This repository hosts the implementation of the knockoff filter method for controlled variable selection, based on the "Controlling the False Discovery Rate via Knockoffs" paper from 2015. The method is designed for high-dimensional data settings to effectively control the false discovery rate while preserving statistical power.
Note: Much of this implementation was crafted either from scratch or without relying on high-level libraries. This approach was chosen primarily for educational purposes (mostly self-educational purposes), allowing for a deeper understanding and exploration of the underlying algorithms. While this method involves some "reinventing the wheel," it might have some educational value. Future development may include the integration of more specialized libraries to enhance functionality and performance.
- Implementation of the knockoff filter method for feature selection.
- Synthetic and GWAS-like data generators for evaluation and testing.
- Lasso regression integration for feature importance assessment.
KnockoffOrigins
is available on PyPI and can be installed using either pip or Poetry.
You can install KnockoffOrigins
directly using pip:
pip install KnockoffOrigins
This command will download and install the latest version of KnockoffOrigins along with its dependencies. Using Poetry
If you are using Poetry for your project, you can add KnockoffOrigins to your project as follows:
poetry add KnockoffOrigins
This will handle the installation and also update your pyproject.toml and poetry.lock files to reflect the change.
If you prefer to install from source or want to contribute to the package, first ensure Poetry dependency management is installed:
pip install poetry
Then clone the repository and install the dependencies:
git clone https://github.com/jrazi/KnockoffOrigins.git
cd KnockoffOrigins
poetry install
To generate knockoff features based on your original data, you can use the generate_knockoff_features
function from the knockofforigins.knockoff_construct
module.
import numpy as np
from knockofforigins.knockoff_construct import generate_knockoff_features, choose_s_vector
# Load your original feature matrix X (n x p)
# ...
# Choose the s vector for knockoff construction
s = choose_s_vector(np.cov(X.T))
# Generate knockoff features
X_knockoff = generate_knockoff_features(X, s)
After generating knockoff features, you can perform feature selection using the Lasso regression model with the augmented design matrix (original and knockoff features concatenated).
from knockofforigins.lasso import compute_feature_importance
# Load your response vector y (n x 1)
# ...
# Compute feature importance statistic W
alpha = 0.1 # Regularization parameter for Lasso
W = compute_feature_importance(X, X_knockoff, y, alpha)
# Select features based on W
selected_features = np.argsort(-W)[:num_features_to_select]
If you need to generate synthetic GWAS data for testing purposes, you can use the InfluentialFeatureGWASDataGenerator class.
from KnockoffOrigins.knockoff_construct import generate_knockoff_features, choose_s_vector
from KnockoffOrigins.gram_matrix import generate_gram_matrix
# Calculate the covariance matrix Sigma of the original features X
Sigma = generate_gram_matrix(X)
# Choose the vector 's' for knockoff feature generation
s = choose_s_vector(Sigma)
# Generate knockoff features using the original features X
X_knockoff = generate_knockoff_features(X, s)
Contributions are welcome, and appreciated!
This project is licensed under the MIT License.
- Fix problems with computing S vector and generating Knockoff data.
- Implement test statistics for feature evaluation.
- Develop FDR control mechanisms as outlined in the original study.
- Implement Lasso feature selection using lower-level libraries.
- Address some of the bugs and implementation issues.
- Create example notebooks demonstrating package usage.
- Replicate experiments from the original 2015 knockoff paper.
- Develop visualization methods for feature selection analysis.
- Implement the KnockOff+ method for enhanced feature selection.