IEKT: 8/2 Data Processing and Training

This repo trains IEKT with a strict 80% train / 20% test split (no validation set).

1. Requirements

• Python 3.8+
• torch
• pandas
• numpy
• tqdm
• scikit-learn

Install:

pip install torch pandas numpy tqdm scikit-learn

2. Raw Data Paths

Place raw CSV files at:

• Data/ASSIST2017/anonymized_full_release_competition_dataset.csv
• Data/Statics2011/AllData_student_step_2011F.csv

3. Preprocess (8/2, No Valid)

Preprocessing outputs:

• history_train.pkl
• history_test.pkl
• problem_skill_maxSkillOfProblem_number.pkl

3.1 ASSIST2017

python Data/preprocess_iekt_assist2017.py \
  --input_csv Data/ASSIST2017/anonymized_full_release_competition_dataset.csv \
  --output_dir Data/iekt_assist2017 \
  --min_seq_len 10 \
  --max_concepts 5 \
  --train_ratio 0.8 \
  --seed 42

3.2 Statics2011

python Data/preprocess_iekt_statics2011.py \
  --input_csv Data/Statics2011/AllData_student_step_2011F.csv \
  --output_dir Data/iekt_statics2011 \
  --min_seq_len 10 \
  --max_concepts 5 \
  --train_ratio 0.8 \
  --seed 42

4. Remove Old Cached Dataset Objects

If you previously created dataset_valid.pkl or old cached datasets, remove them before retraining:

rm -f Data/iekt_assist2017/dataset_*.pkl
rm -f Data/iekt_statics2011/dataset_*.pkl

5. Train

main.py now loads only train and test.

5.1 ASSIST2017

python main.py \
  --run_dir runs/assist2017 \
  --data_dir Data/iekt_assist2017/ \
  --model iekt \
  --device 0 \
  --n_epochs 300 \
  --batch_size 32 \
  --lr 1e-3 \
  --seq_len 200

5.2 Statics2011

python main.py \
  --run_dir runs/statics2011 \
  --data_dir Data/iekt_statics2011/ \
  --model iekt \
  --device 0 \
  --n_epochs 300 \
  --batch_size 32 \
  --lr 1e-3 \
  --seq_len 200

CPU example:

python main.py --run_dir runs/cpu --data_dir Data/iekt_assist2017/ --device -1

6. Per-Epoch Output

Each epoch logs:

• test_auc
• test_acc
• loss

Example:

Epoch: 010, Loss: 0.1234567, test_auc: 0.8123456, test_acc: 0.7456789

Log file:

• <run_dir>/log.txt