Human Context Recognition (HCR) is an important task in Context-Aware (CA), ubiquitous computing systems that often utilize Machine Learning. HCR involves recognizing the user’s context to adapt a context-aware application’s behavior. HCR on smartphones faces an main challenge when data is collected in-the-wild: Imbalanced data as subjects do not visit all contexts equally. This problem cause significant reductions in the performance of HCR machine learning classifiers. To solve this problem, various methods of data augmentation and synthetic data generation have been proposed in prior work. Image-to-image translation Generative Adversarial Networks (GANs) convert images from one domain to another, and prior work has found them effective for augmenting smartphone human activity data, but they have not been explored for HCR data. In this thesis, we systematically studied, rigorously evaluated and compared three state-of-the-art Image-to-Image translation GANs for the task of generating synthetic smartphone HCR data to augment real HCR data to improve HCR performance. Specific state-of-the-art image-to-image translation GANs we explored include StarGAN V2, Gaussian Mixture Model Unsupervised Image-to-Image Translation (GMM-UNIT) and Domain Specific GAN (DOS-GAN). Various quantitative evaluation metrics were employed, including FID score and KL Divergence to evaluate the quality of the data generated. Along with the quantitative measures, the performance of a deep learning classifier trained on synthetic data generated by different GAN models were compared. All three models have been able to create quality and diverse data. The HCR dataset generated using GMM-Unit has achieved the lowest KL Divergence and FID Score of 4.11 and 21.91 respectively. An HCR classifier trained on the synthetic data generated by the GMM-Unit slightly outperformed one trained on real HCR data with minor improvements of 0.72% and 0.3 in accuracy and F1 score respectively.

Creator

• Sankaranarayanan, Ronak

Contributors

• Eltabakh, Mohamed Y.
• Agu, Emmanuel O.

Degree

• MS

Unit

• Data Science

Publisher

• Worcester Polytechnic Institute

Identifier

• etd-67166

Keyword

• Data Augmentation
• Sensor Data
• Human Context Recognition (HCR)
• Generative Adversarial Networks (GAN)
• Time series data

Advisor

• Agu, Emmanuel O.

Defense date

• 2022-04-25

Year

• 2022

Date created

• 2022-05-02

Resource type

• Thesis

Rights statement

• In Copyright

Last modified

• 2023-09-20