ComputeInterestingTuples: Interesting tuples

In MDFS: MultiDimensional Feature Selection

Interesting tuples

Description

Interesting tuples

Usage

ComputeInterestingTuples(
  data,
  decision = NULL,
  dimensions = 2,
  divisions = 1,
  discretizations = 1,
  seed = NULL,
  range = NULL,
  pc.xi = 0.25,
  ig.thr = 0,
  I.lower = NULL,
  interesting.vars = vector(mode = "integer"),
  require.all.vars = FALSE,
  return.matrix = FALSE,
  stat_mode = "MI",
  average = FALSE
)

Arguments

data	input data where columns are variables and rows are observations (all numeric)
decision	decision variable as a binary sequence of length equal to number of observations
dimensions	number of dimensions (a positive integer; 5 max)
divisions	number of divisions (from 1 to 15)
discretizations	number of discretizations
seed	seed for PRNG used during discretizations (NULL for random)
range	discretization range (from 0.0 to 1.0; NULL selects probable optimal number)
pc.xi	parameter xi used to compute pseudocounts (the default is recommended not to be changed)
ig.thr	IG threshold above which the tuple is interesting (0 and negative mean no filtering)
I.lower	IG values computed for lower dimension (1D for 2D, etc.)
interesting.vars	variables for which to check the IGs (none = all)
require.all.vars	boolean whether to require tuple to consist of only interesting.vars
return.matrix	boolean whether to return a matrix instead of a list (ignored if not using the optimised method variant)
stat_mode	character, one of: "MI" (mutual information, the default; becomes information gain when decision is given), "H" (entropy; becomes conditional entropy when decision is given), "VI" (variation of information; becomes target information difference when decision is given); decides on the value computed
average	boolean whether to average over discretisations instead of maximising (the default)

Details

If running in 2D and no filtering is applied, this function is able to run in an optimised fashion. It is recommended to avoid filtering in 2D if only it is feasible.

This function calculates what stat_mode dictates. When decision is omitted, the stat_mode is calculated on the descriptive variables. When decision is given, the stat_mode is calculated on the decision variable, conditional on the other variables. Translate "IG" to that value in the rest of this function's description.

Value

A data.frame or NULL (following a warning) if no tuples are found.

The following columns are present in the data.frame:

• Var – interesting variable index

• Tuple.1, Tuple.2, ... – corresponding tuple (up to dimensions columns)

• IG – information gain achieved by var in Tuple.*

Additionally attribute named run.params with run parameters is set on the result.

Examples

ig.1d <- ComputeMaxInfoGains(madelon$data, madelon$decision, dimensions = 1, divisions = 1,
                             range = 0, seed = 0)
ComputeInterestingTuples(madelon$data, madelon$decision, dimensions = 2, divisions = 1,
                         range = 0, seed = 0, ig.thr = 100, I.lower = ig.1d$IG)

MDFS documentation built on May 31, 2023, 7:31 p.m.