Custom Methods: Using Custom Depth Functions and Classifiers

Description

To use a custom depth function or classifier one has to implement three functions: parameters validator, learning and calculating functions.

Arguments

Details

To define a depth function:

.<NAME>_validate

validates parameters passed to ddalpha.train and passes them to the ddalpha object.

IN:	`ddalpha`
the ddalpha object, containing the data and settings (mandatory)	`<custom parameters>`
parameters that are passed to the user-defined method	`...`
other parameters (mandatory)	OUT:
`list()`	list of output parameters, after the validation is finished, these parameters are stored in the `ddalpha` object

.<NAME>_learn

trains the depth

IN:	`ddalpha`
the ddalpha object, containing the data and settings	MODIFIES:
`ddalpha`	store the calculated statistics in the `ddalpha` object
depths	calculate the depths of each pattern, e.g.
	`for (i in 1:ddalpha$numPatterns) ddalpha$patterns[[i]]$depths = .<NAME>_depths(ddalpha, ddalpha$patterns[[i]]$points)`
OUT:	`ddalpha`

.<NAME>_depths

calculates the depths

IN:	`ddalpha`
the ddalpha object, containing the data and settings	`objects`
the objects for which the depths are calculated
OUT:	`depths`

Usage: ddalpha.train(data, depth = "<NAME>", <custom parameters>, ...)

#### Custom depths ####.MyDepth_validate <- function(ddalpha, mydepth.parameter = "value", ...){ print("MyDepth validating") # validate the parameters if (!is.valid(mydepth.parameter)){ warning("Argument \"mydepth.parameter\" not specified correctly. Default value is used") mydepth.parameter = "value" # or stop("Argument \"mydepth.parameter\" not specified correctly.") } # the values from the return list will be stored in the ddalpha object return (list(mydepthpar = mydepth.parameter)) }.MyDepth_learn <- function(ddalpha){ print("MyDepth learning") #1. Calculate any statistics based on data that .MyDepth_depths needs # and store them to the ddalpha object: ddalpha$mydepth.statistic = "some value" #2. Calculate depths for each pattern for (i in 1:ddalpha$numPatterns){ ddalpha$patterns[[i]]$depths = .MyDepth_depths(ddalpha, ddalpha$patterns[[i]]$points) } return(ddalpha) }.MyDepth_depths <- function(ddalpha, objects){ print("MyDepth calculating") depths <- NULL # The depth parameters are accessible in the ddalpha object: mydepth.parameter = ddalpha$mydepth.parameter mydepth.statistic = ddalpha$mydepth.statistic #calculate the depths of the objects w.r.t. each pattern for (i in 1:ddalpha$numPatterns){ depth_wrt_i = #calculate depths of the objects, as vector depths <- cbind(depths, depth_wrt_i) } return (depths) }ddalpha.train(data, depth = "MyDepth", ...)

To define a classifier:

.<NAME>_validate

validates parameters passed to ddalpha.train and passes them to the ddalpha object

IN:	`ddalpha`
the ddalpha object, containing the data and settings (mandatory)	`<custom parameters>`
parameters that are passed to the user-defined method	`...`
other parameters (mandatory)	OUT:
`list()`	list of output parameters, after the validation is finished, these parameters are stored in the `ddalpha` object. In case of a multiclass classifier the validator must return `methodSeparatorBinary = F` and/or pass `aggregation.method = "none"` to `ddalpha.train`

.<NAME>_learn

trains the classifier. Is different for binnary and mylticlass classifiers.

IN:	`ddalpha`
the ddalpha object, containing the data and settings	`index1`
(only for binary) index of the first class	`index2`
(only for binary) index of the second class	`depths1`
(only for binary) depths of the first class w.r.t. all classes	`depths2`
(only for binary) depths of the second class w.r.t. all classes
	depths w.r.t. only given classes are received by `depths1[,c(index1, index2)]`

for the multiclass classifiers the depths are accessible via `ddalpha$patterns[[i]]$depths`	OUT:
`classifier`	the trained `classifier` object

.<NAME>_classify

classifies the objects

IN:	`ddalpha`
the ddalpha object, containing the data and global settings	`classifier`
the previously trained classifier	`objects`
the objects (depths) that are classified
OUT:	`result`
a vector with classification results
(binary) the objects from class `"classifier$index1"` get positive values

Usage: ddalpha.train(data, separator = "<NAME>", ...)

