Use scipy's decompositions in PCA. (#1080)

* Replace numpy with scipy decomps. * add changelog entry --------- Co-authored-by: Sebastian Raschka <[email protected]>
rasbt · Mar 31, 2024 · 63a2655 · 63a2655
1 parent 8592cc7
commit 63a2655
Show file tree

Hide file tree

Showing 2 changed files with 6 additions and 4 deletions.
diff --git a/docs/sources/CHANGELOG.md b/docs/sources/CHANGELOG.md
@@ -31,7 +31,8 @@ The CHANGELOG for the current development version is available at
 
 ##### Changes
 
-- Add `n_classes_` attribute to stacking classifiers for compatibility with scikit-learn 1.3 ([#1091](https://github.com/rasbt/mlxtend/issues/1091)
+- Add `n_classes_` attribute to stacking classifiers for compatibility with scikit-learn 1.3 ([#1091](https://github.com/rasbt/mlxtend/issues/1091))
+- Use Scipy's instead of NumPy's decompositions in PCA for improved accuracy in edge cases ([#1080](https://github.com/rasbt/mlxtend/issues/1080) via [[fkdosilovic](https://github.com/rasbt/mlxtend/issues?q=is%3Apr+is%3Aopen+author%3Afkdosilovic)])
 
 
 

diff --git a/mlxtend/feature_extraction/principal_component_analysis.py b/mlxtend/feature_extraction/principal_component_analysis.py
@@ -7,6 +7,7 @@
 # License: BSD 3 clause
 
 import numpy as np
+import scipy as sp
 
 from .._base import _BaseModel
 
@@ -152,14 +153,14 @@ def _covariance_matrix(self, X):
 
     def _decomposition(self, mat, n_samples):
         if self.solver == "eigen":
-            e_vals, e_vecs = np.linalg.eig(mat)
+            e_vals, e_vecs = sp.linalg.eig(mat)
         elif self.solver == "svd":
             # Only SVD on mean centered data is equivalent to
             # PCA via covariance matrix (note that computing
-            # the covariance matrix will implicitely center
+            # the covariance matrix will implicitly center
             # the data)
             mat_centered = mat - mat.mean(axis=0)
-            u, s, v = np.linalg.svd(mat_centered.T)
+            u, s, _ = sp.linalg.svd(mat_centered.T, lapack_driver="gesvd")
             e_vecs, e_vals = u, s
             e_vals = e_vals**2 / (n_samples - 1)
             if e_vals.shape[0] < e_vecs.shape[1]: