Online course MOOC learning prediction method based on heterogeneous feature fusion

Abstract

The invention provides an online course MOOC learning prediction method based on heterogeneous feature fusion, and belongs to the field of computer-aided intelligent education. The method comprises the following steps: acquiring an MOOC data set, and preprocessing the data set to obtain a test set; constructing an IHFNet network comprising a multi-agent adaptive static feature selector module, a hierarchical time sequence feature extractor module and a heterogeneous feature fusion module; a multi-agent adaptive static feature selector module screens key static features; the hierarchical time sequence feature extractor module extracts behavior time sequence features with high discrimination ability; the heterogeneous feature fusion module carries out adaptive fusion on the key static features and the behavior time sequence features and carries out classification prediction; an IHFNet network is trained; and collecting MOOC data of a to-be-predicted learner, and inputting the MOOC data to the trained IHFNet network for learning risk prediction. According to the invention, modeling is carried out by fusing the static features of the learner and the behavior time sequence features, the feature representation ability of the learner is enhanced, and the learning risk prediction effect is improved.

Description

Technical Field

[0001] This invention relates to a method for predicting dropout rates in online courses (MOOCs) based on heterogeneous feature fusion, belonging to the field of computer-aided smart education. Background Technology

[0002] With the rapid development of internet technology, online education has become an important component of higher education. Massive Open Online Courses (MOOCs), due to their openness and flexibility, are widely used in various online teaching and learning scenarios. However, in practice, MOOC platforms generally suffer from high dropout rates, with many learners withdrawing midway through their courses. This not only affects the learners' personal development but also seriously impacts the effectiveness of online teaching and the efficiency of educational resource utilization. Therefore, accurately predicting the risk of learners dropping out is of significant practical importance.

[0003] Existing dropout prediction methods mainly include those based on traditional machine learning and those based on deep learning. Traditional methods typically rely on manually constructed features, which struggles to characterize the complex nonlinear relationships in learning behavior and has limited generalization ability in high-dimensional, multi-source learning data scenarios. Deep learning methods often improve prediction performance to some extent by modeling the temporal features of learning behavior; however, most methods focus on temporal modeling of behavior, often neglecting the role of learners' static background features. This results in features with limited dimensionality and insufficient information coverage, making it difficult to accurately characterize learner differences and thus affecting prediction results.

[0004] Furthermore, existing models suffer from weak interpretability, making it difficult to clearly define the specific contributions of key features to dropout prediction, thus limiting their practical application in teaching interventions and learning management. Therefore, it is necessary to propose a method for predicting learner dropout risk that ensures both accuracy and good interpretability. Summary of the Invention

[0005] The purpose of this invention is to provide a method for predicting dropout rates in online courses (MOOCs) based on heterogeneous feature fusion. This method accurately predicts and analyzes the dropout risk of MOOC learners by fusing learners' static features and behavioral temporal features into a model, thereby enhancing learners' feature representation capabilities and improving the effectiveness of dropout risk prediction.

[0006] To achieve the above objectives, the present invention employs the following technical solution: A method for predicting dropout rates in online courses (MOOCs) based on heterogeneous feature fusion includes: Obtain the MOOC dataset and preprocess the dataset to obtain the test set, including data cleaning and binarization of the learner's final learning results. The test set includes static background feature vectors and behavioral time series sequences. An IHFNet network is constructed and trained, comprising a multi-agent adaptive static feature selector module, a hierarchical temporal feature extractor module, and a heterogeneous feature fusion module. The multi-agent adaptive static feature selector module filters key static features from static background features. The hierarchical temporal feature extractor module extracts high-discriminative behavioral temporal features from behavioral temporal sequences. The heterogeneous feature fusion module adaptively fuses the key static features with the behavioral temporal features and performs classification prediction based on the fused features. Train the IHFNet network using the training set; MOOC data of learners to be predicted are collected and input into the trained IHFNet network to predict dropout risk.

[0007] Preferably, the multi-agent adaptive static feature selector module consists of a multi-agent decision layer, a state and reward function layer, a policy optimization layer, and a feature output layer; An agent is configured for each static background feature vector, and each agent performs iterative decision optimization based on the ε-greedy policy and Q-value update mechanism to select a subset of key static features.

[0008] Preferably, the multi-agent adaptive static feature selector module employs a causal reward mechanism based on counterfactual reasoning, where the causal reward for a single agent... for: , in, For overall performance, For counterfactual performance, For intelligent agents exist Actions at any moment This is the counterfactual action vector.

[0009] Preferably, the hierarchical temporal feature extractor module includes multiple sets of parallel one-dimensional convolutional layers, each using a convolutional kernel of a different size; after each set of convolutional branches, a batch normalization layer, a nonlinear activation layer, and a pooling layer are sequentially connected to obtain multi-scale features; after concatenation, the multi-scale features are sequentially input into a fully connected layer and two Transformer encoder layers for processing; the features obtained by the Transformer encoder layer and the concatenated multi-scale features are connected through residuals, and then output as temporal behavior features after layer normalization.

[0010] Preferably, the kernel sizes of the multiple parallel one-dimensional convolutional layers are set to 3×1, 5×1 and 7×1, respectively.

[0011] Preferably, the heterogeneous feature fusion module includes: The feature projection layer includes two parallel multilayer perceptrons, which are used to map the key static features and the behavioral temporal features to the same feature space, respectively. The attention interaction fusion layer is used to weightedly fuse the two types of features after mapping through an attention mechanism; The residual normalization enhancement layer is used to combine residual connections with normalization operations to stabilize the training process. A classifier is used to output dropout risk prediction results based on the fused features.

[0012] Preferably, the IHFNet network uses a combination of weighted cross-entropy loss and feature fusion constraint term as the loss function; the Adam optimization algorithm is used to update the model parameters, and the initial learning rate is set to 0.001; the Dropout regularization mechanism is introduced to suppress overfitting. The dataset was trained and tested using a 10-fold cross-validation method. The dataset was randomly divided into 10 non-overlapping subsets, and training and testing were performed sequentially. The average of the experimental results of each fold was taken as the final prediction performance.

[0013] Preferably, the loss function as follows: , , , in, For weighted cross-entropy loss, For feature fusion constraint terms, To balance the hyperparameters of the main loss and constraint terms; The total number of samples in the training dataset. For the first The true label of each sample The first output of the model The dropout probability of a sample; The weights of the positive dropout samples. Weights for non-dropout negative samples; These are the key static feature vectors after processing by the projection layer. This is the behavioral temporal feature vector after processing by the projection layer. This is the cosine similarity function.

[0014] Preferably, the IHFNet network further includes a SHAP interpretability analysis module, used to perform contribution quantification analysis on the key static features selected by the multi-agent adaptive static feature selector module; The IHFNet network also includes a behavior trend analysis module, which performs learning trend analysis based on the behavior time-series features extracted by the hierarchical time-series feature extractor module, revealing the characteristics of learner activities under different learning trends.

[0015] The advantages of this invention are as follows: This invention adopts heterogeneous feature fusion technology, adaptively selects key static features and combines them with temporal feature modeling, fully captures learner features, avoids the limitations of single features, significantly improves the accuracy and generalization ability of dropout prediction, and can meet the actual application needs of online education platforms.

[0016] To address the black-box nature of existing models, this invention introduces the SHAP interpretability analysis mechanism to quantify the contribution of key static features. Simultaneously, it analyzes typical learning behaviors, clearly revealing the causes of dropout risk and providing a clear basis for educational intervention, thus solving the problem of traditional models only predicting but not explaining. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0018] Figure 1 This is a flowchart illustrating the implementation of the present invention.

[0019] Figure 2 This is a schematic diagram of the main framework of the network structure proposed in this invention.

[0020] Figure 3 This is a schematic diagram of the multi-agent adaptive static feature selector submodule of the network structure proposed in this invention.

[0021] Figure 4 This is a schematic diagram of the network structure hierarchical temporal feature extractor submodule proposed in this invention.

[0022] Figure 5 This invention presents a bar chart displaying the SHAP visualization results of key static features.

[0023] Figure 6 This invention presents a swarm diagram that visualizes the results of SHAP on key static features.

[0024] Figure 7 This is a visualization graph showing the results of the analysis of behavioral temporal features by the present invention.

[0025] Figure 8 This is a visualization of the DDD course results obtained by analyzing the temporal characteristics of behavior according to the present invention.

