A learning ability evaluation model construction method for personalized learning of an online learning platform

Abstract

The application belongs to the technical field of data processing, and provides a learning ability evaluation model construction method for individual learning of an online learning platform, which comprises the following steps: firstly, collecting behavior data and learning resource metadata of users in a learning process; constructing an individual behavior heterogeneous graph, a knowledge system hypergraph and a knowledge point sequential graph based on the collected data; then, presetting a learning scene attention matrix, combining the knowledge point sequential graph and determining a context perception mechanism; finally, constructing a learning ability evaluation model according to the context perception mechanism, the individual behavior heterogeneous graph and the knowledge system hypergraph. The application realizes accurate modeling of the complex interaction relationship among the user, the learning resource and the knowledge point by constructing the individual behavior heterogeneous graph, the knowledge system hypergraph and the knowledge point sequential graph, deeply fusing the behavior data generated by the user in the learning process with the metadata such as the type, the difficulty and the associated knowledge point of the learning resource, fully showing the systematicness and the dynamics of the knowledge system, and effectively improving the accuracy of the learning ability evaluation.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, and in particular relates to a method for constructing a learning ability assessment model for personalized learning on an online learning platform. Background Technology

[0002] As the wave of digital education sweeps the globe, online learning platforms have become a crucial channel for hundreds of millions of learners to acquire knowledge. Significant differences exist among users in terms of learning progress, knowledge base, and cognitive styles, highlighting the growing demand for personalized learning. To meet this need, learning ability assessment models have emerged, aiming to accurately determine learners' knowledge mastery and learning abilities by analyzing their behavioral data, thereby providing tailored learning solutions. However, current technological development still faces numerous bottlenecks. Most existing learning ability assessment models rely on single-modal data, failing to capture the rich information generated during the learning process; in feature processing, they employ fixed weight allocation methods, making it difficult to flexibly adapt to the personalized learning patterns of different learners; and in knowledge relationship modeling, they are often limited to simple bipartite graph structures, insufficiently characterizing the complexity of knowledge systems and lacking awareness of learning scenarios and higher-order dependencies of knowledge points. These problems result in the failure to fully exploit the value of data, making it difficult for models to cope with diverse learning scenarios, and the presentation of knowledge systems remains superficial, failing to accurately grasp the intricate connections between learners' behavior and knowledge structures. Summary of the Invention

[0003] In view of this, the present invention provides a method for constructing a learning ability assessment model for personalized learning on an online learning platform, aiming to solve the problem that existing learning ability assessment models are inaccurate in calculating the complex relationship between learners' behavior and knowledge structure.

[0004] A first aspect of this invention provides a method for constructing a learning ability assessment model for personalized learning on an online learning platform, comprising:

[0005] Collect user behavior data and learning resource metadata during the learning process;

[0006] Based on behavioral data and learning resource metadata, construct an individual behavior heterogeneous graph, a knowledge system hypergraph, and a knowledge point sequence graph;

[0007] A learning scenario attention matrix is ​​preset, and a context awareness mechanism is determined based on the learning scenario attention matrix and the knowledge point sequence diagram;

[0008] A learning ability assessment model is constructed based on context awareness mechanisms, heterogeneous graphs of individual behaviors, and hypergraphs of knowledge systems.

[0009] In one possible implementation, based on behavioral data and learning resource metadata, an individual behavior heterogeneity graph, a knowledge system hypergraph, and a knowledge point sequence graph are constructed, including:

[0010] Based on the learning resource metadata, determine the heterogeneous graph nodes, hypergraph nodes, and logical relationships between knowledge points;

[0011] Based on behavioral data, determine the edges of the heterogeneous graph and construct an individual behavior heterogeneous graph by combining the nodes of the heterogeneous graph;

[0012] The edges of the hypergraph are determined based on the logical relationships between knowledge points, and the knowledge system hypergraph is constructed by combining the hypergraph nodes.

[0013] Based on the knowledge system hypergraph and the logical relationships between knowledge points, a knowledge point sequence graph is created.

[0014] In one possible implementation, the edges of the heterogeneous graph are determined based on the behavioral data, and the individual behavioral heterogeneous graph is constructed by combining the nodes of the heterogeneous graph, including:

[0015] Define user-learning resource edges, user-knowledge point edges, and learning resource-knowledge point edges;

[0016] Set a first-order time decay function next to the user-learning resource side, and set an exponential decay function next to the user-knowledge point side;

[0017] Connect the user-learning resource edge, user-knowledge point edge, and learning resource-knowledge point edge to the corresponding nodes of the heterogeneous graph to obtain the individual behavior heterogeneous graph.

