Adaptive Learning with Attention-Based Knowledge Tracing and Risk Prediction for Improved Student Outcomes

Abstract

The objective of adaptive learning environments is to offer personalized learning experiences, and many current systems are constrained by fixed instructional strategies, noisy learner-interaction data, and a lack of attention to equity across diverse student groups. In this paper, a proposal is presented for an intelligent, machine-learning-based adaptive learning framework, incorporating Attention-based Knowledge Tracing (AKT) to estimate continuous mastery, a calibrated gradient-boosted model to predict risk in early learning, and a contextual bandit policy to recommend adaptive learning content. The suggested closed-loop architecture dynamically examines learner interactions, such as correctness, time-on-task, and engagement indicators, to provide personalized and fair learning interventions. The framework is assessed using 1,200 students and 185,000 interaction records from a classroom-scale dataset. The experimental evidence shows better results than baseline models with an accuracy of 92, a smaller knowledge tracing error (RMSE = 0.17), a better calibration of the probability (ECE = 0.031) and much less fairness gap (F1-gap = 0.05). The results of these studies demonstrate that the suggested framework is effective in improving learning outcomes and minimizing learning differences, underscoring its applicability to online and blended learning.