[0026] Figure 9 This is a visualization of the EEE course results obtained by analyzing the temporal characteristics of behavior according to the present invention.

[0027] Figure 10 This is a visualization of the FFF course results obtained by analyzing the temporal characteristics of behavior according to the present invention. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] Example 1 like Figures 1-2 As shown, a method for predicting dropout rates in online courses (MOOCs) based on heterogeneous feature fusion includes the following steps: S1: Obtain the MOOC dataset (OULAD) and preprocess the dataset, including data cleaning and binarization of the learners' final learning results to obtain a test set, which includes static background feature vectors and behavioral time series sequences; S2: Construct and train an IHFNet network comprising a multi-agent adaptive static feature selector module, a hierarchical temporal feature extractor module, and a heterogeneous feature fusion module; the multi-agent adaptive static feature selector module filters key static features from static background features; the hierarchical temporal feature extractor module extracts high-discriminative behavioral temporal features from behavioral temporal sequences; the heterogeneous feature fusion module adaptively fuses the key static features with the behavioral temporal features and performs classification prediction based on the fused features.

[0030] S3: Train the IHFNet network using the training set; S4: Collect MOOC data of learners to be predicted and input it into the trained IHFNet network to predict dropout risk.

[0031] As a refinement of the above embodiments, dataset preprocessing includes performing systematic data cleaning operations to remove missing values, duplicate samples, and abnormal learning records, ensuring data quality and analytical reliability. Subsequently, the learner's final learning result (final_result) is binarized for use in training the network model.

[0032] This dataset contains learning records from 32,593 students across seven modules (AAA–GGG), covering STEM and social science subjects. Combining learners' demographic information with clickstream data from virtual learning environments (VLE), preliminary analysis reveals that dropout rates for STEM courses are generally higher than for social science courses. This trend may be related to the higher cognitive demands, complex concepts, and emphasis on practical skills in STEM courses. Based on this, this study selected three representative STEM courses (DDD, EEE, and FFF) for model development and evaluation, aiming to accurately predict dropout rates and reveal the potential reasons for the high dropout rate in STEM courses, thereby providing empirical evidence for reducing dropout rates, curriculum design, and learning interventions.

[0033] As a refinement of the above embodiments, in response to the problem that existing multi-agent feature selection methods generally adopt global rewards and have difficulty distinguishing the contribution of individual features, this invention introduces a multi-agent adaptive static feature selector and designs a causal reward mechanism based on counterfactual reasoning to characterize the marginal impact of individual agent decisions on overall prediction performance, thereby achieving a more refined and fair reward allocation.

[0034] like Figure 3 As shown, the multi-agent adaptive static feature selector module adopts a "one agent, one feature" configuration mode, consisting of two or more feature selection agents and a collaborative decision-making unit. Each agent selects and explores features based on an ε-greedy strategy, and continuously optimizes the decision-making strategy through a Q-value update mechanism. This automatically filters out a subset of key static features that contribute significantly to dropout prediction under global collaborative constraints, reducing interference from redundant features. The final output is the filtered key static features.

[0035] Specifically, the multi-agent adaptive static feature selector module consists of a multi-agent decision layer, a state and reward function layer, a policy optimization layer, and a feature output layer.

[0036] In this method, an independent agent is assigned to each static feature, and the agents make decisions in parallel in a shared environment, with their actions taking a binary form. This indicates whether the corresponding feature is selected in the current iteration, with 1 indicating selection and 0 indicating non-selection. It represents the joint action vector of all agents at the same time. These features together constitute a candidate static feature subset, and the prediction model is trained and its performance evaluated based on this feature subset to obtain global performance feedback. When an agent's action changes from its previous action, a counterfactual action vector is constructed only for that agent. Its definition is: , That is, while keeping the actions of other agents unchanged, only the current decision of this agent is rolled back. At that time, only the target intelligent agent The action of one agent rolls back to the previous moment, while the actions of the other agents remain in their current state; when At that time, all other intelligent agents except the target intelligent body... The actual actions at the current moment are retained. A corresponding feature subset is generated based on this counterfactual action vector, and the model performance is re-evaluated to obtain the counterfactual performance. .

