Unsupervised classification is illustrated on the titanic
dataset. It is a data.frame with 1309 observations and 8 variables containing information on the passengers of the Titanic. Each observation represents a passenger described by a set of real variables: age in years (age
), ticket price in pounds (fare
), a set of counting variables: number of siblings/spouses aboard (sibsp
), number of parents/children aboard (parch
) and a set of categorical variables: sex
, ticket class (pclass
), port of embarkation and a binary variable indicating if the passenger survived (survived
). Furthermore, the dataset contains missing values for three variables: age
, fare
and embarked
.
## pclass survived sex age sibsp parch fare embarked
## 1 1st 1 female 29.0 0 0 211.3375 S
## 16 1st 0 male NA 0 0 25.9250 S
## 38 1st 1 male NA 0 0 26.5500 S
## 169 1st 1 female 38.0 0 0 80.0000 <NA>
## 285 1st 1 female 62.0 0 0 80.0000 <NA>
## 1226 3rd 0 male 60.5 0 0 NA S
First, the dataset must be converted in the MixtComp format. Categorical variables must be numbered from 1 to the number of categories (e.g. 3 for embarked
). This can be done using the refactorCategorical
function that takes in arguments the vector containing the data, the old labels and the new labels. Totaly missing values must be indicated with a ?
.
titanicMC <- titanic
titanicMC$sex <- refactorCategorical(titanic$sex, c("male", "female"), c(1, 2))
titanicMC$pclass <- refactorCategorical(titanic$pclass, c("1st", "2nd", "3rd"), c(1, 2, 3))
titanicMC$embarked <- refactorCategorical(titanic$embarked, c("C", "Q", "S"), c(1, 2, 3))
titanicMC$survived <- refactorCategorical(titanic$survived, c(0, 1), c(1, 2))
titanicMC[is.na(titanicMC)] = "?"
head(titanicMC)
## pclass survived sex age sibsp parch fare embarked
## 1 1 2 2 29 0 0 211.3375 3
## 2 1 2 1 0.9167 1 2 151.55 3
## 3 1 1 2 2 1 2 151.55 3
## 4 1 1 1 30 1 2 151.55 3
## 5 1 1 2 25 1 2 151.55 3
## 6 1 2 1 48 0 0 26.55 3
The dataset is splitted in 2 datasets for illustrating learning and prediction.
indTrain <- sample(nrow(titanicMC), floor(0.8 * nrow(titanicMC)))
titanicMCTrain <- titanicMC[indTrain, ]
titanicMCTest <- titanicMC[-indTrain, ]
Then, as all variables are stored as character in a data.frame, a model
object indicating which model to use for each variable is created. In this example, a gaussian model is used for age
and fare
variables, a multinomial for sex
, pclass
, embarked
and survived
, a Poisson for sibsp
and parch
.
We choose to run the clustering analysis for 1 to 20 clusters with 3 runs for every number of clusters. These runs can be parallelized using the nCore
parameter.
summary
and plot
functions are used to have an overview of the results for the best number of classes according to the chosen criterion (BIC or ICL). If this number is not the one desired by the user, it can been changed via the parameter nClass
.
The summary
displays the number of clusters chosen and some outputs as the discriminative power indicating the variables that contribute most to class separation and parameters associated with the 3 most discriminant variables.
