Novelty Detection in Time Series Through Self-Organizing Networks: An Empirical Evaluation of Two Different Paradigms

This paper addresses the issue of novelty or anomaly detection in time series data. The problem may be interpreted as a spatio-temporal classification procedure where current time series observation is labeled as normal or novel/abnormal according to a decision rule. In this work, the construction of the decision rules is formulated by means of two different self-organizing neural network (SONN) paradigms: one builds decision thresholds from quantization errors and the other one from prediction errors. Simulations with synthetic and real-world data show the feasibility of the two approaches.