Neural networks forecasting and classification-based techniques for novelty detection in time series

AUTOR(ES)
DATA DE PUBLICAÇÃO

2004

RESUMO

Novelty detection can be defined as the identification of new or unknown data that a machine learning system is not aware during training. Novelty detection algorithms are designed to classify input patterns as normal or novelty. These algorithms are used in several areas such as computer vision, machine fault detection, network security and fraud detection. The behavior of many systems can be modeled by time series. Recently, the problem of detecting novelties in time series has received great attention, with a number of different techniques being proposed and studied, including techniques based on time series forecasting with neural networks and on classification of time series windows. Forecasting-based time series novelty detection has been criticized because of the not so good performance. Moreover, the small amount of data available in short time series makes forecasting an even harder problem. This is the case of some important auditing problems such as accountancy auditing and payroll auditing. Alternatively, a number of classification-based methods have been recently proposed for novelty detection in time series, including methods based on artificial immune system, wavelets and one-class support vector machines. This thesis proposes a number of neural networks methods for novelty detection in time series. The methods proposed here were developed aiming to detect more subtle novelties that can be of great concern in fraud detection in financial systems. The first method aims to enhance forecasting-based novelty detection by using robust confidence intervals. These intervals are used to overcome the main limitation of forecasting-based novelty detection, namely, the selection of a suitable threshold for detecting a novelty. The proposed method was applied to some real-world time series with good results. Two methods based on classification are also proposed in this thesis. The first method is based on negative samples whereas the second method is based on RBF-DDA neural networks and does not need negative samples for training. The simulation results on a number of real-world time series show that the RBF-DDA based method outperforms the negative samples one. Moreover, the classification performance of the RBF-DDA method does not depend on the test set size whereas the performance of the negative samples method depends on the size of the test set. In addition to the novelty detection methods, this thesis proposes four different methods for improving the generalization performance of RBF-DDA neural networks. The proposed methods are evaluated using six benchmark classification datasets and the results show that they considerably improve RBF-DDA performance and also that they outperform MLPs and AdaBoost and achieve results similar to k-NN. These methods were also used in conjunction with the method for novelty detection in time series based on negative samples and the results show that they are also valuable for improving performance of novelty detection in time series, which is the main subject of this thesis

ASSUNTO(S)

redes rbf (radial basis functions) â mÃtodos de treinamento construtivos sÃries temporais â detecÃÃo de novidades e fraudes redes neurais artificiais ciencia da computacao rbf (radial basis functions) networks time series neural networks

ACESSO AO ARTIGO

Documentos Relacionados