Pitfalls of Overemphasizing Model Accuracy

Welcome to today’s data science interview challenge! Here it goes:

Question 1: What can go wrong if we try hard to shoot for high confidence or accuracy when building models?

Question 2: In a computer vision problem of classifying types of fish (Kaggle fishery monitoring competition), if you realize that your model is recognizing boats rather than really understanding the difference among fish types, what went wrong and how should you think about fixing it?

Source

Here are some tips for readers’ reference:

Question 1:

There’s a common misconception that a higher confidence score means that your model is more accurate or better. That’s not always necessarily the case. In fact, if you are consistently seeing very high confidence, your model is probably overfitting.

Trying to maximize model confidence is a common error that could result in overfitting.

Let’s see how Rasa Senior Developer Advocate Rachael Tatman explains it: