Data Generators for Learning Systems Based on RBF Networks.

Area of Science:

• Computer Science
• Machine Learning
• Data Mining

Background:

• Scarcity and high cost of real-world data hinder algorithm development.
• Need for reliable methods to generate synthetic data for robust testing and optimization.

Purpose of the Study:

• To propose and evaluate a novel semi-artificial data generator.
• To enable large-scale experimentation and simulation for data mining algorithms.
• To assess the quality and usability of generated data.

Main Methods:

• Utilized radial basis function networks to learn Gaussian kernels.
• Employed kernels in a generative mode to produce new data points.
• Evaluated data quality via statistical properties, structural, and predictive similarity.
• Conducted large-scale validation on 51 datasets using supervised and unsupervised learning.

Main Results:

• Generated data exhibited considerable similarity to original datasets.
• Demonstrated effectiveness in development and simulation scenarios.
• Confirmed the utility of the generator across diverse datasets.

Conclusions:

• The proposed generator is a valuable tool for data mining research and practice.
• Synthetic data can augment real datasets to improve model performance.
• The method shows promise for high-dimensional data and various classification tasks.