Date of Award

Winter 2024

Project Type

Dissertation

Program or Major

Computer Science

Degree Name

Doctor of Philosophy

First Advisor

Laura Dietz

Second Advisor

Marek Petrik

Third Advisor

Samuel Carton

Abstract

The application of time series data has a rich history, and its relevance continues to expand due to recent advancements in artificial intelligence (AI) and autonomous systems. These technological developments have driven the collection and utilization of time series data to new heights, propelling its significance to unparalleled levels. Deriving insights from time series data has become indispensable for effective decision-making in various domains, as it enables the identification of patterns, trends, and fluctuations that inform strategic actions and optimize outcomes. The core challenge lies in developing precise and reliable predictive models that effectively leverage the input data for accurate predictions of the desired outputdata. This dissertation advances the state of the art in time series modeling by developing novel methodologies specifically tailored to exploit the inherent characteristics of time series data. These methodologies excel in constructing representations of input data that more accurately predict the desired output. In pursuit of constructing effective representations, we explore three distinct and innovative approaches. First, we introduce a novel method designed to identify and leverage periods of similar behavior between input and output across different time scales, improving the detection of otherwise non-obvious similarities. Inspired by image-denoising techniques, our second approach enhances time series forecasting by proposing a compound forecast-blur-denoise framework that encourages a division of labor between coarse- and fine-grained feature learning. Lastly, we improve time series clustering by developing mutual/shared time series representations that facilitate the identification of similar time series that potentially belong to the same cluster by highlighting common features. These efforts collectively contribute to the advancement of temporal modeling, benefiting researchers and practitioners. Researchers can use these advanced methodologies to gain deeper insights into complex temporal relationships, while practitioners across various industries can implement these techniques to improve forecasting, clustering, classification, anomaly detection, and more.

Download

Share

COinS