Evaluation of topological features’ influence on a node classification task using different GNN architectures
A detailed description of the task, methodology, and the performed experiments is provided in the report.pdf file.
The folder colab notebooks contains the three Google Colab notebooks we used to run all of the experiments related to this project. In particular, it contains:
- Experiments_GAT.ipynb: used to run the experiments for the GAT architecture;
- Experiments_GCN.ipynb: used to run the experiments for the GCN architecture;
- Experiments_linear.ipynb: used to run the experiments for the linear models Logistic Regression, SVM and Decision Tree.
To run a specific .ipynb file on your personal computer, just move it to the root folder and execute its cells. To run it on Google Colab you can do the following:
- upload the current repo to Google Drive under
drive/MyDrive/Colab Notebooks; - open the desired .ipynb with Google Colab;
- load the GDrive partition using the code:
from google.colab import drive
drive.mount('/content/drive')- change the working directory to the repo directory in GDrive with the command:
%cd drive/MyDrive/Colab\ Notebooks/{repo_name}Note: due to stochasticity present in some of the models we employed, results replicated with the notebooks above may have small differences from the ones in the original report.
The folder models contains the implementations of the GCN and GAT models in Pytorch Geometric, along with the MLP and ensemble methods used for topological features integration.
The folder features contains the code used to compute the structural features using NetworkX, and the positional features using an implementation of Node2Vec in Pytorch Geometric.
The folders additional and plots contains helper functions (e.g. datasets loading, normalization functions etc.) and plots of the results, respectively.