treeClust: An R Package For Tree-Based Clustering Dissimilarities

This paper describes treeClust, an R package that produces dissimilarities useful for clustering. These dissimilarities arise from a set of classification or regression trees, one with each variable in the data acting in turn as a the response, and all others as predictors. This use of trees produces dissimilarities that are insensitive to scaling, benefit from automatic variable selection, and appear to perform well. The software allows a number of options to be set, affecting the set of objects returned in the call; the user can also specify a clustering algorithm and, optionally, return only the clustering vector. The package can also generate a numeric data set whose inter-point distances relate to the treeClust ones; such a numeric data set can be much smaller than the vector of inter-point dissimilarities, a useful feature in big data sets.