Adaptive learning path recommendation method based on hypergraph neural network and knowledge tracking

Abstract

The invention discloses an adaptive learning path recommendation method based on a hypergraph neural network and knowledge tracking, and relates to the field of learning path recommendation, and the method comprises the steps: determining an incidence relation between learning resources, and enabling the incidence relation to serve as an edge of a learning resource undirected graph; the features of the learning resources serve as embedded feature vectors of all nodes of the learning resource undirected graph; updating the embedded feature vector by using a graph neural network to obtain a resource embedded vector, and taking the resource embedded vector as a node feature of a learner hypergraph structure; performing iterative aggregation on the learner hypergraph structure by using a hypergraph neural network to obtain a dynamic resource embedding and learner behavior sequence, generating an initial recommendation list, further generating a candidate learning path set, and generating a Pareto frontier solution set by using a non-dominated sorting genetic algorithm II; and calculating a comprehensive score of each path in the solution set based on a dynamic weight distribution strategy and a comprehensive utility function, and determining an optimal learning path. According to the invention, the accuracy and effectiveness of learning path recommendation are improved.

Description

Technical Field

[0001] The present application relates to the field of learning path recommendation, and particularly to an adaptive learning path recommendation method based on hypergraph neural network and knowledge tracing. Background Art

[0002] Currently, online education faces a significant contradiction between a vast amount of learning resources and the personalized needs of learners. Traditional learning path recommendation methods have the following limitations:

[0003] 1. Disconnection between static modeling and dynamic requirements: Traditional recommendation systems (such as collaborative filtering, knowledge graph) mostly model based on static learner features or resource logical relationships, and cannot capture the dynamic changes of knowledge state, interest preference, and resource association during the learning process. For example, collaborative filtering relies on historical interaction data, has the cold start problem and ignores the temporal evolution; although the knowledge graph can represent the prerequisite relationship of knowledge points, it lacks the real-time response ability to the learner's behavior sequence.

[0004] 2. Imbalance between single-objective optimization and comprehensive utility: Most methods focus on a single objective (such as accuracy or diversity), and it is difficult to balance the short-term effect (such as the correct rate of answering questions) and long-term value (such as knowledge gain) of the learning path. Although reinforcement learning can dynamically adjust the path, its reward function design is complex and it is easy to fall into local optimum, making it difficult to take into account multi-dimensional requirements.

[0005] 3. Insufficient modeling of high-order interaction relationships: The interaction between learners and resources usually presents multi-to-many high-order characteristics (such as a learner is simultaneously associated with multiple resources such as videos, exercises, and quizzes), and traditional graph neural networks (GNN) can only model point-to-point relationships and cannot effectively capture such complex dependencies.

[0006] Existing learning path recommendation methods show obvious limitations in multiple key areas. In terms of dynamic modeling, it is difficult for them to adapt to the continuous changes of knowledge state and interest preference during the learning process; in terms of multi-objective optimization, methods often focus on the improvement of a single indicator and ignore the comprehensive consideration of learning effects, such as the balance between accuracy and diversity, and the consideration of both short-term results and long-term development. These problems are intertwined, not only resulting in the rigidity of the recommended path and making it difficult to meet the personalized needs of learners, but also restricting the continuous improvement and in-depth expansion of learning effects. Summary of the Invention

[0007] The purpose of the present application is to provide an adaptive learning path recommendation method based on hypergraph neural network and knowledge tracing to improve the accuracy and effectiveness of learning path recommendation.

[0008] To achieve the above purpose, the present application provides the following solutions:

[0009] In a first aspect, the present application provides an adaptive learning path recommendation method based on a hypergraph neural network and knowledge tracing, including:

[0010] Obtain the historical behavior data of the learner;

[0011] Determine the association relationship between learning resources; the association relationship serves as the edge of the undirected graph of learning resources; the resource type, knowledge point label, and difficulty level of the learning resources serve as the embedding feature vectors of each node in the undirected graph of learning resources; the undirected graph of learning resources includes nodes and edges; one node corresponds to one learning resource; the edge is the logical relationship between various learning resources; the logical relationship includes a prerequisite relationship and a knowledge point association relationship;

[0012] Use a graph neural network to update the embedding feature vectors of each node in the undirected graph of learning resources to obtain resource embedding vectors; the resource embedding vectors serve as the node features of the learner hypergraph structure; the learner hypergraph structure is constructed based on the historical behavior data of the learner, and the learner hypergraph structure includes hypergraph nodes and hyperedges; the hyperedge is the learner; one hypergraph node corresponds to one learning resource interacted by the learner;

[0013] Use a hypergraph neural network to perform iterative aggregation on the learner hypergraph structure to obtain dynamic resource embeddings and the learner behavior sequence;

[0014] Generate an initial recommendation list based on the dynamic resource embeddings and the learner behavior sequence; the initial recommendation list includes multiple learning resources sorted according to the learning resource recommendation probability;

[0015] Generate a candidate learning path set based on the initial recommendation list;

[0016] According to the candidate learning path set, use the non-dominated sorting genetic algorithm II to determine the Pareto front solution set through a multi-objective optimization function; the optimization objectives of the multi-objective optimization function include an accuracy characterization parameter, an adaptability characterization parameter, an effectiveness characterization parameter, and a diversity characterization parameter;

[0017] According to the Pareto front solution set, use a dynamic weight allocation strategy and a comprehensive utility function to determine the optimal learning path; the comprehensive utility function is a function of the accuracy characterization parameter, the adaptability characterization parameter, the effectiveness characterization parameter, and the diversity characterization parameter.

