PEBG (PyTorch)

This project is now organized into 3 steps only:

1. Data Preprocessing
2. Pretraining (PEBG)
3. Training (PEBG+DKT)

Supported datasets:

• assist09
• assist17
• statics2011
• xes3g5m

Requirements

• Python 3.8+
• PyTorch
• NumPy
• SciPy
• pandas
• scikit-learn
• tqdm

Install:

pip install torch numpy scipy pandas scikit-learn tqdm

Step 1: Data Preprocessing

Use the unified preprocessing entry:

python preprocess.py --dataset <dataset_name>

Examples:

python preprocess.py --dataset assist09
python preprocess.py --dataset assist17
python preprocess.py --dataset statics2011
python preprocess.py --dataset xes3g5m

What Step 1 does internally:

• Runs dataset-specific raw preprocessing script in Data/
• Converts intermediate files to PEBG format (prepare_dataset.py)
• Builds graph similarity files (extract.py)

Step 1 output directory:

• datasets/assist09
• datasets/assist17
• datasets/statics2011
• datasets/xes3g5m

Each prepared dataset directory includes:

• <dataset_name>.npz
• pro_feat.npz
• pro_skill_sparse.npz
• pro_pro_sparse.npz
• skill_skill_sparse.npz
• skill_id_dict.txt

Optional args for Step 1:

python preprocess.py --dataset assist17 --max-len 200 --min-len 3 --split all

Step 2: Pretraining (PEBG)

Run PEBG pretraining:

python pebg.py --data-dir datasets/<dataset_name> --epochs 200

Example:

python pebg.py --data-dir datasets/assist17 --epochs 200

Outputs:

• datasets/<dataset_name>/embedding_200.npz
• datasets/<dataset_name>/pebg_model/pebg_*.pt

For large datasets, start with faster settings:

python pebg.py --data-dir datasets/assist09 --epochs 20 --batch-size 2048 --neg-k 10 --skill-sample-size 512

Step 3: Training (PEBG+DKT)

Train downstream KT model:

python pebg_dkt.py \
  --data-dir datasets/<dataset_name> \
  --dataset-name <dataset_name> \
  --use-pretrain \
  --embedding-file embedding_200.npz \
  --epochs 200 \
  --save-root runs

Example:

python pebg_dkt.py --data-dir datasets/assist17 --dataset-name assist17 --use-pretrain --embedding-file embedding_200.npz --epochs 200 --save-root runs

Per epoch output:

• train loss
• test AUC
• test ACC

Saved model/logs (per dataset):

• runs/<dataset_name>/best_model.pt
• runs/<dataset_name>/metrics.csv

best_model.pt is selected by best test AUC.

Minimal End-to-End Example

Using assist17:

python preprocess.py --dataset assist17
python pebg.py --data-dir datasets/assist17 --epochs 200
python pebg_dkt.py --data-dir datasets/assist17 --dataset-name assist17 --use-pretrain --embedding-file embedding_200.npz --epochs 200 --save-root runs

Common Issues

1. Missing raw CSV files in Data/<DATASET_DIR>/

• Put dataset files into expected locations under Data/ first.

2. Missing embedding_200.npz

• Finish Step 2 before Step 3.
• If you used different pretrain epochs, set matching --embedding-file.

3. Slow pretraining on very large datasets

• Reduce epochs for smoke test
• Increase --batch-size if memory allows
• Reduce --neg-k and --skill-sample-size