An online course dropout prediction method based on context-aware timing deep network

Abstract

The application provides an online course dropout prediction method based on context-aware timing deep network, and belongs to the technical field of computer-aided intelligent education. A prediction data set containing static features and dynamic behavior features is collected; global statistical features are extracted from the dynamic behavior features, and a dynamic behavior matrix is constructed according to the dynamic behavior features; an online course dropout prediction network is constructed; the online course dropout prediction network is trained by using a training set, a binary cross-entropy loss is used as a loss function, and an L2 regularization term is introduced; data of a learner to be predicted is collected and input into the network for dropout risk prediction. The application realizes learner-course context and timing behavior feature fusion based on early data to improve the dropout risk prediction effect.

Description

Technical Field

[0001] This invention relates to an online course dropout prediction method based on context-aware temporal deep networks, belonging to the field of computer-aided smart education technology. Background Technology

[0002] With the development of information technology, online education has become a significant trend. MOOCs (Massive Open Online Courses) attract a large number of learners with their high-quality resources and flexible models, forming a vast learning ecosystem. However, despite the significant educational value and development potential of online courses, the high dropout rate exposed during their rapid development has become a key bottleneck restricting their quality improvement and sustainable development. This not only results in a serious waste of high-quality educational resources but also weakens their educational effectiveness and social credibility, posing a severe challenge to the sustainable development of online education. Therefore, solving this problem is urgent, and conducting research on online course dropout prediction is not only a key path to addressing the industry's core pain points but also possesses important theoretical and practical value.

[0003] Existing research often uses methods such as learning behavior analysis to alleviate the high dropout rate of online courses, but these methods often rely on behavioral characteristics in the middle and later stages of learning or use only a single characteristic, which has obvious lag and cannot fully explore information, making it difficult to achieve timely and accurate identification of dropout risks.

[0004] Therefore, it is necessary to design a dropout prediction model based on learners' prior learning behavior data and constructed using multi-source features, so as to achieve early identification of high-risk learners while ensuring prediction accuracy. Summary of the Invention

[0005] The purpose of this invention is to provide an online course dropout prediction method based on context-aware temporal deep networks, which integrates learner-course context and temporal behavioral features based on early data to improve the dropout risk prediction effect.

[0006] To achieve the above objectives, the present invention employs the following technical solution:

[0007] A method for predicting online course dropout based on context-aware temporal deep networks, characterized by the following steps:

[0008] A prediction dataset containing static features and dynamic behavioral features is collected, and the prediction dataset is preprocessed and divided into a training set and a test set; the static features include learner information and course information, and the dynamic behavioral features include activity logs.

[0009] Global statistical features are extracted from dynamic behavioral features, and a dynamic behavior matrix is ​​constructed based on the dynamic behavioral features. The dynamic behavior matrix aggregates the frequency of each student's various behaviors in each specific time step.

[0010] An online course dropout prediction network is constructed, comprising an encoder module, a contextual feature interaction module, a global statistical feature extraction module, a temporal modeling module, and a multi-feature fusion module. The encoder encodes static features and global statistical features using a weighted embedding method to obtain learner information codes, course information codes, and global statistical codes. The contextual feature interaction module fuses learner feature codes and course feature codes using a deep cross-network to obtain contextual interaction features. The global statistical feature extraction module processes the global statistical codes using one-dimensional convolution to obtain behavioral collaboration features. The temporal modeling module sequentially feeds the dynamic behavior matrix into a feature selector and a daily pattern selector to obtain filtered daily activity pattern features and attention-weighted features. The multi-feature fusion module concatenates and fuses the contextual interaction features, behavioral collaboration features, daily activity pattern features, and attention-weighted features based on an attention mechanism.

[0011] An online course dropout prediction network was trained using the training set. The loss function adopted was binary cross-entropy loss, and an L2 regularization term was introduced.

[0012] Data on learners to be predicted is collected and input into the network for dropout risk prediction.

[0013] Preferably, the specific processing procedure of the encoder module is as follows:

[0014] A weighted embedding method is used to map static features and global statistical features to a low-dimensional space, transforming them into feature indexes and feature values; the static features include learner features and course information features.

[0015] The feature indexes are converted into vector representations using the embedding matrix;

[0016] By weighting the eigenvalues ​​and vector representations element by element, we obtain learner information encoding, course information encoding, and global statistical encoding.