[0037] Based on the above counterfactual reasoning process, the causal reward of the agent is defined as the difference between the model performance under the actual action and the counterfactual action: , This reward directly reflects the marginal contribution of the agent's current decision to the overall prediction performance. When A value greater than 0 indicates that the feature decision improved the model performance; conversely, a value greater than 0 indicates that the decision had a negative impact on performance.

[0038] After receiving causal rewards, each agent updates its policy based on the Q-learning mechanism and enters the next round of decision-making. Through the iterative process of "feature selection decision-making - counterfactual performance evaluation - causal reward allocation - policy update", the multi-agent system gradually converges, ultimately obtaining the optimal static feature subset that can balance prediction accuracy and the interpretability of feature contributions.

[0039] As a refinement of the above embodiments, such as Figure 4 As shown, the hierarchical temporal feature extractor module is used to model the dynamic features of learner behavior over time; it includes multiple sets of parallel one-dimensional convolutional layers, each using convolutional kernels of different sizes, and using convolutional kernels of different scales to extract local patterns from the temporal behavior; after each set of convolutional branches, a batch normalization layer, a nonlinear activation layer, and a pooling layer are sequentially connected to obtain multi-scale features. To enhance feature representation and reduce noise impact, multi-scale local temporal features under different receptive fields were obtained. These multi-scale features were then concatenated to form a unified local temporal feature representation. The data is sequentially input into a fully connected layer and two Transformer encoder layers for processing. The features obtained from the Transformer encoder layers are then processed. After concatenating the multi-scale features and performing residual connections, the resulting temporal behavior features are normalized through layer normalization to further capture long-range dependencies in the behavior sequence, thus outputting highly discriminative temporal behavior features. The calculation formula is as follows: , , , , in, The convolution kernel weight matrix is... The input is a time sequence matrix. This is the bias vector for the convolutional layer. This indicates a pooling operation. This represents the activation function. This indicates a normalization operation. This indicates a feature concatenation operation. This indicates a multi-head self-attention mechanism. Normalization function, , , These are query, key, and value vectors, respectively. Let be the dimension of the key vector. Presentation layer normalization operation.

[0040] The kernel sizes of the multiple parallel one-dimensional convolutional layers are set to 3×1, 5×1 and 7×1, respectively.

[0041] As a refinement of the above embodiments, the heterogeneous feature fusion module is used to realize deep interactive modeling between static features and behavioral temporal features. It includes a feature projection layer, an attention interaction fusion layer, and a residual normalization enhancement layer. The feature projection layer maps features from different modalities to a unified feature space. Then, the attention interaction fusion layer performs weighted fusion of the two types of features, and the residual connection and normalization enhancement layers stabilize the training process. The fused features are finally input into the classifier, outputting the learner's dropout risk prediction result.

[0042] Specifically, the function of each layer is as follows: The feature projection layer includes two parallel multilayer perceptrons, which are used to map the key static features and the behavioral temporal features to the same feature space, respectively. The attention interaction fusion layer is used to weightedly fuse the two types of features after mapping through an attention mechanism; The residual normalization enhancement layer is used to combine residual connections with normalization operations to stabilize the training process. A classifier is used to output dropout risk prediction results based on the fused features.

[0043] The specific formula is as follows: , , , in, Represents the key static feature vector. It represents the temporal characteristics of behavior. This represents a multilayer perceptron mapping network. These are the key static feature vectors after processing by the projection layer. This is the behavioral temporal feature vector after processing by the projection layer. This represents the fusion feature representation.

[0044] As a refinement of the above embodiments, the IHFNet network further includes a SHAP interpretability analysis module, which is used to perform contribution quantification analysis on the key static features selected by the multi-agent adaptive static feature selector module. The IHFNet network also includes a behavior trend analysis module, which performs learning trend analysis based on the behavior time-series features extracted by the hierarchical time-series feature extractor module, revealing the characteristics of learner activities under different learning trends.

[0045] By quantifying the contribution of key static features to dropout risk through SHAP analysis and further identifying activity features in three typical learning trajectories through behavioral feature mining analysis, the interpretability of the model is enhanced. The model achieves both prediction and interpretability, providing a more comprehensive solution for learner support systems in the education field.

