Chapter 9. Exploring advanced methods

 

This chapter covers

  • Reducing training variance with bagging and random forests
  • Learning non-monotone relationships with generalized additive models
  • Increasing data separation with kernel methods
  • Modeling complex decision boundaries with support vector machines

In the last few chapters, we’ve covered basic predictive modeling algorithms that you should have in your toolkit. These machine learning methods are usually a good place to start. In this chapter, we’ll look at more advanced methods that resolve specific weaknesses of the basic approaches. The main weaknesses we’ll address are training variance, non-monotone effects, and linearly inseparable data.

To illustrate the issues, let’s consider a silly health prediction model. Suppose we have for a number of patients (of widely varying but unrecorded ages) recorded height (as h in feet) and weight (as w in pounds), and an appraisal of “healthy” or “unhealthy.” The modeling question is this: can height and weight accurately predict health appraisal? Models built off such limited features provide quick examples of the following common weaknesses:

9.1. Using bagging and random forests to reduce training variance

9.2. Using generalized additive models (GAMs) to learn non-monotone relationships

9.3. Using kernel methods to increase data separation

9.4. Using SVMs to model complicated decision boundaries

9.5. Summary