Evolutionary optimization of radial basis function classifiers for data mining applications.

Area of Science:

• Data Mining and Machine Learning
• Computational Intelligence
• Pattern Recognition

Background:

• Feature and model selection are critical, yet computationally intensive, tasks in classification problems.
• Existing methods often require separate optimization steps, increasing complexity and runtime.
• Radial Basis Function (RBF) classifiers require careful selection of input features and network structure.

Purpose of the Study:

• To develop an efficient evolutionary algorithm (EA) for simultaneous feature and model selection in RBF classifiers.
• To accelerate and enhance EA performance through integrated optimization techniques.
• To demonstrate the algorithm's effectiveness across diverse data mining applications.

Main Methods:

• Developed a novel evolutionary algorithm (EA) for joint feature and model selection.
• Integrated techniques including hybrid RBF network training, lazy evaluation, penalty terms for soft constraints, and adaptive EA control.
• Validated the approach on four distinct data mining problems: intrusion detection, biometric verification, customer acquisition, and chemical process optimization.

Main Results:

• Achieved significant reductions in runtime, up to 99%, compared to previous EA-based RBF optimization methods.
• Demonstrated substantial decreases in error rates, up to 86%, across various applications.
• The proposed EA is application-independent, suggesting broad transferability to other classifier paradigms.

Conclusions:

• The developed EA offers a highly efficient and effective solution for simultaneous feature and model selection in RBF classifiers.
• Integrated optimization techniques markedly improve EA performance, reducing computational burden.
• The algorithm's versatility and performance gains make it a valuable tool for diverse data mining challenges.