[0018] In one possible implementation, the edges of the hypergraph are determined based on the logical relationships between knowledge points, and the knowledge system hypergraph is constructed by combining the hypergraph nodes, including:

[0019] Based on the logical relationships between knowledge points, determine the hierarchical inclusion relationship, horizontal parallel relationship, and vertical progressive relationship of knowledge points;

[0020] Based on hierarchical inclusion, horizontal parallel, and vertical progressive relationships, the knowledge point matching score is calculated and used as the edge attribute value of the hypergraph.

[0021] Connect the edges of the hypergraph to the corresponding nodes of the hypergraph to obtain the knowledge system hypergraph.

[0022] In one possible implementation, a knowledge point sequence graph is established based on the knowledge system hypergraph and the logical relationships between knowledge points, including:

[0023] Extract directed hyperedges representing vertical progression from the knowledge system hypergraph, and filter the extracted directed hyperedges based on the logical relationships between knowledge points;

[0024] The selected directed superedges are converted into directed edges in the knowledge point sequence graph, where the direction of the directed edges is consistent with the direction of the superedges, and the edge weights retain the dependency strength, thus obtaining the knowledge point sequence graph.

[0025] In one possible implementation, the context-aware mechanism is determined based on the learning scenario attention matrix and the knowledge point sequence graph, including:

[0026] Based on the learning scenario attention matrix, determine the attention weight of each knowledge point in the knowledge point sequence graph under different scenarios;

[0027] The attention weight of each knowledge point in different scenarios is added as a node attribute to the knowledge point sequence graph to form a context-aware mechanism.

[0028] In one possible implementation, a pre-defined attention matrix for the learning scenario is included, comprising:

[0029] Multiple three-dimensional feature vectors are set; among them, the three-dimensional feature vectors include user state features, resource attribute features, and time environment features; each three-dimensional feature vector is used for a type of learning scenario;

[0030] A learning scene attention matrix is ​​formed based on multiple three-dimensional feature vectors.

[0031] In one possible implementation, a learning ability assessment model is constructed based on context-aware mechanisms, heterogeneous graphs of individual behaviors, and knowledge system hypergraphs, including:

[0032] By using the context-aware mechanism as the scene sub-network, the individual behavior heterogeneous graph as the behavior sub-network, and the knowledge system hypergraph as the knowledge sub-network, a cross-view graph neural network is constructed as a learning ability evaluation model.

[0033] In one possible implementation, a cross-view graph neural network is constructed using a context-aware mechanism as a scene sub-network, an individual behavior heterogeneous graph as a behavior sub-network, and a knowledge system hypergraph as a knowledge sub-network, serving as a learning ability evaluation model, including:

[0034] The learning ability evaluation model is obtained by using the context awareness mechanism as the scene sub-network, the individual behavior heterogeneous graph as the behavior sub-network, and the knowledge system hypergraph as the knowledge sub-network, while using the contrast loss function based on positive and negative sample pairs as the total network loss. The three-view embeddings are weighted and fused through the self-attention mechanism.

[0035] In one possible implementation, user behavior data and learning resource metadata are collected during the learning process, including:

[0036] Collect raw data from online learning platforms;

[0037] The raw data is preprocessed to obtain behavioral data and learning resource metadata.

[0038] The method for constructing a personalized learning ability assessment model for an online learning platform provided in this invention first collects user behavior data and learning resource metadata during the learning process; then, based on the behavior data and learning resource metadata, it constructs an individual behavior heterogeneity graph, a knowledge system hypergraph, and a knowledge point sequence graph; next, it presets a learning scenario attention matrix and determines a context-aware mechanism based on the learning scenario attention matrix and the knowledge point sequence graph; finally, it constructs a learning ability assessment model based on the context-aware mechanism, the individual behavior heterogeneity graph, and the knowledge system hypergraph. This invention, by constructing an individual behavior heterogeneity graph, a knowledge system hypergraph, and a knowledge point sequence graph, deeply integrates the user's behavior data generated during the learning process with metadata such as the type, difficulty, and related knowledge points of learning resources. This achieves accurate modeling of the complex interactive relationships between "user-resource-knowledge points," comprehensively showcasing the systematic and dynamic nature of the knowledge system, and effectively improving the accuracy of learning ability assessment. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the 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.

[0040] Figure 1 This is an application scenario diagram of the learning ability assessment model construction method for personalized learning on an online learning platform provided in this embodiment of the invention;

[0041] Figure 2 This is a flowchart illustrating the implementation of the learning ability assessment model construction method for personalized learning on an online learning platform provided in this embodiment of the invention. Detailed Implementation

[0042] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of the invention. However, those skilled in the art will understand that the invention can be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of the invention with unnecessary detail.

[0043] Figure 1 This is an application scenario diagram of the learning ability assessment model construction method for personalized learning on an online learning platform provided in this embodiment of the invention. For example... Figure 1 As shown, the scenario includes a server 11 and at least one terminal 12.

