Entropy and Confidence-Based Undersampling Boosting Random Forests for Imbalanced Problems.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Imbalanced datasets pose significant challenges in machine learning, often leading to biased models and poor generalization.
• Traditional undersampling methods risk discarding informative majority class samples, hindering model performance.

Purpose of the Study:

• To propose a novel framework, entropy and confidence-based undersampling boosting (ECUBoost), to effectively address imbalanced learning problems.
• To enhance the generalization performance of ensemble models by integrating an advanced undersampling technique.

Main Methods:

• ECUBoost employs both entropy and confidence metrics to guide the undersampling process, ensuring the preservation of valid and structurally representative majority samples.
• The framework is designed to be compatible with non-iterative algorithms, such as decision trees, by leveraging confidence-based undersampling.
• Random forests are utilized as the base classifiers within the ECUBoost framework.

Main Results:

• Experimental evaluations on both artificial and KEEL datasets demonstrate the effectiveness of the proposed ECUBoost method.
• The approach successfully mitigates the negative impact of data imbalance without compromising the integrity of the majority class data.

Conclusions:

• ECUBoost offers a robust and effective solution for handling imbalanced datasets in machine learning.
• The method's ability to preserve informative samples and its compatibility with various algorithms make it a valuable contribution to the field of imbalanced learning.