This is the official PyTorch implementation for the paper:

Chen Wang, Ziwei Fan, Liangwei Yang, Mingdai Yang, Xiaolong Liu, Zhiwei Liu, Philip Yu. Pre-Training with Transferable Attention for Addressing Market Shifts in Cross-Market Sequential Recommendation. KDD 2024.

In this study, we introduce the Cross-market Attention Transferring with Sequential Recommendation (CAT-SR) framework, tailored specifically for cross-market recommendation (CMR) scenarios. CMR poses unique challenges such as strict privacy regulations that limit data sharing, lack of user overlap, and consistent item sets across different international markets. These aspects are further compounded by market-specific variations in user preferences and item popularity, known as market shifts.

To effectively address these hurdles and enhance recommendation accuracy across disparate markets, CATSR employs a sophisticated approach that leverages a preconditioning strategy focusing on item-item correlations and incorporates an innovative selective self-attention mechanism. This mechanism facilitates the transfer of focused learning across markets. Additionally, the framework enhances adaptability through the integration of query and key adapters, which are designed to capture and adjust to market-specific nuances in user behavior.

recbole==1.1.1
python==3.8.5
cudatoolkit==11.3.1
pytorch==1.12.1
pandas==1.3.0
transformers==4.18.0

Put data file into data directory. For example: data/ca_5core.txt

Category: Electronics

Data: metadata

Put dataset into data/Amazon/metadata directory. For example data/Amazon/metadata/meta_Electronics.json.gz

cd data
python data_process.py

python pretrain.py

Take finetune Canada(ca) as an example

python finetune.py --dataset ca