Programm Potsdam-Berlin-Kolloquium, 11.10.2019
Tilman Hartwig (Tokyo)
I will present decision trees as very efficient machine learning method to classify astronomical data. A labelled training sample is split according to available features by requiring that each split minimises the information entropy of the assigned classes. This elegant mathematical formulation allows us to construct decision trees with supervised learning, which can then be applied to classify new observations. Eventually, I will present recent results of my own research: by classifying the chemical abundance patterns of metal-poor stars in the Milky Way, we can derive the multiplicity of the first generation of stars in the Universe. Furthermore, this approach provides the feature importance to identify the most informative chemical elements to classify metal-poor stars, which can be used to optimise future spectroscopic surveys of Milky Way stars.
