Projective ART With Buffer For Clustering In High Dimensional Spaces And An Application

With the development of Computer Science and Information Revolution, large volumes of data are eventually coming out. Data mining is one developing interdisciplinary technology when people are facing with such data. Data mining, also known as knowledge-discovery in databases ( KDD ), is the practice of automatically searching large stores of data for patterns. As one important stage of the process of data mining, data clustering is an unsupervised process of classifying patterns into groups ( Clusters ), aiming at discovering structures hidden in a data set.However, Unlike to traditional clustering algorithms such as hierarchical and partitional algorithm always fail to deal with very large databases, an improved neural network architecture, Projective ART ( PART ), is developed to cluster subsets of data sets in high dimensional spaces. But the success of PART algorithm depends on both the accurate parameters and the satisfied order of input data sets. These disadvantages prevent PART from applied to realtime databases.In this paper, we propose a modification to PART that introduces the buffer management as well as the average similar degree which is provided to successfully work with high-similar noise data sets and partly achieve order-independent objective without the correct parameters. The buffer management mechanism allows the data sets not to immediately cluster to one group. And the average similar degree has a good attribute of parameter-tolerance. Namely, the clustering result doesn't depend on the precise choice of the input parameter, and different parameter values have the exact same clustering result including the dimensions associated with clusters.Also, simulations on high dimensional synthetic data and comparisons are reported and an application by using our algorithm to find stock concurrence association rules is given finally.