AttEAGNN: Attention Based Edge-Aware GNN Applied to Network Load Prediction

Requirements

• python3
• pip

Initial Configuration

Install project dependecies:

pip install -r requirements.txt

Project Structure

The project structure is organized as follow:

• data: Contains the RNP graph, traffic data and geographical data.
• results: Contains the model weights, training information. Additionally, it includes comparative graphics between the models using different metrics.
• src: Contains the source code.
• configs: Contains the configurations used for each model.

Reproducibility

Training and Evaluation Flow of the Models

The training and evaluation of the models consist of the following steps:

1. First, the models are configured using the Config class in src/config.py and then added to the train_models.py file. Sample configurations can be viewed in the configs/ directory.

2. The configurations are loaded using the ModelLauncher class in src/model_launcher.py. This class creates the model with the respective configurations and then performs training, saving the results such as losses, final load prediction, model weights, and model name in the results/ directory.

3. Finally, the script plot_graphics.py is executed. The data saved in the previous step is loaded, and metrics such as MAE (Mean Absolute Error), MSE (Mean Squared Error), and R²(Coefficient of Determination) are calculated. The graphics containing these metrics are saved in the results/ directory.

Reproduction of the Experiments

To simplify the reproduction of the experiments, a shell script (run_experiments.sh) has been provided in the root folder of the project. This script will execute the Python script train_models.py, and at the end, the script plot_graphics.py is run to evaluate the models and generate the graphics. The training progress can be monitored through the terminal.

To execute the shell script, make sure it has execution permission, and if so, run:

./run_experiments.sh

If this file does not have execution permission, execute the command below, and then run the previous command:

chmod +x run_experiments.sh

Tool Customization

Train, Evaluate, and Compare Your Own Model

Our tool allows the inclusion of custom GNN models for training and testing on the RNP network by following these steps:

1. Create your model in src/models/.

2. Define your model and its configuration in the configs/ directory using the Config class. Note that you should choose a class to encapsulate your data.

   2.a. If your model is a traditional GNN, i.e., it only needs node features and edge indices, you can use the GNNData class.

   2.b. If it also uses edge features, use the EdgeGNNData class.

   2.c. Otherwise, you can create your own data class from the DataInterface interface, as done for the CAGNNData class. These classes can be found in src/entities.

   2.d. Note also that you need to define a function to load the network graph data, which will be used to initialize the data class (an extension of DataInterface). Use the implemented configurations as a reference.

3. Add your configuration class to train_models.py.

4. If you have already run the experiments, execute the training in train_models.py only for your model and run the plot_graphics.py script. Otherwise, run the run_experiments.sh including your model.

5. View the graphics in the results/ folder.