Artificial IntelligenceMachine Learning

Machine Learning and Fraud: Why Artificial Intelligence Isn’t Enough

Machine-learning is all the rage in fraud detection, with industry analysts, academics, businesses and technology media examining the advantages of algorithms and big data in the fight against e-commerce fraud. Especially for fraud analysts working in companies with small budgets , machine-learning tools are seen as a cost-effective way to tighten fraud controls while maintaining fast decision times, as Forrester noted in its 2015 cross-channel fraud report. There’s no question that machine-learning tools can be an effective component of fraud reduction program, but relying on them to save staffing costs may not be cost-effective in the long run.

That’s because while machine learning is an invaluable tool in the fight against fraud, it relies on human input and insight to create a comprehensive solution that yields the best results.

Overreliance on automated screening leads to more false declines

Algorithms are useful for identifying potential fraud quickly, but due to variability in consumer behavior – such as making online purchases while traveling abroad — some transactions will be falsely flagged for decline. The costs associated with false declines are too high to ignore. US merchants lose much more money on false declines than on confirmed fraud — $118 billion in false declines, compared to $9 billion in actual fraud, according to MasterCard and Javelin research.

What hasn’t been quantified is the cost of the customer relationships ended by false declines. MasterCard and Javelin found that 32% of customers who received a false transaction decline never shopped with that merchant again. Considering the cost of lost future purchases, as well as the higher relative cost of attracting new customers compared to retaining existing ones, this likely has a considerable impact on merchants.

The solution that protects merchants from fraud and lost business is to combine machine-learning algorithms with data collected by human analysts. Writing about machine-learning and card fraud for The Conversation, Penn State associate professor Jungwoo Ryoo noted that “people can still play a role – either when validating a fraud or following up with a rejected transaction.” This human intervention can reduce the number of falsely declined transactions in the short term, and when the analysts add those transaction outcomes into their data sets, it makes the automated tools smarter.

What machines need to learn varies by segment and merchant

The most effective algorithms will take into account the particular fraud patterns found within the merchant’s segments and geographic markets, as well as the changes that occur in those spaces. For example, the PYMNTS Global Fraud Attack Index found that in 2013, the digital goods segment faced high rates of suspected botnet fraud, while friendly fraud was a problem in the luxury goods segment.

More specifically, different merchants within the same segment may be subject to different mixes of fraud attempts or specific fraud patterns that algorithms must learn to detect. Experienced analysts who’ve worked extensively within a particular segment or who have long-term relationships with specific clients will have the detailed information needed to augment and improve algorithmic fraud screening at the segment and client level.
Besides historical knowledge, human analysts are the best protection against new types of fraud attempts that may launch on a small scale before ramping up to a larger and more damaging attack. These “observers on the battlefield” can raise the alert and ensure that the new data becomes part of the algorithm’s database.

What machines can’t do – yet

Algorithms are one of the technological tools that make modern e-commerce possible and relatively safe, but they can’t stand alone as a defense against fraud perpetrated by determined criminals. The advantages that human analysts bring to the process for the foreseeable future include creative problem-solving, deep knowledge of client and segment fraud landscapes, the ability to communicate directly with customers involved in flagged transactions, and the experience and intuition to pick out new fraud patterns as they develop. As long as humans are the ones perpetrating fraud against e-commerce merchants, it will ultimately be up to humans – and their smart technology – to thwart them.

Like this article? Subscribe to our weekly newsletter to never miss out!

Previous post

Three Unexpected Uses for 3D Printing in Big Data

Next post

"I expect the market to become much more fragmented as some areas of FinTech"... - Interview with LaterPay's Cosmin Ene