Exponential growth of ICT, MOOCs, and Learning Management Systems (LMS), online learning resources have exploded

The gap

Evidence profile

Sourced from the future work and recommendations of the source papers, classified as general, spanning 3 journals.

Supporting evidence — 6 representative gaps

• An Intelligent E-Learning Framework for Personalized Educational Recommendations in Secondary Education through Deep Q-Learning Optimization (2026) · Zenodo (CERN European Organization for Nuclear Research) · doi

  This research set out to design and evaluate an intelligent e-learning framework using DQN to provide personalized educational recommendations for secondary-level students. The motivation stemmed from the fact that traditional learning approaches often fail to adapt to individual learners’ needs, resulting in disengagement, low retention, and inefficiency in learning outcomes. Through careful modeling of the learning environment as a MDP with states representing different learning behaviors and actions reflecting educational strategies the framework enabled adaptive decision-making that evolved over time. The experimental results provided strong evidence of the effectiveness of DQN compared to Tabular Q-Learning and Rule-Based methods. DQN consistently achieved higher rewards across all milestones (10,000 to 50,000 episodes), stabilizing around 12-14 reward points, while Q-Learning remained around 5-6 and the Rule-Based agent around 5. This performance gap highlights the value of deep function approximation in handling complex, dynamic learning environments where traditional RL methods fail to scale. In an educational context, this means that the DQN-powered system can better understand student behaviors, adapt to changes in motivation or knowledge gaps, and provide personalized recommendations that promote engagement and mastery. Unlike static rule-based strategies, which cannot evolve, or Q-Learning’s limited state-action handling, DQN offers a scalable solution aligned with real-world personalized learning needs. Several directions for future research can expand upon this work: • Integration with Real Student Data: Incorporating real learning behavior datasets to validate and fine-tune the framework. • Hybrid Models: Combining DQN with other techniques (e.g., LSTM for sequential learning or attention mecha- nisms) to better capture long-term dependencies in student behavior. • Adaptive Reward Mechanisms: Designing more dynamic reward functions that account for learner satisfaction, emotional state, and long-term academic growth. • Deployment in E-Learning Platforms: Implementing the framework into actual digital learning environments to evaluate its effectiveness in real classrooms. • Cross-Level Generalization: Extending the system beyond secondary education to higher education or profes- sional learning contexts.

  generalfuture work
  Keywords: learning framework real personalized educational rule based around reward student evaluate provide recommendations secondary level
• Retraction notice: Efficient Course Recommendation using Deep Transformer based Ensembled Attention Model. (2026) · ICST Transactions on e-Education and e-Learning · doi

  learner [14][15][16] preferences has become the major research endeavour to address these difficulties, as the proliferation of online courses has made it difficult for students to choose those that best suit their needs and interests. CF algorithms [19,20] are used in conventional course recommendation systems [17,18] to uncover latent information about a user's interests. Using neural recommendation algorithms grounded in deep learning techniques [21,22], these approaches deliver solid results. The Neural Attentive Recommendation Model (NARM) [23] is one such model that can track users' sequential actions and extract their primary goals for taking the class. In addition, the fundamental recommendation model based on attention networks is trained in tandem with the Hierarchical Reinforcement Learning (HRL) model [24], which reduces noisy courses through learner profile model. When consumers have signed up for a wide variety of classes, the course recommendation performances have room to grow. However, HRL fails to deliver satisfactory results since it disregards the user's stated preferences. The adaptability of the recommendation model allows for a number of methods to be used in assessing learner behaviour [25] on e-learning systems. For instance, Chen et al.'s [26] hybrid recommender model is one such example. It does this by employing item-based CF to unearth groups of pertinent are subsequently filtered by a sequential pattern mining algorithm in consideration of commonalities in the sequence of study. Furthermore, Wan and Niu [27] proposed a recommendation model, wherein learning objects are simulated as intelligent object entities using the self-organization theory, and the objects in these simulations naturally interact with one another. self-organization-based things, which Despite their widespread use in course recommendation, the aforementioned approaches all share a critical flaw: they fail to account for students' ever-evolving tastes and requirements as they progress through courses. In other words, these techniques aren't great at extracting the user's choice in every interaction, especially if the learner is interested in a wide variety of subjects and their preferences shift over time. In this scenario, the adaptivity of these techniques are not very good at keeping up with consumers' shifting tastes. the recommendation model low since is The remainder of the paper is structured as follows: In Section 2, a concise review of course recommendation is provided. The proposed technique is then described in detail in Section 3. We conducted experiments using actual data and reported the outcomes in Section 4. Section 5 concludes the document. literature