[0017] Preferably, the context feature interaction module includes cross-network branches, deep network branches, and a combined output layer;

[0018] The cross-network branch learns the combination relationship between features by performing a layer-by-layer cross operation on the concatenated learner information encoding and course information encoding, thus obtaining an explicit feature interaction representation.

[0019] The deep network branch consists of multiple fully connected layers. It captures the complex feature combination patterns between the spliced ​​learner information encoding and course information encoding through nonlinear transformation to obtain implicit high-order representation.

[0020] The combined output layer concatenates explicit feature interaction representations and implicit higher-order feature representations and feeds them into a fully connected layer to obtain context interaction features.

[0021] Preferably, the specific processing procedure of the global statistical feature extraction module is as follows:

[0022] Rearrange the global statistical codes along the dimensions;

[0023] The features after dimensionality rearrangement are input into a one-dimensional convolutional network. A sliding window mechanism is used to scan the global statistical encoding along the feature dimension. The combination relationship between behavioral features is modeled through learnable convolutional kernels.

[0024] After the convolution operation, a nonlinear activation function is introduced to enhance the feature representation capability, resulting in a collaborative feature representation that integrates statistical information from multiple behaviors.

[0025] Preferably, the time-series modeling module includes a feature selector module and a day pattern selector module:

[0026] The feature selector processing procedure is as follows:

[0027] The dynamic behavior matrix is ​​flattened and input into a linear layer containing an activation function to generate a feature mask;

[0028] The feature mask is then multiplied element-wise with the dynamic behavior matrix to obtain a weighted and filtered feature representation.

[0029] Preferably, the processing procedure of the day mode selector module is as follows:

[0030] The filtered feature representations are input into a long short-term memory network to obtain a higher-order temporal representation.

[0031] Learnable position coding is introduced to inject position information into each time step of the high-order temporal representation;

[0032] The high-order temporal representation of the fused location information is input into the attention mechanism to calculate the importance weight of each time step;

[0033] Based on importance weights, an attention-based dual-branch aggregation strategy is adopted to obtain attention-weighted features and daily activity pattern features. The global branch obtains attention-weighted features by weighted summation, while the local branch extracts the most representative key time steps as daily activity pattern features through a Top-K selection mechanism.

[0034] Preferably, the multi-feature fusion module adopts a multi-view gating fusion strategy, and the specific process is as follows:

[0035] The contextual interaction features, behavioral collaboration features, daily activity pattern features, and attention-weighted features are mapped and aligned to a unified hidden layer feature space.

[0036] Extracting micro-dynamic representations of daily activity patterns under global guidance through a cross-attention mechanism;

[0037] A joint feature is constructed by splicing contextual interaction features, behavioral collaboration features, and daily activity pattern features, and then adaptively filtered and denoised using a gating network;

[0038] The filtered and denoised joint features are nonlinearly cross-mapped using a multilayer perceptron to output the dropout risk probability value.

[0039] Preferably, the loss function for training the online course dropout prediction network is:

[0040] ,

[0041] ,

[0042] in, For binary cross-entropy loss, The L2 regularization coefficient is... Represents the space of all learnable parameters of the model. The square of the L2 norm, This represents the total number of samples in the current training batch. For the first The true label of each sample The first output of the model The predicted probability of early school dropout in a sample.

[0043] The advantages of this invention are as follows: It achieves full feature coverage by collaboratively mining multidimensional associations and statistical patterns through multiple modules; supplemented by a Top-k screening mechanism to eliminate redundancy, it balances representational completeness and model operating efficiency. This invention can accurately identify learners at risk of dropping out of school in advance, providing educators with reliable decision-making basis, facilitating timely personalized educational interventions, optimizing learning support strategies, and reducing learner dropout rates. Attached Figure Description

[0044] 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.

[0045] Figure 1 This is a flowchart illustrating the implementation of the present invention.

[0046] Figure 2 This is a structural diagram of the online course dropout prediction network of the context-aware temporal deep network proposed in this invention.

[0047] Figure 3 This is a schematic diagram of the feature selector submodule in the network structure temporal modeling module proposed in this invention.

[0048] Figure 4This is a schematic diagram of the Japanese mode selector submodule in the network structure temporal modeling module proposed in this invention.

[0049] Figure 5 A visualization analysis of the differences in video stopping behavior between dropouts and non-dropouts in the XuetangX dataset.

