The Basics of Classifier Evaluation, Part 2

A previous blog post, The Basics of Classifier Evaluation, Part 1, made the point that classifiers shouldn’t use classification accuracy — that is, the portion of labels predicted correctly — as a performance metric. There are several good reasons for avoiding accuracy, having to do with class imbalance, error costs, and changing conditions. The next part in this series was going to go on to discuss other evaluation techniques such as ROC curves, profit curves, and lift curves. This follows the approximate track of our book, Data Science for Business, in Chapters 5 and 6. However, there are several important points to be made first. These are usually not taught in data science courses, or they are taught in pieces. In software packages, if they’re present at all they’re buried in the middle of…


Link to Full Article: The Basics of Classifier Evaluation, Part 2