[0018] Optionally, determining the association relationship between learning resources and constructing the undirected graph of learning resources specifically includes:

[0019] Classify the learning resources according to the resource type, knowledge point label, and difficulty level to generate learning resource features;

[0020] Determine the association relationship between learning resources according to the characteristics of the learning resources;

[0021] Construct the association relationship between learning resources into an undirected graph of learning resources.

[0022] Optionally, generate an initial recommendation list based on the dynamic resource embedding and the learner behavior sequence, specifically including:

[0023] According to the dynamic resource embedding and the learner behavior sequence, use a gated recurrent unit to determine the hidden state sequence;

[0024] According to the hidden state sequence, use an attention mechanism to determine the weighted context vector;

[0025] According to the hidden state sequence and the weighted context vector, use a fully connected layer to calculate the learning resource recommendation probability;

[0026] Sort all learning resources according to the learning resource recommendation probability to generate the initial recommendation list.

[0027] Optionally, generate a candidate learning path set based on the initial recommendation list, specifically including:

[0028] Based on the initial recommendation list, generate the candidate learning path set through replacement, insertion, and swapping operations.

[0029] Optionally, according to the candidate learning path set, use the non-dominated sorting genetic algorithm II to determine the Pareto front solution set through a multi-objective optimization function, specifically including:

[0030] Use the non-dominated sorting genetic algorithm II to perform non-dominated sorting on the candidate paths in the candidate learning path set through a multi-objective optimization function to obtain the sorted candidate learning path set;

[0031] Calculate the crowding degree of each candidate learning path in the sorted candidate learning path set;

[0032] Based on the crowding degree and non-dominated solutions, determine the Pareto front solution set.

[0033] Optionally, according to the Pareto front solution set, use a comprehensive utility function to determine the optimal learning path, specifically including:

[0034] According to the set learning stage of the learner, use a dynamic weight allocation strategy to dynamically allocate weights to the accuracy characterization parameter, adaptability characterization parameter, effectiveness characterization parameter, and diversity characterization parameter to obtain the weights of the accuracy characterization parameter, adaptability characterization parameter, effectiveness characterization parameter, and diversity characterization parameter;

[0035] Calculate the path score of the candidate learning paths in the Pareto front solution set by using a comprehensive utility function according to the weights of the accuracy characterization parameter, the adaptability characterization parameter, the effectiveness characterization parameter, and the diversity characterization parameter;

[0036] Determine the optimal learning path according to the path score.

[0037] Optionally, the comprehensive utility function is:

[0038] S(LP) = λ1·Hit@K + λ2·Adaptivity(LP) + λ3·E p (LP) + λ4·Diversity(LP);

[0039] where S(LP) is the path score of the candidate learning path; Hit@K is the accuracy characterization parameter of the candidate learning path; λ1 is the weight corresponding to the accuracy characterization parameter; Adaptivity(LP) is the adaptability characterization parameter of the candidate learning path; λ2 is the weight corresponding to the adaptability characterization parameter; E p (LP) is the effectiveness characterization parameter of the candidate learning path; λ3 is the weight corresponding to the effectiveness characterization parameter; Diversity(LP) is the diversity characterization parameter; λ4 is the weight corresponding to the diversity characterization parameter.

[0040] In a second aspect, the present application provides a computer device, including: a memory, a processor, and a computer program stored on the memory and executable on the processor, where the processor executes the computer program to implement the adaptive learning path recommendation method based on a hypergraph neural network and knowledge tracing described in any one of the above.

[0041] In a third aspect, the present application provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the adaptive learning path recommendation method based on a hypergraph neural network and knowledge tracing described in any one of the above.

[0042] In a fourth aspect, the present application provides a computer program product, including a computer program, and when the computer program is executed by a processor, it implements the adaptive learning path recommendation method based on a hypergraph neural network and knowledge tracing described in any one of the above.

