EdGCL: Disentangling Social and Cognitive Homophily in Graph-based Educational Recommender Systems

This repository contains the official implementation of our paper "EdGCL: Disentangling Social and Cognitive Homophily for Educational Recommendation", accepted by AAAI 2026 Special Track on AI for Social Impact.

EdGCL is a graph-based framework designed to model learning motivations from both social and cognitive perstictives.

Environments

We provide a requirements.txt file to specify all necessary dependencies for reproducibility. The experiments were conducted on CUDA 12.1.0 and conda 24.5.0. The main package versions are as follows:

• python=3.9.1
• scipy=1.13.1
• numpy=2.0.2
• torch=2.4.0+cu121
• tqdm=4.67.1

Code Structure

EdGCL
├── data.zip # datasets
├── model
│   ├── gt_base.py
│   ├── gt_hetero.py
│   ├── model.py
|   ├── tokenizer_base.py
|   ├── tokenizer_hetero.py
|   ├── transformer_layer.py
│   └── utils.py
├── early_stops.py
├── global_utils.py
├── metric.py
├── README.md
└── run.py

Dataset

We provide the datasets to validate the effectiveness of our method. Dataset statistics are summarized as follows:

Dataset	MOOCCubeX	MOOPer
Students	10,893	6,057
Resources	66,891	3,784
Social Edges	3,589,654	1,316,667
Social Graph Density	3.03%	3.59%
Learning Edges	2,150,392	422,602
Learniong Graph Density	0.29%	1.63%
Avg. Learning Sequence Length	191	68

Train and Test

Please unzip the dataset before running the code.

unzip data.zip

Then, run the model using the following command.

python run_edgcl.py -d mooper -hop 5 -cog_coef 0.05 -lp_coef 0.1 -ssl_coef 1e-4