[0046] As a refinement of the above embodiments, the specific method for training the IHFNet network using the training set is as follows: (1) Loss Function and Optimization Strategy. During the model training phase, the cross-entropy loss function was used as the objective function to measure the difference between the model's predicted dropout probability distribution and the true label. The model parameters were updated using the Adam optimization algorithm, with an initial learning rate set to 0.001 to ensure stable convergence of the model during training. Simultaneously, Dropout regularization was introduced during network training to effectively suppress overfitting.

[0047] The loss function as follows: , , , , in, The weighted cross-entropy loss is used to address the imbalance of positive and negative samples in the dataset of MOOC dropout prediction tasks. This is a feature fusion constraint term used to ensure consistency between the projected key static features and the behavioral temporal features before fusion. The hyperparameters, which balance the main loss and constraint terms, range from 0.01 to 0.1. The total number of samples in the training dataset. For the first The true label of each sample The first output of the model The dropout probability of a sample; The weights of the positive dropout samples. Weights for non-dropout negative samples; These are the key static feature vectors after processing by the projection layer. This is the behavioral temporal feature vector after processing by the projection layer. This is the cosine similarity function, used to calculate the similarity between two heterogeneous feature vectors. The number of dropout samples in the training dataset.

[0048] (2) Data Partitioning and Training Strategy. To comprehensively evaluate model performance and improve the reliability of results, this invention employs a 10-fold cross-validation method for training and testing the dataset. Specifically, the MOOC learner dataset is randomly divided into 10 non-overlapping subsets. In each round, one subset is selected as the test set, and the remaining subsets are used as the training set. This process is repeated for all folds, and the average of the experimental results from each fold is taken as the final prediction performance. During training, the batch size is set to 32 to balance computational efficiency and storage resource consumption.

[0049] (3) Evaluation Metrics. To comprehensively evaluate the model's performance in the learner dropout prediction task, this invention uses AUC, accuracy, recall, and F1 score as evaluation metrics. AUC measures the model's overall ability to distinguish between dropouts and non-dropout learners at different discrimination thresholds; Accuracy reflects the consistency between the model's prediction results and the true labels, calculated as (TP+TN) / (TP+TN+FP+FN); Recall measures the model's ability to identify real dropout learners, calculated as TP / (TP+FN); and the F1 score comprehensively evaluates the performance of the binary classification model, calculated as 2TP / (2TP+FP+FN). Wherein, TP (True Positives) represents the number of samples correctly predicted as dropout learners, TN (True Negatives) represents the number of samples correctly predicted as non-dropout learners, FP (False Positives) represents the number of non-dropout learners incorrectly predicted as dropouts, and FN (False Negatives) represents the number of dropout learners incorrectly predicted as non-dropouts. The above evaluation metrics comprehensively verify the effectiveness of the method of this invention from dimensions such as overall predictive performance, model discriminative ability, and sensitivity to identifying learners at risk of dropping out.

[0050] This invention proposes a novel method for dropout prediction and interpretability analysis based on heterogeneous feature fusion. This method can simultaneously model learners' static background features and behavioral temporal features, making it applicable to different course scenarios and learning behavior patterns, and exhibiting strong predictive stability across various learning scenarios. The method achieved high predictive performance on the OULAD dataset test set, as shown in Table 1, with an accuracy of 91.43%, recall of 90.19%, AUC of 91.94%, and F1 score of 88.79%, effectively identifying learners at high dropout risk. Comparative experimental results demonstrate that the method of this invention has significant advantages in predictive performance, as shown in Table 1. Compared with traditional dropout prediction methods, this invention avoids complex manual feature design and single-feature modeling, exhibiting significant advantages in prediction accuracy and model generalization ability. Furthermore, it introduces an interpretability analysis mechanism, providing clear and understandable decision-making basis for dropout risk identification results.

[0051] Table 1 Comparative Experiments Regarding the SHAP interpretability analysis module: Most existing deep learning models exhibit "black box" characteristics, lacking sufficient interpretability and exhibiting a "predicts but doesn't explain" phenomenon. This limits the practical application of these models, especially in the field of education, where teachers and policymakers need to understand which characteristics play a crucial role in predicting dropout risk in order to implement targeted interventions. Therefore, we conduct interpretability analyses on learners' static and behavioral characteristics, respectively.

