A Time-Sequence Awareness-Based Dynamic Early Warning Method for Academic Risks in Online Learning Communities

Abstract

This invention discloses a time-series-aware dynamic early warning method for academic risks in online learning communities, belonging to the field of data processing technology. It adaptively adjusts the time window according to the teaching rhythm, constructing a standardized behavior sequence matrix from learners' original behavior logs. Next, CNN-LSTM and GraphSAGE are used to extract individual cognitive and social interaction features respectively, and behavioral regularity and social activity are dynamically fused to generate fused features. Based on this, deep embedding clustering is used to dynamically identify learning communities within each window, and the evolutionary path of the community is reconstructed through matching scores and the Hungarian algorithm, dividing the evolutionary pattern and life cycle stage, and quantifying the intrinsic risk index. Subsequently, individual fused features are combined with community risk indicators to construct a community embedding time-series sequence, which is input into an LSTM model with an attention mechanism to predict the probability of academic risks. Finally, a graded early warning is generated based on the risk probability, and intervention strategies are matched to achieve dynamic early warning and decision support.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a time-series-aware dynamic early warning method for academic risks in online learning communities. Background Technology

[0002] With the deep integration of information technology and higher education, online learning has become a core form reshaping the global higher education landscape. It breaks down the limitations of time and space, providing learners with ubiquitous and personalized learning channels, and greatly expanding the reach of higher education. However, the inherent temporal and spatial separation of online learning scenarios also brings systemic industry challenges such as learners' strong sense of isolation, weak group interaction, insufficient sustained course participation, and persistently low course completion rates. The risk of academic failure has become a core bottleneck restricting the high-quality development of online education. How to accurately and early warn of academic risks for online learners and provide effective support for teaching intervention has become a key technical problem that urgently needs to be solved in the field of online education.

[0003] Academic performance is deeply correlated with social interaction during the learning process. Social constructivist theory clearly points out that social interaction is the core driving force of learners' cognitive development. Online learning communities (OLCs), as the core analytical unit connecting individual learning and social interaction, can effectively characterize the social learning context in which learners find themselves, providing a new research perspective for analyzing the online learning process and identifying academic risks. Currently, research on online learning community modeling and academic risk early warning mainly falls into three branches. While each branch has made some technological progress, significant technological limitations still exist. First, there's the static modeling technique for online learning communities. This technique is primarily based on social constructivism and the theory of practical communities. On one hand, it has developed various measurement tools, such as community perception scales, to quantify the subjective perception dimensions of online learning communities. On the other hand, it widely applies social network analysis (SNA) methods to construct static social networks based on learners' interaction data, characterizing community interaction structures and identifying core and peripheral participants. Some research also analyzes the interaction patterns of community experience construction and collaborative problem-solving in blended learning environments. However, this technique has fundamental technical limitations: it largely relies on cross-sectional data or static network snapshots for analysis, effectively describing the community state at a specific moment. It completely fails to adapt to the inherent temporal and dynamic nature of online learning processes, unable to capture the continuous trajectory of online learning communities evolving with the course teaching cycle, and struggling to reveal the deep correlation between dynamic changes in community structure and individual academic risks, thus failing to provide effective support for dynamic early warning of academic risks.

[0004] Second, dynamic community evolution tracking techniques. To capture the dynamic evolutionary characteristics of community structures, computational sociology has developed a series of dynamic community discovery methods. Early techniques tracked group evolution through faction filtering algorithms or designed community matching methods based on member overlap to achieve basic tracking of community evolution trajectories. In recent years, the development of Dynamic Graph Neural Networks (DGNNs) has provided more powerful technical tools for dynamic community modeling. Representative techniques include DySAT, which learns dynamic graph representations through a self-attention mechanism; EvolveGCN, which models network evolution by recursively updating graph convolutional network parameters; and TGN, which builds a general dynamic graph learning framework. The aforementioned dynamic modeling techniques provide important methodological references for this field, but when directly applied to online education scenarios, two major technical challenges remain: First, most of these methods are designed for general social networks, whose inherent evolutionary dynamics differ significantly from the evolutionary patterns of online learning communities driven by clear teaching rhythms and task cycles, making them unsuitable for the specific characteristics of online education scenarios. Second, existing work focuses only on the technical aspects of community trajectory tracking, lacking a systematic summary and educational semantic characterization of typical evolutionary patterns such as community formation, development, decline, and reorganization in educational scenarios. This results in a serious disconnect between the analysis results and the explanatory needs in teaching practice, failing to directly serve teaching decisions.

[0005] Third, academic risk early warning technology. This field has long focused on predicting learners' learning outcomes through educational big data analysis. Early technologies combined Hidden Markov Models (HMMs) with Long Short-Term Memory Networks (LSTMs) to capture learning state transitions and long-term behavioral dependencies. With the development of deep learning technology, predictive models based on multi-feature fusion and attention mechanisms have become mainstream. Research has also introduced architectures like Informer to handle long-sequence learning behavior prediction. Cutting-edge research has further expanded to concept drift coping and multimodal data fusion (text / video) to achieve a more comprehensive assessment of learners' learning status. However, existing early warning models generally suffer from core technical flaws: most studies follow an "egocentric" analytical paradigm, treating learners as isolated individuals. The model input features completely ignore the dynamic community environment in which learners exist, making it impossible for early warning systems to distinguish whether academic risk stems from individual factors such as insufficient learning input or from macro-level community dynamics such as the dissipation of community cohesion and the breakdown of learning support networks. This not only severely weakens the interpretability of early warning results but also leads to a lack of targeted subsequent teaching interventions, failing to achieve precise personalized learning support.

[0006] In summary, current technologies related to online learning community modeling and academic risk early warning face three major technological bottlenecks: First, most adopt static analysis frameworks, making it difficult to capture the dynamic evolutionary characteristics of online learning communities during the teaching process; second, they lack a systematic analysis framework, failing to fully track the entire lifecycle of online learning communities from formation to dissipation, and unable to systematically summarize community evolution patterns in educational scenarios; third, academic risk early warning models generally focus on individual behavioral characteristics, neglecting the core impact of learners' structural position and evolutionary trajectory within the community on academic performance. These technological bottlenecks prevent existing technologies from answering the core question of "significant differences in academic performance among different learners with similar learning input," and also hinder early and accurate early warning of academic risks from the perspective of social interaction, severely restricting the practical application of online learning analytics technologies and the systematic improvement of online education quality. Summary of the Invention

