2.1 Learning analytics

Learning analytics is a rapidly evolving field that leverages data-driven techniques to analyze and optimize educational processes [24]. By collecting and interpreting data from various sources, such as learning management systems [25], online platforms, and student interactions, it provides valuable insights into student behaviors, performance, and engagement. The core principle of learning analytics lies in transforming raw educational data into actionable knowledge, enabling educators and institutions to make informed decisions. This approach bridges the gap between data science and education, fostering a deeper understanding of learning dynamics and outcomes.

From a technical perspective, learning analytics employs a wide range of methodologies, including statistical analysis [26], machine learning [27], and data mining [28]. These techniques are used to identify patterns, predict outcomes, and classify students based on their learning trajectories. Advanced methods, such as deep learning [29] and natural language processing [30], are increasingly being adopted to handle complex and high-dimensional educational data. The diversity of these techniques allows learning analytics to address various educational challenges, from personalized learning to curriculum design and institutional planning.

Learning analytics can be broadly categorized into descriptive, diagnostic, predictive, and prescriptive analytics. Descriptive analytics focuses on summarizing past and current learning behaviors, while diagnostic analytics aims to uncover the reasons behind observed patterns. Predictive analytics uses historical data to forecast future outcomes, such as student performance or dropout risks. Prescriptive analytics goes a step further by recommending interventions or strategies to optimize learning experiences. Each category serves a unique purpose, contributing to a comprehensive understanding of educational ecosystems.

The advantages of learning analytics are manifold. It enables personalized learning by tailoring educational content and strategies to individual student needs, thereby improving engagement and outcomes. It also supports early intervention by identifying at-risk students and providing timely support. For institutions, learning analytics offers insights into resource allocation, curriculum effectiveness, and overall educational quality. Additionally, it fosters a data-driven culture in education, encouraging evidence-based decision-making and continuous improvement.

Despite its potential, learning analytics faces challenges related to data privacy, ethical considerations, and the interpretability of complex models. Ensuring the responsible use of data and addressing biases in algorithms are critical for maintaining trust and equity in educational systems. As the field continues to evolve, interdisciplinary collaboration and advancements in technology will play a pivotal role in overcoming these challenges and unlocking the full potential of learning analytics.

2.2 CRNN model for educational feature extraction

The CRNN [31] is a hybrid deep learning architecture that combines convolutional neural networks and recurrent neural networks for sequential data processing. In educational contexts, CRNN models are particularly effective for extracting meaningful features from multimodal educational data [39], including handwritten notes [40], scanned documents [41], and digital learning materials [42]. The CNN component automatically learns spatial hierarchies of features from input images, while the RNN component processes these features sequentially to capture temporal dependencies, making it suitable for analyzing educational content with both visual and sequential characteristics.

CRNN models offer several advantages for educational applications. The architecture inherently handles variable-length inputs [43], which is crucial for processing diverse educational materials of different formats and lengths. Unlike traditional OCR systems, CRNNs can recognize text without explicit character segmentation, making them more robust for analyzing poorly scanned documents or handwritten student work. The end-to-end trainable nature of CRNNs eliminates the need for handcrafted feature engineering, allowing the model to automatically discover relevant patterns in educational content across different languages and subject domains.

The practical applications of CRNN models in education are numerous and impactful [44]. They can be deployed for automated grading systems that process handwritten exam answers, digitization of historical educational archives, or real-time transcription of lecture notes. In language learning applications [45], CRNNs can assist in handwriting recognition for character-based languages or analyze the spatial organization of mathematical equations. The technology also enables accessibility features, such as converting scanned textbooks into machine-readable formats for visually impaired students or automatically generating searchable indexes for educational video content.

Compared to alternative approaches, CRNN models demonstrate superior performance in handling the noisy and varied nature of educational materials. Traditional computer vision techniques often struggle with the imperfect quality of student-generated content, while pure CNN architectures lack the sequential modeling capability needed for contextual understanding. CRNNs bridge this gap by maintaining spatial awareness while incorporating sequential processing, making them particularly suitable for educational environments where materials range from neatly typeset textbooks to hastily written student notes with varying handwriting quality. Despite their advantages, implementing CRNN models in educational settings presents certain challenges that require consideration. The models typically demand large annotated datasets for training, which can be resource-intensive to create for specialized educational content. Computational requirements may also pose limitations for real-time applications in resource-constrained educational institutions. Furthermore, the interpretability of CRNN decisions remains an ongoing research area, which is particularly important in high-stakes educational assessments. Future developments may focus on creating more efficient architectures and incorporating domain-specific knowledge to better serve diverse educational needs while addressing these limitations.

