Two models have identical class predictions but different log loss. What extra behavior does log loss distinguish?
Classic MLLevel: IntermediateConfident mistakes
Two models have identical class predictions but different log loss. What extra behavior does log loss distinguish?
Hints
- Look at probabilities before thresholding.
- The logarithm magnifies near-zero true-class probability.
Answer
It evaluates assigned probability to the true class. A confident wrong prediction is penalized much more than an uncertain one, which hard-label accuracy cannot distinguish.
Source and licence: scikit-learn 1.7.2 — model evaluation · BSD-3-Clause · answer: Quantoros editorial