Genetic programming automated machine learning: architecture and performance evaluation

Abstract

Genetic programming-based automated machine learning (GP-AutoML) auto mates machine-learning pipelines through evolutionary search, but its imple mentation requires programming knowledge and configuration of multiple soft ware components. This study presents a web-based platform that integrates the tree-based pipeline optimization tool (TPOT) to support GP-AutoML through a graphical modeling workflow. The platform incorporates conventional machine learning modules, GP-AutoML configuration, result visualization, user authen tication, and cloud-based storage, supported by a flask backend, scikit-learn, TPOT, and Firebase services. A comparative evaluation was conducted be tween the proposed web-based platform and Google colab Python scripting using five binary-classification datasets with varying numbers of records, fea tures, and class distributions. Testing accuracy, Class 1 precision, recall, F1 score, and time-to-completion were evaluated under predefined GP parameter settings. Google colab obtained higher testing accuracy for all datasets. How ever, the differences were small, ranging from 0.01 to 0.06. This indicates comparable predictive performance by the web-based GUI platform. For the moderately balanced dataset, both environments achieved stronger Class 1 pre cision, recall, and F1-score. Although the web-based platform required a longer time-to-completion for every dataset, the evaluated runs remained practical for rapid model development. The findings demonstrate the feasibility of deliver ing TPOT-based GP-AutoML through an integrated web-based architecture that supports accessible visual modeling without direct programming or local soft ware installation.