International Journal for Applied Information Management

Decision Policy Optimization for Human–AI Collaboration Using Off-Policy Reinforcement Learning from Logged Interaction Data

2026-06-16T20:23:37+07:00

This paper investigates offline policy optimization for adaptive learning using logged student interaction traces, targeting reliable improvement without online exploration. A conservative offline reinforcement learning pipeline is implemented with calibrated behavior-policy propensities and doubly robust off-policy evaluation. Using 128,640 student trajectories (2.94 million events) with a 32-dimensional state representation and 12 pedagogical actions, the optimized policy achieved a +0.042-return improvement over a supervised next-item baseline under doubly robust estimation, with a bootstrap confidence width of ±0.021. Self-normalized estimators produced consistent rankings, reporting a +0.041 improvement with comparable uncertainty. Performance gains were horizon-stable and concentrated in medium horizons, where improvement increased from +0.012 at 1 step to +0.055 at 5 steps and remained positive through 10 steps. Safety analysis showed a shift toward bettersupported actions, increasing mean action support from 0.31 to 0.44 and reducing the low-support decision rate from 0.18 to 0.06. Uncertainty pruning activated on 11% of decisions, decreasing the high-uncertainty rate from 0.22 to 0.08 and reducing the maximum importance weight from 14.7 to 9.3, while effective sample size increased by 908. Student-level stratification indicated the strongest gains for mid mastery and mid engagement learners (mean improvement 0.046, median 0.044), with smaller but consistent benefits for high mastery learners driven by reduced repetition rather than correctness shifts. Ablation results confirmed that conservatism and pruning are complementary: removing conservatism increased tail risk and widened confidence intervals, while removing pruning increased evaluation variance despite similar mean return. These findings demonstrate that evidence-constrained offline reinforcement learning can produce deployable adaptive policies with measurable improvements and quantifiable safety guarantees under logged-data constraints.

Enterprise Knowledge-Grounded Human–AI Collaborative Feedback Generation Using Retrieval-Augmented Models

2026-06-16T20:54:00+07:00

Adaptive feedback generated by large language models often suffers from limited auditability and inconsistent pedagogical intent, which constrains trust in real learning deployments. This study proposes an explainable adaptive feedback framework that combines retrieval-augmented generation with pedagogical rationale tracing, linking learner-state signals, retrieved evidence, and instructional move sequencing into a traceable rationale artifact. Across four conditions, the full model improves faithfulness to evidence from 0.66 to 0.88 and reduces scope violations from 7.4% to 1.9%. Learning proxies improve concurrently, with next-item accuracy increasing from 71.2% to 78.9% and revision uptake rising from 34.6% to 47.8%. Human rubric scores confirm higher instructional usefulness, where actionability increases from 3.6 to 4.3 on a 5-point scale while tone remains stable from 4.1 to 4.2. Ablation results show retrieval as the dominant driver of grounding, with faithfulness dropping to 0.62 and scope violations rising to 8.1% when retrieval is removed, whereas removing rationale tracing mainly degrades actionability from 4.3 to 3.7 and revision rate from 47.8% to 39.5%. Stratified analysis indicates the strongest benefits for low mastery and high frustration learners, where next-item accuracy improves from 61.0% to 70.0% and revision rate increases from 45.0% to 58.0%, alongside persistence gains from 68.3% to 79.5%. Mitigation controls further reduce evidence mismatch from 9.6% to 4.1% and rationale incoherence from 6.8% to 2.9%. The findings indicate that grounded generation and explicit pedagogical rationale tracing jointly improve effectiveness, accountability, and deployment readiness of adaptive feedback systems.

