Sramana Mitra: And what does it take on your end – on the technology or services end – to deliver this?
Oliver Downs: We take dynamic feeds of customer transaction information. This includes individual calls, SMS and data session records, balance recharges, account balance data points, and other types of customer status information. We maintain an ongoing history of that information. We are represented as sequences and patterns at the individual customer level. This is where our big data technology comes in.
SM: How do you frame it? What is the data structure and the algorithm to make something like this happen?
OD: The data structure is on a para-individual basis. A customer representative represents a series of events that has happened to a customer through the course of their relationship with him or her. The thing that is very different is the state of the art in play. It is much more common to represent the customer as a row of data instead of as a snapshot of current attributes and all these cumulative sequences of events that have happened to him or her.
Such systems come to have this framing of each customer as a stream of events. This allows us to look for temporal and sequenced patterns in behavior of each individual’s data and to look for commonalities among groups of individuals. Given the commonalities among groups that are sufficiently large – say five to fifteen percent of a carrier’s customer base – we can identify, based on those commonalities, opportunities among that group of customers to interact with them. We formulate a set of triggers that allow us to center on the specific behavior we see in that group of individuals. Then we submit back to the carrier’s messaging and offer the customers promotions – or, rather, the rules and the flagging of customers who should get particular marketing interactions and promotions. Those are then dynamically filled in real-time for each customer.
SM: Is there any learning technology involved in how you do this?
OD: Yes, there is. The nice thing is we have a very rich framework for doing complex experimental design. That framework allows us to measure individuals in the same specific context, such as who have been treated in the same marketing context, and it gives us very strong feedback signals. What we are able to do is learn and adapt how we [handle] individual messaging parameters, to refine the specific combinations of behavioral sequences that perform best in terms of garnering responses to those campaigns. We can actually learn those dynamically based on the observations we make over a sequence of events the customer conducts, which is followed by us messaging them or followed by us seeing a context occur for them and then intentionally not messaging them.
SM: Would it be fair to say that you are both dynamically clustering and dynamically messaging?
OD: That is correct.
SM: And you are learning on both sides?