Constraint-aware machine learning

Most ML research focuses on advancing predictive performance. However, in practice, there are many more dimensions that have to be considered to make an ML application successful, such as fairness, privacy, safety, efficiency, and explainability.
We propose the paradigm of constraint-aware ML where a user defines all constraints in a declarative way and a system automatically generates an ML pipeline that satisfies these specified constraints.
To this end, we make three major contributions. First, we leverage feature selection to satisfy user-specified ML application constraints and propose an optimizer that chooses the most promising feature selection method for this task. Second, we build an AutoML system that adapts its own parameters to search efficiently for an ML pipeline that satisfies the user-specified constraints. Third, we conduct an in-depth holistic analysis of the sustainability and efficiency of AutoML systems and propose multiple approaches to reduce their energy consumption.