如何基于sklearn PCA得到非均值中心化的成分与变换结果？

Awesome question—this is exactly the kind of tweak you need when you want to wrap PCA's inverse transform into a single matrix multiplication, no separate mean addition required. Let's walk through the linear algebra and practical implementation step by step.

First, Let's Formalize the Problem with Notation #

Let's define our variables clearly to avoid confusion:

• $X_{\text{raw}} \in \mathbb{R}^{N \times D}$: Your original non-centered data (N samples, D features)
• $\mu = \text{PCA.mean_} \in \mathbb{R}^{1 \times D}$: The mean vector sklearn automatically subtracts to center your data
• $Z = \text{PCA.fit_transform}(X_{\text{raw}}) \in \mathbb{R}^{N \times K}$: The transformed output from PCA (K = number of principal components)
• $W = \text{PCA.components_} \in \mathbb{R}^{K \times D}$: The principal component vectors from sklearn

We know sklearn's inverse transform uses an affine operation:

$\hat{X}_{\text{raw}} = Z \cdot W + \mu$

Your goal is to find $X_1$ (modified version of Z) and $W_1$ (modified version of PCA.components_) such that:
$\hat{X}_{\text{raw}} = X_1 \cdot W_1$

Linear Algebra Solution #

The key insight is to embed the mean offset into our transformed data and component matrix, turning the affine transform into a pure linear transform in a slightly higher-dimensional space.

Step 1: Modify the Transformed Data ($X_1$) #

Append a column of all ones to your PCA-transformed output $Z$. This adds a "constant feature" to our transformed space:

X1 = np.hstack([Z, np.ones((Z.shape[0], 1))])

In math terms:
$X_1 = \begin{bmatrix} Z & \mathbf{1}_N \end{bmatrix} \in \mathbb{R}^{N \times (K+1)}$
Where $\mathbf{1}_N$ is an $N \times 1$ vector of all ones.

Step 2: Modify the Principal Components ($W_1$) #

Stack the original PCA components with the mean vector $\mu$ as an additional row. This maps our new constant feature directly to the original data's mean:

components_1 = np.vstack([pca.components_, pca.mean_])

In math terms:
$W_1 = \begin{bmatrix} W \ \mu \end{bmatrix} \in \mathbb{R}^{(K+1) \times D}$

Step 3: Verify the Result #

Multiplying $X_1$ and $W_1$ gives exactly the same result as PCA.inverse_transform(Z):
$X_1 \cdot W_1 = \begin{bmatrix} Z & \mathbf{1}_N \end{bmatrix} \cdot \begin{bmatrix} W \ \mu \end{bmatrix} = Z \cdot W + \mathbf{1}_N \cdot \mu$
The $\mathbf{1}_N \cdot \mu$ term creates an $N \times D$ matrix where every row is the mean vector, which matches the offset in sklearn's inverse transform.

Practical Code Example #

Here's a concrete implementation to test this:

import numpy as np
from sklearn.decomposition import PCA

# Generate sample non-centered data
np.random.seed(42)
X_raw = np.random.rand(10, 3)  # 10 samples, 3 features

# Fit PCA and get transformed data
pca = PCA(n_components=2)
Z = pca.fit_transform(X_raw)

# Construct X1 and components_1
X1 = np.hstack([Z, np.ones((Z.shape[0], 1))])
components_1 = np.vstack([pca.components_, pca.mean_])

# Compare results
inv_transform_result = pca.inverse_transform(Z)
dot_result = X1 @ components_1

# Check if they match (within floating-point error)
print(np.allclose(inv_transform_result, dot_result))  # Output: True

Key Notes #

• This works for both dimensionality reduction (K < D) and full-dimensional PCA (K = D)
• The intuition is that we're adding a constant basis vector to our transformed space, which lets us encode the mean offset as part of the linear transformation instead of a separate addition.

内容的提问来源于stack exchange，提问作者Ray Chang