A classifier ensemble approach for the missing feature problem.

Area of Science:

• Machine Learning
• Data Science
• Statistical Modeling

Background:

• Missing data is a common challenge in classification problems, necessitating effective data imputation techniques.
• Evaluating various statistical and machine learning imputation methods is crucial for improving model performance.

Purpose of the Study:

• To evaluate the performance of existing imputation methods and introduce a novel multiple imputation ensemble approach.
• To assess the effectiveness of the proposed method across different datasets and compare it with state-of-the-art techniques.

Main Methods:

• A novel multiple imputation method based on random subspace and fuzzy clustering was developed.
• State-of-the-art classifiers, including support vector machines and input decimated ensembles, were tested with various imputation methods.

Main Results:

• The proposed clustering-based multiple imputation ensemble approach outperformed other state-of-the-art methods.
• The novel approach showed comparable performance to using original data (without missing values) even when over 20% of data was missing.
• Coupling imputation methods with the proposed cluster-based imputation enhanced base method performance.

Conclusions:

• The proposed method demonstrates robust performance even with a high percentage of missing features (≥30%), assuming feature set redundancy.
• The developed multiple imputation ensemble approach offers a valuable solution for classification problems with missing data.
• MATLAB code for the best approach is publicly available for further research and application.