1. Bayesian Networks
A complete framework to
understand consumer perceptions
and their links with external data
Fabien CRAIGNOU, Repères
Carole JEGOU, Kraft R&D

2. Context & Objectives 2
 PRODUCT TESTING in France.
 20 market products tested in the confectionary sector
 Sequential monadic procedure
 N= 200 respondents
 Overall liking, JAR questions
 SENSORY PANEL
 40 significant sensory attributes
 Covering aroma, texture, flavour, aftertaste and after sensation.
 ANALYTICAL MEASURES
 20 key variables
 GAIN UNDERSTANDING OF CONSUMER PERCEPTIONS AND DRIVERS OF LIKING,
AND PROVIDE THE R&D WITH GUIDELINES FOR PRODUCT DEVELOPMENT.
Sensometrics meeting – Rotterdam – July 2010

3. Data processing workflow: 3-step analysis 3
STEP 1 : UNDERSTANDING STEP 2 : SIMPLIFYING
CONSUMER PERCEPTIONS SENSORY & TECHNICAL
INFORMATION
« internal » model
consumer attributes Identification of main
(JAR scales) dimensions
STEP 3 : INFLUENCE OF
TECHNICAL DIMENSIONS ON
CONSUMER PERCEPTIONS
« external » model
sensory attributes,
analytical data
Sensometrics meeting – Rotterdam – July 2010

4. Data processing workflow: 3-step analysis 4
STEP 1 : UNDERSTANDING
CONSUMER PERCEPTIONS
« internal » model
consumer attributes
(JAR scales)
Sensometrics meeting – Rotterdam – July 2010

5. Understanding consumer perceptions STEP 1 5
AUTOMATIC LEARNING - discovering STRUCTURE and PARAMETERS
Colour JAR
41% Mutual information
Taste of X JAR 36% OVERALL
Texture 1 JAR
LIKING
46% 27% 25%
42% 49%
54%
Texture 2 JAR 32% Consistency JAR
14% Sweetness JAR
Aftertaste JAR
HEURISTIC SEARCH ALGORITHM TO TEST DIFFERENT STRUCTURES
QUALITY OF THE POSSIBLE NETWORKS IS ASSESSED BY A SCORE TAKING INTO ACCOUNT
The fit of the model to the data
The complexity of the structure
CROSS VALIDATION IS USED TO ENSURE ROBUSTNESS
Sensometrics meeting – Rotterdam – July 2010

6. Understanding consumer perceptions STEP 1 6
OUTPUT – relative weights in overall liking
Sweetness 24%
Relative Weights in
overall Liking
(Mutual information
Aftertaste 22% normalized to sum 100%)
Texture 20%
Colour 18%
Taste of X 16%
ALL DRIVERS ARE VERY CLOSE IN TERMS OF IMPACT ON TASTE LIKING
(contrary to other markets where 1 or 2 drivers prevail).
OVERALL LIKING requires good performances on…
 … TASTE dimensions (aftertaste, sweetness, taste of…)
 … TEXTURE perception
 …COLOUR perception
Sensometrics meeting – Rotterdam – July 2010

7. Understanding consumer perceptions STEP 1 7
OUTPUT – detailed impact of intensity balance
Impact of sweetness balance Impact texture balance
Probability that Liking >= 8 Probability that liking >= 8
80% 80%
60% 57% 60%
40%
40% 40%
too light too light
20% sweetness 20% meltiness
11% JAR 5% JAR
5% Intensity 2% Intensity
0% too strong 0% too strong
Impact of Aftertaste balance Impact of colour intensity balance
Probability that Liking >= 8 Probability that Liking >= 8
80% 80%
60% 60%
43% 43%
40% 40%
too light too light
20% aftertaste 20% 13% colour
4% JAR JAR
1% Intensity 3% balance
0% too strong 0% too dark
Sensometrics meeting – Rotterdam – July 2010

8. Data processing workflow: 3-step analysis 8
STEP 1 : UNDERSTANDING STEP 2 : SIMPLIFYING
CONSUMER PERCEPTIONS SENSORY & TECHNICAL
INFORMATION
« internal » model
consumer attributes Identification of main
(JAR scales) dimensions
Sensometrics meeting – Rotterdam – July 2010

9. Simplifying sensory/analytical information STEP 2 9
Handling sensory and analytical variables
Sensory and Analytical variables have been discretized into 3 levels using a K-Means
procedure, in order to adapt each discretization to the distribution of the variable.
CHECKING CORRESPONDENCE WITH
GROUPINGS BASED ON THE SENSORY PANEL
Falvour Attribute 1
Sensory Score Flavour Attribute 1
Product 1 29.55 a
Product 2 26.75 a
Product 3 25.94 ab
Product 4 25.75 ab
Product 5 22.38 b c
Product 6 21.98 b c d
Product 7 21.73 b c d
Product 8 21.36 c d
Product 9 21.33 c d
Product 10 21.11 c d
Product 11 21.02 c d
Product 12 20.31 c d
Product 13 19.91 c d
Product 14 19.20 c d e
Product 15 19.15 c d e
Product 16 18.58 c d e
Product 17 18.57 c d e
Product 18 17.95 c d e
Product 19 17.25 d e
Product 20 14.94 e f
Number of
observations
Sensometrics meeting – Rotterdam – July 2010

10. Simplifying sensory/analytical information STEP 2 10
Identifying main dimensions Unsupervised learning
Hierarchical Clustering based on KL divergence
Aroma 1
Aroma 1 IMPORTANCE OF CROSS-VALIDATION
Mouthfeel 3
Mouthfeel 3
Descriptor 8
Descriptor 8 Descriptor 11
Descriptor 11
Descriptor 9 Descriptor 12
Descriptor 9 Descriptor 12
Descriptor 13
Descriptor 13
Mouthfeel 1
Mouthfeel 1 Descriptor 10
Descriptor 15
Descriptor 15
Descriptor 10 Descriptor 14
Descriptor 14
Descriptor 6 Descriptor 7
Descriptor 6 Descriptor 7 Flavour 4
Flavour 4 Descriptor 16
Descriptor 16
Descriptor 5 Flavour 3
Flavour 3
Descriptor 5 Descriptor 18
Descriptor 18
Descriptor 33 Descriptor 34
Descriptor 33 Descriptor 34
Descriptor 17
Descriptor 19 Descriptor 17
Descriptor 19
Descriptor 37 Descriptor 35
Descriptor 4 Descriptor 37 Descriptor 35
Descriptor 4
Descriptor 36
Descriptor 36
Descriptor 3
Descriptor 3
Descriptor 2 Descriptor 29
Descriptor 1 Descriptor 2 Descriptor 29 Descriptor 36
Descriptor 1 Descriptor 20 Descriptor 36
Descriptor 30 Descriptor 20
Descriptor 30
Mouthfeel 2
Mouthfeel 2 Descriptor 31 Descriptor 21
Descriptor 26 Descriptor 31 Descriptor 21
Descriptor 26
Descriptor 32 Descriptor 23 Descriptor 22
Descriptor 27 Descriptor 32 Descriptor 23 Descriptor 22
Descriptor 27
Descriptor 28 Flavour 2
Flavour 2
Descriptor 28 After sensation 1 Descriptor 24
After sensation 1 Descriptor 24
Aroma 2
Aroma 2 Descriptor 25
Descriptor 25 Flavour 1
Flavour 1
Sensometrics meeting – Rotterdam – July 2010

11. Data processing workflow: 3-step analysis 11
STEP 1 : UNDERSTANDING STEP 2 : SIMPLIFYING
CONSUMER PERCEPTIONS SENSORY & TECHNICAL
INFORMATION
« internal » model
consumer attributes Identification of main
(JAR scales) dimensions
STEP 3 : INFLUENCE OF
TECHNICAL DIMENSIONS ON
CONSUMER PERCEPTIONS
« external » model
sensory attributes,
analytical data
Sensometrics meeting – Rotterdam – July 2010

12. Sensory expectations of consumers STEP 3 12
Key issues of the modeling workflow
Consumer attributes Sensory
Dimensions
In order to let the search algorithm
c1 s1
Consumer 1
cn sk focus on the links between Consumer
Data and Sensory Data…
Product 1
Constant for 1 product
Consumer 200
FIXING the arcs between consumer
dimensions (already discovered)
FORBID the arcs between sensory
dimensions (links are too obvious: for
each product => 200 times the same
sensory variables)
Consumer 1
Product 20
Consumer 200
Sensometrics meeting – Rotterdam – July 2010

13. Sensory expectations of consumers STEP 3 13
Structural model
Colour JAR
Taste of X JAR OVERALL LIKING
Texture 1 JAR
Flavour 4
Flavour 4
Texture 2 JAR Consistency JAR
Aftertaste JAR Sweetness JAR
Flavour 2 Flavour 3
Flavour 3
Flavour 2 Mouthfeel 3
Mouthfeel 3 Mouthfeel 1
Mouthfeel 1
After
After
sensation 1
sensation 1
Aroma 2 Mouthfeel 2
Mouthfeel 2
Aroma 2
Flavour 1
Flavour 1
Aroma 1
Aroma 1
Sensometrics meeting – Rotterdam – July 2010

14. Sensory expectations of consumers STEP 3 14
OUTPUT: importance of sensory dimensions on SWEETNESS perception
* Flavour 3
Bitterness
Mutual Information :
non-linear measure of impact.
* After sensation 1
Astringent aftersensation Total sample
Cluster 1 Cluster 2
* Flavour 2
Sourness
Aroma 2
Cocao aroma
Flavour 1
Fruity
Flavour 4
Chocolate flavour
Aroma 1
Sweet aroma
% Mutual
0.0 1.0 2.0 3.0 4.0 5.0 6.0 7.0
information
Sensometrics meeting – Rotterdam – July 2010

15. Sensory expectations of consumers STEP 3 15
OUTPUT: expected sensory levels (top-6 descriptors)
Probability that sweetness is perceived as JAR by consumers
FLAVOUR 3 - descriptor 1 FLAVOUR 3 - descriptor 5 FLAVOUR 3 - descriptor 7
70% 70% 70%
60% 60% 60%
62% 58% 62% 60%
56%
56%
50% 48%
50% 48%
50%
48%
40% 40% 40%
30% 30% 30%
20% 20% 20%
11 - 18 18 - 23 23 - 28 12 - 19 19 - 22 22 - 29 8 - 10 10 - 14 > 20
AFTERSENSATION 1 FLAVOUR 2 -descriptor 2 AFTERSENSATION 1
descriptor 8 descriptor 12
70% 70% 70%
60% 63%
60% 60%
59% 60% 59%
60% 59%
50% 50% 50%
47% 46% 47%
40% 40% 40%
30% 30% 30%
20% 20% 20%
17 - 22 22 - 24 24 - 26 6 - 10 10 - 12 12 - 18 7 - 13 13 - 17 17 - 23
Significant positive impact
Significant negative impact
Sensometrics meeting – Rotterdam – July 2010

16. Sensory expectations of consumers STEP 3 16
Preference Clusters
OUTPUT: expected sensory levels (top-6 descriptors) Cluster 1 Cluster 2
Following 2 of the preference segments identified by KRAFT
FLAVOUR 3 - descriptor 1 FLAVOUR 3 - descriptor 5 FLAVOUR 3 - descriptor 7
70% 70% 70%
63% 63% 63%
65% 64% 65%
60% 60% 60%
58% 58%
61% 61% 58%
53% 57%
50% 50% 50%
48%
48%
40% 40% 40%
38% 38% 38%
30% 30% 30%
20% 20% 20%
11 - 18 18 - 23 23 - 28 12 - 19 19 - 22 22 - 29 8 - 10 10 - 14 > 20
AFTERSENSATION 1 FLAVOUR 2 -descriptor 2 AFTERSENSATION 1
descriptor 8 descriptor 12
70% 70% 70%
64% 63% 63%
65% 65% 60%
60% 60% 60%
62%
58% 59%
53% 58% 53%
50% 54% 50% 51% 50% 54%
40% 40% 40%
41% 41% 41%
30% 30% 30%
20% 20% 20%
17 - 22 22 - 24 24 - 26 6 - 10 10 - 12 12 - 18 7 - 13 13 - 17 17 - 23
Significant positive impact
Significant negative impact
Sensometrics meeting – Rotterdam – July 2010

17. Sensory expectations of consumers STEP 3 17
OUTPUT: Summary of expected sensory levels Main sample
Cluster 1
KEY DESCRIPTORS Cluster 2
Descriptor 1
Descriptor 1
Descriptor 2
Descriptor 2
Monitoring…
Descriptor 3
Descriptor 3
SWEETNESS
AFTERTASTE
Descriptor 4
Descriptor 4
TASTE OF X
Descriptor 5
Descriptor 5
Descriptor 6
Descriptor 6
Descriptor 7
Descriptor 7
MOUTHFEEL
Descriptor 8
Descriptor 8
Sensometrics meeting – Rotterdam – July 2010

18. SUMMARY – ADDED VALUE / LIMITS on the methodology 18
 STRUCURAL LEARNING
possible to discover the links between variables
enables to have a good overview of all consumer perceptions simultaneously
 NON-LINEAR RELATIONS
very important when dealing with links between consumer & sensory
 USE OF CROSS-VALIDATION
to enhance confidence into the models
 Discretizing sensory/analytical variables: tough (not natural?) job,
need to check correspondence with sensory panel significant
differences.
 Non-linear relations: exercise caution, as sometimes weird relations
can be discovered (U-shape like relation for example) => needs to be
cleaned.
 Causality ?
Sensometrics meeting – Rotterdam – July 2010

19. PLS PATH MODELLING – BAYESIAN NETWORKS 19
SOME IMPORTANT DIFFERENCES TO REMEMBER
BAYESIAN PLS PATH MODELLING
Discover the structure Impose the structure
Latent variables: entirely Latent variables
dependent on the constructed to explain
explanatory variables TARGET
Many observations Few observations
Sensometrics meeting – Rotterdam – July 2010