Chapter 6. Predictions and Performances

It is time to make some predictions! In Chapter 4, Loading and Preparing the Dataset, we did split the Titanic dataset into two subsets, the training and held-out subsets, respectively consisting of 70% and 30% of the original dataset randomly shuffled. We have used variations of the training subset extensively in chapter 5 Model Creation, to train and select the best classification model. But so far, we have not used the held-out subset at all. In this chapter, we apply our models to this held-out subset to make predictions on unseen data and make a final assessment of the performance and robustness of our models.

Amazon ML offers two types of predictions: batch and streaming. Batch prediction requires a datasource. The samples you want to predict are given to the model all at once in batch mode. Streaming, also known as real-time or online predictions, requires the creation of an API endpoint and consists of submitting sequences of samples, one by...