Over the past decade, e-learning systems have grown rapidly because they are flexible, accessible, and can deliver learning at scale to millions of users around the world. Worldwide evidence has demonstrated an increased use of online learning in higher education and corporate settings, with more than 110 million learners registered on major MOOC platforms as of 2022 (Kizilcec et al., 2013). Despite the increase in the number of students joining MOOCs, learner disengagement is still a common issue—MOOCs have very low completion rates that range between 5 and 10% (Wong, 2020). Moreover, studies have revealed that 90% of students drop out during different stages of an online class because they lack a personalization experience, little interaction with their peers, and have low support (Linn and Clark, 2019). These results also indicate a significant drawdown in the state-of-the-art systems: recommender techniques need to be better customized to meet the changing needs, behaviors, and contexts of learners.

Traditional e-learning recommenders are commonly based on collaborative filtering (CF), such as users’ interests and history about items. Although CF techniques have been effective for content recommendation, they are inherently predicated on the stability of preferences and rarely take into account contextual, temporal, and behavioral variability that influence learning paths (Jannach and Manzoor, 2020; Chen and Xu, 2021). In reality, learning is not a static process, and students’ experience, goal orientation, motivation, and cognitive patterns evolve over time, with such changes affecting the way they can or cannot benefit from collaborative learning environments. Meanwhile, the current e-learning systems also regard learners as a whole-oriented group without considering job type, level of knowledge, skill, and learning style or even approaching pattern (Bower, 2019; Koren et al., 2019). Research in the area of digital pedagogy suggests that successful online learning is heavily reliant on cooperation, social presence, and sensitive support with respect to the context, not simply similar content (Zhang et al., 2019; de Jong and van der Linden, 2020). These gaps reinforce the need for adaptive, context-sensitive, and behavior-aware recommendation systems.

To address these limitations, in this paper, we implement a hybrid CF model, which is a combination of collaborative filtering and content-based recommendations to recommend peers and study material by utilizing historical interactions, preferences, and performance of learners. The existing dynamic model is augmented with interaction-based clustering, which assimilates learners based on their interactions, learning styles, and development needs within the same cluster. The model promotes the creation of context-oriented study groups, both true to real-time behavior and the demand of learners, tailored toward individual learning progress.

Our model uses BERT embeddings to enhance the semantic understanding of the content and thereby the quality of recommendations. Bidirectional encoder representations from transformers (BERT) is a widely used pre-trained language model that learns semantic meaning from text data. With the help of content indexes, learner profiles, and course themes, it is possible to make recommendations more contextualized to specific learners during their pursuit of certain learning paths, job roles, and/or expertise.

We also propose a new way of measuring the progress of knowledge evolution and learning over time. Watch percentage, time on content, and ratings have been considered by our system to dynamically adapt learning paths and difficulty levels, so it always keeps learners challenged without overwhelming them. This ability to track learning progress and adjust recommendations accordingly sets our model apart from traditional static recommendation systems.

This paper makes the following contributions:

Previous studies have predominantly used static or similarity-based grouping of learners, without considering changes in learner behavior over time and the impact of context (such as job role, expertise level, and course difficulty) on peer compatibility. Our approach brings in an interaction-driven clustering mechanism that evolves peer groups depending on temporal learning dynamics and situational properties. This allows for more meaningful and pedagogically relevant peer recommendations that fill a key gap in prior CF-based systems, which assume learners are a homogeneous, immutable population.

Text-based recommenders either rely on the shallow textual features or keyword overlap between documents that may not reflect the underlying semantic relations for a personalized e-learning. BERT-based embeddings are combined with other contextual information (theme, job role, software proficiency), resulting in a multi-faceted representation of learners and content. Where the limitations of previous works that utilize isolated content features without deep semantic information are thus overcome, and a more accurate peer/learning resource recommendation solution is achieved.

Several previous systems do not take into consideration the dynamic hints of progress learning, such as watch percentage and time spent, rating updated. Our model leverages these behavioral signals as a basis to scale content and personalize learning paths. This contribution will help address the limitation in previous recommendation models that are unable to personalize recommendations according to progresses of learners’ proficiency and engagement.

