10.6 Random forests

A random forest fits many decision trees using resampled observations and randomly selected features, then combines their predictions. The method often improves predictive stability and accuracy relative to one tree, especially when relationships are nonlinear or involve interactions.

The cost is reduced simplicity. Feature-importance measures can help describe the model, but they should be interpreted carefully. Importance may be affected by correlated features, variable scale, and the chosen measure. Partial-dependence or accumulated-effect displays describe model behaviour, not causal intervention effects.

Compare the random forest with a baseline and simpler model. If performance improves only slightly, the simpler model may be easier to explain, maintain, and monitor.