Toward Personalized Meditation Recommendations for Digital Well-Being

Digital meditation platforms are widely used to support mental well-being, yet many still rely on generic recommendations that overlook users' emotional states and situational context. In this work, we explore the use of interpretable machine learning to generate personalized meditation suggestions. We frame the task as a classification problem that predicts fine-grained meditation styles from textual descriptions of user prompts and contextual information, enriched with emotion labels. Two models, Logistic Regression and Random Forest, are evaluated using embedding representations of the input text. Experimental results indicate that user prompts and contextual information provide the most informative signals for identifying suitable meditation techniques, whereas emotion labels alone contribute limited predictive value. To enhance transparency, we also perform token-level attribution analysis, highlighting how specific words and phrases influence model predictions and revealing interpretable connections between user language, emotional cues, and recommended meditation practices.