Labels, features, and train/test splits

Supervised learning starts with labeled examples. In a support-ticket router, the label might be the team that should handle the ticket. The features are the information you are allowed to use before that decision is known, such as the subject and body.

The practical danger is leakage: a feature that secretly contains the answer. If a training row includes final_agent_team as an input feature, the model may look impressive in practice data and fail when new tickets arrive.

A train/test split is a simple review habit: learn from one set of examples, then inspect performance on held-back examples the learner did not use to make the rules.