[0044] Server 11 is equipped with an online learning platform. Students access server 11 through terminal 12 to learn. Server 11 records students' learning content and behavior in real time, and builds a personalized learning ability assessment model for each student based on the recorded content and the server's knowledge base. This allows for a more dynamic assessment of each student's learning ability. Therefore, it can provide more accurate assessment content for each student, enabling the development of better learning plans or more suitable course recommendations.

[0045] Figure 2 This is a flowchart illustrating the implementation of a learning ability assessment model construction method for personalized learning on an online learning platform provided in this embodiment of the invention. Figure 2 As shown, in some embodiments, the method for constructing a learning ability assessment model for personalized learning on an online learning platform includes:

[0046] S210 collects user behavior data and learning resource metadata during the learning process;

[0047] In this embodiment of the invention, behavioral data refers to dynamic interactive data generated by the user during the learning process, specifically including learning progress data, learning duration, interaction records, operational behaviors, evaluation feedback, etc.

[0048] Learning resource metadata refers to static data that describes the inherent attributes and structure of learning resources, including:

[0049] Basic attributes: resource type (e.g., courses, ebooks, audio, journals, AR / VR content), format (e.g., Epub files, panoramic videos), difficulty level, and applicable knowledge areas.

[0050] Knowledge point association: The knowledge points contained in the resource and their logical relationships with other resources (such as hierarchical inclusion, horizontal parallelism, and vertical progression; see the description of knowledge system hypergraph construction).

[0051] Metadata standards: resource ID, title, author, publication time, category tags (such as system-preset course templates and topic resource tags).

[0052] Structured information: the three-level directory structure of the course outline, the directory hierarchy of the e-book, and the triggering conditions for AR resources (such as 2D / 3D object scanning and recognition rules).

[0053] Evaluate relevant attributes: the matching score of the knowledge points corresponding to the resources, and the advancement conditions in the knowledge advancement map (such as stage test papers and ability analysis reports).

[0054] In some embodiments, collecting user behavior data and learning resource metadata during the learning process includes: collecting raw data from the online learning platform; and preprocessing the raw data to obtain behavior data and learning resource metadata.

[0055] In this embodiment of the invention, the collected data is first processed for missing values, outlier filtering, and data deduplication. Then, key features are extracted from the original data and converted into structured data that can be used for model building. Different methods can be used to structure data from different modalities and sources, and no specific method is limited here.

[0056] S220 constructs an individual behavior heterogeneous graph, a knowledge system hypergraph, and a knowledge point sequence graph based on behavioral data and learning resource metadata.

[0057] S230: Pre-set the learning scenario attention matrix, and determine the context awareness mechanism based on the learning scenario attention matrix and the knowledge point sequence diagram;

[0058] S240 constructs a learning ability assessment model based on context-aware mechanisms, heterogeneous graphs of individual behaviors, and hypergraphs of knowledge systems.

[0059] In this embodiment of the invention, by constructing an individual behavior heterogeneous graph, a knowledge system hypergraph, and a knowledge point sequential graph, the behavioral data generated by users during the learning process is deeply integrated with metadata such as the type, difficulty, and related knowledge points of learning resources. This achieves accurate modeling of the complex interactive relationship between "user-resource-knowledge point", fully demonstrates the systematic and dynamic nature of the knowledge system, and effectively improves the accuracy of learning ability assessment.

[0060] In some embodiments, constructing an individual behavior heterogeneous graph, a knowledge system hypergraph, and a knowledge point sequence graph based on behavioral data and learning resource metadata includes: determining heterogeneous graph nodes, hypergraph nodes, and logical relationships of knowledge points based on learning resource metadata; determining the edges of the heterogeneous graph based on behavioral data and constructing the individual behavior heterogeneous graph by combining the heterogeneous graph nodes; determining the edges of the hypergraph based on the logical relationships of knowledge points and constructing the knowledge system hypergraph by combining the hypergraph nodes; and establishing a knowledge point sequence graph based on the knowledge system hypergraph and the logical relationships of knowledge points.

[0061] In this embodiment of the invention, when determining heterogeneous graph nodes based on learning resource metadata, three types of nodes are explicitly included: user nodes, which correspond to various types of users on the platform (such as students, teachers, and administrators), and their attributes can be associated with user tags, learning ability assessment results, etc.; learning resource nodes cover multiple types such as courses, e-books, audio, journals, AR / VR resources, and need to be labeled with resource ID, type, difficulty, tags, and other metadata; knowledge point nodes are the smallest knowledge units extracted from resource metadata, such as the third-level directory knowledge points of a course outline, and need to record the knowledge point ID, the domain to which it belongs, and the associated resources.

