Skip to content

Latest commit

 

History

History
90 lines (69 loc) · 4.51 KB

CLAUDE.md

File metadata and controls

90 lines (69 loc) · 4.51 KB

CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

Project Overview

ICTC (Intra-class Connection based Triadic Closure) is a research implementation of bipartite link prediction methods using graph neural networks and matrix factorization. It implements and compares three approaches: ICTC (GAE+LGAE combination), BiSPM, and SRNMF.

Python 3.0+ required. No build system, test framework, or linter is configured.

Running the Models

Before running, select a dataset by uncommenting the desired line in ictc/config.py (and commenting out others). Available datasets: gpcr, enzyme, ionchannel, malaria, drug, sw, nanet, movie100k.

# ICTC (runs GAE, LGAE, and combined ICTC)
python train.py

# BiSPM (Bipartite Spectral Perturbation Method)
python run_bispm.py

# SRNMF (set args.similarity in run_srnmf.py first: 'srnmf_cn', 'srnmf_jc', or 'srnmf_cpa')
python run_srnmf.py

Dependencies

torch, networkx, scipy, scikit-learn, numpy, matplotlib, seaborn, nimfa, sparsesvd. No requirements.txt exists.

Package Structure

ICTC/
├── data/                        # Datasets
├── ictc/                        # Main package
│   ├── __init__.py
│   ├── config.py                # Hyperparameters and dataset selection (module-level variables)
│   ├── models/
│   │   ├── __init__.py
│   │   ├── gae.py               # GAE, LGAE, VGAE classes + GraphConvSparse, glorot_init, dot_product_decode
│   │   ├── bispm.py             # getBrAndBtriangle(), getBiSPM() functions
│   │   └── srnmf.py             # getXY() function
│   ├── data/
│   │   ├── __init__.py
│   │   ├── loading.py           # Data loaders: load_data, load_data_drug, load_data_citation, etc.
│   │   └── preprocessing.py     # Graph preprocessing, edge splitting, get_data()
│   └── evaluation/
│       ├── __init__.py
│       └── metrics.py           # Scoring: get_scores, get_precision, similarity scores (CN, JC, AA, CPA)
├── train.py                     # Entry point for ICTC
├── run_bispm.py                 # Entry point for BiSPM
├── run_srnmf.py                 # Entry point for SRNMF
├── data_analysis.py             # Standalone analysis script
├── CLAUDE.md
└── README.md

Architecture

Configuration (ictc/config.py)

All hyperparameters and dataset selection live here as module-level variables. Other modules import via from ictc import config as args. Key settings: dataset, model1/model2, learning_rate, num_epoch, hidden1_dim/hidden2_dim, numexp (number of experiment repetitions), num_test (test set ratio), device (CUDA device ID).

Models (ictc/models/gae.py)

Three graph autoencoder variants, all using GraphConvSparse layers and dot_product_decode (sigmoid of Z @ Z^T):

  • VGAE: Variational, samples from learned Gaussian (mean + logstd). Adds KL divergence to loss.
  • GAE: Deterministic two-layer encoder (ReLU then linear).
  • LGAE: Single linear layer encoder.

Training Pipeline (train.py)

run() is the entry point (guarded by if __name__ == '__main__'). For each of numexp experiments:

  1. Loads data and splits edges into train/val/test
  2. Trains model2 (GAE) and model1 (LGAE) via learn_train_adj()
  3. Combines predictions: A_pred = (adj_norm @ A_lgae + transpose) / 2
  4. Evaluates with ROC-AUC, Average Precision, Precision@k
  5. Reports mean +/- standard error across all experiments

State is managed through global variables shared between run() and learn_train_adj().

Data Pipeline

  • ictc/data/loading.py: Loaders for bipartite edge lists (load_data), drug data (load_data_drug), citations (load_data_citation), and other formats. Datasets stored in data/bipartite/.
  • ictc/data/preprocessing.py: preprocess_graph() does symmetric normalization (D^{-1/2} A D^{-1/2}). mask_bipartite_perturbation_test_edges() splits edges into train/val/test with negative sampling. get_data() orchestrates loading and splitting.

Evaluation (ictc/evaluation/metrics.py)

  • get_scores(): ROC-AUC and Average Precision
  • get_precision(): Precision@k on top predictions
  • Similarity scores for SRNMF: Common Neighbors, Jaccard, Adamic-Adar, CPA

Alternative Methods

  • ictc/models/bispm.py: SVD-based matrix completion with spectral perturbation.
  • ictc/models/srnmf.py: NMF with similarity regularization using multiplicative update rules.