What is classification?
Classification is the supervised learning task of predicting which category a record belongs to. The categories, called classes, are fixed in advance, and the model learns from examples whose class is known. The output is usually a probability or score for each class, which a confidence threshold turns into a decision. Regression is the sibling task for predicting numbers.
Classification is the most common form of machine learning in business because so many decisions are sorting decisions, such as approve or refer, defective or acceptable, and likely to churn or not. Any algorithm that produces these decisions is called a classifier.
Types of classification
- Binary classification chooses between two classes, such as fraud or legitimate, and is measured with precision, recall, and the ROC curve.
- Multiclass classification chooses exactly one of several classes, such as which of eight support queues should receive a ticket.
- Multilabel classification allows several classes at once, such as tagging a contract with every clause type it contains.
- Hierarchical classification assigns a record to a tree of categories, such as a product taxonomy, level by level.
Classification in practice
A managed service provider receives several hundred tickets a day by email and portal. A text classifier trained on two years of resolved tickets assigns each new one to a queue (network, endpoint, identity, applications, billing) and a priority. Tickets the model scores with low confidence go to a dispatcher. The team reviews a confusion matrix each month, which shows that most mistakes confuse network and identity tickets, and adds clarifying fields to the intake form in response. Response times improve because tickets reach the right engineer on the first hop.
What to watch for
- Accuracy alone misleads when classes are unbalanced, since a model that labels every transaction legitimate scores high accuracy and catches no fraud.
- Look at the confusion matrix to learn which classes the model confuses, since those errors often point to unclear category definitions.
- Set thresholds from the cost of each error type, and route low-confidence cases to a person.
- Class definitions drift as the business changes, so retrain when new products or ticket types appear.