Supervised Learning bei komplexen Stichproben oder unvollständigen Daten
Prof. Dr. Thomas Augustin
Ludwig-Maximilians-Universität München (LMU Munich)
Die Keynote wird in deutscher Sprache mit englischsprachigen Folien gehalten werden. Hier der vorliegende englische Abstract.
The keynote will be held in German with English-language slides. Here is the English abstract.
Abstract
Complex sampling with unequal selection probabilities, measurement error, and partial identifiability from incomplete observations are crucial topics of classical survey statistics. The talk investigates how far these issues remain relevant in the rigorously prediction-oriented setting of supervised machine learning and to what extent classical adjustment procedures can be extended. Concretely, we consider classification and regression trees as essential ingredients for more complex methods like random forests, discussing attempts to adjust them for complex sampling designs or measurement error. Finally, we speculate on the power of set-valued classification in partially identifiable settings.