You can download the data from here.

1.Use the following command to create a new conda enviroment.

conda env create -f conda_env.yml

2.Install the modified Deepctr-torch[1] package

cd other_projects/DeepCTR-Torch
pip install -e .

To improve the generalization of models, we added several adversarial training methods to this package.

See track1 inference.ipynb, you need to modify the test_data_dir we use three models to predict

The test_data_dir folder's structure is follwing:

├── Data
│   ├── CodeStates
│   │   └── CodeStates.csv
│   ├── DatasetMetadata.csv
│   ├── LinkTables
│   │   └── Subject.csv
│   └── MainTable.csv
├── early.csv
└── late.csv

The score maybe close to 0.7965.

See track2 inference.ipynb, you need to modify the test_data_dir, they use the same folder in Track1 inference.

The score maybe close to 161.

In this task, we try to use methods from the recommender system to predict future students' performance. The first is extracting features, see the jupyter notebook in track1_extract_features.ipynb. Then, we build two types of models to predict the test datasets. Finally, we tried different strategies to merge different models' results.

• single model: we try almost all models in DeepCTR.
• multi huge model: we merge multi models in DeepCTR into one huge model and add multi-dropout and adversarial training to improve the performance in test datasets.

You can train by the command python deepctr.py

You can train by the command python deepctr_huge.py

We use the wandb to manage the results of each model. All config can be found in wandb_configs

You can start one sweep by the following command (one group params search)

wandb sweep wandb_configs/deepctr_5cv.yaml -p csedm_2021

In this trace, we extracted many features from the first 30 questions' interactions to predict each student's final grade. To be specific, first, we build two types of models to model each student's performance. Second, we try to use the different strategies to merge the result from each model.

• one stage: see track2_one_stage.py
• multi stage: see track2_multi_stage.ipynb

1. deepctr-torch
2. CSEDM2021 official website
3. https://github.com/haradai1262/NeurIPS-Education-Challenge-2020

Email: [email protected] (Jiahao Chen)