Some Recent Advances in Exceptional Model Mining: Tales of Potatos, Boris Johnson, and Atrial Fibrillation

Wouter Duivesteijn, TU Eindhoven

   16.04.2024

   2:00 pm - 3:30 pm

   LMU Main Building, Room M001 Geschwister-Scholl-Platz 1, 80539 München

Exceptional Model Mining (EMM) strives to find subgroups in a dataset that behave somehow exceptionally. We partition the columns of the dataset into two sets. The one set is used for defining candidate subgroups (by making conjunctions of conditions on individual attributes); a challenge lies in efficiently traversing the search space of candidate subgroups. The other set is used for evaluating exceptionality of subgroup behavior: we define a kind of interaction between these target columns, and deem subgroups to be exceptional if parameters of this interaction have exceptional values. We will illustrate this concept with three types of interaction. The first, regression, has implications for the price of rice in China (traditional linear regression), dot counting skills of Finnish school children (piecewise regression), and potato growth (mixed models). The second, exceptional trends in repeated cross-sectional data, illustrates how local deviations can make Boris Johnson appear in a graph of trends. The third, exceptional signals in EGCs, has the potential to improve detection of atrial fibrillation. If time remains, we will discuss how EMM relates to clustering; there are clear connections and clear differences between the tasks, but are there fruitful avenues for cross-pollination?