How edogawa-liang/113-2_research works
Project Type and Purpose
• This is a research‐oriented pipeline for graph‐based machine learning.
• Purpose:
• Stage 1: train baseline GNN/MLP on graph datasets (Cora, Citeseer, Amazon).
• Tools & analysis for distance metrics, coverage, plotting, and result summarization.
Top‐Level Components
a. Data Ingestion & Preprocessing
• Modules: data/dataset_loader.py, data/data_modifier.py, data/prepare_split.py, data/split_unknown_to_test.py, data/feature2node.py, data/node2feature.py
• Purpose: load raw graph, split into train/val/test, optionally convert each feature into a new node (feature‐to‐node).
b. Stage 1 – Baseline Model Training
• Entry point: training_main.py
• Modules: trainer/gnn_trainer.py (for GCN), trainer/mlp_trainer.py (for MLP)
• Models: models/basic_GCN.py, models/basic_mlp.py
• Entry point: stage2_expsubg.py
• Subcomponents (in subgraph_selector/):
• random_selector.py
• random_walk_selector.py
• explainer_selector.py (wraps GNNExplainer, CFExplainer)
• cf_remaining_graph.py, remaining_graph.py
• CFExplainer code under models/cf_explanation/
• Tools: tools/stage2_node_share.py, tools/check_cf_summary2.py
• Purpose: summarize which nodes are shared across explanations, verify explainer success.
e. Stage 3 – Training on Remaining Graph
• Entry point: train_remaining_main.py (and test_remaining_main.py)
f. Analysis & Visualization
• Modules in analysis/: scripts to compute distances (distance_from_test_to_selected.py, etc.), degree change tests, coverage tests.
• Tools in tools/: plot_stage2.py, result_summary.py, cal_node_imp.py.
• Final results saved as Excel in “Final Result/0601”, “Final Result/0614”.
g. Command Scripts
• command/*.sh and command*.txt orchestrate multi‐dataset runs, parameter sweeps, feature-to-node experiments.
Relationships & Data Flow
• Raw Dataset → Data Ingestion → Split into train/val/test →
→ Stage 1 Trainer → Trained Baseline Model →
→ Stage 2.5 Storage & Summaries →
→ Performance Metrics → Analysis & Visualization → Excel/Plots in Final Result.
Architectural Patterns & Principles
• Pipeline stages (linear, modular)
• Separation of concerns: data vs. model vs. explanation vs. analysis
• Reusability: the same trainer and model code is reused in Stage 1 and Stage 3