In earlier studies, even when they do, the benchmark was fairly small or partial evaluation metrics were used, and our approach compares the hybrid model against popular methods such as collaborative filtering (CF), matrix factorization (MF), and neural collaborative filtering (NCF). We demonstrate that the hybrid model consistently outperforms standard approaches by analyzing accuracy, precision, recall, and F1-score. This assessment demonstrates the utility of incorporating behavioral, contextual, and sematic elements; an aspect not well investigated in previous studies.

This concept is our effort to improve the peer learning experience and learner engagement, thereby improving the overall success of e-learning platforms by addressing these keys aspects. The rest of this paper is organized as follows: Section 2 reviews related work in the field of recommendation systems for e-learning, Section 3 outlines the methodology used to develop and evaluate the model, Section 4 presents the experimental setup and results, and Section 5 concludes the paper with future research directions.

A Systematic Review: Deep Learning-based e-Learning Recommendation System Perform a comprehensive review study (Bhanuse and Mal, 2021) on deep learning utilization and limitations in e-learning recommendation systems. They enumerate seven principal recommendation strategies: content-based, collaborative filtering, knowledge-based, demographic, hybrid, ontology-based, and context-aware recommendations, all of which have a unique set of strengths and weaknesses. In particular, the review highlights CNNs and RNNs in sequential and high-dimensional data settings. The authors explain that CNNs are effective in recognizing and extracting patterns from large data sets, a quality that positions them well for applications where the recommendation quality depends on identifying subtle variations in learner behaviors and preferences.

In “Personalized Recommender System for e-Learning Environment,” Khanal et al. (2019) introduced a recommendation system designed to address the static nature of traditional e-learning platforms, which often overlook individual learner differences. NPR_eL adopts a hybrid approach that combines collaborative and content-based filtering techniques, enabling it to tailor recommendations based on a learner’s preferences, interests, and background knowledge. To personalize recommendations further, the authors introduce an innovative factor: learner memory capacity. Memory span is assessed through a pre-test, allowing NPR_eL to account for each learner’s cognitive load and ability to retain information. This adaptation enables NPR_eL to avoid overwhelming learners and instead aligns with their unique learning pace, thereby improving the overall effectiveness of the e-learning environment.

Prior studies explained using processing power-heavy CNNs and RNNs, along with providing only the usual features (Liu et al., 2022). Our project takes a step forward by applying BERT embeddings that can learn fine-grained, multi-dimensional relationships between user interactions—content descriptions—contextual factors for any image-based recommendation task, leading to better recommendation precision. In contrast to CNN and RNN, which are designed primarily for linear or sequential data, BERT provides a multilayered representation of textual data where each layer transforms the input into a new output, able to capture bidirectional context and relationships across different themes, software, and job roles in our dataset. This allows our model to provide a contextualized recommendation that is aligned with an individual learning path preference and peer group preferences, but the latter is an important factor in e-learning (not represented directly in taxonomy ) (Bhanuse and Mal, 2021).

By enhancing upon the deep learning approaches described in their paper by incorporating peer-group clustering and creating ongoing skill tracking, which are missing from their taxonomy, our work advances classical solutions. Though CNNs and RNNs are good at pattern recognition and sequential learning, our model also has real-time clustering as well as multi-factor embeddings (Zhang et al., 2021) to adapt the system to changes in user preferences over time and learning progression. Our methodology overcomes static data and cold-start limitations through the embedding of diverse user attributes and evolving learning needs into our recommendation system, which provides integrated, holistic learning experiences tailored to different aspects of the complexity profiles of each user.

NPR_eL provides a rich personalization within the scope of memory capacity and cognitive load for each learner; however, our project takes this one step further with an innovative multi-factor model that includes contextual factors such as job roles, software skills and theme based interest(s), as well as flexibility to dynamically group peers based on their preferred role in learning (Ibrahim et al., 2020). Our approach enables the system to recommend a path for learning by clustering learners with similar behavior and goes beyond the memory-based transfer of knowledge that NPR_eL possesses. This evolution allows a continuous pathway for recommendation. The use of BERT embeddings also embodies a diverse set of user features in a compact and independent vector space, enabling more accurate recommendations that consider the context surrounding some choice or event (NPR_eL does not tackle the learner evolution).