[0007] To address the core technical problems of existing online learning community modeling and academic risk early warning technologies, such as the inability of static analysis frameworks to capture the dynamic evolution characteristics of communities, the lack of adaptability of general dynamic community modeling methods to educational scenarios, the neglect of the core impact of the community environment on academic risk by the individualistic early warning paradigm, and the disconnect between community modeling and risk early warning, leading to insufficient early warning accuracy and pedagogical interpretability, this invention provides a time-series-aware dynamic early warning method for academic risk in online learning communities. The technical solution is as follows: On the one hand, a time-series-aware dynamic early warning method for academic risks in online learning communities is provided. This method includes: preprocessing the original behavior logs of learners in online courses, constructing a standardized individual behavior sequence matrix for each learner based on an adaptive time window mechanism driven by teaching rhythm; based on the individual behavior sequence matrix, extracting the learner's individual cognitive features and social interaction features through a multimodal feature learning framework, and generating a fused feature vector for each learner through an adaptive feature fusion mechanism; based on the fused feature vectors of all learners, dynamically identifying the online learning community within each time window through a deep embedding clustering algorithm, and generating window-level community segmentation results. Based on the community segmentation results of adjacent time windows, the evolutionary path of the online learning community is reconstructed by matching the scoring function and the optimal matching algorithm. Evolutionary events are identified and evolutionary patterns are classified and life cycle stages are divided. The intrinsic risk index of different community types and evolutionary stages is quantified. The fusion feature vectors of learners and the intrinsic risk index of the corresponding communities are integrated to construct the temporal feature sequence of community embedding. A risk prediction model is constructed based on a long short-term memory network with attention mechanism to output the academic risk probability of each learner. Based on the academic risk probability and dynamic early warning indicators, a graded early warning signal is generated and matched with corresponding teaching intervention strategies to complete the dynamic early warning and decision support of academic risks.

[0008] Beneficial effects The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: 1. By preprocessing the original behavioral logs of online course learners and constructing a standardized individual behavior sequence matrix based on the adaptive time window mechanism driven by the teaching rhythm, the behavioral characteristics of learners at different stages can be accurately captured, thereby improving the time sensitivity and data reliability of academic risk warning.

[0009] 2. By extracting learners’ individual cognitive features and social interaction features through a multimodal feature learning framework, and generating a fused feature vector through an adaptive feature fusion mechanism, we can comprehensively characterize learners’ learning status and social participation, thereby achieving a multi-dimensional and accurate representation of learners’ academic risks.

[0010] 3. By dynamically identifying online learning communities within each time window through deep embedding clustering algorithms, and combining matching score functions with optimal matching algorithms to reconstruct the community evolution path and quantify the community's inherent risk index, individual risk and community risk are correlated and integrated, thereby realizing the community embedding and dynamic evolution of academic risk early warning.

[0011] 4. By integrating learner feature vectors with corresponding community-based intrinsic risk indices to construct temporal feature sequences of community embedding, and building a risk prediction model based on a long short-term memory network with attention mechanism, we can accurately capture the temporal evolution of academic risk and key influencing factors, thereby achieving accurate output of learner academic risk probability.

[0012] 5. By constructing a multi-task joint loss function, the quality of online learning community identification and the accuracy of risk prediction are simultaneously optimized, thereby balancing the rationality of community segmentation and the accuracy of risk prediction, and thus achieving synergistic optimization of dynamic early warning of academic risks and decision support for teaching intervention. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 A flowchart of the online learning community time-series perception dynamic early warning method for academic risks provided in this application embodiment. Detailed Implementation

[0015] The following provides explanations for some of the terms used in this application. It should be noted that these explanations are for the convenience of those skilled in the art and do not constitute a limitation on the scope of protection claimed in this application.

[0016] like Figure 1The diagram shows a flowchart of the online learning community time-series perception-based dynamic early warning method for academic risks provided in this application embodiment. The method includes the following steps: preprocessing the original behavior logs of learners in online courses. First, the original behavior logs are cleaned to remove invalid behavior records and fill in zero-value behavior records with missing dates. Core behavior types strongly correlated with learning engagement and knowledge construction are selected. These core behavior types include resource access, page navigation, video viewing, forum posting, forum replying, homework submission, and quiz attempts. Using days as the smallest time granularity, the number of occurrences of each type of behavior for each learner is aggregated and statistically analyzed daily to generate the original multidimensional behavior time series for each learner. Based on the teaching rhythm-driven... An adaptive time window mechanism divides the original behavior time series into windows, obtaining behavior subsequences within each window. A logarithmic transformation is applied to the original behavior counts within each window to eliminate the long-tailed distribution of behavior counts and improve feature distribution stability. Based on historical window statistics, the logarithmically transformed behavior data is recursively standardized, ultimately constructing a standardized individual behavior sequence matrix for each learner within their corresponding window. This completes the full-process preprocessing of the original behavior logs. Based on the teaching rhythm-driven adaptive time window mechanism, a standardized individual behavior sequence matrix is ​​constructed for each learner. Using this individual behavior sequence matrix, a multimodal feature learning framework is used to extract the learner's individual cognitive and social characteristics. Interactive features are used to generate a fused feature vector for each learner through an adaptive feature fusion mechanism. Based on the fused feature vectors of all learners, a deep embedding clustering algorithm is used to dynamically identify online learning communities within each time window, generating window-level community partitioning results. The deep embedding clustering algorithm is based on an autoencoder architecture, consisting of an encoder and a decoder. First, the fused feature matrices of all learners in the current window are input into the autoencoder. Unsupervised pre-training minimizes the reconstruction loss of the autoencoder, learning the low-dimensional embedding representation of the fused features and initializing the model parameters. The K-means algorithm is then used to cluster the low-dimensional embedding features obtained from the pre-training, generating initial cluster centers and initial soft clusters for each learner. Assignment probabilities are defined, with soft assignment probabilities representing the probability that a learner will be assigned to each cluster. Using the student t-distribution as the kernel function, the similarity between the learner's embedded features and each cluster center is calculated to update the learner's soft assignment probability to each cluster. Simultaneously, a target distribution is defined to improve the discriminative power of the clustering results. A clustering loss based on KL divergence is constructed to measure the difference between the soft assignment probability and the target distribution. This is combined with the reconstruction loss of the autoencoder to construct the total clustering loss function. Backpropagation iteratively optimizes the autoencoder parameters and cluster centers, simultaneously optimizing feature representation and cluster assignment results until the clustering loss converges. Finally, a unique cluster label is assigned to each learner, completing the clustering of the online learning community within the current window.Based on the community segmentation results of adjacent time windows, the evolutionary path of the online learning community is reconstructed through matching score functions and optimal matching algorithms. Evolutionary events are identified, and evolutionary pattern classification and life cycle stage division are completed. The intrinsic risk index of different community types and evolutionary stages is quantified. The fusion feature vectors of learners and the corresponding intrinsic risk indexes of communities are integrated to construct a temporal feature sequence of community embedding. A risk prediction model is built based on a long short-term memory network with attention mechanisms, outputting the academic risk probability of each learner. Based on the academic risk probability and dynamic early warning indicators, a graded early warning signal is generated and matched with corresponding teaching intervention strategies to complete dynamic early warning and decision support for academic risks. Based on the distribution of historical academic risk probabilities, the 30th and 70th percentiles were set as two risk thresholds, classifying academic risks into three levels: low risk (concern level), medium risk (warning level), and high risk (alarm level). Combining the abnormalities of dynamic warning indicators, the final warning level was determined. If multiple warning indicators deteriorated simultaneously and the academic risk probability exceeded the 70th percentile, it was classified as high risk (alarm level). A mapping relationship was established between warning levels and teaching intervention strategies: one-on-one academic consultation and community integration guidance intervention were matched for high risk (alarm level); targeted feedback and interactive enhancement intervention were matched for medium risk (warning level); and routine monitoring and peer effect guidance intervention were matched for low risk (concern level).