[0043] According to the specific embodiments provided by the present application, the present application has the following technical effects:

[0044] The present application provides an adaptive learning path recommendation method based on hypergraph neural network and knowledge tracing. The method includes: obtaining historical behavior data of learners; determining the association relationships between learning resources; using the association relationships as the edges of an undirected graph of learning resources; using the resource types, knowledge point labels, and difficulty levels of learning resources as the embedding feature vectors of each node in the undirected graph of learning resources; the undirected graph of learning resources includes nodes and edges; using a graph neural network to update the embedding feature vectors of each node in the undirected graph of learning resources to obtain resource embedding vectors; using the resource embedding vectors as the node features of the learner hypergraph structure; the learner hypergraph structure includes hypergraph nodes and hyperedges; using a hypergraph neural network to perform iterative aggregation on the learner hypergraph structure to obtain dynamic resource embeddings and learner behavior sequences; generating an initial recommendation list based on the dynamic resource embeddings and learner behavior sequences; the initial recommendation list includes multiple learning resources sorted according to the recommendation probabilities of learning resources; generating a candidate learning path set based on the initial recommendation list; using the non-dominated sorting genetic algorithm II to determine the Pareto front solution set through a multi-objective optimization function according to the candidate learning path set; determining the optimal learning path according to the Pareto front solution set by using a dynamic weight allocation strategy and a comprehensive utility function; the comprehensive utility function is a function of accuracy characterization parameters, adaptability characterization parameters, effectiveness characterization parameters, and diversity characterization parameters. The present application uses a hypergraph neural network to achieve dynamic modeling, effectively capturing the high-order interaction relationships between learners and learning resources, thereby breaking through the limitation of traditional graph neural networks that are only limited to pairwise relationships. At the same time, by combining the non-dominated sorting genetic algorithm II with dynamic weight allocation, multi-objective optimization is achieved, successfully balancing multiple objectives such as accuracy, adaptability, effectiveness, and diversity, and improving the accuracy and effectiveness of learning path recommendation. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for use in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0046] Figure 1 It is a schematic flowchart of an adaptive learning path recommendation method based on hypergraph neural network and knowledge tracing provided by an embodiment of the present application;

[0047] Figure 2 It is a flowchart of the adaptive learning path recommendation method based on hypergraph neural network and knowledge tracing of the present application in practical applications;

[0048] Figure 3 It is a schematic structural diagram of a computer device provided by an embodiment of the present application. Detailed implementation manners

[0049] The following will clearly and completely describe the technical solutions in the embodiments of the present application with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present application without creative efforts shall fall within the protection scope of the present application.

[0050] To make the above objects, features, and advantages of the present application more obvious and understandable, the present application will be further described in detail below with reference to the accompanying drawings and specific implementation manners.

[0051] The present application is committed to solving the technical problems existing in current learning path recommendations, and proposes an innovative solution of a multi-objective optimization strategy that combines hypergraph dynamic modeling and knowledge tracing technology (i.e., an adaptive learning path recommendation method based on hypergraph neural networks and knowledge tracing). The core of this solution is that it can not only dynamically capture and respond to various changes in the learning process to ensure the adaptability of the recommended path; moreover, through a carefully designed multi-objective optimization mechanism, while pursuing the accuracy of recommendations, this solution also fully considers the diversity of the learning experience, the depth and breadth of knowledge mastery, and strives to find the optimal solution in multiple aspects such as the dynamic changes in the learning process, the accuracy and coverage of the recommendation results, so that each adjustment of the learning path can maximize the all-round development of the learner's knowledge and skills, thereby greatly improving the learning experience and learning effectiveness as a whole.

[0052] The present application selects an online education platform as an example object. As an important carrier of modern education, the online education platform has a wide variety of learning resources (such as videos, exercises, quizzes, etc.), and there are significant differences in the knowledge backgrounds, learning progress, interest preferences, etc. of learners. The overall of these learning resources has the characteristics of multi-modal, multi-dimensional, and dynamic changes. At the same time, due to the individual needs of learners and the differences in the learning environment, the recommendation of learning paths needs to comprehensively consider various factors such as resource types, knowledge point associations, and learning behavior time series. Considering the contradiction between the massive learning resources and the individual needs of learners, how to recommend learning paths more effectively and accurately and make them dynamically adapt to the learner's knowledge state and learning goals is an urgent problem to be solved. The learning path recommendation method adopted by the present application can effectively solve this problem.

[0053] In an exemplary embodiment, as Figure 1 and Figure 2 shown, there is provided an adaptive learning path recommendation method based on hypergraph neural networks and knowledge tracing, including the following steps:

[0054] S1: Obtain the historical behavior data of the learner.

[0055] S2: Determine the association relationships between learning resources; the association relationships serve as the edges of the undirected graph of learning resources; the resource type, knowledge point labels, and difficulty levels of learning resources serve as the embedding feature vectors of each node in the undirected graph of learning resources; the undirected graph of learning resources includes nodes and edges; one node corresponds to one learning resource; the edges are the logical relationships between various learning resources; the logical relationships include prerequisite relationships and knowledge point association relationships.

[0056] In practical applications, first, data collection and preprocessing are carried out, that is, multi-modal historical behavior data of learners is obtained, and the association information (association relationships) between learning resources is determined.

[0057] Obtaining the multi-modal features of learning resources is to obtain the data type and corresponding feature representation for the multi-modal learning resources (such as videos, exercises, quizzes) output by the online education platform. The specific steps are as described in S2.

[0058] As an optional implementation manner, S2 specifically includes:

[0059] S21: Classify learning resources according to the resource type, knowledge point labels, and difficulty levels to generate learning resource features.

