As mentioned in previous article, if we want to make instant decision about a transaction, application or claims, we need to implement a rule engine such as FICO’s blaze advisor, IBM’s ODM, SAS or Redhat’s decision manager. Before we implement the rule engine, we need a mechanism to generate the rules. There has been different AI software from different companies such IBM, SAS or FICO. But each of these comes at an additional cost in addition to rule engine. There is extra cost of support for them. Below table gives a price indication of it.
|AI Engine||Price Range||Support cost||Ease of rules Generation|
|FICO||Between 100K to 200K depending upon usage||800 GBP per day||Simple|
|SAS||Between 100K to 200K depending upon usage||800 GBP per day||Easy|
|IBM||Between 100K to 200K depending upon usage||800 GBP per day||Easy|
|Fraud Analyser**||Included in Fraud solution, free of cost||Included in Fraud solution, free of cost||Easy with Data Scientist support|
**Fraud Analyser is our fraud solution.
These rules are implemented by the expert knowledge based on logical conclusions derived in history and from AI engine if customer purchase it. These rules are updated on yearly basis. So, there is good chance in one-year duration by acquiring the knowledge of the system and invent new techniques to do frauds. Incidentally, fraud happens, and it goes undetected. So, there is requirement of solutions that are updated dynamically every quarter based on rules detected from recent frauds.
We have implemented the solutions based on machine learning techniques for few of our clients in Africa based on machine learning techniques. Rules are detected on monthly basis using statistical techniques. There are two kinds of rules i) based on past trends ii) based on current trends. The rules based on past trends are discovered by building a fraud score card on last 1 to 3 years historical data. The rules based on current trends are discovered using the anomaly detection techniques such as outlier detections.
To detect the past trends using historical data, we collect the historical data of close to last two years. We use the observation windows of one year and performance window of 6 months to 1 year depending upon when the fraud rate stabilizes. We derive lots of trend variables using the data in 1-year observation window. Then we build the behavioural score card using latest machine learning techniques such as random forest, neural network and gradient boost method. The performance parameter such as accuracy ratio and area under curve for score card is maximized. The performance parameter such as false positives and false negatives are also minimized.
To detect the current trends using current data, one of anomaly detection techniques such as fraud rings or outlier detection is used. Fraud rings describes the association between fraudsters and association between non fraudsters. If claim is found to be part of fraud network, it is thoroughly investigated. Outlier detection happens at claim level first. Once a claim is detected an outlier using un-supervised techniques such as k-means then it is investigated further based on key variable’s data to find the reason of being anomaly.
Both kind of rules based on past trends and current trends are applied to a new claim and results are reported to claim department. If a claim is reported to be fraud, then its investigation report is passed to fraud investigation team. Below is our comparison of Fraud Analyser with other tools in the market.
|AI Engine||Rules update time||Trend detected||Cost of update|
|Fraud Analyser**||Monthly or quarterly||Past and current||Included with subscription|
**Fraud Analyser is our fraud solution.
Both kinds of rules based on past trends and current trends are checked for consistency on monthly basis. PSI and CSI of key variables is compared in the latest data with data used for building the score card, if there is no significant difference then actual fraud rate is compared predicted fraud rate. If there is no significant change then we keep the same score card and rules are not updated. In case there is significant change in the both the parameters then we must build a new scorecard and rules are updated to keep the score card and rules relevant.
By using the latest machine learning techniques and keeping the process dynamic, we managed to achieve excellent results for our clients in Africa. We managed to detect 30% more fraud than rule-based management system. The machine learning techniques can complement the existing techniques based on expert knowledge, but they cannot replace the years of industry experience. Hence our solutions are a complement to existing solutions.