Metric Overview

• effective performance of the model in a real setting: it tells you if the model works as expected.
• wrong conclusions when when we deal with Imbalanced Data: don’t use it when classes are highly imbalanced. In this case the F1 score might be better.
  • Example: if a certain class makes up just 10% of the data, a predictor that predicts nothing but the majority class will be 90% accurate, proving itself quite useless in spite of the high accuracy.

<aside> 💡 No reason to over complicate thinks: You can attempt to improve the usability of the accuracy by considering the accuracy relative to each of the classes and averaging them, but you will find it more useful to rely on other metrics such as precision, recall, and the F1-score.

</aside>

You always need to look at a combination of Precision and Recall. They show you what's really happening. But don't make the mistake of looking at them separately!

Kaggle

• Cassava Leaf Disease Classification
• Text Normalization Challenge - English Language