Say you have created a mood classifier for several different domains. Say movies, music DVDs and electronics. It’s easy to build high quality classifiers because there are tons of training data that you have cleared from Amazon. Along with each classifier, you also create an affinity detector that tells you about a given piece of text, how it looked like a dataset on which each of the classifiers is trained.
Now you want to find the mood of some text from an unknown domain or one in which there is no such excellent dataset for training. Well, what about the fact that we include a balanced set of classifications of the three high-quality classifiers that we already have. If we try to classify a dishwasher overview (for example, there is no giant enclosure of dishwasher reviews), it is probably the most similar to electronics, and therefore the electronics classifier will have the greatest weight. On the other hand, if we try to classify a TV show review, perhaps the movie classifier will do a better job.