Code for our work: Effective and Efficient Training for Sequential Recommendation Using Cumulative Cross-Entropy Loss
The following is a brief directory structure and description of this example:
├── data
├── data_process.py # the data process file.
├── generate_test.py # generate the negative example for BCE loss.
├── Beauty_item2attributes.json # Item attributes for S3-Rec pre-training task.
├── Beauty_sample.txt # Negative sampling for binary cross-entropy loss in advance.
├── Beauty.txt # The user interaction records after data process.
├── ...
├── final_result # the model training log folder.
├── GRU4Rec # unzip the GRU4Rec.zip file here.
├── S3Rec # unzip the S3Rec.zip file here.
├── SASRec # unzip the SASRec.zip file here.
├── RQ2_data # This data is downloaded by the tensorbord of the above files.
├── imgs
├── Intruduction.svg # image in paper introduction.
├── RQ1.png
├── RQ2.svg
├── RQ3.svg
├── README.md
├── datasets.py
├── loss.py
├── models.py
├── modules.py
├── pre_exam.ipynb # generate the experiments result of RQ3.
├── run_finetune_full.py
├── run_pretrain.py # for the pre-training task of S3-Rec.
├── SR_fig_process.ipynb # draw image file.
├── trainers.py
├── utils.py
Note that the GRU4Rec.zip, S3Rec.zip, and SASRec.zip files are the training log run by ourselves. And due to the limitation of file size, we upload separately.
torch == 1.11.0 numpy==1.19.1 scipy==1.5.2 tqdm==4.48.2
python 3.7.3
os: windows/linux
Here we give a example:
python run_finetune_full.py --data_name LastFM --output_dir final_result/ --model_name GRU4Rec --istb 1 --loss_type CE_ALLThis example code means using CCE loss train GRU4Rec in LastFM datasets.
For data_name here are five choice: [Sports_and_Outdoors, Toys_and_Games, Yelp, Beauty, LastFM].
For loss_type [CE, PointwiseCE, CE_ALL], where CE reprensent the CE loss; PointwiseCE represent the BCE loss; CE_ALL represent the CCE loss.
#GRU4Rec
python run_finetune_full.py --data_name LastFM --output_dir final_result/ --model_name GRU4Rec --istb 1 --loss_type CE_ALL
python run_finetune_full.py --data_name LastFM --output_dir final_result/ --model_name GRU4Rec --istb 1 --loss_type CE
python run_finetune_full.py --data_name LastFM --output_dir final_result/ --model_name GRU4Rec --istb 1 --loss_type PointwiseCE
#To get the full result of GRU4Rec, need to change the data_name. This is similar to SASRec and S3Rec.
#SASRec
python run_finetune_full.py --data_name LastFM --output_dir final_result/ --model_name SASRec --istb 1 --loss_type CE_ALL
python run_finetune_full.py --data_name LastFM --output_dir final_result/ --model_name SASRec --istb 1 --loss_type CE
python run_finetune_full.py --data_name LastFM --output_dir final_result/ --model_name SASRec --istb 1 --loss_type PointwiseCE
#S3Rec
python run_finetune_full.py --data_name LastFM --output_dir final_result/ --model_name S3Rec --istb 1 --loss_type CE_ALL --ckp 150
python run_finetune_full.py --data_name Beauty --output_dir final_result/ --model_name S3Rec --istb 1 --loss_type CE_ALL --ckp 150
python run_finetune_full.py --data_name Sports_and_Outdoors --output_dir final_result/ --model_name S3Rec --istb 1 --loss_type CE_ALL --ckp 100
python run_finetune_full.py --data_name Toys_and_Games --output_dir final_result/ --model_name S3Rec --istb 1 --loss_type CE_ALL --ckp 150
python run_finetune_full.py --data_name Yelp --output_dir final_result/ --model_name S3Rec --istb 1 --loss_type CE_ALL --ckp 100
# Note that here has an additional parameter ckp, which depends on the pre-training model in reproduce folder.
# Besides, those pre-training models are offered by the source code of S3Rec.Then, the training log is all saved in the final_result folder. And the experiment results in Table 2 and Fig 2 all came from the training log.
Just run the code in pre_exam.ipynb, then will get the .csv file at final_result folder.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent