A Hybrid Recommendation Approach for Adaptive Worksheet Generation Using Pedagogically Structured Learning Objects

Abstract

Adaptive recommendation mechanisms are widely used to personalise digital learning environments; however, many existing approaches prioritise algorithmic optimisation while providing limited insight into how recommendation behaviour aligns with pedagogically structured instructional artefacts, such as worksheets. To address this gap, this paper proposes a hybrid recommendation approach for adaptive worksheet generation that integrates content-based and collaborative filtering with explicit pedagogical constraints derived from Bloom’s Revised Taxonomy. The system ranks and selects learning and evaluation objects across cognitive levels by combining learner profiles, behavioural signals, and similarity-based information within a unified scoring framework. A simulation-based evaluation was conducted to examine the internal behaviour, stability, and instructional alignment of the recommendation engine under controlled conditions, using Bloom-aligned worksheets and synthetic learner profiles. The analysis focuses on expected–actual alignment and adaptive variation across cognitive levels rather than learning outcomes. Results indicate strong alignment with the intended instructional structure at lower cognitive levels, while bounded and interpretable adaptive variation emerges at higher levels. Evaluation object recommendations showed high agreement with the instructional design, exceeding 95% across simulated conditions. Overall, the study demonstrates how hybrid recommendation mechanisms can support adaptive content selection in pedagogically structured learning scenarios, offering a transparent and robust foundation for information-driven educational systems.

1. Introduction

Digital learning environments increasingly incorporate adaptive mechanisms to address learner diversity and support more personalised forms of instruction [1]. While advances in educational technology have enabled large-scale access to digital content, many instructional activities—particularly practice-oriented materials such as worksheets—remain largely static in their design [2]. As a result, educators are often required to manually adapt learning materials to accommodate differences in learners’ cognitive readiness, preferences, or modes of engagement, a process that is difficult to sustain within everyday teaching practice [3,4].

In response, adaptive learning and educational recommender systems have been widely explored as means of personalising learning experiences [5]. Such systems typically rely on learner-related signals, including performance, preferences, or behavioural patterns, to tailor content recommendations [6]. Although these approaches have demonstrated potential in supporting engagement and resource selection, they frequently prioritise technical adaptivity over instructional structure, resulting in personalised outputs that are responsive to learner data yet weakly anchored to clearly articulated learning objectives or pedagogical intent [7].

Bloom’s taxonomy provides a well-established framework for organising learning activities according to cognitive complexity, from basic knowledge recall to higher-order processes such as evaluation and creation [8]. Despite its central role in instructional design, Bloom’s taxonomy is rarely operationalised explicitly within adaptive worksheet generation systems [9]. In many implementations, adaptive mechanisms select or rank learning resources without systematically accounting for the cognitive level at which learners are expected to engage, particularly across sequences of practice and assessment activities. This disconnect limits the extent to which adaptivity can support both personalisation and coherent instructional progression [2].

Recent research in educational recommender systems has increasingly focused on hybrid approaches that combine multiple sources of evidence, including content characteristics, learner behaviour, and pedagogical constraints [10]. By integrating rule-based logic with similarity-based or data-driven components, hybrid systems aim to introduce adaptivity while preserving instructional structure and transparency [5]. Within educational contexts, this balance is particularly important, as purely data-driven adaptation may lead to recommendations that are difficult for educators to interpret or align with instructional goals. Nevertheless, empirical examinations of how hybrid recommendation mechanisms behave when applied to structured instructional artefacts—such as worksheets organised across distinct cognitive levels—remain limited [10,11].

The present study builds on this line of work by examining a hybrid recommendation approach for adaptive worksheet generation that is explicitly structured around Bloom’s taxonomy. The system generates learning object and evaluation object recommendations for each cognitive level, with the aim of balancing instructional consistency and controlled forms of adaptivity. Rather than treating adaptivity as an end in itself, the focus is placed on how adaptive variation emerges across different Bloom dimensions and whether such variation remains pedagogically interpretable and aligned with instructional intent.

To investigate these questions, the study adopts a simulation-based evaluation focusing on the internal behaviour of the recommendation engine rather than on learner outcomes [12]. Using a controlled set of worksheets and synthetic learner profiles, the analysis examines expected versus actual recommendations across Bloom levels, with particular attention to instructional alignment and the emergence of adaptive variation. By centring on system behaviour under stable conditions, the study provides an initial and transparent assessment of how hybrid adaptivity operates within the context of structured instructional materials.

The primary objective of this study is to design and technically examine a hybrid recommendation framework that integrates explicit cognitive structuring into adaptive worksheet generation within blended learning environments. Existing educational recommender systems typically rely on behavioural similarity and performance signals, yet they rarely incorporate formal instructional constraints that maintain cognitive progression across Bloom levels.

Within this context, the study contributes in three ways. First, it proposes a pedagogically structured hybrid recommendation model in which Bloom’s Revised Taxonomy is embedded directly into the scoring process, enabling adaptive variation while preserving cognitive alignment. Second, it provides a system-level evaluation of recommendation behaviour through expected–actual alignment analysis under controlled simulation conditions, focusing on structural consistency and interpretability rather than solely predictive accuracy. Third, it illustrates how such a hybrid engine can operate within a blended learning architecture while maintaining instructional coherence and educator oversight.

