Time-Series Representation Learning through Dynamic Temporal Reordering and Test-Time Adaptation
Loading...
Authors
Grover, Shivam
Date
2025-07-29
Type
thesis
Language
eng
Keyword
Time series , Representation learning , Test time adaptation , Forecasting , Classification , Adaptation , Temporal learning
Alternative Title
Abstract
Time-series applications span domains such as energy, healthcare, finance, and climate modeling, where the ability to learn rich temporal representations and remain robust under distribution shift is critical. However, existing models often rely on fixed temporal structures that overlook non-local dependencies and are typically non-adaptive when deployed in dynamic, real-world environments. This thesis addresses these challenges through two key contributions. First, we propose S3, a lightweight, architecture-agnostic network layer that enhances time-series representation learning by learning to reorder non-overlapping segments of the input sequence. S3 allows models to discover optimal task-relevant reordering of the time-steps, resulting in more effective learned representations. Extensive experiments across classification, forecasting, and anomaly detection tasks show that integrating S3 into state-of-the-art models yields consistent and significant performance gains of up to 39.6% improvement in classification accuracy and over 68% reduction in forecasting error. Second, we introduce DynaTTA, a dynamic test-time adaptation framework for time-series forecasting under distribution shift. DynaTTA monitors prediction error and embedding drift to estimate shift intensity and adjusts its adaptation rate and influence of adaptation accordingly. This enables models to adapt cautiously under mild shifts and aggressively under severe ones. To support rigorous evaluation, we also present TTFBench, a benchmark suite with diverse synthetic perturbations emulating real-world shifts. Experiments demonstrate that DynaTTA outperforms prior adaptation strategies, delivering robust performance across perturbed and long-horizon settings.
Description
Citation
Publisher
License
Queen's University's Thesis/Dissertation Non-Exclusive License for Deposit to QSpace and Library and Archives Canada
Intellectual Property Guidelines at Queen's University
Copying and Preserving Your Thesis
This publication is made available by the authority of the copyright owner solely for the purpose of private study and research and may not be copied or reproduced except as permitted by the copyright laws without written authority from the copyright owner.
Attribution 4.0 International
Intellectual Property Guidelines at Queen's University
Copying and Preserving Your Thesis
This publication is made available by the authority of the copyright owner solely for the purpose of private study and research and may not be copied or reproduced except as permitted by the copyright laws without written authority from the copyright owner.
Attribution 4.0 International