The aim of this study is to compare two kinds of data collection. Napping is a recent method in which allows a
direct products comparison on a tablecloth, whereas, the classical method analysis is based on the characterization of products.
We are comparing these two methods, and present the differences and similarities in the final conclusion.
This work begins with the napping method, and then followed by the classic method. Preference mapping is also used in both as complement of the analysis.
The data sets used here refers to 16 different cocktails evaluated by 12 panellists (11 for the Napping).
They can be divided into two data sets: “Napping expert” and “Classical expert”.
For the data frame “Napping expert” each row represents one of the 16 products. Each couple (Xi,Yi) represents the coordinates of the cocktail positioned on a tablecloth for each panellist. We have also panelists’ impressions for each juice. This data set named “AFM_nappes.csv” can be downloaded here.
For the data frame “Classical expert” each row represents one of the 16 product evaluated by one panellist at a given session. This data set named “Experts.csv” can be downloaded here.
First, we would want to determine the links and distances between our 16 different types of juices. Which beverages are similar? Different?
The napping data informs us on the distances between the juices and the vocabulary used by the panelist to describe each product.
The classical sensorial data, on the other hand, is constituted of the tasters’ judgments for all the descriptors such as the color intensity,
the mango flavor, the acidity, the pulpiness, etc… (See. Data set presentation).
The napping data set is as follows:
|Juices||Panelist N°1||...||Panelist N°11|
|Mb||6.3||36.7||pineapple, sugar, mango...||...||10.5||29.2||banana, sugar, mango...|
|obm*||12.4||28.2||pineapple, banana, lemon...||...||41||34.2||banana, sweet, orange...|
|BCm||23.95||21.55||acidty, banana, lemon...||...||52.6||34.2||acidty, banana, sugar...|
The non-scaled Multiple Factorial Analysis (MFA) below reveals small distances between the products MCo, OCb, Moc*, MCb*; Mb, obm* and Bmc*, Bo;
which mean these products are perceived similarly by the 11 professional panelists.
Their names give also an indication on their composition. For example, a juice with more Orange juice than mango juice or banana juice is coded as Omb*.
The first two axes have a total inertia of 59.55%, which is enough to interpret both axes because of the number of panelists.
Axis 1 contrasts juice Bmc*, which contains a lot of banana juice, to juice OCm; a juice with more citrus fruit (Orange, Lemon). On the right is the citrus taste and on the left is the banana taste, which represents the comparison between the lemons’ acidity and the bananas’ sweetness.
Axis one has an eigenvalue of 7,21 (1,43 for the second one) and it is common to all the panelists.
Axis 2 compares products with mango juice (Mb, obm*) to products without mango juice (BCo*).
To emphasize the juice groups, we chose to launch a Hierarchical Classification on Principal Components (HCPC).
There are 6 major groups emphasized.
Each group of the MFA corresponds to one taster and one X,Y coordinate.
The graphs below are obtained from the napping data MFA.
We notice that panelists 5 (Contribution to axis 2: 31,99%), 8 (Contribution to axis 2: 22,30%), and 10 (Contribution to axis 2: 20,14%) are bidimensional. They are different from the rest of the tasters, which means that they do not have the same opinions of the juices as panelists 2, 11, or 3 do.
These differences between the panelists could be validated by the representation of panelist n°5’s tablecloth (black) in comparison with the average tablecloth (green) and by the partial individual plot.
As we said, panelist n°5 is much different than the rest of the tasters.
In order to characterize all of the products, we chose to add the vocabulary information. Then we launched a Correspondence Analysis (CA)
in SPAD which allows one to obtain specific words for each juice.
The WORDS method creates a vocabulary base and then the VOSPEC method finishes the characterization.
The table below shows the descriptors which were most used for each juice (P-Value "<" 0,05).
|Juices||POSITIVE DESCRIPTORS (V-TEST ++)||NEGATIVE DESCRIPTORS (V-TEST --)|
|Mb||SUGAR, PINEAPPLE, MANGO||SOUR|
|Bmc*||BANANA, MOUTH, THICK|
|obmc*||GREEN, LIME, ORANGE|
|Om*||ORANGE, GRAPEFRUIT, YELLOW, BRIGHT|
|OCb*||BITTER, MOUTH ULCER, SOUR|
|Obc||ORANGE ACIDTY, ORANGE|
We notice again, the opposition between the words “Banana” and “Sweet” versus “Lemon”, “Sour”, and “Orange”.
You can download the R-code here.
The sensory descriptions are obtained from 11 trained panellists. Each panellist tastes 16 fruit juices. These juices are made of banana,
mango, lemon and/or orange juices in different proportions. 13 descriptors had been evaluated on a 10-point scale during 2 sessions: colour intensity,
olfactory intensity, first mouth intensity, sweetness, sourness, bitterness, orange flavor, banana flavor, mango flavor, lemon flavor, taste persistence,
|Juices||Panelists||Color intensity||Olfactory intensity||.....||Persistence||Pulpy||Thickness|
In each session, panellists have tasted all the juices. The followings analysis is based on an average data table for each taster which allows a
multi-dimensional point of view of the products' space. The panellipse function from SensoMineR is also used.
Although there are some missing values, the panelist function estimates them before performing PCA.
The scaled Principal Component Analysis (PCA) on the Products*Descriptors matrix reveals that the first two principal components account for 76.7% of the
total variance (PC1: 52.2%, PC2: 24.5%).
Axis 1 has an eigenvalue of 6.8 (3.2 for the second axis). It compares juices Bmc* and Bo (Both contain a lot of banana juice) to juices OCb* and OCm* (Both with more citrus fruits juices). This axis could be characterized with the opposition between bitterness and sourness versus sweetness and taste of banana. We remark that variable Hedonic (Illustrative variable) strongly correlates with Sweet and Banana. Therefore we can conclude that panelists prefer sweet and banana flavored juices.
Axis 2 contrasts products with mango juice (MCb*) from juices without mango juice (Obc). It is possible that citrus juice flavor is very strong and whites-out the others.
Then the HCPC function (Hierarchical Classification on Principle Components) from FactoMineR is launched in order to performe a hierarchical classification on the
principal components of the PCA above. We only used the first two axis of the PCA.
There are 3 juices clusters that are emphasized (juices are colored by cluster in the factor map below) :
The hedonic descriptor (Global appreciation of each juice) shows that black colored juices are preferred by the panelists,
especially Bmc*, Bo and BCm (Contribution to axis 1: 18.6, 16.3% and 10.4%). On the opposite side, OCm and Ocb* are disliked.
Finally, we constructed confidence ellipses for each product. We generate virtual panels using bootstrap techniques in order to display the confidence ellipses around the 16 beverages. Thus, it reveals the uncertainty of the positions of each juice. The dataset has been re-sampled 500 times.
Confidence ellipses contain 95% of the simulated data. Products are not distinguishable from one another. It is not possible to distinguish the more similar juices, nevertheless, differences are perceived by the panellists. (e.g: BCm and OCm)
Similarly, to get an impression of the variables’ variability, the same procedure is carried out. 500 variable projections are shown on the variables factor map plot.
As we can see from this figure, some attributes are very stable (Color intensity, sour, mango aroma) while others are more spread out (Odor intensity particularly)
We chose a different way to characterize our juices; indeed, we will not use the CA method, otherwise correlations between juices and variables.
The table below allows us to compare all the juices using the 13 descriptors.
|Juices\Descriptors||Sugary||Banana flavor||Appreciation||Thickness||Pulpy||Olfactiv intensity||Color intensity||Lemon flavor||Persistence||Mango flavor||Orange flavor||First intensity||Bitterness||Sourness|
In white: marks borderline; in pink: marks down on the borderline; in blue: marks across the borderline.
First, we notice that the most preferred juice is obm* (Best Appreciation mark: 6.91). This juice is more sugary (7,17) and flashier than the others. Furthermore, its bitterness, sourness, and orange flavor are less prominent. It has no lemon juice and is made of 46,7% orange juice, 26,7% mango juice, 26,7% banana juice.
Contrary to obm*, OCb* is the liked the least. (Appreciation mark: 316). It is very sour, watery and sugarless because of its amount of citrus fruit juices. Although it is not made with mango juice, panelists noticed a strong mango favor.
Then, we can characterize each products as we did with obm* and OCb*.
In general, the panelists’ performances were good. They managed to discriminate between the juices (Table below). The flavor persistence was not a discriminating descriptor.
You can download the R-code here
. Back up
"The preference mapping methods are commonly used in the fields of marketing research and research and development in order to explore and understand
the structure and tendencies of consumer preferences, to link consumer preference information to other data and to predict the behaviour
of consumers in terms of acceptance of a given product."
"This function refers to the method introduced by M. Danzart. A response surface is computed per consumer; then according to a certain threshold preference zones are delimited and finally superimposed."
In order to perform preference mapping, it is necessary using a new data set made with coordinates of the products in the first plan
(Dimensions 1 and 2) and hedonics marks attributed by each consumers as shown below:
This data frame has the dimension (16,102): each row represents one of the 16 juices.
The two first variables are the coordinates of the products in the first plan. They are obtained from the Multiple Factor Analysis (MFA) or Principal Component Analysis (PCA) obtained in the third and the fourth part with Napping or Classical profile.
The napping preference mapping data set can be downloaded here.
Preference mapping shows a kind of products' appreciation degree.
The opposite preference map reveals that 95% of the tasters liked Mb and obm*; which means these products have been marked highter than the others. At the other hand, OCb* was disliked by the tasters.
It is important to remark that the more juices have citrus juice the more they are unappreciated. Juices with banana juice are the most liked.
The preference mapping data set for the classical method can be downloaded here.
This preference mapping reveals that Mb, BCm and obm* are the most prefered.
Like the napping preference mapping, juice OCb* and OCm are the less appreciated.
You can download the R-code for napping profile prefernece mapping here.
You can download the R-code for classical profile prefernece mapping here.
In conclusion,the results are similar for both methods. Juices Mb and obm* are the most liked by consumers
and juices OCb* and OCm the least appreciated by consumers.
As we said in the first part about napping profile, Mb is sugary, orangey, pineappley and acidic like obm*. The features of these juices show the consumers’ preferences.
This similarity is interesting to notice. Napping, contrary to classical analysis, is not restricted to several descriptors and is the quickest method. Panelists class juices according to their own perceptions.
Napping, like the classic method, permits to characterize the juices with the taster’s own vocabulary which allows a preference mapping to be built.