The self-organizing map of trees

In the standard version of the Self-Organizing Map, each neuron is associated with a vector. An extension using trees instead of vectors is presented. Compared to vectors, trees provide remarkably more degrees of freedom. The essential points of self-organization, the distance function and the learning rule, are adapted to trees by means of graph matching. In order to avoid exhaustive searching in tree matching an efficient heuristic is introduced. The results of the experiments are promising: the proposed methods apply elegantly in the process of self-organization.