[0060] In practical applications, learning resources are classified according to the resource type (videos, exercises, quizzes), knowledge point labels, difficulty levels, etc. to generate multi-modal feature representations (learning resource features).

[0061] S22: Determine the association relationships between learning resources according to the learning resource features.

[0062] S23: Construct the association relationships between learning resources into an undirected graph of learning resources.

[0063] In practical applications, secondly, the static relationships of learning resources are modeled through a graph neural network. Specifically, a graph convolutional network (Graph Convolutional Networks, GCN) can be used to capture the static logical relationships between learning resources.

[0064] Capturing the static logical relationships between learning resources enables GCN to update resource embeddings through the relationship graph between resources. The specific steps are as follows:

[0065] Take learning resources as graph nodes Construct the logical relationships between resources (such as prerequisite relationships, knowledge point associations) into edges ε to construct an undirected graph

[0066] Through the message passing mechanism of GCN, the embedding feature vector of each node (resource) Iteratively update based on the features of its neighbor nodes to generate a resource embedding vector, where d represents the dimension of the embedding vector.

[0067]

[0068] Among them, is the resource embedding vector of node v, and l is the layer index of the GCN; W g is a learnable parameter, is all the neighbor nodes of node v, and u is the u-th neighbor node. is the feature vector of node u at layer l - 1.

[0069] S3: Use a graph neural network to update the embedding feature vector of each node in the undirected graph of learning resources to obtain a resource embedding vector; the resource embedding vector is used as the node feature of the learner hypergraph structure; the learner hypergraph structure is constructed based on the historical behavior data of the learner, and the learner hypergraph structure includes hypergraph nodes and hyperedges; the hyperedges are learners; one hypergraph node corresponds to a type of learning resource with which the learner interacts.

[0070] In practical applications, before constructing the learner hypergraph structure, perform data cleaning, standardization processing, and normalization processing on the historical behavior data of the learner to obtain the processed historical behavior data.

[0071] Clean and standardize the historical behavior data of the learner, remove abnormal data (such as records with a question answering time exceeding the threshold), and normalize features such as learning duration and question answering accuracy rate to obtain the processed historical behavior data.

[0072] S4: Use a hypergraph neural network to perform iterative aggregation on the learner hypergraph structure to obtain dynamic resource embeddings and the learner behavior sequence.

[0073] In practical applications, model the historical interaction behavior through a hypergraph neural network. Specifically, a hypergraph attention network (HGAT) can be used to construct a high-order interaction relationship model between the learner and the resources.

[0074] Constructing a high-order interaction relationship model is to dynamically capture the complex dependence relationship between the learner and the multi-modal learning resources through HGAT. The specific steps are as follows:

[0075] Abstract the learner as a hyperedge e ∈ ε h , and the multiple learning resources with which it interacts are used as the nodes connected by the hyperedge Construct a hypergraph structure

[0076] Generate dynamic resource embeddings and learner behavior sequences by iteratively aggregating node and hyperedge features through HGAT.

[0077] Node-to-hyperedge aggregation: The feature h of hyperedge e e Is weighted aggregated by associated node embeddings:

[0078]

[0079] Where Is the feature representation of hyperedge e at the l-th layer; α v,e Is the attention weight of node v to hyperedge e; Is the node feature of node v at the (l - 1)-th layer; Is the feature representation of hyperedge e at the (l - 1)-th layer.

[0080] Hyperedge-to-node aggregation: The node feature e v Is updated to the weighted sum of associated hyperedge features:

[0081]

[0082] Where β e,v Is the attention weight of hyperedge e to node v; W a ,W b Are attention weight matrices, and ∥ represents vector concatenation.

[0083] S5: Generate an initial recommendation list based on the dynamic resource embeddings and the learner behavior sequence; the initial recommendation list includes multiple learning resources sorted by the learning resource recommendation probability.

[0084] In practical applications, after the historical interaction behavior modeling is completed, an initial path is generated, that is, the dynamic evolution in the student learning process is modeled to generate an initial recommendation list. Modeling the dynamic evolution in the student learning process is to introduce a Gated Recurrent Unit (GRU) and a self-attention mechanism to capture the short-term dependencies and long-term evolution patterns of the learning sequence.

[0085] As an optional implementation manner, S5 specifically includes:

[0086] S51: Determine the hidden state sequence using a gated recurrent unit according to the dynamic resource embeddings and the learner behavior sequence.

[0087] In practical applications, introduce a GRU time series modeling module, combine the dynamic resource embeddings generated by the hypergraph with the learner behavior sequence, and output the hidden state sequence {h1,…,h T}:

[0088] h t = GRU(et , h t-1 ).

[0089] Among them, e t is the dynamic resource embedding at time t; h t-1 is the learner behavior sequence at time t - 1.

[0090] S52: According to the hidden state sequence, use the attention mechanism to determine the weighted context vector.

[0091] Calculate the weighted context vector c t , enhancing the ability to capture the evolution of the learning progress.

