A Cognitive Trust-Aware Framework for Minimax Optimization of Personalized Learning Paths in Higher Education

Abstract

Adaptive learning in higher education depends on two capabilities that the literature has developed in isolation. One is an accurate estimate of what a learner knows. The other is a policy that decides what the learner should do next. Knowledge tracing has advanced the first, most recently by suppressing non-cognitive noise such as slips and guesses through adaptive dynamic programming. Learning path optimization has advanced the second, formulating path planning as a sequential decision process over a knowledge graph and solving it with value-based deep reinforcement learning. Stacked naively, the path optimizer inherits a contaminated knowledge state. A tracing model pays a bounded one-step penalty for that contamination, whereas a reinforcement learning agent compounds the error along the entire trajectory. This article proposes the Cognitive Trust-Aware Personalized Learning Framework (CTPLF) to close the interface. An Adaptive Cognitive State Refinement (ACSR) module recovers the coherent cognitive representation using a p mechanism driven by an adaptive dynamic programming solver, and it also emits a per-concept Cognitive Trust Score derived from the density of optimal control corrections the solver applied. A Cognitive Trust Guided Learning Path Network (CTGLPN) consumes that signal at three gates covering replay sampling, exploration scale, and reward shaping, so the agent neither learns from nor rewards gains the cognitive layer cannot vouch for. An Educational Decision Intelligence Layer (EDIL) aggregates trajectories into segment level indicators. We evaluate the framework end to end on the Open University Learning Analytics Dataset against the existing causal reinforcement learning system validated on the same corpus.