2.3 Deep learning for learning analytics

Deep learning has revolutionized learning analytics by enabling the analysis of complex and high-dimensional educational data through advanced neural network architectures [32]. Techniques such as Convolutional Neural Networks, Recurrent Neural Networks, and Transformer models are widely used to extract spatial, temporal, and contextual features from diverse data sources, including student interactions, assessment records, and learning logs. These models can automatically learn hierarchical representations of data [33], eliminating the need for manual feature engineering and providing more accurate insights into learning behaviors and outcomes. The adaptability of deep learning allows it to handle unstructured data [34], such as text, images, and sequential records, making it a versatile tool for modern educational research.

The following five deep learning models are commonly used for learning path analysis:

Long Short-Term Memory networks [10] are a specialized form of Recurrent Neural Networks (RNNs) designed to capture long-term dependencies in sequential data, making them highly effective for modeling student learning paths. By incorporating memory cells and gating mechanisms, LSTMs can retain relevant information over extended periods, enabling accurate predictions of future learning behaviors based on historical data. Their ability to handle variable-length sequences and temporal patterns is particularly advantageous for analyzing complex learning trajectories. However, LSTMs are computationally intensive and require large datasets for training, which can be a limitation in educational contexts with limited data availability. Additionally, their “black-box” nature makes it challenging to interpret the underlying decision-making process, reducing their transparency in educational applications.

Convolutional Neural Networks are traditionally used for image processing but have been adapted for learning path analysis by transforming sequential data into structured formats, such as matrices or grids [19]. CNNs excel at extracting local patterns and hierarchical features from data, which can be useful for identifying critical learning milestones or interactions. Their efficiency in processing high-dimensional data and ability to generalize well with limited training samples are notable advantages. However, CNNs are less effective at capturing long-term dependencies compared to RNN-based models, limiting their applicability in scenarios where temporal relationships are crucial. Furthermore, their reliance on structured data representations may introduce preprocessing complexities, making them less flexible for diverse educational datasets.

Gated Recurrent Units are a simplified variant of LSTMs that combine the benefits of capturing temporal dependencies with reduced computational complexity [20]. GRUs use update and reset gates to control information flow, enabling efficient modeling of sequential learning behaviors. Their lower parameter count and faster training times make them more practical for real-time applications or environments with limited computational resources. However, GRUs may struggle with extremely long sequences compared to LSTMs, potentially compromising their performance in analyzing extended learning paths. Additionally, like LSTMs, GRUs suffer from interpretability issues, which can hinder their adoption in educational settings requiring transparent decision-making.

Transformer models [13], based on self-attention mechanisms, have revolutionized sequential data analysis by capturing global dependencies without relying on recurrent connections. Their ability to process entire sequences in parallel and focus on relevant features makes them highly effective for learning path analysis, particularly in identifying key learning patterns or predicting outcomes. Transformers excel in handling large-scale datasets and can be fine-tuned for specific tasks, offering flexibility and scalability. However, their computational demands and large memory requirements can be prohibitive for smaller educational institutions. Moreover, the complexity of attention mechanisms often reduces interpretability, posing challenges for educators seeking actionable insights.

Autoencoders are unsupervised learning models that compress input data into a lower-dimensional latent space and reconstruct it, making them useful for identifying patterns or anomalies in student learning paths [21]. By reducing data dimensionality, autoencoders can uncover hidden structures or clusters within learning behaviors, enabling personalized interventions or recommendations. Their ability to work with unlabeled data is particularly advantageous in educational contexts where labeled datasets are scarce. However, autoencoders may struggle with capturing temporal dependencies, limiting their effectiveness in modeling sequential learning processes. Additionally, their performance heavily depends on the quality and representativeness of the training data, which can vary widely across educational settings.

The advantages of deep learning in learning analytics are manifold. Its ability to model complex relationships and patterns in data enables personalized learning experiences by tailoring recommendations and interventions to individual student needs. Deep learning models can also predict student performance, identify at-risk learners, and optimize curriculum design based on data-driven insights. Furthermore, the scalability of these models allows them to process large-scale datasets efficiently, supporting institutional decision-making and resource allocation. By automating data analysis, deep learning reduces the reliance on human expertise and enhances the efficiency of educational systems.

Deep learning is applied across various scenarios in learning analytics, including adaptive learning systems, intelligent tutoring, and educational data mining. For instance, it powers recommendation engines that suggest personalized learning paths, analyzes student engagement patterns to improve course design, and detects anomalies in learning behaviors for early intervention. Additionally, deep learning supports the development of virtual learning environments and chatbots, enhancing student interaction and support. Despite its potential, challenges such as data privacy, model interpretability, and computational costs remain, requiring ongoing research and ethical considerations to ensure its responsible and equitable use in education.

