README.md

In adrienPAVOINE/ClustCheck: Clustering check and interpretation

ClustCheck

The ClustCheck package is intended to provide tools to check and analyse the results of a clustering from a machine learning algorithm. It offers a combination of evaluation metrics and graphical visualisations to understand what variables drive the structure of the clusters and to test the quality of the clustering.

Installing the package

devtools::install_github("adrienPAVOINE/ClustCheck")

Tutorial for package usage

Loading the library

Once the package is installed, the library can be load using the standard commands from R.

library(ClustCheck)

Dataset Import

First of all, you need to import a dataset (with numerical, categorical variables or both). In this example, we’ll be using the BankCustomer dataset. This dataset is included in the ClustCheck package.

BankCustomer <- ClustCheck::BankCustomer

The BankCustomer dataset includes information about bank customers and contains 4 categorical variables and 4 numerical variables. To test the package with this dataset, you have to call the Dataset() function to create your class object. The function takes a dataframe and a vector of predicted cluster groups as input. A vector of real cluster groups can be used as optional input (to be used within the EvaluateC() function).

cbank <- ClustCheck::Dataset(BankCustomer, BankCustomer$Cluster)

## Class correctly instanciated. 
## The dataset contains  4  categorical variables and  4  numerical variables.

Now that our dataset has been loaded to the object cbank, we can proceed with the cluster analysis. If you have any doubt about a function, use help(name_function), exemple : help(Dataset).

Note: In the following code, cbank is the object instanciated by the Dataset() function previously shown.

Analysing the clustering with single variables

Let’s start our analysis by looking at how single variables influence the clustering. We can deal with two possible variable types ; categorical variables and numerical ones. We’ll be starting with the former.

Categorical variables

• Cramer’s V

Here, the Cramer’s V values will be computed for the 4 categorical variables of our dataset.

ClustCheck::vcramer(cbank)

##                 Cramer
## profession  0.78276275
## epargne     0.21795453
## carte_bleue 0.06733646
## pea         0.14985464

To return the Cramer’s V values only between the cluster groups and the variable profession for example, the variable needs to be entered as a function input.

ClustCheck::vcramer(cbank, var = BankCustomer$profession)

## Cramer value between the cluster vector and   BankCustomer$profession  =  0.7827627

A bar graph of the values can be plotted with the function plotVCramer().

ClustCheck::plotVCramer(cbank)

• V-Test

The Test values can be computed either against the modes of a categorical variable (i.e. profession) with the function tvalue_cat or against the numerical variables with the function tvalue_num(). We use tvalue_cat() in our example below.

ClustCheck::tvalue_cat(cbank, var = BankCustomer$profession)

##    var
##           AGR        ART        CAD        EMP        INA        INT 
##   1 -0.5975237 -1.4574992 12.2065556 -2.6794878 -1.1582863 -2.9255381
##   2 -0.5975237  1.7145135 -4.1415099 -1.0845546  1.4669505  0.8574853
##   3  1.0320864 -0.7121871 -4.4387475  3.0042742 -0.6722920  1.2013965
##    var
##           OUV 
##   1 -1.6645303
##   2  1.0601473
##   3  0.1008806

• Phi-value

The Phi values can be computed against the modes of a categorical variable (i.e. profession) with the function phivalue().

ClustCheck::phivalue(cbank, var = BankCustomer$profession)

##    var
##        AGR     ART     CAD     EMP     INA     INT     OUV 
##   1 -0.0951 -0.1713  1.0690 -0.2320 -0.1502 -0.2413 -0.1840
##   2 -0.1381  0.3093 -0.3404 -0.0873  0.2876  0.1440  0.1818
##   3  0.2580 -0.0921 -0.3670  0.4309 -0.1031  0.1864  0.0151

Bar graphs can be plotted with the function plotphi().

ClustCheck::plotphi(cbank, var = BankCustomer$profession)

• Correspondance Analysis

The clustering can be analysed through visualisation by plotting the clusters against the modes of a categorical variable (i.e. profession). It shows the frequency of the modes in each cluster as well as a plot of the coordinates of the clusters centers against the modes of the variable.