[0017] The core inventive concept of this embodiment lies in breaking through the core limitations of static analysis and individual-centered paradigm in existing online academic risk warning, and constructing a full-chain technical framework of "time-sequence perception - community embedding - dynamic early warning - teaching decision". This method first preprocesses behavioral data using an adaptive time window driven by the teaching rhythm, aligning with the inherent characteristics of the weekly teaching cycle of online courses and solving the problem that fixed windows cannot adapt to changes in behavioral patterns at different teaching stages. Next, it collaboratively models learners' individual cognitive and social interaction attributes through a multimodal feature learning framework, compensating for the incomplete characterization of learners by single features. Then, it dynamically identifies online learning communities through a deep embedding clustering algorithm, achieving continuous and accurate capture of community structure. Furthermore, it completes full lifecycle modeling of the community through evolutionary tracking and pattern recognition, quantifying the intrinsic relationship between community structure and academic risk, and establishing a mapping bridge from macro-level community dynamics to micro-level individual risk. Finally, it integrates community risk characteristics into individual time sequences through a community embedding mechanism, combining this with a long short-term memory network with an attention mechanism to complete risk prediction, completely breaking the egocentric paradigm of traditional early warning models. Finally, through tiered early warning and intervention strategy matching, it transforms the algorithm's prediction results into actionable teaching decision support, providing a systematic solution for academic risk early warning in online learning environments that combines high accuracy and high interpretability.

[0018] Furthermore, based on the teaching rhythm-driven adaptive time window mechanism, the steps for constructing a standardized individual behavior sequence matrix for each learner include: calculating the total behavioral activity and historical average activity within the current time window, and calculating the optimal length of the current window using a dynamic adjustment formula, which is: ; in, Let be the length of the t-th window. The historical average activity level The total activity level within the current window, β=3 / The design of this coefficient ensures that when the activity level... At that time, the window length was exactly the maximum of 10 days; when = When the window length is equal to the baseline length of 7 days; when =2 When learning activity is abnormally high, the window length is shortened to a lower limit of 5 days. Therefore, when learning activity is below the historical average, the observation window is automatically extended to obtain a more robust estimate; when learning activity is abnormally high, the window is automatically shortened to capture subtle patterns of change during periods of high activity.

[0019] After performing a logarithmic transformation on the learners' raw behavior counts, a recursive standardization was performed based on statistics from a historical window. This process was first applied to the raw behavior counts. Perform a logarithmic transformation: To improve the stationarity of the feature distribution; and then based on historical windows to Standardize by calculating the mean and standard deviation: ;in, Let the standardized behavioral value of learner i on day d in window t be the value of behavior k, which quantifies the learner's learning engagement at the corresponding time and behavioral dimension. and standard deviation Based solely on historical data, generate a behavioral sequence matrix for each learner within the corresponding window: .

[0020] In this embodiment, the adaptive time window mechanism and behavior sequence standardization steps are further defined. Its core design revolves around the teaching time-series characteristics and causal modeling requirements of online education scenarios. Regarding the window length design, a 7-day baseline length is set based on the weekly teaching cycle commonly used in online courses, while a 5-10 day scaling range is also included, balancing the core contradiction between noise robustness and sensitivity to change in time-series analysis. Through an activity-driven dynamic adjustment formula, the window length can adaptively change with the behavioral activity level during the teaching phase. During periods of exceptionally high behavioral activity, such as pre-exam review, the window is automatically shortened to 5-7 days, improving the temporal resolution of critical phases. During periods of low behavioral activity, the window is automatically extended, ensuring the robustness of feature estimation and completely resolving the problem that fixed windows cannot adapt to the teaching pace or accommodate the modeling needs of different stages. In the sequence standardization stage, logarithmic transformation is first used to improve the distribution stability of the original behavioral count features, followed by recursive standardization based solely on historical window statistics. This completely avoids future data leakage, strictly ensuring the causal validity of the early warning model and providing high-quality, unbiased standardized behavioral sequence input for subsequent multimodal feature learning and time-series risk prediction.