3.1 Overview

This study introduces an innovative framework for analyzing and optimizing student learning paths by integrating three advanced modules: a sequential feature extraction module based on CRNN, a knowledge point association modeling module based on Transformer, and a learning path optimization module based on Reinforcement Learning (RL). The CRNN module serves as the foundational component, extracting temporal features from student learning behavior data, such as timestamps, course progress, learning duration, and interaction records. By capturing the sequential patterns of individual learning paths, this module generates embedded representations of temporal features, which are essential for identifying key learning patterns and enhancing personalized recommendations. The Transformer module builds upon these temporal embeddings by modeling the relationships between knowledge points using a predefined knowledge graph, enabling the prediction of future learning paths and the generation of tailored learning routes. This module leverages the self-attention mechanism to establish deep associations between learning behaviors and knowledge structures, thereby improving the precision of personalized recommendations. Finally, the RL module optimizes the learning path recommendations by continuously adjusting model parameters based on real-world learning data and predefined reward mechanisms, such as learning completion rates and test score improvements. By iteratively refining the recommendations through RL, the framework adapts dynamically to individual student needs, ensuring the delivery of intelligent and context-aware learning strategies. Together, these modules form a cohesive system that addresses the challenges of personalized learning path analysis and optimization, leveraging the strengths of CRNN, Transformer, and RL to provide a comprehensive solution for educational applications. Fig 2 shows the model pipeline.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. The pipeline of the proposed model.

https://doi.org/10.1371/journal.pone.0331491.g002

3.2 Sequential feature extraction with CRNN

The Sequential Feature Extraction module in this study employs a CRNN to extract temporal patterns from student learning behavior data, such as timestamps, course progress, learning duration, and interaction records. The CRNN architecture combines the strengths of Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) to capture both spatial and sequential features from raw data. The CNN component processes the input data to generate feature maps, which are then flattened and passed to the RNN component for sequential analysis. This hybrid approach enables the model to effectively represent the temporal dynamics of learning behaviors, which is crucial for understanding individual learning paths. The output of this module is a set of temporal feature embeddings, denoted as , where h_t represents the feature vector at time step t, and T is the total number of time steps. These embeddings serve as a compact and meaningful representation of the student’s learning trajectory, capturing both local and global patterns in the data.

The CNN component of the CRNN is defined as , where X is the input data matrix and F is the resulting feature map. The CNN applies a series of convolutional filters to extract local patterns from the data, followed by pooling layers to reduce dimensionality. Mathematically, the convolution operation can be expressed as:

where X is the input data, K is the convolutional kernel, and F_i,j is the feature value at position (i,j). The pooling operation, typically max pooling, is defined as:

where P_i,j is the pooled feature value. The output of the CNN is then reshaped into a sequence of feature vectors, , which are fed into the RNN component for sequential analysis.

The RNN component processes these feature vectors sequentially, updating its hidden state at each time step according to the equation:

where h_t is the hidden state at time step t, and h_t−1 is the hidden state from the previous time step. In this study, a Long Short-Term Memory (LSTM) variant of RNN is used to address the vanishing gradient problem and capture long-term dependencies in the data. The LSTM equations are as follows:

Here, i_t, f_t, and o_t are the input, forget, and output gates, respectively; C_t is the cell state; and W and b are the weight matrices and biases. These equations enable the LSTM to selectively retain and update information over long sequences, making it highly effective for modeling learning behaviors.

The innovation of this module lies in its ability to combine the spatial feature extraction capabilities of CNNs with the sequential modeling strengths of RNNs, providing a robust and efficient solution for temporal feature extraction in learning analytics. By generating temporal feature embeddings that capture both local and global patterns in the data, the CRNN module addresses the challenge of identifying key time-dependent features that influence learning outcomes. These embeddings are then passed to the Knowledge Point Association Modeling module, where they are used to establish relationships between learning behaviors and knowledge points. This seamless integration of modules ensures that the temporal dynamics of learning are preserved and utilized in subsequent stages of the framework. The CRNN module not only enhances the accuracy of personalized learning path recommendations but also provides a foundation for further analysis and optimization, making it a critical component of the proposed framework. Fig 3 illustrates the technical principle of the CRNN model.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. The technical principle of the CRNN model.

https://doi.org/10.1371/journal.pone.0331491.g003

3.3 Knowledge point association modeling with transformer

The Knowledge Point Association Modeling module in this study utilizes a Transformer-based architecture to establish deep associations between student learning paths and knowledge points, leveraging the temporal feature embeddings generated by the CRNN module. The Transformer model is particularly well-suited for this task due to its ability to capture long-range dependencies and contextual relationships within sequential data. The input to this module consists of the temporal feature embeddings, , and a predefined knowledge point association graph, which encodes hierarchical relationships between course chapters and topics. By integrating these inputs, the Transformer model predicts future learning paths and generates personalized learning routes based on the relationships between knowledge points.