Related Work

Adaptive learning and educational recommender systems have been widely investigated as mechanisms for supporting personalisation in digital learning environments. Within this research area, adaptive learning approaches typically utilise learner-related signals, such as performance, preferences, or behavioural patterns, to adjust instructional content and learning pathways dynamically [13,14]. Systematic reviews indicate sustained growth in this field, while also documenting substantial heterogeneity in how adaptivity is designed, implemented, and evaluated across educational contexts [10,15].

A significant portion of prior work relies on artificial intelligence and machine learning techniques to support adaptive decision-making. Bibliometric analyses highlight a sharp increase in AI-driven educational research over the last decade, particularly in relation to personalised learning and recommendation tasks [16,17]. These approaches have demonstrated improvements in recommendation accuracy and learner engagement; however, their performance is often shaped by practical constraints, including data sparsity, cold-start effects, and platform-specific data availability [18,19].

Alongside purely data-driven solutions, hybrid recommender systems have become prevalent in educational settings. By combining content-based and collaborative filtering techniques, hybrid approaches seek to balance recommendation accuracy with robustness across diverse learner profiles [20,21]. Such systems frequently incorporate additional criteria, including learning object metadata, difficulty levels, and learner similarity, when computing relevance scores [22,23]. While hybrid models are often positioned as a means of mitigating the limitations of individual recommendation strategies, existing implementations differ considerably in terms of pedagogical transparency and interpretability.

Recent surveys further suggest that many educational recommender systems focus primarily on technical optimisation, with comparatively less emphasis on the explicit structuring of instructional artefacts [10,24]. As a consequence, recommended resources may align closely with learner preferences or behavioural patterns, while remaining only loosely connected to predefined instructional objectives or cognitive progression. This tendency is particularly evident in practice-oriented learning materials, such as worksheets, which are seldom examined as structured sequences of learning and assessment activities within adaptive systems [2].

Bloom’s revised taxonomy continues to serve as a foundational framework for organising learning activities according to cognitive complexity, from lower-order recall to higher-order reasoning and creation [8]. Although widely adopted in instructional design, relatively few adaptive or recommender-based systems explicitly operationalise Bloom’s taxonomy as part of their recommendation logic [9,25]. Where Bloom-related concepts are considered, they are most often treated descriptively rather than embedded as structural constraints guiding content selection and sequencing.

Recent studies have successfully employed machine learning techniques to automate the classification of e-content and learning outcomes according to Bloom’s Taxonomy [26,27]. Furthermore, natural language processing and semantic tagging have been increasingly utilized to automatically generate and curate static question banks based on Bloom’s cognitive levels [28]. Finally, some deep learning approaches have attempted to combine Bloom’s Taxonomy with learning styles to recommend general educational resources [29]. However, while these approaches effectively operationalize Bloom’s Taxonomy for semantic tagging, static content generation, or opaque neural network recommendations, they rarely integrate these pedagogical labels as active, explicit structural constraints within a transparent recommendation engine.

Recent advances in machine learning have explored the integration of predictive models with optimisation processes, giving rise to the predict-and-optimise paradigm. In such approaches, learning is guided not only by prediction accuracy but also by the quality of the resulting decisions [30]. Similarly, end-to-end frameworks have been proposed, where optimisation is embedded within the learning process to directly improve decision outcomes [31]. More broadly, the combination of machine learning and optimisation has been recognised as a promising direction for data-driven decision-making [32]. Related learning-assisted optimisation frameworks have also been explored in domains such as energy systems [33]. While these approaches are applied in different domains, they highlight the broader potential of integrating learning signals with structured optimisation.

As summarized in

Table 1. Comparative overview of representative recent studies in educational recommender systems and automated content generation.

Study	Recommendation Approach/Core Method	Pedagogical Structuring	Target Output	Evaluation Method
Demertzi & Demertzis [20]	Hybrid: Ontology-matching and Rule-based	No explicit pedagogical constraint	General e-learning pathways	Empirical observation
Wan & Yu [21]	Hybrid: Cognitive Maps and CF	Cognitive Maps	General learning sequences	Empirical testing
Dai et al. [25]	Hybrid: Adaptive Testing and CF	Item Response Theory (IRT)	Adaptive test items	Empirical testing
Lin et al. [23]	Hybrid: Deep Neural CF	No explicit pedagogical constraint	Online course recommendations	Simulation-based testing
Thomas & Chandra [26]	Machine Learning: Random Forest	Bloom tagging (descriptive use)	Tagged e-content	Model Accuracy
Shaikh et al. [27]	Machine Learning: LSTM	Bloom tagging (descriptive use)	Tagged learning outcomes	Model Accuracy/F1 Score
Patil et al. [28]	NLP Pipeline and Rule-based	Bloom tagging (descriptive use)	Static question generation	System functional testing
Shen [29]	Deep Learning: Neural Networks	Bloom and Learning Styles (descriptive use)	General educational resources	System Accuracy Metrics
This study	Hybrid: Content-based, CF and Rules	Explicit structural enforcement (Bloom + VARK)	Adaptive, multi-level Worksheets	Simulation-based testing

, in contrast to purely data-driven models that prioritize algorithmic optimization over instructional coherence (often acting as “black boxes”), the proposed approach stands out by moving beyond descriptive pedagogical tagging. Through the explicit structural enforcement of Bloom’s Taxonomy and VARK profiles, the system ensures that the generated worksheets maintain a strict pedagogical progression across cognitive levels. This enables bounded, personalised adaptivity, shifting the paradigm from simple cognitive classification or static generation to the dynamic creation of multi-level, adaptive worksheets.

