Personalized Learning Path Optimization Based on Generative Artificial Intelligence

Learning paths personalized by behavior have become common among personalized learning systems. However, the traditional methods are usually inefficient at building coherent paths as the learner's knowledge state changes dynamically, and the prerequisite structure exists. In light of this problem, we present a generative learning-path optimization framework based on EdNet-KT1 interaction dataset, Transformer knowledge tracing model, Qwen2.5-7B-Instruct candidate generator, and constraint-aware reranking. Mastery profiles of learners are extracted, and then five ten-resource candidate paths are generated and sorted according to learning gain, prerequisites, difficulty deviation, and repetition. We compare our method to collaborative filtering, GRU4Rec, and knowledge-tracing greedy recommendation on offline experiments. The generative optimizer gets NDCG@10 of 0.589, future-concept coverage of 0.721, and prerequisite violation rate of 0.031, and thus outperforms the best baseline on all three metrics. Ablation analysis shows that the prerequisite violation rate rises to 0.091 without constraints, and the learning gain drops to 0.104 without the learning gain term, respectively. Moreover, the full model still achieves better NDCG@10 performance on all learner history categories than other baselines, costing 0.84 s per learner with the memory peak of 13.6 GB. This research demonstrates the effectiveness of combining the generation of learning paths with pedagogical constraints.