ClustCheck::vizAFC(cbank, var = BankCustomer$profession)

Numerical Variables

Lets’s focus now on the numerical variables of our dataset.

• Correlation

Correlation ratios can be computed for all numerical variables.

ClustCheck::corr_ratios(cbank)

##        age anciennete     revenu      score 
##  0.2980303  0.3505494  0.8464643  0.1119814

A bar graph of the ratios can be plotted with the function plotcorr().

ClustCheck::plotcorr(cbank)

- V-Test call

Similar to the categorical variables, test values can be computed for the numerical variables with the function tvalue_num().

ClustCheck::tvalue_num(cbank)

##                    1         2         3
## age         1.168583  5.278941 -6.362284
## anciennete -2.448253  7.143744 -4.941884
## revenu     11.150443 -4.234765 -5.898917
## score       4.048566 -2.232756 -1.444048

A bar graph of the values can be plotted with the function plottvalue().

 ClustCheck::plottvalue(cbank)

• Effect size

Effect size is another way to measure the strength of the relationship between variables and cluster groups. Cohen’s magnitude description can be used as a useful scale to evaluate this strength. In our example below, we can see the overwhelming influence of revenue in the cluster group 1. https://en.wikipedia.org/wiki/Effect_size

ClustCheck::effectsize(cbank)

##                     1          2          3
## age         0.2200033  0.9957201 -1.2732108
## anciennete -0.4683712  1.4997969 -0.9224839
## revenu      5.2245778 -0.7676589 -1.1504799
## score       0.8045345 -0.3859882 -0.2480678

Bar graphs can be plotted with the function plotsizeeff().

ClustCheck::plotsizeeff(cbank)

Analysing the clustering with multiple variables

Now we want to look at the influence on our clustering of the combination of multiple variables. We will need here to use the principles of Principal Component Analysis. https://en.wikipedia.org/wiki/Principal_component_analysis

Categorical variables

The function get_MCA() offers a graphic visualisation of a Multiple Component Analysis on the categorical variables of the dataset.

ClustCheck::get_MCA(cbank)

Numerical variables

The function get_PCA() offers a graphic visualisation of a Principal Component Analysis on the numerical variables of the dataset.

ClustCheck::get_PCA(cbank)

Mixed variables

The function get_FAMD() offers a graphic visualisation of a Factorial Analysis of Mixed Data. This function is intended for datasets with mixed data only, like in our case example.

ClustCheck::get_FAMD(cbank)

Evaluation metrics

The packages offers the choice of 3 metrics to evaluate the quality of the clustering : - Silhouette coefficient - Davies-Boulding index - Dunn index

The silhouetteC function returns the silhouette coefficient for each cluster as well as a computed mean for the whole partition. The silhouette coefficient varies from -1 to +1. A coefficient close to 1 indicates a good clustering. On the contrary a coefficient close to -1 indicates a poor clustering. https://en.wikipedia.org/wiki/Silhouette_(clustering)

ClustCheck::silhouetteC(cbank)

## $cluster_silhouette
## [1] 0.4542508 0.1458115 0.1696981
## 
## $mean_silhouette
## [1] 0.2347641

The Davies_bouldin index is always positive. Zero is a perfect score and indicates the best clustering. The higher the score, the poorer the clustering otherwise. https://en.wikipedia.org/wiki/Davies%E2%80%93Bouldin_index

ClustCheck::davies_bouldinC(cbank)

## [1] 1.840466

The Dunn index is similar to the Davis-Bouldin in its measure of clustering quality. The index value should be interpreted differently though. In this case, the higher the score, the better the clustering. https://simple.wikipedia.org/wiki/Dunn_index

ClustCheck::dunn_indexC(cbank)

## [1] 0.915563

That’s it!

You’ve completed an overview of the package main functions. Hope it will provide useful tools for happy clustering evaluation!

adrienPAVOINE/ClustCheck documentation built on Dec. 31, 2020, 6:45 p.m.