The core mechanism of the Transformer is its self-attention mechanism, which computes attention scores between all pairs of elements in the input sequence. The self-attention operation is defined as:

where Q, K, and V are the query, key, and value matrices, respectively, and d_k is the dimensionality of the key vectors. The query, key, and value matrices are derived from the input embeddings through linear transformations:

where W_Q, W_K, and are learnable weight matrices. The attention scores determine the importance of each element in the sequence relative to others, enabling the model to focus on relevant knowledge points and their relationships. This mechanism allows the Transformer to capture complex dependencies between learning behaviors and knowledge structures, which is critical for accurate prediction and recommendation.

The Transformer architecture also incorporates positional encoding to account for the order of elements in the sequence, as the model itself is permutation-invariant. The positional encoding is added to the input embeddings:

where is the enhanced input embedding matrix, and T is the sequence length. The positional encoding is defined using sine and cosine functions of different frequencies:

where pos is the position in the sequence, and i is the dimension index. This encoding ensures that the model retains information about the temporal order of learning behaviors, which is essential for modeling sequential data.

The output of the Transformer module is a set of contextualized embeddings, , which represent the enriched understanding of learning paths in relation to knowledge points. These embeddings are used to predict future learning paths and generate personalized recommendations. The prediction is formulated as:

where Y is the predicted probability distribution over future learning paths, and W_y and b_y are learnable parameters. The recommendations are tailored to individual students by considering their unique learning patterns and the relationships between knowledge points, ensuring high precision and relevance.

The innovation of this module lies in its ability to integrate temporal feature embeddings from the CRNN module with a structured knowledge graph, enabling the Transformer to model complex relationships between learning behaviors and knowledge points. This approach not only improves the accuracy of personalized recommendations but also provides a deeper understanding of the factors influencing learning outcomes. The output of this module is subsequently used in the Reinforcement Learning module to further optimize learning path recommendations, creating a cohesive and adaptive framework for personalized education. By leveraging the strengths of the Transformer architecture, this module addresses the challenge of establishing meaningful associations between learning paths and knowledge structures, making it a key component of the proposed framework. Fig 4 shows how knowledge points are organized using the Transformer Model.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. How knowledge points are organized using the Transformer Model.

https://doi.org/10.1371/journal.pone.0331491.g004

3.4 Learning path optimization with reinforcement learning

The Learning Path Optimization module in this study employs a Reinforcement Learning (RL) framework based on the Proximal Policy Optimization (PPO) algorithm to adaptively optimize learning paths for individual students. PPO is chosen for its stability, efficiency, and ability to handle continuous action spaces, making it well-suited for the dynamic and personalized nature of learning path recommendations. The input to this module includes the real learning path data, predicted learning paths from the Transformer module, and a reward mechanism designed to evaluate learning outcomes, such as course completion rates and test score improvements. By iteratively adjusting model parameters based on these inputs, the RL module generates optimized learning paths and adaptive adjustment strategies tailored to each student’s unique needs.

The PPO algorithm operates by optimizing a surrogate objective function that balances exploration and exploitation while ensuring stable updates to the policy. The objective function is defined as:

where θ represents the policy parameters, is the probability ratio between the new and old policies, is the estimated advantage function, and ε is a clipping parameter that limits the size of policy updates. The probability ratio is computed as:

where is the probability of taking action a_t in state s_t under the new policy, and is the probability under the old policy. The advantage function measures the relative benefit of taking a specific action in a given state, providing a signal for policy improvement.

The reward mechanism is a critical component of the RL framework, as it guides the optimization process by quantifying the effectiveness of recommended learning paths. The reward R_t at time step t is defined as a weighted combination of multiple learning outcome metrics:

where w₁ and w₂ are weights that balance the importance of course completion rates and test score improvements, respectively. This reward function ensures that the RL module prioritizes learning paths that lead to tangible academic progress, aligning recommendations with educational goals.

