tppis

Open-source Python implementation of Truncated Preconditioned Profiled Independence Screening (TPPIS) from Tanaka and Matsui (2023), plus the baselines SIS, FPSIS (and FPSIS-BIC), and PPIS.

The method screens variables in high-dimensional regression when predictors are strongly multicollinear. A factor-analysis transform removes the common factors; TPPIS then truncates the tail of that transform so the unique-factor signal is not washed out.

Maintainer: Mohammed Aswath M · mohammed.aswath07@gmail.com · issues: github.com/Mohammed-Aswath/tppis

Install

pip install tppis

The package requires Python 3.8+. On 3.9+ it uses scikit-learn 1.6+; on 3.8 it uses scikit-learn 1.3. The TPPIS API is the same.

From a clone of this repository:

python -m venv .venv
source .venv/bin/activate   # Windows: .venv\Scripts\activate
pip install -e ".[dev]"

Requires Python 3.8+, NumPy, SciPy, and scikit-learn.

Quickstart

from tppis import TPPIS, make_example1

data = make_example1(n=100, p=80, phi=0.7, seed=0)
est = TPPIS().fit(data.X, data.y)

est.get_support(indices=True)   # selected columns, ranking order
est.d_, est.alpha_, est.k_      # BIC-chosen parameters
est.bic_
est.grid_                       # every (d, alpha, k, bic) evaluated

Fixed parameters, no search:

TPPIS(d=20, alpha=0.4, k=10).fit(data.X, data.y)

Inside a scikit-learn pipeline:

from sklearn.linear_model import LinearRegression
from sklearn.pipeline import Pipeline
from tppis import TPPIS

pipe = Pipeline([
    ("select", TPPIS(d=8, alpha=0.6, k=10)),
    ("lm", LinearRegression()),
])
pipe.fit(data.X, data.y)

The functional entry point screen(X, y, method="tppis") returns the same fitted estimator.

When to use which method

TPPIS at alpha=1 is exactly PPIS. That identity is tested.

Citation

If you use the method, cite the paper. If you use this software, cite the package as well.

Tanaka, S. and Matsui, H. (2023). Variable screening using factor analysis
for high-dimensional data with multicollinearity. arXiv:2306.05702.

See CITATION.cff.

License

MIT. See LICENSE.