Cross-Domain Sequential Recommendation with Representation-Level User Transfer

1 Introduction

• Sparse or incomplete domain-specific histories limit even expressive sequential models, motivating cross-domain use of shared users’ behavioral signals.Existing approaches may require aligned item spaces, merged interaction data, or unrestricted access to raw interactions.Source: §1 · pp. 1–2
• Practical CDSR must handle restricted data sharing and partial user overlap, including users with no source-domain history.The paper frames restricted exchange of intermediate representations as a vertical federated learning setting.Source: §1 · p. 2
• FLADRec transfers learned user representations rather than raw interactions or shared item embeddings, and integrates a projected source representation into target-sequence modeling.The fusion mechanism aligns the source representation to the target embedding space.Source: §1 · p. 2
• Users without source histories default to a standard single-domain sequential recommender while retaining benefits from joint training with users who have cross-domain information.The introduction says recommendation quality improves for users observed exclusively in the target domain.Source: §1 · p. 2
• Experiments on multiple CDSR benchmarks show consistent gains over strong single-domain and cross-domain baselines.The introduction states that improvements include target users with and without source-domain activity.Source: §1 · p. 2

2.1 Problem Description

• The task predicts a user’s next target-domain item from chronological histories while allowing only learned representations—not raw interaction sequences—to cross domains.The framework also handles users who have no source-domain history.Source: §2.1 · p. 2

2.2 Encoder-Agnostic User Representations

• The framework can use any sequential encoder that produces a fixed-dimensional representation of a user’s interaction history.The source and target sequences are processed in their respective domain-specific pipelines.Source: §2.2 · p. 2

2.3 Fusing layer for cross-domain transfer

• A Fusing Layer normalizes and projects the source-user representation into the target latent space, then adds it element-wise to each target item embedding.The enriched sequence is passed to the target encoder to produce the representation used for next-item prediction.Source: §2.3 · pp. 2–3

2.4 Handling missing source-domain histories

• When a user has no source history, fusion is skipped and the target encoder receives unmodified target embeddings.Joint training includes both fused and unfused sequences, and the paper describes this as supporting improved recommendation quality for target-only users.Source: §2.4 · p. 3

2.5 Vertical Federated Learning

• Vertical federated learning fine-tunes the source encoder for the target objective by propagating gradients from the target side without sharing raw interactions.Only embeddings and their gradients are exchanged; raw interaction data and item IDs remain private to each domain.Source: §2.5 · pp. 3–4, Figure 2 · p. 4
• Source and target parameters are updated jointly, while batches without shared users revert to target-only training.The learned source representations are aligned with the target recommendation task under a distributed training protocol.Source: §2.5 · p. 4

3 Experiment Setup

• The setup evaluates FLADRec in complementary settings to balance comparison with prior work against realism in deployment conditions.Source: §3 · p. 4

3.1 Data and Splitting

• Amazon-derived Food-Kitchen, Beauty-Electronics, and Movie-Book benchmarks enable prior-work comparisons, but impose post-hoc domains, full user overlap, sampled evaluation, and potentially leaky or inconsistent splits.The benchmark protocol retains users appearing in both domains and uses last-item testing with sampled evaluation.Source: §3.1 · p. 4, §3.1.1 · p. 4
• TenRec and Zvuk&MegaMarket provide industrial-scale, naturally partial-overlap data with million-scale catalogs, full-catalog ranking, and strict temporal splits.TenRec histories are split sequentially within each user; Zvuk&MegaMarket uses global temporal weeks for validation and testing.Source: §3.1 · p. 4, §3.1.2 · p. 5

3.2 Training

• The task is next-item prediction; benchmarking uses full Cross-Entropy, while large-catalog scalability experiments use Scalable Cross-Entropy in pretraining and adaptation.Encoders start from pretrained single-domain SASRec models and are adapted with vertical federated learning.Source: §3.2 · p. 5

3.3 Evaluation and Metrics

• Evaluation reports NDCG@k, HR@k, and Coverage@k, using 999 sampled negatives for Amazon benchmarks but full-catalog ranking for industrial datasets.Models are tuned on validation NDCG@10 and retrained on combined training and validation data before test evaluation.Source: §3.3 · p. 5

3.4 Baselines

• Comparisons include target-only SASRec and BERT4Rec baselines, alongside cross-domain models including FedDCSR, CD-SASRec, CD-ASR, MGCL, C2DSR, DREAM, and ABXI.The single-domain baselines use Single-Target and Dual-Target hyperparameter-selection settings.Source: §3.4 · p. 5

4.1 Benchmarking results

• FLADRec achieves the best overall results across benchmark dataset pairs and metrics with limited remote interaction access, except for second place on Book HR@10.Table 3 covers Food-Kitchen, Beauty-Electronic, and Movie-Book.Source: §4.1 · p. 6, Table 3 · p. 6
• Against full-access methods, FLADRec remains broadly competitive under stricter interaction-sharing constraints, while mixed-data SASRec and BERT4Rec offer a viable compromise when joint training is allowed.The mixed-data single-domain baselines trail top performers ABXI and FLADRec; the passage attributes additional gains to adaptation mechanisms.Source: §4.1 · p. 6, Table 4 · p. 7

