Informatik-Kolloquium
Am Montag, 17.02.2020 hält Prof. Dr. Hans Kestler (Universität Ulm) einen Vortrag im Rahmen des Informatik-Kolloquiums. Der Titel lautet „Interpretable classifiers for high-dimensional molecular data„.
Der Vortrag findet im Raum 01.150-128 (Cauerstraße 11) statt.
Abstract:
The interpretability of classification models is essential in the process of selecting biomarkers and developing diagnostic models. In high-dimensional settings, the interpretability of a model is directly related to the construction of a feature set (signature) a classifier is operating on and on the theoretical properties of the chosen concept class. This signature can give hints towards the processes leading to a particular categorisation, the concept class can be used to impose or interpret decision rules.
Nevertheless, purely data driven feature selection is often affected by different forms of uncertainty and the derived signatures do not perfectly fit a given high-level interpretation. External information about the dependencies of measurements can be incorporated to increase a signature interpretability. In addition theoretical invariance properties of the classifiers can support the analysis of these signatures.
In this talk we will first report four linked interpretable linear concept classes/models with distinct invariance properties for high-dimensional molecular classification. In the second part we will show how to incorporate semantic information into the training process of multi-classifier systems. Finally we will give an outlook how to combine both approaches.