Novelty detection for short time series with neural networks
Andriano L. I. Oliveira, Fernando Buarque de Lima Neto, Sílvio R. L. Meira · Hybrid Intelligent Systems · 2003
Novelty detection is an important problem in many application fields, such as scene analysis, machine failure detection, and auditing. There has been an increasing interest in time series novelty detection with novel techniques recently developed. An approach to this problem uses time series forecasting by neural networks. However, time series forecasting is a difficult problem, thus, the use of this technique for time series novelty detection is sometimes criticized. Moreover, the short length of the time series available in several important problems makes forecasting an even harder problem. This is the case of some important auditing problems such as accountancy auditing and payroll auditing. In this work we propose a classification-based approach for short time series novelty detection. The idea is to increase the number of patterns of data sets by adding both normal and novelty random patterns. The proposed approach has been evaluated on four real world time series and has shown promising results.