Visual Exploration of Feature-Class Matrices for Classification Problems

Abstract

When a classification algorithm does not work on a data set, it is a non-trivial problem to figure out what went wrong on a technical level. It is even more challenging to communicate findings to domain experts who can interpret the data set but do not understand the algorithms. We propose a method for the interactive visual exploration of the feature-class matrix used to represent data sets for classification purposes. This method combines a novel matrix reordering algorithm revealing patterns of interest with an interactive visualization application. It facilitates the investigation of feature-class matrices and the identification of reasons for failure or success of a classifier on the feature level. We discuss results obtained by applying the method to the Reuters text collection.