Introduction

In previous chapters, we've tackled a range of data challenges using advanced techniques. In each case, we've applied our techniques to datasets with reasonable success.

In many regards, though, we've had it pretty easy. Our data has been largely derived from canonical and well-prepared sources so we haven't had to do a great deal of preparation. In the real world, though, there are few datasets like this (except, perhaps, the ones that we're able to specify ourselves!). In particular, it is rare and improbable to come across a dataset in the wild, which has class labels available. Without labels on a sufficient portion of the dataset, we find ourselves unable to build a classifier that can accurately predict labels on validation or test data. So, what do we do?

The common solution is attempt to tag our data manually; not only is this time-consuming, but it also suffers from certain types of human error (which are especially common with high-dimensional datasets, where a human...