Online cloud teaching user data quantification method and system, and storage medium

Abstract

The invention discloses an online cloud teaching user data quantification method and system and a storage medium, and the method comprises the following steps: S1, collecting user data, and carrying out the preprocessing of the user data; s2, constructing a heterogeneous time sequence diagram, and generating a node embedding representation by adopting a dual-channel attention mechanism and fusing a position coding structure; s3, extracting behavior pattern features of different levels by using a multi-scale image convolution operation, and coding a behavior migration relationship of the user in a continuous learning process; s4, establishing a positive and negative sample pair, generating a label distribution vector, and outputting a corresponding state clustering label; s5, constructing a course content index, establishing a mapping relation between a course node and a state clustering label, and generating a behavior quantification matrix; and S6, inputting the behavior quantification matrix into a cloud quantitative analysis engine, and generating a multi-dimensional user behavior quantification result. According to the invention, quantification and evaluation of online cloud teaching user behaviors can be realized, and the accuracy of learning process monitoring and the real-time performance of teaching feedback are improved.

Description

Technical Field

[0001] The present invention relates to the field of online education technology, and in particular to a method, system and storage medium for quantifying user data of online cloud teaching. Background Art

[0002] With the continuous advancement of educational informatization, online cloud teaching platforms have gradually become one of the mainstream teaching models, demonstrating significant advantages in improving the efficiency of teaching resource sharing and expanding the time and space boundaries of learning. In particular, in scenarios such as multi-user concurrency, remote access, and large-scale course distribution, cloud computing-based teaching architectures can provide students with highly flexible, personalized learning services and provide comprehensive data support for teachers and administrators. However, current mainstream cloud teaching platforms still have significant shortcomings in analyzing and understanding user behavior data. Traditional learning process evaluation methods often rely on static log analysis, test result statistics, and manual annotation, which makes it difficult to truly reflect the dynamic behavior evolution and state migration process of users during the learning process.

[0003] In existing technologies, some platforms attempt to crudely model learning behavior using data features such as click volume, browsing history, and video playback progress. However, such approaches often fail to distinguish the semantic differences behind user behavior and lack a deep understanding of temporal behavioral logic and learning motivation. Furthermore, common data quantification methods mostly employ static, rule-driven strategies and lack data-driven structured analysis mechanisms. This makes it difficult to capture individual differences in user characteristics across different course structures, learning tasks, and behavioral paths, thus impacting the accuracy of teaching feedback and the reliability of intelligent recommendation systems.

[0004] At the same time, traditional user modeling methods generally lack the ability to model the temporal dependencies between user behaviors, making them unable to effectively characterize the continuity of learning behavior and the state transition trajectory. Some studies that use clustering or classification algorithms to segment users typically analyze based on fixed-dimensional feature vectors, failing to fully explore the graph structure and contextual associations in user behavior. Multi-stage and multi-level learning activities exhibit significant hierarchical and stage-specific behavioral trajectories, but existing approaches have yet to establish a unified modeling framework that can integrate graph structure temporal information with multi-scale behavioral features.

[0005] Furthermore, existing clustering mechanisms often rely on static distance metrics or simple similarity functions, failing to incorporate residual feedback from the evolution of user behavior and the relationship between label distributions for self-supervision and correction. This results in insufficiently representative cluster labels and blurred classification boundaries, severely hindering the scientific and feasible implementation of subsequent behavior quantification and teaching interventions. Furthermore, when it comes to generating behavioral outcomes, most platforms only output basic statistical results and lack structured data analysis methods that integrate both user and course dimensions. This makes it impossible to implement dynamic decision-making and refined profiling based on behavioral data.

[0006] Existing methods, especially when faced with large-scale users and heterogeneous course structures, suffer from poor scalability and generalization, making them unable to adapt to the behavioral expression needs of various teaching scenarios. More importantly, there is currently no method that systematically combines user time-series behavioral data with course content structure to model, and through quantitative mechanisms, generate highly interpretable, clearly structured, and dynamically updated behavioral analysis outputs to meet the needs of cloud teaching platforms for real-time teaching quality monitoring, personalized recommendations, and learning strategy optimization.

[0007] Therefore, how to provide an online cloud teaching user data quantification method, system and storage medium is a problem that technical personnel in this field urgently need to solve. Summary of the Invention

[0008] One purpose of the present invention is to propose a method for quantifying user data of online cloud teaching. The present invention makes full use of technical means such as graph attention network, multi-scale graph convolution, self-supervised comparative clustering and cloud-based quantitative analysis, and describes in detail the whole process of extracting structured features from raw user behavior data, dynamically generating state labels and constructing multi-dimensional behavior quantification results. It has the advantages of high quantification accuracy, strong behavior expression ability and strong ability to adapt to complex teaching scenarios.

[0009] A method for quantifying online cloud teaching user data according to an embodiment of the present invention includes the following steps:

[0010] S1. Collect user data from the online cloud teaching platform and pre-process it;

[0011] S2. Input the preprocessed user data into the graph attention network to construct a heterogeneous time series graph with users as nodes and event sequences as directed edges. A dual-channel attention mechanism is used and integrated with a position encoding structure to generate a node embedding representation that includes time series weights and behavioral semantics.

