Binary tree design using fuzzy isodata

Abstract

A procedure for designing a fuzzy binary decision tree using unlabeled samples is developed. At each node, the data is split into two dissimilar groups using the fuzzy ISODATA algorithm. All the available features are used while clustering. Then, the best feature among these available features is selected based on some separation index. This repeated in a depth-first manner till further dissection of the data at any node is impossible to a predefined measure of separability of the clusters. The decision tree so designed can be used for the classification of future sample patterns. The procedure indicated gives an initial tree configuration amenable for further optimization. Another useful feature of this method is that it can be very effectively used to find the number of clusters underlying the data. This is achieved without resorting to the normal procedure of repeating the clustering procedure by varying the number of clusters and then selecting that configuration which has the optimum partition index. An example from taxonomy is given to illustrate the method.