External preference segmentation with additional information on consumers.

Transcription

1 External preference segmentation with additional information on consumers. Evelyne VIGNEAU Mathilde CHARLES Mingun CHEN

2 External preference mapping and panel segmentation Consumers products Sensory experts liing scores Internal preference segmentation Sensory profiles Regression model of the segment s mean

3 Clustering around Latent Variables (CLV) with co-variables measured on samples objective : merge together consumers who have similar drivers of preference segmentation Y Liing scores X Sensory profiles Simultaneously, define : groups of consumers and Modelisation in each group, the linear combination of the sensory attributes which explains at much as possible the liing scores. Alternative strategy : Fuzzy C-means (FCM) on the residuals of a regression on sensory or design variables (Naes,Kubberod,Sivertsen, FQP, Johansen, Herlseth, Naes, FQP, 010, Menichelli, Olsen, Meyer, Naes, FQP, 01).. see next presentation. 3

4 CLV with co-variables measured on samples maximize c 1 Y c X S K p 1 j 1 j cov( y j, t ) c 3 with t X a a ' a 1 y j : scores of liings for consumer j (j=1,,p) c : latent variable in group (=1,,K) j=1 if x j belongs to G, =0 otherwise Solution K segments in each segment, t is the first PLS regression component of y on X 4

5 Specificity of «statistical» clustering method in hedonic studies (crisp algorithm) Each consumer belongs to one, and only one, group BUT not all consumers are well represented by their group s mean («non typical» or «spurious» liings) OR some consumers are almost between two groups (degree of neighborhood between segments) possible difficulties in the interpretation of each segment specifically by means of external consumers attributes 5

6 Taing account of additional information on consumers Socio-demographic, usage and attitude attributes Woring at the individual level : and sum up the consumers in a segment. after discarding those consumers with low cluster contribution and/or high between-cluster position R with the own cluster, R with the next nearest cluster ( silhouette indices) cluster membership s values from fuzzy clustering. woring on the segment level On the basis of latent components = central tendencies in the hedonic space associated with the consumers attributes and to the products attributes. L-CLV approach 6

7 Clustering around Latent Variables (CLV) with co-variables measured on products and additional information on consumers L-shaped data Information collected by means of a questionnaire on consumers Z (M x p) Liing scores Sensory attributes Y (n x p) X (n x Q) 7

8 L-CLV maximize K ~ Z SX 1 cov( c, t ) with Z c 1 c c 3 c P u u ' P u Y 1 Z P interaction between Y and Z t 1 t X a a ' a 1 Y t X t 3 ~ S Z X 1 n K 1 u ' Z ' Y ' X a 8

9 Apple Case study (COSI-VEG ) 31 apple cultivars locally produced (Loire Valley, France) Consumer questionnaire - Frequency of Consumption, - Apple cultivars nown - Important sensory attributes, - Modalities of consumption (peeled/during meal/ ) - Purchase criteria - Supply location - 14 questions «eater style» (liert scale) - 7 questions «opinion on apple» (liert scale) - Age, gender, professional activity. Sensory descriptive analysis 15 assessors, 15 attributes Hedonic test 4 regular apple consumers -Liing score on a 9-points -5 sessions during 3 wees Crunchy Juicy Fondant Sweet Acid Odour intensity Aroma intensity A_Pineapple/Banana A_Sweet/Rose A_Woody/Earthy A_Rustic A_Lemon A_White flowers A_Ripe fruit A_Green 9

10 Apple Case study (COSI-VEG ) 31 apple cultivars locally produced (Loire Valley, France) Consumers attributes categorical : dummy variables, globally scaled numerical : centered and unit scaled 4 Z 65 Liing scores centered and unit scaled 4 Sensory attributes centered and unit scaled 30 Y X X²

11 L-CLV : choice of the number of segments Segment L3-1 8 consumers (37%) Segment L3-96 consumers (43%) Segment L consumers (0%) 11

12 L-CLV : Consumers segments represented on the internal preference mapping Fondant Sweet A_Pineapple/Banana Aroma intensity A_green Juicy Crunchy 1

13 L-CLV: Consumers segments represented on the internal preference mapping Crunchy Juicy A_Green Fondant Acid A_Lemon 13

14 L-CLV :Loadings associated to the sensory attributes (a ) Comp1 Comp Comp3 Crunchy Juicy Fondant Sweet Acid Odour Int Aroma Int A_Pineapple/Banana A_Sweet/Rose A_Woody/Earthy A_Rustic A_Lemon A_White flowers A_Ripe fruit A_Green t 1 t t 3 t 1 t t 3 Comp1 Comp Comp3 Crunchy _sq Juicy _sq Fondant _sq Sweet _sq Acid _sq Odour Int _sq Aroma Int _sq A_Pineapple/Banana _sq A_Sweet/Rose _sq A_Woody/Earthy _sq A_Rustic _sq A_Lemon _sq A_White flowers _sq A_Ripe fruit _sq A_Green _sq

15 L-CLV :Loadings associated to the sensory attributes (a ) Segments L3-1 and L3- : similar sensory eydrivers (texture crunchy and juicy, sweet flavor, «pineapple/banana» aroma) slight differences for acidity, «lemon» aroma. Segment L3-3 : do not reject fondant texture, appreciate more «rustic» and «ripe-fruit» aroma than «pineapple/banana» aroma, clearly reject acidity, «green» aroma. 15

16 L-CLV :Loadings associated to the consumers attributes (u ) the most discriminant modalities associated to the consumers attributes are highlighted. (chi-square test/v.test) 16

17 L-CLV :Loadings (u ) associated to agreement measurements (liert-type scale, centered data) on apple Clear opposition between segment and segment 1 Segment 1 : apple is not a «good fruit», not for every day, don t eat at lot of fruits and vegetables Segment 3 : don t lie acid fruits, prefer sweet fruits but don t add sugar, are very not fond of apples 17

18 Comparison with CLV on Y with external X-bloc without Z-bloc 0 18

19 Comparison regarding the sensory ey-drivers L- CLV on Y with external X- bloc and Z-bloc Partition in 3 groups n Main sensory drivers L Sweetness + Juicy + Crunchy + A_Pineapple/banana + A_Sweet/rose L A_Pineapple/banana ++ Sweetness + Aroma intensity 0 A_Lemon L A_green -- Acidity - A_lemon + Fondant 0 A_Pineapple/banana Correlation coeff. between the latent variables t CLV on Y with external X-bloc (without Z-bloc) Partition in 4 groups n Main sensory drivers YX Juicy ++ Crunchy + Sweetness + A_Pineapple/banana YX A_Pineapple/banana - A_Rustic + Aroma intensity + A_Lemon YX A_green -- Acidity - A_lemon + Fondant 0 A_Pineapple/banana YX Sweetness ++ A_Sweet/rose + A_Pineapple/banana - A_green 19

20 q-value (<-> p-value) Comparison regarding the consumers attributes partition much more For each from discriminant partition, CLV (YX) information the most : well discriminant associated for the with partition consumers the declared from important L-CLV (as attributes sensory expected). are characteristics highlighted (chi-square (internal /anova) consistency) and the eater-style segmentation «acid according fruit lover» to the age category, opinion and nowledge on apple, purchase criteria Age Nb Important varieties sensory nown charact. Purchase criteria Opinion on apples Eater-style 0

21 Conclusion In external preference mapping/segmentation, by taing into account only the external information on products, no relevant information is necessarily gained with the subsequent use of the consumers attributes. Taing into account simultaneously external information on products attributes and consumers attributes maes it possible to reveal a segmentation of consumers interpretable in terms of sociological and behavioural parameters in relation with the sensory ey-drivers. L-CLV method is suitable for this purpose (mareting research). 1

22 Than you for your attention!

23 3

24 Comment : Clustering in hedonic studies 4

25 5 Comment : Clustering in hedonic studies Each consumer belongs to one, and only one, group (crisp algorithm) BUT not all consumers are well represented by their group s mean («non typical» or «spurious» liings) low «cluster contribution» OR some consumers are almost between two groups (degree of neighborhood between segments) high «between-cluster position» R R Cluster contribution own, j max 0,r(y j G,c ) nearest, j max 0, max r(y j G,cl ) l 1 between-cluster position R ratio j 1 1 R own, j nearest, j R

26 Alternative approaches CLV with Y without external X-bloc and Z-bloc 6

27 Interpretation of the segments from CLV on Y with external X-bloc, without or with cleaning cleaning = discard «spurious» and/ or «between-clusters» consumers, select the consumers near their group s center. Here selection criterion 1-R²ratio <0.90 R own, j max 0,r(y j G,c ) 115 selected / 109 excluded R nearest, j max 0, max r(y j G,cl ) l 1 R ratio j 1 1 R own, j nearest, j R No clear improvement by the cleaning process The segments can t be better been explained by the consumers attributes, except «Age category» and «Origin» for purchase criteria which become more discriminant. 7