2. Materials and Methods

2.1. Architectural Design

The adaptive educational platform follows a microservices architecture to secure maintainability, scalability, and extensibility by decomposing the system into six autonomous services. This approach extends a previously developed adaptive blended learning platform architecture [34], where a worksheet generation engine is introduced as a core component. The present work focuses on the design and evaluation of the recommendation mechanism that supports the adaptive selection of learning and evaluation objects. The system consists of services for learning objects, learning styles, learning methods, learning subjects, user profiling, and worksheet generation. Each service is responsible for a specific domain, enabling third-party access, independent development, and flexible evolution. At the core, the worksheet service dynamically generates personalised worksheets by communicating with the other services, supporting a data-driven and adaptive learning experience. Each microservice operates with isolated business logic and a dedicated database, reducing interdependencies and safeguarding data integrity. Services communicate through lightweight API calls to ensure reliable and efficient interaction. All services are implemented in NodeJS v22 following clean architecture principles, separating domain, application, and infrastructure layers. This clear separation of concerns improves testability, simplifies maintenance, and allows the platform to adapt easily to evolving educational requirements. Finally, a DevOps layer supports modular and reliable development aligned with the microservices architecture, enabling independent versioning and deployment. Using Docker v29 for containerization and Jenkins v2.479 for CI/CD automation, it ensures consistent automated build, test, and deployment across environments with minimal manual effort. Figure 1 illustrates the architectural structure of the proposed adaptive worksheet recommendation framework and clarifies the orchestration logic among the core services. The worksheet engine functions as the central coordination component, integrating pedagogical metadata, learner profiling information, and collaborative interaction signals during worksheet generation. A hybrid scoring mechanism combines pedagogical constraints, learner modality preferences, and Bloom’s taxonomy alignment to rank and select appropriate learning and evaluation objects. The modular separation of services supports scalability and independent system evolution while maintaining coordinated communication between the content services, the recommendation engine, and the learner-facing interfaces.

2.2. Microservices Intercommunication and Data Flow

The worksheet engine serves as the central orchestration service, communicating with the user profiling, learning styles, learning methods, learning subjects, and learning object services through lightweight RESTful APIs. These services provide the required inputs for generating adaptive and personalised worksheets. The worksheet service retrieves learning styles, pedagogical methods, curriculum structure, and Bloom’s taxonomy levels, forming a learning skeleton used to fetch relevant learning objects. Next, the service obtains the learner’s profile and requests the user profiling service to grade the learning objects based on user fit. The highest-ranked objects are then selected for learning and assessment activities. During worksheet execution, learner progress, preferences, and submitted evaluation data are stored and made available to teachers. After the teacher evaluation, grades are saved, and the learner’s performance is sent back to the user profiling service to update the user profile accordingly.

2.3. Data Storage

By following a microservice-based architecture, each service operates independently with its own database. This approach enables the use of different database technologies (SQL and NoSQL) based on the specific data and business logic requirements of each service. Services such as Learning Objects, Learning Subjects, and User Profiling manage highly structured, relational data with strong consistency and complex query needs, making PostgreSQL a suitable choice. These services store metadata-rich educational resources and interrelated learner performance metrics. In contrast, services such as Worksheets, Learning Styles, and Learning Methods handle semi-structured and frequently changing data, including dynamically generated worksheets and questionnaire results. For these use cases, MongoDB is used to provide schema flexibility and efficient handling of evolving data structures. The platform stores both static and user-generated educational content, as well as automatically generated learner data. All data is securely stored with no third-party access. Users may request account deletion, after which sensitive data are removed and remaining records anonymized to preserve platform integrity.

2.4. Recommendation Engine

The recommendation engine operates by combining pedagogical structure, individual learner profiling, and collaborative influence within a unified scoring framework. At each worksheet generation step, candidate learning objects are first filtered according to the instructional skeleton defined by the worksheet design. These candidates are then evaluated through a weighted aggregation of cognitive alignment (Bloom level), learner modality preferences, prior performance indicators, together with similarity-based signals derived from behaviourally related learners. This layered approach ensures that adaptive variation occurs within predefined instructional boundaries rather than emerging solely from behavioural correlation patterns.

