Prediction and Observation Timing in Time Series Classification

Hartvigsen, Thomas

Etd

Prediction and Observation Timing in Time Series Classification

公开

Time series data mining is crucial to a wide variety of domains such as healthcare, weather prediction, seismology, and astronomy. Classifying time series is a challenging and important problem with applications from clinical diagnosis, natural disaster effect estimation, to stellar object detection. With a huge amount of time series data collected every day, solving rapidly-evolving, complex time series classification problems is essential. In this work, we introduce and study two notions of timing in time series classification: (1) The timing of predictions, or how early in time a classification is made without seeing the complete time series, is often a key factor in the usefulness of a time series classifier, and (2) The timing of observations, or timestamped values produced by various data sources, is often irregular across different data dimensions, thus introducing technical challenges to the classifier model. In the first part of this dissertation, we study prediction timing. How quickly a classifier comes to a decision can strongly effect its usefulness when integrated into a decision support system. This intuition underlies early classification of time series, where we find a timestep in an ongoing time series at which a classifier stops and makes its prediction early without having seen the complete time series, typically with a preference for consuming only a few early timesteps and accurate predictions. These two goals of earliness and high accuracy contradict one another. We investigate two directions for tackling prediction timing: 1. Tunable Early Classification of Time Series. A tunable early classifier lets a user choose how much importance to put on each of the contradictory goals of earliness and accuracy. Since no prior works have tackled this challenge, in this work we characterize propose the tunable early classification problem. We then develop a solution strategy based on a recurrent neural network time series model mixed with a reinforcement learning-based halting policy that chooses, at each timestep, whether or not to stop and classify. 2. Early Multi-label Classification of Time Series. Many time series classification problems are naturally modeled as the more general case of multi-label classification, where each instance is associated with a subset of all possible labels. Classifying such series early is an open problem, which we refer to as Early Multi-label Classification. In this dissertation, we develop its first solution: An integration of recent classifier chain approaches with multi-label classification models as well as with adaptive-halting policy networks. In the second part of this dissertation, we study observation timing. Time series are often collected with irregular spacing between different observations (data values). Classifying such irregular time series has many important applications from clinical diagnosis to seismology. However, modeling these data is tremendously challenging because sampling rates can differ between variables, values may be missing, variables evolve over time without being observed, and data generation functions may or may not be correlated. In this dissertation, we develop classifiers that learn directly from irregular time series, studying two challenges related to directions for observation timing: 1. Continuous-Time Attention for Irregular Time Series Classification. Classifying time series often depends on finding discriminative subsequences in time series data. A model that explicitly finds these moments-of-interest in continuous time will be more robust to noise by disregarding irrelevant portions the timeline. In this work, we leverage the power of attention to find relevant moments-of-interest. 2. Classifying Irregular Time Series Early. Once we can find moments-of-interest, a natural follow-on question arises, namely, "can we find them early to support time-sensitive applications?" We investigate this question, developing a model that classifies irregular time series as early as possible without first observing all timesteps. This is the first work to combine observation timing with prediction timing. For all tasks, we compare our proposed models against state-of-the-art alternatives from the literature and verify that our proposed methods outperform them consistently and significantly. In these studies, we measure performance using established metrics on a wide variety of publicly-available datasets.

Creator