Rahayu et al. (2022) builds on NPR_eL but extends its adaptation class to incorporate not only “what the user remembers” (i.e., cognitive factors) but also “why the user needs what they remember” (i.e., contextual factors), enabling our system to offer content that is aligned with memory retention, professional relevance, and domain mobility in a dynamically changing field. Unlike NPR_eL, our model provides cross-domain recommendations due to the combination of multi-factor BERT embeddings and collaborative filtering that we integrated; meanwhile, learners can seek interdisciplinary content based on their real-time preferences and peer interactions. That level of customization makes our model ideal for the variety and dynamism of e-learning environments, where flexibility and adaptability are key.

The adaptation class of NPR_eL is only limited to the “what the user remembers” (i.e., cognitive factors), while our approach also includes the “why the user needs what he/she recalled” (i.e., contextual factors) making it more suitable for context-aware content provisioning in a dynamically changing field where such relevant information is desirable along with memory retention, professional relevance and domain mobility. Different from NPR_eL, our model offers cross-domain recommendations that are facilitated by the merged multi-factor BERT embeddings and collaborative filtering adopted together in Ali et al. (2022); while allowing learners to pursue interdisciplinary content via their word-sensitive demand and peer abilities. It was this level of customization that makes our model the best fit for e-learning environments, considering the diversity and ever-changing nature of such contexts.

While, as highlighted ontology contribution to organizing both user and content knowledge, our model utilizes BERT representations to obtain an equally expressive representation of user and content characteristics on multiple levels without the effort behind manually built ontologies. Using BERT in this way enables our system to reason over the different learner characteristics—language, job type, and software choices, for example—allowing matching of knowledge with similar efficiency but increasedflexibility (Jeevamol and Renumol, 2021). Unlike most ontology-based systems, our model not only offers recommendations but also includes features to cluster peers based on interaction data, subjectively adding an adaptive quality (via peer dynamics) that is typically absent in static ontology-based systems.

Bhaskaran et al. (2021) and Kulkarni et al. (2020) represented an advance over ontology-based methods through adapting to the behavior of learners changing and peer groups changing in real time. This means fine-grained semantic nuances and more intricate user profiles are covered using BERT-based embeddings with significantly less human effort than ontologies would require, producing a solution that is scalable in nature for online education systems and adaptable to new learners and content. Additionally, our solution minimizes the cold-start and sparsity problems by merging contextual embeddings with collaborative filtering to provide users with personal recommendations that elicit change as the profile of the user grows and evolves—providing novel dynamic point-based personalized recommendations that traditional ontology-based models struggle to achieve (Jordan, 2014).

This section introduces the full approach to creating the hybrid recommendation model of dynamic peer learning within e-learning platforms. We combine hybrid CF and interaction-based clustering (Palvia et al., 2018) with BERT embeddings, as well as dynamic peer grouping, to provide context-aware adaptive learning recommendations, as shown in Figure 1. It continuously evolves in accordance with individual learning profiles, thus keeping the suggestions contextual throughout the user’s learning lifespan.

This section explains the experimental setup for evaluating our hybrid recommendation system model. Implementation details, dataset preprocessing, and evaluation metrics are reported in detail, along with baseline models used to compare our method across multiple settings.

The effectiveness of four recommendation methods—collaborative filtering (CF), BERT embeddings (BE), peer group recommendations (PGR), and hybrid model—was measured based on our top-K metrics visuals, i.e., Precision@5, Recall@5, and F1@5. Although it is in general not recommended to trust accuracy for recommendation systems due to the sparsity of user-item interactions, we also present it as a complementary evaluation metric. In this specific case, accuracy is also a further measure of model reliability as it measures the overall concordance of predicted interactions to predictions made on the whole dataset when compared to precision (an indirect measurement for relevance).

