Recently, Michael Schmidt blogged on the topic of Eureqa and finding causation in data, which can be found here. I found myself remembering my first exposure to insightful thinking on this subject.
Many years ago, I worked at an MPP database. At the time, the "Big Data" revolution was just beginning to take shape. One of the customers I was privileged to work with was an online retailer. The journey they began—more than any other influence—opened my eyes to the future direction of marketing. They initiated the use of behavioral analysis in retail marketing while many other firms were still relying upon mere demographics. In retrospect, demographic analysis now seems like a comparatively crude tool, a little like looking at an old computer with vacuum tubes.
I met a gentleman we will call Kyle on the fourth floor of a converted warehouse space in NYC. In line with the modern fashion sense of New York, the space had been turned into a hip, tech savvy retailer location. Open Air and collaborative teams met to debate and implement best practices for reaching their target audiences. Kyle led me to a makeshift conference room where he apologized for the noise. They were growing so quickly that they had no conference rooms left.
Yeah, tough problem to have.
Over the next couple of hours, I learned that the analysis of demographics is a completely outdated concept. Kyle and his team explained that creating a hidden variable, which is typically thought of as demographics, allows the arbitrarily assignment of users and behaviors to that variable. After fitting their behavioral patterns using a number of techniques such as clustering, they "fill in the gaps" of items that people haven't purchased. Now, for those of you familiar with collaborative preferences, this is only part of the problem. They were segmenting their user base to thousands of clusters before applying a collaborative preference technique. Looking from a lay perspective, this simply means that the groups of individuals were arbitrary, based on similarities rather than demographics. What they found was that typical demographic characteristics have little or no effect on buying patterns.
Kyle and I concluded the meeting but eventually went on to do business together and we have stayed in contact since then. I carried with me the feeling that they were lightyears ahead of their time. More than that, I realized they understood business is constantly changing and their own value would rapidly diminish if they were not always adapting to stay ahead. Over the years, nearly every marketer has begun using the technique they showed me.
Recently RoundTower has been working with Eureqa to define and analyze the problem even further. Our work is proving that retail sales not only require clustering but hundreds, if not thousands of models for effective results. Eureqa can create these varied models and it has the ability to avoid bias while uncovering real and meaningful signals.
I was intrigued by Michael Schmidt's latest blog and it left me thinking. Kyle's phone is going to ring tonight.