Hierarchical Clustering Method Using Probabilistic Similarity Measure

In this paper, we propose a new hierarchical clustering method, which is useful to find appropriate clusters of attributes from given dichotomous or frequency data. Important features of our method are 1) the similarity between two attributes is defined as a probability of their pattern vectors being observed under the hypothesis of independence, 2) for each generated cluster, one pattern vector is defined in a natural manner, and 3) it can be used freely without distinguishing the frequency data from the dichotomous one.

A typical frequency data is analyzed to illustrate how our method works effectively. The discussion on similarities among objects is also included to propose a new similarity measure based on our clustering method.