4.2 Scalability Evaluation

• FLADRec yields 3.68×–11.57× speedups over ABXI on benchmark datasets despite requiring four training stages.The comparison uses Table 6; the speedups are reported across different domain pairs.Source: §4.2 · p. 6
• Across TenRec and Zvuk&MegaMarket, FLADRec outperforms SASRec on all ranking metrics and cutoffs, with relative gains of 19–24% on TenRec.On Zvuk&MegaMarket, NDCG gains are 7.4–9.8%, and HitRate gains reach 20.4% at HR@50.Source: §4.2 · pp. 6–7
• On large-scale datasets, Coverage@10 rises by 6.6%, while coverage declines at cutoffs 50 and 100; evaluation uses partial user overlap and full-catalog ranking without negative sampling.The evaluation also uses strict temporal splits for Zvuk&MegaMarket, the dataset with timestamps.Source: §4.2 · p. 7

5.1 Effect of cross-domain embedding transfer

• The ablation compares full transfer, random injection, and no transfer, with no transfer bypassing fusion at inference while retaining source knowledge learned in target-encoder weights during joint training.This isolates explicit embedding transfer from implicit transfer through model weights.Source: §5.1 · pp. 7–8
• Full transfer performs best in both tuned- and frozen-source settings; for unshared users, full and no-transfer results match, while random embeddings can degrade performance.With a tuned source encoder, random embeddings slightly outperform no embedding for shared users; with a frozen encoder, random embeddings perform worse than no embeddings.Source: §5.1 · p. 8, Table 7 · p. 8

5.2 Ablation of source encoder training

• Freezing the source encoder degrades performance across user subsets and transfer variants, with particularly pronounced NDCG@10 drops for shared users under full transfer.Projecting fixed pretrained representations, even with adaptive fusion, is insufficient for optimal cross-domain learning.Source: §5.2 · p. 8
• Most performance gain is attributed to joint training of the proposed architecture rather than to changing the transfer regime or source-encoder status alone.The comparison contrasts full, random, and no transfer with trainable versus frozen source encoders.Source: §5.2 · p. 8

Related work

• CD-SASRec, DCDIR [3], and MIFN fuse domain sequences, while C2DSR [6], ABXI [4], and AutoCDSR [13] use representation alignment, parameter adaptation, or attention.These approaches typically rely on joint training across domains and merged interactions.Source: §6 · p. 9
• Federated recommendation work includes PPCDR, which shares locally differentially private user embeddings but assumes full user overlap, and FedCDR, which addresses rating rather than sequential prediction.Source: §6 · p. 9
• FLADRec transfers distilled user representations without exchanging raw interactions or item mappings, and supports partial overlap by reverting to single-domain recommendation when source data is absent.The framework also reports benefits for target-domain-only users through joint training with shared parameters.Source: §6 · p. 9

7.1 Privacy Analysis

• FLADRec lacks a dedicated defense against user-history reconstruction attacks, although formal differential privacy could be added to federated training.The architecture is designed for services in a shared ecosystem, and its distributed setting has mild privacy restrictions.Source: §7.1 · p. 9

7.2 Toward A Fair, Sustainable CDR Benchmarking Protocol

• Benchmarking pipelines contain five prevalent pitfalls that undermine practical reliability, and the scalability experiments avoid as many of them as possible.Source: §7.2 · p. 9

7.2.1 Origin of Data.

• Datasets split from one platform differ from datasets collected on genuinely separate platforms, whose models are independently produced.The passage identifies MegaMarket–Zvuk and TenRec as examples of the latter.Source: §7.2.1 · p. 9

7.2.2 Domain Overlap Ratio.

• Real-world overlap is limited, but many CDR evaluations assume unrealistically high or full overlap and underreport differences between shared and non-shared users.Source: §7.2.2 · p. 9

7.2.3 Dataset Size and Scalability.

• Comparisons limited to small, heavily processed datasets may have limited value because replicating many such models in real-world settings is computationally infeasible.Source: §7.2.3 · p. 9

7.2.4 Sampled Metric Evaluation.

• Ranking against sampled negatives can distort model comparisons, despite its faster computation.Prior work has proposed corrections for this bias.Source: §7.2.4 · pp. 9–10

7.2.5 Temporal Leakage in Train/Test Split.

• Applying last-item evaluation to non-temporal or temporally inconsistent splits risks temporal leakage; the reported experiments retain a 50% overlap and TenRec lacks timestamps for strict temporal splitting.The authors argue that partial overlap and scalability are the most critical conditions to prioritize, while optimal conditions across all considerations are demanding for fast-paced research.Source: §7.2.5 · p. 10

• The framework transfers learned user representations without sharing raw interaction data or item embeddings across domains.It combines an SASRec-based architecture, a lightweight fusing layer, and vertical federated fine-tuning of the source encoder.Source: §8 · p. 10
• Experiments across five datasets, including two industrial-scale and three academic benchmarks, show consistent improvements over strong single-domain and cross-domain baselines for shared and target-only users.Source: §8 · p. 10
• Future work includes multi-source and multi-task settings, alternative fusion mechanisms and backbones, and privacy-preserving mechanisms for exchanged data during learning.Source: §8 · p. 10

• Claude was used to rephrase parts of the manuscript and improve the readability of the open-source code.Source: GenAI Usage Disclosure · p. 10
• The authors report that experimental design, data collection, statistical analysis, and interpretation were performed without GenAI.Source: GenAI Usage Disclosure · p. 10
• The authors reviewed and edited the AI-assisted text and code changes and take responsibility for the final content.Source: GenAI Usage Disclosure · p. 10