Research map / Label-efficient and robust learning / Recognition and long-tail
Learning Accurate and Bidirectional Transformation via Dynamic Embedding Transportation for Cross-Domain Recommendation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes JPEDET, a dual-target cross-domain recommender for non-overlapped users that fuses ratings and reviews and transports user embeddings bidirectionally with neural ODE-based dynamic transport flow guided by barycentric Wasserstein paths.
Problem
Cross-domain recommendation when users and items are non-overlapped across domains, so no explicit bridge exists for knowledge transfer, and embedding discrepancy between domains hurts performance.
Approach
Dynamic transport flow combines neural ODEs with discrete optimal transport (Sinkhorn coupling) for a matching regularization stage, then adds a moving correction stage with barycentric Wasserstein paths and linear interpolation guidance to obtain straight, constant-speed, invertible transformations between source and target user embeddings.
Key results
- JPEDET achieves RMSE 1.267 and MAE 0.879 on Book->CD, versus 1.456/1.031 for the runner-up SER.
- JPEDET improves over runner-up models by roughly 7.4% to 17.9% across tasks, per the paper.
- Replacing the dynamic embedding transport module with MMD or adversarial training yields worse results than JPEDET (e.g., JPEDET-M 1.714 vs JPEDET 1.267 RMSE on Book->CD).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use JPEDET instead of SER | Cross-domain recommendation · Amazon tasks (Book↔CD, Book↔Movie, Movie↔CD, Phone↔Elec) | RMSE: JPEDET achieves improvement from 7.4% to 17.9% | Helped |
| Add dynamic embedding transportation to joint preference exploration (ablation) | JPEDET · Book->Movie | RMSE: 2.264 -> 1.206 | Helped |
| Use dynamic transport flow instead of MMD (ablation) | JPEDET · Book->Movie | RMSE: 1.521 -> 1.206 | Helped |
| Use dynamic transport flow instead of domain-adversarial training (ablation) | JPEDET · Book->Movie | RMSE: 1.469 -> 1.206 | Helped |
| Use dynamic transport flow instead of OT-Flow (ablation) | JPEDET · Book->Movie | RMSE: 1.358 -> 1.206 | Helped |
| Use dynamic transport flow instead of TPR (ablation) | JPEDET · Book->Movie | RMSE: 1.362 -> 1.206 | Helped |
| Use dynamic transport flow instead of Rectified Flow (ablation) | JPEDET · Book->Movie | RMSE: 1.310 -> 1.206 | Helped |
Surprising findings
- Adversarial training with domain discriminators is reported as unstable and hard to train, degrading cross-domain performance.
- JPEDET-B (without transport module) performs poorly, with RMSE 3.685 on Book->CD, close to single-domain baselines.
Limitations
- Discrete OT with Sinkhorn has O(N^2) complexity per the paper.
- Only evaluated on Amazon datasets with rating-based RMSE/MAE; ranking metrics not reported.
- Overlapped-user extension evaluated only on one task with 5% overlap.
Experimental setup
Datasets: Amazon Movies and TV Amazon Books Amazon CDs and Vinyl Amazon Cell Phones and Accessories Amazon Electronics
Benchmarks: Amazon Books-CD Amazon Books-Movie Amazon Movie-CD Amazon Phone-Electronics
Metrics: RMSE MAE
Baselines: NeuMF DeepCoNN VCM NARRE UBERT ESCOFILT RC-DFM GA-DTCDR Rec-DAN DDTCDR TDAR DisAlign
Builds on: neural ODE OT-Flow rectified flow Sinkhorn algorithm DeepCoNN
Closest papers by meaning
- Towards Unbiased Information Extraction and Adaptation in Cross-Domain Recommendation (AAAI 2025 · similarity 0.87)
- Graph Disentangled Contrastive Learning with Personalized Transfer for Cross-Domain Recommendation (AAAI 2024 · similarity 0.86)
- DisCo: Graph-Based Disentangled Contrastive Learning for Cold-Start Cross-Domain Recommendation (AAAI 2025 · similarity 0.86)
- Domain-Level Disentanglement Framework Based on Information Enhancement for Cross-Domain Cold-Start Recommendation (AAAI 2025 · similarity 0.86)
- Learning Multiple User Distributions for Recommendation via Guided Conditional Diffusion (AAAI 2025 · similarity 0.85)
- Reducing Item Discrepancy via Differentially Private Robust Embedding Alignment for Privacy-Preserving Cross Domain Recommendation (ICML 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.