[0092]

[0093] Among them, α i,j is the attention weight of the target concept j to the current learning concept i; W Q , W K , W V are the projection matrices of Query, Key, and Value respectively, with dimensions of

[0094] S53: According to the hidden state sequence and the weighted context vector, use the fully connected layer to calculate the learning resource recommendation probability.

[0095] Calculate the recommendation probability of each learning resource v through the fully connected layer:[[]]

[0096] p(v) = Softmax(W p [c t ||h t +b p ).

[0097] Among them, W p , b p are learnable parameters.

[0098] S54: Sort all learning resources according to the learning resource recommendation probability to generate the initial recommendation list.

[0099] Generate the initial recommendation list by sorting according to probability

[0100] S6: Generate a candidate learning path set based on the initial recommendation list.

[0101] As an optional implementation manner, S6 specifically includes:[[]]

[0102] Generate the candidate learning path set based on the initial recommendation list through replacement, insertion, and swapping operations.

[0103] After the initial path is generated, candidate path generation and multi-objective optimization are carried out through knowledge tracking and state prediction, that is, based on the initial recommendation list Expand the candidate path through replacement, insertion, and swapping operations to generate a diverse set of candidate learning paths

[0104] S7: According to the set of candidate learning paths, use the Non-dominated Sorting Genetic Algorithm II (NSGA-II) to determine the Pareto front solution set through a multi-objective optimization function; the optimization objectives of the multi-objective optimization function include accuracy characterization parameters, adaptability characterization parameters, effectiveness characterization parameters, and diversity characterization parameters

[0105] Use NSGA-II (Non-dominated Sorting Genetic Algorithm II) for multi-objective optimization and select through Pareto-optimized paths. The specific steps are as follows

[0106] Introduce a Pareto front generation strategy and use a multi-objective optimization function to perform non-dominated sorting on the candidate paths

[0107] Calculate the crowding distance of each candidate learning path in the sorted set of candidate learning paths, and preferentially retain the solutions with sparse distribution to enhance diversity

[0108] Select the Pareto front solution set. First, perform iterative screening. Starting from the first layer, select solutions layer by layer until the preset number (such as Top-10) is reached. If a certain layer cannot accommodate all, select according to the crowding distance from high to low. The final Pareto front solution set includes all non-dominated solutions and some solutions with high crowding distances

[0109] As an optional implementation, S7 specifically includes

[0110] S71: Use the Non-dominated Sorting Genetic Algorithm II to perform non-dominated sorting on the candidate paths in the set of candidate learning paths through a multi-objective optimization function to obtain a sorted set of candidate learning paths

[0111] In practical applications, use the Non-dominated Sorting Genetic Algorithm II to evaluate the multi-objective optimization function values of each candidate path in the set of candidate learning paths, perform non-dominated sorting, and obtain a sorted set of candidate learning paths

[0112] The multi-objective optimization function is

[0113] max F(LP) = [Hit@K(LP), Adaptivity(LP), E p (LP), Diversity(LP)].

[0114] Accuracy characterization parameter (Hit@K):

[0115]

[0116] where n hit is the number of times the target learning resource appears in the Top-k learning resource list predicted by the model; N is the number of test data.

[0117] Adaptivity characterization parameter (Adaptivity):

[0118]

[0119] where |M| is the length of the list; δ is the current ability value of the student,

[0120] , T is the size of the historical window (such as the last 10 answers), ∈ is a smoothing term (to prevent the denominator from being 0), default ∈ = 1e-5. Dif i is the labeled difficulty of the i-th learning resource.

[0121] Effectiveness characterization parameter (E p ):

[0122]

[0123] where E e is the student's mastery of the target concept after the learning path LP; E s is the student's mastery of the target concept before the learning path; E sup is the upper limit of the mastery of the target concept; Gain i is the knowledge gain.

[0124] In practical applications, the knowledge tracking module analyzes the learner's knowledge mastery status in real time.

[0125] The knowledge tracking module quantifies the learner's knowledge gain potential through multi-source data fusion and dynamic prediction. The specific steps are as follows:

[0126] Input the learner's historical behavior data and the dynamic resource embedding e v generated by HGAT, and construct the knowledge state vector s t :

[0127] s t = SLTM(e v ∥r t∥t t ,s t-1 )。

[0128] Among them, r t is the answering accuracy rate, and t t is the learning duration.

[0129] Predict the answering accuracy rate of the learner for the unlearned resources:

[0130] p(r i ) = σ(W p [s t ||e v +b p ).

[0131] Among them, σ is the Sigmoid function.

[0132] Identify the weak knowledge points according to the prediction results, and quantify the knowledge gain Gain i :

[0133]

[0134] Among them, is the prediction accuracy rate before the student learns the resource q i , and is the prediction accuracy rate after the student learns the resource q i . Through denominator adjustment, the gains of students with different initial levels are made comparable.

[0135] Diversity representation parameter (Diversity):

[0136]

[0137] Among them, q i is the i-th learning resource in the path LP; q j is the j-th learning resource in the path LP; e i is the embedding vector of the resource q i ; e j is the embedding vector of the resource q j ; sim() is the cosine similarity.