2.4.1. Learning Object Score

For clarity and notation consistency, v denotes the normalized value of component x and w denotes the corresponding weighting coefficient used in the scoring aggregation. For each existing learning object of the worksheet skeleton, we compute a score that is based on the following formula:

$${\mathit{L}\mathit{O}}_{\mathit{s}\mathit{c}\mathit{o}\mathit{r}\mathit{e}}=({\mathit{v}}_{\mathit{l}\mathit{s}} \times {\mathit{w}}_{\mathit{l}\mathit{s}})+({\mathit{v}}_{bds} \times {\mathit{w}}_{\mathit{b}\mathit{d}\mathit{s}})+({\mathit{v}}_{\mathit{s}\mathit{o}\mathit{l}} \times {\mathit{w}}_{\mathit{s}\mathit{o}\mathit{l}})+({\mathit{v}}_{los} \times {\mathit{w}}_{\mathit{l}\mathit{o}\mathit{s}})$$

The learning object grading formula is based on four weighted components that estimate the suitability of each object for a given learner. LS value (${\mathit{v}}_{\mathit{l}\mathit{s}}$), reflects the learner’s VARK profile by comparing the student’s preferred LS with the target dimension assigned to each LO; scores range from 0 to 10 and are weighted by 0.1 (${\mathit{w}}_{\mathit{l}\mathit{s}}$). The Bloom Dimension (Bds), represented by ${\mathit{v}}_{\mathit{b}\mathit{d}\mathit{s}}$, captures the learner’s cognitive level by classifying learning objects according to Bloom’s Taxonomy (remember, understand, apply, analyze, evaluate, and create); this component is scored on a 0–10 scale and assigned a weight of 0.4 (${\mathit{w}}_{\mathit{b}\mathit{d}\mathit{s}}$) to emphasize cognitive progression. Similar Object Liking (sol), represented by ${\mathit{v}}_{\mathit{s}\mathit{o}\mathit{l}}$ models learner engagement and preference based on interactions with similar learning objects. Behavioural signals in the proposed framework are derived from interaction-based indicators, such as user preference (e.g., liking scores), performance feedback (EO results), and consistency patterns. These types of signals are commonly used in recommender systems to model user engagement and support personalised recommendations [18,22]. Values are rated on a 0–5 scale, normalized to align with the other components, and weighted by 0.2 (${\mathit{w}}_{\mathit{s}\mathit{o}\mathit{l}}$). Finally, Similar Students’ LO Score (slos), represented by ${\mathit{v}}_{\mathit{l}\mathit{o}\mathit{s}}$, integrates collaborative information from learners with similar profiles and is computed as follows:

$${\mathit{v}}_{\mathit{l}\mathit{o}\mathit{s}}=({\mathit{v}}_{\mathit{l}\mathit{o}\mathit{l}} \times {\mathit{w}}_{\mathit{l}\mathit{o}\mathit{l}})+({\mathit{v}}_{\mathit{e}\mathit{o}\mathit{g}} \times {\mathit{w}}_{\mathit{e}\mathit{o}\mathit{g}})$$

where ${\mathit{v}}_{\mathit{l}\mathit{o}\mathit{l}}$ denotes peers’ liking of learning objects weighted at 0.3, and ${\mathit{v}}_{\mathit{e}\mathit{o}\mathit{g}}$ represents their average performance on related evaluation objects weighted at 0.7. The resulting Sss component contributes to the overall grading score with a weight of 0.3 (${\mathit{w}}_{\mathit{l}\mathit{o}\mathit{s}}$). The weighted linear aggregation scheme adopted in this framework follows an established hybrid recommendation strategy in which multiple pedagogical and behavioural signals are combined through explicitly defined coefficients [35]. In educational systems, such linear formulations are often preferred because they preserve structural interpretability and allow the contribution of each component to remain transparent [36]. In the current configuration, the weighting coefficients are fixed to reflect the pedagogical priorities embedded in the platform design. Specifically, Bloom-dimension alignment is assigned the highest weight (w_bds = 0.4), followed by collaborative learner similarity (w_los = 0.3), behavioural preference signals (w_sol = 0.2), and learning-style alignment (w_ls = 0.1). These coefficients were defined a priori to express instructional priorities rather than to maximise predictive performance through parameter optimisation.

The coefficients in the present model were selected to reflect the instructional hierarchy embedded in the platform design. Greater emphasis is assigned to Bloom-level alignment as a structural cognitive constraint, while collaborative and preference-based indicators contribute to adaptive variation within these bounds. The aim at this stage is not coefficient optimisation but controlled examination of system behaviour under a fixed pedagogically prioritised configuration.

2.4.2. Evaluation Object Score

$${\mathit{E}\mathit{O}}_{\mathit{s}\mathit{c}\mathit{o}\mathit{r}\mathit{e}}={(\mathit{v}}_{\mathit{l}\mathit{s}}\times {\mathit{w}}_{\mathit{l}\mathit{s}})+({\mathit{v}}_{eog}\times {\mathit{w}}_{eog})+({\mathit{v}}_{\mathit{e}\mathit{o}\mathit{l}}\times {\mathit{w}}_{\mathit{e}\mathit{o}\mathit{l}})+({\mathit{v}}_{los}\times {\mathit{w}}_{\mathit{l}\mathit{o}\mathit{s}})+({\mathit{v}}_{\mathit{b}\mathit{d}\mathit{s}}+ {\mathit{v}}_{\mathit{e}\mathit{o}\mathit{d}})$$