[0050] Figure 6 A visualization analysis of the differences in video pause behavior between dropouts and non-dropouts in the XuetangX dataset.

[0051] Figure 7 A visualization analysis of the differences in click behavior of courseware between dropouts and non-dropouts in the XuetangX dataset.

[0052] Figure 8 This is a mesh graph showing the learning path transitions of non-dropouts in the XuetangX dataset.

[0053] Figure 9 A mesh graph showing the learning path transitions of dropouts in the XuetangX dataset. Detailed Implementation

[0054] 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.

[0055] like Figure 1 As shown, an online course dropout prediction method based on context-aware temporal deep networks is characterized by the following steps:

[0056] S1: Collect a prediction dataset containing static features and dynamic behavioral features, and preprocess the prediction dataset to divide it into a training set and a test set; the static features include learner information and course information, and the dynamic behavioral features include activity logs.

[0057] S2: Extract global statistical features from dynamic behavioral features, and construct a dynamic behavioral matrix based on the dynamic behavioral features. The dynamic behavioral matrix aggregates the frequency of each student's various behaviors in each specific time step.

[0058] S3: Construct an online course dropout prediction network comprising an encoder module, a contextual feature interaction module, a global statistical feature extraction module, a temporal modeling module, and a multi-feature fusion module. The encoder encodes static features and global statistical features using a weighted embedding method to obtain learner information encoding, course information encoding, and global statistical encoding. The contextual feature interaction module fuses learner feature encoding and course feature encoding using a deep cross-network to obtain contextual interaction features. The global statistical feature extraction module processes the global statistical encoding using one-dimensional convolution to obtain behavioral collaboration features. The temporal modeling module sequentially feeds the dynamic behavior matrix into a feature selector and a daily pattern selector to obtain filtered daily activity pattern features and attention-weighted features. The multi-feature fusion module concatenates and fuses contextual interaction features, behavioral collaboration features, daily activity pattern features, and attention-weighted features based on an attention mechanism.

[0059] S4: Train the online course dropout prediction network using the training set. The loss function adopts the binary cross-entropy loss and introduces the L2 regularization term.

[0060] S5: Collect data on learners to be predicted and input it into the network to predict dropout risk.

[0061] As a refinement of the above embodiment, the XuetangX dataset can be used in step S1.

[0062] The preprocessing includes:

[0063] (1) Remove missing values ​​and correct outliers;

[0064] (2) Based on the data characteristics, the features are divided into two categories: categorical and numerical. The numerical features are standardized to conform to a standard normal distribution with a mean of 0 and a standard deviation of 1.

[0065] (3) A unified key-value pair encoding scheme is used to process static features: For categorical features in static features, each categorical value is mapped to a continuous integer index and assigned a fixed value of 1.0. For numerical features in static features, the same index encoding method is retained, and the standardized numerical value is used as the value.

[0066] This method establishes a standardized key-value pair representation for all static features.

[0067] Learner information (such as age, gender, and education level) and course information (such as subject category and difficulty level) that remain unchanged throughout the course learning cycle are collectively referred to as static information. On the other hand, activity logs (such as video viewing records, exercise answering behavior, and forum interactions) are constantly updated and changed as the learning process progresses, and are therefore referred to as dynamic behavioral characteristics.

[0068] As a refinement of the above embodiments, the global statistical features in step S2 include the total number of all learner behaviors, the individual number of each type of specific behavior, and the maximum and average number of each behavior.

[0069] Construct a dynamic behavior matrix based on dynamic behavior characteristics: aggregate the frequency of each student's various behaviors in each specific time step, and map the original discrete logs into a two-dimensional dynamic behavior matrix X.

[0070] Here, a day is used as a fixed time step. Within the time window, for a specific day... Each course selection record (learner-course pair) has a dynamic behavior matrix that can be represented as follows: :

[0071] ,

[0072] in, Indicates the first A dynamic behavior matrix of course selection records. The sequence length representing the time step. This represents the feature dimension (the number of different interaction behaviors). Representing the The course selection record is in the first The first time step occurs Frequency of this type of behavior.

[0073] As a refinement of the above embodiments, the network structure in step S3 is as follows: Figure 2 As shown:

[0074] (1) The specific processing procedure of the encoder module is as follows:

[0075] Learner features, course features, and global statistical features are represented as corresponding feature indices and feature value formats, respectively.

[0076] Discrete feature indices are mapped to low-dimensional dense vector representations using their respective embedding matrices.

