Integration of Multikinds Imputation With Covariance Adaptation Based on Evidence Theory

Area of Science:

• Machine Learning
• Data Science
• Computer Science

Background:

• Missing attribute values in incomplete data classification are typically handled by imputation methods.
• Imputation can alter data distributions, leading to degraded classification performance.

Purpose of the Study:

• To propose a novel framework, MICA (integration of multikinds imputation with covariance adaptation) based on evidence theory (ET), for classification with incomplete training data.
• To address the distribution differences introduced by imputation and effectively combine results from multiple classifiers.

Main Methods:

• Employing multiple imputation methods to create diverse imputed training datasets.
• Utilizing a covariance adaptation module (CAM) to minimize distribution discrepancies between imputed and test datasets.
• Combining soft classification results from multiple classifiers using evidence theory, weighted by dataset reliability.

Main Results:

• MICA significantly improves classification performance compared to existing methods.
• The proposed weighting scheme accounts for the varying reliability of imputed datasets.
• Experimental results across several datasets validate the effectiveness of MICA.

Conclusions:

• MICA offers a robust framework for handling incomplete data in classification tasks.
• The integration of imputation, covariance adaptation, and evidence theory enhances classification accuracy.
• The method provides a significant advancement in dealing with data imputation challenges in machine learning.