Robust framework to combine diverse classifiers assigning distributed confidence to individual classifiers at class...

Area of Science:

• Machine Learning
• Data Mining
• Artificial Intelligence

Background:

• Class label noise and imbalanced datasets pose significant challenges in machine learning model training.
• Existing ensemble methods may not effectively address these issues, leading to suboptimal performance.

Purpose of the Study:

• To develop a robust classification framework that combines heterogeneous classifiers while mitigating the impact of noisy training data.
• To enhance the accuracy and reliability of ensemble models, particularly in the presence of class label noise and imbalanced classes.

Main Methods:

• An extended m-Mediods based modeling approach was used to identify and filter noisy data.
• Noise-free data was utilized to train Gaussian Mixture Models (GMM) and Support Vector Machines (SVM).
• A genetic algorithm-based weight learning method was employed to optimize classifier ensemble weights for improved accuracy.

Main Results:

• The proposed ensemble method demonstrated superior performance compared to standard ensemble techniques (Adaboost, Bagging, Random Subspace Methods).
• Significant improvements in classification accuracy were observed, especially when dealing with datasets containing class label noise and imbalanced classes.
• The framework effectively filtered noisy training data, leading to more reliable model learning.

Conclusions:

• The developed ensemble classification framework offers a robust solution for handling noisy and imbalanced datasets.
• The integration of m-Mediods, GMM, SVM, and genetic algorithms provides a powerful approach for building accurate and resilient machine learning models.
• This method presents a valuable advancement for real-world classification tasks where data quality is a concern.