SEE-Classify: Simple Evolutionary Exploration Tool to Search Classifiers and their Hyper-parameters

Anonymous

Keywords: meta-learning, supervised machine learning, classification, research software, research support software, genetic algorithms

TL;DR: This work presents a software tool that helps researchers select supervised learning algorithms via a simple genetic algorithm.

Abstract:

Genetic algorithms (GAs) find solutions to search problems through a process
inspired by evolution. Possible solutions to a problem are randomly selected and
tested using a fitness function. The best solutions undergo changes (mutations) over
multiple iterations (generations) to try and find better solutions. There have been
several studies that use GAs to search over hyperparameters of machine learning
algorithms to learn values that work well for specific problems. In this work an
existing GA framework was extended to search over different classifiers and their
hyperparameters. This will allow scientists from any field to search a classifier
"algorithm space" to find a specific classifier (Support Vector Machine, Forest of
Decision trees, Neural Networks, etc.) that works well for their problem. This paper
demonstrates the feasibility of the SEE-Classify system by testing the system on
well-known classification examples provided with the Scikit-learn Python Library
and reproducing results from a previous study that performs simlar hyper-parameter
genetic search over diagnostic breast cancer data.