Mclust: Order of input parameters affecting clustering results

user_3068797 0
Viewed 1309

I am using mclust to see various clusters in my data set using various numbers of input (X,Y,Z,R, and S in the script below):

e.g.

elements<-cbind(X,Y,Z,R,S)
dataclust<-Mclust(elements)

I just find out that the order of the input parameters matters and affect the results; in other words elements <- cbind(X,Y,Z,R,S) gives a different clusters than say elements-<cbind(Y,Z,X,R,S). My understanding is that all the input parameters have the same weight and importance in the clustering analysis. am I wrong or is it a bug?

I have seen that in R 2.15.3 and 2 other R versions.

Any comment on or explanation of the above is appreciated.

cluster-analysis r
4 Answers

Often, Gaussian Mixture Model clustering is initialized randomly, as it will only find a local maximum.

Don't expect it to return the same result all the time.

user_10603500

I notice this is a very old thread but I think it's still worth to post an (official) answer. The issue was addressed in mclust 5: Clustering, Classification and Density Estimation Using Gaussian Finite Mixture Models from the R Journal with suggested solutions ([https://journal.r-project.org/archive/2016/RJ-2016-021/RJ-2016-021.pdf][1]), Pages 305-307. In short, "A problem with the MBHAC approach may arise in the presence of coarse data, resulting from the discrete nature of the data or from continuous data that are rounded when measured. In this case, ties must be broken by choosing the pair of entities that will be merged. This is often done at random, but regardless of which method is adopted for breaking ties, this choice can have important consequences because it changes the clustering of the remaining observations. Moreover, the final EM solution may depend on the ordering of the variables."

user_153777920
Related