RAM-GNN

Relation-Aware Multiplex Graph Neural Network

Requirements & Setup

This project uses Python's built-in venv for managing the virtual environment instead of Conda.
Please follow the steps below to set up the environment using requirements.txt.

1. Install Python (≥ 3.13.5)

Ensure that Python version 3.13.5 or later is installed on your system.
You can download it from the official Python website: https://www.python.org/downloads/

2. Clone the Repository

Use Git to clone the repository and navigate into the project directory:

git clone https://github.com/sybeam27/RAM-GNN.git
cd RAM-GNN

3. Create and Activate a Virtual Environment

Use Python's venv module to create and activate a virtual environment:

# Create virtual environment
python3.13 -m venv venv

# Activate the virtual environment (Linux/macOS)
source venv/bin/activate

# Or activate it on Windows
venv\Scripts\activate

4. Install Required Dependencies

Install all required Python packages using the provided requirements.txt file:

pip install --upgrade pip
pip install -r requirements.txt

Datasets

Below are the details and download links for datasets used in our experiments:

[Dataset details]

1. O*NET (Download): O*NET is a standardized U.S. Department of Labor database covering nearly 1,000 occupations. It provides detailed descriptors—skills, abilities, knowledge, work activities—based on hundreds of validated items in a content model.

2. BLS (OES/OEWS) (Download): The BLS Occupational Employment & Wage Statistics (OEWS, formerly OES) offers semi‑annual estimates of employment and wages across 800+ U.S. occupations, segmented by industry and region, from non‑self‑employed workers data.

3. Automation Risk (Download): This dataset provides automation probability scores for 702 U.S. occupations, based on Frey & Osborne’s study using a Gaussian process classifier, which found that approximately 47% of occupations were at high risk of automation.

4. AIOE (AI Occupational Exposure) (Download): The AIOE dataset by Felten, Raj, and Seamans (2021) links 52 O*NET‑based job abilities to 10 AI application types, producing occupation‑level AIOE scores, plus industry (AIIE) and geographic (AIGE) exposures.

Run Models

Run Detection (Shell Script)

./detection.sh

Run Detection (Python Script)

python ./utils/detection.py --model <model_name> 

Optional Arguments

  --model               (Required) Model name to use (e.g., Linear, RandomForest, XGBoost, MLP, GCN, GAT, GraphSAGE, RGCN, HAN, MuxGNN, RAM-GNN)
  --graph_path          Path to the saved graph_data.pt file (default: ../data/detection/graph_data.pt)
  --save_path           Directory to save the trained model (default: ../model/detection/)
  --res_path            Directory to save evaluation results (default: ../res/detection/)
  --hidden_dim          Hidden layer dimension for neural network models (default: 64)
  --num_classes         Number of output classes (default: 2)
  --num_layers          Number of GNN layers or MLP layers (default: 2)
  --dropout             Dropout rate for regularization (default: 0.3)
  --lr                  Learning rate (default: 1e-3)
  --weight_decay        L2 regularization factor (default: 1e-4)
  --epochs              Number of training epochs (default: 500)

Run Prediction (Shell Script)

./prediction.sh

Run Prediction (Python Script)

python ./utils/prediction.py --model <model_name> 

Optional Arguments

  --model               (Required) Model name to use (e.g., Linear, RandomForest, XGBoost, MLP, GCN, GAT, GraphSAGE, RGCN, HAN, MuxGNN, RAM-GNN)
  --graph_path          Path to the saved graph_data.pt file (default: ../data/prediction/graph_data.pt)
  --save_path           Directory to save the trained model (default: ../model/prediction/)
  --res_path            Directory to save evaluation results (default: ../res/prediction/)
  --hidden_dim          Hidden layer dimension (default: 64)
  --num_layers          Number of GNN or MLP layers (default: 2)
  --dropout             Dropout rate for regularization (default: 0.3)
  --lr                  Learning rate (default: 1e-3)
  --weight_decay        Weight decay (L2 regularization factor) (default: 1e-4)
  --epochs              Number of training epochs (default: 500)