You are building an app that can recognize handwriting. To train your model, you use a very large dataset with millions of handwritten letters.
You calculate the macro accuracy of your handwriting app and get a value of 0.7. Then you calculate the micro accuracy, but now you get a value of 0.1.
What is causing this huge difference?
Choose the correct answer from the options below.
Please select at least one answer!
Congratulations, you are correct!
I'm sorry, your answer is not correct.
The dataset is unbiased is incorrect. A gap between micro- and macro accuracies means that the dataset is biased.
The dataset is biased, and the accuracy on the popular classes is much higher than the accuracy on the rare classes. is incorrect. For this to be true, the micro accuracy would have to be higher than the macro accuracy.
The dataset is biased, and the accuracy on the popular classes is the same as the accuracy on the rare classes. is incorrect. For this to be true, the micro accuracy would have to be roughly equal to the macro accuracy.
The dataset is biased, and the accuracy on the popular classes is much lower than the accuracy on the rare classes. is correct
Check out my video lesson to learn more about multiclass classification.