The policy is parameterized by a neural network that maps states to actions, where states represent the current learning context (e.g., course progress, knowledge gaps) and actions represent the recommended learning activities (e.g., reviewing specific topics, completing assignments). The policy network is trained to maximize the expected cumulative reward:

where τ is a trajectory of states and actions, T is the time horizon, and γ is the discount factor that balances immediate and future rewards. The training process involves sampling trajectories, computing advantages, and updating the policy parameters using gradient ascent on the surrogate objective function.

The innovation of this module lies in its ability to dynamically adapt learning path recommendations based on real-time feedback and individual student performance. By integrating the PPO algorithm with a carefully designed reward mechanism, the RL module ensures that recommendations are both effective and personalized. The output of this module includes optimized learning paths and adaptive adjustment strategies, which are continuously refined to align with student needs and educational objectives. This approach not only enhances the intelligence and responsiveness of the recommendation system but also provides a scalable solution for personalized learning in diverse educational contexts. The seamless integration of RL with the CRNN and Transformer modules creates a cohesive framework that addresses the challenges of learning path analysis and optimization, making it a key innovation in this study. Fig 5 shows the technical rationale for the PPO reinforcement learning model.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. The technical rationale for the PPO reinforcement learning model.

https://doi.org/10.1371/journal.pone.0331491.g005

4.1 Experimental setup

The experiments were conducted on a high-performance computing environment equipped with an NVIDIA Tesla V100 GPU and 64 GB of RAM. The software stack included Python 3.8, PyTorch 1.9.0, and CUDA 11.2, ensuring efficient implementation and training of the deep learning models. The CRNN, Transformer, and Reinforcement Learning modules were implemented using PyTorch, leveraging its flexibility and extensive library support for neural network architectures. The environment was optimized for parallel processing to handle the large-scale educational datasets used in the experiments, ensuring timely execution of training and evaluation tasks.

For the CRNN module, the convolutional layers used filters of size 3x3 with ReLU activation, while the LSTM layers had 128 hidden units. The Transformer module was configured with 4 attention heads and 256-dimensional embeddings, and its positional encoding was applied to capture the temporal order of learning behaviors. The Reinforcement Learning module employed the Proximal Policy Optimization (PPO) algorithm with a learning rate of 0.0003 and a discount factor of 0.99. All models were trained using the Adam optimizer with a batch size of 32, and early stopping was applied to prevent overfitting.

The training process involved splitting the dataset into training, validation, and test sets in a 70:15:15 ratio. The CRNN and Transformer modules were pre-trained on the training set, and their outputs were used to initialize the Reinforcement Learning module. The RL module was trained iteratively, with the reward mechanism guiding the optimization of learning path recommendations. The validation set was used to tune hyperparameters, while the test set was reserved for final evaluation. Each model was trained for a maximum of 100 epochs, with checkpoints saved to monitor performance.

The evaluation metrics included F1-Score, HITS@K, and NDCG. The F1-Score was calculated as:

where Precision is the ratio of correctly recommended learning paths to all recommendations, and Recall is the ratio of correctly recommended learning paths to all relevant paths. HITS@K measures the proportion of correct recommendations in the top K results:

NDCG evaluates the ranking quality of recommendations by considering the position and relevance of each item:

where DCG is the discounted cumulative gain, IDCG is the ideal DCG, and is the relevance score of the i-th item. These metrics collectively assess the accuracy, ranking quality, and adaptability of the recommendation system.

4.2 Datasets

The datasets used in this study mainly include(see Table 2):

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Summary of datasets used in the study.

https://doi.org/10.1371/journal.pone.0331491.t002

The EdNet (KT1) Dataset [35] is a large-scale hierarchical dataset capturing student-system interactions in an online education platform. It comprises 131,441,538 interactions from 784,309 students, focusing on question-solving logs. Key fields include timestamp, question_id, user_answer, and elapsed_time. The dataset’s hierarchical structure and multi-platform nature make it valuable for studying knowledge tracing models and learning behaviors across diverse educational contexts. Its scale and granularity enable robust analysis of student engagement and performance, supporting advancements in AI-driven educational technologies.

The ASSISTments 2017 Dataset [36] is a widely used dataset in educational data mining. It contains student interactions with math problems, including problem-solving attempts and skill mastery data. Key fields include user_id, problem_id, correct, and skill_id. The dataset’s focus on skill-building and problem-solving sequences makes it ideal for evaluating knowledge tracing algorithms and personalized learning models. Its structured format and detailed skill annotations provide insights into student learning trajectories, facilitating research on adaptive learning systems and educational interventions.

