How daehoum1/fisf works
Project type and purpose
• This is a research‐oriented Machine Learning codebase (“FISF”) for graph‐based feature imputation, implemented in PyTorch and PyTorch Geometric (PyG).
• It supports two main experimental workflows: semi-supervised node classification (run_node.py) and link prediction (run_link.py), based on the paper “Propagate and Inject…”.
File structure
• Top‐level scripts:
• run_node.py – orchestrates the node classification experiment
• run_link.py – orchestrates the link prediction experiment
• train.py – generic training loop (called by run_*.py)
• evaluation.py – metrics calculation
• seeds.py – experiment reproducibility
• Core modules:
• data_loading.py / data_utils.py – graph data ingestion, preprocessing & batch construction
• models.py – PyTorch model architectures (GNN encoders, decoders, imputation heads)
• fisf.py – implementation of the Propagate & Inject algorithm logic
• Utility modules:
• utils.py – logging, config parsing, checkpointing
• utils_link.py – link‐prediction–specific helpers
• External dependencies and config are implicitly defined by the README (torch==1.10.2, pyg==2.0.3).
Extract insights from the README
• Confirms two experiment types (node classification and link prediction).
• Emphasizes that this builds on Twitter’s Feature Propagation codebase.
Detail
• Dependencies on external frameworks (torch, pyg, feature‐propagation).
• Branching for the two workflows (node vs. link) early in the Experiment Layer.
– Shows all 12+ modules distinctly
– Illustrates how data flows from scripts through the FISF algorithm into evaluation
– Highlights separation of concerns and external integrations
– Can be extended later to include hyperparameter stores or deployment endpoints if needed.