Method and system for predicting learning conditions of learners based on meta-paths, and electronic device

Abstract

The present invention relates to the technical field of knowledge tracing, and in particular to a method and system for predicting learning conditions of learners based on meta-paths, and an electronic device. In the method, after learner data, exercise book data, and learner answer interaction sequence data are obtained, firstly, a learner behavioral graph is constructed, and learning content preferences of learners are extracted; secondly, learning abilities of the learners are calculated; thirdly, features of the learning content preferences and features of the learning abilities are fused, and a learner heterogeneous information network based on meta-paths is constructed; next, embedding learning is performed on the leaner heterogeneous information network by using a layered graph attention mechanism, to extract leaner fused features; and finally, learning conditions of the learners are predicted by using a GRU-based knowledge tracing model. Compared with the prior art, the present invention has the advantages of more accurately predicting the probability that a learner will answer a next exercise correctly, evaluating and predicting a learning outcome of the learner, and thus providing the learner with multi-dimensional feedback and guidance.

Description

Learner learning situation prediction method, system and electronic device based on meta-path Technical Field

[0001] The present invention relates to the field of knowledge tracking technology, and in particular to a method, system and electronic equipment for predicting a learner's learning situation based on a meta-path. Background Art

[0002] In big data research on smart education, accurately evaluating and predicting learners' learning outcomes has always been a key focus and hot topic. The Knowledge Tracing (KT) model tracks changes in a learner's knowledge state in real time and predicts future learning performance. This core technology, which estimates and predicts a learner's knowledge mastery and mines their learning information, not only provides direct services and assistance for personalized education but also effectively supports multiple key educational processes or scenarios, such as automated assessment, learning resource recommendations, educational interactive robots, and test generation. This improves classroom teaching models and learner evaluation methods, providing technical support for establishing a new model of data-based education governance.

[0003] Currently, research on knowledge tracing primarily focuses on mining and analyzing learners' answering behavior data. Learners' answering behavior data is sequential, and artificial intelligence techniques can be used to transform learners' characteristics into digital, vectorized forms. However, existing research has rarely considered converting learners' answering behavior sequence data into graph-structured data to capture local and global features. Furthermore, commonly used knowledge tracing models use learners' answer records as input data. However, due to the limited number of learning features, this input data is sparse and does not take learners' individual attributes into comprehensive consideration.

[0004] In recent years, deep learning has achieved remarkable results in many fields. Many scholars have begun applying it to the field of knowledge tracing, evolving knowledge tracing from traditional statistical methods and machine learning models to deep learning models. This has resulted in a significant amount of research, such as knowledge tracing-based learning path recommendations, automated assessment of subjective questions, and validation of classic educational theories. Among these, the most groundbreaking work is the Deep Knowledge Tracing (DKT) model. The DKT model is the first to combine deep learning with knowledge tracing. Its input vector is a binary tuple of learner response behavior records. It uses a recurrent neural network (RNN) to simulate the learner's implicit knowledge mastery and predict their future learning performance. However, the DKT model has certain technical shortcomings: 1. A lack of graph-relational data; 2. Data sparsity; and 3. A lack of learner assessment. Therefore, improving the accuracy of predictions of learner performance has become a pressing issue in this field.

[0005] Summary of the Invention

[0006] The purpose of the present invention is to provide a method, system and electronic device for predicting learners' learning status based on meta-path in order to overcome the defect of the traditional DKT model in the prior art that the accuracy of predicting learners' learning status needs to be improved.

[0007] The purpose of the present invention can be achieved by the following technical solutions:

[0008] According to a first aspect of the present invention, a method for predicting a learner's learning status based on a meta-path is provided, comprising the following steps:

[0009] S1, obtain learner data, exercise set data and learner answer interaction sequence data;

[0010] S2, constructing a learner behavior graph based on the learner's answering interaction sequence data, and extracting each learner's learning content preference from the learner behavior graph using a self-organizing map neural network;

[0011] S3, calculating the learning ability of each learner based on the learner data, the exercise set data and the learner answering interaction sequence data;

[0012] S4, integrating the characteristics of the learning content preference and the characteristics of the learning ability to construct a learner heterogeneous information network based on a meta-path, wherein the meta-path is selected by analyzing the relationship between the learning content preference and the learning ability;

[0013] S5, using a hierarchical graph attention mechanism, embedding the learner's heterogeneous information network into learning and extracting the learner's fusion features;

[0014] S6, based on the learner fusion features, using the GRU-based knowledge tracking model to predict the learning status of each learner.

[0015] As a preferred technical solution, the exercise set data is obtained by performing a vector product of a knowledge point vector and a corresponding exercise difficulty coefficient.

[0016] As a preferred technical solution, the learner's question-answering interaction sequence is composed of multiple groups of corresponding exercise vectors and answer results in sequence.

[0017] As a preferred technical solution, the process of extracting each learner's learning content preference from the learner behavior graph using a self-organizing map neural network includes:

[0018] Based on the learner behavior graph, multi-head graph attention and mean aggregation are used to obtain the behavior graph embedding vector of each learner;

[0019] Based on the behavior graph embedding vector, the self-organizing map neural network is used to cluster the behavior of each learner into different learning content preference categories to obtain the learner's learning content preference category, and the learning content preference category is added to the learner's characteristics.

[0020] As a preferred technical solution, the learning ability is obtained by calculating the difference between the proportion of correctly answered exercises and the proportion of incorrectly answered exercises.

[0021] As a preferred technical solution, the meta-path includes the learner-learning content preference-learner meta-path and the learner-learning ability-learner meta-path.

[0022] As a preferred technical solution, the hierarchical graph attention mechanism includes a node-level attention mechanism and a meta-path-level attention mechanism.

[0023] As a preferred technical solution, the S6 specifically includes:

[0024] Splicing the learner fusion feature and the historical data of the learner's answering interaction sequence into a triple vector;

[0025] Initialize the GRU-based knowledge tracking model;

[0026] Based on the triplet vector, the initialized GRU-based knowledge tracking model is used to obtain the prediction results of each learner's learning situation.

[0027] According to a second aspect of the present invention, a system for predicting a learner's learning situation based on a meta-path is provided. The system comprises a signal-connected data acquisition module, a feature engineering module, a feature fusion module, and a prediction module.

[0028] The data acquisition module is used to acquire learner data, exercise set data and learner answering interaction sequence data;

[0029] The feature engineering module is used to construct a learner behavior graph based on the learner's question-answering interaction sequence data, extract each learner's learning content preference from the learner behavior graph using a self-organizing map neural network, and calculate each learner's learning ability based on the learner data, the exercise set data, and the learner's question-answering interaction sequence data;

[0030] The feature fusion module is used to fuse the features of the learning content preference and the features of the learning ability to construct a learner heterogeneous information network based on a meta-path. The meta-path is selected by analyzing the relationship between the learning content preference and the learning ability, and a hierarchical graph attention mechanism is used to embed learning in the learner heterogeneous information network to extract learner fusion features.

[0031] The prediction module is used to predict the learning status of each learner based on the learner fusion features and using a GRU-based knowledge tracking model.

[0032] According to a third aspect of the present invention, an electronic device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] 1. The learning situation prediction method proposed in the present invention is a meta-path-based learner heterogeneous information network knowledge tracing (Heterogeneous Information Network Embedding with Meta-path Knowledge Tracing, MPHINE-KT) method. This method uses a hierarchical graph attention mechanism to embed learning in learner heterogeneous information networks and extract learner fusion features. That is, it considers the fusion and utilization of information in learner heterogeneous information networks and adopts a hierarchical attention learning representation method to complete feature extraction of learner heterogeneous information networks, capturing the semantics between nodes in the network and the topological structure information of the network. Compared with the traditional DKT model, it makes full use of graph relational data, can better capture the associations at different levels in heterogeneous information networks, more effectively integrate various information in the network, improve the model's representation ability and understanding of learner behavior, and thus improve the accuracy of predicting the learning situation of different learners;

[0035] 2. This invention uses learning content preferences and learning abilities to construct a meta-path-based learner heterogeneous information network, effectively solving the problem of how to integrate more effective features and evaluating learner attributes to further improve the accuracy of predicting learners' future learning performance.

[0036] 3. The present invention uses a GRU-based knowledge tracking model to predict the learning status of each learner. The knowledge tracking model is embedded and uses the learner's heterogeneous information network as a constraint. It can solve the data sparsity in the traditional DKT model and improve the predictive effect of the knowledge tracking model on predicting the learner's future performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] FIG1 is a schematic flow chart of a method according to an embodiment of the present invention;

[0038] FIG2 is a schematic diagram of the structure of a learner heterogeneous information network based on meta-paths in an embodiment of the present invention;

[0039] FIG3 is a schematic diagram of the structure of the system in an embodiment of the present invention. DETAILED DESCRIPTION

[0040] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0041] Example

[0042] As shown in FIG1 , this embodiment provides a method for predicting a learner's learning status based on a meta-path, comprising the following steps:

[0043] Step S1, obtain learner data, exercise data and learner answer interaction sequence data. That is, obtain learner learning interaction data. Specifically, define the learner set S = {s1, s2, ..., s n}, knowledge point set K = {k1, k2, ..., k M}, exercise set Q = {q1, q2, ..., q m}.

[0044] Among them, k n Represents the nth knowledge point vector. Each exercise contains one or more knowledge points, so the knowledge point represented in the data set may be one or more. The representation of the exercise is q i , by transforming the vector k n It is obtained by vector multiplication of the difficulty coefficient of the exercise. i It represents the difficulty coefficient of the exercise, and its formula is as follows:

[0045] Among them, t i N represents the number of learners who answered exercise i correctly. i The total number of people who answered exercise i. The difficulty coefficient of each exercise is between 0 and 1. The larger the difficulty coefficient, the easier the question, and the smaller the difficulty coefficient, the more difficult the question.

[0046] The learner's answering interaction sequence is expressed as:

[0047] R{r1, r2, ..., r t-1}={(q1, a1), (q2, a2),..., (q t-1 , a t-1 )} (2)

[0048] The above formula shows that the learner's answering interaction sequence is a set consisting of t-1 exercise vectors and answer results, t represents the order in which the learner answers questions in the online learning system, where a t =1 means the learner answered correctly, a t =0 means the learner answered incorrectly.

[0049] Step S2: construct a learner behavior graph based on the learner's answering interaction sequence data, and use a self-organizing map neural network to extract the learning content preference of each learner from the learner behavior graph.

[0050] In this embodiment, the learner's learning content preference indicates that the learner has a special tendency to choose certain learning contents during the learning process. t-1}Build a behavior graph. A behavior graph is an undirected weighted graph. The nodes represent the questions that appear in the answer sequence, the edges represent the transfer between two questions, and the weight represents the number of transfers.

[0051] Use G b ={V b , E b} represents the learner behavior graph, G b Each node v∈V b represents the exercises q∈Q answered by the learner during the test-taking process, and each edge e ij ∈E b Indicates q i and q j The correlation between ij , d ij Indicates that node q i From the starting node to the ending node q j The number of edges.

[0052] Using the graph attention network, we calculate the node q in the learner behavior graph jIn order to assign different weights to different neighbor nodes, the calculation formula of the attention value of each node in the learner behavior graph is as follows:

[0053] Among them, the weight matrix W and parameter α are trainable parameters, V b Represents Exercise q i The set of neighbor nodes, q k Represents Exercise q i A neighbor node of .

[0054] In this embodiment, multi-head graph attention is used to update the representation of the node, so that it includes the characteristics of neighboring nodes. The representation of all nodes is aggregated by the mean to obtain the representation of the graph embedding, that is, the learner's learning content preference representation I. Specifically, combined with the obtained learner behavior graph node attribute information and behavior graph topology information, the attention between different nodes in the graph attention network and the attention between different node edges can be combined to form the attention coefficient g ij , which is calculated as follows:

[0055] g ij =α ij θ ij (5)

[0056] Among them, d ji For node q i To node q j The number of edges, d i For all points to q i The number of edges.

[0057] Based on the attention coefficients captured by the graph attention network, the feature vectors of the nodes in the behavior graph are updated, and multi-head attention is used to enhance the updating of node features and representation learning. After updating the nodes with multi-head attention, the relevant information is aggregated using mean aggregation to obtain the embedding vector of each learner based on the behavior graph, which is the learner's learning content preference representation I. The formula is as follows:

[0058] Among them, q i is the representation of exercise i after the single-head attention mechanism, q i (H) is the representation of exercise i after the H-head attention mechanism, σ is the sigmoid function, W i is a trainable parameter, H is the number of attention maps, represents the weight parameter of the h-th attention head, mean is the mean function, V b is the set of all points in the behavior graph.

[0059] The learning content preference representation I of all learners is input into a self-organizing map neural network (SOM), and the learners' answering behaviors are clustered into different learning content preference categories, thereby obtaining each learner's learning content preference category. The learning content preference category is added to the learner's features as a discrete custom constant to construct a heterogeneous information network. The output layer in the self-organizing map neural network is also called the competition layer. The entire learning process can be divided into competition, collaboration, and adaptation processes. In the competition process, the input vector I and weight W are vector normalized. In the collaboration process, the winning neuron selects neighboring neurons and updates the weight of the winning neuron. In the adaptation process, the learning rate is used to control convergence. The clustering calculation formula is as follows:

[0060] in, is the weight matrix, η is the learning rate with respect to the training time t and the winning neural unit topology distance dis. If the learning rate η≤η min Or when the preset number of iterations is reached, the network training process ends.

[0061] Step S3: Calculate the learning ability of each learner based on the learner data, the exercise set data, and the learner's interactive sequence data.

[0062] The learning ability of learners reflects their ability to learn knowledge. By calculating the correct answer ratio of exercises, Correct(q ij ) and the proportion of incorrectly answered questions Incorrect(q ij ) is used to represent the learner's learning ability. The learner's learning ability characteristic is represented by Ability, and its calculation formula is:

[0063] Among them, N ij Representative learners i Try to answer q j The number of times Correct(q ij ) and Incorrect(q ij ) represent learners s i Answer the exercises correctly or incorrectly j The probability of b. ij Representative learners i Whether the question is answered correctly or incorrectly at time t, the value is 0 or 1, 0 represents an incorrect answer and 1 represents a correct answer.

[0064] As mentioned above, the learning ability of learners ranges from -1 to 1. Learners are divided into five ability groups. Each learner is assigned a learning ability label, and the average learning ability of each ability group is calculated. The learner is then assigned the average learning ability value of that ability group. Specifically, if the average learning ability is <= 0.4, the learning ability level is poor; if the average learning ability is 0.5-0.6, the learning ability level is fair; if the average learning ability is 0.7-0.8, the learning ability level is medium; if the average learning ability is 0.8-0.9, the learning ability level is good; and if the average learning ability is 0.9-1.0, the learning ability level is excellent.

[0065] Step S4 integrates the characteristics of learning content preferences and learning ability to construct a learner heterogeneous information network based on meta-paths. Meta-paths analyze the relationship between learning content preferences and learning ability to explore different semantic relationships between learners. Meta-paths are the connections defined between objects in the learner heterogeneous information network.

