Cost-Effective Multitask Active Learning in Wearable Sensor Systems.

Area of Science:

• * Mobile Health
• * Machine Learning
• * Wearable Sensors

Background:

• * Supervised multitask learning (MTL) models enhance performance but require extensive labeled data.
• * Multitask active learning (MTAL) addresses data scarcity but requires effective query strategies, especially in mobile health.
• * Existing MTAL strategies are underexplored in wearable sensor systems, necessitating research into efficient label acquisition.

Purpose of the Study:

• * To investigate multitask active learning (MTAL) query strategies for wearable sensor systems.
• * To evaluate the effectiveness of different sampling methods in mobile health applications.
• * To propose and validate a novel sampling approach for MTAL in resource-constrained settings.

Main Methods:

• * Investigated rank-based sampling and other traditional MTAL query strategies.
• * Utilized activity recognition and emotion classification datasets from wearable sensors.
• * Proposed and implemented a Clustered Stratified Sampling (CSS) method integrated with MTAL.

Main Results:

• * Rank-based sampling demonstrated superior performance, particularly for highly correlated tasks.
• * Solely relying on informativeness for sample selection can introduce model bias.
• * The proposed CSS method, combined with rank-based querying, achieved up to 9% accuracy improvement with a 2000-query budget.

Conclusions:

• * Effective MTAL query strategies are crucial for successful mobile health applications using wearable sensors.
• * The proposed Clustered Stratified Sampling (CSS) method enhances MTAL efficiency and accuracy.
• * CSS optimizes budget utilization and mitigates bias, offering a promising solution for label-deficit scenarios.