Social media's pervasive adoption among students has increasingly been linked to academic procrastination and reduced productivity, yet existing studies remain largely correlational and lack predictive robustness. This paper proposes an interpretable ensemble learning framework to forecast the impact of social media usage on students' academic productivity. Using a primary survey dataset of 1,501 students from Dhaka, Bangladesh, a rigorous machine learning pipeline is developed incorporating stratified train-test splitting, training-only preprocessing, and SMOTE-based class balancing to prevent data leakage and address class imbalance. The proposed AdaBoost-DT ensemble synergistically combines three heterogeneous base learners, a Multi-Layer Perceptron Classifier, Gradient Boosting, and SVM with RBF kernel, through adaptive weighting. Evaluated against eleven baseline classifiers and four ensemble strategies, the model achieves a state-of-the-art accuracy of 98.5%, further validated through 5-fold stratified cross-validation, yielding a mean accuracy of 97.8%±0.4%, confirming stable generalization. An ablation study isolates the contribution of each component, demonstrating that SMOTE and adaptive weighting each contribute meaningfully to the final performance. SHAP-based Explainable AI provides global and instance-level interpretability, identifying study delay, platform preference, and time spent on social media as the most influential predictors. The framework offers a reliable decision-support tool for educators and policymakers designing evidence-based digital wellness interventions.