  generalrecommendations
  Keywords: recommendation model learner course learning preferences courses user using techniques based students interests algorithms used
• Retraction notice: Efficient Course Recommendation using Deep Transformer based Ensembled Attention Model. (2026) · ICST Transactions on e-Education and e-Learning · doi

  model, wherein learning objects are simulated as intelligent object entities using the self-organization theory, and the objects in these simulations naturally interact with one another. self-organization-based things, which Despite their widespread use in course recommendation, the aforementioned approaches all share a critical flaw: they fail to account for students' ever-evolving tastes and requirements as they progress through courses. In other words, these techniques aren't great at extracting the user's choice in every interaction, especially if the learner is interested in a wide variety of subjects and their preferences shift over time. In this scenario, the adaptivity of these techniques are not very good at keeping up with consumers' shifting tastes. the recommendation model low since is The remainder of the paper is structured as follows: In Section 2, a concise review of course recommendation is provided. The proposed technique is then described in detail in Section 3. We conducted experiments using actual data and reported the outcomes in Section 4. Section 5 concludes the document. literature

  generalrecommendations
  Keywords: recommendation model objects using self organization course tastes techniques wherein learning simulated intelligent object entities
• Retraction notice: Efficient Course Recommendation using Deep Transformer based Ensembled Attention Model. (2026) · ICST Transactions on e-Education and e-Learning · doi

  exponential growth of ICT, MOOCs, and Learning Management Systems (LMS), online learning resources have exploded. Due to disparities in cognitive ability and preferences, learners are unable to rapidly identify and select the learning resources in which they are interested and required [28]. Therefore, it is imperative to develop an intelligent model that accurately and efficiently recommends useful and engaging learning resources to students. In this section, we examine extant relevant research in the following fields: (1) Conventional algorithms for course recommendation consist of context-based, content-based, collaborative filtering (CF), and hybrid recommendations. recommendations based on Artificial (2) Course 2 EAI Endorsed Transactions on e-Learning | Volume 9 | 2023 |RETRACTED Efficient Course Recommendation using Deep Transformer based Ensembled Attention Model (AI) and Machine Learning

  generalrecommendations
  Keywords: learning based resources course model recommendation recommendations exponential growth moocs management systems online exploded disparities
• Retraction notice: Efficient Course Recommendation using Deep Transformer based Ensembled Attention Model. (2026) · ICST Transactions on e-Education and e-Learning · doi

