What is classification?
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.
Classification means assigning something to one or more defined categories. In AI, a classification model is a system trained to make those assignments using examples. The categories could be digits from zero to nine, types of customer request or whether an email should be treated as unwanted.
Recognizing a handwritten digit makes the task concrete. The system receives a picture of a number and proposes a category, such as “7.” It does not need to write a description of the picture or explain the person’s handwriting. Its job has a constrained set of possible answers.
During training, the learning process, examples are usually paired with their intended categories. Those answers are called labels. By comparing its predictions with the labels, the model’s internal settings can be adjusted. The aim is to handle new examples, not merely remember the pictures already supplied.
Some tasks have two categories, such as unwanted email and ordinary email. This is binary classification: a choice between two defined classes. Others have several alternatives. A separate possibility is that more than one label can apply at once. A photograph can contain both a bicycle and a pedestrian; forcing it into only one category may misrepresent the task.
A classifier may produce scores before selecting a category. A decision threshold is the point at which a score triggers a particular choice. For example, a message might be held for review only when its unwanted-mail score exceeds a chosen level. Moving that threshold changes the balance between missed unwanted messages and ordinary messages incorrectly held back.
The categories themselves deserve attention. “Other” may contain many different situations, and some inputs may not fit any available label. A useful system may need an “uncertain” option or a route to human review rather than being forced to choose confidently.
For a clinic’s incoming administrative messages, classification could assign labels such as appointment change, billing query or prescription renewal request. More than one label might apply, and uncertainty may need its own route. Predicting the total number of messages tomorrow would be regression; discovering unanticipated groups in the messages would be clustering. These are different tasks. Before evaluating the classifier, agree what each label means and how a mistaken assignment would affect the staff member or patient waiting for a response.
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