I have a data set of transactions and want to build a fraud detection model (classifier). Only 3 variables are given that could be used as input features. The number of transactions during past 3, 6 and 12 months. These three features are highly correlated. I want to use the information content in all three features as much as possible. What is the best way to handle their great amount of correlation.

  • $\begingroup$ My thinking is you really haven't defined the problem well. Provide more info. Are you looking to do Regression or Classification? Do you have only 3 features or more? Are you looking to reduce the number of features or use partial least squares to your advantage? $\endgroup$ – oaxacamatt Oct 6 '20 at 2:18
  • $\begingroup$ @oaxacamatt Thanks for your comment. I elaborated accordingly. $\endgroup$ – user2348674 Oct 6 '20 at 2:24
  • 1
    $\begingroup$ @user2348674 if there are only 3 variables and they are all highly correlated then I'm afraid you can't build any model. Correlation detection and treatment is a dimensionality reduction method. Say you set a cut-off point of 0.75 and find two out of three variables meet the criteria. Then are you going to build a model using just one variable? $\endgroup$ – mnm Oct 6 '20 at 8:51