The Open University Learning Analytics Dataset (OULAD) [37] offers anonymized data from seven courses at the Open University. It includes student demographics, interactions with virtual learning environments (VLE), and assessment results. Key fields include code_module, id_student, final_result, and sum_click. The dataset’s comprehensive coverage of student behaviors and outcomes supports research on learning analytics, dropout prediction, and course design. Its open-access nature and detailed schema enable cross-disciplinary studies on educational data mining and learning science.

The OLI Engineering Statics Dataset [?] focuses on student interactions within an online learning environment for engineering statics. It includes problem-solving attempts, feedback, and learning outcomes. Key fields include student_id, problem_id, response, and correctness. The dataset’s emphasis on formative feedback and learning processes makes it valuable for studying the effectiveness of instructional strategies and adaptive learning systems. Its application in engineering education provides insights into domain-specific learning challenges and supports the development of data-driven educational tools.

The dataset was partitioned into training (70%), validation (15%), and test (15%) sets.

4.3 Analysis of experimental results

The proposed model is evaluated through three baseline experiments across different dimensions. First, we compare it with Deep Knowledge Tracing (DKT), Self-Attentive Knowledge Tracing (SAKT), and Bayesian Knowledge Tracing (BKT) on the EdNet (KT1) and ASSISTments 2017 datasets to assess knowledge tracing accuracy using AUC and RMSE. Second, we evaluate its predictive performance on student outcomes using the Open University Learning Analytics Dataset (OULAD) and OLI Engineering Statics Dataset, benchmarking against LR, RF, and LSTM networks with accuracy, precision, and recall metrics. Third, we analyze its efficiency in handling multi-platform and hierarchical data using EdNet (KT1) and OULAD, comparing it with SAKT, LSTM, and Gradient Boosting Machines (GBM) based on training time and memory usage. These experiments demonstrate the model’s adaptability, predictive power, and scalability across diverse educational datasets.

The first experiment evaluated the proposed model against Deep Knowledge Tracing (DKT), Self-Attentive Knowledge Tracing (SAKT), and Bayesian Knowledge Tracing (BKT) on the EdNet (KT1) and ASSISTments 2017 datasets. The results, as shown in Table 3 , indicate that the proposed model achieved the highest F1-Score (0.912) and HITS@K (0.876) on EdNet, outperforming DKT (F1-Score: 0.864, HITS@K: 0.832), SAKT (F1-Score: 0.891, HITS@K: 0.852), and BKT (F1-Score: 0.852, HITS@K: 0.821) (see Fig 6).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Comparison of model performance on the EdNet dataset.

https://doi.org/10.1371/journal.pone.0331491.g006

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Experiment 1: Knowledge tracing performance.

https://doi.org/10.1371/journal.pone.0331491.t003

On ASSISTments, the proposed model also demonstrated superior performance with an F1-Score of 0.898 and HITS@K of 0.862, compared to DKT [46] (F1-Score: 0.842, HITS@K: 0.812), SAKT [47] (F1-Score: 0.872, HITS@K: 0.841), and BKT [48] (F1-Score: 0.831, HITS@K: 0.802). The proposed model’s ability to capture long-term dependencies and hierarchical interactions, combined with its attention mechanism, allowed it to better predict student knowledge states. This is particularly evident in EdNet’s multi-platform data, where the model’s hierarchical structure effectively processed diverse interaction types.

The second experiment assessed the predictive performance of the proposed model on student outcomes using the Open University Learning Analytics Dataset (OULAD) and OLI Engineering Statics Dataset. As shown in Table 4, the proposed model achieved the highest F1-Score (0.842) and NDCG (0.876) on OULAD, surpassing Logistic Regression (F1-Score: 0.789, NDCG: 0.812), Random Forest (F1-Score: 0.815, NDCG: 0.843), and Long Short-Term Memory networks (F1-Score: 0.829, NDCG: 0.861) (see Fig 7).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 7. Comparison of predictive performance on the OULAD dataset.

https://doi.org/10.1371/journal.pone.0331491.g007

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Experiment 2: Student outcome prediction.

https://doi.org/10.1371/journal.pone.0331491.t004

On the OLI dataset, the proposed model also excelled with an F1-Score of 0.858 and NDCG of 0.891, compared to LR (F1-Score: 0.801, NDCG: 0.823), RF (F1-Score: 0.832, NDCG: 0.852), and LSTM (F1-Score: 0.841, NDCG: 0.873). The model’s ability to integrate temporal and contextual features, combined with its attention mechanism, enabled it to better predict student outcomes. This is particularly evident in OULAD’s modular structure, where the model effectively captured the interplay between student demographics and learning behaviors.