[0077] Feature values ​​are introduced to weight the embedding vector element by element to characterize the differences in importance of different features;

[0078] Learner codes, course codes, and global activity statistics codes are obtained respectively, thus forming a unified encoder feature representation, which provides structured input for subsequent models.

[0079] Specifically, it is expressed as follows:

[0080] ,

[0081] in, Represents learner feature index, Embedding matrices for learners, This indicates element-wise multiplication. Represents learner feature values. For transpose operation, Encode representations for learners, Indicates the course feature index. Embedding the course into a matrix, Represents the course feature value. This is represented by course codes. Represents the global statistical feature index. For global statistical embedding matrix, Represents global statistical characteristic values. This is a global statistical encoding representation.

[0082] (2) The context feature interaction module includes cross-network branches, deep network branches, and combined output layers; the processing method is as follows:

[0083] The learner encoding representation output by the encoder is concatenated with the course encoding representation along the feature dimension to form a unified context input representation;

[0084] A deep cross-network is introduced to perform explicit high-order feature interaction modeling of spliced ​​features, and the combination relationship between user attributes and course information is captured through layer-by-layer cross-operations;

[0085] A deep branch consisting of a multi-layer fully connected network is constructed to perform nonlinear transformation on the spliced ​​features in order to learn implicit high-order semantic representations;

[0086] The outputs of the cross-network and the deep network are concatenated and fused, and the final context interaction feature representation is obtained through linear mapping, thereby realizing the collaborative modeling of explicit and implicit feature interactions.

[0087] The specific processing procedure is as follows:

[0088] ,

[0089] ,

[0090] ,

[0091] ,

[0092] ,

[0093] ,

[0094] in, For context input representation, After being flattened , For the first in the cross network The feature output of the layer, This is the output of the cross-network branches, i.e., an explicit feature interaction representation. This is the output of the deep network branch, i.e., the implicit higher-order representation. For contextual interaction features, This indicates a feature concatenation operation. This represents the flattening operation, which maps three-dimensional features to two-dimensional vectors. Indicates the first The output of the layer cross network, For learnable weights, For bias terms, This represents a cross-network module composed of the aforementioned stacked cross layers. This represents a feedforward neural network composed of multiple linear transformations and nonlinear activation functions. Used to map the concatenated features to the target dimension.

[0095] (3) The global statistical feature extraction module is used to model the collaborative relationship between various behavioral statistical features; the processing method is as follows:

[0096] The global statistical encoding representation is rearranged in dimension so that the embedding dimension is used as the channel dimension to input a one-dimensional convolutional network.

[0097] A sliding window mechanism is used to scan the behavioral statistical sequence in terms of feature dimension, and the combination relationship between behavioral features is modeled in a local range through learnable convolutional kernels;

[0098] After the convolution operation, a non-linear activation function is introduced to enhance the feature representation capability; a collaborative feature representation that integrates statistical information from multiple behaviors is obtained, thereby improving the model's ability to characterize global behavioral patterns.

[0099] The specific processing procedure is as follows:

[0100] ,

[0101] ,

[0102] in, The features are after dimensional rearrangement. This represents a dimension reordering operation, transforming the input features into a shape that satisfies the input format of a one-dimensional convolution. To integrate behavioral collaboration features from multiple behavioral statistical information, Represents a non-linear activation function. This represents a one-dimensional convolution operation.

[0103] (4) The time series modeling module sends the dynamic behavior matrix into the feature selector module and the daily pattern selector module to obtain the daily activity pattern features and attention-weighted features; Feature selector: adaptively weights the original input features; Daily pattern selector: models the time series dependency relationship through LSTM and selects the most representative K time steps by combining position encoding and attention mechanism.

[0104] like Figure 3 As shown, the feature selector assigns learnable weights to each input feature dimension and constrains them to the [0,1] interval through a nonlinear mapping, thereby forming a soft mask to measure the importance of each feature; the processing procedure is as follows:

[0105] The input feature tensor is flattened in the time dimension, and then the nonlinear relationship between features is extracted through a two-layer fully connected network. The weight mask is generated using the Sigmoid function.

[0106] The feature mask is multiplied element-wise with the dynamic behavior matrix to enhance key features and suppress redundant information, resulting in the filtered feature representation.

