Nearest neighbor imputation algorithms: a critical evaluation.

Area of Science:

• Data Science
• Machine Learning
• Statistical Analysis

Background:

• Nearest neighbor (NN) imputation is an efficient method for handling missing data.
• NN imputation aims to replace missing values with plausible estimates while preserving data structure and variable distribution.
• The impact of NN imputation methods on data structure is not well understood.

Purpose of the Study:

• To evaluate the performance of NN imputation algorithms in terms of imputation accuracy and data structure preservation.
• To compare the effects of different NN imputation variants (1NN, kNN, wkNN) across various machine learning frameworks.
• To identify optimal NN imputation strategies for minimizing imputation error and maintaining data integrity.

Main Methods:

• Simulations were performed on synthetic datasets with varying missingness patterns and degrees.
• Evaluated NN imputation with one neighbor (1NN) and k neighbors (kNN, wkNN).
• Tested imputation methods within different learning frameworks including ReliefF filtering, bagging, and Random-Forest-like approaches.

Main Results:

• kNN generally outperformed 1NN in imputation precision and reduced inferential errors.
• 1NN was the only method that preserved data structure; distortion occurred even with small k values.
• Resampling schemas exacerbated data distortion.

Conclusions:

• The combination of three neighbors with ReliefF filtering offers the best trade-off between imputation error and data structure preservation.
• These findings were validated on the SPECTF heart dataset with artificially introduced missing data.