The proposed hybrid model was specifically developed by focusing on scalability and computational efficiency. Collaborative filtering works on a precomputed user-item interaction matrix, BERT embeddings are computed offline for content items, and learner peer groups are dynamically K-mean clustered—such scaling linearly in the number of learners. The Annoy index allows searching through content embeddings fast enough to fetch recommendations in near real time. As a result, the approach can be used for recommendations in large-scale e-learning systems with low latency while guaranteeing the trade-off between recommendation quality and applicability.

The overall low recall values across models (0.01–0.06) are expected given the sparse and heterogeneous context of e-learning data, in which users access a very limited portion of content items. This is not specific to our work but rather a known issue in educational recommender systems. The challenge we encompass to solve is tackled by our multi-factor hybrid model that encompasses user interaction data, semantic embeddings, and peer grouping in order to maximize precision at the expense of interpretability.

Finally, this study makes enhancements to the literature at several levels. The contribution of this paper is 2-fold. First, it shows that hybrid and multi-factor recommendations can be successful in educational environments, where content relevance is more important than covering all available items. Second, it adopts dynamic peer grouping and BERT-based content embeddings to diversify its recommendations. Third, we present our results based on precision, recall, F1 score, and accuracy to address these gaps in prior studies with an evaluation framework that is both transparent and reproducible. In sum, these methodological and practical inputs underscore the robustness and real-world relevance of the hybrid model.

In summary, the hybrid model we propose has strong precision as well and is relevant and methodologically novel; for its value and publishability, it deserves to be developed in future research to improve recall. The model is especially useful in adaptive e-learning systems where it is more beneficial to recommend highly related content rather than covering all possible corpus.

Table 1 presents paired t-test results comparing performance metrics between the hybrid model and the baseline model. Reported values include mean difference, t-statistics, degrees of freedom (df), p-values, 95% confidence intervals, and Cohen’s d effect sizes. Significant p-values (<0.05) indicate meaningful improvements.

Table 2 reports top-K ranking performance metrics across four peer recommendation models. Precision@5 and Precision@10 indicate early precision, Recall@10 measures the fraction of relevant peers recovered, and NDCG@10 reflects ranking quality based on graded relevance. The hybrid model consistently outperforms all baselines across all top-K metrics, demonstrating superior relevance and ranking effectiveness.

We have introduced a new hybrid dynamic peer-learning recommendation system in e-learning settings that combines collaborative filtering (CF), interaction-based clustering, and BERT embeddings. The current research work addresses the limitation by proposing a system that adapts peer grouping dynamically at run-time based on varying contextual features such as difficulty level of learning, job role, software proficiency, and temporal learning trend adapting to context, thereby providing a personalized and context-aware experience. Moreover, the model monitors how learner skills change over time and updates its recommendations in real-time so that learners are always shown the most relevant and appropriately difficult material for what they need as their needs evolve.

Experimental results show that our method achieves superior performance over conventional sequential recommendation approaches. In particular, we had a higher accuracy, precision, recall, and F1-score than the best results so far, showing that our system was providing more correct and pertinent recommendations. Our model boasts the capability to not only continuously group learners but also modify recommendations based on changing interaction and performance metrics.

This research makes the following key contributions:

The implications of the findings of this study are impactful for e-learning platforms in providing adaptive and personalized experiences based on individual as well as group profiles. Utilizing a hybrid, context-aware recommendation system allows educational platforms to enrich personalized learning, enhance collaboration, and optimize academic performance.

There are several avenues toward future studies in this area that could help improve the proposed recommendation system. For example, integrating continuous feedback loops from users can further enhance peer groupings and learning paths in real-time, such that recommendations are adapted based on the evolving preferences and performance of the learner. Even more immediate and tailored suggestions could be achieved with in-the-moment adaptation by incorporating fine-grained data regarding user interaction and performance. It should also check the scalability of the system to handle a large-scale dataset and quickly recommend as more users and learning content are created. Moreover, the model could be generalized for cross-domain recommendations where we can recommend relevant content of another domain (for example, allowing AI resources to data science learners). Finally, supporting multi-modal data sources such as video, textual, and behavioral data is another option to boost the system’s understanding of user needs and recommendation performance.