Accurate Time Series Classification Using Deep Learning and Partial Observations

Over the past decade, there has been a significant increase in interest in classifying time series data. Typically, there are two objectives of classifying time series data, efficiency (speed) and effectiveness (accuracy). In this dissertation, we propose a framework to detect optimal partial observations to classify time series data. Further, we propose a few deep learning architectures to improve the classification accuracy of univariate and multivariate time series data. Finally, we present an architecture that successfully generates adversarial samples on a couple of time series classification models. These adversaries possess a security concern and would require further research to defend against.