How cgao-comp/gamc works
Project type and purpose
• GAMC is a machine-learning pipeline for unsupervised fake-news detection. It builds graph representations of news items, trains a graph autoencoder with masking and contrastive learning, and evaluates anomaly/fake-news scores.
File structure
• Top-level
• configs.yml: hyperparameters, model & dataset settings
• main_new.py (or main_graph.py): entry point that parses arguments, loads config, kicks off training/evaluation
• code/gamc package
• datasets/
– data_util.py: functions to load raw data, build DGL graphs
• models/
– edcoder.py: encoder + decoder logic for graph autoencoder
– gat.py, gin.py: Graph Attention Network & Graph Isomorphism Network backbones
– loss_func.py: masking / contrastive loss implementations
• evaluation.py: computes metrics / fake-news scores
• utils.py: logging, seed setting, helper routines
README for insights
• Dependencies: Python, PyTorch, DGL, PyYAML
• Invocation: python main_graph.py --dataset DATASETNAME --use_cfg
• External datasets hosted on Google Drive
Detail
• Break the pipeline into as many discrete steps as make sense (e.g., raw CSV loader, feature normalization, graph builder)
• Config parsing as its own component
• Both training and evaluation branches