What is Preference Mapping?

This tutorial was inspired by [Schlich P, McEwan J.A. (1992). Cartographie des Préférences. Un outil statistique pour l'industrie agro-alimentaire. Sciences des aliments, 12, pp 339-355] and by the exchanges between the Addinsoft development team and some experts in this domain.

We distinguish two types of Preference Mapping methods:

Internal Preference Mapping

This method corresponds to MDPREF (Multidimensional Analysis of Preference Data) and is based on a PCA (Principal Component Analysis) performed on preference data with as observations the products, and as variables the consumers. The data are ratings given by the consumers for each product. The preference map is the biplot (two or three dimensional) of the observations and the variables.

As the capacity to summarize of the preference map diminishes with the number of consumers (the number of axes to interpret increases), a non-metric PCA is sometimes used to reduce the number of necessary axes. The non-metric PCA consists of a monotone transform of the data so that the variability explained by the k (k=2 or 3) first axes is maximized. This transform implies that we consider the ratings have an ordinal meaning, and that the distances or ratios between the ratings are not important. To reduce the number of axes you might also want to group the consumers and to perform the PCA with the groups as variables.

Internal preference mapping allows to generate a map on which one can identify the consumers' or groups of consumers preferences represented as vectors.

External Preference Mapping

This method allows to relate the preferences shown by the consumers to some phisico-chemical, sensory or economical characteristics of the products. This approach is essential as it gives a reliable basis to the marketing and R&D teams for adapting or creating products that will correspond to the consumers' expectations.

This method requires an additional table that describes the products with a series of criteria. Contrary to what is done with internal preference mapping, the first step consists in mapping the products on the basis of their characteristics. This can be obtained by running a Principal Component Analysis (PCA), a Correspondence Analysis (CA) or a Generalized Procrustes Analysis (GPA). This first visualization is named sensory map. By applying the PREFMAP method, we then model for each consumer (or group of consumers) the ratings that have been given to the products, using as explanatory variables the characteristics of the products, with the aim of representing the consumers on the sensory map. The full model writes:

Y = Si aiXi + Si biXi² + Sij cijXiXj

The PREFMAP method uses four sub-models:

- Vector model: bi and cij are null. The model is a hyperplan. This model allows to display the observations on a sensory map as vectors. The size of the vectors can be related to the R² of the model ; in that case the longer the vector, the better the underlying model. The consumer's preference increases the further you go in the direction of the vector. Interpreting the preference can be done by projecting the products on the various vectors (product preference). The drawback of that model is that it neglects that for some criteria such as saltiness or temperature, the preference can increase until an optimal value, and then decrease;
- Circular ideal point model: the bi are equal and the cij are null. The model corresponds to a hyperquadric hypersurface. If the surface has a maximum in terms of preference we will talk of an ideal point. If this surface has a minimum we will talk of anti-ideal point. With the circular model it is possible to trace isopreference circular lines around the ideal or anti-ideal point.
- Elliptical ideal point model: les cij sont nuls. The model corresponds to a hyperquadric hypersurface. In this case the isopreference lines are ellipses, which makes the interpretation of the distances of the products to the ideal or anti-ideal points more complex. If the bi have opposite signs, there is no ideal or anti-ideal point but only a saddle point which interpretation is tricky.
- Quadratic surface model: this model corresponds to the complete model which shape is a hypersurface. This model allows to take into account interactions between the characteristics (the cijXiXj.

How do I run Preference Mapping with XLSTAT-MX?

The following example shows how one can create a preference map using the PREFMAP method. An Excel sheet with both the data and the results can be downloaded by clicking here. The data used here are composed by:
The consumer acceptability data: 99 consumers rated 10 different commercial samples of potato crisps. These data have been obtained from the article by Schlich and McEwan (1992). The ratings have been discritized from 1 to 30 (30 corresponds to the highest acceptability). These data are stored in a 99 x 10 table.
The average ratings computed from the ratings given by 8 experts to the 10 crisps samples for 4 texture attributes and 7 flavor attributes. These data, simulated for the purpose of teaching by the author of this tutorial on the basis of the article by Schlich and McEwan (1992), make up a 10 x 11 table.

Step 1: Creating the sensory map

We first create the sensory map by applying a PCA on the 10 x 11 table. This gives us a two-dimensional visualization of the crisps. As a tutorial is dedicated to PCA, we do not elaborate on that subject here. The dialog box of the PCA has been filled in as shown below.

The display options have been set as follows:

The map obtained, which quality is fine (69.3% of the variability is displayed), allows us to notice that the products are well differentiated by the experts.

The criteria seem to be little redundant given their dispersion on the correlations circle.

Step 2. Grouping the consumers

We now focus on the ratings given by the 99 consumers. As the number of consumers is significant, we decided to group them into homogeneous groups in order to make the PREFMAP results easier to interpret. We chose the Agglomerative Hierarchical Clustering (AHC). As a tutorial is dedicated to AHC, we do not elaborate on that subject here. The dialog box of the AHC has been filled in as shown below.

The "Center / Reduce" option has been left activated in order to diminish the differences between the judgment scales of the consumers. By looking at the dendrogram, it makes sense to decide to work with 8 groups (we truncate the dendogram at the level 32).

We then re-run the AHC while specifying that we request 8 clusters. The dialog box has been filled in as displayed below.

We then store the centroids of the clusters for the last step of the analysis. The table is copied and paste (Edit / Paste special with Transposed option) in a new sheet named "Clusters' pref.".

Step 3: Creating the preference map using the PREFMAP method

In this section we apply the PREFMAP method, using the coordinates of the crisps in the two-dimensional factor space and the ratings given by the consumers summarized by the standardized ratings for the 8 clusters.

To activate the Preference Mapping dialog box, start XLSTAT, then select the XLSTAT/XLSTAT-MX/Preference Mapping command, or click on the corresponding button of the XLSTAT-MX toolbar (see below).

When you click on the button, the Preference Mapping dialog box will appear. Select the data on the Excel sheet. To the "Attributes (Y)" correspond the ratings of the 8 clusters. The "Configuration (X)" corresponds to the factor scores of the crisps obtained from the PCA. We decided to use the vector model.

At the end of the treatments, a second

The computations begin once you have clicked on the "OK" button. They stop to let you choose the options for the Preference map. We selected the R² (coefficient of determination) to determine the length of the vectors on the preference map.

The results we obtain (see below) show that the vector model is well fitted for the clusters 1, 3 and 6. For the other clusters it will be more risky to interpret the results.

The preference map allows to quickly interpret the results. When looking at both the map and the correlations circle, we see that the consumers of the cluster 1prefer crisps that are greasy and melting, not crispy to the contrary of the Cluster6. Consumers of the cluster 3 prefer crisps that are salty and sweet, not sticky and melting.

The preference orders for the various groups of consumers are displayed.

We notice that the crisp 4, characterized by an earthy only little sweet and salty taste, is preferred by the clusters 7 and 8, while it is the most unaccepted by the clusters 3, 4 and 5. The marketing and R&D teams will be able to take this information into account to direct their creation of new crisps towards the right directions.

Click here for other tutorials.