无法安装额外库时，如何用sklearn的kNN手动填补缺失值？

Great question! Let's tackle both of your concerns clearly:

Absolutely, you can build your own KNN-based missing value imputer using just scikit-learn's built-in tools—no external libraries or GitHub code required. Here's how it works, along with a practical implementation:

Core Idea #

For each sample with a missing value in a feature:

• Find its k nearest neighbors using the non-missing features (we'll handle scaling since KNN is sensitive to feature scales).
• Fill the missing value with the mean (for numerical features) or mode (for categorical features) of that feature across the k neighbors.

Code Implementation #

import numpy as np
from sklearn.neighbors import NearestNeighbors
from sklearn.preprocessing import StandardScaler

def knn_impute(X, k=5):
    # Standardize features first—critical for KNN performance
    scaler = StandardScaler()
    X_scaled = scaler.fit_transform(X)
    
    # Create a copy to store imputed values
    X_imputed = X.copy()
    
    # Iterate over each feature column
    for col in range(X.shape[1]):
        # Get indices of samples with missing values in this column
        missing_idx = np.where(np.isnan(X[:, col]))[0]
        if len(missing_idx) == 0:
            continue  # Skip if no missing values
        
        # Get all complete samples (no missing values in any feature) to use as neighbors
        complete_mask = ~np.isnan(X).any(axis=1)
        X_complete = X_scaled[complete_mask]
        y_complete = X[complete_mask, col]
        
        # Handle case where we have fewer complete samples than k
        if len(X_complete) < k:
            fill_val = np.mean(y_complete)
            X_imputed[missing_idx, col] = fill_val
            continue
        
        # Fit NearestNeighbors on complete samples
        nn_model = NearestNeighbors(n_neighbors=k)
        nn_model.fit(X_complete)
        
        # Impute each missing sample
        for idx in missing_idx:
            # Create a temporary version of the sample, filling other missing features with column means
            temp_sample = X_scaled[idx].copy()
            for c in range(X.shape[1]):
                if np.isnan(temp_sample[c]):
                    temp_sample[c] = np.mean(X_scaled[complete_mask, c])
            
            # Find k nearest neighbors
            _, neighbor_indices = nn_model.kneighbors(temp_sample.reshape(1, -1))
            # Fill with mean of neighbors' values for the current column
            fill_val = np.mean(y_complete[neighbor_indices[0]])
            X_imputed[idx, col] = fill_val
    
    return X_imputed

Notes #

• Scaling: We use StandardScaler because KNN relies on distance metrics—features with larger scales would dominate the distance calculation otherwise.
• Categorical Features: If you have categorical data, modify the code to use mode instead of mean for filling, and consider encoding categorical features (e.g., one-hot encoding) before calculating distances.
• Edge Cases: The code handles scenarios where there aren't enough complete samples to find k neighbors by falling back to the column mean.

Short answer: No, scikit-learn's KNeighborsClassifier and KNeighborsRegressor cannot handle missing values directly.

These models calculate distances between samples (e.g., Euclidean, Manhattan) to find neighbors, and any NaN in the feature data will result in NaN distance values. This will cause errors during fitting, prediction, or both.

You must first impute all missing values (using the manual KNN method above, or simpler methods like mean/median/mode imputation) before using scikit-learn's KNN models on your data.

内容的提问来源于stack exchange，提问作者Kevin