Dr. John M. Ennis is Vice President of Research Operations at The Institute for Perception. John received his PhD in mathematics from the University of California at Santa Barbara where he conducted post-doctoral studies in the UCSB Department of Psychology. An active researcher, he has published in statistics, mathematics, psychology, and sensory science. John has a strong interest in the widespread adoption of best practices throughout sensory science, serves on the editorial board of the Journal of Sensory Studies, and is chair of the ASTM subcommittee E18.04 - “Fundamentals of Sensory.”

Dr. Joey (Chih-ying) Lu is a Sensory Scientist at GlaxoSmithKline Consumer Healthcare. She serves as point contact of flavor houses, has responsibility for flavor brief, and leads sensory and consumer testing on Wellness products. Joey received her PhD in food science from Rutgers University focusing on flavor creation and flavor chemistry. She has 6 years of experience in sensory and consumer testing including time spent in the food industry and consumer healthcare industry.

Tom Carr, founder of Carr Consulting, has over 30 years experience in applying statistical techniques to research on consumer products. He is involved in continuing education in the US and internationally. Tom co-authored “Sensory Evaluation Techniques, 4rd Ed.” and “Sensory Evaluation in Quality Control,” serves on the editorial board of the Journal of Sensory Studies, and contributes as a peer reviewer. He also serves as chairman and delegate for the ASTM, and is vice-chairman of the US delegation to ISO. 7629 Hull Street Road . Richmond, VA 23235

www.IFPress.com

Optimum Product Selection for a Drivers of Liking® ProjectJohn Ennis, Daniel Ennis and Charles Fayle

Reprinted from IFPress (2010) 13(1) 2,3

either similar or not-similar. Our goal then becomes that of finding twelve products that are all not-similar to one another.

Table 1. A portion of the expert descriptive data on the eighteen products (7-point scale).

In this simplified language, we see that when we represent our eighteen products visually, with products connected by a line segment when they are considered similar, we obtain a graph as shown in Figure 1. This perspective transforms our problem to one in graph theory, allowing us to use well developed mathematical tools2,5.

Figure 1. A graph representing similarity of product pairs.

Figure 1 shows eight of the eighteen products as nodes on a graph that are connected when the corresponding products are considered similar. Our goal now is to find

Background: A project team selecting products for a category appraisal must contend with a number of competing objectives. One of these objectives is to obtain information on the performance of its own products and those of its competitors. This usually leads to extensive discussion and debate among the product development and market research members regarding the inclusion of different products in a study with a limited number of product slots. A second objective, often overlooked and of more interest to the research analysts, is the need to span the drivers of liking sensory space so that a robust account of the resulting maps is obtainable. Such an account is essential if the model that underlies the landscape will be used to predict the acceptability of products not tested in the original study1.

Products under consideration for inclusion in such an appraisal can be described on multiple attributes and there are numerous ways of considering the similarities and differences among the products. The attributes are often obtained from expert descriptive panels or from analytical data such as fat content, sucrose levels, or acidity. If there is existing knowledge of the relationship between these attributes and liking, then it is desirable to focus attention on those variables that are most likely to emerge as drivers of liking. For the purpose of this report, we assume that a set of likely drivers of liking has been chosen. Using this set we examine candidate products for consumer testing with the goal of finding a subset of products that spans the space of those variables as completely as possible. Although conventional solutions to this problem exist, including the inspection of principal component plots and cluster analysis of the products based on their attribute values, these methods are often descriptive and their interpretations are somewhat subjective. In this report we describe a novel method that provides definitive guidance by way of optimal solutions obtained using concepts from the mathematical field of graph theory.

Scenario: You are interested in developing a landscape map of carbonated orange-flavored beverages and you have eighteen candidate products under consideration for inclusion in the consumer testing component of your project. Budget and practical testing constraints require you to choose twelve products for testing and you are interested in finding the best twelve of the eighteen to choose that will likely span the resulting landscape map. For each product you have expert descriptive mean data, a portion of which is shown in Table 1. In order to have the best likelihood of capturing all relevant drivers of liking and of spanning the future landscape map, you wish to find the twelve products among the eighteen that are most different from one another.