## ############### MixtCompLearn Run ###############
## nClass: 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
## Criterion used: BIC
## 1 2 3 4 5 6 7
## BIC -14402.77 -12807.94 -12081.82 -11861.55 -11672.19 -11619.08 -11449.07
## ICL -14402.77 -12841.96 -12096.52 -11921.96 -11724.80 -11686.01 -11537.20
## 8 9 10 11 12 13 14
## BIC -11520.15 -11489.33 -11469.35 -11472.09 -11375.19 -11425.32 -11460.93
## ICL -11579.53 -11550.58 -11538.77 -11551.15 -11491.42 -11568.60 -11567.78
## 15 16 17 18 19 20
## BIC -11501.76 -11584.42 -11442.32 -11598.08 -11536.71 -11559.30
## ICL -11615.10 -11710.87 -11606.24 -11756.62 -11652.34 -11680.97
## Best model: 12 clusters
## ########### MixtComp Run ###########
## Number of individuals: 1047
## Number of variables: 8
## Number of clusters: 12
## Mode: learn
## Time: 0.436 s
## SEM burn-in iterations done: 50/50
## SEM run iterations done: 50/50
## Observed log-likelihood: -10836.28
## BIC: -11375.19
## ICL: -11491.42
## Discriminative power:
## fare pclass parch sibsp embarked sex age survived
## 0.595 0.423 0.185 0.167 0.140 0.138 0.136 0.130
## Proportions of the mixture:
## 0.077 0.075 0.048 0.089 0.048 0.054 0.054 0.047 0.161 0.187 0.034 0.125
## Parameters of the most discriminant variables:
## - fare: Gaussian
## mean sd
## k: 1 28.062 10.350
## k: 2 84.347 33.233
## k: 3 27.792 1.793
## k: 4 12.366 1.262
## k: 5 15.222 8.547
## k: 6 39.076 15.930
## k: 7 41.565 23.455
## k: 8 209.304 110.168
## k: 9 7.833 0.114
## k: 10 7.808 0.731
## k: 11 68.820 20.934
## k: 12 15.540 4.704
## - pclass: Multinomial
## modality 1 modality 2 modality 3
## k: 1 0.000 1.000 0.000
## k: 2 1.000 0.000 0.000
## k: 3 1.000 0.000 0.000
## k: 4 0.000 1.000 0.000
## k: 5 0.000 0.581 0.419
## k: 6 0.000 0.000 1.000
## k: 7 0.565 0.267 0.167
## k: 8 1.000 0.000 0.000
## k: 9 0.000 0.000 1.000
## k: 10 0.000 0.000 1.000
## k: 11 1.000 0.000 0.000
## k: 12 0.000 0.000 1.000
## - parch: Poisson
## lambda
## k: 1 0.951
## k: 2 0.282
## k: 3 0.000
## k: 4 0.000
## k: 5 0.000
## k: 6 2.569
## k: 7 0.000
## k: 8 1.109
## k: 9 0.000
## k: 10 0.000
## k: 11 0.220
## k: 12 0.744
## ####################################
The plot
function displayed the values of criteria, the discriminative power of variables and the parameters of the three most discriminative variable. More variables can be displayed using the nVarMaxToPlot
parameter.
## $criteria
##
## $discrimPowerVar
##
## $proportion
##
## $fare
##
## $pclass
##
## $parch
The most discriminant variable for clustering are fare
and pclass
. The similarity between variables is shown with the following code:
## fare age pclass survived sex embarked sibsp parch
## fare 1.00 0.37 0.41 0.38 0.37 0.39 0.37 0.37
## age 0.37 1.00 0.59 0.75 0.74 0.72 0.70 0.72
## pclass 0.41 0.59 1.00 0.60 0.57 0.59 0.54 0.54
## survived 0.38 0.75 0.60 1.00 0.82 0.75 0.71 0.72
## sex 0.37 0.74 0.57 0.82 1.00 0.74 0.71 0.72
## embarked 0.39 0.72 0.59 0.75 0.74 1.00 0.69 0.69
## sibsp 0.37 0.70 0.54 0.71 0.71 0.69 1.00 0.71
## parch 0.37 0.72 0.54 0.72 0.72 0.69 0.71 1.00
The greatest similarity is between survived
and sex
, this relation is well-known in the dataset with a great proportion of women surviving compared to men. On the contrary, there is few similarity between fare
and other variables.
Getters are available to easily access some results: getBIC
, getICL
, getCompletedData
, getParam
, getProportion
, getTik
, getPartition
, … All these functions use the model maximizing the asked criterion. If results for an other number of classes is desired, the extractMixtCompObject
can be used. For example:
## k: 1 k: 2 k: 3 k: 4 k: 5 k: 6 k: 7
## 0.07729008 0.07538168 0.04770992 0.08874046 0.04770992 0.05438931 0.05438931
## k: 8 k: 9 k: 10 k: 11 k: 12
## 0.04675573 0.16125954 0.18702290 0.03435115 0.12500000
## k: 1 k: 2
## 0.4746896 0.5253104
Once a model is learnt, one can use it to predict the clusters of new individuals.
The probabilities of belonging to the different classes and the associated partition is given by:
## [,1] [,2] [,3] [,4] [,5]
## [1,] -Inf 0.000000 -Inf -Inf -Inf
## [2,] -Inf -4.381198 -Inf -0.012589273 -Inf
## [3,] -Inf -5.473653 -Inf -0.004204703 -Inf
## [4,] -Inf 0.000000 -Inf -Inf -Inf
## [5,] -Inf -4.191544 -Inf -0.015238431 -Inf
## [6,] -Inf 0.000000 -Inf -Inf -Inf
## [1] 2 4 4 2 4 2