Advanced genetic programming techniques for machine learning - A comparative analysis with state-of-the-art automated machine learning methods in the context of imbalanced binary classification

Abstract

The number of available machine learning methods and tools is increasing rapidly, with one recent trend being the usage of advanced genetic programming concepts and automated machine learning tools. However, through the rising number of upcoming innovations, it has become a challenge for machine learning applicants to keep up with all the new opportunities and to identify their potentials. While emerging methods are typically compared to conventional standard machine learning algorithms upon their initial introduction, research is still scarce on comparisons of the performances between the new concepts themselves. Therefore, this thesis provides a comparative analysis of two novel genetic programming techniques, differentiable Cartesian genetic programming for artificial neural networks and geometric semantic genetic programming, alongside three state-of-the-art automated machine learning tools, Auto-Keras, Auto-PyTorch and Auto-sklearn, with regard to their relative performances in the machine learning subfield of imbalanced binary classification. In this analysis, the five methods are tested against each other on 20 benchmark datasets, primarily regarding their average and maximum performance, and subsequently the most successful technique is applied to the real-world problem of fraud detection. The purpose of this thesis is not only to familiarize machine learning users with these methods, but above all to determine whether the novel genetic programming techniques can compete with the more established automated machine learning tools, and to identify the overall best performing method.