Imbalanced dataset

A dataset in which some classes or conditions occur much more often than others.

Imbalance can reflect the real world, but it can also hide weak performance on rare events. Overall accuracy may look strong when the common class dominates. Evaluate relevant slices and error types before deciding whether to resample, collect additional data or change the metric.

In practice

A review process that accepts every label looks accurate when errors are rare, yet catches none of them.