Characterizing Sensory Segmentation using Machine Learning
Russ, W. J. and Ennis, J. M. (2018). IFPress, 21(4) 3-4.
A revolution is happening in business worldwide as companies increasingly leverage computational and algorithmic power to solve problems that previously might never have even been considered. For example, who would have expected ten years ago that large numbers of people would routinely awaken to countertop virtual assistants turning on their lights and greeting them with music, telling them the news and weather for the day, checking their calendars and messages, and eventually summoning people they’ve never met to ferry them to their destinations? Yet such actions are now commonplace. Similarly, in sensory and consumer science, advances in computation are reaching an inflection point of supporting behaviors that were previously inconceivable. For example, just a few years ago it would have required a supercomputer to run a full TURF analysis on a dataset of even moderate size. But now such analyses can be run instantly - it is now possible to evaluate the entire space of possible TURF solutions on a standard desktop computer and to find the “best” solution according to a variety of criteria. From a computational perspective, several of our previous technical reports have discussed the ubiquitous sensory and consumer science problem of finding best combinations - this problem may appear in the guise of finding the best combinations of features, benefits, ingredient, images, and/or products in a portfolio.
In this report we turn our attention to a different combina-torial problem, the problem of characterizing consumer segments using the best combination of descriptors for the segments. These descriptors may be demographic, behavioral, psychographic, or could encapsulate any other information known about the respondents. Just as with other combinatorial problems, an explicit search for solutions is intractable. But, just as with other combinatorial problems, recent advances in computing and algorithmic power now allow us to find solutions, this time using a method known as machine learning.
Figure 1. 2D projection of LSA map with subject ideals color coded according to segments from Step 2.