Abstract

Rising energy demands have intensified the pursuit of intelligent building systems capable of adapting to occupant behavior to optimize energy consumption. Occupancy Estimation and Activity Recognition play a central role in enabling data-driven, energy efficient control strategies that enhance sustainability while maintaining comfort. However, existing approaches often rely on large, labeled datasets collected from a single environment. When deployed in new settings, their performance degrades due to domain shifts caused by differences in sensor layouts, room
configurations, and occupant behavior. Acquiring sufficient labeled data for each environment is costly, time-consuming, and raises privacy concerns. This thesis addresses these challenges by developing explainable, label-efficient domain adaptation frameworks that enable models to generalize across buildings. First, we introduce Imbalance–Structure Expansion Reduction (IMB-SER), an interpretable tree-based domain adaptation method that mitigates class imbalance. Second, we propose Uncertainty aware Transfer via Augmentation for Decision Trees (UTA-DT), a source-free framework that augments target data with uncertainty measures from pretrained source model to achieve cross-domain robustness. Lastly, we develop a novel semi-supervised framework, the Semi-Supervised Mixture of Probabilistic Principal Component Analyzers (Semi-MPPCA), which leverages unlabeled data and is fine-tuned with minimal labeled samples to model human behavior
effectively. Building on this foundation, we extend the model by integrating an unsupervised domain adaptation framework, collectively referred to as Probabilistic Hypothesis Anchored Domain Adaptation (PHADA). A comprehensive analysis is conducted under varying label availability scenarios to evaluate performance, robustness, and label efficiency. Together, these contributions advance the development of transparent, label-efficient, and privacy-preserving domain adaptation methods.