March 22, 2017 Christina Göpfert

Title: Feature Relevance Analysis using Relevance Intervals

Abstract:

In classification tasks, the relevance of each feature for the
classification strongly impacts performance and plays an important role
in gaining insight into the underlying processes. Applications such as
gene expression analysis in the biomedical domain, or error pattern
recognition in motion tracking, generate data with high-dimensional and
strongly correlated features, many of which are likely to be redundant,
but not irrelevant on their own. In my talk, I present a relevance
taxonomy and the concept of feature relevance intervals, which can be
used to structure features according to said taxonomy. For the case of
linear classification, I introduce and illustrate a method for computing
relevance intervals based on linear problems.