Personalized exercise recommendation systems aim to enhance learning efficiency by dynamically guiding students toward content aligned with their evolving knowledge states. Among various approaches, Reinforcement Learning (RL) has emerged as an effective framework for modeling student-environment interactions as sequential decision-making processes. However, most existing RL-based methods typically reward recommendations that target unmastered knowledge concepts.Such reward-driven strategies are prone to local optima, restricting exploration of unattempted or less familiar knowledge concepts. In addition, they often overlook key learning factors, such as forgetting dynamics and exercise difficulty, leading to suboptimal outcomes. To address these issues, we propose a novelIntegratingEntropyRegulation andDual-objectiveOptimizationExerciseRecommendation (IERDO-ER)method. Specifically, we introduce an entropy-based function to encourage broader exploration in the exercise space. We also design an end-to-end policy network that generates candidate exercises as next-step actions for the recommendation agent. To enable more adaptive and pedagogically sound recommendations, we develop a dual-objective reward mechanism that incorporates both anti-forgetting and gentleness objectives. This mechanism continuously balances policy optimization through a synergy of rewards and penalties. Experiments on three real-world educational datasets show that our approach consistently outperforms strong baselines, validating its effectiveness.

Group recommendation services are an important way to foster collaborative learning among users and enhance learning effectiveness in online education. However, existing group recommendation methods fail to consider the diverse learning behaviors and collaborative relationships among members (users), which brings challenges in capturing group consensus. Group consensus is key to improving the accuracy of recommendations and capturing the real needs of members. In response to this issue, we propose a Multi-behavior Enhanced Group Recommendation (MEGR) model for smart educational services, which captures the multi-behavior features and collaborative relationships of members to model group consensus. Specifically, we design specialized interaction subgraphs to capture multi-behavioral features for learning members’ preferences. Additionally, we model the independence of each behavior to avoid over-relying on any single one. Subsequently, we model intra-group collaboration and potential dependencies by projecting the member’s preference representations into multiple different spaces, which are used to capture group consensus. The experimental results on the two real-world educational datasets, MOOCCube and EdNet, show that our MEGR improves group recommendation performance by an average of 0.97% and 2.10%, respectively, compared to the best baseline methods.

Recommender system is regarded as a crucial strategy for realizing personalized e-learning. To better enhance the learning outcomes for learners, exercise recommendation plays a vital role in this system. However, current modeling methods are insufficient to fully capture the dynamic process of learners’ knowledge concept forgetting. Moreover, recommending learning resources based solely on the learner’s current level, without exploring the learner’s learning intentions, leads to inefficient utilization of learning resources. To address these issues, this paper proposes a novel Intention-aware Exercise Recommendation enhanced by deep Forgetting modeling for e-learning (IERF). Specifically, we integrate the forgetting mechanism into multi-concept sequences, model the contextual dependencies of sequences, and design behavioral balancing factors in multiple dimensions to optimize the forgetting prediction of the learning process. Additionally, we construct a heterogeneous information network (HIN) of the learning process and design multi-attention mechanisms to highlight high-order learning relationships between networks and perceive the learner’s learning intentions. Experimental results on three public real-world datasets show that the proposed model outperforms state-of-the-art baselines, and enhances the interpretability under the exercise recommendation.

Knowledge Tracing (KT) aims to leverage students’ learning interactions to trace their knowledge state and accurately predict future learning performance. We observe that the learning process is carried out in stages, with each stage having different learning intention that drive students’ behavior and performance. In fact, within and across these learning stages, students’ knowledge state may vary due to variations in their learning intention. Most existing KT methods consider students’ learning interactions as a continuous process, overlooking the staged variations in students’ learning intention, leading to inconsistent representation of students’ actual knowledge state. To address this problem, we explore a new paradigm of KT and propose a novel model named Learning Intention-Aware Knowledge Tracing for Learning Stage (ISKT), which perceives the learning intention of staged variations to trace the students’ knowledge state. Specifically, we have designed a hierarchical intention-aware network, which separately mines the interaction relations within learning stages and the stage relations between the learning stages, to perceive learning intention at both the interaction and stage levels. This network also provides an effective representation of intention for the entire learning process by adaptively integrating these dual levels of learning intention. Additionally, to represent the staged knowledge state, we utilize Knowledge gain within learning stages and knowledge forgetting across learning stages to model the staged learning progress. We design an intention fusion method, which learns the fusion coefficient between learning intention and staged learning progress, and then performs fusion based on this coefficient. Extensive experimental results on public datasets demonstrate that ISKT outperforms state-of-the-art baseline models in predicting students’ future performance.

Massive Open Online Courses (MOOCs) are playing a key role in improving educational ways. Abundant learning resources make it difficult for online users to find suitable learning content. The current personalized service in the field of online education relies more on course recommendations. However, coarse-grained recommendations cannot help users discover the defects of their knowledge network effectively. In this paper, we propose a Knowledge-Aware Meta-Concept (KAMC) framework to provide fine-grained recommendation services. We innovatively incorporate Knowledge Graph (KG) into the field of educational recommendation to provide abundant auxiliary information. However, simply combining knowledge graphs with educational recommender systems cannot improve the performance of existing recommendation models, and may even weaken the performance of the models. Because the modeled KG ignores the enhancements on the user side and only considers the enhancements on the item side. We further propose to enrich the semantic representation of users with collaborative information in user-item interactions, and at the same time enrich the semantic representation of items with information in KG. Furthermore, to provide users with more accurate and fine-grained personalized recommendation services, we propose a user-based attention mechanism to capture users’ fine-grained semantic information. Our method is experimentally validated on three real-world datasets. In the three datasets, KAMC’s AUC evaluation index is $$6.2\%$$ , $$6.9\%$$ , and $$2.2\%$$ higher than the latest baseline method (KGAN), respectively. Experimental results show that the KAMC method outperforms the current state-of-the-art baseline methods.

Abstract Online education platform urgently needs recommendation methods to service learning groups. The existing group recommendation methods rely on member preference aggregation. However, potential factors such as teachers and schools have an important impact on the decision-making of learning groups. Due to the different ways in which potential factors operate, it is very challenging to model the impact these potential factors have on groups. In this paper, we propose a group recommendation method for online education (GEMst) that innovatively designs two modules to capture different types of potential factor embedding. In the subjective guidance module, we construct a heterogeneous information network based on the relationship between guiders (teacher, leader, etc.) and groups, analyze the influence of the guider on the group, and obtain the group representation from this perspective. In the objective environment hypergraph module, we consider the influence mode of the environment (school, platform, etc.) on the group, and propose the objective relationship hypergraph convolution. Compared with hypergraph convolution, it directly learns the influence between groups in the same environment. We provide information fusion strategies that GEMst can collaboratively consider group members’ opinions and the impact of potential factors. In addition, to enhance the learning effect of the model, we design a pre-training strategy based on the user grouping relationship that obtains more accurate user embedding representations. We test GEMst on two real-world datasets, and the results show that GEMst outperforms the current baseline group recommendation method.