Multi-label Data-Stream Framework

This page contains software combining the MEKA and MOA frameworks for carrying out multi-label experiments on data streams. It currently relies heavily on the MEKA framework for classifiers.

UPDATE Feb 22, 2016 In particular, the code relates to a paper Read, Bifet, Pfahringer, Holmes - Scalable and Efficient Multi-label Classification for Evolving Data Streams. The original code was put together at a time when MOA and MEKA began to evolve quickly and they were not integrated early enough for this package to survive intact, so the original experiments (from the paper) became difficult to replicate. Several people contacted us regarding this issue. Finally we have finished the re-integration!

Instructions (Updated Feb 25, 2016)

Following the next releases of MOA and MEKA, this code will be integrated in the official MOA repository. In the meantime, to get the code running:

Update (Feb 25, 2016): Now the master branch https://github.com/Waikato/moa is up to date (regarding step 1.), but the extra files (meka-1.9.1-SNAPSHOT.jar, test.sh) need to be obtained from the multilabel branch.

1. Checkout the multilabel branch fork of MOA. (Note: this will be merged into the main moa repository for the next release of MOA).
2. Change the version tag in the moa/pom.xml to read <version>2012.09-SNAPSHOT</version> -- otherwise it may not compile (fixed now on the multilabel branch)
3. Compile, either using your favourite IDE, or: mvn -DskipTests -Dmaven.javadoc.skip=true package
4. (Optional) Edit the header of the script test.sh with any options you want to change
5. Download a dataset (A-TMC7-REDU-X2-500.arff is the default one used in the script).
6. Run the script ./test.sh 0 (option 0 if you want to test the code from the paper).

Notes

• This version has been built almost from scratch after years of changes to both MOA and MEKA (and even WEKA). Results may not match exactly those from before but will be significantly similar enough to use as a benchmark.
• The version of MEKA downloaded by Maven is not recent enough, which is why meka-1.9.1-SNAPSHOT.jar is needed (for now).
• Evaluation metrics only include exact match and hamming score at the moment.
• We are using weka.classifiers.moa.HoeffdingTree at the moment as a base classifier to MEKA classifiers.
• Error/Warning messages are visible on the command line output -- these occur with the initial instance/s in the stream that must be classified prior to building a model

OLD INSTRUCTIONS

UPDATE (March 12, 2012): Updloaded a first complete edition: moa+.tar.bz2

Requirements

• weka.jar: WEKA
• meka.jar: MEKA: Multilabel Extension for WEKA
• mice.jar: MEKA's Incremental Classification Extension
• moa-ml.jar: MOA with Multi-Label

RECOMMENDED: download everything (jars and moa-ml source files) moa+.tar.bz2

If you have any trouble, contact: jesse AT tsc dot uc3m dot es, or abifet AT cs dot waikato dot ac dot nz

Numerous examples of how to generate data and run classifiers are included in the bash script (.sh) files in moa+.tar.bz2.