Evaluation of ordered attributes
The method evaluates the quality of ordered attributes specified by the formula with ordEval algorithm.
ordEval(formula, data, file=NULL, rndFile=NULL, variant=c("allNear","attrDist1","classDist1"), ...)
formula: Either a formula specifying the attributes to be evaluated and the target variable, or a name of target variable, or an index of target variable.data: Data frame with evaluation data.file: Name of file where evaluation results will be written to.rndFile: Name of file where evaluation of random normalizing attributes will be written to.variant: Name of the variant of ordEval algorithm. Can be any of "allNear", "attrDist1", or "classDist1"....: Other options specific to ordEval or common to other context-sensitive evaluation methods (e.g., ReliefF).The parameter formula can be interpreted in three ways, where the formula interface is the most elegant one, but inefficient and inappropriate for large data sets. See also examples below. As formula one can specify:
formula: used as a mechanism to select features (attributes) and prediction variable (class). Only simple terms can be used and interaction expressed in formula syntax are not supported. The simplest way is to specify just response variable: class ~ .. In this case all other attributes in the data set are evaluated. Note that formula interface is not appropriate for data sets with large number of variables.data are used as predictors.data, all the other columns are used as predictors.In the data frame data take care to supply the ordinal data as factors and to provide equal levels for them (this is not necessary what one gets with read.table). See example below.
The output can be optionally written to files file and rndFile, in a format used by visualization methods in plotOrdEval.
The variant of the algorithm actually used is controlled with variant parameter which can have values "allNear", "attrDist1", and "classDist1". The default value is "allNear" which takes all nearest neighbors into account in evaluation of attributes. Variant "attrDist1" takes only neighbors with attribute value at most 1 different from current case into account (for each attribute separately). This makes sense when we want to see the thresholds of reinforcement, and therefore observe just small change up or down (it makes sense to combine this with equalUpDown=TRUE in plot.ordEval function). The "classDist1" variant takes only neighbors with class value at most 1 different from current case into account. This makes sense if we want to observe strictly small changes in upward/downward reinforcement and has little effect in practical applications.
There are some additional parameters (note ... ) some of which are common with other context-sensitive evaluation methods (e.g., ReliefF). Their list of common parameters is available in helpCore (see subsection on attribute evaluation therein). The parameters specific to ordEval are:
ordEvalNoRandomNormalizers: type: integer, default value: 0, value range: 0, Inf,
number of randomly shuffled attributes for normalization of each attribute (0=no normalization). This parameter should be set to a reasonably high value (e.g., 200) in order to produce reliable confidence intervals with `plot.ordEval`. The parameters `ordEvalBootstrapNormalize` and `ordEvalNormalizingPercentile` only make sense if this parameter is larger than 0.
ordEvalBootstrapNormalize: type: logical, default value: FALSE
are features used for normalization constructed with bootstrap sampling or random permutation.
ordEvalNormalizingPercentile: type: numeric, default value: 0.025, value range: 0, 0.5
percentile defines the length of confidence interval obtained with random normalization. Percentile `t` forms interval by taking the $nt$ and $n(1-t)$ random evaluation as the confidence interval boundaries, thereby forming $100(1-2t)$% confidence interval (`t`=0.025 gives 95% confidence interval). The value $n$ is set by `ordEvalNoRandomNormalizers` parameter.
attrWeights: type: character,
a character vector representing a list of attribute weights in the ordEval distance measure.
Evaluation of attributes without specifics of ordered attributes is covered in function attrEval.
The method returns a list with following components: - reinfPosAV: a matrix of positive reinforcement for attributes' values,
reinfNegAV: a matrix of negative reinforcement for attributes' values,
anchorAV: a matrix of anchoring for attributes' values,
noAV: a matrix containing count for each value of each attribute,
reinfPosAttr: a vector of positive reinforcement for attributes,
reinfNegAttr: a matrix of negative reinforcement for attributes,
anchorAttr: a matrix of anchoring for attributes,
noAVattr: a vector containing count of valid values of each attribute,
rndReinfPosAV: a three dimensional array of statistics for random normalizing attributes' positive reinforcement for attributes' values,
rndReinfPosAV: a three dimensional array of statistics for random normalizing attributes' negative reinforcement for attributes' values,
rndAnchorAV: a three dimensional array of statistics for random normalizing attributes' anchoring for attributes' values,
rndReinfPosAttr: a three dimensional array of statistics for random normalizing attributes' positive reinforcement for attributes,
rndReinfPosAttr: a three dimensional array of statistics for random normalizing attributes' negative reinforcement for attributes,
rndAnchorAttr: a three dimensional array of statistics for random normalizing attributes' anchoring for attributes.
attrNames: the names of attributes
valueNames: the values of attributes
noAttr: number of attributes
ordVal: maximal number of attribute values
variant: the variant of the algorithm used
file: the file to store the results
rndFile: the file to store random normalizations
The statistics used are median, 1st quartile, 3rd quartile, low and high percentile selected by
ordEvalNormalizingPercentile, mean, standard deviation, and expected probability according to value distribution. With these statistics we can visualize significance of reinforcements using adapted box and whiskers plot.
#prepare a data set dat <- ordDataGen(200) # evaluate ordered features with ordEval est <- ordEval(class ~ ., dat, ordEvalNoRandomNormalizers=100) # print(est) printOrdEval(est) plot(est)
Marko Robnik-Sikonja, Koen Vanhoof: Evaluation of ordinal attributes at value level. Knowledge Discovery and Data Mining, 14:225-243, 2007
Marko Robnik-Sikonja, Igor Kononenko: Theoretical and Empirical Analysis of ReliefF and RReliefF. Machine Learning Journal, 53:23-69, 2003
Some of the references are available also from http://lkm.fri.uni-lj.si/rmarko/papers/
Marko Robnik-Sikonja
plot.ordEval, CORElearn, CoreModel, helpCore, infoCore.