PCA降维后如何评估原始特征的重要性及排序？

Absolutely, there are straightforward, reliable methods to figure out the importance ranking and corresponding scores of your original features (A-E) after performing PCA. Let’s break down the most practical approaches:

1. 加权载荷得分法（最常用且严谨） #

这个方法 combines the variance weight of each principal component (PC) with the original feature's loading on that PC, giving you a precise measure of how much each feature contributes to the overall variance explained by PCA.

• 核心逻辑：Each PC's variance contribution rate represents its "importance," while the squared loading of a feature on that PC represents its share of contribution to that PC. Weighted summation of these values gives the feature's comprehensive importance score.
• 具体步骤：
  • Extract the loading matrix from your PCA model: each row corresponds to an original feature, each column to a PC, and the value is the coefficient of the feature on that PC.
  • Extract the variance contribution rate of each PC (i.e., explained_variance_ratio_, where the first PC has the highest value).
  • For each feature, calculate: Comprehensive Score = Σ( (Feature's loading on PC j)² × PC j's variance contribution rate )
  • Sort features by their scores in descending order to get the importance ranking.

2. 主成分载荷直接分析（快速直观） #

If you care more about features' contributions to the most critical PCs, you can simplify the analysis:

• Focus on the absolute loading values of original features on the top 1-2 PCs (those with the highest variance): the larger the absolute value, the more the feature contributes to the core variance explanation.
• You can also calculate the average or sum of absolute loadings across all PCs as a quick importance indicator.

3. 代码实现（以Python scikit-learn为例） #

Here's ready-to-run code to compute scores and rankings quickly:

import numpy as np
from sklearn.decomposition import PCA
from sklearn.preprocessing import StandardScaler

# Assume X is your original feature dataset (n samples, 5 features A-E)
# Standardize features first (PCA is sensitive to scale—this step is mandatory!)
scaler = StandardScaler()
X_scaled = scaler.fit_transform(X)

# Fit PCA model to reduce to 3 dimensions
pca = PCA(n_components=3)
pca.fit(X_scaled)

# Get loading matrix (5 features × 3 PCs)
loadings = pca.components_.T
# Get variance contribution rate of each PC
var_contributions = pca.explained_variance_ratio_

# Calculate importance score for each original feature
feature_scores = np.sum(loadings ** 2 * var_contributions, axis=1)
# Corresponding feature names
feature_names = ['A', 'B', 'C', 'D', 'E']

# Sort features by score in descending order
sorted_feature_ranking = sorted(zip(feature_names, feature_scores), 
                                key=lambda x: x[1], reverse=True)

# Output results
print("Original Feature Importance Ranking & Scores:")
for idx, (name, score) in enumerate(sorted_feature_ranking, 1):
    print(f"{idx}. {name}: {score:.4f}")

Key Notes #

• Always standardize features: If original features have vastly different scales (e.g., A ranges 0-100, B ranges 0-1), PCA will bias toward high-variance features, distorting loadings and scores. Standardization ensures all features have equal variance, making results reliable.
• This "importance" is variance-explanation based: Unlike feature importance from tree-based models, it measures how much a feature contributes to preserving information during data compression—perfect for PCA-specific analysis.

内容的提问来源于stack exchange，提问作者Navdeep