Unsupervised Domain Adaptation With Source Data for Estimating Occupancy and Recognizing Activities in Smart Buildings
Jawher Dridi, Manar Amayri, Nizar Bouguila · IEEE Transactions on Artificial Intelligence · 2025
In this paper, we create and develop several unsupervised domain adaptation (UDA) methods that use source data to estimate the number of occupants and recognize activities in smart buildings. The created and developed methods have direct access to the source data, and their goal is to mitigate domain shift between source and target environments and to deal with the challenge of costly and time-consuming data labeling. We create our own novel UDA method called ATDOC-DDC that combines two powerful domain invariant methods Auxiliary Target Domain-Oriented Classifier (ATDOC) and Deep Domain Confusion (DDC) for better adaptation and performance. We introduce the Virtual Adversarial Domain Adaptation (VADA) model which is a combination of domain adversarial training and a penalty term. On the one hand, domain adversarial training trains a model on source and target data, and a discriminator to distinguish between data coming from different domains. The goal of the model is to fool the discriminator by creating features that generalize for different domains. On the other hand, the penalty term is added to punish the violation of the cluster assumption. Moreover, we introduce Sliced Wasserstein Discrepancy (SWD) for UDA which uses task-specific decision boundary and Wasserstein metric to perform domain alignment between source and target environments. The SWD distance is used to measure the difference between data distributions of the two domains so that the model can update its weights to minimize the considered distance (data alignment). Also, we consider three Adaptive Feature Norm (AFN) approaches that adapt the feature norms of the models to allow information transferability between domains. The adapted methods are Hard Adaptive Feature Norm (HAFN), Stepwise Adaptive Feature Norm (SAFN), and SAFN with entropy minimization. In addition, data poisoning has been employed to evaluate the robustness of the considered methods using mislabeled data. The obtained scores prove the efficiency and robustness of the proposed method on activity recognition and occupancy estimation datasets.