[0012] S3. Based on node embedding representation, multi-scale graph convolution operations are used to extract behavioral pattern features at different levels, encode the user's behavioral migration relationship in the continuous learning process, and output a stage-by-stage state vector;

[0013] S4. Input the phased state vector into the self-supervised contrastive clustering network to construct positive and negative sample pairs. By maximizing the similarity between similar states and minimizing the similarity between different states, a label distribution vector is generated and the corresponding state cluster label is output.

[0014] S5. Construct a course content index based on the preprocessed user data, establish a mapping relationship between course nodes and status cluster labels, and generate a behavior quantization matrix that represents the distribution of user status labels under each course node;

[0015] S6. Input the behavior quantification matrix into the cloud-based quantitative analysis engine to generate multi-dimensional user behavior quantification results and output a structured data table containing course dimensions and user dimensions.

[0016] Optionally, the user data includes course browsing data, interactive behavior data, learning time data, resource access data and teaching progress data, and the preprocessing includes data cleaning, timestamp alignment and data standardization.

[0017] Optionally, the S2 specifically includes:

[0018] S21. Construct an event set based on the pre-processed user data, expressed as E = {e1, e2, ..., e n}, where e i ∈E represents the i-th event that occurred to the user, n represents the total number of events, and the events are arranged in the order of their occurrence;

[0019] S22. Based on the event set E, with users as nodes and the order of events as directed edges, a heterogeneous time series graph G = (V, A, T, F) is constructed, where the node set V represents the set of all users, and A represents the edge set that satisfies A = {a ij}, a ij Represents the slave node v i Points to node v j The event sequence relationship, T represents the time set, satisfying T={t ij}, t ij Represents edge a ij The associated event timestamp, F represents the behavior feature set, satisfying F={f ij}, indicating that the slave node v i Points to node v j The event behavior type code;

[0020] S23. Use graph attention mechanism to model user nodes and construct dual-channel attention weight matrix Calculated by the following formulas:

[0021]

[0022] in, represents the attention score based on behavioral semantics, represents the attention score based on the time interval, W f represents the behavioral feature weight matrix, f ij represents the behavioral feature vector, b f represents the bias term, σ represents the activation function, γ represents the time attenuation coefficient, t ik Represents edge a ik The associated event timestamp, t i(k-1) Represents edge a i(k-1) The associated event timestamp, t i(j-1) Represents edge a i(j-1) The associated event timestamp, exp(·) represents the natural exponential function with base e;

[0023] S24, will and Perform attention fusion to obtain fusion representation features. The attention fusion formula is as follows:

[0024]

[0025] Among them, b i Indicates event e i The corresponding fusion representation vector in , α and β represent the fusion weight coefficients of the behavior semantic channel and the time channel respectively, i Indicates event e i The feature encoding vector of n represents the total number of events.

[0026] S25. Introduce the position coding structure in the heterogeneous timing diagram and define the position coding vector Get all position encoding sets {p i1 ,p i2 ,...,p in}, directly concatenate the position encoding set with the fusion representation vector to obtain the node embedding vector Where ω is the frequency adjustment constant and sin(·) represents the sine function.

[0027] Optionally, the S3 specifically includes:

[0028] S31. Embed each user's node into a vector Input into the multi-scale graph convolution structure, use graph convolution kernels with different receptive fields to extract the user's feature representation at multiple behavior levels, build a scale set S, and perform graph convolution calculations on the node features at each scale;

[0029] S32, fuse the graph convolution outputs at multiple scales to obtain a unified multi-scale behavior feature representation, the fusion operation is achieved by weighted splicing, and the output node v i The behavioral pattern feature vector g i :

[0030] S33, according to the behavior pattern feature vector g i , combined with the total number of events n and the time interval sequence ΔT i ={t i2 -t i1 ,t i3 -t i2 ,...,t in -t i(n-1)}, construct the behavioral transfer sensitivity, and the calculation formula is:

[0031]

[0032] Among them, r i Represents node v i The behavioral migration sensitivity of , δ represents the time smoothing coefficient, ρ represents the normalization function, |t ij -t i(j-1) | represents the absolute value of the time interval between adjacent events, and log2(·) represents the logarithmic function;

[0033] S34, through the fusion coefficient θ, the behavior pattern feature vector g i and behavioral transfer sensitivity r i Perform weighted fusion to obtain the stage state vector s i .

[0034] Optionally, the scale set S = {s1, s2, s3}, which represent the local behavior scale, the mid-range behavior scale and the global behavior scale respectively, for each scale s k ∈S performs graph convolution calculation on the node features, and defines the graph convolution output at the kth scale as:

[0035]

[0036] in, Indicates scale s k Next node v i behavioral pattern characteristics, Indicates scale s k The adjacency weight under k Represents the graph convolution weight matrix of the corresponding scale, η k represents the scale adjustment factor, σ represents the nonlinear activation function, Represents the adjacent node v j The node embedding vector of n represents the total number of events.

