Before any learning there is a table. Suppose we want to predict which subscription customers will cancel within 30 days. Each row is one customer, one example. Each column that describes them, such as months subscribed, support tickets opened and plan price, is a feature.
One more column is special: the label, the thing we want to predict, here "cancelled" or "stayed". Our table has 5,000 rows. The model never sees anything beyond this table, so whatever is wrong with it, the model will learn.
💡Good to know
The model only knows what the table shows. Make the table worth learning from.