Fraud Analysis and the Value of Layered Modeling

Identifying fraud prior to its occurrence is an imperative for the industry of ecommerce to survive and be profitable. Problems with identifying fraud are threefold: fraudsters don’t commit fraud all the time, identifying prior fraudsters can be difficult since they make attempts to change their identity, and fraud is a rare but impactful event that typically costs several times more on average than a standard transaction. This creates a problem of high false positive rates and identification of fraud for standard logistic models difficult since it is easy for a model to have a seemly high accuracy output due to the conflation of rare events and the smoothing of data to lose sensitive signal analysis. Python holds several key packages within TensorFlow, FuzzyWuzzy, and Smote that in combination allow for the measurement of signal analysis when fuzzy logic is necessary to identify a previous fraudster and reduce the false positive rate when dealing with new customers. This, in combination with a collaborative filtering process, to identify fraudsters rather than fraud events, and a Bayesian weighted system for multiple model inputs, gives a high accuracy rate that overcomes the issue of identifying rare events with minimal false positive impact to customers.