[0037] Optionally, the S4 specifically includes:

[0038] S41, the phase state vector s i Input the self-supervised contrast clustering network to construct the positive sample vector pair With negative sample vector in Indicates that i Another similar state vector in the behavioral feature space, Indicates that i Another state vector that is not similar in the behavioral feature space is mapped to a clustering vector z using the projection function φ(·) i =φ(s i ),

[0039] S42. Construct a contrast loss function in the behavioral feature space, maximize the similarity for positive sample pairs, and minimize the similarity for negative sample pairs. Define the contrast loss function and perform an average operation on the event set:

[0040]

[0041] in, Represents node v i The contrastive clustering loss is τ, which represents the temperature scaling factor, ‖z i ‖, Represents cluster vector z i , The Euclidean norm of Represents the overall contrast clustering loss, n represents the total number of events, and the optimization goal is to minimize

[0042] S43, according to each cluster vector z i The spatial distribution of the cluster is Euclidean distance from the cluster center, and the nearest cluster center method is used to assign state labels. The cluster center set is C = {c1, c2, ..., c K}, where K represents the number of clusters, and the state label assignment function is defined as:

[0043]

[0044] Among them, y i Represents node v i The state cluster label, ‖z i -c k ‖ 2 Represents the cluster vector and cluster center c k ∈C, λ is the residual adjustment factor, ψ(c k) means clustering center c k The reconstructed vector after mapping to the behavioral feature space, Represents the reconstruction vector and the behavior pattern feature vector g i The squared residual.

[0045] Optionally, the S5 specifically includes:

[0046] S51. Construct a course content index set based on the pre-processed user data, expressed as L = {l1, l2, ..., l q}, each course l j ∈L corresponds to the course node index d j , each node v i The corresponding state cluster label is y i , map the status cluster labels of each user on the course node, and construct a label mapping matrix M, where the elements M ij Represents node v i In the course j The state label index under;

[0047] S52, based on the event set E, the label mapping matrix M and the corresponding course node index d j , count the frequency of each state cluster label under each course node, and construct a three-dimensional label distribution B, where B ijk Represents node v i In the course j The frequency of behaviors assigned to the k-th state cluster label;

[0048] S53. Based on the three-dimensional label distribution B, construct a behavior quantization matrix Q. The calculation formula is as follows:

[0049]

[0050] Among them, Q ij Represents node v i In the course j The behavior quantization score, ∈ is a very small constant to prevent division by zero errors, Represents the reconstruction vector ψ(c k ) and the behavioral pattern feature vector g i The squared residual of express The average residual under the k-th state cluster label, Represents the reconstruction vector ψ(c s ) and the behavioral pattern feature vector g i The squared residual of express The average residual under the s-th state cluster label, Bijk Represents node v i In the course j The frequency of behaviors assigned to the r-th state cluster label, max represents the maximum function, and K represents the number of clusters.

[0051] Optionally, the S6 specifically includes:

[0052] S61. Input the behavior quantification matrix Q into the cloud-based quantitative analysis engine, perform vector aggregation operations on the behavior quantification scores of each user in all course nodes, and generate a user dimension feature vector;

[0053] S62. Based on the user dimension feature vector, calculate the behavior mean, deviation coefficient, and status label ratio of all users at each course node to construct the course dimension feature vector;

[0054] S63. Index-align the user dimension feature vector and the course dimension feature vector to construct a structured data table, which records the user ID, course ID, behavior quantification score, state label number, cluster residual value and time index information.

[0055] An online cloud teaching user data quantification system according to an embodiment of the present invention includes:

[0056] The data collection module is used to collect user data from the online cloud teaching platform and perform preprocessing;

[0057] The heterogeneous fusion module is used to input preprocessed user data into the graph attention network to construct a heterogeneous time series graph with users as nodes and event sequences as directed edges. It uses a dual-channel attention mechanism and integrates a position encoding structure to generate a node embedding representation that contains temporal weights and behavioral semantics.

[0058] The state encoding module is used to extract behavioral pattern features at different levels, encode the user's behavioral migration relationship during the continuous learning process, and output the stage state vector;

[0059] The contrastive clustering module is used to input the stage state vector into the self-supervised contrastive clustering network, construct positive and negative sample pairs, generate label distribution vectors by maximizing the similarity between similar states and minimizing the similarity between different states, and output the corresponding state cluster labels;

[0060] The index mapping module is used to construct the course content index, establish the mapping relationship between course nodes and status cluster labels, and generate a behavior quantification matrix that represents the distribution of user status labels under each course node;

[0061] The quantitative analysis module is used to input the behavior quantitative matrix into the cloud-based quantitative analysis engine, generate multi-dimensional user behavior quantitative results, and output a structured data table containing course dimensions and user dimensions.

[0062] According to an embodiment of the present invention, a computer-readable storage medium is provided, on which a computer program is stored. When the program instructions are executed by a processor, a method for quantifying user data of online cloud teaching is implemented.

[0063] The beneficial effects of the present invention are:

[0064] First, the present invention constructs a heterogeneous time series graph with users as nodes and behavioral event time series as directed edges, and introduces a dual-channel attention mechanism and position encoding structure to achieve dynamic modeling and precise expression of users' multi-dimensional behavioral data, overcoming the problems of vague behavioral semantics and lack of temporal relationships in the existing technology, and providing a more discriminative embedding representation for subsequent state recognition and clustering analysis.

[0065] Secondly, relying on multi-scale graph convolution to extract behavioral pattern features, and combining with self-supervised comparative clustering strategies, the positive and negative sample construction and residual-driven label assignment methods are introduced, which effectively improves the discrimination and expression stability of user state labels, solves the technical bottlenecks of unclear state labels and unclear clustering boundaries in traditional methods, and significantly enhances the accuracy and generalization ability of state recognition and behavior evolution modeling.