[0066] Specifically, a learner heterogeneous information network is constructed to represent the relationship between the learner (s), the learning content preference (c), and the learning ability (a), i.e., the three objects in the heterogeneous graph. Two types of connections are established for the three objects: the "learner-learning content preference" connection and the "learner-learning ability" connection. The objects in the learner heterogeneous information network can be connected through defined connection methods, i.e., the "learner-learning content preference-learner" (SCS) meta-path and the "learner-learning ability-learner" (SAS) meta-path, forming the heterogeneous graph G. l-HIN =(V G ,E G ), V b ={c,a,s},

[0067] In step S5, a hierarchical graph attention mechanism is used to embed the learner’s heterogeneous information network and extract the learner’s fusion features.

[0068] Specifically, a hierarchical attention mechanism is used to embed the learner's heterogeneous information network, using node-level and meta-path-level attention mechanisms. The node feature vector dimension needs to be transformed so that the projected features of all nodes share the same dimension. The formula is as follows:

[0069] in, are the original feature vectors of different types of nodes, is the latent eigenvector after projection, is a learnable weight matrix.

[0070] (1) Update node weights using node-level attention mechanism.

[0071] The target node is encoded using a node-level attention mechanism. First, the same type of node-level attention mechanism is used to calculate the influence of the learning content preference node on the target node and the influence of the learning ability node on the target node, and the embedding of the target node is updated. Then, different types of node-level attention mechanisms are used to calculate the attention scores of the two different types of nodes, learning content preference and learning ability, on the target node. Finally, the node-level embedding vector of the target node is obtained by calculating the inner product of the attention scores of different nodes and the embedding vector of the target node. Specifically:

[0072] For neighbor nodes of the same type, their contributions to the target node are different. The calculation formula for the node's attention score is as follows:

[0073] Among them, a type is the same type of node-level attention parameter vector, which is a learnable parameter, α ij is the importance weight of nodes of the same type to the target node, Updated vector for nodes of the same type.

[0074] Considering that different types of nodes contribute differently to the target node, after completing the weight calculation of the attention of nodes of the same type to the target node, the attention weight of each node type to the target node is calculated. The formula is as follows:

[0075] in, is the attention parameter vector of different types of node levels, and are the weight matrix and offset value, which are learnable parameters. is the updated vector obtained according to the above formula The linear change value of is used to calculate the weight of each type of node. Normalization is done to get Represents the contribution of different types of neighbor nodes to the target node, and then and Perform product operation to get the embedded node level vector of the target node

[0076] (2) Use the meta-path level attention mechanism to learn the importance of different meta-path instances in affecting the target node.

[0077] The target node is encoded using a node-level attention mechanism. First, the same type of node-level attention mechanism is used to calculate the influence of the learning content preference node on the target node and the influence of the learning ability node on the target node, and the embedding of the target node is updated. Then, different types of node-level attention mechanisms are used to calculate the attention scores of the two different types of nodes, learning content preference and learning ability, on the target node. Finally, the node-level embedding vector of the target node is obtained by calculating the inner product of the attention scores of different nodes and the embedding vector of the target node. Specifically:

[0078] Defining two symmetric meta-paths in learner heterogeneous information networks Considering the different contributions of different meta-paths to the target node, a meta-path-level attention mechanism is used to pass messages from the path to the target node. First, the nodes in the meta-path are updated. The formula is as follows:

[0079] in, is the updated node vector on the meta-path, d i and d j is the degree of node i and node j, are the neighbor nodes of target node i under different meta-path conditions.

[0080] Calculate the target node i and all meta paths The importance is calculated using the attention mechanism, and then the feature vector of the target node based on the meta-path level is obtained through normalization. The formula is as follows:

[0081] in, is the meta-path level attention parameter vector, and are the weight matrix and the offset value respectively, Based on The obtained linear change value is normalized to obtain It is used to represent the contribution value of neighbor nodes on different meta-paths to the target node. and Perform product operation to get the embedded node level vector of the target node

[0082] (3) The obtained node-level attention embedding vector and meta-path-level attention embedding vector are concatenated as the final embedding representation of the learner’s heterogeneous information network. The formula is as follows:

[0083] Step S6: Based on the learner fusion features, the GRU-based knowledge tracking model is used to predict the learning status of each learner.