[0107] Specifically, it is expressed as follows:

[0108] ,

[0109] ,

[0110] ,

[0111] in, , For a learnable parameter matrix, Represents a feature mask. This is a one-dimensional feature vector obtained after flattening it along the time dimension. For dynamic behavior matrix, The filtered feature representation, This is a shape restoration operation.

[0112] like Figure 4 As shown, the goal of the day pattern selector is to capture students' learning rhythm over consecutive time periods and identify the K most representative key dates (time steps) for subsequent dropout prediction tasks. This module models the learning rhythm through multi-stage processing, consisting of an LSTM time series construction layer, positional encoding injection, attention mechanism, and time step selection.

[0113] The processing procedure of the day mode selector module is as follows:

[0114] The filtered feature representations are input into a long short-term memory network to capture the temporal dependencies in the behavioral sequence and obtain a high-order temporal representation.

[0115] Learnable positional encoding is introduced to inject positional information into each time step, thereby enhancing the model's ability to perceive temporal order and periodic patterns.

[0116] The filtered feature representation after fusing location information is input into the attention mechanism to calculate the importance weights at each time step;

[0117] An attention-based dual-branch aggregation strategy is adopted, in which the global branch obtains the overall behavioral representation through weighted summation, and the local branch extracts the most representative key time steps through the Top-K selection mechanism, thereby achieving joint modeling of continuous and intermittent learning modes.

[0118] Specifically, it is expressed as follows:

[0119] ,

[0120] ,

[0121] ,

[0122] ,

[0123] ,

[0124] in, This represents the learnable positional encoding matrix. This is the feature representation after fusing location information. For attention score, This represents attention-weighted features. Indicates the characteristics of daily activity patterns. For higher-order timing representation, For index-based feature extraction operations, Based on attention scores Select the one with the largest value. Operations on the position index corresponding to each score. The number of critical time steps to be retained. For vectors The Attention score corresponding to each time step For higher-order timing representation The The feature vector corresponding to each time step.

[0125] (5) The multi-feature fusion module is used to perform unified modeling of multi-source features and generate the final prediction result. This module achieves deep integration of complex features through asymmetric interaction and adaptive noise reduction. It consists of several parts: heterogeneous feature alignment layer, global guidance local interaction layer, feature-level gating network and nonlinear prediction layer.

[0126] The behavioral collaboration features, contextual interaction features, attention-weighted features, and daily activity pattern features are aligned to a unified feature space through linear mapping.

[0127] A cross-attention mechanism is introduced, using attention-weighted features as query vectors to perform attention calculation and weighted aggregation on daily activity pattern features, thereby extracting micro-dynamic representations guided by the global state.

[0128] The four types of features are concatenated along the feature dimension to form a joint representation, which is then input into a feature-level gating network for adaptive feature selection and noise reduction.

[0129] The selected high-order fusion features are input into a multilayer perceptron for nonlinear cross-mapping, and finally the predicted probability of early school dropout is output through the Sigmoid function.

[0130] The specific processing procedure is as follows:

[0131] ,

[0132] ,

[0133] ,

[0134] ,

[0135] ,

[0136] in, This represents the cross-attention score between the global and local levels. Represents the context features after mapping alignment. This represents the collaborative features of behavior after mapping alignment. This represents the attention-weighted features after mapping alignment. This represents the microscopic dynamic representation of the fused global information after mapping alignment. For the combined features after splicing, This is the higher-order fusion representation after being filtered by a gating network. This represents the Sigmoid activation function. To predict the probability of eventually dropping out of school, The global state features after mapping and alignment. For learnable query projection matrix, The number of critical time steps to be retained. For the mapped and aligned daily activity pattern features, The learnable bond projection matrix. Scaling factor Weight vector The first in One element, The first characteristic of daily activity patterns a time step vector For the learnable weight matrix of the gated network, This refers to the learnable bias term for the gated network.

[0137] As a refinement of the above embodiments, step S4 specifically includes:

[0138] (1) Loss Function and Optimization Strategy. During the model training phase, the binary cross-entropy loss function (BCELoss) is used as the main objective function to measure the difference between the model's predicted dropout probability and the true label. To effectively prevent overfitting during training, an L2 regularization term is introduced into the total loss to constrain the complexity of the model parameters. In terms of parameter optimization, the model calculates the gradient of the objective function with respect to each parameter through the backpropagation algorithm and updates the parameters based on the gradient using the Adam optimization algorithm to ensure the efficiency and stability of the training process. In addition, a dynamic learning rate adjustment and early stopping mechanism are introduced during the training process: after each training round, the model calculates the AUC index on the validation set and dynamically decays the learning rate accordingly; when the AUC of the validation set does not improve within several consecutive rounds and triggers the early stopping condition, the model will automatically terminate training and obtain the final model.