[0066] Finally, the present invention constructs a behavior quantification matrix based on the status label distribution and course content index, and then generates a structured user behavior data table through a cloud-based quantitative analysis engine, realizing two-way quantitative modeling of user dimension and course dimension, filling the technical gap of the existing cloud teaching platform in dynamic behavior evaluation, group behavior profiling and intelligent teaching feedback, and providing an efficient and reliable data foundation for improving teaching quality and recommending personalized learning paths. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0068] Figure 1 This is a flow chart of a method for quantifying user data of online cloud teaching proposed by the present invention;

[0069] Figure 2 This is a schematic diagram of constructing a heterogeneous time series diagram for the online cloud teaching user data quantification method proposed by the present invention;

[0070] Figure 3 This is a schematic diagram of the self-supervised comparative clustering process of the online cloud teaching user data quantification method proposed in the present invention;

[0071] Figure 4 This is a module structure diagram of the online cloud teaching user data quantification system proposed by the present invention. DETAILED DESCRIPTION

[0072] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0073] refer to Figure 1-3 , a method for quantifying online cloud teaching user data, comprising the following steps:

[0074] S1. Collect user data from the online cloud teaching platform and pre-process it;

[0075] S2. Input the preprocessed user data into the graph attention network to construct a heterogeneous time series graph with users as nodes and event sequences as directed edges. A dual-channel attention mechanism is used and integrated with a position encoding structure to generate a node embedding representation that includes time series weights and behavioral semantics.

[0076] S3. Based on node embedding representation, multi-scale graph convolution operations are used to extract behavioral pattern features at different levels, encode the user's behavioral migration relationship in the continuous learning process, and output a stage-by-stage state vector;

[0077] S4. Input the phased state vector into the self-supervised contrastive clustering network to construct positive and negative sample pairs. By maximizing the similarity between similar states and minimizing the similarity between different states, a label distribution vector is generated and the corresponding state cluster label is output.

[0078] S5. Construct a course content index based on the preprocessed user data, establish a mapping relationship between course nodes and status cluster labels, and generate a behavior quantization matrix that represents the distribution of user status labels under each course node;

[0079] S6. Input the behavior quantification matrix into the cloud-based quantitative analysis engine to generate multi-dimensional user behavior quantification results and output a structured data table containing course dimensions and user dimensions.

[0080] This paper realizes deep structural modeling of online cloud teaching user behavior data by constructing heterogeneous time series graphs, introducing graph attention mechanism and multi-scale graph convolutional network, and effectively improves the accuracy of user behavior pattern recognition and the precision of quantitative expression.

[0081] In this embodiment, the user data includes course browsing data, interactive behavior data, learning time data, resource access data and teaching progress data, and the preprocessing includes data cleaning, timestamp alignment and data standardization.

[0082] This invention introduces multiple types of raw data such as course browsing, interactive behavior, and learning time, and implements data cleaning, time alignment, and standardization in the preprocessing stage, ensuring that the data foundation for subsequent graph structure modeling is more complete, unified, and computable.

[0083] In this embodiment, S2 specifically includes:

[0084] S21. Construct an event set based on the pre-processed user data, expressed as E = {e1, e2, ..., e n}, where e i ∈E represents the i-th event that occurred to the user, n represents the total number of events, and the events are arranged in the order of their occurrence;

[0085] S22. Based on the event set E, with users as nodes and the order of events as directed edges, a heterogeneous time series graph G = (V, A, T, F) is constructed, where the node set V represents the set of all users, and A represents the edge set that satisfies A = {a ij}, a ij Represents the slave node v i Points to node v j The event sequence relationship, T represents the time set, satisfying T={t ij}, t ij Represents edge a ij The associated event timestamp, F represents the behavior feature set, satisfying F={f ij}, indicating that the slave node v i Points to node v j The event behavior type code;

[0086] S23. Use graph attention mechanism to model user nodes and construct dual-channel attention weight matrix Calculated by the following formulas:

[0087]

[0088] in, represents the attention score based on behavioral semantics, represents the attention score based on the time interval, W f represents the behavioral feature weight matrix, f ij represents the behavioral feature vector, b f represents the bias term, σ represents the activation function, γ represents the time attenuation coefficient, t ik Represents edge a ik The associated event timestamp, t i(k-1) Represents edge a i(k-1) The associated event timestamp, t i(j-1) Represents edge ai(j-1) The associated event timestamp, exp(·) represents the natural exponential function with base e;

[0089] S24, will and Perform attention fusion to obtain fusion representation features. The attention fusion formula is as follows:

[0090]

[0091] Among them, b i Indicates event e i The corresponding fusion representation vector in , α and β represent the fusion weight coefficients of the behavior semantic channel and the time channel respectively, i Indicates event e i The feature encoding vector of n represents the total number of events.

[0092] S25. Introduce the position coding structure in the heterogeneous timing diagram and define the position coding vector Get all position encoding sets {p i1 ,p i2 ,...,p in}, directly concatenate the position encoding set with the fusion representation vector to obtain the node embedding vector Where ω is the frequency adjustment constant and sin(·) represents the sine function.

[0093] This paper proposes a dual-channel graph attention mechanism and integrates the position encoding method, which can simultaneously capture the semantic characteristics and temporal dynamic information of user behavior, significantly enhancing the distinguishability and temporal correlation of node embedding expression.