Continual Learning for Human–AI Collaborative Learning Analytics under Behavioral Drift

2026-06-16T21:12:57+07:00

Semester-to-semester non-stationarity undermines the reliability of adaptive learning analytics, particularly when predictive models are deployed without explicit drift monitoring and controlled updating. This study analyzes a 14-semester longitudinal panel constructed from learning management system traces and assessment records, covering 18–21 distinct courses per semester and 812–936 active students per term. Drift is concentrated in performance-relevant behavioral channels, with the strongest intensity observed in practice attempts, submission timeliness, and session regularity, alongside a pronounced regime shift around the mid-sequence semester. Under semester-forward evaluation, a static model yields mean macro-F1 of 0.706 with a worst-semester macro-F1 of 0.652 and high volatility across semesters (std 0.030). Periodic retraining improves mean macro-F1 to 0.724 and worst-semester macro-F1 to 0.681 (std 0.022) but remains sensitive to update timing. Drift-aware continual learning achieves the highest and most stable performance, improving mean macro-F1 to 0.742 and worst-semester macro-F1 to 0.711 while reducing temporal variance (std 0.015) and increasing mean AUROC to 0.812. Reliability gains are reflected in lower expected calibration error (ECE 0.039 versus 0.056 for static) and improved decision quality at fixed intervention capacity, raising risk precision from 0.62 to 0.69 and risk recall from 0.48 to 0.56 while reducing alert volatility (CV 0.14 versus 0.29). Fairness robustness improves under drift-aware updating, reducing mean subgroup recall gap from 0.118 to 0.082 and lowering the maximum recall gap from 0.172 to 0.121. Ablation shows that intermediate drift thresholds balance robustness and governance load, sustaining worst-semester performance with approximately 1–2 updates per semester and diminishing returns beyond moderate replay memory.

Measuring the Impact of Human–AI Collaborative Personalized Interventions through Temporal Causal Inference

2026-06-16T21:30:01+07:00

Adaptive learning platforms frequently report performance improvements, yet many evaluations remain vulnerable to time-varying confounding because interventions are triggered by evolving learner states. This study evaluates three intervention families, adaptive sequencing, targeted hints, and remediation triggers, using a longitudinal causal framework with horizon-locked outcomes and learner-level cross-fitting. The analytic cohort includes 2,480 learners and 118,640 decision points observed across 12 instructional weeks, with median 41 decisions per learner. Intervention exposure rates per 100 decisions are 38.6 for sequencing, 24.1 for hints, and 8.9 for remediation, with higher targeting intensity in low-mastery strata. Causal estimates show distinct temporal signatures by intervention mechanism. Targeted hints yield the largest same-session improvement, increasing mastery by 2.4 points, but effects attenuate at 7 days (1.3 points) and 14 days (0.9 points). Adaptive sequencing provides more stable medium-horizon benefits, improving mastery by 1.6 points same-session, 2.8 points at 7 days, and 2.2 points at 14 days. Remediation triggers demonstrate delayed consolidation, increasing mastery by 1.1 points same-session, 3.4 points at 7 days, and 4.1 points at 14 days, albeit with wider uncertainty consistent with lower overlap and late-course concentration. Heterogeneity analyses at the 7-day horizon indicate sequencing peaks for mid-mastery learners, reaching 3.9 points under high engagement versus 3.4 under low engagement, while hints are most effective for low mastery with low engagement (1.6 points) and decline sharply for high mastery with high engagement (0.4 points). Remediation remains meaningful across strata, reaching 3.6 points for mid mastery with high engagement and 2.3 points for high mastery with high engagement, supporting a diagnostic targeting interpretation rather than uniform escalation. Robustness and diagnostic checks support internal validity. After weighting, standardized mean differences for key confounders fall to 0.05–0.09, and placebo effects on pre-decision outcome change remain near zero in magnitude (absolute value ≤0.05) across all intervention types. Overlap trimming of the lowest 5% support preserves the ranking of interventions, with only modest attenuation for remediation, and effective sample size remains adequate for sequencing and hints while declining for remediation in late decision indices. These findings justify a tiered deployment strategy where sequencing is the default optimization lever, hints are constrained to high-instability episodes and paired with post-hint practice allocation, and remediation is gated by high-confidence misconception signals with overlap and effective-sample-size monitoring.