To operate in a customer-centric manner, businesses must be able to respond to customer actions as they take place. In many situations, data analysis and data-driven decisioning must occur in real time or near real time, based at least partly on the stream of data coming in from mobile phones, ATMs, online activity, point-of-sale (POS) devices, sensors, etc.
The value of streaming analytics has been well-proven in credit card fraud management. Here, models detect unusual patterns of cardholder behavior and instantly generate a ranked score indicating how suspicious the transaction it is. These analytics not only drive real-time customer decisions, but can access and dynamically update profiles for individual cardholders, ATMs, POS devices and other entities involved in transactions.
The requirement to build models that don’t require supervised training with historical data is likely to become more common in the era of Big Data. Although vast amounts of data are being amassed, the dynamic nature of business today means it will not always fit modeling objectives and timeframe. As new ways of interacting with customers emerge and companies want to analyze data in innovative ways that produce a competitive advantage, relevant historical data may be limited, problematic or non-existent. Where the data does exist, customer behavior may be changing so rapidly that traditionally trained models still need the ability to learn on the fly.
We see the need emerging in banking, an increasingly dynamic industry. A South African bank, for instance, has seen a sudden rise in fraud on account-to-account transactions. With limited data at hand, the company turned to FICO for streaming self-learning analytics to solve the problem.
These streaming analytics are based on a technology we developed called self-calibrating outlier models. They compare the behavior of peers (e.g., similar customers, similar accounts, similar mobile connections) in order to detect outliers and score them for degree of deviance from the norm. The patented technology is critical to streaming analytic problems where the algorithms must continuously update, in real time, estimates of feature distributions so that detection of outliers is always based on current distributions.
This kind of model is an example of a good marriage between human expertise and machine learning—another analytic imperative I covered last week on the FICO Labs Blog. As with supervised models, the first step in building an unsupervised self-calibrating outlier model is decidedly human. Determining which customer characteristics are highly predictive and how to incorporate them in the model requires deep human understanding of the domain and problem to be solved. But while supervised models are trained with months of historical data to recognize normal and abnormal values for those characteristics, in self-calibrating models, machine learning takes over, inferring the values from the stream of transactions.
I’ll be continuing my discussion of analytic imperatives next week on the FICO Labs Blog. For more information on this topic, I invite you to check out my recent Insights white paper: "When Is Big Data the Way to Customer Centricity?" (registration required).