[0084] Specifically, as shown in Figure 2, first, the original data answering interaction records of the DKT model (q t ,a t ) (i.e., historical data of learners’ interactive sequences of answering questions) and the embedding vector e of the learner’s heterogeneous information network, and the concatenated triple vector c is obtained. t The formula is as follows:

[0085] c t =Concat(q t ,a t ,e) (25)

[0086] Then, the data is passed to the output layer through the calculation of the GRU neural network, and the output y t is a numerical value that predicts the probability that the learner will answer the exercise correctly. The formula is as follows:

[0087] in, is the output weight matrix, the bias term are the model parameters to be trained.

[0088] In this embodiment, the construction of the knowledge tracking model based on GRU adopts the cross entropy loss function. t You can get the next question q t+1 The accuracy rate y t+1 , and the true label a t+1 Calculate the cross entropy function and the loss function of the entire sequence of questions for a single learner, as shown in the following formula:

[0089] The gradients of all learnable parameters are calculated according to the above loss function values, and the learnable parameters are updated until the learnable parameters converge iteratively on the sample data set, thereby obtaining a GRU-based knowledge tracking model.

[0090] Furthermore, this embodiment also provides a learner learning situation prediction system based on meta-path, as shown in Figure 3. The system includes a signal-connected data acquisition module, a feature engineering module, a feature fusion module, and a prediction module. The data acquisition module is used to acquire learner data, exercise data, and learner answer interaction sequence data; the feature engineering module is used to construct a learner behavior graph based on the learner answer interaction sequence data, and use a self-organizing map neural network to extract each learner's learning content preference from the learner behavior graph. At the same time, based on the learner data, exercise data, and learner answer interaction sequence data, the learning ability of each learner is calculated; the feature fusion module is used to fuse the features of learning content preference and learning ability to construct a learner heterogeneous information network based on meta-path. The meta-path is selected by analyzing the relationship between learning content preference and learning ability, and uses a hierarchical graph attention mechanism to embed learning into the learner heterogeneous information network and extract learner fusion features; the prediction module is used to predict the learning situation of each learner based on the learner fusion features using a GRU-based knowledge tracking model. The specific workflow of each module is the same as the implementation process of the aforementioned learner learning situation prediction method based on meta-path, and will not be repeated here.

[0091] Furthermore, this embodiment provides an electronic device. The electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the program to implement the aforementioned meta-path-based learner learning status prediction method. The electronic device may also include a communication interface for communication between the memory and the processor. The memory is used to store the computer program executable on the processor. The memory may include high-speed RAM (Random Access Memory) memory, or may also include non-volatile memory, such as at least one disk storage device. If the memory, processor, and communication interface are implemented independently, the communication interface, memory, and processor may be interconnected via a bus and communicate with each other. The bus may be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, among others. Buses may be classified as address buses, data buses, control buses, and the like. Optionally, in a specific implementation, if the memory, processor, and communication interface are integrated on a single chip, the memory, processor, and communication interface may communicate with each other via an internal interface. The processor may be a CPU (Central Processing Unit), or an ASIC (Application Specific Integrated Circuit), or one or more integrated circuits configured to implement the embodiments of the present application.

[0092] The above describes in detail the preferred embodiments of the present invention. It should be understood that those skilled in the art can make numerous modifications and variations based on the concepts of the present invention without inventive effort. Therefore, any technical solutions that can be derived by those skilled in the art through logical analysis, reasoning, or limited experimentation based on the concepts of the present invention and the prior art should be within the scope of protection defined by the claims.

