prune: Classifier Builder

Description

An implementation of the CBA-CB M1 algorithm (Liu et al, 1998) adapted for R and arules package apriori implementation in place of CBA-RG.

Usage

prune(
  rules,
  txns,
  classitems,
  default_rule_pruning = TRUE,
  rule_window = 50000,
  greedy_pruning = FALSE,
  input_list_sorted_by_length = TRUE,
  debug = FALSE
)

Value

Returns an object of class rules. Note that `rules@quality` slot has been extended with additional measures, specifically `orderedConf`, `orderedSupp`, and `cumulativeConf`. The rules are output in the order in which they are assumed to be applied in classification. Only the first applicable rule is used to classify the instance. As a result, in addition to rule confidence -- which is computed over the whole training dataset -- it makes sense to define order-sensitive confidence, which is computed only from instances reaching the given rule as \(a/(a+b)\), where \(a\) is the number of instances matching both the antecedent and consequent (available in slot `orderedSupp`) and \(b\) is the number of instances matching the antecedent, but not matching the consequent of the given rule. The cumulative confidence is an experimental measure, which is computed as the accuracy of the rule list comprising the given rule and all higher priority rules (rules with lower index) with uncovered instances excluded from the computation.

Arguments

rules: object of class rules from arules package
txns: input object with transactions.
classitems: a list of items to appear in the consequent (rhs) of the rules.
default_rule_pruning: boolean indicating whether default pruning should be performed. If set to TRUE, default pruning is performed as in the CBA algorithm. If set to FALSE, default pruning is not performed i.e. all rules surviving data coverage pruning are kept. In either case, a default rule is added to the end of the classifier.
rule_window: the number of rules to precompute for CBA data coverage pruning. The default value can be adjusted to decrease runtime.
greedy_pruning: setting to TRUE activates early stopping condition: pruning will be stopped on first rule on which total error increases.
input_list_sorted_by_length: indicates by default that the input rule list is sorted by antecedent length (as output by arules), if this param is set to false, the list will be resorted
debug: output debug messages.

References

Ma, Bing Liu Wynne Hsu Yiming. Integrating classification and association rule mining. Proceedings of the fourth international conference on knowledge discovery and data mining. 1998.

Examples

Run this code

 #Example 1
  txns <- as(discrNumeric(datasets::iris, "Species")$Disc.data,"transactions")
  appearance <- getAppearance(datasets::iris,"Species")
  rules <- apriori(txns, parameter = list(confidence = 0.5,
  support= 0.01, minlen= 2, maxlen= 4),appearance = appearance)
  prune(rules,txns, appearance$rhs)
  inspect(rules)

#Example 2
 utils::data(Adult) # this dataset comes with the arules package
 classitems <- c("income=small","income=large")
 rules <- apriori(Adult, parameter = list(supp = 0.3, conf = 0.5,
 target = "rules"), appearance=list(rhs=classitems, default="lhs"))
 # produces 25 rules
 rulesP <- prune(rules,Adult,classitems)
 rulesP@quality # inspect rule quality measured including the new additions
 # Rules after data coverage pruning: 8
 # Performing default rule pruning.
 # Final rule list size:  6