Projective ART for clustering data sets in high dimensional spaces.

A new neural network architecture (PART) and the resulting algorithm are proposed to find projected clusters for data sets in high dimensional spaces. The architecture is based on the well known ART developed by Carpenter and Grossberg, and a major modification (selective output signaling) is provided in order to deal with the inherent sparsity in the full space of the data points from many data-mining applications. This selective output signaling mechanism allows the signal generated in a node in the input layer to be transmitted to a node in the clustering layer only when the signal is similar to the top-down weight between the two nodes and, hence, PART focuses on dimensions where information can be found. Illustrative examples are provided, simulations on high dimensional synthetic data and comparisons with Fuzzy ART module and PROCLUS are also reported.