Max Kuhn and Kjell Johnson wrote this book to fill a gap between machine learning texts that focus on theory and those that read like software manuals. It is about the practice of modeling when predictive accuracy is the main goal: it gives intuitive explanations of how models work, shows the practical side of training and tuning them, and provides the data and code needed to reproduce every analysis. The book won the 2014 Technometrics Ziegel Prize for Outstanding Book.
It serves both as an introduction to predictive models and as a guide for practitioners who want to broaden their skills.
Equations are kept to a minimum, though readers should know basic statistical ideas such as correlation and linear regression, and some mathematical background helps with the advanced topics.
Most chapters close with a computing section in R, an appendix introduces R and the caret package, and the data sets are distributed in the companion AppliedPredictiveModeling package.
Case studies draw on real problems such as cell segmentation, compound solubility, concrete strength, grant applications, and job scheduling.
Topics include:
- Data preprocessing, transformations, and missing values
- Overfitting, data splitting, resampling, and model tuning
- Measuring performance for regression and classification models
- Linear regression, partial least squares, and penalized models
- Nonlinear models such as neural networks, MARS, support vector machines, and K-nearest neighbors
- Tree and rule-based models, including random forests, boosting, Cubist, and C5.0
- Remedies for severe class imbalance
- Predictor importance, feature selection, and factors that affect model performance