[SOLVED] KMeans and Nominal Measures

annbraannbra Member Posts: 3 Contributor I
edited July 2019 in Help
Hi,

I am new in the forum and I have a question about kMeans.
I have read in the forum that it is not possible to use polynominal data for kMeans algorithm. But in RapidMiner 5.2.008, it is possible to choose nominal measures (e.g. nominal distance) and to determine clusters with centroids of polynominal data. There is only a warning that it is not possible to use polynominal data by kMeans, but it is possible because I get "good" results.
How does RapidMiner calculate the clusters? Is it possible to have a look on the different steps during the calculation?

Greetings,
Anne
Tagged:

Answers

  • MariusHelfMariusHelf RapidMiner Certified Expert, Member Posts: 1,869 Unicorn
    Hi Anne,

    it is not possible to see intermediate steps of the algorithm. But yes, k-Means is capable of handling nominal attributes with the nominal or mixed measures. NominalDistance e.g. is 0 if two strings match exactly, and 1 otherwise. The issue about the warning (which is obviously wrong) has just been fixed and will disappear in the next release.

    Best regards,
    Marius
  • annbraannbra Member Posts: 3 Contributor I
    Hi Marius,

    thank you very much for your answer.
    The calculation of distance is now clear but how does the algorithm determine the centroid. A centroid of one cluster is a linear combination of the examples, and for this you need numeric data. Does the algorithm transfer the polynominal data in natural numbers and determine the centroid by "normal" summation?

    Greetings,
    Anne
  • MariusHelfMariusHelf RapidMiner Certified Expert, Member Posts: 1,869 Unicorn
    Actually it seems so. It may be better to use Nominal to Numerical with coding_type=dummy_coding before actually applying k-Means. That way you'll be on the safe side.
  • annbraannbra Member Posts: 3 Contributor I
    Thank you very much.
Sign In or Register to comment.