Unsupervised learning might come in many shapes and forms, but the goal is always to convert original data into a richer, more meaningful representation, whether that means making it easier for humans to understand or easier for machine learning algorithms to parse.

Some common applications of unsupervised learning include the following:

• Dimensionality reduction: This takes a high-dimensional representation of data consisting of many features and tries to compress the data so that its main characteristics can be explained with a small number of highly informative features. For example, when applied to housing prices in the neighborhoods of Boston, dimensionality reduction might be able to tell us that the indicators we should pay most attention to are the property tax and the neighborhood's crime rate.
• Factor analysis: This tries to find the hidden causes or unobserved components that gave rise to the observed data. For example, when applied to all of the episodes of the 1970s TV show, Scooby-Doo, Where Are You!, factor analysis might be able to tell us that (spoiler alert!) every ghost or monster on the show is essentially some disgruntled count playing an elaborate hoax on the town.

• Cluster analysis: This tries to partition the data into distinct groups of similar items. This is the type of unsupervised learning we will focus on in this chapter. For example, when applied to all of the movies on Netflix, cluster analysis might be able to automatically group them into genres.

To make things more complicated, these analyses have to be performed on unlabeled data, where we do not know beforehand what the right answer should be. Consequently, a major challenge in unsupervised learning is to determine whether an algorithm did well or learned anything useful. Often, the only way to evaluate the result of an unsupervised learning algorithm is to inspect it manually and determine by hand whether the result makes sense.

That being said, unsupervised learning can be immensely helpful, for example, as a preprocessing or feature extraction step. You can think of unsupervised learning as a data transformation—a way to transform data from its original representation into a more informative form. Learning a new representation might give us deeper insights into our data, and sometimes, it might even improve the accuracy of supervised learning algorithms.