The MEKA project provides an open source implementation of methods for multi-label classification. It is based on the WEKA Machine Learning Toolkit from the University of Waikato.

I’ve been involved with the project on a software engineering basis, rather than a researcher one. The project only offered command-line calls, so I added a simple interface for exploration (inspired by the WEKA Explorer) and experimentation (in some sense simpler but also more flexible than the WEKA Experimenter). Furthermore, I added a junit-based testing framework to ensure that the code is working properly and the algorithms don’t suddenly change (regression tests). In the process, I also changed the build system from ant to maven.





  1. zhao

    There are something I am puzzled with. For instance, there are five classes {0, 1, 2, 3, 4} in the three attributes in the file ‘solar_flare.arff’, which is provided by MEKA. Multi-target methods could not predict class {2,3,4} , which means there are only {0,1} in the predicted results. But {2,3,4} appearance when I modify the .arff data to obtain more {2, 3, 4}.

    Is there something wrong with me or some logical explanation in multi-target learning methods?

    1. fracpete (Post author)

      Please use the MEKA mailing list for questions regarding MEKA. See MEKA homepage for details.


Leave a Comment

Your email address will not be published. Required fields are marked *