ILPnet2 Toolbox

 System ILPnet2 Toolbox Pointers pena@iws.cs.uni-magdeburg.de Stefan Wrobel

Dear ILPnet member,

You can put this package to several uses, and we can certainly recommend it as a method to allow students to quickly get a brief hands-on impression of an algorithm in a graphical environment:

• its primary use in our view is teaching. We are using the platform as parts of the hands-on exercises for students in our classes. We ask the students to work their way through the step-by-step introduction to Kepler, and then ask them to follow the step-by-step instructions for playing with Progol. Both of these are very well tested out, and our students are reliably able to carry them out.
• for Midos, there is not yet a step-by-step introduction, but there is a section in chapter 4 of the relational data mining (Dzeroski/Lavrac) book that you might find useful.
• you can of course in the same fashion use Kepler to allow people to quickly play with other included algorithms, such as decision trees,... Some of them are already included in the basic Kepler step-by-step section in the manual.

For the ILP practitioner and/or researcher, Kepler offers potential but also some limitations which means you should have a look yourself whether you find it useful for the following purposes:

• we have found Kepler useful for preparing ILP data. You might want to have a look in particular at the "Power Query" operator which graphically allows complex conjunctive queries to be formulated, including construction of new attributes and sampling. There is also a "Prolog Rules" operator which allows new relations to be defined based on a Prolog rule, but there is currently no documentation available on this operator. Once you have prepared interesting data in Kepler, you can always export them in a simple Prolog-like format to keep working with the data outside of Kepler (see also next point). See the Kepler manual on how to do that.
• finally, Kepler of course can be used for actually analyzing real world data, and we find that useful for the basic and well tested algorithms such as C4.5 and Midos. As for Progol, we have thoroughly tested the step-by-step guide, but for real down and dirty work with such an ILP algorithm have found that it is probably best to work directly with the algorithm on the command line, both due to the unavoidable bugs which are still present, and the additional complexity of preparing complex input with a graphical tool. From a machine learning point of view, Kepler is also lacking the required testing facilities (for example, you cannot really do accuracy testing or crossvalidation unless offered by the plug-in tool itself).
• as a long-term goal, we are planning to make further ILP tools available as plug-ins to this platform, and to offer step by step guides as we currently have it for Progol. In principle, everybody can develop the necessary wrappers to make their algorithm available as a plug-in in Kepler, and there is a guide (written by Mathias Kirsten a while ago) on how to do that. Since this has turned out more difficult to use than anticipated, we do not recommend trying this until a better version of the guide is available. However, this will require significant additional manpower, so it may not actually materialize.

In sum, we hope you find this package useful for teaching and perhaps also for experimenting yourself!

(Research group Knowledge discovery and Machine learning, Otto-von-Guericke-Univ. Magdeburg)

P.S. A word of caution. We are providing this package as a service to the community, and have done our best to test the step-by-step guide and make it work reliably. We will try to fix any error in the step by step guide and the accompanying data, but we cannot fix bugs in the basic Kepler platform. We are discussing with Fraunhofer AIS whether someone could be found for doing that, but we have not found a solution yet. P.P.S Questions about the ILP Toolbix should be addressed to Lourdes Pena at pena@iws.cs.uni-magdeburg.de

back to index