KTbench

KTbench provides solutions to the label leakage problem in Knowledge Tracing models.

This project is based on the following papers:

• Addressing Label Leakage in Knowledge Tracing Models
• Enhancing Knowledge Tracing through Leakage-Free and Recency-Aware Embeddings

Installing

Create a virtual environment then run the following from the project folder:

pip install .

Reproducing the label leakage paper

Run the following from the project folder:

python baseline/run.py

By default a ".ktbench" folder is created, containing the experiment logs:

.ktbench
└── dataset_name
    └── model_name
        └── training_time_stamp
            ├── test.yaml          # Contains results on the test set.
            └── valid_fold_k.yaml  # Contains validation results on the kth fold during training.

Usage

An example of training and evaluating a DKT with basic KC-expanded sequence

from ktbench import Pipeline, bench_model
from dataclasses import dataclass
from ktbench.model.dkt.dkt import DKT

@dataclass
class Cfg:
    model_cls = DKT
    window_size: int = 150
    is_unfold = True
    eval_method = Pipeline.EVAL_UNFOLD_KC_LEVEL
    kfold = 5

@dataclass
class Traincfg:
    batch_size = 128
    eval_batch_size = 128
    n_epoch = 100
    lr = 0.001

bench_model(Cfg(), Traincfg(), datasets = ['assist2009'])

Resources

Useful tools for knowledge tracing:

• EduStudio: A Unified Library for Student Cognitive Modeling including Cognitive Diagnosis (CD) and Knowledge Tracing (KT).
• pykt-toolkit: A Python Library to Benchmark Deep Learning based Knowledge Tracing Models

Some data preprocessing and model implementations were adapted from these tools.