[0021] Furthermore, the steps for extracting learners' individual cognitive features and social interaction features using a multimodal feature learning framework include: employing a CNN-LSTM hybrid network as an individual cognitive encoder to encode the learner's behavioral sequence matrix; extracting local behavioral patterns through one-dimensional convolutional layers in the encoder, including regular login cycles, resource access frequency features, assignment submission timing patterns, and test attempt behavior features; modeling long-term temporal dependencies through LSTM layers to capture learners' long-term learning strategies around teaching nodes, including phased learning planning, learning sprint behavior before assignment deadlines, and pre-exam review behavior patterns; and outputting the learner's individual cognitive feature vector, which fully represents the learner's learning engagement level, self-regulation ability, and cognitive learning strategies. For each time window, a weighted interaction graph is constructed among learners, with learners as nodes and interaction relationships between learners as edges. Edge weights are calculated by comprehensively considering interaction frequency, interaction depth, and timeliness, defining the weighted interaction graph as... ; where the node set For all learners within the current time window, the edge set Effective interactions among learners include replying to and being replied to in forum posts, interactions in collaborative tasks, and sharing and commenting on learning resources; this applies to edge sets. Each interactive edge The interaction frequency, interaction depth, and timeliness of the two learners were calculated separately. Interaction frequency is the total number of effective interactions between learners i and j within the window; interaction depth is the average text length and interaction level of the interactive content between the two learners within the window; and timeliness is the time decay coefficient calculated based on the distance between the time of the interaction and the current teaching node. After normalizing the three dimensions, the edge weights were obtained through linear weighted summation. The higher the edge weight, the stronger the interaction relationship and the higher the quality of the interaction among learners. Based on the node set, edge set and edge weight, a weighted interaction graph within the current time window is constructed, providing a structured network data foundation for the extraction of learners' social interaction characteristics. GraphSAGE graph neural network is used as the social interaction encoder to aggregate the neighbor information of nodes and generate a social interaction feature vector for each learner. Specifically, firstly, a fixed number of multi-order neighbor nodes are sampled for each target node through the neighbor sampling layer of GraphSAGE to avoid the complexity explosion of graph computation. Then, an aggregation layer is used to aggregate the neighbor information of the target node using a mean aggregation function with edge weights to generate the neighbor aggregation feature of the target node. During the aggregation process, the neighbor node with the higher the edge weight has a greater contribution to the aggregation feature. After concatenating the initial feature of the target node with the neighbor aggregation feature, the feature transformation is performed through a fully connected layer and a non-linear activation function to generate the embedding vector of the target node. After multi-layer graph convolution operation of GraphSAGE, the final output is a fixed-dimensional social interaction feature vector. The neighbor information includes the feature information of the target node's preset-order neighbor nodes in the weighted interaction graph, the edge weight information between the target node and each neighbor node, including the interaction strength of first-order neighbors, the network transitivity interaction information of second-order neighbors, and the distribution of behavioral features of neighbor nodes.

[0022] In this embodiment, the multimodal feature learning framework is specifically defined, and a dual-encoder architecture is constructed from two core dimensions: individual cognition and social interaction, to comprehensively characterize learners' learning and social attributes. The individual cognition encoder employs a CNN-LSTM hybrid network architecture. It extracts local high-frequency behavioral patterns in the temporal dimension through one-dimensional convolutional layers, such as learners' regular logins, resource access frequency, and assignment submission cycles—short-term behavioral patterns. Then, it models long-term temporal dependencies through LSTM layers, capturing learners' long-term learning strategies around assignment deadlines, such as learning sprints and phased learning plans, thus fully representing learners' cognitive investment and self-regulation capabilities. The social interaction encoder, based on the dynamic social network characteristics of the online learning environment, constructs a weighted interaction graph for each time window, integrating interaction frequency, depth, and timeliness. Then, it uses a GraphSAGE graph neural network to generate social embeddings of nodes through neighbor information aggregation, accurately characterizing learners' structural position, interaction intensity, and social influence in the interaction network. This overcomes the shortcomings of traditional methods that can only characterize individual behavior and cannot represent the social interaction context, providing a two-dimensional, highly discriminative feature foundation for subsequent online learning community identification and risk prediction.

[0023] Furthermore, the steps for generating the fused feature vector for each learner via the adaptive feature fusion mechanism include: calculating the learner's behavioral regularity and social activity within the corresponding window, where behavioral regularity is obtained by normalizing the standard deviation of the total number of daily behaviors within the window, and social activity is obtained by normalizing the learner's node degree centrality in the interaction graph; and calculating the dynamic fusion weights using the Softmax function, the formula for which is: ; in, Let i be the fusion weight of learner i in the t-th window. For behavioral regularity, For social activity, The temperature parameter is used; based on dynamic fusion weights, individual cognitive features and social interaction features are weighted and fused to generate a fused feature vector. The fusion formula is as follows: ; ; , ; in, Let be the standardized individual behavior sequence matrix of learner i within the t-th time window. For individual cognitive encoders, i.e., CNN-LSTM hybrid networks, It is the set of learnable parameters of an individual cognitive encoder. Let i be the individual cognitive feature vector of learner i within the t-th time window. The individual cognitive feature vector is located in 3D real vector space, As a dimension of individual cognitive characteristics, This is a graph of empowered interactions among learners within the t-th time window. For learner i, the corresponding interaction graph node. For social interaction encoders, namely GraphSAGE graph neural networks, For the set of learnable parameters of the social interaction encoder, Let i be the social interaction feature vector of learner i within the t-th time window. The social interaction feature vector is located 3D real vector space, As a dimension of social interaction characteristics, Let i be the fused feature vector of learner i within the t-th time window. The location of the fused feature vector 3D real vector space, To integrate the dimensions of features, .

[0024] In this embodiment, the present invention further refines the adaptive feature fusion mechanism. Addressing the heterogeneity of online learner participation patterns, a dynamic weighted feature fusion scheme is designed, resolving the core issue that fixed-weight fusion cannot adapt to different types of learners. This mechanism first obtains a learner's behavioral regularity index by normalizing the standard deviation of the total daily behavior within a window, and obtains a social activity index by normalizing the degree centrality of nodes in the interaction graph, quantifying the learner's dominant behavioral attribute and socially driven attribute respectively. Then, a Softmax function with a temperature parameter τ is used to calculate the dynamic fusion weights. For learners with high social activity, the weight of social interaction features is automatically increased; for socially isolated learners with strong behavioral regularity, the weight of individual cognitive features is automatically increased, achieving an adaptive balance between the two types of features. The design of the temperature parameter τ effectively avoids the weight polarization problem, ensuring the robustness of the fusion process. This fusion scheme not only achieves a silhouette coefficient of 0.62 and a modularity of 0.64 in clustering tasks, significantly outperforming single-feature schemes, but also the fusion weights themselves can serve as an effective indicator for distinguishing learner participation types, providing an additional interpretable dimension for subsequent differentiated teaching interventions, and realizing the organic unity of feature fusion and instructional semantics.

