Visualization of Regression Models Using Discriminative Dimensionality Reduction

Alexander Schulz, Barbara Hammer

Abstract: Although regression models offer a standard tool in machine learning, there exist barely possibilities to inspect a trained model which go beyond plotting the prediction against single features. In this contribution, we propose a general framework to visualize a trained regression model together with the training data in two dimensions. For this purpose, we rely on modern nonlinear dimensionality reduction (DR) techniques. In addition, we argue that discriminative DR techniques are particularly useful for the visualization of regression models since they can guide the projection to be more sensitive for those aspects in the data which are important for prediction. Given a data set, our framework can be utilized to visually inspect any trained regression model.