CRoP: Context-wise Robust Static Human-Sensing Personalization.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Human-Computer Interaction

Background:

• Deep learning and IoT enable human sensing, but distribution shifts challenge generic models.
• Existing personalization methods often neglect intra-user data heterogeneity across contexts.
• Limited data in clinical settings exacerbates generalization and personalization challenges.

Purpose of the Study:

• To address the intra-user generalization challenge in human sensing personalization.
• To introduce CRoP, a static personalization approach enhancing intra-user robustness.
• To improve model performance in diverse contexts, particularly in healthcare.

Main Methods:

• Leveraging off-the-shelf pre-trained models as a foundation.
• Employing adaptive pruning on a minimal sub-network to capture user-specific traits.
• Incorporating generic knowledge into the remaining network parameters.

Main Results:

• CRoP demonstrated superior personalization effectiveness and intra-user robustness.
• Validation across four human-sensing datasets, including two real-world health domains.
• Empirical justification provided through gradient analysis, ablation studies, and baseline comparisons.

Conclusions:

• CRoP offers a robust solution for personalized human sensing, addressing intra-user heterogeneity.
• The approach shows significant practical and social impact, especially in data-scarce clinical applications.
• CRoP enhances generalizability and personalization by effectively balancing user-specific and generic knowledge.