Evaluating Value Weighting Schemes in the Clustering of Categorical Data

The majority of the algorithms in the clustering literature utilize data sets with numerical values. Recently, new and scalable algorithms have been proposed to cluster data sets with categorical data, data whose inherent ordering is not obvious. However, these algorithms deem all data values present in the data sets as equally important. Thus, the resulting clusters may be influenced by values that appear almost exclusively and reflect non-natural groupings. In this paper, we present a set of weighting schemes that allow for an objective assignment of importance on the values of a data set. We use well established weighting schemes from information retrieval, web search and data clustering to assess the importance of whole attributes and individual values. To our knowledge, this is the first work that considers weights in the clustering of categorical data. We perform clustering in the presence of importance for the values within the LIMBO framework, a new and scalable algorithm to cluster categorical data. Our experiments were performed in a variety of domains, including data sets used before in clustering research and three data sets from large software systems. We report results as to which weighting schemes show merit in the decomposition of data sets.