Modelling of Clusters
A usual pattern recognition problem in chemistry is to determine whether or not a compound is of a given type. The assumption often made that all patterns belonging to the same class form a distinct cluster is often unrealistic. An asymmetric situation occurs if one class does not form a proper homogeneous group . In Figure 40 the patterns of class 1 form a rather compact cluster while class 2 is scattered throughout the whole pattern space. In a practical example, class 1 may correspond to “good” samples and class 2 to “bad” samples (equivalent to outliers whose features differ markedly from standard values). In such a case, it is useful to construct a model of class 1. The model is utilized either for the classification of unknown patterns or for the interpretation of the data structure. A confidence region can be constructed around the class model. Patterns outside this region are considered to belong to the “unstructured” class.
Unable to display preview. Download preview PDF.