13.07.2015 Views

Data Mining: Practical Machine Learning Tools and ... - LIDeCC

Data Mining: Practical Machine Learning Tools and ... - LIDeCC

Data Mining: Practical Machine Learning Tools and ... - LIDeCC

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

414 CHAPTER 10 | THE EXPLORERLBR (for Lazy Bayesian Rules) is a Bayesian classifier that defers all processingto classification time. For each test instance it selects a set of attributes forwhich the independence assumption should not be made; the others are treatedas independent of each other given the class <strong>and</strong> the selected set of attributes.It works well for small test sets (Zheng <strong>and</strong> Webb 2000).LWL is a general algorithm for locally weighted learning. It assigns weightsusing an instance-based method <strong>and</strong> builds a classifier from the weightedinstances. The classifier is selected in LWL’s object editor: a good choice is NaïveBayes for classification problems <strong>and</strong> linear regression for regression problems(Section 6.5, pages 251–253). You can set the number of neighbors used, whichdetermines the kernel b<strong>and</strong>width, <strong>and</strong> the kernel shape to use for weighting—linear, inverse, or Gaussian. Attribute normalization is turned on by default.Miscellaneous classifiersThe misc. category includes two simple classifiers that were mentioned at theend of Section 4.7 (page 136). Hyperpipes, for discrete classification problems,records the range of values observed in the training data for each attribute <strong>and</strong>category <strong>and</strong> works out which ranges contain the attribute values of a test instance,choosing the category with the largest number of correct ranges. VFI(voting feature intervals) constructs intervals around each class by discretizingnumeric attributes <strong>and</strong> using point intervals for nominal ones, records classcounts for each interval on each attribute, <strong>and</strong> classifies test instances by voting(Demiroz <strong>and</strong> Guvenir 1997). A simple attribute weighting scheme assignshigher weight to more confident intervals, where confidence is a function ofentropy. VFI is faster than Naïve Bayes but slower than hyperpipes. Neithermethod can h<strong>and</strong>le missing values.10.5 Metalearning algorithmsMetalearning algorithms, listed in Table 10.6, take classifiers <strong>and</strong> turn them intomore powerful learners. One parameter specifies the base classifier; othersspecify the number of iterations for schemes such as bagging <strong>and</strong> boosting <strong>and</strong>an initial seed for the r<strong>and</strong>om number generator. We already met FilteredClassifierin Section 10.3: it runs a classifier on data that has been passed through afilter, which is a parameter. The filter’s own parameters are based exclusively onthe training data, which is the appropriate way to apply a supervised filter totest data.Bagging <strong>and</strong> r<strong>and</strong>omizationBagging bags a classifier to reduce variance (Section 7.5, page 316). This implementationworks for both classification <strong>and</strong> regression, depending on the base

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!