Perché l'accuratezza inganna con dati sbilanciati

Accuracy feels like the obvious score: how many predictions you got right. On balanced data it works fine, but imbalance quietly breaks it.

A dataset is imbalanced when one class is far rarer than the others. Think fraud, disease, or churn: the interesting event is the rare one.

Imagine 99 normal transactions for every 1 fraud. A model that screams normal every single time scores a glorious 99% accuracy. ⚠️

That 99% model never flagged a single fraud. The number looks elite, but the model is useless for the exact job you hired it to do.

Accuracy averages over all rows, so the huge majority drowns out the tiny minority. The rare class can be a total failure and barely move the score.

Always compare against the majority-class baseline: just predict the most common label. If your model cannot beat it, it learned nothing useful.

Before celebrating, ask one thing: what fraction of rows is the majority class? That number is the accuracy a lazy model gets for free.

One line reveals how lopsided your labels are. If one value dwarfs the rest, accuracy alone will mislead you.

Missing the rare class usually hurts most: an undetected tumor or fraud is costly. Accuracy treats every mistake as equal, but reality does not.

The fix is not a magic metric yet, it is a mindset. Stop trusting a single number and start asking how the minority class actually did.

Soon you will meet precision, recall, and PR curves built for rare events. First, internalize why accuracy earned your distrust here.

Why can accuracy mislead on imbalanced data?

On imbalance, accuracy flatters lazy models that ignore the rare class. Beat the majority baseline and judge how the minority class truly did. 🎯