Health and well-being prediction plays an essential role in mental healthcare and well-being-aware computing. The complex nature of well-being, resulting from its dependency on a person’s physiological health, mental state, and surroundings, makes its prediction a challenging task. In this work, we utilize mobile sensing data to predict self-reported well-being metrics. In addition to a person’s physiology, we incorporate the environment’s impact through weather and social network data. To this end, we leverage phone data to construct social networks and develop a machine learning architecture that aggregates information from multiple users within the graph network and integrates it with the temporal dynamics of data to predict well-being outcomes for all users.

To address the dynamic nature of social networks, we introduce GEDD (Graph Extraction for Dynamic Distribution), an approach that automatically adapts to fluctuating network sizes. GEDD utilizes graph properties, including connectivity and components, to transform variable-sized graphs into a standardized format, ensuring no user data is discarded. The proposed architecture supports online learning, making it feasible to scale to large networks without adding ecological momentary assessments (EMAs) or additional data collection burdens, thus preserving user privacy.

Through extensive evaluations, we show that social network incorporation improves prediction accuracy, although node influence, especially in users with high eigenvector centrality, can amplify noise. To address this, we propose a robust system that leverages attention and social contagion in well-being behaviors through graph networks and integrates it with physiological and phone data from ubiquitous mobile and wearable devices. This system is designed to predict well-being outcomes, such as sleep duration and other health metrics while mitigating the challenges posed by noisy and incomplete data.

Finally, we further leverage the graph structure to reduce the user burden associated with collecting health and well-being metrics, which are often captured at a much lower resolution than sensing data through surveys and EMAs. To this end, we introduce a benchmark framework to evaluate existing state-of-the-art graph-based active learning (AL) strategies in dynamic sensing environments. Our framework assesses AL strategies in terms of adaptability to real-time, user-centric data by evaluating performance over time in a stream-based setting. We also introduce new metrics, including sampling entropy, coverage ratio, and time-gap analysis, to quantify user burden, sampling diversity, and generalization performance. These metrics provide a holistic view of the AL strategies’ effectiveness, helping to identify those that best balance predictive accuracy and user engagement. This comprehensive evaluation framework supports scalable and efficient health prediction systems, facilitating practical, large-scale deployment.