Probabilistic Hypothesis Anchored Domain Adaptation for Modeling Human Behavior
Year: 2026
Type: conference-paper
Abstract: Energy-efficient smart buildings depend on accurate modeling of human behavior, such as occupancy and activity patterns, to manage resources intelligently and reduce overall energy consumption while maintaining occupant comfort. However, achieving robust performance in such models remains challenging due to scarce labeled data, variations in sensor configurations and occupant behavior across buildings, and the high-dimensional, non-linear nature of multimodal sensor data. Furthermore, privacy and data-sharing constraints further limit access to raw building data. To address these challenges, we propose PHADA (Probabilistic Hypothesis-Anchored Domain Adaptation), a source-free framework that integrates probabilistic modeling with hypothesis transfer to enable robust unsupervised domain adaptation. PHADA leverages a Semi-MPPCA backbone and introduces an unsupervised adaptation stage with a novel component-anchor alignment mechanism that maintains coherent latent structure as the model transfers across domains. This design improves performance and reduces class collapse under domain shift. Experiments on multi-environment smart building datasets show that PHADA outperforms state-of-the-art baselines under domain shift and label scarcity, while ablation studies demonstrate the impact of the proposed design choices. We further include a post-hoc feature-level analysis to examine how domain adaptation affects model decision behavior on the target domain. Code will be released upon acceptance.
Institution: Concordia University
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7165165161