Mobileum’s Jaskaran Singh, Head of Data Science, on bringing AI to real -world use cases

This article first appeared in European Comms magazine:

In a data-hungry digital world, with increasingly diverse mobile connections driven by IoT and 5G networks, there is a critical need for a more cost-effective, digital way of managing the whole telco ecosystem. Artificial Intelligence (AI) is pivotal to this transformation, in combination with other strategic data science investment.

Blending AI analytics with big-data and service delivery platforms, supported by specialist data science and domain expertise, is the key to delivering tangible business impact, says Jaskaran Singh, Senior VP for Big Data & Analytics at Mobileum.

AI has the potential to have a major impact on all aspects of the CSP business. From optimization and automation of the network, to improving customer engagement, to workforce automation there are very few areas where advanced anaytics and AI will not have a role to play. Mobileum uses AI today to analyse and predict customer behaviour on the network, and act on the insight to improve campaign performance, to identify complex network security threats, and to prevent sophisticated usage fraud. In this article Jaskaran Singh explores how AI is transforming the Fraud problem for CSPs.

European Communications: Bigdata analytics and AI is a hot topic across many markets at the moment. Where does Mobileum see this kind of advanced data science having the biggest appeal and impact for telcos at the moment?

Jaskaran Singh: AI has great potential, but we focus on the areas where we can deliver the most value. The key to delivering value is the ability to combine domain expertise, data science and the technology to take actions in a single platform. We deliver solutions based on our Active Intelligence platform which combines all three. Our heritage is in networks and roaming which is like a microcosm of the whole CSP business and we have leveraged that deep undertanding of that domain to deliver new value with AI powered solutions in customer segmentation and targeting, in signalling security and in fraud.

Given all the negative publicity around security breaches in telecoms lately, is analytics being embraced more widely now?

It ought to be, given how rampant threats are now and the potential damage CSPs are risking. In US dollars, global fraud loss is estimated to be costing the industry $29.2 billion annually now, accounting for 1.27 percent of global telecom revenues. Operators could be stemming millions in losses by developing more sophisticated fraud detection and prevention capabilities. If a customer suddenly finds a €500 charge on their bill because their phone has been hijacked, the ramifications are huge. One post to Twitter and confidence in a brand can slide sharply.

What are the costliest issues for operators currently?

Interconnect bypass (for example SIM Box) events are the most common problem, accounting for $4.27 billion, but International Revenue Share Fraud (IRSF) is the costliest issue, adding up to $6.10 billion in losses, according to the Communications Fraud Control Association (CFCA). The trouble is that fraudsters are experts at avoiding detection, quickly finding new ways round the latest security measures. Legacy systems can’t keep up with that. Being rules-based, they’re good at spotting and reacting to known issues but not at picking up emerging threats. Also, they’re limited by finite databases which at best are likely to be able to hold up to three months’ worth of data – that’s no use for tracking discrete patterns of potentially fraudulent activity over time. You may need to be able to go back several years to do this. That’s where big data analytics, using AI, neural networks and machine learning, comes in: very high capacity, intelligent systems which are quick to learn what constitutes potentially suspicious activity. By performing very complex analyses, the technology can pick up the subtlest patterns in a vast range of data – to a degree that hasn’t previously been possible.

How do traditional fraud detection methods fall short?

It’s about having to second-guess what fraud might look like: this is what rulesbased systems rely on. You might set the system to generate an alert if more than 60 minutes of usage is incurred by a subscriber after 11pm from a blacklisted cell site, for instance. Such rules invariably have a very high rate of false positives – greater than 50 percent in some cases. They might for example flag up heavy users on unlimited bundles (service abusers rather than fraudsters), business users, service numbers, telemarketers and so on. This indiscriminate event capture leads to a lot of time being spent by fraud analysts to weed out the false positives, significantly increasing the latency of fraud detection. And actually instances of fraud tend to be few, yet very high impact. There’s little value in picking them up at some point after the fact, as by then most of the loss has already occurred. So operators need to be much more targeted in their efforts, and machine learning-based fraud analytics makes that possible. Today’s detection rules tend to be based on thresholds too but, as fraud evolves, those thresholds are changing so there’s a need to monitor new usage attributes in order to detect fraud early.

Isn’t there a danger in relying on machine intelligence?