[0052] SHAP visualization analysis was used to analyze a subset of key static features, quantifying the contribution of each feature to dropout prediction. Figure 5 The data shows that the studied credits feature has the highest importance score, indicating that this feature has the most significant impact on the final prediction results. Socioeconomic background (IMD_band), disability status (disability_Y), and highest education level (highest_education_PostGraduateQualification) also show high contributions, indicating that learners' external environment and individual physical differences have a significant impact on dropout risk. Although the region feature has a lower importance, it persists in the feature subset, indicating that geographical background still influences dropout prediction results to some extent.

[0053] Figure 6The graph clearly depicts the quantitative relationship between each key static feature and the dropout risk prediction results, clarifying the positive or negative impact of different feature values ​​on dropout risk. As shown in the graph, higher studied credits correspond to lower dropout probabilities, while lower credits indicate higher dropout risk. This suggests that students with fewer credits may face more challenges or lack sufficient motivation, leading to a higher dropout risk. A higher value for the feature "imd_band_90-100%" (the top 10% of socioeconomic backgrounds) corresponds to a significantly negative SHAP value, indicating that this feature has a protective effect against dropout risk and effectively reduces the probability of learners interrupting their studies. This may be attributed to the fact that affluent economic conditions provide material security and support for continued learning. Conversely, a higher value for the feature "imd_band_0-10%" (the bottom 10% of socioeconomic backgrounds) corresponds to a significantly positive SHAP value, indicating that this feature increases dropout risk. This is consistent with the negative impact of multiple deprivation factors such as economic pressure and insufficient educational resources on learning continuity. Higher disability (disability_Y) eigenvalues ​​did not correlate with a significant increase in dropout rates. This suggests that while students with disabilities face multiple real-world challenges in learning, including physical and environmental adaptation difficulties, they may possess strong psychological resilience and proactive learning strategies. Students with the highest education (Post Graduate Qualification) were less likely to drop out, demonstrating stronger self-directed learning abilities. Regionally, students in London, Scotland, and the south had lower dropout risks, while students in Yorkshire had higher dropout risks as the eigenvalue increased. This may be related to differences in digital infrastructure, educational support, and community learning cultures across regions.

[0054] Regarding the trend analysis module: To explore the relationship between learning behavior and dropout risk, this paper categorizes learners' behavior into three patterns based on the slope and variance of their clickstream time series: stable, fluctuating, and declining. Stable learners exhibit relatively stable or gradually increasing behavioral intensity; fluctuating learners show periodically changing behavioral intensity; and declining learners show a continuously decreasing behavioral intensity.

[0055] Further integrate learners' weekly behavioral data from the three behavioral models to analyze changes in learners' weekly behavior, such as Figure 7 As shown, stable learners exhibited significantly higher engagement than other behavioral patterns in the early stages of the course. Although engagement declined somewhat in the initial phase (first ten weeks) and showed a gradual downward trend as the course progressed, engagement remained at a relatively high level throughout the learning process, significantly higher than that of fluctuating and declining learners.

[0056] It is noteworthy that the engagement trajectories of the three groups of learners began to diverge significantly at the beginning of the course, suggesting that early behavioral patterns may play a crucial role in subsequent learning persistence. Therefore, we further examined the flow patterns of learners' activities from a temporal evolution perspective, particularly the migration of different learning activities from the first 10 days to the last 20 days. This analysis aims to reveal the dynamic adjustment of early learning priorities and engagement strategies under different learning trends, such as... Figures 8-10 As shown: In the DDD curriculum, stable learners showed higher turnover rates in collaborative activities (ouwiki, 3.78) and virtual classrooms (ouelluminate, 1.38), indicating increased interactivity and engagement over the following 20 days. In contrast, participation in oucontent, glossary, and external quiz remained stable. This behavioral pattern depicts a learning process from knowledge construction and internalization to deepening interaction, highlighting learners' progressively increasing active participation and interactive behavior. Fluctuating learners also showed higher turnover rates in collaborative activities (ouwiki, 2.61) and virtual classrooms (ouelluminate, 2.00), as well as a higher turnover rate in external quiz (external quiz, 1.06). However, compared to stable learners, their turnover rates in other activities were generally lower, indicating greater volatility and lower motivation in their behavioral patterns. Declining learners have higher turnover rates in collaborative activities (ouwiki, 2.76) and glossary (2.07), indicating that they may actively participate in knowledge construction and understanding key terms in the early stages of learning, but gradually reduce their participation in other learning activities over time, suggesting that their learning motivation or sustained engagement begins to decline.

