A Visualization to Investigate and Give Feedback to Classifiers

Classification is a core technique of data mining. The performance of a classifier might be improved by user feedback especially when a stable classifier cannot be generated from the available training data. Classification processes and classifier decisions are sometimes hard to grasp. However, for giving feedback users need to understand a classifier’s decision and the results. Visualizations can boost the understanding of classification processes and results even for non-expert users. In the past, several classifier visualizations were proposed, however they are either restricted to a limited set of classifiers [Chodos and Zaiane 2008] or no feedback mechanisms are supported [Rheingans and desJardins 2000].
In this work we present a visualization system to investigate and give feedback to all classifiers whose predictions can be interpreted as probability distribution over classes. We provide an intuitive and simple application to observe classification processes and results even for non-expert users. Users can give feedback to classifiers directly in the visualization with well-known interaction techniques like drag and drop.