Alternative Clustering

Alternative Clustering

Clustering is a common unsupervised exploratory pattern recognition approach that partitions data objects into n regions{C₁,C₂,...,C_n} on the basis of similarity measure (Jain & Dubes, 1988). Clustering algorithms attempt to generate a suitable bunching of the input data objects. Often you may find the clustering, which may not particularly useful and actionable and wishes to find an alternative of it. Look at a clustering problem of loan request applications to find out worse loan request applications but the clustering cannot draw a line between acceptable and non-acceptable loan request applications. We can discover various optimum alternative clusterings with good amount of accuracy. Because of high dimensional data space searching alternative clusterings is quite obvious (Davidson & Qi, 2008).

Figure 1.

Graphical representation of clusters

Generally, clustering algorithms are designed on the idea of grouping similar data points together and form clusters of similar features. Every criterion has its own merits and demerits, and given that, different clusterings are right for different intentions, so we cannot say any clustering is optimum for all applications (Hartigan, 1985). Traditional clustering methods present a single clustering or a single view of the dataset. In hard complicated applications, various interesting views of bunching data points can lie in the dataset. Consequently, user asks for alternative descriptions of data to get various views. There are various methods evolved for generating the alternative cluster sets (Truong & Battiti, 2012). Alternative clustering is a variant of constrained clustering. Its main objective is to find all the possible groups from various perspectives of quality within the cluster and isolation with other groups (Saha & Mitra, 2014).