DISTANCE CLASSIFIERS Clause Samples

DISTANCE CLASSIFIERS. Distance classifiers, which use models of data positions in Euclidean space and match based on distance to these positions, are useful for two reasons. Firstly, the process of creating models of the data enables clustering and position of class objects to be observed. Likely success or failure of the classifier can be gleaned from this model, as poor clustering of object types or poor classification boundary creation problems can be observed at an early stage. In the case of poor clustering, the data collection method or measurement type may be changed to obtain a better model. Secondly, there exists many ways of separating and clustering data based on this type of approach, therefore, even data which is poorly clustered my still be used in some instances, as long as a powerful decision boundary technique (for example, support vector machines or linear discriminate analysis) is used. In this instance, the model used was based on centroid positions for each class, constructed from class measurements of width and height. The use of centroids simplified the model and clearly showed the clustering and inter­class separation. Decision boundaries were then created using both Euclidean based and Gaussian fitting techniques. Other techniques for clustering class measurement data exist, for example k­means clustering. This technique creates clusters from data by altering inter­ and intra­class distances until they are optimal. This technique was not used in this case as the class types and juxtapositions are known and should be preserved in order to function as a useful model.