Photo by Maksim Goncharenok from Pexels

5 the most common evaluation metrics used by a machine learning & deep learning scientists that you should know in depth.

Evaluation metrics are the foundations of every ML/AI project. The main goal is to evaluate performance of a particular model. Unfortunately, very often happens that certain metrics are not completely understood — especially with a client side.

In this article I will introduce 5 most common metrics and try to show some potential idiosyncratic* risks they have.

Accuracy

Accuracy metric shows the percentage of good classified predictions in a given dataset versus all predictions. Simply if all classes are categorised correctly the accuracy is 100%. It seems like an easy and good metric, but it also has its caveats.

For example:

We have a model designed to predict fraud cases in a bank. There is a label “1” for fraud transactions and “0” for no frauds. The formula for accuracy is as follows:

accuracy = all_correct_predictions / all_predictions

or:

accuracy = (TP + TN) / all_predictions
TP — True positives (correctly predicted cases — a true frauds)
TN — True negatives (correctly predicted cases as not frauds — non fraud transactions)