INTERPRETABLE PREDICTION OF STUDENT HAPPINESS USING SUPPORT VECTOR REGRESSION AND SHAP EXPLANATIONS

Abstract

Based on the UN Sustainable Development Goal 3 (Good Health and Well-Being), the proposed study constructs a machine-learning framework, which can be interpreted in a person-oriented way, to forecast Happiness level in a cohort of 1,500 university students using a psychosocial and demographic dataset obtained from Kaggle. Unlike previous research, in which most aggregate national well-being indexes are predicted based on black-box models, we model individual prediction and incorporate explainable AI to determine practical drivers of student well-being. Evaluation was done by the 5-fold cross-validation. Support Vector Regression (SVR) demonstrated the best generalization performance of indirect regressors (MAE = 0.0740, MSE = 0.0088) as far as RMSE is concerned (RMSE = 0.0937, R² = 0.6664) and adjusted R² = 0.6566). To make predictive accuracy policy-relevant, we utilized SHAP to measure the contributions of features. Social Support, Work-Life Balance, Work-related factors and Academic Stress emerged as the most influential predictors and Generosity and Financial Status generated lesser positive effects, whereas, Anxiety, Depression and Isolation had negative effects. Demographic factors (i.e. age and gender) did not have significant influence and it is thought that modification of psychosocial conditions plays the biggest role in explanatory power in this cohort. The implications of these findings would make a clear and implementable impact on the universities: a prioritization in stress-reduction programs, a reinforcement of peer-support structures, and a provision of focused financial support are likely to result in quantifiable improvement in the level of student well-being. In addition to the current use, the suggested architecture also proves the transferable prediction-to-intervention pipeline to evidence-based decision-making in education and population health setting in accordance with SDG-3.