Dinesh,

One of the problems that fraud presents is that by its very nature it is hidden. A common assumption is that fraudulent activity is an outlier, but this may not be the case (that is why fraud is so difficult to detect). That being said, once you have the specific type of fraud it may be a matter of encoding the specific fraud signature for detection.

To show you the difficulty of detecting a transaction fraud, imagine that you are targeting a fraudster that is trying to evade collection and he passes ownership of his most valuable asset to a relative , how would you encode that in a vector? one possible solution is to encode it as passed asset to relative , time of collection notice, time of ownership transfer. While this may be a valid encoding the question is: is it a good detection technique for this kind of fraudulent transaction, and are there other efficient ways to evade detection? (did the person convert the money in cash and dug a hole to hide the money? did he file banckruptcy)

While I have no doubt that any classifier can correctly classify the vectors described above, the problem lies with the coverage of the samples and the unique ways that fraudsters can use to evade the classifier once the detection technique is known ( assume that you are up against an adversary trying to evade the detector).

Hope this helps

I totally agree with Arturo, fraud detection is often treated as an outlier detection problem where a normal model will be trained based on the history of user's valid transaction. This model can be probabilistic which outputs a high likelihood for a normal test data, but whenever any fraudulent transaction is made, or any in case of anomalous activity it will output a very low likelihood which is indicative of a possible fraud. This can easily be performed using a Gaussian mixture model. Alternatively, you can use deterministic models or classifier like one class SVM which creates a sphere around normal data, and any fraud lent activity would lie outside of this sphere when passed onto this model.. Hope it helps..