XLSTAT: Clustering tables by means of CLUSTATIS in Excel

 

Dataset for Example

The data come from a projective mapping/Napping study carried out in Rennes by AGROCAMPUS OUEST. Eight smoothies were tasted by 24 subjects (panelists) and then placed on a tablecloth. The coordinates were collected for the STATIS analysis. If a panelist considers two products to be similar, the latter ones are placed closed by on the tablecloth, so they have similar coordinates. The original file can be obtained with the R SensoMineR package.

This dataset is also used in our tutorial on STATIS, in order to analyze smoothies directly without subject clustering.

The aim here is to perform a three-step data analysis:

1. Segment the tables (named here configurations) of data, i.e. the subjects herein, according to their perceptions of the products. The CLUSTATIS method will be used to create the most homogeneous classes possible.

2. Analyse each class of subjects using the STATIS method in order to determine the differences in perceptions of smoothies between classes.

3. Determine clustering quality indices.

 

Setting Up

• Open XLSTAT.

• Click on the Sensory profiling menu and select the CLUSTATIS feature.

 

• In the CLUSTATIS dialog box, under the General tab:

  • Select the configuration data (each configuration corresponds to one subject’s coordinates).

  • Enter the number of configurations (24 in this case).

  • Since each configuration has 2 variables, select the Equal option.

  • If variable counts vary, select a column specifying the number for each.

  • Activate Variable labels and Object labels (here, the smoothies).

 

In the Options tab:

• Do not scale (reduce) the variables, as all variables within each configuration are on the same scale.

• Leave the number of classes selection to automatic.

• Enable consolidation of the classes from the hierarchical algorithm to refine the cluster analysis.

 

In the Charts tab:
Check “Display charts on the first two axes” for automatic plots on the first two factorial axes.
If unchecked, a window will prompt you to select the axes manually.

• Click on the OK button to launch the computations.

• The results are displayed on a new Excel sheet.

 

Interpreting the Results

The dendrogram below shows how the algorithm groups configurations and their clusters. The dotted line marks the automatic cut, resulting in three groups.

The following table shows the classes of the different configurations before and after consolidation. It appears that only subject 10 changed class with the consolidation.

 

Next, each class is analyzed by representing the objects (smoothies). Differences in perception appear, especially for Carrefour_SB, which is grouped with Innocent_SB by class 1, with Innocent_PBC and Casino_PBC by class 2, and with Carrefour_MP by class 3.

Analysis of cluster 1:

 

Analysis of cluster 2:

 

Analysis of cluster 3:

 

The following graph shows the proximity of configurations to their class consensus using the RV coefficient (0 to 1). In class 2, subject 2 is farthest from the consensus, indicating a differing perception. This kind of problem can be addressed by adding a "K+1" class in the Options tab, which is an additional class designed to set aside atypical configurations.

 

Finally, the homogeneity indexes assess the quality of the cluster analysis; values closer to 1 indicate more homogeneous classes. Here, class 1 is slightly less homogeneous, while the overall homogeneity is acceptable but could improve by adding a "K+1" class.