[0062] The hypergraph nodes are built around knowledge points, supplemented by resource nodes. Knowledge point nodes need further annotation of hierarchical attributes (e.g., beginner / intermediate / advanced) and logical relationship types (hierarchical inclusion / horizontal parallel / vertical progression). For example, in the "Knowledge Advancement Map Subsystem," resource packages from different knowledge domains need to be organized according to knowledge point hierarchy. Hypergraph nodes can be directly mapped to stage knowledge points in the knowledge advancement map, and their attributes need to match advancement conditions (e.g., "Complete the preceding knowledge points to unlock subsequent content"). Resource nodes, as carriers of knowledge points, are associated with corresponding knowledge point nodes, strengthening the mapping relationship between the knowledge system and resources.

[0063] The logical relationships between knowledge points are extracted from the structured information in the resource metadata, encompassing three core types: hierarchical inclusion relationships, reflecting the subordinate structure of knowledge points (e.g., "Mathematics" includes "Algebra" and "Geometry"); horizontal parallel relationships, indicating the parallel association of knowledge points at the same level (e.g., "Python" and "Java" under "Programming Languages"); and vertical progressive relationships, representing the dependency between knowledge points (e.g., "Basic Calculus" → "Multivariable Calculus"). The modeling of these relationships is based on the syllabus's directory structure, the progression rules of the knowledge advancement map, and the knowledge point association annotations in the resource metadata.

[0064] In some embodiments, the edges of the heterogeneous graph are determined based on behavioral data, and an individual behavior heterogeneous graph is constructed by combining the nodes of the heterogeneous graph. This includes: defining user-learning resource edges, user-knowledge point edges, and learning resource-knowledge point edges; setting a first-order time decay function on the user-learning resource edges and an exponential decay function on the user-knowledge point edges; and connecting the user-learning resource edges, user-knowledge point edges, and learning resource-knowledge point edges to the corresponding nodes of the heterogeneous graph to obtain the individual behavior heterogeneous graph.

[0065] In this embodiment of the invention, the user-learning resource edge represents a user's direct operation on learning resources, such as accessing, downloading, favorited, or caching resources. Specifically, this can be obtained from user operation records in the behavior log, such as "User ID + Resource ID + Operation Time + Operation Type (Browse / Download, etc.)".

[0066] In this embodiment of the invention, the user-knowledge point edge reflects the interaction between the user and the knowledge point, such as the knowledge point learning time, the accuracy rate of answering questions, and the collection of incorrect questions. Specifically, it is determined by the evaluation data during the learning process, such as "user ID + knowledge point ID + mastery score + answering time".

[0067] In this embodiment of the invention, the learning resource-knowledge point edge characterizes the static association between learning resources and knowledge points, such as the list of knowledge points included in a course or the knowledge points corresponding to chapters in an e-book. Specifically, this is determined by structured information in the resource metadata, such as the knowledge points corresponding to the three-level directory of the course outline.

[0068] In this embodiment of the invention, the first-order time decay function is:

[0069]

[0070] Among them, w u-r Let λ be the edge weight, t be the operation time, T be the current time, and λ be the decay coefficient. The weight decays linearly over time, emphasizing that recent user actions on resources better reflect the current learning status. For example, a resource downloaded by a user one day ago has a weight of 0.9, which decays to 0.5 after 7 days.

[0071] In this embodiment of the invention, the exponential decay function is:

[0072] w u-k (t)=e -λ(T-t)

[0073] Among them, w u-k The edge weights decay exponentially over time, reinforcing the influence of short-term memory on the mastery of knowledge points. For example, a knowledge point mastered by a user one day ago has a weight of 0.9, which decays to 0.5 (λ = 0.3) after 7 days, reflecting the forgetting pattern more rapidly than first-order decay.

[0074] Whenever a user generates a new action, the decay weight of the corresponding edge is recalculated to ensure that the graph structure reflects the user's latest learning dynamics in real time, thereby guaranteeing the continuous updating of the cross-view graph neural network.

[0075] In some embodiments, the edges of the hypergraph are determined according to the logical relationships of knowledge points, and the knowledge system hypergraph is constructed by combining the hypergraph nodes. This includes: determining the hierarchical inclusion relationship, horizontal parallel relationship, and vertical progressive relationship of knowledge points according to the logical relationships of knowledge points; calculating the knowledge point matching degree score according to the hierarchical inclusion relationship, horizontal parallel relationship, and vertical progressive relationship, and using it as the edge attribute value of the hypergraph; and connecting the edges of the hypergraph to each hypergraph node to obtain the knowledge system hypergraph.

[0076] In this embodiment of the invention, three core types can be determined based on the logical relationships between knowledge points:

[0077] Hierarchical inclusion relationship: This reflects the subordinate structure of knowledge points, such as "mathematics" including "algebra" and "geometry".

