Features & Labels

Every column gets a job

Open any dataset and each column must be assigned a role. Features (X) are what the model sees at prediction time: tenure, monthly bill, support tickets, email text. The label (y) is the answer to learn: churned or not, spam or ham, price. This split sounds trivial and causes half of all beginner disasters — because a column in the wrong role poisons everything downstream.

Features must exist at prediction time

The golden rule: a feature is legitimate only if you'll actually have its value when predicting. cancel_date when predicting cancellation, refund_issued when predicting refunds — these are leakage: the answer smuggled into the inputs. Models with leaked features post miraculous 99% scores and fail completely in production, where the smuggled column doesn't exist yet.

Choosing features is the actual skill

Domain sense beats algorithms here. For churn: declining usage matters more than raw usage (trends over snapshots), support-ticket sentiment beats ticket counts, tenure interacts with everything. Two habits of strong practitioners: engineer a few smart features before trying a fancier model, and drop features that are proxies for the label in disguise.

Remember this

• Features = inputs available at prediction time. Label = the answer. Never confuse them.
• "Would I know this when predicting?" — ask it about every column.