[0025] Furthermore, the steps for dynamically identifying online learning communities within each time window using a deep embedding clustering algorithm include: constructing a fusion feature matrix from the fused feature vectors of all learners within the current window and inputting it into the deep embedding clustering algorithm; jointly optimizing feature representation and cluster assignment by optimizing KL divergence clustering loss and autoencoder reconstruction loss, assigning a unique cluster label to each learner; adaptively determining the optimal number of clusters based on silhouette coefficient and modularity index, generating window-level online learning community partitioning results that satisfy completeness and mutual exclusivity. The silhouette coefficient is a quantitative indicator measuring cluster cohesion and separation; for a single learner, its silhouette coefficient is the ratio of the learner's cohesion to its clustering degree. The average similarity of other samples within a cluster, minus the average similarity between the learner and all samples in the nearest neighbor cluster, and then divided by the maximum of the two, yields the overall silhouette coefficient of the clustering result, which is the average of the silhouette coefficients of all learners. The silhouette coefficient ranges from -1 to 1; the closer the value is to 1, the higher the intra-cluster similarity and the better the inter-cluster discrimination. Modularity is a quantitative indicator for measuring the quality of the network community partitioning structure. It is used to quantify the difference between the sum of edge weights within a community and the expected sum of edge weights in a random network. The modularity ranges from -1 to 1; the closer the value is to 1, the tighter the interaction connections within the community and the more significant the community structure. The search range for the preset number of clusters is […]. , Within the search range, all possible cluster numbers are traversed. For each cluster number, the overall silhouette coefficient and modularity are calculated. A comprehensive evaluation function for cluster quality is constructed, and the silhouette coefficient and modularity are summed with equal weights. The cluster number corresponding to the maximum value of the comprehensive evaluation function is selected as the optimal cluster number for the current window. Based on the optimal cluster number, a window-level online learning community partitioning result is generated. This partitioning result ensures that each learner is assigned to a unique community through a hard allocation mechanism, satisfying the completeness that all learners have a place to belong, and the mutual exclusion that there is no member overlap between any communities.

[0026] In this embodiment, the present invention specifically defines the steps for identifying online learning communities based on deep embedding clustering, solving the core problems of feature mismatch and clustering target, and subjective bias due to fixed cluster numbers, in the traditional two-stage method of "feature extraction first, then clustering". This step inputs the fused feature matrix of all learners into the deep embedding clustering (DEC) algorithm. By jointly optimizing the clustering loss based on KL divergence and the autoencoder reconstruction loss, feature representation and cluster assignment are simultaneously optimized in the low-dimensional feature space of neural network learning. This allows feature learning to directly serve the community partitioning target, significantly improving the quality of clustering results. Simultaneously, through a comprehensive evaluation function of silhouette coefficient and modularity, the optimal number of clusters is adaptively determined within a preset range, enabling the community partitioning results to dynamically adapt to the natural evolution of community structure under different teaching stages, avoiding partitioning bias caused by the assumption of a fixed number of clusters. The algorithm's hard allocation mechanism assigns a unique cluster label to each learner, ensuring the completeness and mutual exclusivity of the community partitioning results. This precisely matches the core definitions of cohesion and behavioral homogeneity in online learning community structure, providing a reliable and stable window-level community partitioning foundation for subsequent community evolution tracking and risk quantification.

[0027] Furthermore, by reconstructing the evolutionary path of the online learning community through a matching score function and an optimal matching algorithm, the steps for identifying evolutionary events include: constructing a matching score function, integrating the overlap of members and feature similarity between adjacent windows of the community, and quantifying the degree of matching between the two communities. The matching score function is as follows: ; in, The matching score is the score between community a in the previous window and community b in the current window. For Jaccard similarity, , These are the average feature vectors of the two communities, To balance hyperparameters, For the online learning community numbered a within the (t-1)th time window, Let b be the online learning community within the t-th time window; A matching score matrix for adjacent window communities is generated based on a matching score function. The optimal matching relationship is then solved using the Hungarian algorithm. The Hungarian algorithm constructs a complete bipartite graph model by treating all communities from the previous window as the left vertex set and all communities from the current window as the right vertex set. The matching score between each community in the left vertex set and each community in the right vertex set is calculated based on the matching score function, generating a dimension of... × The matching score matrix, where This represents the number of communities in the previous window. The number of communities in the current window represents the number of communities. A higher matching score indicates a higher degree of matching between the two communities. The matching score matrix is ​​converted into the cost matrix required by the Hungarian algorithm. By taking the negative value of the matching score, the problem of maximizing the total matching score is transformed into the problem of minimizing the total cost. The Hungarian algorithm is used to perform iterative solution, including four core steps: row reduction, column reduction, covering zero elements, and augmenting path search, to solve for the maximum weight perfect matching of the bipartite graph and obtain the globally optimal one-to-one matching relationship between the communities in the previous window and the communities in the current window. For isolated communities that do not match a corresponding community, they are marked as unmatched nodes, providing a complete matching relationship foundation for subsequent community evolution event identification. Based on the optimal matching relationship, four basic evolution events of communities are identified: continuation, splitting, merging, and dissipation. The determination condition for continuation events is: community a in the previous window and community b in the current window are a one-to-one optimal match, and the Jaccard similarity of the members of the two communities is greater than the preset continuation threshold. Community a is determined to continue as community b, belonging to the continuous evolution of the same community. Splitting events... The criteria for determining a split event are as follows: If a single community 'a' in the previous window forms an optimal match with multiple communities in the current window, and the overlap of members between community 'a' and each matched community is greater than a preset split threshold, then community 'a' is determined to have split, and it differentiates into multiple communities in the current window. The criteria for determining a merge event are as follows: If multiple communities in the previous window form an optimal match with a single community 'b' in the current window, and the overlap of members between each preceding community and community 'b' is greater than a preset merge threshold, then multiple preceding communities are determined to have merge, and they merge into community 'b' in the current window. The criteria for determining a dissipation event are as follows: If community 'a' in the previous window has no optimal match in the current window, and the Jaccard similarity of its members with all communities in the current window is less than a preset dissipation threshold, then community 'a' is determined to have dissipated, and its lifecycle terminates. Based on the evolution event identification results of all adjacent windows throughout the entire course cycle, a continuous evolution link from formation to termination is established for each online learning community, recording all evolution events it experiences, and reconstructing the complete evolution path of all online learning communities.

[0028] In this embodiment, the present invention specifically defines the steps for tracing the community evolution path, achieving a complete and accurate reconstruction of the full-cycle evolution trajectory of an online learning community. This step first constructs a matching score function that integrates member overlap and behavioral pattern stability. The Jaccard similarity metric is used to quantify the continuity of member overlap between adjacent window communities, and the Euclidean distance in the feature space is used to quantify the stability of group behavioral patterns. Compared to traditional matching methods that rely solely on member overlap, this method can more comprehensively and accurately measure the matching degree between two communities, avoiding matching errors caused by minor member movements. Then, based on the matching score matrix, the Hungarian algorithm is used to solve for the globally optimal matching relationship between adjacent window communities, rather than a local greedy matching, ensuring the global accuracy of the evolution path reconstruction. Finally, based on the optimal matching results, the system identifies four basic evolutionary events of the community: continuation, splitting, merging, and dissipation. This fully reconstructs the continuous evolutionary process of the online learning community from its formation to its dissipation, bridging the key link from static community snapshots to dynamic evolutionary analysis, and providing core temporal evolutionary data support for subsequent evolutionary pattern classification and structural risk quantification.

