I am playing with a credit fraud detection dataset at Kaggle. An imbalanced dataset with about 0.1% of fraud transaction. The features are 28 PCs out from a PCA exercise done by the data publisher + time & txn amount and a class variable of 0/1 for legit/fraud txn.
From my brief understanding, collinearity should have been dealt with during PCA. However, I found that PCs are still correlated among the fraud cases (if you break the dataset up into legit/fraud cases). What should be a good approach to minimise that effect for fraud detection using a Naive Bayes classifier?
Another thing is that I have been taught in DS101 to deal with outliers. However, I don't seem to think removing outliers is a wise choice given that fraud can be an outlier by itself. What are some common approach to deal with outliers while not removing them?
P.S. I am fairly new to Data Science so any good directions on the above topics would be welcome. It is just not as clear cut as I have seen in the introductory text.