KNNLearner don´t use similarities

fjcuberosfjcuberos Member Posts: 18 Maven
edited November 2018 in Help
The new implementation in 4.2 of KNNLearner has changed and the use of SimilarityUtil.resolveSimilarityMeasure has been dropped. Now only the EuclideanDistance is defined as selectable.
¿how I can use my similarity measure in a KNNLeraner? ¿or anyone in the com.rapidminer.operator.similarity package?
Any idea will be accepted  ;)


Thanks in advance.

F.J. Cuberos
Tagged:

Answers

  • IngoRMIngoRM Administrator, Moderator, Employee, RapidMiner Certified Analyst, RapidMiner Certified Expert, Community Manager, RMResearcher, Member, University Professor Posts: 1,751 RM Founder
    Hi,

    yeah, we know. We completely re-implemented NaiveBayes and KNN since both were ridicously slow on larger data sets. During this re-implementation, we unfortunately had to remove the ID-based similarity measures from Michael Wurst (which were one of the main reasons why the KNN learner was slow) and started a new similarity measure hierarchy. Adding all other similarity measures again and also improving the clustering schemes by letting them use the new similarity measures is at the top of our todo list and will certainly be finished until the next release.

    So for now, there are only two workarounds:
    • stick to RM 4.1 until version 4.3 is published which will then contain all similarity measures.
    • add the KNN implementation from version 4.1 to version 4.2 yourself (code and in operators.xml) and recompile RapidMiner
    Sorry for the inconvenience this have caused but at the end we will have a much more efficient system (just as an info: the improvement for the KNN learners was factor 13 on average).

    Cheers,
    Ingo
  • fjcuberosfjcuberos Member Posts: 18 Maven
    OK. I´d thought the reason was a redesign when found the com.rapidminer.tools.math.container.GeometricDataCollection class used in the KNNLearner.
    I´ve just  added the 4.1 version to my plugin and I'll wait for the 4.3 version.

    Thanks again, Ingo.
    Regards.

    F.J. Cuberos
  • IngoRMIngoRM Administrator, Moderator, Employee, RapidMiner Certified Analyst, RapidMiner Certified Expert, Community Manager, RMResearcher, Member, University Professor Posts: 1,751 RM Founder
    Thanks for your understanding.

    Cheers,
    Ingo
Sign In or Register to comment.