Independent Sets and Graph Theory: To simplify the problem of finding twelve maximally different products we first suppose that we have classified all pairs of products as

Product Sweetness Bitterness Oiliness1 4.07 3.93 3.29 …2 6.78 1.50 3.61 …3 3.49 4.43 3.69 …4 5.84 2.45 3.36 …5 6.33 2.04 3.14 …6 3.08 4.59 3.33 …7 3.07 4.70 3.71 …8 3.34 4.44 3.58 …9 3.16 4.57 3.05 …… … … …

14

5

23

6

7 8

To download additional papers and technical reports, visit www.ifpress.com/publications and become a colleague.

Reprinted from IFPress (2010) 13(1) 2,3

Figure 2. An independent set corresponding to the selected products (1,2,6,8).

You conduct your study with the twelve products determined above. Using the liking data for the twelve products, you conduct a Landscape Segmentation Analysis®. The twelve products span the space and the drivers of liking explain the space well. Using the expert descriptive information you place the remaining six products on the map and obtain predictions for these additional product locations as was shown in a previous technical report1.

Conclusion: The choice of the products to be included in a product landscape study has a large effect on the future utility of the map and practical considerations limit the number of products that can reasonably be included in any such study. In many cases, techniques used to guide product selection have not consistently provided clear guidance. Using methods from the field of graph theory we now have a method that operates deterministically to provide clear guidance. This new method allows us to select a collection of products that are maximally different from each other and that will have the best chance of spanning the product landscape. Once the landscape is obtained, drivers of liking can be determined and the locations of other products not selected for original inclusion can be predicted. From this information, their acceptability to particular consumer segments can be estimated without incurring the additional cost of consumer testing.

References1. Rousseau, B., and Ennis, D.M. (2008). Improving the cost and

speed of product innovation. IFPress 11(1), 2-3.2. Valiente, G. (2002). Algorithms on Trees and Graphs. NY:

Springer.3. Lawler, E., Lenstra, J.K. and Kan, H.G.R. (1980). Generating all

maximal independent sets: NP-hardness and polynomial-time algorithms. SIAM Journal on Computing. 9, 558-565.

4. Cazals, F. and Karande, C. (2008). A note on the problem of reporting maximal cliques. Theoretical Computer Science. 407 (1-3), 564-568.

5. Ennis, J.M., Fayle, C.M. and Ennis, D.M., 2009. Reductions of letter displays. IFPress Research Papers 903, 1-44.

twelve vertices that are all not connected to each other. This problem is well known in graph theory as the problem of finding independent sets. Although this problem is extremely computationally intensive in general, state-of-the-art algorithms typically allow us to find the complete collections of maximal independent sets within reasonable time for graphs with fewer than 100 nodes3,4.

Similarity of Product Pairs: Returning to your problem, you now seek a way to classify product pairs as either similar or not-similar. For this you first compute the standardized distances between the products, obtaining a distance matrix, part of which is shown in Table 2. You then require a threshold distance to be applied to the distance matrix that yields at least one independent set of size twelve but none of size thirteen. You will use this threshold to declare pro- ducts that are within the threshold of each other to be similar and those further apart than the threshold to be not-similar. Recent research on the theory of independent sets has facilitated solutions to problems of this type and they will be used here to find the best set of twelve products to choose for the category appraisal.

Table 2. A portion of the standardized distance matrix for the eighteen products.

Determining an Optimal Collection of Products: Using this method you find a threshold for which there exists a collection of twelve products that are all not-similar to each other, but for which there is no such collection of thirteen products. Thus, you determine an optimal collection of twelve maximally different products for testing. This collection is shown in Figure 2 for four of the twelve products. The optimality of this solution avoids the arbitrariness introduced when other methods are used to address the same problem. In your case, there happened to be a unique collection, but more generally there could be multiple collections of maximally different products. In this case you could choose any of the sets of twelve, but most likely they would be chosen on the basis of other practical considerations such as availability, cost or competitiveness with the knowledge that you are selecting among equally acceptable optimal alternatives from the standpoint of the category landscape.

