Introduction

This repository contains code for the experiments in the paper Statistical Estimation from Dependent Data [arxiv]. The code base is derived from the code base of the following paper: Graph Random Neural Network for Semi-Supervised Learning on Graphs [arxiv]. There are three graph datasets we work with:

• Cora
• Citeseer
• Pubmed

We study the improvements obtained by the maximum pseudo-likelihood estimator (MPLE) analyzed in our paper on inference from graph-dependent data. We also compare the performance of MPLE with a contemporary GNN based approach for the problem and observe competitive performance (as of February 2021).

Requirements

• Python 3.7.3
• Please install other pakeages by pip install -r requirements.txt

Code Structure

The train_*.py scripts perform the training of different variants on any of the three datasets. The run_*.sh and run100_*.sh bash scripts have been set up to call the appropriate train_*.py based on the command line arguments passed. The run100_*.sh scripts perform multiple Monte-Carlo executions across a range of hyperparameter settings. The output of the runs is set to be recorded in folders cora/, pubmed/, citeseer/. result_*.py files compile the results of the Monte-Carlo runs generated above and generate plots.