For Ricardo, the security of the platform has always been essential. That's why they put hours of manual work into identifying fraud attempts and fraudulent offers. Caroline Nef from Ricardo and Dominic Herzog from TX Group joined forces to implement an efficient and automated process.
Objectives and Implementation
At first, the problem had to be identified and described. The focus was mainly on the time required to manage the previous security tool, which was too high at around 60 hours per week.
They wanted to reduce manual work by at least 30% while identifying and blocking at least the same amount of fraud attempts.
An in-house solution was the only option for Ricardo. This way the costs could be kept lower and the internal data processing know-how could be secured.
Human and Machine Resources
Part of the challenge was to optimize the existing security and fraud tools to identify fraud attempts. In order to further develop these tools efficiently, the long-standing know-how of the employees was taken into account. In this way, already identified patterns could be added to the tool's algorithm and it did not have to learn them first. This procedure enabled the machine learning to concentrate on the recognition of additional patterns, and rule adjustments no longer had to be made manually.
Ricardo could not predict to what extent the chosen methods could actually help achieve the set goals. It was also important that all customer service employees could understand the new rules. Therefore the process had to be documented understandably.
Decision Trees Meet the Target
Improvements have been made in the area of account linking, which makes it more difficult for a banned user to create a new account. Additional rules have been created for behaviour that experienced customer service representatives have identified as indicative of fraud. For example, potential fraud accounts often do not use their own images of the item being offered, do not accept cash on collection and often sell trendy items at disproportionately low prices.
The internally trained "Random Forest" model (for more information on the model: link), analyzes user behaviour on the platform. This "Random Forest" approach is a classification procedure and consists of a large number of individual, uncorrelated decision trees that work together. Each tree in the Random Forest generates a prediction about the group to which a user behaviour belongs. The user behaviour is assigned to the group with the most predictions.
The model's input data includes buying and selling activity, profile data changes, ratings, browsing behavior, and more. Through this approach, 80% of potential fraud cases are successfully assessed by the model in an automated manner and only about 10 cases per hour need to be forwarded to customer service for manual review.
Automation made the fraud process more efficient and potential fraud attempts are evaluated more quickly.
Together Against Fraudsters
The hybrid approach has proven to be very efficient. Ricardo was able to build on the knowledge of the customer service staff and transform their human pattern recognition into meaningful rules. Subsequently, the training of the model could specialize in assessing the conspicuous user accounts. Using the Random Forest model, the algorithm's decisions could be reproduced. This was very important for the customer service, especially in the beginning, since the employees have to justify the blocking of a user account to the customer. Last but not least, the whole process was characterized by a close collaboration between the data science team of TX Group and Ricardo's customer service. Short feedback loops made it possible to continuously adjust parameters without flooding the essential customer service tools. This way, even in case of unplanned events, customers were not affected at all.