Data Mining: Practical Machine Learning Tools and Techniques, Second Edition

Supervised attribute filters

Discretize,highlighted in Figure 10.17, uses the MDL method of supervised dis-

cretization (Section 7.2). You can specify a range of attributes or force the dis-

cretized attribute to be binary. The class must be nominal. By default Fayyad

and Irani’s (1993) criterion is used, but Kononenko’s method (1995) is an

option.

402 CHAPTER 10 | THE EXPLORER

(a) (b)

Figure 10.17Using Weka’s metalearner for discretization: (a) configuring FilteredClas-

sifier, and (b) the menu of filters.

Table 10.3 Supervised attribute filters.

Name Function

AttributeSelection Provides access to the same attribute selection methods as the

Select attributespanel

ClassOrder Randomize, or otherwise alter, the ordering of class values

Discretize Convert numeric attributes to nominal

NominalToBinary Convert nominal attributes to binary, using a supervised method

if the class is numeric

Table 10.4 Supervised instance filters.

Name Function

Resample Produce a random subsample of a dataset, sampling with

replacement

SpreadSubsample Produce a random subsample with a given spread between

class frequencies, sampling with replacement

StratifiedRemoveFolds Output a specified stratified cross-validation fold for the dataset