Abstract:
Presentation of machine learning (ML) model results plays an important role in decision makers’ trust and use. Yet, there has been little agreement on how information should be visualized to present models' evaluations. The purpose of this thesis is to formulate an approach to visualize the results of classification model’s evaluation to increase decision makers’ trust. This work proposes a dashboard that visualizes supervised ML model performance in a dashboard which is split into three main sections: statistical measures, feature importance, sensitivity analysis. Three sample dashboards were generated and evaluated using a survey by ten faculty members and students from George Mason University most of which said that the dashboard provides useful information and gives a better understanding of the model behavior than other methods they have experienced.
Model evaluation strategies differ based on the prediction problem considered. However, a consistent representation of evaluation results may increase decision makers’ trust in the models. The next step of this project is to visualize the difference between multiple models.
