USE CASE I: Proactive Credit Card Fraud Management
The goal of this use case is to forecast fraudulent activity and make decisions in order to prevent the financial loss. Fraud is currently detected after it takes place. Decisions about fraudulent activities are made in a matter of minutes after fraud detection. We will forecast fraud and support subsequent sub-second decisions about blocking or reviewing the corresponding transactions.
Fraud detection and forecasting is a needle in the haystack problem as fraudulent transactions constitute at most 0.1% of the total number of transactions. In 2010, fraud in the Single Euro Payments Area (that includes 27 EU member states) was estimated at 1.26 billion Euros. Furthermore, fraud is continuously evolving and therefore the fraud patterns are constantly updated (new fraud patterns appear on almost a weekly basis). Perfect recall (finding all fraud cases) and perfect precision (never raise a false alarm) are out of reach – the state-of-the-art recall and precision rates are about 60% and 10% respectively. At the same time, raising false alarms (that is, unnecessarily calling customers or blocking cards) is very costly in time and customer relationships. Missing true alarms is also very costly (in terms of lost money).
Credit card fraud forecasting requires the analysis of very large noisy data streams storming from all over the world, as well as massive amounts of historical data. Fraud forecasting has to be performed on up to 10 thousand events (transactions)/sec streaming from all over the world, and about 700 million events, if we take into account the history of just 6 months. The event patterns expressing fraudulent activity are highly complex involving 100s of rules and performance indicators.
SPEEDD will be able to successfully detect and forecast the ever evolving fraudulent activities as it will incorporate machine learning techniques supporting the continuous refinement of event patterns expressing fraud. The SPEEDD software implementations will be rigorously tested in the environment of a professional organization that has a clear stake in the solution and a clear path to deploying it.
The dashboard is comprised of three main sections. At the top we can see the stats section. Here is where information about the overall state of system is shown (i.e. total number of transactions investigated, Total number of transactions flagged, average amount and average volume). These numbers update as more transactions are investigated. Clicking the ‘eye’ icon on the top right corner of these boxes will make color the regions on the map based on the respective values. For example, if the eye on ‘Average amount’ box is pressed, the regions on the map will turn green. The color intensity increases as the average amount in the region increases. Clicking on a region on the map and clicking the examine button afterwards brings up more information about the region selected.
The window on the right hand side called ‘Event list' shows the list of patterns flagged by the system. These can vary in certainty as further information is found by the automated system in the background. Selecting a particular pattern from the Event list will highlight (yellow border) the regions on the map where the corresponding transactions have occurred. Moreover, clicking on the Explain button brings up a window containing more information about the pattern in question.
As the demo progresses, transactions are being investigated and, in the top left hand corner, we can see the number of Transactions investigated being incremented. After a short time patterns will appear in the event list.
USE CASE II: Proactive Traffic Management
Traffic detection and forecasting requires the analysis of massive noisy data streams storming from various sensors, including fixed sensors installed in highways and mobile sensors such as smart phones and GPS traces, as well as large amounts of historical data.
The goal of this use case is to forecast traffic congestions before they happen and make decisions in order to attenuate them. We will forecast traffic congestions 5-20 minutes before they happen, and make decisions within 30 seconds of the forecast about adjustment of traffic light settings and speed limits. This will be achieved by enabling the fusion and assimilation of a multi-technology sensor network for real-time traffic data collection.