[0029] Furthermore, the steps to classify evolutionary patterns and life cycle stages, and quantify the inherent risk index of different community types and evolutionary stages, include: based on the life cycle characteristics of communities, including duration of existence, member turnover rate, and interaction density, classifying online learning communities into four evolutionary patterns: long-term stable, short-term task-oriented, long-term mobile, and short-term loose. Specifically, the three core life cycle characteristics of community duration of existence, member turnover rate, and interaction density are selected as the core basis for evolutionary pattern classification. Duration of existence is the total time window spanned by the community from its formation to its dissolution, and member turnover rate is... The turnover rate is the percentage of community members changing between adjacent windows, and the interaction density is the average edge weight and interaction frequency among members within the community. Based on the lifecycle feature dataset of all communities throughout the entire course cycle, cluster analysis is used to classify online learning communities into four evolutionary modes: long-term stable, short-term task-oriented, long-term fluid, and short-term loose. The criteria for long-term stable are: a duration greater than 50% of the total number of course windows, a member turnover rate less than 10%, and an interaction density higher than 4 points (out of 5). The criteria for short-term task-oriented are: a duration less than the total number of course windows. The characteristics of the long-term mobility type are: the duration of the course exceeds 50% of the total number of course windows, the member turnover rate is 40%-70%, and the interaction density is 2-3 points; the characteristics of the short-term loose type are: the duration of the course is less than 20% of the total number of course windows, the member turnover rate is higher than 80%, and the interaction density is lower than 2 points. This classification method achieves standardized and quantifiable classification of all online learning community evolution patterns, providing a classification basis for quantifying the inherent risks of the community; and statistically analyzes the community members corresponding to each evolution pattern. The historical academic failure rate was calculated, and the inherent risk index of the corresponding community type was assigned. Using historical teaching data of the courses as a statistical sample, the evolutionary pattern classification results of all online learning communities in the sample were extracted, as well as the final academic outcome labels of all learners in the corresponding communities. The academic outcome labels were divided into two categories: academic success and academic failure. Academic failure includes failing the course and dropping out midway. For the four evolutionary patterns, the total number of all learners in the corresponding communities and the number of learners who ultimately failed were counted. The historical academic failure rate of each evolutionary pattern community was calculated using the following formula: ,in, This represents the set of all learners who have historically belonged to the OLC of type T. learner Academic outcome tags Representing learners' ultimate academic failure, and based on historical academic failure rates, an intrinsic risk index is assigned to four evolutionary community models. The short-term loosely structured model has the highest intrinsic risk index, while the long-term stable model has the lowest. The intrinsic risk indices of the four community models exhibit a continuous gradient distribution. Simultaneously, based on the community's life cycle stage, a corresponding short-term risk coefficient is matched to the intrinsic risk index to correct for the community's immediate risk. The short-term risk coefficients for the evolution and dissipation stages are higher than those for the formation and stability stages. Finally, the ultimate intrinsic risk index for each community within its corresponding time window is obtained, achieving dynamic and precise quantification of community structural risk. The life cycle of online learning communities is divided into four stages: formation, stability, evolution, and dissipation. The life cycle stage of a community is determined based on the changing trends of member turnover and interaction density, and a corresponding short-term risk coefficient is matched to each stage.

[0030] In this embodiment, the present invention further defines the steps of evolution pattern recognition, life cycle division and structural risk quantification, and constructs a systematic community evolution analysis system. For the first time, it provides a complete typological framework and empirical evidence for the quantitative correlation between "community structure and academic risk" in online learning scenarios. This step, based on three core lifecycle characteristics of online learning communities—existence duration, member turnover rate, and interaction density—classifies online learning communities into four typical evolutionary models: long-term stable, short-term task-oriented, long-term mobile, and short-term loose. By statistically analyzing the academic failure rates of members in each type of community from historical data, an intrinsic risk index for different community types is quantified. Empirical verification demonstrates a strong negative correlation between community structural stability and academic risk. The proportion of high-risk members in short-term loose communities reaches 62.4%, while in long-term stable communities it is only 5.1%, clearly demonstrating the significant impact of community structure on individual academic performance. Simultaneously, the community lifecycle is divided into four continuous stages: formation, stability, evolution, and dissipation. Stage determination is completed by analyzing the changing trends of member turnover rate and interaction density, clarifying that the stable stage has the lowest risk, while the evolution and dissipation stages have the highest immediate risk. This step achieves a quantitative mapping from macro-level community evolutionary models to micro-level individual risks, enabling the early warning model to incorporate community environmental factors into the risk assessment system, fundamentally improving the contextual awareness and interpretability of early warning results.

[0031] Furthermore, based on the fusion feature vectors of fusion learners and the intrinsic risk index of the corresponding communities, the steps for constructing the temporal feature sequence of community embedding include: constructing a dynamic early warning indicator set for each learner in each time window, including participation change rate, behavior pattern drift, social marginalization index, and the intrinsic risk index of the community to which they belong; concatenating the learner's fusion feature vector with the dynamic early warning indicator set to generate the observation vector of the corresponding window; and arranging all observation vectors of the learner from the start of the course to the current window in chronological order to construct the temporal feature sequence of community embedding.

[0032] In this embodiment, the invention specifically defines the steps for constructing the temporal features of community embedding, which directly embodies the core innovation of "community embedding" and completely breaks through the individual-centric paradigm of traditional academic risk warning models. This step constructs an observation vector for each learner in each time window, which not only includes individual features obtained through multimodal fusion but also integrates a dynamic warning indicator set including participation change rate, behavioral pattern drift, social marginalization index, and community risk index. The core innovation lies in directly embedding the inherent risk index of the learner's community into the individual feature sequence, realizing the quantitative implementation of macro-community environmental risks at the individual feature level. The temporal feature sequence constructed in this way not only includes the evolutionary information of the learner's individual learning behavior but also fully integrates the dynamic changes in their social learning context. This enables the warning model to distinguish whether academic risk stems from insufficient individual learning input or from macro-community dynamics such as the dissipation of community cohesion and the breakdown of support networks. This solves the core defects of traditional warning models, such as the inability to attribute risks and poor interpretability, and also gives the final output warning signal clear pedagogical semantics, rather than simply a probability value.

[0033] Furthermore, the steps for constructing a risk prediction model based on a long short-term memory network with an attention mechanism and outputting the academic risk probability for each learner include: inputting the temporal feature sequence of community embedding into the long short-term memory network and extracting the hidden state sequence at each time step; calculating the attention weights for the hidden states at each time step through an attention layer, and generating a context vector by weighted summation. The formula for calculating the attention weights is: ; in, Let i be the attention weight at the j-th time step. Let q be the hidden state corresponding to the time step, q be the learnable query vector, and t be the total number of time windows from the start of the course to the current prediction node. The context vector is input into the fully connected layer, and the Sigmoid activation function outputs the learner's academic risk probability in the current window. The academic risk probability is the predicted probability of the learner's eventual academic failure.

