ViewClassifier: Visual Analytics on Performance Analysis for Imbalanced Fatal Accident Data

The accurate analysis and modeling of high-dimensional imbalanced fatal accident data including its conflicting and overlapping features are significant in terms of public and transportation safety. Traditional visualization techniques and single-valued performance metrics, such as accuracy, obstruct understanding the behavior of machine learning algorithms for imbalanced fatal accident datasets. Therefore, we propose an interactive visualization tool called ViewClassifier to support the analysis of imbalanced fatal accident datasets as well as their prediction results. ViewClassifier enables users to simulate unseen/missing class features through negative sampling by displaying raw data attribute distributions via parallel coordinates. It also allows users to compare multiclass classifiers at first glance and discover model behavior and possible reasons for misclassified results by filtering and examining at instance-level via customized plots. We presented two case studies to show the usage and efficacy of the proposed visualization tool using real-world traffic accident datasets.