Abstract

Learning from time series data is an essential component in the AI landscape given the ubiquitous time-dependent data in real-world applications. To motivate the necessity of learning from time series data, we first introduce different applications, data sources, and properties. These can be as diverse as irregular and (non-)continuous time series data as well as streaming and spatio-temporal data. To introduce the mechanics of learning from time series data, we elaborate on the most renowned convolutional, recurrent and transformer architectures for learning from time series. Then, we discuss essential characteristics of learning with time series. Therefore, we explain deep metric learning, which learns feature representations that capture the similarity between time series data. We further describe time series similarity learning to extract representations that allow comparison between sequences of spatio-temporal data. In addition, we discuss the interpretability of learning methods on time series data that target safety, non-discrimination, and fairness.