Claims

1. A method for predicting the learning situation of learners based on meta-paths, characterized in that, It includes the following steps: S1. Obtain learner data, exercise set data, and learner answer interaction sequence data; S2. Construct a learner behavior graph based on the learner answer interaction sequence data, and use a self-organizing mapping neural network to extract the learning content preferences of each learner from the learner behavior graph; S3. Calculate the learning ability of each learner based on the learner data, the exercise set data, and the learner answer interaction sequence data; S4. Integrate the features of the learning content preferences and the features of the learning ability to construct a learner heterogeneous information network based on a meta-path, and the meta-path is selected by analyzing the relationship between the learning content preferences and the learning ability; S5. Use a hierarchical graph attention mechanism to embed and learn the learner heterogeneous information network, and extract learner fusion features; S6. Based on the learner fusion features, use a GRU-based knowledge tracing model to predict the learning situation of each learner.

2. The method for predicting the learning situation of a learner based on a meta-path according to claim 1, wherein The exercise set data is obtained by taking the vector product of the knowledge point vector and the corresponding exercise difficulty coefficient.

3. The method for predicting the learning situation of a learner based on a meta-path according to claim 1, wherein The learner answer interaction sequence is composed of multiple groups of corresponding exercise vectors and answer results in sequence.

4. The method for predicting the learning situation of a learner based on a meta-path according to claim 1, wherein The process of using a self-organizing mapping neural network to extract the learning content preferences of each learner from the learner behavior graph includes: Based on the learner behavior graph, use multi-head graph attention and mean aggregation to obtain the behavior graph embedding vector of each learner; Based on the behavior graph embedding vector, use the self-organizing mapping neural network to cluster the behaviors of each learner into different learning content preference categories, obtain the learner learning content preference categories, and add the learning content preference categories to the learner features.

5. The method for predicting the learning situation of a learner based on a meta-path according to claim 1, characterized in that The learning ability is obtained by calculating the difference between the proportion of correctly answered exercises and the proportion of wrongly answered exercises.

6. The method for predicting the learning situation of a learner based on a meta-path according to claim 1, wherein The meta-path includes the learner-learning content preference-learner meta-path and the learner-learning ability-learner meta-path.

7. The method for predicting the learning situation of a learner based on a meta-path according to claim 1, wherein The hierarchical graph attention mechanism includes a node-level attention mechanism and a meta-path-level attention mechanism.

8. The method for predicting the learning situation of a learner based on a meta-path according to claim 1, characterized in that The specific steps of S6 include: Concatenate the learner fusion features with the historical data of the learner answer interaction sequence into a triple vector; Initialize the GRU-based knowledge tracing model; Based on the triple vector, use the initialized GRU-based knowledge tracing model to obtain the prediction results of the learning situation of each learner.

9. A learner learning situation prediction system based on a meta-path, characterized in that, The system includes a data acquisition module, a feature engineering module, a feature fusion module, and a prediction module that are signal-connected. The data acquisition module is used to obtain learner data, exercise set data, and learner answer interaction sequence data; The feature engineering module is used to construct a learner behavior graph based on the learner answer interaction sequence data, and use a self- organizing mapping neural network to extract the learning content preferences of each learner from the learner behavior graph, and at the same time calculate the learning ability of each learner based on the learner data, the exercise set data, and the learner answer interaction sequence data. The feature fusion module is used to fuse the features of the learning content preference and the features of the learning ability, construct a learner heterogeneous information network based on a meta-path, the meta-path is selected by analyzing the relationship between the learning content preference and the learning ability, and a hierarchical graph attention mechanism is used to perform embedded learning on the learner heterogeneous information network to extract learner fusion features; The prediction module is used to predict the learning situation of each learner based on the learner fusion features by using a knowledge tracing model based on GRU.

10. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored on the memory and executable on the processor, and the processor executes the program to implement the method according to any one of claims 1-8.

Images