Mining unknown patterns in data when the features are correlated

Abstract

In this paper, a previously introduced data mining technique, utilizing the Mean Field Bayesian Data Reduction Algorithm (BDRA), is extended for use in finding unknown data clusters in a fused multidimensional feature space. In extending the BDRA for this application its built-in dimensionality reduction aspects are exploited for isolating and automatically mining all points contained in each unknown cluster. In previous work, this approach was shown to have comparable performance to the classifier that knows all cluster information when mining up to two features containing multiple unknown clusters. However, unlike results shown in previous work based on lower dimensional feature spaces, the results in this paper are based on utilizing up to twenty fused features. This is due to improvements in the training algorithm that now mines for candidate data clusters by processing all points in a quantized cell simultaneously. This is opposed to the previous method that processed all points sequentially. This improvement in processing has resulted in a substantial reduction in the run time of the algorithm. Finally, performance is illustrated and compared with simulated data containing multiple clusters, and where the relevant feature space contains both correlated and uncorrelated classification information.