Novelty Detection in Time Series Using Self-Organizing Neural Networks: A Comprehensive Evaluation

Abstract

In this survey paper, we report the results of a comprehensive study involving the application of dynamic self-organizing neural networks (SONNs) to the problem of novelty detection in time series data. The study is comprised of three main parts. In the first part, we aim at evaluating how the performances of nonrecurrent dynamic SONNs are influenced by the introduction of different short-term memory kernels, such as Gamma, Gamma II and Laguerre, in the network input. In the second part, we analyze the performances of recurrent dynamic SONNs with the goal of inferring if they possess or not any competitive advantage over nonrecurrent architectures for novelty detection in time series. Finally, in the third part of the study, we introduce an alternative approach for dynamic SONN-based novelty detection by revisiting the operator map framework introduced by Lampinen and Oja (Proceedings of the 6th Scandinavian conference on image analysis (SCIA’89), pp 120–127, 1989) and Lehtimaki et al. (Proceedings of the joint international conference on artificial neural networks and neural information processing (ICANN/ICONIP’2003), pp 622–629, 2003). This framework allows the design of dynamic SONNs whose neurons are regarded as adaptive local linear models. In this case, novel/abnormal patterns are detected based on the statistics of prediction errors of the local models. A comprehensive performance comparison involving several nonrecurrent and recurrent dynamic SONN architectures is carried out using both synthetic and real-world time series data.