A Hierarchical Projection Pursuit Clustering Algorithm
Alexei D. Miasnikov, Jayson E. Rome, Robert M. Haralick · 2004
We define a cluster to be characterized by regions of high density separated by regions that are sparse. By observ-ing the downward closure property of density, the search for interesting structure in a high dimensional space can be reduced to a search for structure in lower dimensional subspaces. We present a Hierarchical Projection Pursuit Clustering (HPPC) algorithm that repeatedly bi-partitions the dataset based on the discovered properties of interest-ing 1-dimensional projections. We describe a projection search procedure and a projection pursuit index function based on Cho, Haralick and Yi’s improvement of the Kittler and Illingworth optimal threshold technique. The output of the algorithm is a decision tree whose nodes store a pro-jection and threshold and whose leaves represent the clus-ters (classes). Experiments with various real and synthetic datasets show the effectiveness of the approach. 1.