kamila: Clustering Mixed-Type Data in R and Hadoop

Abstract

In this paper we discuss the challenge of equitably combining continuous (quantitative) and categorical (qualitative) variables for the purpose of cluster analysis. Existing techniques require strong parametric assumptions, or difficult-to-specify tuning parameters. We describe the kamila package, which includes a weighted k-means approach to clustering mixed-type data, a method for estimating weights for mixed-type data (Modha-Spangler weighting), and an additional semiparametric method recently proposed in the literature (KAMILA). We include a discussion of strategies for estimating the number of clusters in the data, and describe the implementation of one such method in the current R package. Background and usage of these clustering methods are presented. We then show how the KAMILA algorithm can be adapted to a map-reduce framework, and implement the resulting algorithm using Hadoop for clustering very large mixed-type data sets.

Distance Metrics and Clustering Methods for Mixed‐type Data

“…For most data sets with multiple nominal variables, this inevitably leads to small sample sizes within each categorical cell. Consider two typical mixed‐type data sets analysed in Foss & Markatou (): the first, a biomedical data set contains five nominal variables measured on 475 patients, while the second contains five nominal variables measured on about 80 million domestic airline flights in the USA. The distribution of counts within the combinatorial cells for each data set is shown in Figure ; in the biomedical data set, the median number of observations per cell is two, and even in the much larger airline data set, 25% of the cells have count less than 16.…”

“…If an inadequate sample size is suspected, KAMILA incorporates a categorical smoother that can ameliorate these issues in most circumstances. The KAMILA method has been implemented in the R package kamila, as well as in Hadoop, with usage recommendations described in Foss & Markatou ().…”

Distance‐based clustering of mixed data

“…Although Modha–Spangler clustering accounts for variable significance within the algorithm, it is vulnerable to individual noninformative variables, due to the fact that the single weight does not allow individual variables to be up‐ or downweighted (Foss, Markatou, Ray, & Heching, ). The Modha–Spangler algorithm is implemented in R package kamila (Foss & Markatou, ).…”

Investigation on the Clusterability of Heterogeneous Dataset by Retaining the Scale of Variables

“…Clustering heterogenous dataset is a challenging process. The outcome of the analysis gives a significant impact on the interpretation of clusters [1,2,3,4]. Moreover, it demanded excessive computational skills and memory storage due to incorporation of broad categories [5].…”

“…The most common approached in treating heterogeneous data is through converting the variables into a single scale of measurement. However, this method may result in information loss [6,7,4]. Meanwhile, conducting a separate cluster analysis can abandon the connection between the variables which can be inappropriate.…”