#### Custom classifiers ####.MyClassifier_validate <- function(ddalpha, my.parameter = "value", ...){ print("MyClassifier validating") # validate the parameters ... # always include methodSeparatorBinary. # TRUE for the binary classifier, FALSE otherwise return(list(methodSeparatorBinary = T, my.parameter = my.parameter )) }# a binary classifier # the package takes care of the voting procedures. Just train it as if there are only two classes .MyClassifier_learn <- function(ddalpha, index1, index2, depths1, depths2){ print("MyClassifier (binary) learning") # The parameters are accessible in the ddalpha object: my.parameter = ddalpha$my.parameter #depths w.r.t. only given classes are received by depths1[,c(index1, index2)] depths2[,c(index1, index2)] # train the classifier classifier <- ... #return the needed values in a list, e.g. return(list( coefficients = classifier$coefficients, ... )) }# a multiclass classifier .MyClassifier_learn <- function(ddalpha){ print("MyClassifier (multiclass) learning") # access the data through the ddalpha object, e.g. for (i in 1:ddalpha$numPatterns){ depth <- ddalpha$patterns[[i]]$depths number <- ddalpha$patterns[[i]]$cardinality ... } # train the classifier classifier <- ... # return the classifier return(classifier) }# the interface of the classify function is equal for binary and multiclass classifiers .MyClassifier_classify <- function(ddalpha, classifier, depths){ print("MyClassifier classifying") # The global parameters are accessible in the ddalpha object: my.parameter = ddalpha$my.parameter # The parameters generated by .MyClassifier_learn are accessible in the classifier object: classifier$coefficients # here are the depths w.r.t. the first class depths[,classifier$index1] # here are the depths w.r.t. the second class depths[,classifier$index2] # fill results in a vector, so that: # (binary) the objects from class "classifier$index1" get positive values # (multiclass) the objects get the numbers of patterns in ddalpha result <- ... return(result) }ddalpha.train(data, separator = "MyClassifier", ...)

Examples

Run this code


if (FALSE) {

#### example:  Euclidean depth ####

#.Euclidean_validate is basically not needed

.Euclidean_learn <- function(ddalpha){
  print("Euclidean depth learning")
  
  #1. Calculate any statistics based on data that .MyDepth_depths needs
  #   and store them to the ddalpha object: 
  for (i in 1:ddalpha$numPatterns){
    ddalpha$patterns[[i]]$center <- colMeans(ddalpha$patterns[[i]]$points)
  }
  
  #2. Calculate depths for each pattern
  for (i in 1:ddalpha$numPatterns){
    ddalpha$patterns[[i]]$depths = .Euclidian_depths(ddalpha, ddalpha$patterns[[i]]$points)
  }
  
  return(ddalpha)
}

.Euclidean_depths <- function(ddalpha, objects){
  print("Euclidean depth calculating")
  depths <- NULL
  
  #calculate the depths of the objects w.r.t. each pattern
  for (i in 1:ddalpha$numPatterns){
    # The depth parameters are accessible in the ddalpha object:
    center = ddalpha$patterns[[i]]$center
    
    depth_wrt_i <- 1/(1 + colSums((t(objects) - center)^2))
    depths <- cbind(depths, depth_wrt_i)
  }
  
  return (depths)
}

#### example:  binary decision tree ####

library(rpart)

.tree_validate  <- function(ddalpha, ...){
  print("tree validating")
  return(list(methodSeparatorBinary = T))
}

# a binary classifier 
# the package takes care of the voting procedures. Just train it as if there are only two classes
.tree_learn <- function(ddalpha, index1, index2, depths1, depths2){
  print("tree learning")
  
  # prepare the data
  data = as.data.frame(cbind( (rbind(depths1, depths2)), 
                              c(rep(1, times = nrow(depths1)), rep(-1, times = nrow(depths1)))))
  names(data) <- paste0("V",seq_len(ncol(data)))
  names(data)[ncol(data)] <- "O"
  # train the classifier
  classifier <- rpart(O~., data = data)
  
  #return the needed values in a list, e.g.
  return(classifier)
}

# the interface of the classify function is equal for binary and multiclass classifiers
.tree_classify <- function(ddalpha, classifier, depths){
  print("tree classifying")
  
  # fill results in a vector, so that the objects from class "classifier$index1" get positive values
  data = as.data.frame(depths)
  names(data) <- paste0("V",seq_len(ncol(data)))
  result <- predict(classifier, as.data.frame(depths), type = "vector")
  
  return(result)
}




#### checking ####

library(ddalpha)
data = getdata("hemophilia")

ddalpha = ddalpha.train(data = data, depth = "Euclidean", separator = "tree")
c = ddalpha.classify(ddalpha, data[,1:2])
errors = sum(unlist(c) != data[,3])/nrow(data)
print(paste("Error rate: ",errors))

# building the depth contours using the custom depth
depth.contours.ddalpha(ddalpha, asp = T, levels = seq(0.5, 1, length.out = 10))

}

Run the code above in your browser using DataLab

Description

Arguments

Details

See Also

Examples