A Neural Fraud Detection Framework Automatic Rule Discovery
- When?
- Wednesday 15 September 2010, 14:00 to 15:30
- Where?
- 39BB02
- Open to:
- Staff, Students
- Speaker:
- Mr Nick Ryman-Tubb
Fraud is a serious and long term threat to a peaceful and democratic society; the total cost of fraud to the UK alone was estimated by the Association of Chief Police Officers to be at least £14bn a year. One such fraud is payment card fraud - to detect this fraud, organisations use a range of methods, with the majority employing some form of automated rules-based Fraud Management System (FMS). These rules are normally produced by experts and it is often an expensive and time-consuming task, requiring a high degree of skill. This analytical approach fails to address the fraud problem where the data and relationships change over time.
An alternative is the use of inductive learning methods, such as neural networks, that learn relationships in the data based purely on learning from examples. The ability to then induce a set of generalising rules that are human comprehensible and accurate is therefore important. A new algorithm has been developed called the Sparse Oracle-based Adaptive Rule (SOAR) extraction algorithm. Experiments recently completed using SOAR have successfully demonstrated that rules can be automatically extracted from a trained neural network, that outperform a leading decision tree approach. The rules are comprehensible and can be used within a fraud environment. The rules discovered relationships of keen interest to fraud analysts who were familiar with the dataset. The results highlighted the need for a more accurate neural classifier and that the SOAR extraction algorithm was not sufficiently sound which has let to key research questions.
