TransKT

Contrastive Cross-Course Knowledge Tracing via Concept Graph Guided Knowledge Transfer

Considering the relevance and transferability of learners' cross-course learning behaviors, we propose the TransKT framework to provide a more comprehensive and accurate understanding of learners' knowledge states for reliable prediction on the knowledge tracing task.

🌟Overview

The overview of TransKT. First, during preprocessing, TransKT employs LLM to complete the concept-to-concept links both within and between courses, bridging individual concept graphs. Next, TransKT utilizes a semantic-enhanced knowledge transfer module, which leverages an LLM-to-LM pipeline to summarize and extract rich semantic features for GCN-based knowledge propagation. Finally, TransKT employs a contrastive learning objective to enhance the robustness of representations for prediction by maximizing mutual information between single-course and cross-course knowledge states.

For more details, please read our paper: "Contrastive Cross-Course Knowledge Tracing via Concept Graph Guided Knowledge Transfer"

➡️Quick Start

Installation

Git clone this repository and creating conda environment:

conda create -n transkt python=3.8
conda activate transkt
pip install -r requirements.txt 

Create directory

Unzip the TransKT dataset files and create training, testing, and validation sets:

unzip C_DS.zip  # Take C&DS dataset for example.
python split.py

Establishing the directory for the output file of the optimal model:

mkdir trained

Establishing the directory for the output files of training and testing logs:

mkdir logs

Training & Testing

Our model experiments are conducted on a Tesla A800, with actual VRAM requirements of around 10 GB. You can execute it directly using Python commands.

CUDA_VISIBLE_DEVICES=3 python -u main.py --lr 0.0001 --batch_size 64 --model_id ours --num_epoch 200  --optim adamw  --IM 1 --theta1 0.2 --theta2 0.7  \
  --lambda 0.8 --eta 0.4  --warm_epoch 0 --GNNL 2 > log/test.log 2>&1&

You can also choose to directly execute the shell script in the command line:

bash start.sh

📌Detailed information

In this paper, the publicly available PTADisc dataset (sourced from the PTA) is further processed to obtain four sub-datasets, each specifically designed to support the Cross-Course Knowledge tracing (CCKT) task.

Dataset statistics

| Dataset | Java&Python | C&DS | C&Java | CS&MA | |---------------|-------------|--------|--------|--------| | #Records | 1,129,999 | 1,800,066 | 1,501,414 | 282,135 | #Learners | 7,770 | 12,275 | 9,618 |2,431 | #Quetions | 5,734/7,562 | 10,957/9,641 | 11,934/7,624 | 5,870/1,386 | #Concepts | 360/364 | 362/323 | 424/408 |359/140

🎉Acknowledgment

• We further processed the publicly available PTADisc dataset to obtain the cross-course knowledge tracing dataset.