[D] Classification with Few Examples of Required Class

Generally machine learning algorithms seem to assume you have a lot of data. What techniques can be used to train classifiers where you only have a small number of examples of one of the classes? As an example problem, in fraud detection systems you may only have a few examples of the fraud but you can obtain a very large number of examples of not fraud. Or if you want to produce a text classifier that can find movies you might like by reading the synopsis on IMDB, you may only have a couple of dozen examples of movies that are interesting but can potentially list thousands of uninteresting movies.