To perform tuning (i.e. to select meta-parameter values) on Classification Tasks usually use such integral criteria, as F_k_measure ( =
(k+1)/( k/recall +1/precision ) ), G_measure ( = ( precision*recall)^0.5 ), breakeven point (point, where precision = recall), etc. But is it correct? In my opinion, no. See, please, my note
Negative Results Is it Good to use F-measure on Classification Tasks?
What is your opinion? Thanks for answer beforehand. Regards, Sergey.
Conference Paper To tune or not to tune: recommending when to adjust SVM hype...
https://stats.stackexchange.com/questions/49226/how-to-interpret-f-measure-values
https://machinelearningmastery.com/classification-accuracy-is-not-enough-more-performance-measures-you-can-use/
https://arxiv.org/pdf/1209.2501
https://pdfs.semanticscholar.org/0b74/9a1ce53c90a37d4b086f7f2cb8b64ed5407a.pdf
Dear Mohamed AbdElAziz Khamis, thanks a lot for your answer. I will carefully study these articles. Regards, Sergey.
It really depends on the problem. For instance, when dealing with rare-events (previous on seizure prediction, a case of rare events in biosignals), and you are really interested in high sensitivity (e.g. correctly predict seizures), the 'beta' value of the F-score equation can be optimized for this specific context. I guess that the take-away message is to adjust the optimization measure to the each classification problem.
- Badges
- Science topic
- Similar topics
- Data Mining
- Classification