[0139] The loss function is:

[0140] ,

[0141] ,

[0142] in, The binary cross-entropy loss is used to quantify the classification error between the predicted distribution and the true dropout label. This is the L2 regularization coefficient (weight decay parameter), used to control the penalty force of the regularization term on the main loss. Represents the space of all learnable parameters of the model. The squared L2 norm is used to limit excessively large parameter weights, thereby improving the model's generalization ability. This represents the total number of samples in the current training batch. For the first The true label of each sample The first output of the model The predicted probability of early school dropout in a sample.

[0143] (2) Evaluation Metrics. To comprehensively evaluate the model's performance in the learner dropout prediction task, and considering the significant class imbalance in the dataset used in this study, traditional accuracy often becomes biased in such scenarios and fails to fully reflect the model's true performance. Drawing on the experience of the KDDCup2015 competition, which pointed out that the area under the receiver operating characteristic (ROC) curve (AUC) still provides reliable discrimination ability under imbalanced data conditions, it is widely considered one of the most effective evaluation metrics. Based on this consideration, we selected AUC and F1-score as the main evaluation metrics in the experiment to more comprehensively measure the model's classification performance.

[0144] Specifically, the ROC curve is plotted with the True Positive Rate (TPR) on the vertical axis and the False Positive Rate (FPR) on the horizontal axis, and is used to describe the model's classification ability at different thresholds. Its calculation method is as follows.

[0145] ,

[0146] ,

[0147] in, , , , These represent true positives, true negatives, false positives, and false negatives, respectively. The value of AUC is defined as the area under the ROC curve:

[0148] ,

[0149] The value of AUC ranges from [0,1][0,1][0,1], and the closer it is to 1, the stronger the model's ability to distinguish between positive and negative samples.

[0150] On the other hand, the F1-score, as a harmonic average of precision and recall, can balance the importance of both in cases of class imbalance. Its calculation formula is:

[0151] ,

[0152] ,

[0153] ,

[0154] In F1-score, Precision reflects how many samples predicted as positive by the model are actually positive, while Recall measures how many samples that are actually positive were correctly identified by the model. Compared to using Precision or Recall alone, F1-score provides a trade-off and is more suitable for evaluating the model's performance on imbalanced data.

[0155] This invention proposes an online course dropout prediction method based on context-aware temporal deep networks. This method can simultaneously model learners' static features and behavioral temporal features, effectively detecting learners' dropout tendencies early in the course, helping online course platforms and teachers to adjust teaching methods and intervene in a timely manner. The method achieved high prediction performance on the XuetangX dataset test set, as shown in Table 1, with an AUC of 0.8775 and an F1 score of 0.9087, effectively identifying learners at high dropout risk. Comparative experimental results show that the method has a significant advantage in prediction performance, as detailed in Table 1. Compared with previous dropout prediction methods, this invention adopts a multi-dimensional joint perspective in the feature extraction stage, with a superior feature fusion architecture and temporal modeling design, enabling it to more accurately fit the evolutionary trajectory of learners in real online education scenarios.

[0156] Table 1 Comparative Experiment

[0157]

[0158] To further explore the intrinsic link between learning behavior and dropout outcomes, and to enhance the interpretability of model predictions, this invention conducts further statistical and comparative analysis of learners' behavioral characteristics. By comparing the distribution differences in learning activity and core learning behaviors between dropouts and course completers, the behavioral patterns of different groups are intuitively revealed, providing data support for understanding the factors influencing dropout.

[0159] like Figures 5-7 As shown, this invention selects three core learning activities with the highest click-through rates and the most representative characteristics on the platform as the analysis objects. The number of visits by the two groups to each type of activity is statistically analyzed, and a distribution comparison chart is drawn. Through the visual analysis of the differences in behavioral characteristic distribution, the degree of influence of different learning behaviors on dropout outcomes can be intuitively reflected, thus providing a richer behavioral explanation for dropout analysis.

