Andreas Hess // Projects // Triskel // Diagrams :: Table

The Triskel Algorithm

The Triskel algorithm is a novel ensemble learning method that is based on a set of classifiers that are biased towards high precision on a particular class each. For a detailed explanation of how Triskel works you are referred to our publications. This page provides additional experimental results that we had to omit in our papers because of space restrictions. You can also download a reference implementation of Triskel that is based on the well-known tool Weka that was developed at the University of Waikato.

Related Publications

Experimental Results

We evaluated the Triskel algorithm and compared it with AdaBoost, using SMO as the base classifier. Tests were conducted using the Weka framework. We used several multiclass datasets from the well known UCI repository.

Reference Implemenation

You can obtain our reference implementation of Triskel in two ways

and place them in the correct location in your existing Weka installation. Depending on your version of Weka you will have to modify the GenericObjectEditor.props file located in the gui folder of Weka. The lines you have to add for Triskel are:

weka.classifiers.ucd.ensemble.Triskel, \
weka.classifiers.ucd.ensemble.TriskelA, \
weka.classifiers.ucd.ensemble.DivideAndConquer, \
weka.classifiers.ucd.ensemble.RandomClassifier, \
weka.classifiers.ucd.biased.UndersamplingBiasedClassifier, \
weka.classifiers.ucd.biased.OverweightingBiasedClassifier, \
weka.classifiers.ucd.biased.PrecisionThresholdSelector, \
weka.classifiers.ucd.biased.BiasSelector, \
weka.classifiers.ucd.biased.ErrorBiasedClassifier, \
weka.classifiers.ucd.biased.RandomEnsemble, \
weka.classifiers.ucd.biased.RandomBiasedEnsemble, \
weka.classifiers.ucd.biased.Splitter, \
weka.classifiers.ucd.biased.CoverNegatives, \

Our reference implementation of Triskel is licensed under the terms of the GPL.

Undocumented Features

Many of the settings of the algorithms are explained in our papers. But, as you can see from the above list, there are more classifiers in the package than we have described in the paper, and some classifiers that are described have more options than we explained. Feel free to play around with the options, and if you find a setting that performs better than ours, let me know :-)


26 Mar 2007, Andreas Hess, andreas at idirlion dot de