Comparing Models Fairly During Selection

Suppose you have trained a classification model and it scores well on the measure you happened to compute first. It is tempting to treat that number as the answer. The source material pushes back on this habit. The end goal of a machine learning task is rarely a model with high accuracy by itself. A model exists to support some decision or action in the real world, and the way you judge the model should reflect that purpose.

This section is about that choice. Before you can select or compare models, you have to pick the evaluation metric, which is the scoring function that decides which model looks best. The source's advice is direct: make sure the metric you choose is a good stand-in for what the model will actually be used for. The rest of this article follows one prediction at a time, then uses those mistakes to decide which metric to trust.

Take a single email and follow it through a spam filter. Two things determine what happens: what the email really is, and what the filter labels it. Each combination lands in exactly one of four outcomes. Start with an email that really is spam. If the filter labels it as spam, the prediction matches reality and nothing has gone wrong. If the filter labels it as legitimate, the message slipped past the filter.

Now take an email that really is legitimate. If the filter labels it as legitimate, again the prediction matches reality. If the filter labels it as spam, a harmless message has been flagged. Those are the four cells of the picture: a correct flag, a missed spam, a wrongly flagged legitimate message, and a correct pass.

The same classifier can produce either error, so the useful question is not whether it makes mistakes but which mistake you can better tolerate. For a spam filter, a false positive means an important message may be hidden from someone, while a false negative means one spam message reaches the inbox. For a screening model that flags patients as likely to have a condition, the two errors mean very different things for the person being screened. The consequences differ because the situations differ, and the metric should reflect that difference.

The source notes that in practice classification problems rarely have balanced classes. Balanced classes means the categories appear in roughly equal numbers in the data. When one category dominates, a score computed over all predictions can look impressive while hiding the behavior that matters. The dominant category supplies most of the predictions, so getting it right dominates the overall number.

Return to the 100-email inbox. Ninety-nine emails are legitimate and one is spam. A model that never flags anything as spam gets all 99 legitimate emails right and misses the single spam email. Its overall accuracy is 99%, which sounds excellent, yet the category that matters most, spam, is caught zero times out of one. The overall score has hidden the failure entirely.

The practical step is to work backward from the decision. Start by asking what the model will be used for, then ask which errors would hurt in that setting and how much. Only after that should you choose the metric that scores models. The source's instruction is to make sure you understand what the consequences of false positives and false negatives are, and then pick an evaluation metric accordingly. The order matters: the consequences come first, and the metric follows from them.

Applied to the ten-patient screening scenario, the path leads to a clear answer once the consequences are stated. If a sick patient left unflagged is the serious outcome, the reasoning favors the metric that penalizes missed cases, which points toward Model B despite its lower accuracy. If an unnecessary alarm is the serious outcome, the reasoning favors the metric that penalizes false flags, which points away from Model B. Accuracy alone cannot make this choice, because it treats the two errors as interchangeable.