HawkesKT

This is our implementation for the paper:

Chenyang Wang, Weizhi Ma, Min Zhang, Chuancheng Lv, Fengyuan Wan, Huijie Lin, Taoran Tang, Yiqun Liu, and Shaoping Ma. Temporal Cross-effects in Knowledge Tracing. In WSDM'21.

Usage

1. Install Anaconda with Python >= 3.5
2. Clone the repository and install requirements

git clone https://github.com/THUwangcy/HawkesKT

3. Prepare datasets according to README in data directory
4. Install requirements and step into the src folder

cd HawkesKT
pip install -r requirements.txt
cd src

5. Run model

python main.py --model_name HawkesKT --emb_size 64 --max_step 50 --lr 5e-3 --l2 1e-5 --time_log 5 --gpu 1 --dataset ASSISTments_09-10

Example training log can be found here.

Arguments

The main arguments of HawkesKT are listed below.

Args	Default	Help
emb_size	64	Size of embedding vectors
time_log	e	Base of log transformation on time intervals
max_step	50	Consider the first max_step interactions in each sequence
fold	0	Fold to run
lr	1e-3	Learning rate
l2	0	Weight decay of the optimizer
batch_size	100	Batch size
regenerate	0	Whether to read data again and regenerate intermediate files

Performance

The table below lists the results of some representative models in ASSISTments 12-13 dataset.

Model	AUC	Time/iter	Time-aware	Temporal cross
DKT	0.7308	3.8s
DKT-Forgetting	0.7462	6.2s	√
KTM	0.7535	49.8s	√
AKT-R	0.7555	13.8s	√
HawkesKT	0.7676	3.2s	√	√

Current running commands are listed in run.sh. We adopt 5-fold cross validation and report the average score (see run_exp.py). All experiments are conducted with a single GTX-1080Ti GPU.

Citation

Please cite our paper if you use our codes. Thanks!

@inproceedings{wang2021temporal,
  title={Temporal cross-effects in knowledge tracing},
  author={Wang, Chenyang and Ma, Weizhi and Zhang, Min and Lv, Chuancheng and Wan, Fengyuan and Lin, Huijie and Tang, Taoran and Liu, Yiqun and Ma, Shaoping},
  booktitle={Proceedings of the 14th ACM International Conference on Web Search and Data Mining},
  pages={517--525},
  year={2021}
}

Contact

Chenyang Wang ([email protected])