Research map / Label-efficient and robust learning / Domain adaptation and generalization

Enhancing Healthcare Recommendations: A Privacy-Protective and Interpretable Cross-Domain Framework

AAAI 2025 · Label-efficient and robust learning › Domain adaptation and generalization · approach: Cross-domain text and graph transfer · 1 citations (OpenAlex)

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 HCR, a privacy-preserving, interpretable cross-domain recommender that perturbs source-domain user interactions via a cold-item replacement synthesizer and generates patient-friendly explanations with GPT-2 using user/item ID embeddings as continuous prompts.

Problem

Cross-domain recommendation in healthcare (short-video viewers to doctors for online consultation) must protect sensitive user interaction data and produce explanations that lay users can understand, which existing CDR and PPCDR methods do not jointly address, especially for multimodal data.

Approach

Combine a self-supervised modality-aware LightGCN-style encoder with a privacy-preserving synthesizer that replaces low-contribution (cold) items with Gumbel-Softmax-generated virtual items constrained by a privacy hinge loss, plus a GPT-2 explanation generator that takes user and item embeddings as continuous prompts, trained in a two-stage prompt-then-full tuning scheme.

Key results

• On Healthcare Video→Doctor, HCR achieves HR .4453 and NDCG .3560 with privNE .7220, versus PMMRec's HR .4449 and privNE .2803, a roughly 158% privacy-metric improvement.
• Human study with 500 participants found explanations improved relevance ratings, with a larger effect in healthcare (Video→Doctor) than e-commerce (Software→Game).
• Removing the privacy synthesizer (Ours w/o PP) raises some accuracy metrics but drops privNE to about 0.29, showing a trade-off between accuracy and privacy.

Measured effects reported in the paper

Surprising findings

• Removing the self-supervised encoder (w/o SS) gives much higher privacy NE (0.72) but substantially lower recommendation accuracy, so privacy gains came with accuracy loss in that variant.
• Explanations had a larger impact on healthcare recommendation acceptance than on e-commerce, where images were sufficient for engagement.

Limitations

• Relies on public datasets; not tested in real-world healthcare deployments with interoperability constraints.
• HCR shows accuracy loss on some datasets (e.g., Amazon Arts→Fashion) relative to multimodal baselines like PMMRec.
• Explanation-human study uses only 50 users in one setup and 500 participants overall; the collected healthcare dataset is not described in detail regarding ethics approval.

Experimental setup

Datasets: Amazon Arts Amazon Fashion Amazon Software Amazon Game Healthcare Video Healthcare Doctor

Benchmarks: Amazon Arts→Fashion Amazon Software→Game Healthcare Video→Doctor Healthcare Doctor→Video

Metrics: HR@5 NDCG@5 privNE privCPL USR

Baselines: CoNet CCTL SEMI PMMRec PriCDR DPSMRec PPGenCDR Ours w/o SS Ours w/o PP

Builds on: LightGCN InfoNCE Gumbel-Softmax GPT-2 CoNet

Closest papers by meaning

• Enhancing Small Medical Learners with Privacy-preserving Contextual Prompting (ICLR 2024 · similarity 0.85)
• CARES: A Comprehensive Benchmark of Trustworthiness in Medical Vision Language Models (NeurIPS 2024 · similarity 0.82)
• Reducing Item Discrepancy via Differentially Private Robust Embedding Alignment for Privacy-Preserving Cross Domain Recommendation (ICML 2024 · similarity 0.82)
• An Investigation of Memorization Risk in Healthcare Foundation Models (NeurIPS 2025 · similarity 0.81)
• Collaborative Synthesis of Patient Records through Multi-Visit Health State Inference (AAAI 2024 · similarity 0.81)
• Domain-Level Disentanglement Framework Based on Information Enhancement for Cross-Domain Cold-Start Recommendation (AAAI 2025 · similarity 0.81)