Riffle: an R package for Nonmetric Clustering
Georey B. Matthews, Robin A. Matthews · 2006
We present here an R package for Riffle, a nonmetric clustering technique [2]. This is a algorithm for clustering (unsupervised learning) that does not rely on a similarity measure for multivariate data, and uses only nonparametric (order) statistics. It is suitable for mixed nominal, ordinal, etc. attributes, as are often found in environmental data analysis. The current implementation of Riffle in R has a number of improvements over the original implementation [2], utilizing a marginalized expectation maximization (EM) approach to speed up the search. This has advantages in avoiding local minima, as well. Also, a technique for creating useful seed clusterings has been developed (rather than completely random initial clusters, as in [2]), substantially speeding up the clustering and making the final cluster less susceptible to noise. Procedures are also provided to use the resulting clustering to find an optimal subset of the attributes, and to create a naive Bayes classifier. Riffle has been used successfully in a variety of clustering tasks, and we have found it to be a useful, intuitive technique for graduate and undergraduate students in environmental sciences. Although nonmetric clustering is over a decade old, it is not widely known. A recent authoritative survey [1],pp. 541-542, does not include a discussion of them, and laments, “How do we treat vectors