[0138] S72: Calculate the congestion degree of each candidate learning path in the sorted candidate learning path set.

[0139] S73: Determine the Pareto front solution set based on the congestion degree and non-dominated solutions.

[0140] S8: Based on the Pareto front solution set, determine the optimal learning path using a dynamic weight allocation strategy and a comprehensive utility function; the comprehensive utility function is a function of the accuracy representation parameter, adaptability representation parameter, effectiveness representation parameter, and diversity representation parameter.

[0141] The dynamic weight allocation strategy is to allocate weights to the comprehensive utility function according to the set learning stage of the learner, obtaining the comprehensive utility function with allocated weights.

[0142] In practical applications, perform a comprehensive score ranking on the Pareto front solution set.

[0143] According to rank the candidate paths and select the candidate learning path with the highest score as the optimal learning path. Or select the final path from the Pareto front driven by the learner's preferences.

[0144] As an optional implementation manner, S8 specifically includes:

[0145] S81: According to the set learning stage of the learner, use the dynamic weight allocation strategy to dynamically allocate weights to the accuracy representation parameter, adaptability representation parameter, effectiveness representation parameter, and diversity representation parameter, obtaining the weights of the accuracy representation parameter, adaptability representation parameter, effectiveness representation parameter, and diversity representation parameter;

[0146] Dynamically allocate the weights of each objective, determine the total utility function, output the final recommended path (optimal learning path) by calculating the function value, and adjust the recommended path in combination with real-time learning feedback data (i.e., continuous monitoring and updating).

[0147] Balance the short-term effect and long-term value through the dynamic weight allocation strategy.

[0148] Dynamically adjust the objective weights according to the learning stage set by the learner, including accuracy (Hit@K), adaptability (difficulty matching degree), effectiveness (knowledge gain), and diversity (resource dissimilarity), generating paths under different preferences (such as high-accuracy paths, high-gain paths, balanced paths). The dynamic weight allocation strategy is shown in Table 1.

[0149] Table 1 Weight Allocation Table for Different Learning Stages

[0150] Learning stage <![CDATA[λ1]]> <![CDATA[λ2]]> <![CDATA[λ3]]> <![CDATA[λ4]]> Initial stage (weak) 0.5 0.3 0.1 0.1 Middle stage (consolidation) 0.3 0.4 0.2 0.1 Final stage (improvement) 0.2 0.3 0.4 0.1

[0151] Introduce a diversity constraint (such as the maximum margin relevance algorithm) to avoid duplicate path content.

[0152] Dynamic weight allocation strategy, which adjusts the target weight according to the learning stage (e.g., emphasizing accuracy in the initial stage and effectiveness in the later stage).

[0153] S82: According to the weights of the accuracy representation parameter, the adaptability representation parameter, the effectiveness representation parameter, and the diversity representation parameter, use the comprehensive utility function to calculate the path score of the candidate learning paths in the Pareto front solution set.

[0154] S83: Determine the optimal learning path according to the path score.

[0155] As an optional implementation, the comprehensive utility function is:[[]]END]]

[0156] S(LP) = λ1·Hit@K + λ2·Adaptivity(LP) + λ3·E p (LP) + λ4·Diversity(LP);

[0157] Wherein, S(LP) is the path score of the candidate learning path; Hit@K is the accuracy representation parameter of the candidate learning path; λ1 is the weight corresponding to the accuracy representation parameter; Adaptivity(LP) is the adaptability representation parameter of the candidate learning path; λ2 is the weight corresponding to the adaptability representation parameter; E p (LP) is the effectiveness representation parameter of the candidate learning path; λ3 is the weight corresponding to the effectiveness representation parameter; Diversity(LP) is the diversity representation parameter; λ4 is the weight corresponding to the diversity representation parameter.

[0158] Continuously monitor the learning effect of the learner learning according to the optimal learning path to update the recommendation strategy.

[0159] Continuous monitoring and updating maintain the real-time performance and adaptability of the recommendation system through a feedback mechanism. The specific steps are as follows:[[]]END]]

[0160] Real-time collect learner interaction data (such as answering record (answer accuracy rate), learning duration, path deviation behavior).

[0161] Update the parameters of the HGNN and the knowledge tracking module to ensure that the recommended optimal learning path is dynamically optimized with the learning progress.

[0162] For the above-mentioned adaptive learning path recommendation method based on hypergraph neural network and knowledge tracking, the present application also provides a system architecture, including a data collection layer (learning behavior log), a model calculation layer (HGNN + knowledge tracking module), and a recommendation engine layer (multi-objective optimization output).

[0163] The client displays the dynamically adjusted learning path through the interaction interface, supporting real-time feedback from students.