[0078] Horizontal parallel relationship: refers to the parallel association of knowledge points at the same level, such as "Python" and "Java" under "programming languages". Its modeling basis includes the parallel design of peer labels in the resource classification system and course modules.

[0079] Vertical progression relationship: indicates the sequential dependence of knowledge points, such as "basic calculus" → "multivariable calculus", which corresponds to the advancement condition in the knowledge advancement map that "completing the preceding knowledge points is required to unlock the subsequent content" and the "vertical progression directed superedge" mentioned.

[0080] Based on three types of logical relationships, a knowledge point matching score needs to be calculated as a hypergraph edge attribute value: quantifying the degree of logical connection between knowledge points (values ​​[0,1]). For example, the score for the hierarchical inclusion relationship between "Algebra" and "Mathematics" can be set to 0.9, the score for the horizontal parallel relationship between "Python" and "Java" can be set to 0.7, and the score for the vertical progression relationship between "Basic Calculus" and "Multivariable Calculus" can be set to 0.95. The score is determined through methods such as the depth of the course syllabus hierarchy, the co-occurrence frequency of user learning behavior, or expert annotations to ensure that the score reflects the true strength of the logical connection.

[0081] When constructing a knowledge system hypergraph, knowledge points are used as core nodes, supplemented by resource nodes as knowledge carriers:

[0082] Node attributes: Knowledge point node labels include ID, domain, level, and logical relationship type; resource nodes are associated with the corresponding knowledge points.

[0083] Edge connection rules:

[0084] The hierarchy consists of parent knowledge points connected to child knowledge points via directed hyperedges (e.g., "Mathematics" → {"Algebra" "Geometry"}), with the edge weight being the matching score.

[0085] Horizontal parallel arrangement: Knowledge points at the same level are connected by undirected hyperedges or bidirectional directed edges (e.g., The edge weights are the matching scores.

[0086] Vertical progression: Prerequisite knowledge points are directed to subsequent knowledge points via directed hyperedges (e.g., "Basic Calculus" → "Multivariable Calculus"), with edge weights representing the dependency strength score.

[0087] By modeling "one-to-many" and "many-to-many" relationships using higher-order edges, the systematic and dynamic nature of the knowledge system is fully demonstrated, providing a global modeling foundation at the knowledge level for learning ability assessment.

[0088] In some embodiments, a knowledge point sequence graph is established based on the knowledge system hypergraph and the logical relationships between knowledge points, including: extracting directed superedges with vertical progression from the knowledge system hypergraph, and filtering the extracted directed superedges according to the logical relationships between knowledge points; converting the filtered directed superedges into directed edges of the knowledge point sequence graph, wherein the direction of the directed edges is consistent with the direction of the superedges, and the edge weight retains the dependency strength, thus obtaining the knowledge point sequence graph.

[0089] In this embodiment of the invention, when extracting directed hyperedges representing vertical progression from a knowledge system hypergraph, the focus should be on the sequential relationships of dependency between knowledge points. Such directed hyperedges represent the logic of "prerequisite knowledge points must be mastered before subsequent knowledge points can be learned" (e.g., "calculus basics → multivariable calculus"), and should exclude hierarchical inclusion (e.g., "mathematics → algebra") and horizontal parallelism (e.g., ...). Edges representing relationships. The extraction is based on directed hyperedges of the "vertical progression" type already marked in the knowledge system hypergraph, and their matching score reflects the strength of knowledge point dependencies.

[0090] Filtering directed hyperedges must adhere to two conditions: first, logical relationship conformity, retaining only edges explicitly marked as "vertically progressive"; second, dependency strength threshold, setting a matching score ≥ 0.7 (or other reasonable threshold) to filter weakly related edges. For example, in the hypergraph, "Basic Calculus → Multivariable Calculus" (score 0.95) meets the condition and can be retained, while "Algebra → Geometry" (horizontally parallel, score 0.6) is filtered out. Through this filtering, the knowledge point sequence graph is ensured to contain only strong dependencies that can guide the learning order, providing accurate data support for learning path planning.

[0091] When converting the selected directed hyperedges into directed edges of the knowledge point sequence graph, the principles of consistent direction and weight preservation must be followed: the direction of the hyperedge is directly mapped to the direction of the directed edge in the sequence graph (e.g., "knowledge point A → knowledge point B"), and its matching score is used as the edge weight to represent the dependency strength. The constructed knowledge point sequence graph is a directed acyclic graph (DAG), which supports the generation of linear learning paths through topological sorting (e.g., "knowledge point A → knowledge point B → knowledge point C") and can be applied to two main scenarios: one is learning plan navigation, which automatically generates personalized learning paths including stage assessments based on the sequence graph; the other is a knowledge advancement map, which uses the sequence of knowledge points in the sequence graph as advancement conditions, such as "completing the prerequisite knowledge points and achieving the required score to unlock subsequent content".

