Data Mining Algorithms: Explained Using R

Chapter 10

Regression model evaluation

10.1 Introduction

Just like for classification model evaluation addressed in Chapter 7, the evaluation of a regression model is intended to provide a reliable assessment of its predictive performance, i.e., the quality of the target function's approximation it represents. There are several regression performance measures calculated by comparing the model's predictions and true target function values on a particular dataset. These are not only the direct indicators of dataset performance, but—under some conditions—can also serve as estimators of true performance, i.e., their expected values on the whole domain.

10.1.1 Dataset performance

Dataset performance, obtained by calculating one or more selected performance measures on a particular dataset, represents the degree of match between model predictions and target function values on this dataset.

10.1.2 Training performance

Performance measures calculated for a model on the training set used to create the model represent its training performance. It may be useful for diagnostic purposes, but does not provide information on the actual predictive utility of the model.

10.1.3 True performance

The actual predictive power of a model is reflected by its expected performance (with respect to one or more selected performance measures) on the whole domain, which is the model's true performance. Since target function values are generally unavailable, true performance always remains unknown ...