Which is best performance evaluation metric for Classification problems - accuracy or AUC-ROC?

machine_learning

#1

Hi,

which is best performance evaluation metric for Classification problems
is it the accuracy or AUC in ROC curve

Regards,
Tony


#2

@tillutony- This is completely dependent on the data-set used for classification. If data is not biased you can use ROC curve to check your model performance.
If the data is an imbalanced(skewed/biased) data, accuracy is not a good measure for model evaluation. Use Precision-recall curve, Gini Co-efficient, Gain chart or K-S value for model performance evaluation.


#3

Thanks for the reply .

can u please point me to some resource on the above


#4

@tillutony - You can read the following posts, which explain these techniques in detail.


#5

@tillutony- You can also refer the following URL:


#6

It depends on the data set, if data set is very skewed then accuracy is not a good evaluation matric because in your data set if positive labels are very very less compare to negative labels then if your model is predicting negative class all the time then it will look like your model have good accuracy but it will fail when data set is changed so ROC curve is a good option.


#7

Hi JindallSB,

How will you know the data is bias?

Sufyan


#8

Hi blackberry,

Visualize the class, it is acceptable as long as the gap is not that huge.