Scalable Methods for Multivariate Time Series Classification

The availability of low-tech sensors has significantly impacted the goals of industry and academia. These sensors are now used in various health, agriculture, and finance applications. Their usage has led to the proliferation of temporal data, presenting challenges and opportunities. Temporal data is a series of real values collected over time. The changes in the temporal data are used to improve our understanding of how systems work and can also be used to make predictions. For example, data recorded from an athlete's run spanning over a few minutes could help identify a potential injury. The ease of availability of temporal data has opened up new opportunities for research and development. However, it also presents unique challenges. One challenge is that temporal data can be vast and complex, making storing, processing, and analysing difficult. The advancement in data capture technology and machine learning approaches have enabled us to tackle some of these challenges. This thesis focuses on studying and developing machine learning methods for analysing temporal data. In particular, it focuses on a type of temporal data called multivariate time series and a specific machine learning task called time series classification. Considering the above example of an athlete's run, data captured from the run is multivariate as it involves collecting motion data from multiple sources like arms, legs, and waist. Each source, in this case, a body part, is known as a channel. The task of predicting a discrete label, for example, injury type, on this temporal data is called multivariate time series classification. The multiple channels in a multivariate time series classification task can result in more computation time for different classifiers as the data increases proportionately to the number of channels. Therefore, reducing these delays is sometimes as crucial as the accuracy of classifiers. These delays can be broadly mapped to the scalability of classifiers. In this thesis, we examine various aspects of scalability for different types of time series datasets and time series classification algorithms and propose solutions to address them.