Multi-objective Optimization: Theory, Algorithms, and Applications in Machine Learning

In this inaugural lecture, we will focus on the theoretical and algorithmic analysis of specific multi-objective optimization problems. The goal of these problems is to minimize a finite number of conflicting objective functions simultaneously (in the sense of Edgeworth and Pareto) over a finite feasible set. We will highlight their application to support vector machines (SVMs) in supervised machine learning for binary classification. One key outcome will be the derivation of a multi-objective data reduction approach for hard-margin linear SVMs. This approach is particularly useful when the underlying dataset grows over time.

The SVM-related results are based on joint work with Marc Steinbach.