Visualizing Prediction Correctness of Eye Tracking Classifiers

Abstract

Eye tracking data is often used to train machine learning algorithms for classification tasks. The main indicator of performance for such classifiers is typically their prediction accuracy. However, this number does not reveal any information about the specific intrinsic workings of the classifier. In this paper we introduce novel visualization methods which are able to provide such information. We introduce the Prediction Correctness Value (PCV). It is the difference between the calculated probability for the correct class and the maximum calculated probability for any other class. Based on the PCV we present two visualizations: (1) coloring segments of eye tracking trajectories according to their PCV, thus indicating how beneficial certain parts are towards correct classification, and (2) overlaying similar information for all participants to produce a heatmap that indicates at which places fixations are particularly beneficial towards correct classification. Using these new visualizations we compare the performance of two classifiers (RF and RBFN).