How To Use Unsupervised Machine Learning To Prevent Promotion Abuse
Reduce losses, restore campaigns, repair partnerships, and power growth by focusing on what matters most — good customers.
Conducting business in the modern, digital economy is a complicated process that involves, among other challenges, addressing ongoing tensions between powering growth and incurring risk. Online promotional campaigns present an excellent example of this challenge, and many businesses continue to struggle with the question of how to roll out successful promotions without exposing new vulnerabilities.
Fortunately, transformational technologies such as unsupervised machine learning (UML) can be used to build solutions that support growth and mitigate risk — in real time, and at scale.
Promotions and Promotion Abuse
Offering promotions or coupons is a tactic used by companies across many sectors to attract new users, and ideally, convert those users into long-term, loyal customers. Using promotions to drive business growth is particularly common in the travel and hospitality industries. These tactics can be very effective, but an unfortunate side effect of this approach is the invitation for fraudulent users to abuse these promotions for their own financial gain.
Promo Abuse in the Travel Industry
For travel metasearch and reservation sites specifically, promotion abuse can result in losses not just for the site, but for its accommodation partners. Good users suffer as well, as their experiences are negatively impacted. Accommodation partners are hit with no-show reservations and failure to collect payment for a stay. At the same time, good users are cheated out of their ideal reservation opportunities by fraudsters blocking reservations en masse.
The Challenge
Effective solutions for preventing damage associated with promotion abuse have historically been elusive. Existing solutions that leverage rules or supervised machine learning are not able to keep up with ever-changing fraud attacks. These solutions require a large volume of labels to train a model or identify patterns to create rules in order to block fraudsters’ attempts. By the time these systems are implemented, fraud attack patterns have already evolved to get around the new rules in place.
The Solution
DataVisor’s UML-based approach, enhanced with deep learning, natural language processing, and a Global Intelligence Network of consortium data for digital info, is able to identify correlations and connections across events to detect large coordinated fraud attacks. Applying this methodology to the issue of promo abuse through travel sites, we can capture fraudulent users with high precision. Fraudsters are detected by various digital signals — IP address, device info, similar email and naming patterns — in addition to their behaviors.
A Client Success Story
As a recent example, we worked with a travel platform, operating in over 200 countries and territories, that was facing challenges with fraudsters abusing their promotions. Using our UML-powered approach, we were able to detect malicious users at the moment they were attempting to make fraudulent reservations. In this way, the site was able to automatically block these reservations from going through. This approach ensured that bookings were not made by fraudulent users looking to sell these reservations off-platform, and accordingly served to preserve the good user experience.
Among the many problems associated with fraudulent bookings was the fact that they often resulted in “no-shows” that caused financial losses for hotel partners. We encountered one fraud ring of fraudsters that would mass-reserve hotel rooms in the same area for the same dates corresponding to major events (such as New Year’s Eve in NYC) during which hotel availability is already scarce.
Working with DataVisor, our client was able to significantly reduce fraud losses, restore promotional campaigns, and rebuild partner relationships. Today, our client now detects 40% more fraud than before, with over 93% accuracy. Most importantly, they’re capturing more than 70% of all attacks early, before any damage can occur.
Labels to Enhance Performance
UML’s ability to produce actionable insights from raw data almost immediately — without a hindering reliance on pre-existing knowledge, labels, or rules — is one of the technology’s most valuable benefits. However, the use of labels can, in fact, even further optimize performance.
In some instances, for example, a UML model will capture good user groups that exhibit similar behavior or digital signals that are most often associated with fraudulent users. To manage situations like these, we look to our client to provide additional context so that we can correctly identify any false positives, or any good user groups that were initially flagged as suspicious by DataVisor’s model.
Using labels, we can further enhance our model to ensure that nuanced differences in certain good user behaviors are not captured along with fraudulent users. It is vital to incorporate each client’s business knowledge into our model, and we work closely with our clients to incorporate their unique understanding of their business logic with our domain expertise and advanced technologies to provide the best fraud solution possible.
How UML Works to Expose Otherwise Undetectable Fraud
Here is one example of how our solutions were able to surface and expose suspicious activities that would otherwise have gone undetected by more conventional solutions.
After analyzing a number of captured fraud rings that included apparent ‘good’ users (based on their prior reservation history through the travel site), an interesting trend emerged. Groups of users were detected as having made numerous reservations from the same IP address using the same non-standard email provider. Further investigation of the email providers led to various websites offering hotel or flight reservations — many unofficial or small-time fare aggregator sites were routing their reservations through the initial travel site in order to benefit from the promotions. While these sites previously flew under the radar due to their good reservation histories, DataVisor was able to detect this coordinated and fraudulent behavior through the observation of patterns and correlations that would not have been identified through individual account or transaction analysis.
Conclusion
DataVisor’s unique approach to fraud detection makes it possible to identify fraudulent behavior in real-time, at the moment a reservation attempt is made, allowing sites to immediately block that reservation from being confirmed. Not only does this prevent immediate loss and damage, but it also provides an early warning to discourage further attacks.
By analyzing all user history and actions across multiple entities, our solutions can provide a big picture view into a user’s behavior and produce unique insights into suspicious behavior and platform abuse. This approach to fraud detection can even expose previously unknown attack types.
Given the travel industry’s global footprint, its migration online, and the importance it places on positive user experiences, businesses in this industry must be able to provide safe, secure, and friction-free platforms. DataVisor’s advanced, AI-powered fraud management solutions make it possible for organizations to focus their energies on what really matters — good customers.
~