In this chapter, we created predictive models in Amazon ML--from selecting the datasource, applying transformations to the initial data with recipes, and analyzing the performance of the trained model. The model performance exploration depends on the type of prediction problem at hand: binary, multi-classification, or regression. We also looked at the model logs for the Titanic dataset and learned how the SGD algorithm trains and selects the best model out of several different ones with different learning rates.

Finally, we compared several data transformation strategies and their impact on the model performance and algorithm convergence in the context of the Titanic dataset. We found out that quantile binning of numeric values is a key strategy in boosting the convergence speed of the algorithm, which overall generated much better models.

So far, these models and performance evaluation are all obtained on training data. That is data that is fully available to the model from the start...