[0094] In this embodiment, S3 specifically includes:

[0095] S31. Embed each user's node into a vector Input into the multi-scale graph convolution structure, use graph convolution kernels with different receptive fields to extract the user's feature representation at multiple behavior levels, build a scale set S, and perform graph convolution calculations on the node features at each scale;

[0096] S32, fuse the graph convolution outputs at multiple scales to obtain a unified multi-scale behavior feature representation, the fusion operation is achieved by weighted splicing, and the output node v i The behavioral pattern feature vector g i :

[0097] S33, according to the behavior pattern feature vector g i , combined with the total number of events n and the time interval sequence ΔT i ={t i2 -ti1 ,t i3 -t i2 ,...,t in -t i(n-1)}, construct the behavioral transfer sensitivity, and the calculation formula is:

[0098]

[0099] Among them, r i Represents node v i The behavioral migration sensitivity of , δ represents the time smoothing coefficient, ρ represents the normalization function, |t ij -t i(j-1) | represents the absolute value of the time interval between adjacent events, and log2(·) represents the logarithmic function;

[0100] S34, through the fusion coefficient θ, the behavior pattern feature vector g i and behavioral transfer sensitivity r i Perform weighted fusion to obtain the stage state vector s i .

[0101] The present invention extracts feature information of users at different behavioral levels based on a multi-scale graph convolutional structure, and constructs behavioral migration sensitivity in combination with event time intervals to achieve effective modeling of behavioral change trends, thereby improving the expressive ability of behavioral sequences.

[0102] In this embodiment, the scale set S = {s1, s2, s3} represents the local behavior scale, the mid-range behavior scale and the global behavior scale respectively. For each scale s k ∈S performs graph convolution calculation on the node features, and defines the graph convolution output at the kth scale as:

[0103]

[0104] in, Indicates scale s k Next node v i behavioral pattern characteristics, Indicates scale s k The adjacency weight under k Represents the graph convolution weight matrix of the corresponding scale, η k represents the scale adjustment factor, σ represents the nonlinear activation function, Represents the adjacent node v j The node embedding vector of n represents the total number of events.

[0105] The present invention introduces a fusion mechanism of multi-scale adjustment coefficients and adjacency relationship weights in graph convolution calculations, which can improve the ability to recognize local, mid-range and global behavioral structural features, and enhance the hierarchical adaptability and expression depth of user behavior modeling.

[0106] In this embodiment, the S4 specifically includes:

[0107] S41, the phase state vector s i Input the self-supervised contrast clustering network to construct the positive sample vector pair With negative sample vector in Indicates that i Another similar state vector in the behavioral feature space, Indicates that i Another state vector that is not similar in the behavioral feature space is mapped to a clustering vector z using the projection function φ(·) i =φ(s i ),

[0108] S42. Construct a contrast loss function in the behavioral feature space, maximize the similarity for positive sample pairs, and minimize the similarity for negative sample pairs. Define the contrast loss function and perform an average operation on the event set:

[0109]

[0110] in, Represents node v i The contrastive clustering loss is τ, which represents the temperature scaling factor, ‖z i ‖, Represents cluster vector z i , The Euclidean norm of Represents the overall contrast clustering loss, n represents the total number of events, and the optimization goal is to minimize

[0111] S43, according to each cluster vector z i The spatial distribution of the cluster is Euclidean distance from the cluster center, and the nearest cluster center method is used to assign state labels. The cluster center set is C = {c1, c2, ..., c K}, where K represents the number of clusters, and the state label assignment function is defined as:

[0112]

[0113] Among them, y i Represents node v i The state cluster label, ‖z i -ck ‖ 2 Represents the cluster vector and cluster center c k ∈C, λ is the residual adjustment factor, ψ(c k ) means clustering center c k The reconstructed vector after mapping to the behavioral feature space, Represents the reconstruction vector and the behavior pattern feature vector g i The squared residual.

[0114] The present invention introduces a self-supervised contrast clustering method to construct positive and negative sample pairs and designs a contrast loss function. At the same time, it integrates behavioral feature residuals to guide label allocation, thereby improving the expression stability and representation accuracy of state clustering labels.

[0115] In this embodiment, the S5 specifically includes:

[0116] S51. Construct a course content index set based on the pre-processed user data, expressed as L = {l1, l2, ..., l q}, each course l j ∈L corresponds to the course node index d j , each node v i The corresponding state cluster label is y i , map the status cluster labels of each user on the course node, and construct a label mapping matrix M, where the elements M ij Represents node v i In the course j The state label index under;

[0117] S52, based on the event set E, the label mapping matrix M and the corresponding course node index d j , count the frequency of each state cluster label under each course node, and construct a three-dimensional label distribution B, where B ijk Represents node v i In the course j The frequency of behaviors assigned to the k-th state cluster label;

[0118] S53. Based on the three-dimensional label distribution B, construct a behavior quantization matrix Q. The calculation formula is as follows:

[0119]

[0120] Among them, Q ij Represents node v i In the course j The behavior quantization score, ∈ is a very small constant to prevent division by zero errors, Represents the reconstruction vector ψ(c k ) and the behavioral pattern feature vector gi The squared residual of express The average residual under the k-th state cluster label, Represents the reconstruction vector ψ(c s ) and the behavioral pattern feature vector g i The squared residual of express The average residual under the s-th state cluster label, B ijk Represents node v i In the course j The frequency of behaviors assigned to the r-th state cluster label, max represents the maximum function, and K represents the number of clusters.

[0121] The present invention constructs a course content index and three-dimensional label distribution, and integrates residual correction and entropy weighting strategies to achieve high-dimensional, fine-grained, multi-label quantification of user behavior, effectively enhancing the discriminability of user status under course nodes.

