Designing flavors that many people like a lot
Industry Example: Formulated Products, Consumer Products
Who Cares: Product Designers, Project managers, Marketers, Investors
What Enabled: A framework for efficient iterative product design, new formulations, customer segmentation
Collaborators: Kalyan Veeramachaneni and Una-May O'Reilly (Evolutionary Design and Optimization Group @MIT, USA), (Givaudan Flavors Inc)
This study addresses the challenge: How to design popular and well liked flavors, the flavors that many people like a lot. Flavors are combinations of several ingredients taken in different quantities. The complication in sensory evaluation science is that the same ingredients taken in the same proportions but at a higher overall volume might have drastically different evaluation scores. The ingredients influence each other and the overall pleasantness of the flavor in a strongly non-linear fashion. This is why it is so important for sensory scientists to understand which ingredients drive liking, at which quantities, in which direction and to what extent.
This case study contained a limited set of evaluation data of only 40 flavors with seven ingredients evaluated by about 70 people. These people are not experts but are representative of the target population for the new flavor. Each evaluator assigned a liking score to each flavor on a 9-point hedonic scale (with 1 being dislike extremely, and 9 - like extremely).
To understand which ingredients drive liking we used Evolved Analytics' DataModeler to model the liking scores of individual panelists and developed individual predictive ensembles. It takes time to capture individual preferences of all people, but once the models are there - they can be used to instantaneously predict the liking scores for any unseen flavor. Developed ensembles, or 'cyber-panelists' allowed us to very efficiently explore the design space of all combinations of eight ingredients, and identify the following outcomes.
- People's propensity to like is driven by different ingredients. Even if people's liking is driven by the same ingredient - it could be driven in different directions (Vanilla is the driver for both Jack and Jill, but Jack hates it, and Jill loves it).
- Once the robust ensembles are created they can be analyzed for the ingredients that drive liking. This information can be used to cluster the target population based on ingredients that drive liking.
- Ensembles evaluated on thousands and millions of flavors ``in silico'' can be used to cluster the target audience by propensity to like flavors. (three distinct classes of ``hard-to-please'', ``easy-to-please'', and ``medium'' can be seen on a plot below.
- Ensembles can be used to iteratively optimize liking by searching for flavors that are popular (liked by many) and well liked (liked to a high degree).
These discoveries demonstrate clearly that advanced predictive analytics can be used successfully to design better products faster, segment the target population, quantify development investment based on the predicted impact on the target population, and focus product design on a particular target group.
Related Links/ Publications/ Resources:
- Kalyan Veeramachaneni, Ekaterina Vladislavleva, Una-May O’Reilly – Knowledge mining sensory evaluation data: genetic programming, statistical techniques, and swarm optimization Journal on Genetic Programming and Evolvable Machines, March 2012, Volume 13, Issue 1, p.103-133
- Kalyan K. Veeramachaneni, Ekaterina Vladislavleva and Una-May O’Reilly – Feature extraction from optimization samples via ensemble based symbolic regression. In Annals of Mathematics and Artificial Intelligence Journal, 2011, Springer Netherlands, ISSN: 1012-2443
- Katya Vladislavleva, Kalyan Veeramachaneni, Matt Burland, Jason Parcon, Una-May O'Reilly, – Knowledge mining with genetic programming methods for variable selection in flavor design. In Juergen Branke et al. (Editors),GECCO'2010: Proceedings of the 12th annual conference on Genetic and evolutionary computation, pages 941-948, Portland, Oregon, USA, 2010, ISBN13:978-1-4503-0072-8
- Kalyan Veeramachaneni, Katya Vladislavleva, Matt Burland, Jason Parcon, Una-May O'Reilly. – Evolutionary Optimization of Flavors. In Juergen Branke et al. (Editors), GECCO'2010: Proceedings of the 12th annual conference on Genetic and evolutionary computation, pages 1291-1298, Portland, Oregon, USA, 2010, ISBN13: 978-1-4503-0072-8
- Katya Vladislavleva, Kalyan Veeramachaneni, Una-May O'Reilly – Learning a Lot from Only a Little: Genetic Programming for PanelSegmentation on Sparce Sensory Evaluation Data. In A.I.Esparcia-Alcazar et al. (Editors), Proceedings of the 13th European Conference on Genetic Programming, EuroGP2010, Lecture Notes on Computer Science, Volume 6021, pages 244-255, Istanbul, 2010, Springer, ISBN13: 978-3-642-12147-0
Key Words: Formulations, Prediction, Consumer Segmentation, Marketing, Focus Groups, Customer Satisfaction, New Products, Predictive Modeling, Feature Selection, Sparse Data, Research Analytics