When are we satisfied with the performance of a classifier?

Asked Jun 30 '22 at 02:46

Active Jun 30 '22 at 02:46

Viewed 35 times

I am training a multi-head classifier (based on Resnet18) for a multilabel classification task. The dataset I am using to train the classifier is noisy and to improve the accuracy for the worse-performing labels I am grouping classes by looking at the confusion matrix.

Is there a criterium to decide when the accuracy for a selected label is satisfying? I have been told that the accuracy should be above 80% but I cannot find a reference for that. Thanks!

asked Jun 30 '22 at 02:46

randomal

Compare it to a baseline model, whatever that may be for your case, say a simple model or a model with very simple forecasts. – user2974951 Jun 30 '22 at 05:42
1

Does this answer your question? How to know that your machine learning problem is hopeless? – Stephan Kolassa Jun 30 '22 at 05:48
1

Note that accuracy is not a very good evaluation metric. – Stephan Kolassa Jun 30 '22 at 05:49
1

What level of accuracy do you need? Don't let some anonymous, forgotten stranger determine your goals for you! – whuber Jun 30 '22 at 12:16

When are we satisfied with the performance of a classifier?

0 Answers0