GRL-mini-project-2023

Welcome to the git repository for my GRL mini project. Using this codebase, I compared the ability of two architectures to capture correlation between the input features and the target labels in a node classification task. The first studied architecture is GATv2 () \cite{GATv2}, consisting of GATv2 layers interleaved with nonlinearities. The second studied architecture () is an extension thereof, consisting of a slightly modified type of layers interleaved with nonlinearities:

Repository structure

Files (code)

File	Description
dataset.py	Code for generating, loading, and saving the dataset.
gatv2.py	Implementation of the $\mathcal{A}$ and $\mathcal{B}$ single-layer and two-layer architectures.
experiment.py	Code for training each of the four models on all 10 sections of the dataset.
baseline.py	Evaluation of the argmax model on the datasets.
figures.ipynb	Jupyter notebook used to plot figures in the project report. Accuracies were copied from TensorBoard.

Folders (results)

• mini-study = a smaller-scale pilot study I conducted before doing any heavy experiments

  • datasets: Contains datasets for the mini-study.
  • runs: Contains binary log files for each run, viewable with TensorBoard.

• study1, study2, study3 = the three independent runs of the main study, from which the median accuracies were taken

  • datasets: Identical datasets for studies 1, 2, and 3.
  • runs: Contains binary log files for each run, viewable with TensorBoard.