-head attention modules distill explicit (intra-domain) and implicit (cross-domain) interests from auxiliary behaviors, aligned by contrastive InfoNCE losses. Inductive representation mechanisms allow cold-start users in one domain to borrow interest representations by nearest-neighbor mapping from auxiliary domains, even in the absence of user overlap.

d) Variational and Probabilistic Latent Variable Models

i²VAE/IM-VAE (Ning et al., 2024) introduces a variational autoencoder with structured latent variables disentangling domain-specific and cross-domain user interests, augmented with pseudo-sequence generation to support cold-start and long-tailed users. Mutual information-based KL regularizers ensure cross-domain latents carry transferrable information, while denoising regularizes the impact of generated pseudo-sequences.

e) Attention- and Prototype-based Aggregation

MAN (Lin et al., 2023) and similar models feature mixed attention frameworks with domain-specific and global encoders, cross-attention modules, and shared group prototypes learned from both domains, facilitating knowledge transfer at the representation level without explicit user-item overlaps.

f) LLM-Based Semantic Bridging

LLM4CDSR (Liu et al., 25 Apr 2025) utilizes frozen LLMs to generate unified semantic embeddings for all items and synthesizes user-level preference summaries by LLM-based hierarchical profiling. Lightweight adapters and contrastive losses bridge the semantic gap, improving cross-domain alignment in absence of overlapping users.

3. Theoretical Guarantees and Privacy

Ensuring privacy and providing theoretical guarantees are central to the data-isolated setting:

MergeRec (Kim et al., 5 Jan 2026) and WeaveRec (Hou et al., 30 Oct 2025) leverage only model weights and pseudo-user inferences for model merging, which leak no raw user PII.
PriCDSR (Lin et al., 2024) formalizes $\varepsilon$ -sequential differential privacy (SDP) for ordered item-sequence data and introduces a randomized mechanism that perturbs auxiliary sequences before transfer. The mechanism is proven to satisfy $\varepsilon$ -SDP and can operate client-side for local DP guarantees.
WeaveRec supplies generalization error bounds grounded in domain adaptation theory (Ben-David et al.), demonstrating strictly lower expected target-domain error when hybrid training is employed versus source-only or naive merging.

4. Empirical Results and Comparative Evaluation

Experimental studies on Amazon multi-domain, MovieLens, and Douban datasets, as well as large-scale industrial systems, confirm the efficacy of data-isolated cross-domain methods:

MergeRec outperforms both joint-learning and other merging baselines in average Recall@10 by 8.72–17.21%, maintaining superior performance even with no target domain logs (Kim et al., 5 Jan 2026).
WeaveRec improves over target-only and naive merge approaches by 2–5% in NDCG, with each hybrid branch alone matching target-only, but their merge delivering further uplift (Hou et al., 30 Oct 2025).
LLM4CDSR yields gains of up to +31.7% in Hit@10 versus best CDR baselines when only 25% of users overlap, leveraging LLM embeddings to bridge semantic gaps (Liu et al., 25 Apr 2025).
Variational and contrastive-denoising approaches such as i²VAE (Ning et al., 2024) and MACD (Xu et al., 2023) report up to 15.2% improvements for cold-start users and robustness to decreasing overlap ratios.
Online A/B tests with SyNCRec (Park et al., 2024) and AMID (Xu et al., 2023) in real-world environments validate positive click-through and conversion-rate impacts, with strong ablation evidence for the importance of negative-transfer correction and interest-grouping modules.

5. Adaptation to Heterogeneous and Open-World Scenarios

Modern approaches address the challenges of users/items entering unseen domains (open-world), negative transfer, and selection bias:

AMID (Xu et al., 2023) leverages an adaptive multi-interest module (MIM) that forms soft interest-group links between users in different domains based on their sequence similarity, independent of overlapping IDs. Such bridges allow information propagation in fully isolated regimes.
Doubly robust estimators correct selection bias in debiased risk computation, offering lower bias/variance than standard inverse propensity scoring and remaining robust as overlap vanishes.
Mixture-of-experts routing (e.g., SyNCRec (Park et al., 2024)) and adaptive reweighting strategies suppress negative transfer from unrelated domains, leading to consistently positive gains even in highly heterogeneous environments.

6. Practical Implementation and Recommendations

State-of-the-art solutions provide explicit recipes to ensure practical deployment:

Model merging (MergeRec, WeaveRec) requires only access to model weights, not logs; shared base model preparation is critical (Kim et al., 5 Jan 2026, Hou et al., 30 Oct 2025).
Pseudo-user construction using all items as single-interaction proxies allows full utilization of model ensemble strengths without leaking user-level data (Kim et al., 5 Jan 2026).
For privacy, $\varepsilon$ in PriCDSR should be selected contextually: $\varepsilon=[5,20]$ usually balances utility and legal requirements, and the randomized mechanism can be executed locally or in preprocessing (Lin et al., 2024).
Integration with existing sequential recommendation backbones is straightforward, often plug-and-play for MACD, MAN, and AMID (Xu et al., 2023, Lin et al., 2023, Xu et al., 2023).
LLM-based or hybrid approaches (LLM4CDSR) minimize online latency by requiring LLM queries only offline; at inference, prediction is performed by lightweight neural components (Liu et al., 25 Apr 2025).

7. Research Impact and Future Directions

Data-isolated cross-domain sequential recommendation has redefined the practical and theoretical boundaries of transfer learning in recommender systems, making robust, privacy-compliant knowledge transfer feasible in industrial and privacy-sensitive scenarios. Open problems include:

More efficient, scalable, and theoretically grounded indirection methods for large-scale, highly heterogeneous, or dynamically evolving domains.
Enhanced treatment of non-overlapping entities beyond user-level isolation, including cold-start item adaptation and cross-lingual/semantic alignment.
Tighter theoretical analysis of model merging and privacy-preserving mechanisms.
Unified benchmarks and standardized protocols to evaluate emerging data-isolated transfer schemes across diverse recommendation tasks.

Key advances such as model merging with pseudo-user signals (Kim et al., 5 Jan 2026), weaving hybrid adapters with zero inference overhead (Hou et al., 30 Oct 2025), and leveraging LLM semantic alignment (Liu et al., 25 Apr 2025), have already established foundational methodologies in this domain. Ongoing research will likely focus on federated, personalized, and explainable cross-domain recommendation, preserving data sovereignty and enhancing robustness in open-world settings.