[0092] In some embodiments, determining a context-aware mechanism based on the learning scenario attention matrix and the knowledge point sequence graph includes: determining the attention weight of each knowledge point in the knowledge point sequence graph under different scenarios based on the learning scenario attention matrix; adding the attention weight of each knowledge point under different scenarios as a node attribute to the knowledge point sequence graph to form a context-aware mechanism.

[0093] In this embodiment of the invention, the construction of the learning scenario attention matrix is ​​based on three-dimensional feature vectors, with each vector corresponding to a learning scenario. These three-dimensional feature vectors include user status features (such as learning progress, points, and learning ability assessment level), resource attribute features (such as resource type, difficulty, and knowledge point association), and time environment features (such as learning time period and duration). By pre-setting multiple such three-dimensional feature vectors, a learning scenario attention matrix can be formed to quantify the differences in the importance of knowledge points in different scenarios. For example, for the scenario of "novice user + basic resources + nighttime learning," the matrix can assign higher attention weights to basic knowledge points, reflecting the user's more urgent need for basic knowledge in this scenario.

[0094] Based on the learning scenario attention matrix, it is necessary to calculate the attention weight of each knowledge point in the knowledge point sequence graph under different scenarios. Specifically, by mapping the matrix to the node features of the knowledge point sequence graph, the weight value of each knowledge point in the combined scenario of "user state-resource attribute-time environment" can be determined. For example, the knowledge point "variable definition" has a weight of 0.8 in the scenario of "novice user + Python course + weekday daytime", but only 0.3 in the scenario of "advanced user + Java framework + weekend night". This weight calculation process ensures that the model can perceive the impact of the user's current scenario on the importance of the knowledge point.

[0095] By adding the scene attention weights of each knowledge point as node attributes to the knowledge point sequence graph, a context-aware mechanism is formed. At this point, each node in the knowledge point sequence graph not only contains the logical relationships and dependency strengths of the knowledge points themselves, but also carries scene-sensitive attention information. For example, in user learning path planning, the system can automatically increase the weight of corresponding knowledge points based on the current scene, prioritizing the recommendation of relevant resources or evaluation content. This mechanism, by dynamically adjusting the priority of knowledge points through scene integration, achieves the goal of enhancing the model's adaptability to diverse learning scenarios, providing crucial input for subsequent modeling of cross-view graph neural networks (scene sub-networks).

[0096] In some embodiments, the preset learning scenario attention matrix includes: setting multiple three-dimensional feature vectors; wherein the three-dimensional feature vectors include user state features, resource attribute features and time environment features; each three-dimensional feature vector is used for a type of learning scenario; and the learning scenario attention matrix is ​​formed based on the multiple three-dimensional feature vectors.

[0097] In this embodiment of the invention, the three-dimensional feature vector is constructed based on the core influencing factors of the learning scenario, and includes three types of key features:

[0098] User status characteristics: These include the user's learning progress on the platform (such as course completion rate, accumulated points in the points system, learning ability assessment results (such as level, scores in each dimension) and user tags (such as novice / advanced user).

[0099] Resource attribute characteristics include the type of learning resource (such as courses, e-books, AR content), difficulty level (beginner / intermediate / advanced), knowledge point association (such as the knowledge domain and prerequisite knowledge points), and tag classification (such as "high-frequency test points" and "thematic resources").

[0100] Temporal and environmental characteristics: These include the time period during which learning behavior occurs (e.g., weekday daytime / weekend nighttime), duration (e.g., duration of a single learning session), and periodicity (e.g., pre-exam sprint stage, daily learning).

[0101] The mapping between 3D feature vectors and learning scenarios is achieved by combining three types of features, with each vector uniquely identifying a learning scenario:

[0102] Scenario 1: New users + introductory courses + fragmented time

[0103] Vector example: [User level = Novice, Resource difficulty = Beginner, Time = Weekday daytime (≤30 minutes)], suitable for scenarios where users use fragmented time to learn basic knowledge points.

[0104] Scenario 2: Advanced users + complex courses + blocks of time

[0105] Vector example: [User level = Advanced, Resource type = Professional course, Time = Weekend afternoon (≥2 hours)], corresponding to scenarios that require deep learning of complex knowledge points.

[0106] Scenario 3: Test-taking users + high-frequency test point resources + pre-exam stage

[0107] Vector example: [Learning goal = exam preparation, resource tag = high-frequency test points, time = 1 week before the exam], matching the scenario requirements of intensive review before the exam.

