One of the main assumptions of classical machine learning models is that data are generated by a stationary concept. This is often violated due to unexpected events or outliers, sensor drift, or individual user behavior. A particular challenge is present in the agnostic setting, ie the form and strength of drift is not given.
In the talk, we will deal with the question, how to formalize the presence of drift in continuous time, and how to derive robust drift detection and drift localization methods based thereon. In particular, we present a supervised non-parametric drift learning method and demonstrate its use in a number of benchmarks. In addition, we address the challenge of tackling possibly drifting data in an unsupervised way, and we propose novel drift explanation and drift decomposition methods, which enable a closer inspection of such drifting data sets.
invited by Manfred Opper
***Due to the current pandemic this colloquium will be conducted online. We invite you to join and spread the news. We will send out an invitation for a zoom meeting via our email list. If you are not on the mail list already, please send an email up front to liv.heinecke[at]uni-potsdam.de ***