The Impact of Domain Shift on Predicting Perceived Sleep Quality from Wearables

Machine learning models for personal informatics systems are typically trained offline on records of a specific population of users, resulting in population models. These models may suffer performance degradation in real-world settings due to domain shift, i.e., differences in data distributions across users and contexts. Domain adaptation techniques can address this issue by, e.g., personalizing models with user-specific data. In this paper, we quantify the impact of domain shift on the performance of both population and personalized models in a specific scenario: sleep quality recognition. To this end, we also collect and make available to the research community the new BiheartS dataset. Our analysis shows that domain shift causes the accuracy of population models to decrease by up to 18.54 percentage points, when used on new data. Personalized models, instead, show robust performance across datasets. However, crafting personalized models typically requires using new data or user-provided labels, limiting their applicability in real settings. To mitigate the limitations of both population and personalized models, we propose a novel unsupervised domain adaptation approach: the cluster-based population model (CBPM). CBPM achieves accuracy improvements of up to 13.45 percentage points w.r.t. population model without requiring the use of user-specific records or labels.