Selective Equation Constructor: A Scalable Genetic Algorithm

摘要

Efforts to improve machine learning performance begin with defining a valuable feature set. However, datasets with copious amounts of attributes can have relevant information that is obscured by its high dimensionality, which can be caused by repetitive characteristics or irrelevant qualities. Genetic algorithms provide improvements to feature sets through dimensionality reduction and feature construction. Most genetic algorithms follow the theoretical framework of evolutionary theory where a population of features randomly evolves through generations through a series of random operations such as crossover and mutation. While successful, the randomness of feature modification operations and derived constructed features may yield offsprings that under-perform compared to their ancestors, yet their properties are utilized in future generations. We developed a new genetic algorithm called Selective Equation Constructor (SEC) that evolves constructed features selectively in order to limit the shortcomings of other genetic algorithms. The algorithm leads to faster computation and better results compared to similar algorithms. Analysis of the results indicates increases in classification accuracy, decreased run time, and reduction in attribute count.