Lukas Pfannschmidt

Big improvements in classification algorithms in recent years not only brought new application possibilities but also lead to increased feature selection performance. In the context of biomedical data analysis, feature selection or feature relevance weighting constitutes a particularly important part for model interpretation, since it offers possible candidates for biomarkers. While such methods are capable of reliable extracting clear biomarkers, they are less reliable when it comes to subtle biomarkers in particular for highly correlated input data.

In this project we develop novel interactive methods which enable the computation of relevance bounds which highlight especially subtle or weakly relevant features. One goal is the ability for the researcher to incorporate hyphotheses into the feature weighting, which allows iterative refinement of the output.

Supervisors: Barbara Hammer (Bielefeld University)