chenhch8 / rl4rs-dqn

This is a DQN-based recommendation system for item-list recommendation and it finally achieved second place in the competition of RL-based RecSy.

BigData2021-Reinforcement-Learning-based-Recommender-System-Challenge

Our project structure is as follows:

├── dataset
│   ├── dev.pt
│   ├── item_info.csv
│   ├── items_info.pt
│   ├── test.pt
│   ├── track1_testset.csv
│   ├── track2_testset.csv
│   ├── train.pt
│   └── trainset.csv
├── output
│   ├── best_checkpoint.pt
├── process.sh
├── src
└── train.sh

Dependence

pytorch==1.6.0
prettytable
cloudpickle

Data Processing

Download the trainset.csv, track1_testset.csv, track2_testset.csv and item_info.csv into dataset and then run the following shell to process the data to obtain the pickle data.

sh process.sh

Train

Run the following shell to train the DQN, and it will save best_checkpoint.pt in output.

Note that our best checkpoint have been saved in output.

sh train.sh

Eval

Run the following shell to obtain the evaluating result on the track2 dataset, and it will generate test_result.json in output.

sh test.sh

Upload result

Run the following shell to obtain the submission.csv based on the above test_result.json.

sh upload.sh --pred_file output/test_result.json