space-graph 0.5.0

Pure Python SPACE: sparse partial correlation estimation

space-graph

Discover which variables in your dataset are directly related to each other, even after accounting for all other variables. Given a data matrix, SPACE estimates a sparse network of partial correlations -- connections that remain after removing indirect effects. Designed for settings where the number of variables can far exceed the number of samples (e.g. genomics).

Pure Python implementation of SPACE (Sparse Partial Correlation Estimation) from Peng et al. (2009), with no R or C dependencies.

Paper: Sparse Partial Correlation Estimation for High-Dimensional Data

Install

pip install space-graph

Optional Numba (faster inner jsrm loop):

pip install 'space-graph[numba]'

From GitHub:

pip install git+https://github.com/shahrozeabbas/space-graph.git

Usage

import numpy as np
from space_graph import SPACE

X = np.random.randn(20, 5)

model = SPACE(
    alpha=0.7,
    max_outer_iter=2,
    max_inner_iter=500,
    tol=1e-6,
    weight='uniform',
)

model.fit(X)

print(model.partial_correlation_)

Penalty

alpha is overall regularization strength (like sklearn). gamma (γ) in [0, 1] sets the mix: lam1 = alpha * gamma, lam2 = alpha * (1 - gamma). Default gamma=1 matches R space::space.joint default lam2 = 0.

Options

• gamma (default 1): mix γ in [0, 1] for lam1 / lam2; 1 matches R default lam2 = 0.
• tol (default 1e-6): inner coordinate-descent stopping tolerance (and active-set threshold), same scale as the reference C code.
• weight: default uniform (unit weights). Use equal as an alias. Other modes: sig, degree, or a custom positive vector of length p.
• backend (default auto): inner JSRM shooting loop. numpy always uses pure NumPy (no Numba import). auto uses Numba when installed (lazy on first fit), otherwise NumPy. numba requires Numba and raises ImportError if it is missing. The first fit with auto or numba may include JIT compilation time.

Tests

pytest

For a quick manual run after an editable install (or with PYTHONPATH=. from the repo root), see test_space_graph.py.

Optional: build libjsrm_test.so from ../space/src/JSRM.c to run the ctypes cross-check in tests/test_space.py.

License

GPL-3.0-or-later (same family as the original space R package).