[0164] The significant advantage of this application lies in its use of the Hypergraph Neural Network (HGNN) to achieve dynamic modeling, effectively capturing the high-order interaction relationships between learners and resources, thereby breaking through the limitation of traditional Graph Neural Networks (GNNs) which are only limited to pairwise relationships. At the same time, by combining the Pareto front generation strategy with dynamic weight allocation, this application achieves multi-objective optimization and successfully balances multiple objectives such as accuracy, adaptability, effectiveness, and diversity. In addition, this application can provide real-time feedback and drive the recommendation system to maintain high responsiveness and adaptability. This method provides an efficient, accurate, and robust personalized learning path generation and adjustment solution for online education scenarios by dynamically modeling learner-resource interaction relationships, tracking knowledge states in real time, and combining multi-dimensional optimization strategies.

[0165] Advantages of this application over related technologies:

[0166] ① Dynamic adaptability: Real-time update the recommendation path through knowledge tracking to solve the problem of the dynamically changing knowledge states of students.

[0167] ② High-order relationship modeling: HGNN processes many-to-many interactions, and the recommendation accuracy is better than that of traditional GNNs.

[0168] ③ Multi-objective balance: The optimization framework ensures that the recommended path is both challenging and interesting.

[0169] HGNN captures complex relationships, the knowledge tracking module provides real-time state input, and multi-objective optimization ensures comprehensive performance.

[0170] Other alternative solutions that can also achieve the purpose of this application:

[0171] Interaction modeling based on traditional Graph Neural Networks (GNNs): Use traditional GNNs (such as Graph Convolutional Network (GCN), Graph Attention Network (GAT)) to replace the Hypergraph Neural Network (HGNN), and capture the interaction relationships between learners and resources through node representation learning.

[0172] Temporal modeling of the Transformer architecture: Use the self-attention mechanism of the Transformer to replace GRU / LSTM (Long Short-Term Memory) to directly model the long-term dependencies of learning behavior sequences.

[0173] Fusion of Bayesian Knowledge Tracing (BKT) and deep models: Combine the Bayesian probability model with a deep neural network to predict the probability that a learner has mastered unlearned resources.

[0174] Graph-Enhanced Knowledge Tracing (Graph-Enhanced KT): Introduce a knowledge graph in knowledge tracing and fuse the logical relationships between knowledge points through GCN.

[0175] Genetic Algorithm-Driven Multi-Objective Optimization: Use a genetic algorithm to generate a set of candidate paths and explore diverse solutions through crossover and mutation operations.

[0176] Q-Learning-Based Reinforcement Learning Framework: Adopt the Q-learning algorithm to replace the policy gradient and optimize the path selection strategy through a Q-value table or a Deep Q-Network (DQN).

[0177] In an exemplary embodiment, a computer device is provided, including a memory and a processor. A computer program is stored in the memory, and when the processor executes the computer program, the above-mentioned adaptive learning path recommendation method based on a hypergraph neural network and knowledge tracing is implemented.

[0178] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program, and when the computer program is executed by a processor, the above-mentioned adaptive learning path recommendation method based on a hypergraph neural network and knowledge tracing is implemented.

[0179] In an exemplary embodiment, a computer program product is provided, including a computer program, and when the computer program is executed by a processor, the above-mentioned adaptive learning path recommendation method based on a hypergraph neural network and knowledge tracing is implemented.

[0180] In an exemplary embodiment, a computer device is provided. The computer device can be a server or a terminal, and its internal structure diagram can be as Figure 3 shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O), and a communication interface. Among them, the processor, the memory, and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The input / output interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, an adaptive learning path recommendation method based on a hypergraph neural network and knowledge tracing is implemented.

[0181] Those skilled in the art can understand,Figure 3 The structure shown is only a block diagram of some structures related to the solution of this application, and does not constitute a limitation on the computer device to which the solution of this application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine some components, or have different component arrangements.

[0182] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use, and processing of relevant data need to comply with relevant regulations.

[0183] Those of ordinary skill in the art can understand that all or part of the processes of implementing the methods in the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, database, or other medium used in the embodiments provided in this application can include at least one of non-volatile and volatile memories. Non-volatile memories can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memories can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0184] The databases involved in the embodiments provided in this application can include at least one of relational databases and non-relational databases. Non-relational databases can include distributed databases based on blockchain, etc., and are not limited thereto. The processors involved in the embodiments provided in this application can be general-purpose processors, central processors, graphics processors, digital signal processors, programmable logics, data processing logics based on quantum computing, etc., and are not limited thereto.

[0185] The technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this specification.

[0186] Specific examples are used in this text to elaborate on the principles and implementation manners of the present application. The descriptions of the above embodiments are only used to help understand the method and its core idea of the present application; at the same time, for those of ordinary skill in the art, according to the idea of the present application, there will be changes in the specific implementation manners and application scopes. In summary, the content of this specification should not be construed as a limitation to the present application.

