gfcca

Graph-embedded fuzzy kernel canonical correlation analysis (GFCCA) for multimodal / multi-omics feature fusion and classification.

GFCCA extends regularized kernel CCA with:

• Intuitionistic fuzzy scoring — each sample gets a confidence weight derived from its distance to its class centroid (membership) and the proportion of other-class neighbours within a relative distance threshold (non-membership).
• Discriminative graph embedding — a class-aware graph regularizer (Lw - beta * Lb) pulls same-class samples together while pushing different-class samples apart in the fused space.

The fused canonical variables are concatenated into a joint feature matrix that can be fed into any scikit-learn classifier.

Installation

pip install gfcca

Requires Python 3.10+.

Quick start

import numpy as np
from sklearn.ensemble import RandomForestClassifier
from gfcca import build_gfcca_features

# Two views (modalities) of the same samples, plus labels
X1 = np.random.rand(100, 20)  # e.g. eye imaging features
X2 = np.random.rand(100, 50)  # e.g. SNP genotypes
y = np.random.randint(0, 3, 100)

F_train, models = build_gfcca_features([X1, X2], y)
clf = RandomForestClassifier().fit(F_train, y)

Package layout

Experiments

The full multi-modal classification pipeline (5-modality diabetes cohort, CV + SHAP interpretability) lives in experiments/ and is not part of the PyPI package:

uv sync --group experiments
uv run --group experiments python experiments/main.py \
    --modalities eye fur methy pbmc snp \
    --feature gfcca --classifier rf --shap --seed 42

See experiments/datasets/README.md for the expected data layout.

Development

uv sync --group dev
uv run ruff check .
uv run ruff format .
uv run pytest tests/
uv run mypy src/gfcca

License

Apache-2.0