[0122] In this embodiment, S6 specifically includes:

[0123] S61. Input the behavior quantification matrix Q into the cloud-based quantitative analysis engine, perform vector aggregation operations on the behavior quantification scores of each user in all course nodes, and generate a user dimension feature vector;

[0124] S62. Based on the user dimension feature vector, calculate the behavior mean, deviation coefficient, and status label ratio of all users at each course node to construct the course dimension feature vector;

[0125] S63. Index-align the user dimension feature vector and the course dimension feature vector to construct a structured data table, which records the user ID, course ID, behavior quantification score, state label number, cluster residual value and time index information.

[0126] In the quantitative analysis stage, the present invention integrates two-way indicators of user dimension and course dimension to construct a structured quantitative data table, so that the behavioral assessment results can be used in various application scenarios such as teaching evaluation, learning profiling, strategy feedback, etc., with high practicality and scalability.

[0127] refer to Figure 4 , an online cloud teaching user data quantification system, comprising:

[0128] The data collection module is used to collect user data from the online cloud teaching platform and perform preprocessing;

[0129] The heterogeneous fusion module is used to input preprocessed user data into the graph attention network to construct a heterogeneous time series graph with users as nodes and event sequences as directed edges. It uses a dual-channel attention mechanism and integrates a position encoding structure to generate a node embedding representation that contains temporal weights and behavioral semantics.

[0130] The state encoding module is used to extract behavioral pattern features at different levels, encode the user's behavioral migration relationship during the continuous learning process, and output the stage state vector;

[0131] The contrastive clustering module is used to input the stage state vector into the self-supervised contrastive clustering network, construct positive and negative sample pairs, generate label distribution vectors by maximizing the similarity between similar states and minimizing the similarity between different states, and output the corresponding state cluster labels;

[0132] The index mapping module is used to construct the course content index, establish the mapping relationship between course nodes and status cluster labels, and generate a behavior quantification matrix that represents the distribution of user status labels under each course node;

[0133] The quantitative analysis module is used to input the behavior quantitative matrix into the cloud-based quantitative analysis engine, generate multi-dimensional user behavior quantitative results, and output a structured data table containing course dimensions and user dimensions.

[0134] The online cloud teaching user data quantification system proposed in this invention modularly integrates data collection, graph construction, state encoding, comparative clustering, index mapping and quantitative analysis. The system is fully functional and has automatic processing capabilities and cross-platform deployment capabilities.

[0135] A computer-readable storage medium stores a computer program, wherein when the program instructions are executed by a processor, a method for quantifying user data of online cloud teaching is implemented.

[0136] The present invention encapsulates the full-process algorithm in the form of an instruction set through a computer-readable storage medium, can achieve rapid deployment and stable operation in multiple types of processor environments, and has good versatility and platform adaptability.

[0137] Example 1:

[0138] To verify the feasibility of this invention, we applied it to remote learning on an online education platform. We comprehensively collected and processed user behavior data for a structured course on the platform to enable quantitative analysis of the user learning process. In this teaching scenario, the course structure included four teaching components: video explanations, graphic reading, staged quizzes, and interactive Q&A. The platform supported both PC and mobile access, and the user learning path exhibited strong asynchrony, diverse behavioral patterns, and a dispersed pace of task completion.

[0139] Previously, the platform primarily relied on page clicks, login frequency, and test scores to assess user activity and learning status. These results were static, subjective, and coarse-grained, making it difficult to reflect user engagement at each course node. Teachers were slow to identify students' learning difficulties, and the content recommendation system was inadequately responsive to short-term learning fluctuations, resulting in inaccurate feedback and untimely intervention. To address this issue, the present invention collects, structures, models, quantifies, and outputs user behavior data throughout the entire course learning process, fundamentally improving the data's explanatory power and decision-making guidance value.

[0140] In the experiment, the platform selected behavioral data from 2,365 users as input samples. The course content was divided into 17 main nodes and labeled as video, reading, practice, and interactive. The data collection phase covered users' click behavior, dwell time, resource access sequence, interaction frequency, and submission time interval. After data preprocessing, an event set and course index matrix were established. In the graph attention network, user behavior nodes calculate attention weights based on semantic and temporal dual channels, and are embedded in the graph after combining position encoding for representation. Subsequently, multi-scale graph convolution is used to extract local to global behavioral features, and the behavior transition sensitivity is constructed based on the event time difference, outputting a stage-by-stage state vector.

[0141] In the contrastive clustering network, positive and negative sample pairs are constructed to optimize the label generation mechanism. Ultimately, state labels are mapped to course nodes to form a behavior label tensor, and a behavior quantization matrix is ​​generated. A normalization method based on residual weighting and entropy penalty is used to obtain a user's behavior quantization score for each course node. This matrix is ​​input into a cloud-based quantization engine, generating a structured data table containing user and course dimensions for teaching evaluation, user clustering, and dynamic recommendations.

