WhiBo extension available for RapidMiner 5

[Milan Vukicevic]

We announce the release of WhiBo extension for RM5 as presented at the RCOMM 2010.

WhiBo is a platform for White Box (component based) machine learning algorithm design. It is intended to be used by data mining practitioners, research scientists and algorithm developers. The main idea of WhiBo is to offer standardized components for algorithm design which will enable simple design and performance testing, easy extension of the component repository and creation of new generic algorithms.

Currently stable version of Generic Decision tree is provided (with CV operator with log and 5x2CV F test for validation and performance comparison) with detailed documentation.
Generic clustering algorithm is also provided but, stable version and documentation will be released in next few weeks.

Plug-in, user and developer documentation can be found on www.whibo.fon.bg.ac.rs.
We would be thankful for any suggestions, ideas and bug reports (on this RapidMiner forum, our WhiBo forum or by email) before we announce a final version.
 
Evolutionary algorithm for automatic design of decision tree algorithms is designed and evaluated (we got really promising results, submitted to IDA journal) as promised on RCOMM 2010. Operators for automatic algorithm design will be published with next version of WhiBo.

Cheers
Milan

[Ingo Mierswa]

Hi Milan,

those are great news! I am happy that you moved to RapidMiner 5 and will certainly try your extension as soon as possible.

By the way: if you like, we could also distribute your extension with the update mechanism known from other extensions - this way it is much more likely that people get aware of the WhiBo extension and use it.

Cheers,
Ingo

[Milan Vukicevic]

Hi Ingo,

It would be great to distribute WhiBo with the update mechanism. I'm sending you a final version of WhiBo next week, I just have to test some more generic clustering algorithm and write some more manuals and to attach example streams. 

Cheers
Milan

[Ingo Mierswa]

Hi Milan,

great! Of course I am in particular keen on

Evolutionary algorithm for automatic design of decision tree algorithms is designed and evaluated (we got really promising results, submitted to IDA journal) as promised on RCOMM 2010. Operators for automatic algorithm design will be published with next version of WhiBo.

since this is quite connected to my masters thesis where I did something very similar for the preprocessing (feature extraction) from time series data. Here is the MLJ article about this if you are interested:

http://www-ai.cs.uni-dortmund.de/auto?self=$Publication_e3swwc81

Anyway, I am looking forward to distributing this great extension via the update server. Please contact us when you are ready.

Cheers,
Ingo

[awchisholm]

Hello Ingo,

Do you have any news on the various value series capabilities mentioned here?

http://rapid-i.com/rapidforum/index.php/topic,1608.msg6275.html#msg627

regards,

Andrew

[Ingo Mierswa]

Hi Andrew,

no, sorry. Due to other development priorities it is not very likely that we will deliver the automatic feature extraction based on genetic programming any time soon. This might change if one of our students find the time to adapt the classes for RM 5 or if somebody out of the community is willing to provide a new version.

Cheers,
Ingo

[awchisholm]

Hello Ingo,

No problem, I'll pass the hat round ;-)


Andrew

[Milan Vukicevic]

since this is quite connected to my masters thesis where I did something very similar for the preprocessing (feature extraction) from time series data. Here is the MLJ article about this if you are interested:

http://www-ai.cs.uni-dortmund.de/auto?self=$Publication_e3swwc81

Hi Ingo,

that's really great. I will read your paper carefully, from what I saw till now it is the same approach for feature extraction, as we did for algorithm design. I will think about possible integration of these two (maybe some other preprocesing steps as well). These are highly complementary things and could lead to further development of automatic DM workflows. Maybe we can think also about some RM implementation and some joint publication.

Cheers
Milan

[Ingo Mierswa]

Hi Milan,

yes, I also believe that both topics are quite connected and share lots of common properties. I would love to see the idea extended to complete DM processes but I doubt that genetic programming would be the optimal choice for that. Actually, supporting the user in creating complete processes is one of the major goals of our European project e-LICO and I hope that there will be more great results during the next 18 months. I am always in for a publication on those topics although working with me on such a publication can be a test of your patience 

Cheers,
Ingo