What is unsupervised learning?
The Harvard Chan AI Glossary
In our classes, we often pause to explain the ideas behind AI and the language used to describe it. These posts grew out of those conversations. They’re for anyone who would like to understand AI better, with no technical background assumed.
Unsupervised learning is a way for computer systems to discover patterns in information without being given target answers or category labels. The information might be photographs, purchases, songs or written documents. The system looks for regularities, such as items that resemble one another or measurements that tend to change together. It is a method of machine learning, which is a method of achieving artificial intelligence.
Suppose you have a large music collection with no genre labels. You could compare songs using their tempo, loudness and other sound characteristics. An unsupervised method might arrange similar songs into groups. This is called clustering: putting examples together according to a chosen measure of similarity. A person can then listen to the groups and decide what they mean.
The system has not necessarily discovered rock, jazz and classical music. It may have separated quiet recordings from loud ones, or older recording styles from newer ones. What it finds depends on the information it receives and how similarity is calculated. Changing those choices can change the groups, even when the songs themselves remain the same.
Other unsupervised methods simplify complicated information. A collection might describe each song using hundreds of measurements. A method can combine those measurements into a smaller set that preserves useful differences. This is called dimensionality reduction: describing the same examples using fewer numbers. It can make patterns easier to explore, although some information will be lost.
“Unsupervised” does not mean people have no involvement. People choose the information, the method and the purpose of the analysis. They also decide whether the result is useful. A neat-looking diagram is not evidence that the groups reflect a meaningful distinction in the real world.
The contrast is with supervised learning, where examples include target answers. In public health, an unsupervised analysis might group responses to a survey about difficulties attending appointments. Some groups could suggest transport problems, inconvenient opening hours or trouble arranging childcare. These would be starting points for investigation, not established patient categories. Researchers and community members would need to examine the responses and decide whether the groupings reveal meaningful barriers or merely differences in wording.
Sources: scikit-learn — Clustering; scikit-learn — Principal Component Analysis.
If you would like to find out more about AI and improve your leadership skills, consider the AI in Health Care Certificate of Specialization from the Advanced Learning Academy at Harvard T.H. Chan School of Public Health. You can take an individual course or work toward the full certificate.
AI is a fast-moving field. This explanation reflects our class discussions and related material at the time of writing. Terminology, capabilities and tools may change.
For suggested additions or corrections, please email me at tpanch@hsph.harvard.edu.
Dr. Trishan Panch
Program Director, AI in Health Care
Harvard T.H. Chan School of Public Health — Advanced Learning Academy