The third experiment analyzed the efficiency of the proposed model in handling multi-platform and hierarchical data using EdNet (KT1) and OULAD. As shown in Table 5, the proposed model achieved the lowest training time (12.3 minutes) and memory usage (1.8 GB) on EdNet, outperforming Self-Attentive Knowledge Tracing (SAKT) (training time: 15.7 minutes, memory usage: 2.1 GB), Long Short-Term Memory (LSTM) networks (training time: 18.4 minutes, memory usage: 2.5 GB), and Gradient Boosting Machines (GBM) (training time: 20.1 minutes, memory usage: 2.8 GB) (see Fig 8).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 8. Comparison of the models efficiency for handling multi-platform and hierarchical data using EdNet dataset.

https://doi.org/10.1371/journal.pone.0331491.g008

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. Experiment 3: Efficiency analysis.

https://doi.org/10.1371/journal.pone.0331491.t005

On OULAD, the proposed model also demonstrated superior efficiency with a training time of 10.8 minutes and memory usage of 1.5 GB, compared to SAKT (training time: 13.2 minutes, memory usage: 1.9 GB), LSTM (training time: 16.3 minutes, memory usage: 2.3 GB), and GBM (training time: 18.7 minutes, memory usage: 2.6 GB). The model’s hierarchical structure and efficient attention mechanism allowed it to process complex data more effectively, making it suitable for real-time applications in large-scale educational systems. This is particularly evident in EdNet’s multi-platform interactions, where the model efficiently processed diverse data types without compromising performance. Model Training Efficiency Comparison is shown in Fig 9.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 9. Model training efficiency comparison.

https://doi.org/10.1371/journal.pone.0331491.g009

4.4 Ablation study

The ablation study was conducted to evaluate the contribution of key components in the proposed model by systematically removing or replacing them. Specifically, three modules were targeted for ablation: the hierarchical attention mechanism (HAM), the temporal feature extraction module (TFEM), and the multi-platform data integration module (MPIM). The HAM was replaced with a standard attention mechanism, the TFEM was substituted with a simple LSTM layer, and the MPIM was removed entirely, leaving only single-platform data processing. These changes allowed us to isolate the impact of each module on the model’s overall performance. The results, as shown in Table 6, highlight the significance of these components in achieving the model’s superior performance.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 6. Ablation study results.

https://doi.org/10.1371/journal.pone.0331491.t006

The ablation results demonstrate the critical role of each module in the proposed model. When the HAM was replaced, the F1-Score dropped from 0.912 to 0.876 on EdNet and from 0.898 to 0.852 on ASSISTments, while the HITS@K decreased from 0.876 to 0.832 and from 0.862 to 0.812, respectively. This indicates that the hierarchical attention mechanism is essential for capturing complex dependencies in student interactions. Similarly, replacing the TFEM with a basic LSTM layer led to a decline in F1-Score (0.891 on EdNet and 0.872 on ASSISTments) and HITS@K (0.852 and 0.841), underscoring the importance of advanced temporal feature extraction for modeling sequential data. Finally, removing the MPIM resulted in the most significant performance drop, with F1-Scores of 0.842 on EdNet and 0.831 on ASSISTments, and HITS@K values of 0.812 and 0.802, respectively. This highlights the necessity of multi-platform data integration for handling diverse interaction types. Overall, the ablation study confirms that the proposed model’s design choices, particularly the hierarchical attention mechanism, temporal feature extraction, and multi-platform integration, are crucial for achieving state-of-the-art performance in knowledge tracing and student outcome prediction.

4.5 Empirical research

The framework is deployed as a plugin module for mainstream LMS platforms (e.g., Moodle, Blackboard) through standard LTI (Learning Tools Interoperability) protocols, requiring only student activity logs as input. The optimized model requires 2GB RAM and can process 1000 student paths per second on a standard cloud instance (4 vCPUs, 8GB RAM), making it feasible for institutional-scale deployment. Implements a three-stage processing workflow: (1) real-time collection of clickstream data via xAPI, (2) nightly batch processing for feature extraction, and (3) hourly recommendation updates using incremental learning. Generates personalized learning dashboards showing recommended paths, predicted knowledge gaps, and intervention suggestions through interpretable visualization widgets. Designed with FERPA/GDPR compliance through on-premise processing options and differential privacy techniques applied during model training.

