Combining MLP and RBF Neural Networks for Novelty Detection in Short Time Series

Abstract

Novelty detection in time series is an important problem with application in different domains such as machine failure detection, fraud detection and auditing. In many problems, the occurrence of short length time series is a frequent characteristic. In previous works we have proposed a novelty detection approach for short time series that uses RBF neural networks to classify time series windows as normal or novelty. Additionally, both normal and novelty random patterns are added to training sets to improve classification performance. In this work we consider the use of MLP networks as classifiers. Next, we analyze (a) the impact of validation and training sets generation, and (b) of the training method. We have carried out a number of experiments using four real-world time series, whose results have shown that under a good selection of these alternatives, MLPs perform better than RBFs. Finally, we discuss the use of MLP and MLP/RBF committee machines in conjunction with our previous method. Experimental results shows that these committee classifiers outperform single MLP and RBF classifiers.