LS, Slos, and Bds values are calculated in the same way as in LOscore, with the exception that in this case, we do not include a weight in the Bds. This is because we want to keep the bloom dimension the most impactful score in this case. Similar Evaluation Object Liking (Eol), represented by ${\mathit{v}}_{\mathit{e}\mathit{o}\mathit{l}}$, incorporates preference and satisfaction information from learners with similar profiles toward comparable evaluation objects; values are rated on a 0–5 scale, normalized to align with other components, and weighted by 0.2 (${\mathit{w}}_{\mathit{e}\mathit{o}\mathit{l}}$). Similar Evaluation Object Grade (eog), represented by v_eog, is a composite measure defined as:

$${\mathit{v}}_{eog}=({\mathit{v}}_{\mathit{e}\mathit{o}} \times {\mathit{w}}_{\mathit{e}\mathit{o}})+({\mathit{v}}_{\mathit{a}\mathit{g}} \times {\mathit{w}}_{\mathit{a}\mathit{g}})$$

where ${\mathit{v}}_{\mathit{e}\mathit{o}}$ denotes the preference for similar learners, weighted by 0.3, and ${\mathit{v}}_{ag}$ represents their average grades on comparable evaluation objects, weighted by 0.7; this collaborative component contributes to the final score with an overall weight of 0.4 (${\mathit{w}}_{\mathit{e}\mathit{o}\mathit{g}}$). Finally, Evaluation Object Difficulty (Eod), represented by ${\mathit{v}}_{\mathit{e}\mathit{o}\mathit{d}}$, acts as a normalization factor ranging from −0.2 to 0.2, adjusting the final score based on the alignment between the evaluation object’s difficulty level and the learner’s skill level.

2.4.3. Similar Students

To find similar students to the corresponding user, the platform uses the following formula:

$${\mathit{S}\mathit{s}}_{\mathit{s}\mathit{c}\mathit{o}\mathit{r}\mathit{e}}={\mathit{l}\mathit{s}}_{\mathit{s}\mathit{d}}+{\mathit{e}\mathit{o}}_{\mathit{s}\mathit{d}}+{\mathit{b}\mathit{d}}_{\mathit{s}\mathit{d}}$$

where each component represents the absolute difference between the target learner and candidate peers across the corresponding normalised dimensions. For the collaborative filtering component, similarity between learners is computed using the sum of absolute differences. This choice reflects the bounded and low-dimensional nature of the similarity space, which consists of three normalised pedagogical components. In this setting, L1 distance preserves proportional contribution across dimensions while maintaining transparent dimensional influence.

Learning Style Score Difference (${\mathit{l}\mathit{s}}_{\mathit{s}\mathit{d}}$) represents the difference between the current learner’s and peers’ learning style scores as derived from their respective VARK profiles. Evaluation Object Score Difference (${\mathit{e}\mathit{o}}_{\mathit{s}\mathit{d}}$) captures the difference between the current learner’s and peers’ average performance on evaluation objects. Bloom Dimension Score Difference (${\mathit{b}\mathit{d}}_{\mathit{s}\mathit{d}}$) reflects the cognitive level gap between the learner and peers, based on alignment with Bloom’s Taxonomy across their completed learning activities.

Learners most similar to a target learner are identified using the smallest total difference (Ss). The behavioral data of the top five similar peers are used to populate the Ss components of the ${\mathit{L}\mathit{O}}_{\mathit{s}\mathit{c}\mathit{o}\mathit{r}\mathit{e}}$ and ${\mathit{E}\mathit{O}}_{\mathit{s}\mathit{c}\mathit{o}\mathit{r}\mathit{e}}$ calculations. This collaborative filtering layer ensures that recommendations are influenced by both individual behavior and successful patterns observed in similar learners. After a tutor submits the final grading of a worksheet, the system stores the learner’s worksheet score, evaluation object scores, and liking values. All records are stored together with the associated worksheet, Bloom taxonomy dimensions, and learning style types. These data are then used to update the learner profile, capturing performance history, preferences, and satisfaction with learning and evaluation objects.

Profile updates are performed in three steps. First, learning style scores are updated per course using a weighted formula: 70% based on learner performance (grade) and 30% based on liking. The update is positive or negative depending on thresholds (grade > 5 and liking > 2 indicate positive change). This weighting prioritizes performance while still reflecting learner preferences. Second, Bloom dimension scores for the specific course are updated by recalculating the learner’s average worksheet performance since the latest submission. Finally, the learner’s overall Bloom profile is updated by averaging scores from all evaluation objects associated with each cognitive dimension. This continuous update process ensures that future recommendations accurately reflect both learner progress and evolving learning preferences.

Because the hybrid score is computed through explicitly defined components (Bloom alignment, learner performance indicators, preference factors, and similarity-based contributions), recommendation decisions remain structurally interpretable. Each adaptive replacement can be traced to the relative weighting of these components rather than to an opaque optimisation process. This preserves pedagogical transparency and allows educators to understand how instructional constraints and learner signals jointly influence worksheet variation.