  environments, World Wide Web 17 (2) (2014) 271–284. [26] S. Wan, Z. Niu, An e-learning recommendation approach learning resource, based on the self-organization of Knowl.-Based Syst. 160 (2018) 71–87 algorithm adapted in [27] J. Zhang, F. Gu, Y. J. Ji, and J. F. Guo, “Personalized scientific resources technological recommendation based on deep learning,” Journal of Intelligent Fuzzy Systems, vol. 41, no. 2, pp. 2981–2996, 2021. literature and [28] Zhang B, Yang FZ, Fan LN. Query recommendation based on document content user focuses on. Appl Mech Mater 2014;511:385–8 J, Kim for SB. Content-based recommendation systems using multiattribute networks[J]. Expert Syst Appl 2017;89:404–12 [29] Son filtering [30] Ng YK. CBRec: a book recommendation system for children using the matrix factorisation and content-based filtering approaches[J]. Int J Bus Intell Data Min 2020;16(2):129–49. [31] Sakboonyarat S, Tantatsanawong P. Massive open online courses (MOOCs) recommendation modeling using deep learning[C]//. the 2019 23rd international computer science and engineering conference (ICSEC). IEEE; 2019. p. 275–80 In: Proceedings of [32] Verbert K, Manouselis N, Ochoa X. Context-aware recommender systems for learning: a survey and future challenges[J]. IEEE Trans Learn Technol 2012;5(4): 318– 35 [33] Zapata A, Menendez VH, Prieto ME. A hybrid recommender method for learning objects [J]. In: IJCA Proceedings on design and evaluation of digital content for education (DEDCE). 1; 2011. p. 1–7. [34] Salehi M, Kamalabadi IN, Ghoushchi MBG. An effective recommendation learning environments using a learner preference tree and a GA[J]. IEEE Trans Learn Technol 2013;6(4):350–63. framework personal for [35] M.L. Littman, Reinforcement learning improves behaviour from evaluative feedback, Nature 521 (2015) 445–451. 10 [36] D. Silver, T. Hubert, J. Schrittwieser, I. Antonoglou, M. Lai, A. Guez, M. Lanctot, L. Sifre, D. Kumaran, T. Graepel, et al., A general reinforcement learning algorithm that masters chess, shogi, and go through self-play, Science 362 (6419) (2018) 1140–1144. [37] A. Kara, I. Dogan, Reinforcement learning approaches for inventory specifying ordering policies of perishable systems, Expert Syst. Appl. 91 (2018) 150–158. [38] X. Wang, Y. Wang, D. Hsu, Y. Wang, Exploration in interactive personalized music recommendation: A reinforcement learning approach, ACM Trans. Multimed. Comput. Comm. Appl. 11 (1) (2014) 22 [39] F. Liu, X. Li, H. Guo, R. Tang, Y. Ye, X. He, End-to-end deep reinforcement learning based recommendation with supervised embedding, in: Proceedings of the 13th ACM International Conference on Web Search and Data Mining, 2020, pp. 384–392. [40

  generalrecommendations
  Keywords: learning recommendation based reinforcement systems content appl using syst deep ieee proceedings trans wang environments
• Large-Scale Classroom-Based Social Recommendation Model Using Heterogeneous Graph Neural Networks and Social Regularisation (2026) · Contemporary Education and Teaching Research · doi

  the systems mainly traditional recommendation paradigm. Collaborative filtering methods generate recommendations by analyzing groups of learners with similar ratings or learning behaviors. Its basic assumption is that users learning with similar the resources ubiquitous data cold-start problems in MOOCs make the performance of traditional collaborative recommendation methods based on ratings limited, and new learners or new courses lack sufficient historical interaction data to find reliable similar neighbors (Najafabadi et al., 2026). Content-based recommendation focuses on the matching of learning resource metadata and learner profiles, but this method is highly dependent on high-quality and structured resource annotation, which is difficult to guarantee in the open and dynamic MOOCs environment, and easily leads to the lack of diversity and surprise of recommendation results. In order to overcome the shortcomings of a single method, hybrid recommendation strategies trying to combine collaborative have emerged, recommendation, other information sources. For example, some studies have integrated the recommendation process, or used deep learning models from multi-source data (Chen, 2025). representations content-based recognition learning learn style joint into and to Recognizing the key role of social factors in learning, system has recommendation gradually become an independent and active research social 15 a by and prior social circles such as framework introducing (Jdidou et factorization branch. Its core idea is derived from an intuitive sociological principle that users are more likely to accept and adopt the suggestions of trusted members in their al., 2025). Incorporating social relationship information into the recommendation algorithm is proved to not only improve the accuracy of recommendation, but also effectively alleviate the problem of data sparsity and cold start. Early classical works, the trust-based recommendation proposed by Massa and Avesani, use explicit trust networks among users for propagation and prediction (Simone et al., 2014; Avesani et al., 2024). Golbeck’s TidalTrust model calculates indirect trust through the aggregation of trust paths (Seo & Han, 2010). These pioneering studies have laid the foundation for trust-aware recommendation. Subsequently, the combination of matrix social information has led to a series of important models. The Social Regularization (SoReg) model proposed by Sun et al. encodes the social relationship as a smoothing social regularization term in the objective function to constrain the latent feature vectors of friend users to be as similar as possible (Sun et al., 2022). The TrustSVD model further developed by Chen et al. (2025), innovatively considers both the explicit and implicit effects of user ratings and trust relationships, extends on the SVD++ framework, enriches user representations by modeling implicit feedback from trusted users, and demonstrates superior

  generalrecommendations
  Keywords: recommendation social learning trust users similar based collaborative ratings information model traditional learners cold start

Questions about this gap