A biologically inspired neural clustering model for capturing patterns from incomplete data

Data in the real world is seldom complete. Occlusions or temporally unavailable sensors often lead to situations where incomplete data is presented for analysis. Approaches to handle incomplete data have been proposed using neural networks such as fuzzy ARTMAP and back propagation. In this paper we propose a novel approach extending the unsupervised neural network based clustering technique called the Growing Self Organizing Map (GSOM) to address the problem of missing input information. The GSOM has been extensively used for clustering and classification of large datasets, especially in the areas of text mining and bioinformatics. It is mainly used as a data visualization tool since it maps high dimensional input data into a two dimensional output space. The proposed model is biologically inspired and uses hierarchically organized GSOMs incorporated with Bayesian networks to handle missing input values. We demonstrate how missing information is predicted at different levels of abstraction through combining the known information about the input and the previous knowledge about similar inputs, in the manner a human would make inferences about unknown data.