2.5. Study Design

This study employed an early-stage technical evaluation design to investigate how the hybrid recommendation engine performs when generating adaptive worksheet recommendations across six Bloom dimensions. The experiment was conducted using a fixed content space of 54 Learning Objects (LOs) and 98 Evaluation Objects (EOs), organized into six worksheets, each covering all six levels of Bloom’s taxonomy. For each level, the recommendation process produces a ranked set of Learning Objects (LOs), from which the top three candidates are retained. In addition, a single Evaluation Object (EO) is selected to assess learner performance at the corresponding cognitive level. This structure ensures a progressive instructional sequence while allowing limited variation through multiple candidate learning resources. This controlled setup enables consistent expected–actual comparisons and reduced external sources of variation suitable for system behavior analysis.

The evaluation dataset was constructed in three stages. First, a researcher defined 15 synthetic learner profiles that reflect modality tendencies, inspired by the VARK categories. Second, an independent researcher assigned behavioural and performance attributes to each profile, including worksheet-level liking scores, EO grades, and behavioural tendencies (e.g., low or high performer, inconsistent evaluator). Third, an expert educator developed the instructional blueprint for the six worksheets and specified the Expected LO–EO sets using Bloom-aligned tagging. The use of synthetic learner profiles is well established in early validation of adaptive and recommender systems, particularly when the aim is to examine internal decision mechanisms before deployment with real learners [12].

The recommendation engine was executed under two configurations: the proposed hybrid model and a content-based baseline in which collaborative signals were removed from the scoring functions. Both configurations processed all worksheets for all learner profiles, producing 540 Bloom-level recommendation instances (six Bloom levels × six worksheets × fifteen profiles). Each instance generated three LO recommendations and one EO recommendation, yielding a total of 1620 Expected LOs, 1620 Actual LOs, 540 Expected EOs, and 540 Actual EOs (3780 recommendation units overall). To ensure internal validity, a researcher who had not participated in dataset construction manually cross-validated all Expected–Actual outputs to confirm alignment with the intended hybrid scoring logic.

The analysis focused on two main aspects. Instructional alignment was evaluated through LO overlap accuracy and EO matching rates across Bloom levels and generated worksheets. Adaptive behaviour was examined by observing how LO and EO selection patterns varied across learner profiles with different performance and liking signals. As the evaluation relied on synthetic data, all findings are descriptive and intended as an initial validation of system behaviour. While this design limits direct generalisation to classroom contexts, it enables controlled and repeatable testing of system-level decision mechanisms before future studies with real learners.

To further clarify the worksheet generation process, Figure 2 presents an example of the worksheet creation workflow. The process follows a structured sequence of steps. Initially (Step 1), the educator provides general information, such as the subject, course, educational level (e.g., primary or higher education), and thematic unit, ensuring alignment with the corresponding curriculum and instructional context. In Step 2, the educator selects the desired Bloom’s taxonomy levels that define the cognitive structure of the worksheet, along with additional pedagogical parameters such as the targeted 4C skills and other configuration options illustrated in the interface.

In Step 3, specific learning objectives are defined for each selected Bloom level, guiding the instructional focus of the worksheet. In Step 4, the system retrieves and presents a set of candidate Learning Objects (LOs) that satisfy the defined prerequisites and instructional criteria. In Step 5, a corresponding set of Evaluation Objects (EOs) is presented following the same principles. At this stage, the educator may either accept the proposed set or selectively choose the most appropriate items, allowing for teacher-guided refinement of the generated content.

Finally, in Step 6, the worksheet is reviewed and finalised before being provided to learners. This process demonstrates how curriculum constraints, pedagogical goals, and recommendation mechanisms are combined with educator input to generate structured and adaptive worksheets.

3. Results

3.1. Accuracy of Learning Object (LO) and Evaluation Object (EO) Recommendations

To evaluate recommendation accuracy, all 540 Bloom-level recommendation instances were analysed (six Bloom levels × six worksheets × fifteen learner profiles). The analysis compares the behaviour of the proposed hybrid recommender with a content-based baseline configuration in which collaborative signals were removed from the scoring functions. For each Bloom level, the system generated three candidate Learning Objects (LOs) and one Evaluation Object (EO), which were compared against the predefined instructional design (Expected). Because each Bloom level includes three LOs, LO accuracy was assessed using overlap analysis, measuring how many of the expected items (0–3) were present in the actual recommendation set. This approach captures partial alignment and is more appropriate for hybrid adaptive systems than binary accuracy measures.

3.2. LO Overlap Accuracy Across Bloom Levels

Across all Bloom levels, the hybrid recommender achieved a mean LO overlap of 2.55 out of 3, corresponding to a normalized LO accuracy of 85.1%. As shown in

Table 2. LO Overlap Accuracy by Bloom Level.

