Counting the Dots: The Underexploited Potential of “Good Enough” Location Data
August 27, 2018 | Contributed by: Jeffrey Dungen
How does one make sense of customer behavior in the real world? Simply connect the dots.
A variety of technologies can estimate the location of a customer on their journey both towards and ultimately within a brick and mortar store. Much emphasis has been placed on the location accuracy of the competing technologies, the expectation being that more precise dots will yield a more precise representation of the customer journey when connected.
But what does one do in the absence of precision location? What can one achieve with plenty of dots that do not afford the ability to map individual customer journeys in-store?
That was the problem faced by Georges Nassif, a Master’s student at Ecole Polytechnique Montreal whose thesis defense I recently attended. Georges was able to collect plenty of data points from the BLE emissions of smartphones and wearables using four sensors placed within the convenience store of a gas station. However, neither the constrained geometry nor the practical and aesthetic constraints of the store afforded optimal placement of these sensors for precision location. As is typically the case, the technology had to conform to the space, not the other way around.
Georges confirmed that mapping individual customer journeys from sub-optimally positioned dots did not yield meaningful results for analysis, as one would expect. Abandoning the prospect of analyzing collections of individual journeys, he instead tested the predictive potential of the relative occupancy of each zone. In other words, if you can’t connect the dots, count them instead.
Each of the four sensors represented a logical zone in the store (ex: snacks, alcohol, tobacco) and these produced plenty of anonymous customer ‘dots’ which could easily be grouped by zone and time. Using this matrix of data, Georges experiment showed that a decision tree algorithm could, in several product categories, be an excellent predictor of sales, when compared against actual register receipts.
Among the conclusions of his thesis, Georges indicates that “from less precise information, more useful applications came out of [the system] such as smart alerts when a certain pattern is detected.” And this is precisely (pun intended) the point: good enough location data can produce better than good enough results.
We should accept that we’re unlikely to be able collect the real-world customer behavior data we want, but rather receive that data which is readily available. For instance, the anonymous BLE data used in this research can be collected in any store as a steady stream of good enough dots. I would argue that there is a tremendous, undeserved opportunity to refine such available data into a format that lends itself well to interpretation and further analysis by the retailers themselves. What Georges’ work suggests to me is that exploiting this opportunity may be easier than one might think.