Chapter 11 Fine Tune Models
In the previous chapters, we have obtained the knowledge that many prediction models need fine-tune, so they could:
- better fit the training dataset,
- produce better performance for the test dataset (less overfit).
In most data science projects, fine-tune a model is not only necessary but also desirable since it can increase the final outcomes of the projects. In this chapter, we will demonstrate techniques that are used to fine-tune a prediction model.
Fine-tune a model is specific to the model. Different models may have different parameters and different measurements of the performance. In general, A model’s overall performance is affected by the three factors:
- The predictors. The numbers of predictors used in a model and the specific predictors used in the model.
- The training sample. The larger the data sample the better change of model fit. However, there are many methods for dealing with a small data sample. Mostly to enlarge the data sample or make efficient use of available data samples.
- The parameter of the model. The adjustable parameters in a model. Most of the model tuning refers to adjust the parameters of the model.
We can see that the search space defined by the three dimensions is fairly large. Sometimes it is computationally infeasible to search them in one go. In practice, we generally fix one or two dimensions and search for another dimension. In other words, we can fine-tune them one by one.
That is what we are going to do.