Dissertation
Multi-Source Unsupervised Domain Adaptation for Time-Series Sensor Data
Washington State University
Doctor of Philosophy (PhD), Washington State University
01/2021
DOI:
https://doi.org/10.7273/000003114
Handle:
https://hdl.handle.net/2376/122882
Abstract
Real-world problems such as human activity recognition and gesture recognition involve time series data. However, in the case of human activity recognition, for example, a machine learning model trained on one person's data may not perform well on another person's data due to different activity patterns, sensor positions, or sampling rates. This poses a problem, requiring data annotations to be collected for each person individually. Yet collecting such labels for time series may be expensive or infeasible due to the difficulty of interpreting raw time series sensor data. Instead, labels are often collected while performing each activity, but this adds to the burden of data collection. Multi-source unsupervised domain adaptation provides a remedy, by improving the performance of a machine learning model for an unlabeled (target) domain by leveraging ground truth labels in related (source) domains, i.e., leveraging data from additional people. If extra information is available about the target domain such as estimated label proportions, e.g., self-reported estimates of how many hours are spent performing various activities, then weak supervision can further improve performance through regularization. While domain adaptation has been previously studied in other contexts, prior work leaves open a number of avenues that may lead to drastically improved time series adaptation performance. We hypothesize that we can improve time series adaptation performance by (1) improving the neural network architecture used for adaptation, (2) learning a domain-invariant representation across multiple source domains and the target domain, (3) incorporating a weak supervision regularizer if target-domain label proportions are available, and (4) utilizing cross-source label information via contrastive learning. We validate each of these hypotheses through developing new adaptation models and methods that we evaluate on a variety of time series sensor datasets including human activity recognition, electromyography, and synthetic data. We observe significant improvements in unsupervised domain adaptation performance for time series sensor data. Finally, prompting directions for future work, we describe and evaluate a variety of domain adaptation methods that we find are currently inadequate for time series data and additionally collect a more challenging mobile human activity recognition dataset on which domain adaptation methods currently struggle.
Metrics
75 File views/ downloads
149 Record Views
Details
- Title
- Multi-Source Unsupervised Domain Adaptation for Time-Series Sensor Data
- Creators
- Garrett Wilson
- Contributors
- Diane J Cook (Advisor)Janardhan Rao Doppa (Advisor)Lawrence B Holder (Committee Member)
- Awarding Institution
- Washington State University
- Academic Unit
- Electrical Engineering and Computer Science, School of
- Theses and Dissertations
- Doctor of Philosophy (PhD), Washington State University
- Publisher
- Washington State University
- Number of pages
- 330
- Identifiers
- 99900651794201842
- Language
- English
- Resource Type
- Dissertation