Claims

1. An adaptive learning path recommendation method based on hypergraph neural network and knowledge tracing, characterized in that Including: Obtaining the historical behavior data of the learner; Determining the association relationship between learning resources; Taking the association relationship as the edge of the undirected graph of learning resources; The resource type, knowledge point label, and difficulty level of the learning resources are used as the embedding feature vector of each node in the undirected graph of learning resources; the undirected graph of learning resources includes nodes and edges; one node corresponds to one learning resource; the edge is the logical relationship between learning resources; the logical relationship includes prerequisite relationship and knowledge point association relationship; Using a graph neural network to update the embedding feature vector of each node in the undirected graph of learning resources to obtain a resource embedding vector; the resource embedding vector is used as the node feature of the learner hypergraph structure; The learner hypergraph structure is constructed based on the historical behavior data of the learner, and the learner hypergraph structure includes hypergraph nodes and hyperedges; the hyperedge is the learner; one hypergraph node corresponds to one learning resource interacted by the learner; Using a hypergraph neural network to perform iterative aggregation on the learner hypergraph structure to obtain dynamic resource embeddings and the learner behavior sequence; Generating an initial recommendation list based on the dynamic resource embeddings and the learner behavior sequence; The initial recommendation list includes multiple learning resources sorted according to the learning resource recommendation probability; Generating a candidate learning path set based on the initial recommendation list; According to the candidate learning path set, using the Non-dominated Sorting Genetic Algorithm II, determining the Pareto front solution set through a multi-objective optimization function; the optimization objectives of the multi-objective optimization function include an accuracy characterization parameter, an adaptability characterization parameter, an effectiveness characterization parameter, and a diversity characterization parameter; Determining the optimal learning path according to the Pareto front solution set by using a dynamic weight allocation strategy and a comprehensive utility function; The comprehensive utility function is a function of the accuracy characterization parameter, the adaptability characterization parameter, the effectiveness characterization parameter, and the diversity characterization parameter.

2. The adaptive learning path recommendation method based on hypergraph neural network and knowledge tracing according to claim 1, wherein Determining the association relationship between learning resources and constructing an undirected graph of learning resources, specifically including: Classifying learning resources according to the resource type, knowledge point label, and difficulty level to generate learning resource features; Determining the association relationship between learning resources according to the learning resource features; Constructing the association relationship between learning resources into an undirected graph of learning resources.

3. The adaptive learning path recommendation method based on hypergraph neural network and knowledge tracing according to claim 1, characterized in that Generating an initial recommendation list based on the dynamic resource embeddings and the learner behavior sequence, specifically including: According to the dynamic resource embeddings and the learner behavior sequence, using a gated recurrent unit to determine the hidden state sequence; According to the hidden state sequence, using an attention mechanism to determine the weighted context vector; According to the hidden state sequence and the weighted context vector, using a fully connected layer to calculate the learning resource recommendation probability; Sorting all learning resources according to the learning resource recommendation probability to generate the initial recommendation list.

4. The adaptive learning path recommendation method based on hypergraph neural network and knowledge tracing according to claim 1, wherein Generating a candidate learning path set based on the initial recommendation list, specifically including: Generating the candidate learning path set based on the initial recommendation list through replacement, insertion, and swap operations.

5. The adaptive learning path recommendation method based on hypergraph neural network and knowledge tracing according to claim 1, characterized in that, According to the set of candidate learning paths, the non-dominated sorting genetic algorithm II is used to determine the Pareto front solution set through a multi-objective optimization function, specifically including: Using the non-dominated sorting genetic algorithm II, perform non-dominated sorting on the candidate paths in the set of candidate learning paths through the multi-objective optimization function to obtain the sorted set of candidate learning paths; Calculate the crowding degree of each candidate learning path in the sorted set of candidate learning paths; Based on the crowding degree and non-dominated solutions, determine the Pareto front solution set.

6. The adaptive learning path recommendation method based on hypergraph neural network and knowledge tracing according to claim 1, wherein According to the Pareto front solution set, use the dynamic weight allocation strategy and the comprehensive utility function to determine the optimal learning path, specifically including: According to the set learning stage of the learner, use the dynamic weight allocation strategy to dynamically allocate weights to the accuracy representation parameter, adaptability representation parameter, effectiveness representation parameter, and diversity representation parameter, and obtain the weights of the accuracy representation parameter, adaptability representation parameter, effectiveness representation parameter, and diversity representation parameter; According to the weights of the accuracy representation parameter, adaptability representation parameter, effectiveness representation parameter, and diversity representation parameter, use the comprehensive utility function to calculate the path scores of the candidate learning paths in the Pareto front solution set; According to the path scores, determine the optimal learning path.

7. The adaptive learning path recommendation method based on hypergraph neural network and knowledge tracing according to claim 1, wherein The comprehensive utility function is: S(LP) = λ1·Hit@K + λ2·Adaptivity(LP) + λ3·E p (LP) + λ4·Diversity(LP); Among them, S(LP) is the path score of the candidate learning path; Hit@K is the accuracy characterization parameter of the candidate learning path; λ1 is the weight corresponding to the accuracy characterization parameter; Adaptivity(LP) is the adaptivity characterization parameter of the candidate learning path; λ2 is the weight corresponding to the adaptivity characterization parameter; E p (LP) is the effectiveness characterization parameter of the candidate learning path; λ3 is the weight corresponding to the effectiveness characterization parameter; Diversity(LP) is the diversity characterization parameter; λ4 is the weight corresponding to the diversity characterization parameter.

Images