[0057] In the EEE course, stable learners showed high turnover rates in dualpane (2.59), collaborative platforms (oucollaborate, 2.32), quizzes (3.36), and URL access (2.48). Their participation in other activities remained relatively stable. This pattern indicates an open learning process that integrates active resource exploration and collaborative co-creation, and makes reasonable use of platform tools, reflecting an active exploration learning model. Fluctuating learners showed high turnover rates in collaborative activities (ouwiki, 2.67), quizzes (2.75), and URL access (2.52), but their participation in other learning activities was relatively low compared to stable learners. This suggests insufficient utilization of basic learning resources, leading to instability in their learning behavior. Decreasing learners showed relatively high transition rates in quizzes (5.57), collaborative activities (ouwiki, 4.00), and collaborative platforms (oucollaborate, 3.03). This pattern suggests that, despite low overall engagement, these learners still intermittently rely on assessment-driven efforts and collaborative support to compensate for learning deficiencies.

[0058] In the FFF course, all learning trend groups showed high turnover rates in the questionnaires, indicating that these tasks consistently served as core engagement points throughout the 30-day learning cycle. Stable and fluctuating learners showed high turnover rates in collaborative activities (ouwik) and dual panes, demonstrating that collaborative learning was a core strategy and that they effectively utilized multi-task learning tools. However, in other activities, fluctuating learners had relatively lower turnover rates compared to stable learners, reflecting lower motivation and a lack of sustained engagement depth. In contrast, declining learners showed higher turnover rates in resources (0.99) and subpages (0.95), indicating that they relied more on independent research and access to specific learning materials and that they were gradually shifting towards independent learning.

[0059] Example 2 This disclosure also provides an online course (MOOC) dropout prediction device based on heterogeneous feature fusion, including a processor and a memory. Optionally, the device may further include a communication interface and a bus. The processor, communication interface, and memory can communicate with each other via the bus. The communication interface can be used for information transmission. The processor can call logical instructions in the memory to execute the online course (MOOC) dropout prediction method based on heterogeneous feature fusion described in the above embodiments.

[0060] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium.

[0061] Memory, as a computer-readable storage medium, can be used to store software programs and computer-executable programs, such as the program instructions / modules corresponding to the methods in the embodiments of this disclosure. The processor executes the program instructions / modules stored in the memory to perform functional applications and data processing, thereby realizing the MOOC dropout prediction method based on heterogeneous feature fusion in the above embodiments.

[0062] The memory may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the terminal device. Furthermore, the memory may include high-speed random access memory and may also include non-volatile memory.

[0063] This disclosure provides a computer-readable storage medium storing computer-executable instructions configured to execute the above-described MOOC dropout prediction method based on heterogeneous feature fusion.

[0064] The aforementioned computer-readable storage medium may be a transient computer-readable storage medium or a non-transitory computer-readable storage medium.

[0065] The technical solutions of this disclosure can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes one or more instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in this disclosure. The aforementioned storage medium can be a non-transitory storage medium, including: a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, and other media capable of storing program code. It can also be a transient storage medium.

[0066] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting dropout rates in online courses (MOOCs) based on heterogeneous feature fusion, characterized in that, Includes the following steps: Obtain the MOOC dataset and preprocess the dataset to obtain the test set, including data cleaning and binarization of the learner's final learning results. The test set includes static background feature vectors and behavioral time series sequences. An IHFNet network comprising a multi-agent adaptive static feature selector module, a hierarchical temporal feature extractor module, and a heterogeneous feature fusion module is constructed and trained; the multi-agent adaptive static feature selector module selects key static features from static background features; The hierarchical temporal feature extractor module extracts high-discriminative behavioral temporal features from the behavioral temporal sequence; the heterogeneous feature fusion module adaptively fuses the key static features with the behavioral temporal features, and performs classification prediction based on the fused features; Train the IHFNet network using the training set; MOOC data of learners to be predicted are collected and input into the trained IHFNet network to predict dropout risk.