Hybrid Recommender—Our Proposed System
Bloom Level	Mean LO Overlap (0–3)	Normalized LO Accuracy	% 3/3 Overlap	% 2/3 Overlap	% 1/3 Overlap	% 0/3 Overlap
Remember	3.00	100%	100%	0%	0%	0%
Understand	3.00	100%	100%	0%	0%	0%
Apply	3.00	100%	100%	0%	0%	0%
Analyze	2.11	70.3%	26.7%	57.8%	15.6%	0%
Evaluate	2.13	71.1%	24.4%	64.4%	11.1%	0%
Create	2.16	72.0%	28.9%	57.8%	13.3%	0%
Content-Based Recommender
Bloom Level	Mean LO Overlap (0–3)	Normalized LO Accuracy	% 3/3 Overlap	% 2/3 Overlap	% 1/3 Overlap	% 0/3 Overlap
Remember	3.00	100%	100%	0%	0%	0%
Understand	3.00	100%	100%	0%	0%	0%
Apply	3.00	100%	100%	0%	0%	0%
Analyze	2.06	68.67%	25.6%	54.4%	20.0%	0%
Evaluate	2.07	69%	24.4%	57.8%	17.8%	0%
Create	2.08	69.33%	26.7%	55.6%	16.7%	1%

, both the hybrid and the content-based baseline produced identical recommendations at the lower Bloom levels (Remember, Understand, Apply), where full overlap (3/3) was consistently observed. Differences between the two configurations emerged at higher Bloom levels (Analyze, Evaluate, Create), where the hybrid model produced slightly higher overlap values. For these levels, mean overlap values for the hybrid configuration ranged from 2.11 to 2.16, corresponding to normalized accuracy levels between 70.3% and 72.0%, while the content-based baseline produced slightly lower overlap values. No instances of complete mismatch (0/3 overlap) were observed in the hybrid configuration, whereas the content-based baseline produced a very small number of such cases.

3.3. Accuracy Across Worksheets

LO and EO accuracy were also examined at the worksheet level for both recommendation configurations. For the hybrid recommender, normalized LO accuracy ranged from 82.7% to 94.7% across the six worksheets, while EO accuracy consistently exceeded 95% (

Table 3. LO Overlap and EO Accuracy by Worksheet.

Hybrid Recommender—Our Proposed System
Worksheet	Mean LO Overlap (0–3)	Normalized LO Accuracy	EO Accuracy
W1	2.82	94.0%	100%
W2	2.48	82.7%	96.7%
W3	2.51	83.7%	95.6%
W4	2.52	84.0%	95.6%
W5	2.61	86.9%	96.7%
W6	2.84	94.7%	98.9%
Content-Based Recommender
Worksheet	Mean LO Overlap (0–3)	Normalized LO Accuracy	EO Accuracy
W1	2.44	81.5%	86.7%
W2	2.50	83.3%	94.4%
W3	2.60	86.7%	90.0%
W4	2.46	81.9%	90.0%
W5	2.63	87.8%	90.0%
W6	2.57	85.6%	96.7%

). Comparison with the content-based baseline shows that LO overlap remained broadly comparable between the two configurations, with each approach producing slightly higher overlap in different worksheets. In contrast, EO agreement was consistently higher for the hybrid configuration across all worksheets. These results suggest that the inclusion of collaborative signals primarily improves the stability of evaluation object selection while maintaining similar learning-object overlap behaviour across worksheets.

3.4. Simulation-Based Adaptivity Validation

This section reports observable patterns of recommendation behaviour under simulated learner-profile conditions, focusing primarily on the hybrid recommendation configuration. Across the lower Bloom levels (Remember, Understand, Apply), recommendations consistently matched the predefined instructional design, including for learner profiles associated with low performance or negative liking scores.

At higher Bloom levels (Analyze, Evaluate, Create), learner-profile signals were associated with greater variation in LO selection. Profiles characterized by low task performance or erratic liking patterns were more frequently associated with partial overlap in the LO set, typically involving the replacement of one of the three expected learning objects. Profiles that corresponded to consistently high-performing or highly engaged learners more often exhibited full LO overlap. Across all simulated contexts, no extreme mismatches were observed. EO recommendations produced by the hybrid configuration remained closely aligned with the instructional design across all simulated contexts.

3.5. Performance

System performance was examined during worksheet generation to assess suitability for interactive educational use. The average worksheet generation time ranged from 350 to 450 ms, with the slowest cases remaining below 1 s. Latency analysis showed that local processing required approximately 20 ms or less, while more than 98% of total request time was attributable to remote service calls. Requests related to user profiling accounted for the largest share of latency (approximately 90%), followed by learning-object retrieval, which introduced a smaller overhead. Occasional periodic latency spikes associated with remote services were observed; however, the system remained responsive throughout execution.

4. Discussion

The findings of this pilot study offer a clearer understanding of how a hybrid recommendation engine can balance instructional stability with meaningful adaptivity. One notable result is the system’s conservative behaviour at the lower Bloom levels. Given that these levels target foundational knowledge and well-structured tasks, it is pedagogically appropriate and consistent with prior research that the system maintains full alignment with the instructional blueprint [13,37]. In this sense, the absence of variation at these levels should not be interpreted as a lack of adaptivity. Rather, it reflects the broader educational principle that early-stage cognitive tasks require stability to support scaffolding and to avoid unnecessary cognitive load. This observation is further supported by the consistently complete overlap observed at these levels across all worksheets and learner-profile conditions, indicating that structural fidelity was preserved even when learner performance was low or preferences were inconsistent.

