PUB - Publikationen an der Universität Bielefeld

Automated machine learning (AutoML) technologies constitute promising tools to automatically infer model architecture, meta-parameters or processing pipelines for specific machine learning tasks given suitable training data. At present, the main objective of such technologies typically relies on the accuracy of the resulting model. Additional objectives such as sparsity can be integrated by pre-processing steps or according penalty terms in the objective function. Yet, sparsity and model accuracy are often contradictory goals, and optimum solutions form a Pareto front. Thereby, it is not guaranteed that solutions at different positions of the Pareto front share the same architectural choices, hence current AutoML technologies might yield sub-optimal results. In this contribution, we propose a novel method, based on the AutoML method TPOT, which enables an automated optimization of ML pipelines with sparse input features along the whole Pareto front. We demonstrate that, indeed, different architectures are found at different points of the Pareto front for benchmark examples from the domain of systems security.