Optimization of Distributions Differences for Classification.

Area of Science:

• Computer Science
• Machine Learning
• Data Mining

Background:

• Classification algorithms are crucial for data analysis.
• Existing methods face challenges with imbalanced data and generalization.
• Feature space transformation is a key area for improving classification.

Purpose of the Study:

• Introduce a novel classification algorithm, Optimization of Distribution Differences (ODD).
• Develop a method to transform feature spaces for better class separation.
• Evaluate ODD's performance against established and recent classification techniques.

Main Methods:

• Formulated classification as a multiobjective optimization problem.
• Employed a hybrid approach combining evolutionary strategy and quasi-Newton method.
• Experimented with both linear and nonlinear transformation functions in the feature space.

Main Results:

• ODD demonstrated superior performance over eight benchmark and two recent classification methods.
• The algorithm showed reduced sensitivity to imbalanced datasets compared to other methods.
• ODD exhibited enhanced generalization ability across 12 standard datasets.

Conclusions:

• ODD offers a robust and effective approach to classification tasks.
• The algorithm's ability to handle imbalanced data and generalize well is a significant advantage.
• ODD represents a promising advancement in classification algorithm development.