moa-bico-experiment

This is an extension for the MOA framework which provides an experiment environment for the BICO algorithm. In order to use this extension, you must add the moa-bico-experiment.jar to the classpath when launching MOA.

Example:

$ java -cp moa-bico-experiment.jar:moa.jar -javaagent:sizeofag.jar moa.gui.GUI

Experiment with the MOA framework

For a clustering experiment with the MOA framework you should use the following settings.

Stream: class moa.streams.clustering.SimpleCSVStream

• csvFile: path to input file
• splitChar: input CSV split character (optional, typical ,)
• classIndex: true if the last component of an input point is a class label (optional, typical false)
• decayHorizon: number of input points
• decayThreshold: (is not needed)
• evaluationFrequency: number of input points

Algorithm: class moa.clusterers.bico.BICO

• Cluster: number of desired centers
• Dimensions: dimension of an input point
• MaxClusterFeatures: coreset size (typical Cluster * 200)
• Projections: number of random projections used for nearest neighbor search in first level (typical Dimensions)
• evaluateMicroClustering: true if the coreset should be the result (optional, typical false)

Run the MOA framework from the command line

$ java -cp moa-bico-experiment.jar:moa.jar -javaagent:sizeofag.jar moa.DoTask EvaluateClustering -s \(moa.streams.clustering.SimpleCSVStream -f csvFile -s splitChar -c classIndex -h decayHorizon -e evaluationFrequency\) -l \(moa.clusterers.bico.BICO -k Cluster -d Dimensions -n MaxClusterFeatures -p Projections -M evaluateMicroClustering\) -i instanceLimit -d dumpFile

Arguments

• csvFile: path to input file
• splitChar: input CSV split character (optional, typical ,)
• classIndex: true if the last component of an input point is a class label (optional, typical false)
• decayHorizon: number of input points
• evaluationFrequency: number of input points
• Cluster: number of desired centers
• Dimensions: dimension of an input point
• MaxClusterFeatures: coreset size (typical Cluster * 200)
• Projections: number of random projections used for nearest neighbor search in first level (typical Dimensions)
• evaluateMicroClustering: true if the coreset should be the result (optional, typical false)
• instanceLimit: number of input points
• dumpFile: path to summary file (comma-separated)

Stand-alone experiment

This extension also provides three different classes to launch the BICO algorithm without the MOA framework.

"Experiments"

$ java -cp moa-bico-experiment.jar:moa.jar -javaagent:sizeofag.jar moa.clusterers.bico.experiment.Experiments input n k d space output projections [seed]

Arguments

• input: path to input file (space-separated)
• n: number of input points
• k: number of desired centers
• d: dimension of an input point
• space: coreset size (typical k * 200)
• output: path to output file (space-separated)
• projections: number of random projections used for nearest neighbor search in first level (typical d)
• seed: random seed (optional)

"Quickstart"

$ java -cp moa-bico-experiment.jar:moa.jar -javaagent:sizeofag.jar moa.clusterers.bico.experiment.Quickstart input n k d space output projections [splitchar [seed]]

Arguments

• input: path to input file
• n: number of input points
• k: number of desired centers
• d: dimension of an input point
• space: coreset size (typical k * 200)
• output: path to output file (space-separated)
• projections: number of random projections used for nearest neighbor search in first level (typical d)
• splitchar: input CSV split character (optional, typical ,)
• seed: random seed (optional)

"Script"

$ java -cp moa-bico-experiment.jar:moa.jar -javaagent:sizeofag.jar moa.clusterers.bico.experiment.Script input k [r]

With this class the BICO algorithm evaluates all points of the input file, uses k * 200 as coreset size and the dimension of the input points as number of random projections used for nearest neighbor search in first level.

Arguments

• input: path to input file (comma-separated)
• k: number of desired centers
• r: number of runs (optional, typical 1)

Reference

Hendrik Fichtenberger, Marc Gillé, Melanie Schmidt, Chris Schwiegelshohn, Christian Sohler: BICO: BIRCH Meets Coresets for k-Means Clustering. ESA 2013: 481-492 (2013) http://ls2-www.cs.tu-dortmund.de/bico/

License

See License