Short introduction to stream analysis using MOA
Massive Online Analysis
Martin Juřen
Fakulta informatiky Masarykova univerzita
Brno, 2011
Martin Juřen (Masarykova univerzita)
< * ► s -00.0 Brno, 2011 1/10
Content
O Data stream evaluation
Q Work with MOA
• Basic steps
• Data stream generators
• Data stream classifiers
• Data stream clustering
• Data stream algorithm evaluation
Q MOA Demonstration Q Bibliography
Martin Jufen (Masarykova univerzita)
MOA
Data stream evaluation
Requirements on algorithms
• Process an example at a time, inspect it only once
• Limited amount of memory and time q Predict at any time
Martin Jufen (Masarykova univerzita)
< * ► s -00.0 Brno, 2011 3/10
Work with MOA
Basic steps using MOA
Basic steps
Q Choose and configure data stream generator
Q Choose an configure an algorithm
O Choose an configure an evaluation method
□
Martin Jufen (Masarykova univerzita)
MOA
Work with MOA
Data stream generators
Generators
ArffFileStream Input from file
ConceptDriftStream It generates stream with concept drift
FilteredStream:AddNoiseFilter It generates stream with noise
AgrawalGenerator Based on text: Rakesh Agrawal, Tomasz Imielinski, Arun Swami: Database Mining: A Performance Perspective. IEEE Transactions on Knowledge and Data Engineering, 1993.
Some other generators And their concept drift variants
Martin Jufen (Masarykova univerzita)
< * ► s -00.0 Brno, 2011 5/10
Work with MOA
Data stream classifiers
Classifiers
Majority class most frequently observed class
Hoeffding tree (and variants) This algorithm stands on the fact that
small sample could be enough to find an optimal splitting attribute.
Naive Bayes
Decision Stump single-level decision trees
Martin Jufen (Masarykova univerzita)
< * ► s -00.0 Brno, 2011 6/10
Work with MOA
Data stream clustering
Clustering
CobWeb Not quite data stream algorithm
CluStream Temporal extensions of cluster feature vector,
micro-clusters. Storing in snapshots in pyramidal pattern.
ClusTree Parameter-free algorithm. Capable to detect concept drift.
Martin Jufen (Masarykova univerzita)
< * ► s -00.0 Brno, 2011 7/10
Evaluation
Work with MOA
Data stream algorithm evaluation
Holdout Periodically testing the model with one test set
Test-Then-Train The model is firstly tested by new data until then it is trained.
Martin Juřen (Masarykova univerzita)
< * ► s -00.0 Brno, 2011 8/10
Demonstration
MOA Demonstration
• It is a new software, there is a lot of bugs.
• Written in Java, Open source project
• It could be linked to WEKA
• Goal: running experiments, evaluating algorithms, algorithm comparison
Martin Juřen (Masarykova univerzita)
< * ► s -00.0 Brno, 2011 9/10
Bibliography
Bibliography
• MOAteam: moa.cs.waikato.ac.nz
• Biffet A., Kirkby R., Krannen P., Reutemann P.: Massive Online Analysis: Manual. Online at moa.cs.waikato.ac.nz, May 2011.
• Biffet A., Kirkby R.: Tutorial 1. Introduction to MOA: {Mjassive {Ojnline {Ajnalysis. Online at moa.cs.waikato.ac.nz, January 2011.
Martin Jufen (Masarykova univerzita)
► < * ► s -00.0 Brno, 2011 10/10