The Application of Hierarchical Clustering Algorithms for Recognition Using Biometrics of the Hand

In data analysis, the hierarchical clustering algorithms are powerful tools allowing to identify natural clusters, often without any priori information of the data structure, and are quite often used because provide a graphical representation of the resulting partitions, a hierarchy or dendrogram, revealing more information than non-hierarchical algorithms that returns a unique partition. Moreover, it is not necessary specify the number of clusters a priori. Cutting the dendrogram in different levels on the hierarchy produces different partitions and also, the use of different clusters aggregation methods for the same data set can produces different hierarchies and hence different partitions. So, several studies have been concerned with validate the resulting partitions comparing them, for instance, by the analysis of cohesion and separation of their clusters. The work presented here focuses on the problem of choosing the best partition in hierarchical clusterings. The procedure to search for the best partition is made in the nested set of partitions, defined by the hierarchy. In traditional approaches each partition is defined by horizontal lines cutting the dendrogram at a determined level. Was proposed an improved method, SEP/COP, to obtain the best partition, based on a wide set of partitions. In this paper we discuss these two types of approaches and we do a comparative study using a set of experiments using two-dimensional synthetics and real- world data sets, based on the biometrics of the hands. This database is provided from Bosphorus Hand Database, in the context of recognition of the identity of a person by using the features of her hand/biometrics. In the results of the experiments, the SEP/COP showed to be a better partition algorithm in some situations namely regarding to real data, leading to a contribution to identification systems based on the biometrics of the hands shape.