[0108] The learning scenario attention matrix is ​​constructed using multiple three-dimensional feature vectors as row vectors to form a multi-dimensional matrix. Each row of the matrix corresponds to a learning scenario, and the columns represent feature values ​​or weight parameters of user state, resource attributes, and time environment. This matrix, through preset feature combinations, provides a basis for scenario classification for subsequent context-aware mechanisms. This allows the model to extract corresponding feature vectors from the matrix based on the user's current scenario, dynamically adjusting the attention weights of each node in the knowledge point sequence graph, thus achieving the goal of "differentiated evaluation of the importance of knowledge points in different scenarios."

[0109] In some embodiments, a learning ability assessment model is constructed based on a context-aware mechanism, an individual behavior heterogeneous graph, and a knowledge system hypergraph, including: constructing a cross-view graph neural network as a learning ability assessment model by using the context-aware mechanism as a scene sub-network, the individual behavior heterogeneous graph as a behavior sub-network, and the knowledge system hypergraph as a knowledge sub-network.

[0110] In some embodiments, a cross-view graph neural network is constructed as a learning ability evaluation model by using a context-aware mechanism as a scene sub-network, an individual behavior heterogeneous graph as a behavior sub-network, and a knowledge system hypergraph as a knowledge sub-network. This includes: using a context-aware mechanism as a scene sub-network, an individual behavior heterogeneous graph as a behavior sub-network, and a knowledge system hypergraph as a knowledge sub-network, while using a contrastive loss function based on positive and negative sample pairs as the total network loss, and performing weighted fusion of the three-view embeddings through a self-attention mechanism to obtain the learning ability evaluation model.

[0111] In this embodiment of the invention, the learning ability assessment model is constructed based on three sub-networks, corresponding to the modeling of learning scenarios, user behavior, and knowledge systems, respectively:

[0112] The context-aware mechanism serves as a scene-specific sub-network. This sub-network assigns scene-sensitive attention weights to nodes in the knowledge point sequential graph by learning a scene attention matrix. For example, it generates three-dimensional feature vectors based on user state, resource attributes, and temporal context, dynamically adjusting the importance of knowledge points in different scenarios. Its core function is to enable the model to perceive changes in the learning scenario and adapt to diverse learning needs.

[0113] Individual behavioral heterogeneous graphs serve as behavioral subnetworks: these subnetworks model dynamic user behavior data and characterize the real-time interaction relationships between "users, resources, and knowledge points" through three types of edges (user-resource-knowledge point) and a time decay function. For example, resources recently downloaded and frequently accessed knowledge points have higher weights, reflecting the current learning focus.

[0114] The knowledge system hypergraph serves as a sub-network of knowledge: This sub-network constructs a hypergraph structure based on the hierarchical inclusion, horizontal parallelism, and vertical progression of knowledge points, and quantifies the strength of logical connections between knowledge points through a matching score. For example, the edge weight of the vertical progression relationship between "Basic Calculus" and "Multivariable Calculus" is 0.95, reflecting a strong dependency order and supporting the systematic modeling of the knowledge system.

[0115] The integration mechanism across graph neural networks is achieved through the following steps:

[0116] Parallel input of multiple sub-networks: The scene sub-network (context-aware mechanism), behavior sub-network (individual behavior heterogeneous graph), and knowledge sub-network (knowledge system hypergraph) are respectively used as independent views and input to different branches of the graph neural network to extract feature representations of each view.

[0117] Self-attention fusion: The embedding vectors of the three views are weighted and fused using a self-attention mechanism to automatically learn the importance weights of different views in the evaluation. For example, when evaluating the learning ability of novice users, the recent operation records of the behavior subnetwork and the "novice scene" features of the scene subnetwork may be given higher weights.

[0118] Contrastive loss function optimization: A contrastive loss function based on positive and negative sample pairs is introduced as the total network loss, which forces the feature representations of similar samples to be closer in the embedding space and the dissimilar samples to be further apart, thereby improving the model's ability to distinguish differences in users' learning abilities.

[0119] This invention uses graph neural networks to fuse heterogeneous data (behavior + knowledge + scenario), breaking through the limitations of traditional models that rely on single-modal data, and achieving accurate modeling of complex relationships between "user-resource-knowledge points", effectively improving the accuracy of learning ability assessment.