[0160] In the three core learning behaviors of stopping videos, pausing videos, and clicking on courseware, non-dropouts showed a significantly wider distribution of clicks and a higher overall level of engagement, while dropouts' engagement was highly concentrated in the low-activity range. Non-dropouts demonstrated more complete learning conversations and proactive self-regulation strategies through frequent stopping and pausing, and their higher frequency of courseware clicks reflected a stronger willingness to explore content. Dropouts, on the other hand, exhibited insufficient engagement, a lack of proactive intervention, and in-depth exploration across all behaviors. Overall, non-dropouts were more planned, proactive, and exploratory in their learning process, while dropouts showed randomness, passivity, and insufficient engagement. These differences not only provide key behavioral characteristics for identifying early dropout risks but also offer empirical evidence for designing teaching interventions and personalized learning support strategies.

[0161] like Figures 8-9As shown, behavioral data encompassing multiple activity categories was hierarchically categorized, and a heatmap of behavioral transition matrices for dropouts and non-dropouts was created based on high-level behavioral categories. The results show a high degree of similarity in the transition matrices, indicating that both dropouts and non-dropouts exhibit similar behavioral trajectories, with behaviors largely concentrated in activities such as "Answer Checking," "Content Creation," "Interactive Entry," "Learning Process Control," and "Problem Operations." This suggests that regardless of the final learning outcome, learners' primary interaction patterns on the course platform tend to be consistent, and the platform's key functions are widely used across different groups. However, non-dropouts are more active in paths such as Content Creation, Interactive Entry, Learning Process Control, and Progress Controls, especially after completing Problem Operations, often transitioning to progress monitoring and process adjustment. This pattern indicates that non-dropout learners not only focus on mastering knowledge points but also tend to maintain their learning pace through self-checking and adjusting their study plans, demonstrating strong systematic and planned characteristics. In contrast, dropout learners show a higher proportion of activity related to content deletion, course resources, and some learning control transfers, exhibiting behavioral characteristics of dependence on resource-based operations and localized steps, lacking coherence and in-depth development. This suggests that their learning paths often remain at a low level, failing to effectively translate into control and advancement of the overall learning process. In summary, the behavioral patterns of non-dropout learners are more constructive and self-regulating, enabling them to optimize the learning process through progress monitoring and learning control; while dropout learners exhibit fragmentation and passivity, struggling to form stable learning paths. This difference not only reveals the inherent behavioral patterns of the two types of learners but also provides a theoretical basis for explaining differences in learning persistence and designing targeted intervention strategies.

[0162] This disclosure also provides an online course dropout prediction device based on context-aware temporal deep networks, 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 invoke logical instructions in the memory to execute the online course dropout prediction method based on context-aware temporal deep networks described above.

[0163] 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.

[0164] 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 online course dropout prediction method based on context-aware temporal deep networks in the above embodiments.

[0165] 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.

[0166] This disclosure provides a computer-readable storage medium storing computer-executable instructions configured to execute the above-described online course dropout prediction method based on context-aware temporal deep networks.

[0167] The aforementioned computer-readable storage medium may be a transient computer-readable storage medium or a non-transitory computer-readable storage medium.

[0168] 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.

[0169] 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 online course dropout based on context-aware temporal deep networks, characterized in that, Includes the following steps: A prediction dataset containing static features and dynamic behavioral features is collected, and the prediction dataset is preprocessed and divided into a training set and a test set. The static features include learner information and course information, and the dynamic behavioral features include activity logs, which include video viewing records, exercise answering behavior, and forum interactions. Global statistical features are extracted from dynamic behavioral features, and a dynamic behavior matrix is ​​constructed based on the dynamic behavioral features. The dynamic behavior matrix aggregates the frequency of each student's various behaviors in each specific time step. An online course dropout prediction network is constructed, comprising an encoder module, a contextual feature interaction module, a global statistical feature extraction module, a temporal modeling module, and a multi-feature fusion module. The encoder encodes static features and global statistical features using weighted embedding to obtain learner information codes, course information codes, and global statistical codes. The contextual feature interaction module fuses learner feature codes and course feature codes using a deep cross-network to obtain contextual interaction features. The global statistical feature extraction module processes the global statistical codes using one-dimensional convolution to obtain behavioral collaboration features. The temporal modeling module sequentially feeds the dynamic behavior matrix into a feature selector and a daily pattern selector to obtain filtered daily activity pattern features and attention-weighted features. The multi-feature fusion module combines contextual interaction features, behavioral collaboration features, daily activity pattern features, and attention-weighted features based on an attention mechanism. An online course dropout prediction network was trained using the training set. The loss function adopted was binary cross-entropy loss, and an L2 regularization term was introduced. Collect data on learners to be predicted and input it into the network to predict dropout risk; The time-series modeling module includes a feature selector module and a day pattern selector module: The feature selector processing procedure is as follows: The dynamic behavior matrix is ​​flattened and input into a linear layer containing an activation function to generate a feature mask; The feature mask is multiplied element-wise with the dynamic behavior matrix to obtain a weighted and filtered feature representation. The processing procedure of the day mode selector module is as follows: The filtered feature representations are input into a long short-term memory network to obtain a higher-order temporal representation. Learnable position coding is introduced to inject position information into each time step of the high-order temporal representation; The high-order temporal representation of the fused location information is input into the attention mechanism to calculate the importance weight of each time step; Based on importance weights, an attention-based dual-branch aggregation strategy is adopted to obtain attention-weighted features and daily activity pattern features. The global branch obtains attention-weighted features by weighted summation, while the local branch extracts the most representative key time steps as daily activity pattern features through a Top-K selection mechanism.

