Blending ensemble aplicado em reconhecimento de conjunto aberto para classificação de séries temporais

Abstract

This master thesis investigates the application of Open-Set Recognition (OSR) methods to Time Series Classification (TSC), addressing a gap in the existing literature. While TSC involves labeling time series instances, OSR methods are designed to simultaneously classify known samples and detect unknown ones. This is relevant because it enables models to identify known and detect unknown patterns in dynamic and unpredictable environments, enhancing robustness and reliability in real-world applications that have no shortage of time series data. Despite their potential, OSR methods have seen limited application in TSC compared to other domains such as text or image classification. Recent research has introduced innovative methodologies to address this gap, each with its strengths and limitations. However, existing OSR methods in TSC often overlook the utilization of “known unknowns” during training, which could enhance model robustness. Moreover, the transfer-learning approach across various neural network models requires improvement, and experiments with benchmark datasets are limited in scope. In response to these challenges, this research proposes a novel approach that integrates neural networks, models that have a proven record of high performance in TSC, with OpenMax layers, designed to make neural networks capable of working in ope-set recognition, both providing features to train a blending model, while also incorporating known unknown instances during training. To rigorously evaluate the effectiveness of the proposed model, comprehensive and extensive experiments are conducted across various configurations using diverse time series datasets from different domains as well as comparing its performance against previous similar work. The results of these experiments provide valuable insights into the performance and applicability of the proposed approach in real-world scenarios and, contributing to the advancement of both Time Series Classification and Open-Set Recognition methodologies.