[0120] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0121] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for constructing a learning ability assessment model for personalized learning on an online learning platform, characterized in that, include: Collect user behavior data and learning resource metadata during the learning process; Based on the behavioral data and the learning resource metadata, construct an individual behavior heterogeneous graph, a knowledge system hypergraph, and a knowledge point sequential graph; A learning scenario attention matrix is ​​preset, and a context awareness mechanism is determined based on the learning scenario attention matrix and the knowledge point sequence diagram; Based on the aforementioned context-aware mechanism, individual behavior heterogeneity graph, and knowledge system hypergraph, a learning ability assessment model is constructed. Based on the attention matrix and knowledge point sequence graph of the learning scenario, the context-aware mechanism is determined, including: Based on the learning scenario attention matrix, determine the attention weight of each knowledge point in the knowledge point sequence graph under different scenarios; The attention weight of each knowledge point in different scenarios is added as a node attribute to the knowledge point sequence graph to form a context-aware mechanism. Based on the behavioral data and the learning resource metadata, an individual behavior heterogeneity graph, a knowledge system hypergraph, and a knowledge point sequence graph are constructed, including: Based on the learning resource metadata, determine the heterogeneous graph nodes, hypergraph nodes, and logical relationships of knowledge points; Based on the behavioral data, the edges of the heterogeneous graph are determined, and the individual behavioral heterogeneous graph is constructed by combining the nodes of the heterogeneous graph. The edges of the hypergraph are determined based on the logical relationships between knowledge points, and the knowledge system hypergraph is constructed by combining the hypergraph nodes. Based on the knowledge system hypergraph and the logical relationships between the knowledge points, a knowledge point sequence graph is established; Based on the behavioral data, the edges of the heterogeneous graph are determined, and the individual behavioral heterogeneous graph is constructed by combining the nodes of the heterogeneous graph, including: Define user-learning resource edges, user-knowledge point edges, and learning resource-knowledge point edges; A first-order time decay function is set on the user-learning resource edge, and an exponential decay function is set on the user-knowledge point edge; Connect the user-learning resource edge, the user-knowledge point edge, and the learning resource-knowledge point edge to the nodes of each heterogeneous graph to obtain the individual behavior heterogeneous graph. Based on the aforementioned context-aware mechanism, individual behavior heterogeneity graph, and knowledge system hypergraph, a learning ability assessment model is constructed, including: By using the aforementioned context-aware mechanism as a scene sub-network, the individual behavior heterogeneous graph as a behavior sub-network, and the knowledge system hypergraph as a knowledge sub-network, a cross-view graph neural network is constructed as a learning ability evaluation model.

2. The method for constructing a learning ability assessment model for personalized learning on an online learning platform according to claim 1, characterized in that, The edges of the hypergraph are determined based on the logical relationships between knowledge points, and the knowledge system hypergraph is constructed by combining the hypergraph nodes, including: Based on the logical relationships between knowledge points, determine the hierarchical inclusion relationship, horizontal parallel relationship, and vertical progressive relationship of knowledge points; Based on hierarchical inclusion, horizontal parallel, and vertical progressive relationships, the knowledge point matching score is calculated and used as the edge attribute value of the hypergraph. Connect the edges of the hypergraph to the corresponding nodes of the hypergraph to obtain the knowledge system hypergraph.

3. The method for constructing a learning ability assessment model for personalized learning on an online learning platform according to claim 1, characterized in that, Based on the knowledge system hypergraph and the logical relationships between the knowledge points, a knowledge point sequence graph is established, including: The directed hyperedges with vertical progressive relationships are extracted from the knowledge system hypergraph, and the extracted directed hyperedges are filtered according to the logical relationships of the knowledge points. The selected directed superedges are converted into directed edges in the knowledge point sequence graph, where the direction of the directed edges is consistent with the direction of the superedges, and the edge weights retain the dependency strength, thus obtaining the knowledge point sequence graph.

4. The method for constructing a learning ability assessment model for personalized learning on an online learning platform according to claim 1, characterized in that, Preset attention matrix for learning scenarios, including: Multiple three-dimensional feature vectors are set; wherein, the three-dimensional feature vectors include user state features, resource attribute features, and time environment features; each three-dimensional feature vector is used for a type of learning scenario; A learning scene attention matrix is ​​formed based on multiple three-dimensional feature vectors.

5. The method for constructing a learning ability assessment model for personalized learning on an online learning platform according to claim 1, characterized in that, Using the aforementioned context-aware mechanism as a scene sub-network, the individual behavior heterogeneous graph as a behavior sub-network, and the knowledge system hypergraph as a knowledge sub-network, a cross-view graph neural network is constructed as a learning ability evaluation model, including: The context-aware mechanism is used as the scene sub-network, the individual behavior heterogeneous graph is used as the behavior sub-network, and the knowledge system hypergraph is used as the knowledge sub-network. At the same time, the contrast loss function based on positive and negative sample pairs is used as the total network loss. The three-view embeddings are weighted and fused through a self-attention mechanism to obtain the learning ability evaluation model.

6. The method for constructing a learning ability assessment model for personalized learning on an online learning platform according to any one of claims 1-5, characterized in that, Collect user behavior data and learning resource metadata during the learning process, including: Collect raw data from online learning platforms; The raw data is preprocessed to obtain behavioral data and learning resource metadata.

Images