Highly goal-oriented learning path recommendation with the forgetting curve

Learning path recommendation aims to organize learning items in an appropriate order to help learners achieve their learning goals. Reasonable planning of the concept learning sequence and dynamic matching of suitable learning resources are critical for boosting recommendation performance. Existing reinforcement learning-based methods usually select all prerequisites of a target concept as recommendation candidates without screening, leading to redundant or excessively long learning paths. Moreover, exercise recommendation based on knowledge states assessed by knowledge tracing suffers from training instability due to shifting response data distributions across epochs, resulting in a mismatch between the difficulty of recommended exercises and the learner's concept mastery level. To address these issues, this paper proposes a highly goal-oriented method named IFLPR. We design a selection strategy based on the forgetting curve and centrality-based algorithm to filter prerequisite concepts, retaining only those with high contributions to mastering the target concept as candidates. The difference between the initial knowledge state assessed by cognitive diagnosis and knowledge tracing model is introduced as a loss to stabilize model training, ensuring better exercise difficulty-concept mastery level matching. Compared with the SOTA methods, experimental results show that IFLPR improves the goal promotion by 33.24% under a 20-step constraint, while reducing the required recommendation steps to achieve the learning goal by up to 41.45%, highlighting its superior efficiency.