[0142] The results of quantitative analysis show that compared with the traditional evaluation mechanism, the method proposed in this invention is more sensitive in expressing behavioral differences. In course number C03 (stage test), 72.4% of users with low quantitative scores (scores below 0.25) showed behaviors with long time intervals in the middle stage of the course; while in course number C09 (picture and text reading), users with scores above 0.75 had an average single-page residence time of 135 seconds, which was significantly higher than the course average level of 91 seconds. In the state cluster distribution, a total of 5 types of user state labels were extracted, among which the state number Y2 was strongly correlated with the course dropout rate (Pearson coefficient was -0.81), providing a data basis for teaching risk warning. After using the structured results provided by this method, teachers can focus on intervening in students with learning difficulties based on user behavior type, degree of deviation and label characteristics, and adjust the allocation of teaching resources in real time through a dynamic label refresh mechanism. The following is a schematic display of some structured data results:

[0143] Table 1 Statistics of user behavior quantitative structured results

[0144]

[0145]

[0146] As can be seen from the above data, the online cloud teaching user data quantification method constructed by this invention can achieve dynamic identification, state classification, and refined measurement of learning behavior in practical applications. It has the advantages of complete structure, timely feedback, accurate judgment, and strong applicability. This example fully demonstrates the practicality and operability of this invention in complex teaching scenarios.

[0147] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for quantifying online cloud teaching user data, characterized in that: The steps include: S1. Collect user data from the online cloud teaching platform and pre-process it; S2. Input the preprocessed user data into the graph attention network to construct a heterogeneous time series graph with users as nodes and event sequences as directed edges. A dual-channel attention mechanism is used and integrated with a position encoding structure to generate a node embedding representation that includes time series weights and behavioral semantics. S3. Based on node embedding representation, multi-scale graph convolution operations are used to extract behavioral pattern features at different levels, encode the user's behavioral migration relationship in the continuous learning process, and output a stage-by-stage state vector; S4. Input the phased state vector into the self-supervised contrastive clustering network to construct positive and negative sample pairs. By maximizing the similarity between similar states and minimizing the similarity between different states, a label distribution vector is generated and the corresponding state cluster label is output. S5. Construct a course content index based on the preprocessed user data, establish a mapping relationship between course nodes and status cluster labels, and generate a behavior quantization matrix that represents the distribution of user status labels under each course node; S6. Input the behavior quantification matrix into the cloud-based quantitative analysis engine to generate multi-dimensional user behavior quantification results and output a structured data table containing course dimensions and user dimensions.

2. The method for quantifying online cloud teaching user data according to claim 1, characterized in that: The user data includes course browsing data, interactive behavior data, learning time data, resource access data and teaching progress data, and the preprocessing includes data cleaning, timestamp alignment and data standardization.

3. The method for quantifying online cloud teaching user data according to claim 1, characterized in that: The S2 specifically includes: S21. Construct an event set based on the pre-processed user data, expressed as E = {e1, e2, ..., e n }, where e i ∈E represents the i-th event that occurred to the user, n represents the total number of events, and the events are arranged in the order of their occurrence; S22. Based on the event set E, with users as nodes and the order of events as directed edges, a heterogeneous time series graph G = (V, A, T, F) is constructed, where the node set V represents the set of all users, and A represents the edge set that satisfies A = {a ij }, a ij Represents the slave node v i Points to node v j The event sequence relationship, T represents the time set, satisfying T={t ij }, t ij Represents edge a ij The associated event timestamp, F represents the behavior feature set, satisfying F={f ij }, indicating that the slave node v i Points to node v j The event behavior type code; S23. Use graph attention mechanism to model user nodes and construct dual-channel attention weight matrix Calculated by the following formulas: in, represents the attention score based on behavioral semantics, represents the attention score based on the time interval, W f represents the behavioral feature weight matrix, f ij represents the behavioral feature vector, b f represents the bias term, σ represents the activation function, γ represents the time attenuation coefficient, t ik Represents edge a ik The associated event timestamp, t i(k-1) Represents edge a i(k-1) The associated event timestamp, t i(j-1) Represents edge a i(j-1) The associated event timestamp, exp(·) represents the natural exponential function with base e; S24, will and Perform attention fusion to obtain fusion representation features. The attention fusion formula is as follows: Among them, b i Indicates event e i The corresponding fusion representation vector in , α and β represent the fusion weight coefficients of the behavior semantic channel and the time channel respectively, i Indicates event e i The feature encoding vector of n represents the total number of events. S25. Introduce the position coding structure in the heterogeneous timing diagram and define the position coding vector Get all position encoding sets {p i1 ,p i2 ,...,p in }, directly concatenate the position encoding set with the fusion representation vector to obtain the node embedding vector Where ω is the frequency adjustment constant and sin(·) represents the sine function.

4. The method for quantifying online cloud teaching user data according to claim 1, characterized in that: The S3 specifically includes: S31. Embed each user's node into a vector Input into the multi-scale graph convolution structure, use graph convolution kernels with different receptive fields to extract the user's feature representation at multiple behavior levels, build a scale set S, and perform graph convolution calculations on the node features at each scale; S32, fuse the graph convolution outputs at multiple scales to obtain a unified multi-scale behavior feature representation, the fusion operation is achieved by weighted splicing, and the output node v i The behavioral pattern feature vector g i : S33, according to the behavior pattern feature vector g i , combined with the total number of events n and the time interval sequence ΔT i ={t i2 -t i1 ,t i3 -t i2 ,...,t in -t i(n-1) }, construct the behavioral transfer sensitivity, and the calculation formula is: Among them, r i Represents node v i The behavioral migration sensitivity of , δ represents the time smoothing coefficient, ρ represents the normalization function, |t ij -t i(j-1) | represents the absolute value of the time interval between adjacent events, and log2(·) represents the logarithmic function; S34, through the fusion coefficient θ, the behavior pattern feature vector g i and behavioral transfer sensitivity r i Perform weighted fusion to obtain the stage state vector s i .

