About the global seed in kmeans and kmedoids in RapidMiner

amyamy Member Posts: 16  Maven
edited November 2018 in Help
Hi there,
I have a quick question about k-means and k-medoids clustering in
RapidMiner. What does the global seed mean? Is that mean using a dynamic way
to pick different seeds during clustering? Thanks a lot.

Answers

  • amyamy Member Posts: 16  Maven
    Just got reply from Ingo
    k-Means and k-Medoids both take an initial random sample from the data as starting centroids for the clustering. This random selection is guided by a random number generator and this number generator again delivers a sequence of random numbers depending on the random seed. On the one hand, this ensures that you will get a sequence of random numbers, on the other hand for the same random seed number, you will always get exactly the same random numbers. Hence, your experiments can be reproduced as long as the same random number seeds are used.
    Hi Ingo,
    Thanks a million for your kind reply.
    I found a paper with the title The Global K-Means Clustering Algorithm http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.19.5033 Is it the algorithm you have implemented by using a global seed? I know that by setting the random seed to a given random number the same clustering result can be achieved. If we set it to -1 (a global seed), what will happen then?
  • landland RapidMiner Certified Analyst, RapidMiner Certified Expert, Member Posts: 2,525   Unicorn
    Hi,
    since the sequence of random numbers will be different, and hence another initialization will be used, you might gain a different result.
    We don't have this algorithm implemented, it's just an implementation of the usual and very basic K-Means algorithm.
    The local or global only refers to the sequence of used random numbers. You could either use a local generator, giving the operator always the same, deterministic sequence of random numbers, or you could use the global random seed, which numbers will depend on the previous operators.

    Greetings,
      Sebastian
Sign In or Register to comment.