Avaliação de algoritmos de redes neurais artificiais em tarefas de detecção de novidades: uma abordagem unificadora

Abstract

Novelty detection is a pattern recognition task whose goal is to report the occurrence of novel events, or new observations, from a previously computed statistical model of the data. Being the focus of an increasing attention in machine learning and data mining elds, a wide range of methods for novelty detection are available in literature, among

them it is possible to cite the linear and nonlinear statistical discriminants, recursive iden- tication of systems, expert systems, articial neural networks, fuzzy logic, evolutionary

computing, among others. However, few works evaluating the performance of the meth- ods just mentioned have been published. This fact can be, at least, partially explained

by the inherent complexity of comparing dierent modeling methods. Performance com- parisons are important to evaluate, for example, which technique works best on certain

types of data or which one is less sensitive (more robust) to outliers present in the data

used to build the model. Thus, the work developed in this dissertation attempts to con- tribute to the comparative analysis of novelty detection methods, proposing a general

methodology to compare the performance of neural-based novelty detection systems. The rationale underlying the proposed methodology consists in using the statistical models for the data, obtained from various articial neural networks, to compute decision thresholds for novelty detection tests. Such thresholds are based on the the concept of nonparametric bootstrap condence intervals. Finally, regarding the new methodology, it is worth noting that it allows both supervised and unsupervised neural algorithms to be compared under a common framework. Such a generality is illustrated through two distinct applications, namely, anomaly detection in cellular systems and cancer detection in mammographies.