Peformance metrics selection
Question
If you are using machine learning to check whether a particular transaction is fraudelent or legitimate, will you use accuracy as performance metrics? If not then what will you use?
solved
0
Interview Question
4 years
1 Answer
631 views
Master 0
Answer ( 1 )
As most fraud detection data sets are highly imbalanced, accuracy is not a
ideal metric in this scenario. Suppose, you have 5 frauds in a total of 100 transactions.
You can predict every transaction to be not fraud and end up with a accuracy of 95%
and still not detect a single fraudulent transaction.
We can use recall in this scenario which measures that out of the fraudulent transactions,
what percentage of these are correctly identified by our model.
We can use precision to have an estimate that out of the total predicted fraudulent
how many were actually fraudulent.
We can also use F1-score which is the weighted average of both.