Official PyTorch implementation of “Hint-Conditioned Knowledge Tracing: A Causal-Inspired View of Context-Dependent Hint Usage,” published at the 2026 IEEE International Conference on Data Mining (ICDM 2026).
HCKT is a hint-conditioned knowledge tracing framework that models hint-taking as a context-dependent learning-process signal under a causal-inspired chronological view. It constructs a behavior-enriched proxy context, jointly estimates hint-taking tendency and hint-conditioned response patterns, and injects the resulting model-estimated discrepancy between hint and non-hint conditions into the KT backbone through gated feature fusion and hint-aware attention modulation.
Motivating example: similar hint-requested correct responses can correspond to different latent mastery trajectories.
.
├── data/
│ ├── README.md # Data format and preparation notes
│ └── data_preprocessing.ipynb # Reference preprocessing and simulation code
├── src/
│ ├── main.py # Training and evaluation entry point
│ ├── overall_perform.sh # Real-world experiment command manifest
│ ├── simulation.sh # Synthetic experiment command manifest
│ ├── run_exp.py # Overall-experiment runner
│ ├── run_exp_simulation.py # Simulation runner
│ ├── helpers/
│ │ ├── DataReader.py # Sequence construction and fold generation
│ │ └── KTRunner.py # Training, evaluation, and metric reporting
│ ├── models/
│ │ ├── HCKT.py # Proposed model
│ │ ├── BaseModel.py # Shared model interface
│ │ └── ... # Baselines used by the experiment manifests
│ └── utils/
│ └── utils.py
├── hckt_motivation.png # Motivating example used above
├── requirements.txt
└── README.md
The included comparison implementations are DKT, AKT, DKVMN, LBKT, simpleKT, FoLiBiKT, HawkesKT, RobustKT, and LEFOKT-AKT.
Create an isolated Python environment and install the dependencies:
python -m venv .venv
python -m pip install --upgrade pip
python -m pip install -r requirements.txtActivate the environment using the command appropriate for your operating system before running experiments. Training runs on CUDA automatically when a CUDA-enabled PyTorch installation and an available GPU are detected; otherwise it runs on CPU.
Raw datasets are not redistributed in this repository. Prepare each dataset as a tab-separated interactions.csv file under its experiment name:
data/
├── assist09/interactions.csv
├── assist12/interactions.csv
├── assist17/interactions.csv
├── junyi/interactions.csv
├── statics/interactions.csv
└── <synthetic-dataset>/interactions.csv
The dataset aliases must match the --dataset values in the command manifests. See data/README.md for the required fields and preprocessing notes. The first run creates a cached Corpus_<max_step>.pkl beside the processed interaction file.
All commands below assume that the current directory is src, because the experiment code uses paths relative to that directory.
cd src
python main.py \
--model_name HCKT \
--dataset assist09 \
--max_step 50 \
--emb_size 64 \
--num_layer 2 \
--num_head 8 \
--lr 3e-3 \
--l2 5e-5 \
--dropout 0.01 \
--lambda_prop 0.3 \
--lambda_outcome 0.05 \
--random_seed 2025 \
--kfold 5 \
--fold 0This command trains one fold. Change --fold from 0 to 4 for the five folds, or use the bundled runners below to execute a complete command manifest.
Per-run logs are written to log/<model>/, model checkpoints to model/<model>/, and runner summaries to CSV files under log/. These generated artifacts are intentionally excluded from version control.
Part of this codebase is adapted from and extended upon HawkesKT, the official implementation of Temporal Cross-Effects in Knowledge Tracing (WSDM 2021). We thank the original authors for their open-source contribution.
If you use this code, please cite:
Hint-Conditioned Knowledge Tracing: A Causal-Inspired View of Context-Dependent Hint Usage. IEEE International Conference on Data Mining (ICDM), 2026.
The complete BibTeX entry should be taken from the final proceedings record.