ProRel: Prototype-Guided Reliable Transfer for Cross-Domain Recommendation

Official implementation of ProRel, a prototype-guided reliable transfer framework for cross-domain recommendation.

Abstract

Cross-domain recommendation (CDR) mitigates target-domain sparsity by transferring preference signals through overlapping users. However, shared users establish cross-domain connectivity, not guaranteed transferability. Overlap-based bridging indicates that two domains are connected, but it does not specify whether the carried signals are trustworthy for target-domain recommendation. Reliability gaps appear at two stages where overlap-induced connectivity is consumed. During graph propagation, uniformly treating observed edges can amplify weakly supported or domain-specific user-item relations. During representation fusion, transferred representations may retain redundant or domain-conflicting factors and increase negative transfer.

We propose ProRel, a Prototype-guided Reliable transfer framework that applies reliability criteria at these two stages. For graph propagation, Prototype-Guided Reliable Graph Learning (PGRL) estimates edge reliability by measuring the consistency between local user-item relations and prototype-level preference regularities. PGRL assigns larger propagation weights to edges with stronger node-prototype agreement, so unreliable relations are down-weighted before their signals spread through the shared graph.

For representation fusion, Reliability-Aware Selective Fusion (RASF) targets negative transfer after reliable propagation. RASF preserves recommendation-relevant user correspondence across shared, domain-specific, and fused representations, while suppressing redundant cross-domain dependence with HSIC-based regularization. By doing so, RASF reduces the influence of source-specific or conflicting factors before final prediction. Experiments on four Amazon CDR tasks show that ProRel consistently outperforms competitive baselines, achieving an average relative improvement of 7.03% and up to 12.20% over the strongest baseline across HR@10 and NDCG@10.

Environment

• Python 3.9.0
• PyTorch 1.12.0
• NumPy 1.24.3
• SciPy 1.11.1

Datasets

We use the datasets provided by DisenCDR.

Place the processed data under dataset/ with the expected domain-pair folder structure, for example:

dataset/
  electronic_cloth/
  cloth_electronic/
  sport_phone/
  phone_sport/

Running

Run commands from the src/ directory. The shared hyperparameter across the four Amazon tasks is lambda1=0.01. The item prototype count is controlled by k_item for both domains.

Amazon Elec & Cloth:

CUDA_VISIBLE_DEVICES=1 python train_rec.py --dataset electronic_cloth --lambda1 0.01 --lambda2 0.6 --lambda_regularization 0.6 --k_user 4 --k_item 16

Amazon Sport & Phone:

CUDA_VISIBLE_DEVICES=0 python train_rec.py --dataset sport_phone --lambda1 0.01 --lambda2 0.4 --lambda_regularization 0.5 --k_user 4 --k_item 16

Amazon Sport & Cloth:

CUDA_VISIBLE_DEVICES=2 python train_rec.py --dataset sport_cloth --lambda1 0.01 --lambda2 0.6 --lambda_regularization 0.6 --k_user 16 --k_item 16

Amazon Elec & Phone:

CUDA_VISIBLE_DEVICES=2 python train_rec.py --dataset electronic_phone --lambda1 0.01 --lambda2 0.6 --lambda_regularization 0.6 --k_user 4 --k_item 32

Main Arguments

• --model_name: model name, default ProRel
• --use_prototype_denoise: enable PGRL
• --edge_reliability_mode: edge reliability setting, one of learned, uniform, or shuffle
• --k_user: number of user prototypes
• --k_item: number of item prototypes for each domain
• --lambda_regularization: weight for HSIC-based reliability regularization in RASF
• --fusion_mode: user representation fusion strategy, one of ours, mean, or sum

Acknowledgements

This repository builds on resources from:

• LightGCL
• DisenCDR
• ETL

We thank the authors for making their code and datasets publicly available.