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.

418 CHAPTER 10 | THE EXPLORERThe third metalearner, ThresholdSelector, optimizes the F-measure (Section5.7) by selecting a probability threshold on the classifier’s output. Performancecan be measured on the training data, on a holdout set, or by cross-validation.The probabilities returned by the base learner can be rescaled into the full range[0,1], which is useful if the scheme’s probabilities are restricted to a narrow subrange.The metalearner can be applied to multiclass problems by specifying theclass value for which the optimization is performed as1. The first class value2. The second class value3. Whichever value is least frequent4. Whichever value is most frequent5. The first class named yes, pos(itive), or 1.Retargeting classifiers for different tasksFour metalearners adapt learners designed for one kind of task to another. ClassificationViaRegressionperforms classification using a regression method bybinarizing the class <strong>and</strong> building a regression model for each value. Regression-ByDiscretization is a regression scheme that discretizes the class attribute into aspecified number of bins using equal-width discretization <strong>and</strong> then employs aclassifier. The predictions are the weighted average of the mean class value foreach discretized interval, with weights based on the predicted probabilities forthe intervals. OrdinalClassClassifier applies st<strong>and</strong>ard classification algorithms toordinal-class problems (Frank <strong>and</strong> Hall 2001). MultiClassClassifier h<strong>and</strong>lesmulticlass problems with two-class classifiers using any of these methods:1. One versus all the rest2. Pairwise classification using voting to predict3. Exhaustive error-correcting codes (Section 7.5, page 334)4. R<strong>and</strong>omly selected error-correcting codesR<strong>and</strong>om code vectors are known to have good error-correcting properties: aparameter specifies the length of the code vector (in bits).10.6 Clustering algorithmsTable 10.7 lists Weka’s clustering algorithms; the first two <strong>and</strong> SimpleKMeans aredescribed in Section 6.6. For the EM implementation you can specify how manyclusters to generate or the algorithm can decide using cross-validation—inwhich case the number of folds is fixed at 10 (unless there are fewer than 10training instances). You can specify the maximum number of iterations <strong>and</strong> setthe minimum allowable st<strong>and</strong>ard deviation for the normal density calculation.

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

Saved successfully!

Ooh no, something went wrong!