Multi-Source Deep Domain Adaptation with Weak Supervision for Time-Series Sensor Data

Garrett Wilson, Janardhan Rao Doppa, Diane J. CookKDD 2020 · DOI

TL;DR

This paper introduces CoDATS, a novel deep learning architecture for domain adaptation in time-series sensor data, which significantly improves accuracy and training efficiency over existing methods. It also proposes a new "Domain Adaptation with Weak Supervision" (DA-WS) approach that leverages readily available label proportions to further boost performance, especially on imbalanced datasets.

What Problem Does This Paper Address?

Time-series sensor data, like that from wearables or IoT devices, is abundant but often suffers from domain shift – meaning models trained on one set of data (e.g., one person's activity) perform poorly on another (e.g., a different person's activity). Domain adaptation (DA) aims to bridge this gap by reusing labeled data from source domains to improve performance on a target domain with limited or no labels. However, existing DA methods, often designed for image data, are not well-suited for the temporal nature of time-series data, and can be slow to train. Furthermore, obtaining precise labels for time-series data (like human activity recognition) is difficult and expensive, unlike image labeling. This paper addresses these challenges by proposing a time-series-specific DA architecture and a novel method to leverage weaker forms of supervision, like label proportions, which are easier to obtain.

Methods

The core of the paper is the CoDATS (Convolutional deep Domain Adaptation model for Time Series data) architecture. CoDATS builds upon the principles of domain-invariant learning, particularly the Domain-Adversarial Neural Network (DANN) approach. However, it adapts this for time-series data by using a fully convolutional network (FCN) as the feature extractor. This CNN-based approach is chosen for its efficiency and ability to handle sequential data, outperforming recurrent neural networks (RNNs) in terms of training speed and comparable accuracy.

CoDATS supports both single-source and multi-source domain adaptation. In the multi-source setting, it can leverage data from multiple labeled source domains to improve adaptation to the target domain.

The paper also introduces Domain Adaptation with Weak Supervision (DA-WS). This method tackles the challenge of expensive time-series labeling by utilizing target-domain label proportions. Instead of exact labels, DA-WS uses information about the percentage of time spent on each activity. This is incorporated into the training objective as a differentiable regularization term, guiding the model to produce predictions that, on average, match these known proportions.

Datasets

The authors evaluate their methods on several real-world time-series sensor datasets:

  • Human Activity Recognition (HAR) dataset: Data from 30 participants using accelerometers, gyroscopes, and body acceleration sensors.
  • Heterogeneity Human Activity Recognition Dataset (HHAR): Data from 31 smartphones with varying manufacturers and models, capturing accelerometer data.
  • WISDM activity recognition (WISDM AR) dataset: Accelerometer data from 33 participants.
  • uWave dataset: Accelerometer data from 8 participants performing hand gestures.

These datasets are suitable for domain adaptation because they naturally allow for splitting data into different domains based on participant identifiers, reflecting variations in data collection or usage patterns.

Key Insights and Ideas

  • CNNs are a viable and efficient alternative to RNNs for time-series DA: The paper demonstrates that a CNN-based feature extractor can achieve competitive accuracy with significantly faster training times compared to RNN-based methods like R-DANN and VRADA.
  • Multi-source DA offers substantial gains: Leveraging data from multiple source domains consistently improves accuracy, especially when the number of source domains increases. This highlights the benefit of pooling diverse labeled data.
  • Weak supervision (label proportions) is powerful for imbalanced datasets: The DA-WS method shows significant accuracy improvements on datasets with imbalanced class distributions (like WISDM AR), where exact labels are hard to obtain and class frequencies vary widely. This suggests a practical way to enhance DA models without requiring full labels.
  • Domain adaptation is crucial for real-world sensor data: The significant gap between "No Adaptation" and adapted models across datasets underscores the necessity of DA techniques for deploying models in diverse user or environmental conditions.

Key Results

  • Single-Source DA: CoDATS consistently outperformed R-DANN and VRADA in accuracy across multiple datasets, often achieving significantly higher scores. For example, on the HAR dataset, CoDATS achieved an average accuracy of 90.2% compared to 70.0% for R-DANN and 70.2% for VRADA.
  • Training Efficiency: CoDATS demonstrated substantial speedups, requiring only 20-22% of R-DANN's training time and 6% of VRADA's training time.
  • Multi-Source DA: Increasing the number of source domains (n) generally led to improved accuracy for CoDATS, outperforming the "No Adaptation" baseline. For instance, on HHAR with n=25, CoDATS achieved 86.7% accuracy compared to 76.2% for "No Adaptation."
  • Domain Adaptation with Weak Supervision (DA-WS): CoDATS-WS provided further accuracy improvements, particularly on the imbalanced WISDM AR dataset. For n=1, CoDATS-WS achieved 73.3% accuracy, a notable increase from CoDATS's 60.5%. This benefit diminished on more balanced datasets.

Limitations and Future Work

  • Limitations: The paper notes that DA-WS's benefit is most pronounced when label distributions differ between source and target domains. On perfectly balanced datasets, its advantage is minimal. The authors also acknowledge potential negative transfer in multi-source DA when domains are extremely dissimilar, which can slightly degrade performance.
  • Future Work: The authors suggest extending CoDATS to handle heterogeneous feature sets, exploring other forms of weak supervision relevant to time-series data, and incorporating insights from advanced architectures like InceptionTime for further model improvements.

Summary updated

BibTeX
@inproceedings{wilson2020multi,
  author = {Garrett Wilson and Janardhan Rao Doppa and Diane J. Cook},
  title = {{Multi-Source Deep Domain Adaptation with Weak Supervision for Time-Series Sensor Data}},
  booktitle = {ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
  year = {2020},
  doi = {10.1145/3394486.3403228},
}
Figure