Another Adaptive Approach to Novelty Detection in Time Series

This paper introduces a novel approach to novelty detection of every individual sample of data in a time series. The novelty detection is based on the knowledge learned by neural networks and the consistency of data with contemporary governing law. In particular, the relationship of prediction error with the adaptive weight increments by gradient decent is shown, as the modification of the recently introduced adaptive approach of novelty detection. Static and dynamic neural network models are shown on theoretical data as well as on a real ECG signal.