4

Suppose we are building/testing a fraud detection model for a specific credit card/ or a quick cash loan business. We have a lot of data to play with (say past 5years), and after careful preprocessing, model selection, and parameter-tuning, we build a good model to detect/prevent fraud. We thought we did a superb job. However, as we build our model, Con-artists are developing their anti-fraud-detection system/methodology and soon enough, the behavior pattern of frauds become completely deferent. The model we build before become useless,and we need to build new models again...

I have very limited working experiences in building fraud detection models. My question is if there are any machine-learning models/combined models can self-evolve and detect this behavior changing issue, and quickly capture this pattern and adapt? or are there any academic/practical resources regarding this self-evolve AI/machine-learning? Thank you.

DaL
  • 4,462
  • 3
  • 16
  • 27
user6396
  • 423
  • 5
  • 12

2 Answers2

4

The behavior that you describe is called concept drift. Mostly concept drift is due to a natural change in the underlying data. In your case, you are in a game against the fraudsters, so you might want to be even more strict.

In order to know whether your model still performs well, due to a change in the fraud patterns or a natural change, most methods are based on checking it from time to time.

If there is a change, you can relearn if you have enough information. If you don't have enough data (very likely) or if you want to benefit from the previous methods, techniques designed of domain adaptation are very useful.

In case that you want to be very strict in the game, you can assume that your rival have unlimited computational power. Learning in this setting is described at

Learning in the Presence of Malicious Errors by Michael Kearns , Ming Li

DaL
  • 4,462
  • 3
  • 16
  • 27
0

I'm a beginner in machine learning and this is just my opinion but Sequential Bayesian Learning might prove to be an effective approach to this style of problems. Basically you start with some prior probability distribution of the model parameters and in light of the observations your prior distribution changes to a posterior distribution. Now you can discard this used data and choose this posterior to be your new prior and given the new observations in the data this changes to posterior.And this goes on. So as we keep on providing new data our model tries to adapt rather than the likelihood approach where the model parameters are fixed after optimising a loss function.

Sharad
  • 1