Hybrid Data Science Framework for Adaptive Learning Optimization in Digital Education Ecosystems

The rapid expansion of digital education platforms has generated large volumes of heterogeneous learning data, yet many adaptive learning systems remain limited to static personalization rules or prediction-centric analytics that fail to translate insights into effective instructional decisions. This study proposes a Hybrid Data Science Framework that integrates multi-source educational data engineering, hybrid analytical modeling, and adaptive policy optimization to enhance learning personalization in digital education ecosystems. The framework is designed as an end-to-end, deployment-oriented architecture that transforms raw interaction logs, assessment records, content metadata, and contextual engagement signals into optimized adaptive learning interventions. Empirical evaluation demonstrates that the proposed data engineering pipeline reduces missing data rates from 18% to 4%, decreases invalid session patterns from 14% to 2%, and minimizes temporal leakage flags to 1%, thereby ensuring evaluation validity and operational reliability. In predictive modeling experiments, the hybrid model, combining statistical and machine learning components, outperforms single-paradigm baselines, achieving lower RMSE (0.60 vs. 0.74 and 0.66), reduced MAE (0.46), and higher Macro-F1 (0.74), alongside improved calibration. When embedded within a policy-based adaptive optimization loop, the framework yields steadily increasing cumulative rewards across learning episodes, indicating convergence toward stable and effective intervention strategies. Learning outcome analysis reveals that adaptive interventions lead to measurable improvements in learner performance, with normalized gains averaging 13.3% in mid-performing learners and consistent positive gains across low-performing segments. Robustness testing further shows gradual performance degradation under behavioral drift, with Macro-F1 decreasing from 0.74 to 0.67 over a 12-month simulation, while maintaining bounded inference latency (mean ≈ 85 ms). These findings collectively demonstrate that the proposed hybrid framework not only improves predictive accuracy but also delivers tangible learning gains, operational stability, and scalability, positioning it as a viable foundation for next-generation adaptive learning systems.