2. The MOOC dropout prediction method based on heterogeneous feature fusion according to claim 1, characterized in that, The multi-agent adaptive static feature selector module consists of a multi-agent decision layer, a state and reward function layer, a policy optimization layer, and a feature output layer. An agent is configured for each static background feature vector, and each agent performs iterative decision optimization based on the ε-greedy policy and Q-value update mechanism to select a subset of key static features.

3. The MOOC dropout prediction method based on heterogeneous feature fusion according to claim 2, characterized in that, The multi-agent adaptive static feature selector module employs a causal reward mechanism based on counterfactual reasoning, where the causal reward for a single agent... for: ， ， in, For overall performance, For counterfactual performance, For intelligent agents exist Momentary actions This is a counterfactual action vector.

4. The MOOC dropout prediction method based on heterogeneous feature fusion according to claim 1, characterized in that, The hierarchical temporal feature extractor module contains multiple sets of parallel one-dimensional convolutional layers, each using a convolutional kernel of a different size; after each set of convolutional branches, a batch normalization layer, a nonlinear activation layer, and a pooling layer are connected in sequence to obtain multi-scale features; After being concatenated, the multi-scale features are sequentially input into a fully connected layer and two Transformer encoder layers for processing. The features obtained from the Transformer encoder layer are then connected with the concatenated multi-scale features through residual connections, and finally, after layer normalization, temporal behavior features are output.

5. The MOOC dropout prediction method based on heterogeneous feature fusion according to claim 4, characterized in that, The kernel sizes of the multiple parallel one-dimensional convolutional layers are set to 3×1, 5×1 and 7×1, respectively.

6. The MOOC dropout prediction method based on heterogeneous feature fusion according to claim 5, characterized in that, The heterogeneous feature fusion module includes: The feature projection layer includes two parallel multilayer perceptrons, which are used to map the key static features and the behavioral temporal features to the same feature space, respectively. The attention interaction fusion layer is used to weightedly fuse the two types of features after mapping through an attention mechanism; The residual normalization enhancement layer is used to combine residual connections with normalization operations to stabilize the training process. A classifier is used to output dropout risk prediction results based on the fused features.

7. The MOOC dropout prediction method based on heterogeneous feature fusion according to claim 6, characterized in that, The IHFNet network uses a combination of weighted cross-entropy loss and feature fusion constraint term as the loss function; the Adam optimization algorithm is used to update the model parameters, with the initial learning rate set to 0.001; and the Dropout regularization mechanism is introduced to suppress overfitting. The dataset was trained and tested using a 10-fold cross-validation method. The dataset was randomly divided into 10 non-overlapping subsets, and training and testing were performed sequentially. The average of the experimental results of each fold was taken as the final prediction performance.

8. The MOOC dropout prediction method based on heterogeneous feature fusion according to claim 7, characterized in that, The loss function as follows: ， ， ， in, For weighted cross-entropy loss, For feature fusion constraint terms, To balance the hyperparameters of the main loss and constraint terms; The total number of samples in the training dataset. For the first The true label of each sample The first output of the model The dropout probability of a sample; The weights of the positive dropout samples. Weights for non-dropout negative samples; These are the key static feature vectors after processing by the projection layer. This is the behavioral temporal feature vector after processing by the projection layer. This is the cosine similarity function.

9. The MOOC dropout prediction method based on heterogeneous feature fusion according to any one of claims 1-8, characterized in that, The IHFNet network also includes a SHAP interpretability analysis module, which is used to perform contribution quantification analysis on the key static features selected by the multi-agent adaptive static feature selector module. The IHFNet network also includes a behavior trend analysis module, which performs learning trend analysis based on the behavior time-series features extracted by the hierarchical time-series feature extractor module, revealing the characteristics of learner activities under different learning trends.

10. A MOOC dropout prediction device based on heterogeneous feature fusion, comprising a processor and a memory storing program instructions, characterized in that, The processor is configured to execute, when running the program instructions, the MOOC dropout prediction method based on heterogeneous feature fusion as described in any one of claims 1-9.

Images