Single/Multi-Source Black-Box Domain Adaption for Sensor Time Series Data.

Area of Science:

• Machine Learning
• Sensor Data Analysis
• Time Series

Background:

• Unsupervised domain adaptation (UDA) transfers knowledge from labeled source domains to unlabeled target domains.
• Existing UDA methods for sensor time series data face limitations like requiring source data access and ignoring temporal consistency.
• Privacy concerns and storage limitations restrict direct access to source data in many applications.

Purpose of the Study:

• To develop a black-box domain adaptation framework for sensor time series data (B2TSDA).
• To address challenges including privacy constraints, limited source data access, temporal inconsistency, and low signal-to-noise ratio (SNR).

Main Methods:

• A single/multi-source teacher-student learning framework is proposed for knowledge distillation.
• A novel temporal consistency loss is designed using adaptive masking and dynamic thresholding.
• A Shapley-enhanced method is introduced for multi-source black-box domain adaptation to weigh source domain contributions.

Main Results:

• The B2TSDA framework demonstrates superior performance in both single-source and multi-source domain adaptation scenarios.
• The proposed methods effectively maintain temporal information and handle learning difficulties.
• Experimental results show significant improvements over existing black-box UDA methods.

Conclusions:

• The B2TSDA framework offers an effective solution for domain adaptation in sensor time series data, especially under privacy constraints.
• The developed temporal consistency loss and Shapley-enhanced contribution method advance the field of UDA for sensor data.
• This work provides a robust approach for applications like human activity and gesture recognition.