2. The online course dropout prediction method based on context-aware temporal deep networks according to claim 1, characterized in that, The specific processing procedure of the encoder module is as follows: A weighted embedding method is used to map static features and global statistical features to a low-dimensional space, transforming them into feature indexes and feature values; the static features include learner features and course information features. The feature indexes are converted into vector representations using the embedding matrix; By weighting the eigenvalues ​​and vector representations element by element, we obtain learner information encoding, course information encoding, and global statistical encoding.

3. The online course dropout prediction method based on context-aware temporal deep networks according to claim 2, characterized in that, The context feature interaction module includes cross-network branches, deep network branches, and combined output layers; The cross-network branch learns the combination relationship between features by performing a layer-by-layer cross operation on the concatenated learner information encoding and course information encoding, thus obtaining an explicit feature interaction representation. The deep network branch consists of multiple fully connected layers. It captures the complex feature combination patterns between the spliced ​​learner information encoding and course information encoding through nonlinear transformation to obtain implicit high-order representation. The combined output layer concatenates explicit feature interaction representations and implicit higher-order feature representations and feeds them into a fully connected layer to obtain context interaction features.

4. The online course dropout prediction method based on context-aware temporal deep networks according to claim 3, characterized in that, The specific processing procedure of the global statistical feature extraction module is as follows: Rearrange the global statistical codes along the dimensions; The features after dimensionality rearrangement are input into a one-dimensional convolutional network. A sliding window mechanism is used to scan the global statistical encoding along the feature dimension. The combination relationship between behavioral features is modeled through learnable convolutional kernels. After the convolution operation, a nonlinear activation function is introduced to enhance the feature representation capability, resulting in a collaborative feature representation that integrates statistical information from multiple behaviors.

5. The online course dropout prediction method based on context-aware temporal deep networks according to claim 1, characterized in that, The multi-feature fusion module adopts a multi-view gating fusion strategy, and the specific process is as follows: The contextual interaction features, behavioral collaboration features, daily activity pattern features, and attention-weighted features are mapped and aligned to a unified hidden layer feature space. Extracting micro-dynamic representations of daily activity patterns under global guidance through a cross-attention mechanism; A joint feature is constructed by splicing contextual interaction features, behavioral collaboration features, and daily activity pattern features, and then adaptively filtered and denoised using a gating network; The filtered and denoised joint features are nonlinearly cross-mapped using a multilayer perceptron to output the dropout risk probability value.

6. The online course dropout prediction method based on context-aware temporal deep networks according to any one of claims 1-5, characterized in that, The loss function for training the online course dropout prediction network is: ， ， in, For binary cross-entropy loss, The L2 regularization coefficient is... Represents the space of all learnable parameters of the model. The square of the L2 norm, This represents the total number of samples in the current training batch. For the first The true label of each sample The first output of the model The predicted probability of early school dropout in a sample.

7. An online course dropout prediction device based on context-aware temporal deep networks, comprising a processor and a memory storing program instructions, characterized in that, The processor is configured to execute, when running the program instructions, the online course dropout prediction method based on a context-aware temporal deep network as described in any one of claims 1-6.

8. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the online course dropout prediction method based on a context-aware temporal deep network as described in any one of claims 1-6.

Images