Explainable Deep Learning Models for Interpreting Learner Progression in Adaptive Education Systems

Adaptive education systems increasingly rely on deep sequence models to estimate learner progression and trigger risk-sensitive interventions, yet limited transparency constrains instructional oversight and trust. This study proposes an Explainable Deep Learning (XDL) framework that integrates concept-aligned attention, sparse concept gating, and temporally smooth latent states to jointly optimize prediction and interpretation. Experiments used interaction logs from 1,248 learners over a 10-week course, comprising 286,410 timestamped events across 2,016 learning items and 62 syllabus concepts. On the held-out Weeks 9–10 test window, the proposed XDL achieved an AUC of 0.852 for next-item correctness, a weekly mastery MAE of 0.061, and a disengagement-risk macro-F1 of 0.681, improving over a standard Transformer (AUC 0.824, MAE 0.071, macro-F1 0.628). Probabilistic reliability also improved, with Expected Calibration Error (ECE) decreasing from 0.053 (Transformer) to 0.038 (XDL). Explanation evaluation showed stronger faithfulness and coherence: deletion-based AUC drop increased from 0.15 (Transformer) to 0.22 (XDL), insertion-based AUC gain increased from 0.12 to 0.18, and temporal explanation stability rose from 0.78 to 0.87 cosine similarity. Pedagogical alignment improved, with top-concept plausibility increasing from 0.79 to 0.90, indicating that explanations preferentially referenced the concept of the upcoming assessment item or its prerequisite chain. Ablation analysis confirmed that removing sparsity reduced stability from 0.87 to 0.79 while only modestly affecting AUC (0.852 to 0.838), demonstrating that interpretability mechanisms materially shape explanation quality beyond predictive performance. Overall, results indicate that embedding explainability constraints into progression modeling yields accurate, calibrated, and instructionally actionable interpretations suitable for adaptive learning deployment.