Over-Fitting and Model Tuning

Many modern classification and regression models are highly adaptable; they are capable of modeling complex relationships. Each model's adaptability is typically governed by a set of tuning parameters, which can allow each model to pinpoint predictive patterns and structures within the data. However, these tuning parameters can very identify predictive patterns that are not reproducible. This is known as “over-fitting.” Models that are over-fit generally have excellent predictivity for the samples on which they were built, but poor predictivity for new samples. Without a methodological approach to building and evaluating models, the modeler will not know if the model is over-fit until the next set of samples are predicted. In Section 4.1 we use a simple example to illustrate the problem of over-fitting. We then describe a systematic process for tuning models (Section 4.2), which is foundational to the remaining parts of the book. Core to model tuning are appropriate ways for splitting (or spending) the data, which is covered in Section 4.3. Resampling techniques (Section 4.4) are an alternative or complementary approach to data splitting. Recommendations for approaches to data splitting are provided in Section 4.7. After evaluating a number tuning parameters via data splitting or resampling, we must choose the final tuning parameters (Section 4.6). We also discuss how to choose the optimal model across several tuned models (Section 4.8) We illustrate how to implement the recommended techniques discussed in this chapter in the Computing Section (4.9). Exercises are provided at the end of the chapter to solidify concepts.