A Simple Clustering Method for Binary Data based on a Binary Associative Memory

This paper studies clustering methods for binary data based on nonlinear dynamics in a binary associative memory (BAM) characterized by ternary connection parameters and signum activation function. First, as a set of binary data is given, we select several the center candidates. Applying a simple learning rule to the candidates, we obtain a BAM having multiple fixed points. Second, each datum is applied as an initial point and basin of attraction to a fixed point gives a cluster of the datum. Third, the clustering is evaluated as compared with desired distribution on the clusters. Repeating these three steps, the algorithm explores better clusters. Applying the algorithm to typical examples, the algorithm efficiency is confirmed.