Welcome to the OR-Exchange, your site for questions and answers in operations research.

2

I'm interested in hearing methods OR methods for fraud prediction. I'm aware of the classical methods like Logistic Regression. Perhaps there are some other methods.

flag

4 Answers

2

Discriminant models can be used to distinguish fraudulent/suspicious/ok transactions, assuming you have training samples with known classifications. I'm pretty sure data mining techniques are also used for this.

link|flag
What are some good resources and references to OR Discriminant Models? – larrydag Mar 15 at 13:00
I don't know about "OR discriminant models" specifically, just discriminant methods/models in general. In that vein: * <b>Construction and Assessment of Classification Rules</b> by David J. Hand (Wiley) is my favorite reference to discrimination/classification in general. * There was an issue of <b>Annals of Operations Research</b> (Volume 74, 1997) devoted to non-traditional classification methods (which mainly means using a math programming model to directly or indirectly minimize classification errors on a training sample). – Paul Rubin Mar 15 at 19:25
2

Not exactly OR, but Mark Nigrini has shown in his "Benford's Law" book how to detect anomalies in certain data sets that may indicate fraud.

link|flag
2

Some types of fraud have characteristics that make them "abnormal". So a general procedure to detect fraud is to quantify "normality", and then examine the observations that don't fit. This can be done in many different ways, for example

  • Multi-dimensional cluster analysis and then identifying outliers
  • Simple frequency and probability-based rules and thresholds
  • We've even had some fun adapting statistical quality control chart procedures to detecting departures from normal processes that can mean fraud

Of course, this type of approach only works if the fraud cases look different in some way from the non-fraud cases. Many of the discriminant methods suffer from this as well. Clever fraudsters go out of their way to make their actions/transactions indistinguishable from the mass of "normal" data.

A different type of approach to fraud detection that looks very interesting is link analysis, where you examine all the points in common between different entities and try to identify "hubs" that are suspicious. Two examples:

  • Spotting identity fraud, for example on credit card applications. Trying to identify points of overlap between applications, and especially between new applications and known fraudulent applications. Since the fraudster eventually has to get the credit card, there are surprisingly often things like the same address used, or the same last name and switched month and day on the birthday, or the same employer.
  • Analysis of telephone calling circles to see who is an associate of whom, often there are multiple points of contact between people, but at several layers of remove. Eg, we know that Alan and Bob are known fraudsters. They never call each other, but they each regularly call many others, most of whom are probably innocent. However, they both call Charles quite often... maybe we should look more closely at Charles (and maybe not).

Most fraud detection approaches will only be able to narrow down the list of possible frauds from "everyone" to "a smaller subset".

link|flag
I've just noticed that this is a big answer to a simple question, sorry! – David Woods Apr 20 at 15:49
its not a simple answer though so no worries. Keep it coming David! – larrydag Apr 20 at 19:10
It's a good answer and that is what matters – adamo Apr 20 at 20:40
1

This method of Pattern Classification via Linear Programming looks real promising.

http://cgm.cs.mcgill.ca/~beezer/cs644/main.html#Proof

Prof. Godfried Toussaint teaches a Pattern Recognition course at McGill Univ in Quebec. This link is to a student project for the course presented by Bohdan Kaluzny. I'm especially interested in the Breast Cancer Diagnosis example.

link|flag

Your Answer

Not the answer you're looking for? Browse other questions tagged or ask your own question.