CBA_ruleset: Constructor for Objects for Classifiers Based on Association Rules

Description

Objects for classifiers based on association rules have class CBA. A creator function CBA_ruleset() and several methods are provided.

Usage

CBA_ruleset(
  formula,
  rules,
  default,
  method = "first",
  weights = NULL,
  bias = NULL,
  model = NULL,
  discretization = NULL,
  description = "Custom rule set",
  ...
)

Value

A object of class CBA representing the trained classifier with fields:

formula: used formula.
rules: the classifier rule base.
default: default class label (uses partial matching against the class labels).
method: classification method.
weights: rule weights.
bias: class bias vector if available.
model: list with model description.
discretization: discretization information.
description: description in human readable form.

rules returns the rule base.

Arguments

formula: A symbolic description of the model to be fitted. Has to be of form class ~ .. The class is the variable name (part of the item label before =).
rules: A set of class association rules mined with mineCARs() or arules::apriori() (from arules).
default: Default class. If not specified then objects that are not matched by rules are classified as NA.
method: Classification method "first" found rule or "majority".
weights: Rule weights for the majority voting method. Either a quality measure available in the classification rule set or a numeric vector of the same length are the classification rule set can be specified. If missing, then equal weights are used
bias: Class bias vector.
model: An optional list with model information (e.g., parameters).
discretization: A list with discretization information used by predict() to discretize data supplied as a data.frame.
description: Description field used when the classifier is printed.
...: Additional arguments added as list elements to the CBA object.

Author

Michael Hahsler

Details

CBA_ruleset() creates a new object of class CBA using the provides rules as the rule base. For method "first", the user needs to make sure that the rules are predictive and sorted from most to least predictive.

Examples

Run this code

## Example 1: create a first-matching-rule classifier with non-redundant rules
##  sorted by confidence.
data("iris")

# discretize and create transactions
iris.disc <- discretizeDF.supervised(Species ~., iris)
trans <- as(iris.disc, "transactions")

# create rule base with CARs
cars <- mineCARs(Species ~ ., trans, parameter = list(support = .01, confidence = .8))

cars <- cars[!is.redundant(cars)]
cars <- sort(cars, by = "conf")

# create classifier and use the majority class as the default if no rule matches.
cl <- CBA_ruleset(Species ~ .,
  rules = cars,
  default = uncoveredMajorityClass(Species ~ ., trans, cars),
  method = "first")
cl

# look at the rule base
inspect(cl$rules)

# make predictions
prediction <- predict(cl, trans)
table(prediction, response(Species ~ ., trans))
accuracy(prediction, response(Species ~ ., trans))

# Example 2: use weighted majority voting.
cl <- CBA_ruleset(Species ~ .,
  rules = cars,
  default = uncoveredMajorityClass(Species ~ ., trans, cars),
  method = "majority", weights = "lift")
cl

prediction <- predict(cl, trans)
table(prediction, response(Species ~ ., trans))
accuracy(prediction, response(Species ~ ., trans))

## Example 3: Create a classifier with no rules that always predicts
##  the majority class. Note, we need cars for the structure and subset it
##  to leave no rules.
cl <- CBA_ruleset(Species ~ .,
  rules = cars[NULL],
  default = majorityClass(Species ~ ., trans))
cl

prediction <- predict(cl, trans)
table(prediction, response(Species ~ ., trans))
accuracy(prediction, response(Species ~ ., trans))

Run the code above in your browser using DataLab