Codes for SIGIR 2023 Full Paper-Towards Multi-Interest Pre-training with Sparse Capsule Network

Firstly, create a folder named "dataset" in the current directory. Then, within the "dataset" folder, create two subfolders named "Ratings" and "Metadata" respectively. Then you can donwload the dataset from Amazon. The .csv files should be placed in the "Ratings" folder, while meta_*.json.gz files should be placed in the "Metadata" folder. Thanks to UniSRec for providing another link to download datasets from cloud disks

You only needs python process_amazon.py to process the raw datasets if you placed the dataset as required in Step1.

You can pretrain the model by python main.py --dataset=FHCKM --gpu_id=0 --save_step=1 --epoch=3 --stage=pretrain --train_batch_size=2048

Notes: the train_batch_size hyperparameter is important, if you modify this hyperparameter, you may get unexpected results due to overfitting. The FHCKM means the mixed pretrain dataset.

You can test the model by python main.py --dataset=Scientific --gpu_id=0 --epoch=100 --stage=trans --train_batch_size=1024 --load_model_path={pretrain_model_path in Step3}

Notes: the load_model_path parameter is printed in Step3, please use the checkpoint ending with "pretrain-2.pkl"(which pretrain for 3 epochs), such as "./saved_model/2-13-50-10/pretrain-2.pkl"

Thanks to UniSRec for its open source codes and datasets.

Notes:I have run the entire code on my computer to ensure that it can be used, if you encounter problems, you can contact me by email [email protected]