A comparison with the content-based baseline further clarifies this behaviour. While LO overlap remained broadly comparable between the two configurations across worksheets, evaluation-object agreement was consistently higher for the hybrid model. This pattern suggests that collaborative signals contribute primarily to stabilizing assessment selection rather than substantially altering the core instructional content. In other words, the hybrid mechanism appears to reinforce the alignment of evaluation objects with learner performance patterns while preserving the pedagogical structure defined by the worksheet design.

The adaptive behaviour becomes more evident at the higher Bloom dimensions. Here, the system introduces small but targeted deviations, typically modifying one of the three expected learning objects while preserving the core structure of the task set. Importantly, these deviations remain limited in scope and do not result in complete divergence from the instructional design. This pattern is consistent with the literature on hybrid recommender systems, which notes that controlled, interpretable adjustments often emerge when rule-based logic is combined with content and behavioural signals [10,20]. It also aligns with research showing that higher-order tasks naturally allow for greater differentiation due to the open-ended nature of analysis, evaluation, and creation [25]. In this study, the system demonstrates precisely this type of bounded adaptivity, adjusting where meaningful, but without diverging from the instructional intent.

The simulation-based results reinforce this view. By testing the engine across a wide range of synthetic learner profiles, from low-performing and disengaged users to highly consistent and motivated ones, the analysis highlights the robustness of the adaptive logic. Even under behavioural patterns that would typically challenge an adaptive system, such as volatile or consistently negative ratings, the model remained stable and avoided extreme mismatches. Notably, adaptive variation at the learning-object level did not disrupt evaluation alignment, as assessment recommendations remained coherent across all simulated contexts. This robustness aligns with calls in the literature for human-centered adaptive systems that remain resilient to noise and maintain pedagogical safety [22,38]. Crucially, the system’s behaviour indicates that it adapts where differentiation has pedagogical value, while maintaining structural fidelity where instructional coherence is required. From a system-level perspective, the results suggest that the recommendation process can support interactive use, while highlighting the role of remote service dependencies in overall latency.

Overall, the results present an adaptive engine that embodies the principles described in recent research on responsible and interpretable personalization. It is not “adaptive for its own sake”; instead, it adjusts modestly and predictably in the cognitive regions where variation supports deeper thinking, while preserving stability in areas where structure is essential. This equilibrium between consistency and personalization is at the core of current approaches to hybrid adaptive design [21,39]. Although further validation with real learners will be necessary to assess learning impact, current findings offer encouraging evidence that the system operates within well-defined pedagogical boundaries and provides a solid foundation for future deployment in blended learning environments.

Limitations and Future Work

This study represents a prototype-level technical validation conducted under controlled simulation conditions. While this setup enabled systematic examination of the internal recommendation logic and scoring behaviour, it does not directly measure learner outcomes in real classroom environments. The use of synthetic learner profiles allowed the system mechanisms to be examined in isolation before introducing the variability that characterises authentic learning contexts. For this reason, the findings should be interpreted primarily as evidence of structural stability and internal recommendation consistency rather than direct instructional impact.

A second limitation concerns the size and structure of the evaluation dataset. The system was tested using six worksheets and a predefined pool of learning and evaluation objects. This controlled design allowed precise examination of LO–EO alignment but does not fully reflect the diversity and scale of real educational repositories, where content is typically larger and more heterogeneous.

In addition, the current adaptive model relies on a limited set of learner signals including performance indicators, liking values, and similarity-based behavioural patterns. While these indicators are sufficient for early-stage system validation, richer behavioural data such as time on task, longer-term performance trends, or engagement patterns may further improve recommendation precision.

Future work will focus on empirical validation with real learners in blended classroom settings. Such deployment will allow examination of learner progression, engagement behaviour, and instructional impact under authentic conditions. Further development will also explore scalability in larger content repositories and optimisation strategies to reduce latency associated with remote profiling services.

5. Conclusions

This paper presented a hybrid recommendation approach for adaptive worksheet generation that combines pedagogical structuring based on Bloom’s taxonomy with content-based and collaborative filtering mechanisms. The goal of the proposed framework is to support adaptive content selection while preserving the instructional structure defined by the worksheet design.

The simulation-based evaluation indicated strong alignment between generated recommendations and the intended instructional structure. At lower Bloom levels, the system maintained full consistency with the predefined worksheet blueprint, reflecting the pedagogical need for stability at foundational stages of learning. At higher cognitive levels, the engine introduced limited but interpretable variation in learning object selection while preserving the overall task structure.

Comparison with a content-based baseline indicated that the hybrid configuration primarily improves the stability of evaluation object recommendations while maintaining comparable behaviour in learning object selection. This suggests that collaborative signals can enhance assessment alignment without disrupting the pedagogical constraints embedded in worksheet-based instruction.

Overall, the results indicate that hybrid recommendation mechanisms can support adaptive learning activities within structured instructional environments while maintaining transparency and pedagogical coherence. These findings provide a foundation for further investigation of hybrid adaptive systems in blended educational settings.