Classification by Regression Operator vs. Polynomial By Binomial Classification

mbukombuko Member Posts: 10 Contributor II
edited November 2019 in Help
Hi,

I have a multi class approach by using SVM (mySVM) and a operator to enable the multi classification. The problem is, that the two possible operators "Classification by Regression Operator" and "Polynomial By Binomial Classification" lead to different kind of results for the confidence values:

1) Classification by Regression Operator: confidence value element of (-∞, 1]. This seems to be the signed distance to the hyperplane. Is this correct? Why are there no values higher than 1? (1 would mean that it is on the edge of the margin. Might it depend on the kernel function?)

2) Polynomial By Binomial Classification: confidence value element of [0,1]. Is this any kind of probability? Definition?

Unfortunately, I don't find any hint how the confidence values are defined (with regard to the used operators or SVM implementation).

In order to use RapidMiner and the output, I need a clear understanding of the parameters, confidence values and dependencies to the RapidMiner operators.

I would be pleased, if you would help me with these issues!

Best regards,
Mark

PS: I have already opened a similar thread in another category (https://rapid-i.com/rapidforum/index.php/topic,9418.msg31536.html)

Answers

  • MartinLiebigMartinLiebig Administrator, Moderator, Employee, RapidMiner Certified Analyst, RapidMiner Certified Expert, University Professor Posts: 3,503 RM Data Scientist
    confidences should usually be [0,1]. So the case in 1) seems odd to me. I never used classification by regression though.

    ~martin
    - Sr. Director Data Solutions, Altair RapidMiner -
    Dortmund, Germany
  • mbukombuko Member Posts: 10 Contributor II
    @Martin thank you for your reply!

    Unfortunately, a clear description and definition of the output and operators are missing, but I need it in order to work scientifically correct. I am looking forward to an official answer from RapidMiner, but at the moment there is an absence of any reaction.
  • MartinLiebigMartinLiebig Administrator, Moderator, Employee, RapidMiner Certified Analyst, RapidMiner Certified Expert, University Professor Posts: 3,503 RM Data Scientist
    Hi,

    i guess there will be more offical answer than my postings, except if you ask at support.rapidminer.com, but this requieres a licence. I am employed at RapidMiner; That makes it kind of offical?

    What you can do is have a look at the source. The classification by Regression operator can be found at:

    https://github.com/rapidminer/rapidminer-studio/blob/master/src/main/java/com/rapidminer/operator/learner/meta/ClassificationByRegression.java

    There we find the comment:

    /**
    * For a classified dataset (with possibly more than two classes) builds a classifier using a
    * regression method which is specified by the inner operator. For each class {@rapidminer.math i} a
    * regression model is trained after setting the label to {@rapidminer.math +1} if the label equals
    * {@rapidminer.math i} and to {@rapidminer.math -1} if it is not. Then the regression models are
    * combined into a classification model. In order to determine the prediction for an unlabeled
    * example, all models are applied and the class belonging to the regression model which predicts
    * the greatest value is chosen.
    *
    * @author Ingo Mierswa, Simon Fischer
    */
    I checked the attached example process. Apperently confidence(Mine) is [0,1] but confidence(Rock) not. For some reason only one confidence is set "correctly". I will check the code deeper once I have more time

    ~Martin
    - Sr. Director Data Solutions, Altair RapidMiner -
    Dortmund, Germany
  • mbukombuko Member Posts: 10 Contributor II
    @Martin, thank you for your help!

    The comment is also written in the documentation. I have the same problem with the different confidence outputs: only one is set "correctly" (but only with 1 ?).

    What is the definition of the confidence value if it is in [0,1]? (I wanted to avoid looking into the implementations, but it seems to be necessary)
  • MartinLiebigMartinLiebig Administrator, Moderator, Employee, RapidMiner Certified Analyst, RapidMiner Certified Expert, University Professor Posts: 3,503 RM Data Scientist
    usually this is a measure on how much the algorithm trusts it's own calculation. Usually the higher the value the more likely is the algorithm to be right.

    The value is calculated for every algorithm differently. For a k-nn it is the fraction of neighbours with this class (in the unweighted case). For the SVM the value depends on the distance to the seperating hyperplane.
    - Sr. Director Data Solutions, Altair RapidMiner -
    Dortmund, Germany
  • mbukombuko Member Posts: 10 Contributor II
    "For the SVM the value depends on the distance to the seperating hyperplane"
    - usually yes, but it also depends on the SVM implementation (mySVM vs. libSVM) and the wrapping operators like the two mentioned in the thread's title.
    Why are the confidence values for the same SVM algorithm but different operators within a different range (which is not documented)?
    In the case of libSVM there is the possibility to estimate the probability as confidence instead of the distance to the hyperplane. Unfortunately, the specific meanings of the confidence values regarding the algorithms etc. are not documented.

    I think I have to look into the implementions since the academic support does not help either.

    Thanks so far, I would be very pleased about any further support.
  • Muhammed_Fatih_Muhammed_Fatih_ Member Posts: 93 Maven
    Hello together, 
    Hello @mschmitz

    does anybody know on which scientific base the ClassficationbyRegression-operator was implemented? Is there a paper which describes how this ensemble approach is working in detail despite the above mentioned git-code? 

    Thank you in advance for your answer! 

    Best regards! 
  • MartinLiebigMartinLiebig Administrator, Moderator, Employee, RapidMiner Certified Analyst, RapidMiner Certified Expert, University Professor Posts: 3,503 RM Data Scientist
    sorry, I don't have any reference, but it seems to be just a common "trick" which is used all over the place?

    Best,
    Martin
    - Sr. Director Data Solutions, Altair RapidMiner -
    Dortmund, Germany
  • Muhammed_Fatih_Muhammed_Fatih_ Member Posts: 93 Maven
    Hi @mschmitz

    many thanks for the quick response!

    @IngoRM Could you maybe help with a reference? 

    Best regards!
Sign In or Register to comment.