5. The method for quantifying online cloud teaching user data according to claim 4, characterized in that: The scale set S = {s1, s2, s3} represents the local behavior scale, the mid-range behavior scale and the global behavior scale respectively. k ∈S performs graph convolution calculation on the node features, and defines the graph convolution output at the kth scale as: in, Indicates scale s k Next node v i behavioral pattern characteristics, Indicates scale s k The adjacency weight under k Represents the graph convolution weight matrix of the corresponding scale, η k represents the scale adjustment factor, σ represents the nonlinear activation function, Represents the adjacent node v j The node embedding vector of n represents the total number of events.

6. The method for quantifying online cloud teaching user data according to claim 1, characterized in that: The S4 specifically includes: S41, the phase state vector s i Input the self-supervised contrast clustering network to construct the positive sample vector pair With negative sample vector in Indicates that i Another similar state vector in the behavioral feature space, Indicates that i Another state vector that is not similar in the behavioral feature space is mapped to a clustering vector z using the projection function φ(·) i =φ(s i ), S42. Construct a contrast loss function in the behavioral feature space, maximize the similarity for positive sample pairs, and minimize the similarity for negative sample pairs. Define the contrast loss function and perform an average operation on the event set: in, Represents node v i The contrastive clustering loss is τ, which represents the temperature scaling factor, ‖z i ‖, Represents cluster vector z i , The Euclidean norm of Represents the overall contrast clustering loss, n represents the total number of events, and the optimization goal is to minimize S43, according to each cluster vector z i The spatial distribution of the cluster is Euclidean distance from the cluster center, and the nearest cluster center method is used to assign state labels. The cluster center set is C = {c1, c2, ..., c K }, where K represents the number of clusters, and the state label assignment function is defined as: Among them, y i Represents node v i The state cluster label, ‖z i -c k ‖ 2 Represents the cluster vector and cluster center c k ∈C, λ is the residual adjustment factor, ψ(c k ) means clustering center c k The reconstructed vector after mapping to the behavioral feature space, Represents the reconstruction vector and the behavior pattern feature vector g i The squared residual.

7. The method for quantifying online cloud teaching user data according to claim 1, characterized in that: The S5 specifically includes: S51. Construct a course content index set based on the pre-processed user data, expressed as L = {l1, l2, ..., l q }, each course l j ∈L corresponds to the course node index d j , each node v i The corresponding state cluster label is y i , map the status cluster labels of each user on the course node, and construct a label mapping matrix M, where the elements M ij Represents node v i In the course j The state label index under; S52, based on the event set E, the label mapping matrix M and the corresponding course node index d j , count the frequency of each state cluster label under each course node, and construct a three-dimensional label distribution B, where B ijk Represents node v i In the course j The frequency of behaviors assigned to the k-th state cluster label; S53. Based on the three-dimensional label distribution B, construct a behavior quantization matrix Q. The calculation formula is as follows: Among them, Q ij Represents node v i In the course j The behavior quantization score, ∈ is a very small constant to prevent division by zero errors, Represents the reconstruction vector ψ(c k ) and the behavioral pattern feature vector g i The squared residual of express The average residual under the k-th state cluster label, Represents the reconstruction vector ψ(c s ) and the behavioral pattern feature vector g i The squared residual of express The average residual under the s-th state cluster label, B ijk Represents node v i In the course j The frequency of behaviors assigned to the r-th state cluster label, max represents the maximum function, and K represents the number of clusters.

8. The method for quantifying online cloud teaching user data according to claim 1, characterized in that: The S6 specifically includes: S61. Input the behavior quantification matrix Q into the cloud-based quantitative analysis engine, perform vector aggregation operations on the behavior quantification scores of each user in all course nodes, and generate a user dimension feature vector; S62. Based on the user dimension feature vector, calculate the behavior mean, deviation coefficient, and status label ratio of all users at each course node to construct the course dimension feature vector; S63. Index-align the user dimension feature vector and the course dimension feature vector to construct a structured data table, which records the user ID, course ID, behavior quantification score, state label number, cluster residual value and time index information.

9. An online cloud teaching user data quantification system, executing an online cloud teaching user data quantification method according to any one of claims 1 to 8, characterized in that: include: The data collection module is used to collect user data from the online cloud teaching platform and perform preprocessing; The heterogeneous fusion module is used to input preprocessed user data into the graph attention network to construct a heterogeneous time series graph with users as nodes and event sequences as directed edges. It uses a dual-channel attention mechanism and integrates a position encoding structure to generate a node embedding representation that contains temporal weights and behavioral semantics. The state encoding module is used to extract behavioral pattern features at different levels, encode the user's behavioral migration relationship during the continuous learning process, and output the stage state vector; The contrastive clustering module is used to input the stage state vector into the self-supervised contrastive clustering network, construct positive and negative sample pairs, generate label distribution vectors by maximizing the similarity between similar states and minimizing the similarity between different states, and output the corresponding state cluster labels; The index mapping module is used to construct the course content index, establish the mapping relationship between course nodes and status cluster labels, and generate a behavior quantification matrix that represents the distribution of user status labels under each course node; The quantitative analysis module is used to input the behavior quantitative matrix into the cloud-based quantitative analysis engine, generate multi-dimensional user behavior quantitative results, and output a structured data table containing course dimensions and user dimensions.

10. A computer-readable storage medium having program instructions stored thereon, characterized in that: When the program instructions are executed by a processor, the method according to any one of claims 1 to 8 is implemented.

Images