How tianxiangzhao/graphsmote works
Project Type and Purpose
• This is a machine learning research project that implements GraphSMOTE. Its purpose is to address imbalanced classification in graph data, where nodes may belong to minority classes.
• The README explains that it supports experiments on various datasets (BlogCatalog, Cora, and Twitter) and provides multiple baseline models (e.g., vanilla, upsampling, reweighting, and embed-SMOTE) alongside two network architectures (GraphSage and GCN).
Analyzing the File Structure
• Top-level items include:
• README.md – documentation.
• A “checkpoint” directory – holds model checkpoints for different datasets (BlogCatalog and Cora), indicating storage of trained models.
• A “data” directory – expected location for dataset files. It includes placeholder folders for datasets and specific CSV files (e.g., twitter_fake_ids.csv).
• Several Python scripts:
◦ main.py – likely the entry point for training and evaluation.
◦ models.py – contains definitions of the graph neural network models and potentially different architectural choices.
◦ utils.py – contains helper functions including argument parsing (get_parser()) and possibly configuration management.
◦ data_load.py – dedicated to loading the external datasets, possibly including pre-processing and conversion steps.
• This indicates a modular design where data ingestion, model definition, execution logic, and utility functions are separated into distinct components.
Elements to Be Included
• Incorporate labels for different baseline configurations and indicate alternative paths for pretraining vs. fine-tuning.
• If possible, distinguish experimental settings (like oversampling, reweighting, etc.) as sub-options within the main training flow.