A Data Science Central Community
Hi jaap Vink,
Thanks for your good information. Iam making use of this PASW Modeler for predicting fraudulent customers. I have generated a model using around 3000 records of customers along with their fraud history and used this obtained model in a new stream for new single customer data. It is going to give predicted result(i..e T/F) and probability value. Am i doing in correct way in predicting new customer based on previous customers? if not please suggest me a solution. Any suggestion may help us.
Thanks in Advance,
It sounds like what you're doing is correct. Fraud can sometimes be tricky due to the low frequency of fraud (hopefully your business has only a handful of fraudulent customers).
Am I correct in assuming that for the 3000 customer records you're using to build this model, you have BOTH some customers who you know committed fraud and some customers who did not? If this is the case, then you would have a binary (Yes/No) variable that indicates whether the customers committed fraud -- this is your "target" variable. How many fraud and non-fraud customers are in your 3000 customer sample? What types of variables are you using in the predictive model (e.g., is it transaction history, demographics, etc.)? How you code your predictor variables can sometimes be important for both model building and model scoring.
If you don't have a binary fraud target variable, there are still some good things you can do. But I would need to know more about your data and business situation to help you. You can contact me directly at [email protected] if you want to discuss this further. We're a small analytic consulting firm, and we've conducted analytics to identify fraud for clients in several industries.
Good luck! -- Karl Rexer, PhD
Thanks for your quick reply. Iam going to contact you through my personal-id, please respond for that one.
Thanks in Advance,