No abstract

A self-organizing tree (S-TREE) has a self-organizing capability and better performance than previously reported tree-structured clustering. In the S-TREE algorithm, since a tree grows in greedy fashion, a pruning mechanism is necessary to reduce the effect of bad leaf nodes. Extra nodes are pruned when the tree reaches a predetermined maximum size (U ). U is problem-dependent and is therefore difficult to specify beforehand. Furthermore, since U gives the limit of tree growth and also prevents self-organizing of the tree, it may produce unnatural clustering. We are presenting a new pruning algorithm without U . In this paper, we present results showing the performance of the new pruning algorithm using samples generated from normal distributions. The results of computational experiments showed that the new pruning algorithm works well for clustering of those samples.

Self-organizing tree (S-TREE) models solve clustering problems by imposing tree-structured constraints on the solution. It has a self-organizing capacity and has better performance than previous tree-structured algorithms. S-TREE carries out pruning to reduce the effect of bad leaf nodes when the tree reaches a predetermined maximum size (U ), However, it is difficult to determine U beforehand because it is problemdependent. U gives the limit of tree growth and can also prevent self-organization of the tree. It may produce an unnatural clustering. In this paper, we propose an algorithm for pruning algorithm that does not require U . This algorithm prunes extra nodes based on a significant level of cluster validity and allows the S-TREE to grow by a self-organization. The performance of the new algorithm was examined by experiments on vector quantization. The results of experiments show that natural leaf nodes are formed by this algorithm without setting the limit for the growth of the S-TREE.