On improving ROCK-based clustering for categorical data: student research abstract

In the field of data mining, the analysis of categorical data (i.e., data that can assume a limited number of values, such as names) is particularly challenging due to the lack of implicit geometrical properties. Clustering of categorical data is becoming increasingly important, since non-numerical data are ubiquitous and clustering can be used, for example, to optimize an anonymization process [1], or to perform anomaly detection, or in any application where there is the need to automatically recognize the intrinsic structure of data. Various algorithms have been proposed for clustering this kind of data (see [3]), such as ROCK (RObust Clustering using linKs) [4].