Extended WEKA including Ensembles of Hierarchically Nested Dichotomies

Overview for extended WEKA including Ensembles of Hierarchically Nested Dichotomies.

See:
Description

Packages

weka.associations
weka.associations.tertius
weka.attributeSelection
weka.classifiers
weka.classifiers.bayes
weka.classifiers.evaluation
weka.classifiers.functions
weka.classifiers.functions.neural
weka.classifiers.functions.pace
weka.classifiers.functions.supportVector
weka.classifiers.lazy
weka.classifiers.lazy.kstar
weka.classifiers.meta
weka.classifiers.misc
weka.classifiers.rules
weka.classifiers.rules.part
weka.classifiers.trees
weka.classifiers.trees.adtree
weka.classifiers.trees.j48
weka.classifiers.trees.lmt
weka.classifiers.trees.m5
weka.clusterers
weka.core
weka.core.converters
weka.datagenerators
weka.estimators
weka.experiment
weka.filters
weka.filters.supervised.attribute
weka.filters.supervised.instance
weka.filters.unsupervised.attribute
weka.filters.unsupervised.instance
weka.gui
weka.gui.beans
weka.gui.boundaryvisualizer
weka.gui.experiment
weka.gui.explorer
weka.gui.graphvisualizer
weka.gui.streams
weka.gui.treevisualizer
weka.gui.visualize

Overview for extended WEKA including Ensembles of Hierarchically Nested Dichotomies.

The extensions for EHNDs are based on WEKA version 3-4-1 (March 2004) and the classes weka.classifiers.meta.END and weka.classifiers.meta.ND as provided by Eibe Frank.

Newly incorporated classes for covering EHNDs

• weka.classifiers.meta.NestedDichotomy
• weka.classifiers.meta.HND
• weka.core.ClassHierarchy
• weka.core.ClassTree
• weka.core.converters.ClassHierarchyParser
• weka.core.converters.AbstractClassHierarchyParser
• weka.core.converters.ClassTreeParser
• weka.core.converters.ClassTreeFileParser
• weka.core.converters.ClassTreeArffFileParser
• weka.core.converters.HierarchicalCostMatrix

Changes in original WEKA classes, END, and ND for incorporating EHNDs

• weka.classifiers.meta.END:
  • Requires weka.classifiers.meta.NestedDichotomy as base learner instead of weka.classifiers.meta.ND.
  • Respective change of text in thrown Exception.
  • Omission of setHashtable() in preparation of base classifier.
• weka.classifiers.meta.ND: Implements marker-interface weka.classifiers.meta.NestedDichotomy.
• weka.core.Instances:
  • Added attribute: lastRemoved to keep track of last removed attribute-position.
  • Added attribute: classRemoved to provide whether the class-attribute was removed. To be set to false, if the class-attribute was added again or another attribute was deleted afterwards.
  • Changed method: deleteAttributeAt(int position): keeps track of the currently deleted position and whether the class-attribute was removed.
  • Changed method: insertAttributeAt(Attribute att, int position): If the class-attribute was removed and position equals lastRemoved, m_ClassIndex is set to position and classRemoved is set to false.

Usage

Currently neither END, nor ND or HND are enabled to run with graphical user interface. For usage of EHNDs in command-line interface call weka.classifiers.meta.END with the respective parameters, weka.classifiers.meta.HND as base learner.

Planned changes for incorporation into a future WEKA distribution

In theory HND provides the means to run a hierarchical classification with any base classifier. We propose to generalize the class to handle any base learner, and to incorporate it as an independent meta schema.

Currently the hierarchy can be encoded in an arff file as comment. Thus the hierarchy will get lost if a respective new file is created after filtering. Perhaps it could be possible to incorporate a new (optional) key-word to the arff-format denoting a line that provides an encoded hierarchy.

Furthermore, by omitting the call of setHashtable(), END become a merely arbitrary ensemble method, that is not realy specialized to run with ND or HND as base learner only. Thus the restriction by requiring a base learner implementing the marker interface weka.classifiers.meta.NestedDichotomy could eventually be removed. The same effect, however, could perhaps be achieved using RandomCommittee, so END might become somewhat redundant.