I think both of log-loss and f-score can handle the unbalanced data and the f-score is normalized and more interpretable than log-loss. However, is there any advantage for log-loss over f-score?

  • $\begingroup$ They are both answering different questions really. Log loss is an objective function to optimise. f1-score is a measure of classification performance. $\endgroup$ Jan 20 '18 at 16:04
  • $\begingroup$ log-loss measures the quality of probabilistic predictions, while f-score ignores the probabilistic nature of classification. The focus on unbalanced data is specious here. They do very different things. $\endgroup$ Jan 22 '18 at 23:25
  • $\begingroup$ Have a look at stats.stackexchange.com/questions/321333/… for a logloss example. Logloss specifically optimizes the certainty of a prediction and not only the whether it is right or wrong (opposed to f-score). $\endgroup$ Jan 23 '18 at 10:04

The log-loss is a type of loss function. Your model will optimize it (e.g. minimizing) for reaching your goal. Usually it needs to be differentiable, since you could need to calculate the derivative for getting the max or min (these details depend from your optimization algorithm).

The f-score is a metric. A metric is used to evaluate the performance of your model, in this case considering the harmonic mean between precision and recall, but it's not involved in the optimization process and they are not required to be differentiable (so you can't use them for training your model).


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.