AI’s role is to detect the previously undetectable and do it quickly. Operators can still set strict controls and train systems what to look for, what level of sensitivity is appropriate, and what action to take. It’s about making fraud detection more intelligent, targeted, accurate and cost-effective. The ‘machine learning’ part means systems can be trained – and, even better, train themselves – to distinguish real threats from false positives or ‘noise’. So, over a very short period of time, they become highly accurate at telling the difference – only flagging or reacting to scenarios that are genuinely a risk. This improves the hit rate and speed of discovery, and saves fraud analyst teams from wading through reams of false alerts, driving up cost-efficiency. A carefully implemented AI solution that is continuously adapting, and aids human decision-making when required, can be very effective in reducing fraud significantly.

The benefits of this approach seem obvious. So why aren’t more operators using this kind of technology in their fraud strategies?

I think fraud managers have been reticent to ask for new budget after promising that their existing systems were the definitive solution to addressing today’s growing threats. Also, operators’ fraud analysts tend to be finance people, who don’t really understand the technology. Ultimately, fraud control is still viewed as a cost centre. It requires boldness and a bit of rebellion to break away from existing approaches, but the growing risks make this an imperative now.

How are you managing to increase the sense of urgency at the companies you talk to?

We try and remove a lot of the barriers to AI adoption by shielding operators from the complexity (the highly sophisticated mathematical algorithms being applied to complex raw data – as held in operational, billing and CRM systems, for instance). Our Active Intelligence platform automates as much as possible, so that as the software discovers events, it can test their irregularity and act without intervention (for example, by blocking calls) in a closed loop. We also provide access to data scientists who can adapt and refine the algorithms, or make sense of the patterns being detected, so fraud or IT teams stay one step removed from the technicalities. In due course, we’ll also put more of this power into operators’ hands through an ‘analytics workbench’ approach. The idea is that operator fraud teams will be able to control more of this themselves, without having to get down to the technical detail.

How soon will AI and machine learning become the de facto way to manage and guard against fraud for operators?

That’s the million-dollar question. Really there is no time to lose. For now, it’s the only way to comprehensively mine the enormous quantities of data, to find and shut down issues at speed and get ahead of the criminals. And in fact many operators are already using this kind of technology in other areas of the business – for instance as customer care interfaces (eg, chatbots), and in NFV automation. There is no question it should be applied to fraud monitoring as well – as it is in banking. There’s no question that this is the way fraud management is going. And it pays for itself very quickly, in the losses saved.

What other systems and data could AI-driven analytics be applied to for fraud control, beyond those already being scrutinised?

AI can be applied to unlimited sources of data – structured or unstructured. So additional sources could include content data, social media sentiment, deep packet inspection data – there are really no limits.

Usually we hear about big data analytics and AI being handled in the cloud, due to the superior scalability and economics when there are such vast data loads involved. Is cloud central to your solution?

Absolutely, but although our entire solution is cloud ready, most telcos are not. Telecoms operators are bound by greater regulatory control than their counterparts in other industries and so tend to be very nervous about letting sensitive data stray too far from their premises, even with the protection of robust data anonymisation. So we offer a range of options, including on-site and private cloud provision. Of course over-the-top digital service providers – the likes of Google, Amazon etc – use the cloud by default to manage sensitive data, so it’s only a matter of time before telcos adjust, but for now we can support operators using whatever data handling measures they prefer.

What sets Mobileum apart from the competition?

We offer a very comprehensive solution, with several unique strengths. First, there’s our telecoms industry expertise, and the deep knowledge and skills of our data scientists. Second, there’s our unrivalled Active Intelligence platform. This combines a big data and AI capability with a service delivery and test platform, which can take action in the network based on any discovered insight, closing the loop automatically – for superior speed and cost-efficiency. Lastly we can gain real market traction using our deep relationships with our many customers. We’re significantly ahead in applying AI and machine learning to operator problems compared to our competitors, particularly in roaming, security and fraud prevention. We’ve been working with customers on these kinds of solutions for the last two years, so we already have a strong track record.

Is AI the be-all-and-end-all, or a complementary solution to existing investments?

It depends on the application. Our use of machine learning in fraud is much more effective for detecting and protecting against the unknowns, for instance, while known frauds can be still be effectively addressed with simpler rulesbased systems. In the future, especially as more ‘things’ are connected over networks, smarter, learning-based techniques will absolutely be needed. Our solutions discover the fraud other systems can’t – and at high speed. Already, today, this technology is saving operators from millions of euros of losses they would otherwise be incurring. Without doubt, AI-based analytics is the next essential investment in fraud control: the business case is robust.