Classifiers, threshold, and ROC curve

[fog37]

TL;DR

Classifiers, threshold, and ROC curve

Hello,

A classifier is a ML model that can classify between 2 or more classes. Some classifiers are called probabilistic in the sense that they output a probability score that is then compared against a threshold value (usually 0.5) to make the class decision. Other classifiers are not probabilistic...I guess they are called deterministic. We can always plot the ROC curve for a binary classifier. The ROC curve depends on TPR, FPR and various explored threshold values. The TPR and FPR vary for different threshold values...

Do all deterministic classifiers make their decision also based on some set threshold? If so, does it mean that we can plot the ROC curve for any classifier, probabilistic or not?

Thank you!

 

[WWGD]

TL;DR Summary: Classifiers, threshold, and ROC curve

Hello,

A classifier is a ML model that can classify between 2 or more classes. Some classifiers are called probabilistic in the sense that they output a probability score that is then compared against a threshold value (usually 0.5) to make the class decision. Other classifiers are not probabilistic...I guess they are called deterministic. We can always plot the ROC curve for a binary classifier. The ROC curve depends on TPR, FPR and various explored threshold values. The TPR and FPR vary for different threshold values...

Do all deterministic classifiers make their decision also based on some set threshold? If so, does it mean that we can plot the ROC curve for any classifier, probabilistic or not?

Thank you!

I'm not aware of any probabilistic classifier. Usually you just compare the predicted with the actual known value/class of elements in the Testing set., all, like you said, given a threshold, so that, e.g., a threshold of 0.6 will give us a given Confusion Matrix Can you give us examples of probabilistic classifiers?

 

[pbuk]

Some classifiers are called probabilistic in the sense that they output a probability score that is then compared against a threshold value (usually 0.5) to make the class decision.

No, that is not what a probabilistic classifier does: https://en.wikipedia.org/wiki/Probabilistic_classification

Do all deterministic classifiers make their decision also based on some set threshold?

No: first of all the term 'deterministic classifier' is not generally recognised, and secondly you should revise your understanding of this material and consider whether your question makes sense given the diversity of classification algorithms.

If so, does it mean that we can plot the ROC curve for any classifier, probabilistic or not?

Once you have revised this material you should be able to see whether this question is relevant.

 

[WWGD]

No, that is not what a probabilistic classifier does: https://en.wikipedia.org/wiki/Probabilistic_classification


No: first of all the term 'deterministic classifier' is not generally recognised, and secondly you should revise your understanding of this material and consider whether your question makes sense given the diversity of classification algorithms.


Once you have revised this material you should be able to see whether this question is relevant.

Confusingly, Knn is sometimes described as a predictor, some times as a classifier.

 

[pbuk]

Confusingly, Knn is sometimes described as a predictor, some times as a classifier.

Yes, in a field as diverse and dynamic as machine learning categorisation and making generalisations in the way the OP is trying to do is IMHO a waste of time.

 

[WWGD]

Yes, in a field as diverse and dynamic as machine learning categorisation and making generalisations in the way the OP is trying to do is IMHO a waste of time.

Same goes for SVMs, also listed for both Classification and Regression