Mastering Scala Machine Learning

Basic, stratified, and consistent sampling

I've met quite a few data practitioners who scorn sampling. Ideally, if one can process the whole dataset, the model can only improve. In practice, the tradeoff is much more complex. First, one can build more complex models on a sampled set, particularly if the time complexity of the model building is non-linear—and in most situations, if it is at least N* log(N). A faster model building cycle allows you to iterate over models and converge on the best approach faster. In many situations, time to action is beating the potential improvements in the prediction accuracy due to a model built on complete dataset.

Sampling may be combined with appropriate filtering—in many practical situation, focusing on a subproblem ...