by Dr Andy Corbett
Random Forests
9. Interpolation Through a Random Forest
Download the resources for this lesson here.
Random forests interpolate through the functions represented by the constituent weak learners, the decision trees. This means they have the weight of smooth predictors, whilst still maintaining the explainabile structure of the trees they contain.
- Revise a problem studied with a decision tree now with a new model: the random forest.
- Interrogate the individual graphs of the forests trees and compare against a single tree's output.
- Observe the smoothing/interpolation property carried by a bagged ensemble method.
Visualising the interpolation property
We return back to the start of this section: let's see how a random forest performs on our task in Fig. 1, originally pitted against a decision tree.
Figure 1. Compare of the performance of a decision tree (left) with a random forest of decision trees (right).
We began by claiming that decision trees a prone to over fitting. In this example, 50 data points were used to train both a decision tree and a random forest. The forest displays a more reasonable interpolation of the data, closer to the ground truth. Whereas the decision tree thresholds its predictions precisely around each datapoint.
In this video and notebook we also explore the interpreability of random forests in terms of their constituent components. The graphs for each weak learner are often much smaller for the same degree of model accuracy. Thus giving easy insight to non-technically trained individuals.