RankControl-class function

RankControl Class

A virtual S4 class to store control parameters for model fitting. class

Details

RankControl class must be extended to reflect what distance metric should be used. Possibles extensions are RankControlWeightedKendall, RankControlKendall, RankControlPhiComponent, RankControlWtau, RankControlSpearman, RankControlFootrule, RankControlHamming, and RankControlCayley.

The control parameters that start with prefix EM_ are intended for the EM iteration. The ones with prefix SeachPi0 control the behaviour of searching model ranking.

Slots

• EM_limit: maximum number of EM iteration
• EM_epsilon: convergence error for weights and cluster probabilities in EM iteration
• SearchPi0_limit: maximum number of iterations in the local search of pi0.
• SearchPi0_FUN: a function object that gives a goodness of fit criterion. The default is log likelihood.
• SearchPi0_fast_traversal: a logical value. If TRUE (by default), immediately traverse to the neighbour if it is better than the current best. Otherwise, check all neighbours and traverse to the best one.
• SearchPi0_show_message: a logical value. If TRUE, the location of the current pi0 is shown.
• SearchPi0_neighbour: a character string specifying which type of neighbour to use in the local search. Supported values are: "Cayley" to use neighbours in terms of Cayley distance or "Kendall" to use neighbours in terms of Kendall distance. Note that Kendall neighbours are a subset of Cayley neighbours

User-defined Criterion

You can specify user-defined criterion to choose modal rankings. The function object SearchPi0_FUN takes a list as argument. The components in the list include the following. obs: the number of observations. w.est: the estimated weights. log_likelihood: the estimated log_likelihood. With this information, most of the popular information criterion can be supported and customized criterion can also be defined. A larger returned value indicates a better fit. Note that if you are fitting a mixture model the EM algorithm always tries to maximized the log likelihood. Thus the default value should be used in this case.

References

Qian Z, Yu L. H. P (2019) "Weighted Distance-Based Models for Ranking Data Using the R Package rankdist." Journal of Statistical Software, 90 (5), 1-31. doi: 10.18637/jss.v090.i05

See Also

RankData, RankInit