clusterInit: Create an instance of [`ClusterInit`] class

Description

The initialization step is a two stages process: the proper initialization step and some (optionnals) iterations of an algorithm [clusterAlgo].

Usage

clusterInit(
  method = "class",
  nbInit = 5,
  algo = "EM",
  nbIteration = 20,
  epsilon = 0.01
)

Value

a [ClusterInit] object

Arguments

method: Character string with the initialisation method. Possible values: "random", "class", "fuzzy". Default value is "class".
nbInit: integer defining the number of initialization point to test. Default value is 5.
algo: String with the initialisation algorithm. Possible values: "EM", "CEM", "SEM", "SemiSEM". Default value is "EM".
nbIteration: Integer defining the number of iteration in algo. nbIteration must be a positive integer. Default values is 20. if .
epsilon: threshold to use in order to stop the iterations. Default value is 0.01.

Author

Serge Iovleff

Details

There is three ways to initialize the parameters:

random: The initial parameters of the mixture are chosen randomly
class: The initial membership of individuals are sampled randomly
fuzzy: The initial probabilities of membership of individuals are sampled randomly

A few iterations of an algorithm [clusterAlgo] are then performed. It is strongly recommended to use a few number of iterations of the EM or SEM algorithms after initialization. This allows to detect "bad" initialization starting point.

These two stages are repeated until nbInit is reached. The initial point with the best log-likelihood is conserved as the initial starting point.

Examples

Run this code

 clusterInit(method = "class", nbInit=1, algo="CEM",nbIteration=50, epsilon=0.00001)
 clusterInit(nbIteration=0) # no algorithm

Run the code above in your browser using DataLab