Package

weka.filters.unsupervised.attribute

Synopsis

Replaces missing numeric values using Expectation Maximization with a multivariate normal model. Described in " Schafer, J.L. Analysis of Incomplete Multivariate Data, New York: Chapman and Hall, 1997."

Available in Weka 3.7.0 - 3.7.1. Available via the package management system for Weka >= 3.7.2 (EMImputation)

Options

The table below describes the options available for EMImputation.

Option	Description
debug	Turns on output of debugging information.
logLikelihoodThreshold	Log-likelihood threshold for convergence in Expectation Maximization. If the change in the observed data log-likelihood across iterations is no more than this value, then convergence is considered to be achieved and the iterative process is ceased. (default = 0.0001)
numIterations	Maximum number of iterations for Expectation Maximization. EM is used to initialize the parameters of the multivariate normal distribution. (-1 = no maximum)
ridge	Ridge parameter for ridge prior.
useRidgePrior	Use a ridge prior instead of noninformative prior.

Capabilities

The table below describes the capabilites of EMImputation.

Capability	Supported
Class	Empty nominal class, Unary class, Missing class values, String class, Relational class, Nominal class, Date class, No class, Numeric class, Binary class
Attributes	Missing values, Numeric attributes
Min # of instances	0