[0034] In this embodiment, the present invention specifically defines the long short-term memory network risk prediction model based on the attention mechanism, achieving accurate and interpretable temporal prediction of academic risk. This model first inputs the temporal feature sequence of community embedding into an LSTM network, effectively capturing the long-term temporal dependency between learning behavior and community evolution, solving the gradient vanishing problem in long sequence prediction, and adapting to the needs of long-term temporal modeling throughout the entire online course cycle. Then, dynamic weights are calculated for the hidden states at each time step through the attention layer. The weight magnitude directly reflects the contribution of the corresponding time window to the final risk prediction. This not only improves the model's predictive performance through weighted key information but also intuitively reveals the key time windows for learner risk formation through weight distribution. Empirical results show that the attention weights are generally higher in the initial stage of the course and the pre-exam review period, highly consistent with the key stages of risk formation in teaching practice, greatly enhancing the model's interpretability. Finally, the context vector obtained by weighted summation is input into a fully connected layer, and the academic risk probability in the 0-1 interval is output through the Sigmoid activation function, intuitively representing the likelihood of learner academic failure. This model outperforms existing mainstream baseline models in all aspects of early warning performance on the OULAD dataset, providing core algorithmic support for early and accurate early warning of online academic risks.

[0035] Furthermore, the present invention also includes a joint model optimization step, specifically: A multi-task joint loss function is constructed, which integrates the clustering loss of deep embedding clustering and the weighted binary cross-entropy loss of risk prediction. The joint loss function is as follows: ; ; ; in, For the total loss, For clustering loss, To predict losses for risk, It is the final predicted probability at the end of the course. For clustering loss based on KL divergence, For autoencoder reconstruction loss, For clustering loss weights, These are the weighting coefficients for positive samples, i.e., samples of academic failure. Let i be the binary label for the learning outcome. This indicates that the learner ultimately failed academically. This indicates that the learner has ultimately achieved academic success. The weighting coefficients for the reconstruction loss of the autoencoder are used to balance the optimization priorities of clustering loss and reconstruction loss; by minimizing the joint loss function, the community identification quality and risk prediction accuracy are optimized simultaneously, and a unified feature representation is learned.

[0036] In this embodiment, the present invention specifically defines the joint optimization steps of the model. A multi-task joint loss function is used to achieve deep collaborative optimization of the two modules: community modeling and risk warning. This solves the core problem that feature representations cannot simultaneously adapt to both tasks due to phased training. The total loss function constructed in this step integrates the clustering loss of unsupervised deep embedding clustering and the weighted binary cross-entropy loss of supervised risk prediction. By balancing the optimization priorities of the two tasks through hyperparameters, the model can learn a unified feature representation that is beneficial for both community structure identification and effective prediction of academic risks. This achieves collaborative optimization of the two core modules, rather than a simple process concatenation. The clustering loss integrates the clustering optimization objective based on KL divergence and the autoencoder reconstruction loss, balancing the two through weight coefficients to optimize clustering effects while preventing feature degradation. The risk prediction loss adopts a weighted binary cross-entropy form, addressing the class imbalance problem of low proportion of academic failure samples in online education scenarios through positive sample weights, significantly improving the model's ability to identify high-risk learners. By minimizing the joint loss function to optimize the global parameters of the model, the overall framework of this invention achieves end-to-end training optimization, which is one of the core reasons why the model performance surpasses existing staged modeling methods. At the same time, it also greatly simplifies the model training and deployment process.

Claims

1. A time-series-aware dynamic early warning method for academic risks in online learning communities, characterized in that, Includes the following steps: The raw behavioral logs of learners in online courses are preprocessed, and a standardized individual behavioral sequence matrix is ​​constructed for each learner based on an adaptive time window mechanism driven by the teaching rhythm. Based on the individual behavior sequence matrix, the learner's individual cognitive features and social interaction features are extracted through a multimodal feature learning framework, and a fused feature vector for each learner is generated through an adaptive feature fusion mechanism. Based on the fused feature vectors of all learners, the online learning community within each time window is dynamically identified through a deep embedding clustering algorithm, generating window-level community segmentation results. Based on the community segmentation results of adjacent time windows, the evolution path of the online learning community is reconstructed by matching score function and optimal matching algorithm, evolutionary events are identified and evolutionary pattern classification and life cycle stage division are completed, and the inherent risk index of different community types and evolutionary stages is quantified. By combining the fusion feature vectors of learners with the intrinsic risk index of the corresponding communities, a temporal feature sequence of community embedding is constructed. A risk prediction model is built based on a long short-term memory network with an attention mechanism, and the academic risk probability of each learner is output. Based on the academic risk probability and dynamic early warning indicators, a graded early warning signal is generated and matched with corresponding teaching intervention strategies to complete the dynamic early warning and decision support for academic risks.

2. The online learning community time-series-aware dynamic early warning method for academic risks as described in claim 1, characterized in that: The adaptive time window mechanism based on teaching rhythm, which constructs a standardized individual behavior sequence matrix for each learner, includes the following steps: Calculate the total activity level and historical average activity level within the current time window, and then calculate the optimal length of the current window using a dynamic adjustment formula. The dynamic adjustment formula is as follows: ； in, Let be the length of the t-th window. The historical average activity level The total activity level within the current window, β=3 / ; After performing a logarithmic transformation on the learners' original behavior counts, a recursive standardization is performed based on the statistics of the historical window to generate a behavior sequence matrix for each learner within the corresponding window.

3. The online learning community time-series-aware dynamic early warning method for academic risks as described in claim 1, characterized in that: The steps of extracting learners' individual cognitive features and social interaction features using a multimodal feature learning framework include: A CNN-LSTM hybrid network is used as an individual cognitive encoder to encode the learner's behavior sequence matrix. Local behavior patterns are extracted through one-dimensional convolutional layers, and long-term temporal dependencies are modeled through LSTM layers to output the learner's individual cognitive feature vector. For each time window, a weighted interaction graph is constructed among learners, with learners as nodes and the interaction relationships between learners as edges. The edge weights are calculated by combining interaction frequency, interaction depth, and timeliness. The GraphSAGE graph neural network is used as a social interaction encoder to aggregate the neighbor information of nodes and generate a social interaction feature vector for each learner. The neighbor information consists of the feature information of the target node's neighbor nodes of a preset order in the weighted interaction graph and the edge weight information between the target node and each neighbor node.

