Iterative Information Granulation for Novelty Detection in Complex Datasets

Recognition memory in a number of mammals is usually utilised to identify novel objects that violate model predictions. In humans in particular, the recognition of novel objects is foremost associated to their ability to group objects that are highly compatible/similar. Granular computing not only mimics the human cognition to draw objects together but also mimics the ability to capture associated properties by similarity, proximity or functionality. In this paper, an iterative information granulation approach is presented, for the problem of novelty detection in complex data. Two granular compatibility measures are used, based on principles of Granular Computing, namely the multidimensional distance between the granules, as well as the granular density and volume. A two-stage iterative information granulation is proposed in this work. In the first stage, a predefined number of granular detectors are constructed. The granular detectors capture the relationships (rules) between the input-output data and then use this information in a second granulation stage in order to discriminate new samples as novel. The proposed iterative information granulation approach for novelty detection is then applied to three different benchmark problems in pattern recognition demonstrating very good performance.