机器学习缺失值插补咨询：求聚类插补的可用工具包

Hey there! Let's dive into clustering-based missing value imputation and the tooling options you can use, building on the work you've already done with KNN, MICE, and median imputation.

Your Current Work Recap #

• You're expanding beyond basic mean/median/mode imputation, already testing KNN, MICE, and median imputation strategies.
• You're using a smart validation approach: introducing artificial missing values into the complete Iris dataset, which lets you directly compare imputed values to ground truth for accuracy.

Clustering-Based Imputation: Tooling You Can Use #

It’s true that clustering-based imputation isn’t as widely packaged as KNN or MICE, but there are practical ways to implement it without building everything from scratch:

1. Build a Custom Pipeline with Scikit-Learn #

You can leverage scikit-learn’s clustering algorithms (like KMeans, AgglomerativeClustering) to create your own imputation workflow:

• Step 1: Split your data into rows with no missing values and rows with missing values.
• Step 2: Cluster the complete rows using their full feature set.
• Step 3: For each incomplete row, use its non-missing features to assign it to the nearest cluster.
• Step 4: Impute the missing values using the mean/median/mode of the assigned cluster’s corresponding feature.

Here’s a quick code example to kick things off:

import numpy as np
from sklearn.cluster import KMeans
from sklearn.datasets import load_iris

# Load Iris and create artificial missing values (10% missing)
iris = load_iris()
X = iris.data
np.random.seed(42)
mask = np.random.rand(*X.shape) < 0.1
X_missing = X.copy()
X_missing[mask] = np.nan

# Separate complete and incomplete records
complete_rows = X_missing[~np.isnan(X_missing).any(axis=1)]
incomplete_rows = X_missing[np.isnan(X_missing).any(axis=1)]

# Cluster the complete data
kmeans = KMeans(n_clusters=3, random_state=42)
kmeans.fit(complete_rows)

# Impute missing values for incomplete rows
for idx, row in enumerate(incomplete_rows):
    # Get indices of non-missing features
    non_missing_mask = ~np.isnan(row)
    # Predict cluster using available features
    cluster_label = kmeans.predict(row[non_missing_mask].reshape(1, -1))[0]
    # Impute with the cluster's mean for the missing feature(s)
    missing_mask = np.isnan(row)
    incomplete_rows[idx, missing_mask] = np.mean(complete_rows[kmeans.labels_ == cluster_label][:, missing_mask], axis=0)

# Combine back into the full imputed dataset
X_imputed = np.vstack([complete_rows, incomplete_rows])

2. Use Specialized Imputation Libraries #

While not as mainstream, a few libraries include built-in clustering-based imputation:

• impyute: This library has a ClusterImpute method that uses K-Means clustering to impute missing values. Install it with pip install impyute, then use it like this:

  from impyute.imputation.cs import cluster
  
  # Impute the missing Iris data
  X_imputed = cluster(X_missing)
  

• fancyimpute: While best known for KNN imputation, you can combine its utilities with clustering logic (e.g., use its IterativeImputer as part of a clustered workflow) for more customized imputation.

Pro Tips for Evaluating Clustering Imputation #

• Since you’re using the Iris dataset with ground truth, measure accuracy using metrics like mean absolute error (MAE) or root mean squared error (RMSE) between imputed values and the original data.
• Experiment with different clustering algorithms (K-Means vs. hierarchical clustering) to see which performs better for your data’s structure.
• Always scale your features first! Clustering algorithms are sensitive to feature scales, so standardize or normalize your data before clustering.

内容的提问来源于stack exchange，提问作者uharsha33