A tale of two time series methods : representation learning for improved distance and risk metrics

Abstract

In this thesis, we present methods in representation learning for time series in two areas: metric learning and risk stratification. We focus on metric learning due to the importance of computing distances between examples in learning algorithms and present Jiffy, a simple and scalable distance metric learning method for multivariate time series. Our approach is to reframe the task as a representation learning problem -- rather than design an elaborate distance function, we use a CNN to learn an embedding such that the Euclidean distance is effective. Experiments on a diverse set of multivariate time series datasets show that our approach consistently outperforms existing methods. We then focus on risk stratification because of its clinical importance in identifying patients at high risk for an adverse outcome. We use segments of a patient's ECG signal to predict that patient's risk of cardiovascular death within 90 days. In contrast to other work, we work directly with the raw ECG signal to learn a representation with predictive power. Our method produces a risk metric for cardiovascular death with state-of-the-art performance when compared to methods that rely on expert-designed representations.