Developing new fitness functions in genetic programming for classification with unbalanced data.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Machine learning algorithms, including genetic programming (GP), often evolve biased classifiers when faced with unbalanced datasets.
• Class imbalance occurs when one class has significantly fewer training examples (minority class) than others (majority class).
• Traditional training criteria can lead to classifiers with high accuracy on the majority class but poor performance on the minority class.

Purpose of the Study:

• To identify limitations of current genetic programming approaches for unbalanced classification tasks.
• To develop and evaluate novel fitness functions designed for binary classification with imbalanced data.
• To improve classifier performance across all classes without resorting to data resampling techniques.

Main Methods:

• Development of several new fitness functions specifically for genetic programming in binary classification.
• Utilized a range of real-world classification problems characterized by significant class imbalance.
• Employed the original, unbalanced training data directly within the genetic programming learning process.

Main Results:

• The proposed fitness functions successfully evolved classifiers demonstrating good performance on both minority and majority classes.
• Empirical results validated the effectiveness of the new fitness functions across diverse imbalanced datasets.
• The methods avoided the need for artificial data balancing techniques such as over- or under-sampling.

Conclusions:

• The novel fitness functions offer a robust solution for training genetic programming classifiers on unbalanced datasets.
• These approaches enhance the reliability and fairness of machine learning models in real-world imbalanced scenarios.
• The study demonstrates the viability of improving classifier performance without altering the original data distribution.