Consortium data at Ravelin: A nuanced approach

Consortium data – which cross-references data points across a database of merchants – is often used to generate fraud scores. But nuance matters when you’re making decisions that can affect the bottom line, growth and reputation of a company.

We have introduced nuance to Ravelin’s consortium data approach within our AI-native fraud detection, so that our merchants can take advantage of what we know across our network of merchants without falling into common consortium data traps.

This means we’re employing a nuanced approach, where features take into account the context, recency and objectivity of consortium data while also ensuring data is anonymized and hashed.

Instead of treating all this information as of equal importance, we’ve built our fraud prevention to consider which data is most reliable and most recent – ultimately resulting in better-performing machine learning models.

The traditional approach to consortium data

In fraud prevention, consortium has been on a bumpy ride: Touted by some vendors as a be-all-and-end-all, these shared databases have brought in mixed results for merchants. There are even anecdotes about companies misrepresenting a good customer’s legitimacy just to prevent their competitors from selling to them.

At Ravelin, consortium features have long been a building block in our multi-layered, multi-model approach to fraud detection.

Today, our consortium database includes more than 9 billion identity elements, which feed into consortium features, including data such as:

• Payment method type

• Card BINs

• Order end and start location

• ASN score

• Email address

• IP address

• Card issuer

• …and more

Consortium features can help to build up a picture of whether someone is a good customer or not. However, how important consortium features are for each machine learning model varies from client to client and across industries.

A traditional approach to consortium data might have looked at all the requests and combine all different data points as if they were equal – for example, they treat manual reviews and dispute/chargebacks in the same way, although one largely depends on human opinion and the other on fact.

Of course, all this information is useful to have and utilize. But this traditional approach lacks nuance.

And a lack of nuance in consortium features can lead to false positives, one of the biggest risks to a merchant. No merchant wants to miss out on a legitimate sale.

For example, for some models, maybe knowing a dispute score for the latest data point is a more powerful feature? Or perhaps the order-to-refund ratio is a better signal for refund abuse than disputes and manual reviews? And, for shoppers, should they be blocked from buying a television just because their cold food was refunded twice?

This was the starting point for Ravelin’s updated, nuanced approach that puts consortium data in context.

Introducing nuance to consortium data for better recommendations

We recently introduced almost 100 additional consortium features that we can use to inform our fraud recommendations.

But we’ve also fundamentally changed how we calculate consortium features so we can take things like recency, objectiveness and underlying behavior into account.

1. Characteristics instead of static identities

Fraudsters are constantly recreating their personas to evade fraud defenses. Burner email accounts, emulated devices and fake IP addresses, among other techniques, can conceal their identity.

Plus, more and more good customers are becoming privacy-minded, opting for guest checkout and for sharing as little as possible with a merchant.

Therefore, it wouldn’t be so helpful to look at static data in the consortium database.

For Ravelin, it’s the characteristics of these fraudsters that are important, not their static, full identities. Our consortium features analyze these data points to assess which characteristics are more likely to be linked to fraud.

For example, it’s not only the specific email address used by a fraudster that’s important but how risky a specific type of email or email domain has proven to be across our network. Some providers make it easier than others to generate hundreds of email addresses, which can then be used to commit fraud.

Context like this is useful to take into consideration within machine learning recommendations.

2. Splitting objective fact from subjective opinion

We have split consortium features out into two scores: one based on more objective data and the other based on subjective inputs, because not all information in the consortium network is equally reliable.

Some data sources are influenced by human opinion. A manual review can be subjective. The kinds of customers that are manually reviewed as fraudsters – and the reasons for this – can vary across merchants. Plus, there’s potential for false positives and false negatives.

Ravelin’s ML fraud prevention models now take into account which data in the consortium is objective and which subjective instead of treating it all as equally reliable.

• Objective data and features: Based on historic behavior linked to hard data. For instance, when a customer has disputed a payment and is requesting a chargeback from their card-issuing bank. Here, there’s no gray area – the claim has been received.
  
  

• Subjective data and features: Manual review conducted by human analysts has the potential to be wrong. And labels can be used differently across merchants. This information is therefore subjective, and the features that make use of it ought to be considered as such.

Ravelin’s ML models make a distinction between these two types of consortium data, to the benefit of more accurate fraud scoring.

For example, perhaps for some clients or verticals, manual review-based scores are a stronger indicator than for others. This update allows us to take advantage of this fact.

3. Recency is important – so is the latest state

We look at all the data points for each customer, giving us a historic view of their behavior. But recency is also important, so we have a separate score that considers just the latest data point.

It’s not just about whether someone has been linked to fraud in the past. It’s also important to consider how long ago this was, and what has happened since.

Recency of consortium data is key. A stale consortium database can misrepresent the situation and a customer’s intentions, resulting in false positives and thus lost sales. So we have specific cut-off points for storing our data, also subject to applicable legislation.

It is also important for consortium features to reflect the latest state. Disputes can be forgiven, reviews can be undone, and good orders can be placed by customers who have previously been suspected of committing refund abuse.

It would not be accurate or fair if this was not reflected in the score. This means consumers are not penalized based on outdated inputs.

4. Transparency of consortium scoring to empower analysts

In updating and adding nuance to our approach to consortium data, we have ensured we can provide as much context and transparency as possible.

Ravelin’s Dashboard also now surfaces information about how much consortium features have contributed to fraud scores and recommendations.

In fact, you can select the Consortium megafamily on the Dashboard to view a further breakdown of the consortium data points which contributed to a customer’s overall score, and by how much. You can see an interactive demo of this on the Ravelin platform here.

Note that we only use the data we need to build features. Any personally identifiable information (PII) is anonymized and hashed.

Features don’t indicate which client the data came from, and we have ensured no company would ever be able to use the data we provide to attack or sabotage its competitors.

5. More consortium features for improved results

True to our mission to continue improving and updating our solutions, we have introduced new types of consortium features.

The 100 new features that we recently added to Ravelin’s machine learning include:

• Refund-specific consortium features: Refund abuse is on the rise and eating into companies’ profits. These ML consortium features help catch refund fraud and abuse, and can indicate whether a first party is more likely to commit friendly fraud and opportunistic abuse.
  
  

• Email-focused consortium features: These features identify shopper email addresses linked to historical fraud across Ravelin’s network of customers. But beyond this, they also give us a general idea of the fraud rates of a certain data point across the board – indicating whether it’s part of a setup that fraudsters prefer over others.
  
  

• Breached credential features: It’s important to know when a customer’s PII has been leaked online, as it’s more likely that their online shopping history would involve fraud – perpetrated by bad actors who took over accounts or stole cards. These consortium features are built around Ravelin’s internal breached credentials database.

Related resources

• AI fraud detection with ML & NLP: How Ravelin does it
• Case study presentation: How Spotify uses machine learning to prevent fraud
• Why there’s no single indicator for fighting fraud
• How to measure the success of your fraud prevention tools