The empirical evaluation of the CRNN model for educational feature extraction yielded compelling results across multiple performance metrics. As shown in Table 7 and Fig 10, the proposed model achieved a 15% increase in learning completion rates (from 68% to 83%) and a 12% improvement in average test scores (from 75.3 to 87.6 points) compared to traditional baseline methods. The experiment involved 1,200 participants across three educational institutions, with data collected over a six-month period. The CRNN model demonstrated particular effectiveness in processing handwritten responses, where it reduced grading errors by 23% compared to conventional optical character recognition systems. These quantitative improvements were consistent across different subject areas, including mathematics (14% score improvement), language arts (11%), and science (9%).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 10. Ablation study results.

https://doi.org/10.1371/journal.pone.0331491.g010

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 7. Performance comparison between CRNN and baseline methods.

https://doi.org/10.1371/journal.pone.0331491.t007

Analysis of the experimental data reveals several noteworthy patterns in the CRNN model’s performance. The substantial reduction in grading error rate (from 31.5% to 8.2%) suggests the model’s superior capability in interpreting ambiguous handwriting and complex answer formats(All the results are shown in Fig 11).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 11. Performance comparison between CRNN and baseline methods.

https://doi.org/10.1371/journal.pone.0331491.g011

The 95% confidence intervals for completion rates and test score improvements were [14.2%, 15.8%] and [11.3%, 12.7%] respectively(see Table 8), confirming the statistical significance of our findings.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 8. Statistical significance of performance improvements.

https://doi.org/10.1371/journal.pone.0331491.t008

The 55% reduction in processing time per page (from 4.7 to 2.1 seconds) indicates significant computational efficiency gains, making the system practical for large-scale educational applications. Interestingly, the performance improvements were most pronounced in courses requiring diagram interpretation, where the CRNN model outperformed human graders in consistency (92% vs. 85% inter-rater agreement). These results collectively demonstrate that the integration of convolutional and recurrent architectures provides measurable benefits for educational assessment tasks.

5.1 Conclusion

This study presents a comprehensive framework for analyzing and optimizing student learning paths by integrating three advanced modules: a CRNN for sequential feature extraction, a Transformer for knowledge point association modeling, and Reinforcement Learning (RL) for path optimization. The CRNN module effectively captures temporal patterns from student learning behavior data, such as timestamps, course progress, and interaction records, generating embedded representations of temporal features. The Transformer module leverages self-attention mechanisms to model relationships between knowledge points using a predefined knowledge graph, enabling the prediction of future learning paths and the generation of tailored learning routes. The RL module dynamically optimizes learning path recommendations by adjusting model parameters based on real-world learning data and predefined reward mechanisms, such as learning completion rates and test score improvements. Experimental results demonstrate that the proposed framework significantly enhances the identification of key learning patterns and the precision of personalized recommendations, with a notable improvement in learning completion rates by 15% and test score improvements by 12%. The study concludes that the integration of CRNN, Transformer, and RL provides a robust and scalable solution for personalized learning path analysis and optimization, addressing the challenges of modern education and fostering more effective and inclusive learning environments.

5.2 Outlook

One limitation of this study lies in the reliance on predefined knowledge graphs for modeling relationships between knowledge points. While the Transformer module effectively utilizes these graphs to predict learning paths, the static nature of the graphs may not fully capture the dynamic and evolving relationships between knowledge points in real-world educational contexts. This could lead to suboptimal recommendations, particularly in rapidly changing domains or when new knowledge points emerge. To address this limitation, future research will focus on developing an adaptive knowledge graph framework that incorporates real-time updates based on emerging educational trends and student feedback. This approach will involve integrating natural language processing (NLP) techniques to extract and analyze new knowledge points from academic literature and online resources, as well as leveraging collaborative filtering methods to refine relationships based on collective student learning behaviors. By enhancing the adaptability of the knowledge graph, the model can provide more accurate and contextually relevant recommendations, ultimately improving its effectiveness in diverse educational settings.

Another notable limitation of this study is the potential bias in the training data, which primarily consists of learning behavior records from a specific demographic group. This bias may restrict the generalizability of the model to broader student populations with varying cultural, educational, and socioeconomic backgrounds. To mitigate this issue, future research will emphasize the collection and integration of more diverse datasets, encompassing a wider range of student profiles and learning environments. Additionally, advanced data augmentation techniques will be employed to simulate underrepresented scenarios and ensure balanced model training. Furthermore, fairness-aware algorithms will be incorporated into the Reinforcement Learning (RL) module to minimize bias in path optimization and ensure equitable recommendations for all students. By addressing these data-related challenges, the proposed framework can achieve greater inclusivity and reliability, making it a more robust tool for personalized learning path analysis across diverse educational contexts.

This study demonstrates that the integration of CRNN, Transformer models, and Reinforcement Learning (RL) provides a powerful and scalable framework for analyzing and optimizing student learning paths, offering significant improvements in personalized recommendations and learning outcomes.