In this quick read, the article’s author Raymond Willey notes that, in most classification tasks, the general aim is to simply maximize some measure of accuracy, whether it’s an F1 Score, Balanced Accuracy, etc. Of course, in these cases, we seek to understand the errors for the sole purpose of minimizing their frequency in the future and we want to separate datasets into as clear and distinct groups as possible.
Raymond Willey then asks, but what if we want to do the opposite? What if we have data that is already clearly distinct, but we want to understand how they fit together?
In these cases, Willey notes, we can potentially learn more from the errors than we can from the accuracy levels of predictions. Read the full article at Towards Data Science.