4. The online learning community time-series-aware dynamic early warning method for academic risks as described in claim 3, characterized in that: The steps for generating the fused feature vector for each learner via the adaptive feature fusion mechanism include: The learner's behavioral regularity and social activity within the corresponding window are calculated. The behavioral regularity is obtained by normalizing the standard deviation of the total number of daily behaviors within the window, and the social activity is obtained by normalizing the degree centrality of the learner's nodes in the interaction graph. The dynamic fusion weights are calculated using the Softmax function, and the formula for calculating the dynamic fusion weights is as follows: ； in, Let i be the fusion weight of learner i in the t-th window. For behavioral regularity, For social activity, For temperature parameters; Based on dynamic fusion weights, individual cognitive features and social interaction features are weighted and fused to generate a fused feature vector. The fusion formula is as follows: ； ； ， ； in, Let be the standardized individual behavior sequence matrix of learner i within the t-th time window. For individual cognitive encoders, i.e., CNN-LSTM hybrid networks, It is the set of learnable parameters of an individual cognitive encoder. Let i be the individual cognitive feature vector of learner i within the t-th time window. The individual cognitive feature vector is located in 3D real vector space, As a dimension of individual cognitive characteristics, This is a graph of empowered interactions among learners within the t-th time window. For learner i, the corresponding interaction graph node. For social interaction encoders, namely GraphSAGE graph neural networks, For the set of learnable parameters of the social interaction encoder, Let i be the social interaction feature vector of learner i within the t-th time window. The social interaction feature vector is located 3D real vector space, As a dimension of social interaction characteristics, Let i be the fused feature vector of learner i within the t-th time window. The location of the fused feature vector 3D real vector space, To integrate the dimensions of features, .

5. The online learning community time-series-aware dynamic early warning method for academic risks as described in claim 1, characterized in that: The steps for dynamically identifying online learning communities within each time window using a deep embedding clustering algorithm include: Construct a fusion feature matrix from the fused feature vectors of all learners in the current window, and input it into the deep embedding clustering algorithm; By optimizing the KL divergence clustering loss and the autoencoder reconstruction loss, feature representation and cluster assignment are jointly optimized to assign a unique cluster label to each learner. The optimal number of clusters is adaptively determined based on the silhouette coefficient and modularity index, generating window-level online learning community partitioning results that satisfy completeness and mutual exclusivity.

6. The online learning community time-series-aware dynamic early warning method for academic risks as described in claim 1, characterized in that: The steps of reconstructing the evolutionary path of the online learning community by matching the score function and the optimal matching algorithm, and identifying evolutionary events, include: A matching score function is constructed to quantify the degree of matching between two communities by combining the overlap of members and feature similarity between adjacent windows. The matching score function is as follows: ； in, The matching score is the score between community a in the previous window and community b in the current window. For Jaccard similarity, , These are the average feature vectors of the two communities, To balance hyperparameters, For the online learning community numbered a within the (t-1)th time window, Let b be the online learning community within the t-th time window; The matching score matrix of adjacent window communities is generated based on the matching score function, and the optimal matching relationship is solved by the Hungarian algorithm. Based on the optimal matching relationship, the four basic evolutionary events of the community are identified: continuation, splitting, merging, and dissipation, and the complete evolutionary path of all online learning communities is reconstructed.

7. The online learning community time-series-aware dynamic early warning method for academic risks as described in claim 6, characterized in that: The steps for completing the classification of evolutionary patterns and life cycle stages, and quantifying the inherent risk index of different community types and evolutionary stages include: Based on the life cycle characteristics of communities, including duration of existence, member turnover rate, and interaction density, online learning communities are divided into four evolutionary modes: long-term stable, short-term task-oriented, long-term mobile, and short-term loose. Statistically analyze the historical academic failure rate of community members corresponding to various evolutionary patterns, and calculate and assign the intrinsic risk index of the corresponding community type; The life cycle of online learning communities is divided into four stages: formation, stability, evolution, and dissipation. The stage of the life cycle of the community is determined based on the changing trends of member turnover and interaction density, and the corresponding short-term risk coefficient is matched.

8. The online learning community time-series-aware dynamic early warning method for academic risks as described in claim 7, characterized in that: The steps for constructing the temporal feature sequence of community embedding based on the fusion feature vector of the fusion learner and the intrinsic risk index of the corresponding community include: For each learner, a dynamic early warning indicator set is constructed for each time window, including participation change rate, behavior pattern drift, social marginalization index, and the inherent risk index of the community to which they belong; The learner's fused feature vector is concatenated with the dynamic early warning indicator set to generate the observation vector for the corresponding window; Arrange all observation vectors of the learner from the start of the course to the current window in chronological order to construct a temporal feature sequence of community embedding.

9. The online learning community time-series-aware dynamic early warning method for academic risks as described in claim 1, characterized in that: The steps for constructing a risk prediction model based on a long short-term memory network with an attention mechanism and outputting the academic risk probability for each learner include: The temporal feature sequence embedded in the community is input into the Long Short-Term Memory network to extract the hidden state sequence at each time step; Attention weights are calculated for the hidden states at each time step using an attention layer, and then weighted and summed to generate a context vector. The formula for calculating the attention weights is as follows: ； in, Let i be the attention weight at the j-th time step. Let q be the hidden state corresponding to the time step, q be the learnable query vector, and t be the total number of time windows from the start of the course to the current prediction node. The context vector is input into the fully connected layer, and the Sigmoid activation function outputs the learner's academic risk probability in the current window. The academic risk probability is the predicted probability of the learner's eventual academic failure.

10. The method for dynamic academic risk early warning of online learning communities based on time-series awareness and community embedding as described in claim 1, characterized in that, It also includes a joint optimization step for the model, specifically: A multi-task joint loss function is constructed, which integrates the clustering loss of deep embedding clustering and the weighted binary cross-entropy loss of risk prediction. The joint loss function is as follows: ； ； ； in, For the total loss, For clustering loss, To predict losses due to risk, It is the final predicted probability at the end of the course. For clustering loss based on KL divergence, For autoencoder reconstruction loss, For clustering loss weights, These are the weighting coefficients for positive samples, i.e., samples of academic failure. Let i be the binary label for the learning outcome. This indicates that the learner ultimately failed academically. This indicates that the learner has ultimately achieved academic success. These are the weighting coefficients for the autoencoder reconstruction loss, used to balance the optimization priorities of clustering loss and reconstruction loss; By minimizing the joint loss function, the quality of community identification and the accuracy of risk prediction are simultaneously optimized, and a unified feature representation is learned.

Images