Interpretable Decision Tree for Predicting High Self-Reported Happiness Using Digital Behavior and Lifestyle Indicators

Abstract

Research on digital well-being is dominated by text-derived features and post hoc explanations, leaving limited evidence on whether compact behavioral data can support an inherently interpretable happiness classifier without obscuring predictive trade-offs. This study develops and audits an inherently interpretable decision tree for classifying high self-reported happiness from age, screen time, sleep quality, stress, offline days, exercise, gender, and social-media platform. The contribution is a leakage-safe evaluation that combines nested cross-validation, comparison with Random Forest, logistic regression, and radial-basis-function support vector machine, out-of-fold discrimination and calibration, a one-standard-error tree-selection rule, bootstrap stability, and sensitivity analyses. Across five outer folds, the decision tree obtained balanced accuracy 0.779 ± 0.049, macro-F1 0.778 ± 0.050, ROC-AUC 0.869 ± 0.045, and Brier score 0.146 ± 0.028. Random Forest achieved the highest mean balanced accuracy (0.820), whereas logistic regression achieved the highest ROC-AUC (0.913) and lowest Brier score (0.122). The selected depth-three tree retained eight leaves and used daily screen time and stress as its primary decision pathways. In 200 bootstrap samples, stress was selected in 100.0% of trees and screen time in 86.5%; however, the root alternated between stress (53.5%) and screen time (45.0%), indicating stable relevance but unstable ordering. Results remained similar with the four-feature subset, while performance changed across happiness thresholds. The model therefore offers transparent decision rules at a modest predictive cost, but the cross-sectional, single-dataset design supports predictive association rather than causal or clinical interpretation