At accommodation platform HousingAnywhere, the safety of our users is paramount. The platform has been engineered in such a way that, provided business is carried out within the secure confines of the platform, scammers cannot cheat users of their money. One of our central tenets is to safeguard users against fraud via our 48-hour policy. When tenants pay the first month’s rent to secure their accommodation, the money doesn’t go directly to the advertiser; first, HousingAnywhere securely holds the money, and only transfer the funds to the advertiser 48 hours after the tenant has moved in. This gives tenants the opportunity to report any fraudulent behaviors.
Fraudsters can be persistent though and will continue to create fake listings in the hope of attracting eager searchers away from the safety of the platform. Like many other platforms, the past years saw us manually identifying and removing scammer profiles to maintain the quality of our service. Unfortunately, too much precious time was lost on this; our customer service team sometimes spent between 70 and 80 hours a week tackling scammers, and with around 200 fake ads popping up per month, it could take ten hours for a listing to even be marked as fraudulent. All this time, scammers could try to convince users to send payments or personal data beyond the safety of the platform. Determined to improve the protection we offer our users, and inspired by the other technological advancements we are making on our platform, we started to track and tackle scammers more efficiently.
The build vs. buy dilemma: should you invest in proprietary tech?
The topic of scammers and fraud prevention is certainly not new. Several companies offer this prevention as part of their service, often in connection with a SaaS platform. There are two widely-adopted approaches: human moderators or automatic vetting. While human moderators can offer up to a 100% success rate, it is hard to find a good provider. Also, this manual approach is inherently expensive and (for HousingAnywhere’s purposes) not scalable enough to match the rapid growth of our platform.
A technical solution was our immediate preference, but off-the-shelf packages are either too generic or based on static filtering, which doesn’t combat the dynamic nature of fraudulent behavior. We realized we needed to build a customized solution, for a perfect fit with the nature of our platform and the complexity of our data. The outlook of this was exciting to us– after all, we’re a tech company armed with the necessary knowledge, expertise, and passion to make this happen. Enter Penelope.
Who is Penelope?
Penelope helps our customer service staff keep scammers away from the platform more efficiently than humans alone could dream of. Penelope, developed in-house, is programmed by artificial intelligence; we trained her with data collected over the past three years, used to document the behavior of scammers on the platform.
Working with data generated by a living platform has proven to be tough (messy inconsistencies, anyone?). We were presented with challenges completely different to the ones you usually face when working on online data science competitions — where someone prepares and cleans the data for you. Simply aiming to actually deploy an AI-driven product has allowed us to grow a lot, both as a team and as a company. In essence, Machine Learning isn’t intrinsically valuable; the value comes only once it’s used to build products for people that actually contribute.
Penelope studies various parameters of newly-created listings, such as the location of the provider, and the location and price of the property on offer. She also studies a multitude of behavioral red flags, to determine if a listing is fraudulent; for example, if a landlord signs up in a city different from that in which their property is located, this increases the likelihood of an attempted scam. There are dozens of indicators of fraudulent practices (which we won’t disclose in full here, as you will understand), that are all taken into Penelope’s assessment. If the probability of a scam is higher than 80 percent, Penelope automatically marks the profile as a scammer and it will be immediately deleted. A probability between 20 and 80 percent sends a notification to the service desk, so that a manual check of the listing can be done. Penelope: the perfect synergy between artificial and human intelligence!
Blocking increasingly sophisticated scams: adding AI layers to detect new behaviors
During testing, when Penelope was still in her infancy, we soon noticed that the algorithms — though particularly effective — seemed to lose some of their efficacy over time. We studied the data, and discovered that the behavior of scammers changed whenever we made adjustments to the platform. In short, scammer strategy evolved at the same rate we did! To account for this, we now train the AI every week, giving it access to an increasingly extensive set of behavioral characteristics. It works with different models, each relating to a specific period in the platform’s past. For instance, there is a model based on behavior learned three years ago which is still relevant today; another model works with data from last year, and a third works with current data, only days or weeks old. From all these models, Penelope calculates a weighted average, which ensures optimum efficiency in blocking out new, suspicious listings, while better predicting future scammer behavior.
Penelope reduced our human workload by >80% and our response time to under just one hour
Since Penelope went live on the HousingAnywhere platform in January this year, our service desk staff only need to check ten to fifteen percent of the original number of suspicious listings. Instead of 80 hours, they now only spend ten to fifteen hours a week fighting scammers. Moreover, the time taken to identify a scammer has been reduced from ten hours to under one, meaning the timeframe a scammer has to contact and convince a tenant, has narrowed; usually, the listings are removed before they even get the chance. Don’t meet Penelope’s standard? She will name and shame you (by sending a notification to our customer service). To this day, a malicious scammer has yet to outsmart her…