Product 1 2 3 4 …1 0 7.5 2.2 4.5 …2 7.5 0 8.1 5.2 …3 2.2 8.1 0 4.9 …4 4.5 5.2 4.9 0 …… … … … … …

1 4

5

23

6

7 8

Attend a short-course for more in-depth understanding! Visit www.ifpress.com/short-courses to see course details.

April 8 - 10, 2013 The Greenbrier in West Virginia DAT E LO C AT I O N

Designed for attorneys, product developers, and advertising and marketing executives, this CLE accredited course provides an overview of the basic principles of what is required to support advertising claims. Drs. Daniel Ennis, John Ennis, and Benoît Rousseau will discuss the design and interpretation of surveys and product tests designed to support product claims. The two-and-a-half-day course will cover National Advertising Division (NAD) cases and litigated cases discussed by invited speakers from diverse legal backgrounds who represent internal and external legal opinion, the corporate scientific view and the NAD perspective.

Course Calendar for Fall 2012 and Spring 2013Challenge your thinking with the latest ideas in

sensory product testing and consumer preference investigation.Go online to www.ifpress.com/short-courses for course information and registration

November 5 - 9, 2012 The Greenbrier in West Virginia DAT E LO C AT I O N

2-Day Course:

Combinatorial Tools for Product and Brand Optimization2.5 Day Course:

Drivers of Liking® and Emotion MappingIn Course 1 you will discover how to find optimal combinations of ingredients, components and products from a potentially astronomical number of possibilities by applying new graph theory techniques. You will explore in Course 2 the concepts of similarity, Drivers of Liking,® and Landscape Segmentation Analysis® (LSA) to achieve a greater understanding of how your consumers “see” the market. Invited speakers include Dr. Jean-Marc Dessirier of Unilever, Frank Rossi of Kraft Foods, and Anthony Manuele of MillerCoors who will show applications of LSA in new introductions.

Spring 2013 (date and hotel TBA) Napa Valley, CA DAT E LO C AT I O N

2 Day Course: A Powerful Framework for Improved Sensory Measurement

2.5 Day Course: Descriptive Analysis and Panel Training

In Course 1, you will achieve a deeper understanding of traditional discrimination and rating methods by learning a common framework in which to interpret results across methodologies. In Course 2, you will learn new product profiling techniques through an in-depth exploration of descriptive panel methodologies. Invited speakers include Dr. Michael O’Mahony of UC Davis, Frank Rossi of Kraft Foods, and Dr. F. Gregory Ashby of UC Santa Barbara.

June 10 - 14, 2013 Radisson BLU EU - Brussels, Belgium DAT E LO C AT I O N

2 Day Course: A Powerful Framework for Improved Sensory Measurement2.5 Day Course:Drivers of Liking® and Emotion MappingIn Course 1, you will achieve a deeper understanding of traditional discrimination and rating methods by learning a common framework in which to interpret results across methodologies. Course 2 will help you learn to “see” the market from your consumers’ perspective as you explore the concepts of similarity, Drivers of Liking,® and Land-scape Segmentation Analysis® (LSA). You will also be introduced to recently developed combinatorial tools.Speakers include Pieter Punter of OP&P Product Research, Bert Borggreve of H.J. Heinz - Europe, and Frank Rossi of Kraft Foods.

www.ifpress.com mail@ifpress.com

804-675-2980 804-675-2983

7629 Hull Street Road Richmond, VA 23235

T o C o N T a C T u s :

W H aT W e D o :

Client Services: Provide full-service product and concept testing for product development, market research and legal objectives

Short Courses: Offer internal and external courses on product testing, sensory science, and advertising claims support

IFPrograms™: License proprietary software to provide access to new modeling tools

Research: Conduct and publish basic research on human perception in the areas of methodology, measurement and modeling

re

GIs

Te

r a

s a

Co

LL

ea

Gu

e o

N o

ur

We

Bs

ITe

To

D

oW

NL

oa

D

Pa

Pe

rs

a

ND

r

eP

or

Ts

F

ro

M

IFP

re

ss

.Co

M

▲ ▲ ▲