Personalized learning path recommendation method based on knowledge relationship mining and graph embedding driving

Abstract

The application belongs to the technical field of recommendation algorithm, and discloses a personalized learning path recommendation method based on knowledge relationship mining and graph embedding driving, and the specific steps are as follows: step one: mining knowledge point dependency relationship and constructing a weighted directed graph, data analysis is carried out on the public data set ASSISTments and Junyi, the hidden dependency relationship between knowledge points is mined through the improved Apriori algorithm, the time dynamics and sequential dependency characteristics in the learning process are fully considered, and the weighted directed graph between knowledge points is constructed according to the dependency relationship and the weight value obtained by mining; the application mines the hidden dependency relationship between knowledge points and constructs a weighted directed graph through an innovative method, combines a unique self-defined embedding layer to integrate the graph embedding technology into the recommendation model, compared with the existing recommendation method based on simple association rules or shallow network, the mining and expression ability of the complex logical relationship between knowledge points is significantly improved, and the recommendation omission rate of sparse knowledge points can be effectively reduced.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of recommendation algorithm, and particularly relates to a personalized learning path recommendation method based on knowledge relationship mining and graph embedding driving. BACKGROUND

[0002] In the current learning path recommendation algorithm research, there are various technical means to realize personalized learning path design, among which, a common method is to generate a recommended path based on a knowledge graph through pre-set constraint rules, and this method uses the logical relationship between knowledge points to provide a structured learning order for students; another widely used technology is to predict the concept that a student should learn next based on the student's historical learning sequence combined with a long short-term memory network (LSTM), and this method attempts to dynamically infer the learning needs of the student by analyzing the student's learning behavior pattern; in recent years, knowledge tracking technology has been gradually introduced into the learning path recommendation field, which can assess the student's mastery of the current recommended knowledge point in real time and adjust the subsequent learning path accordingly; many studies have attempted to combine knowledge tracking and learning path recommendation models in order to more accurately reflect the student's learning state and optimize the recommendation effect.

[0003] However, the above methods have the following problems: the learning path recommendation method based on the knowledge point graph excessively relies on pre-defined knowledge constraints, making it difficult to flexibly adapt to the personalized needs of learners and limiting its application effect in diversified learning scenarios; the deep learning-based method (such as LSTM and GRU) only considers the learning preferences of students, ignoring the relationship between knowledge points, and has high computational complexity, affecting large-scale real-time applications; some methods use Transformer to explore the relationship between knowledge points and integrate DKT knowledge tracking modules to predict the student's mastery of concepts, but the logical relationship of knowledge points is complex, and it is difficult to reveal its deep cognitive structure relying on the original attention mechanism and neural network, therefore, knowledge graph and graph embedding technology need to be introduced to enhance the exploration ability of the relationship between knowledge points. SUMMARY

[0004] The purpose of the present application is to provide a personalized learning path recommendation method based on knowledge relationship mining and graph embedding driving to solve the problems raised in the background art.

[0005] In order to achieve the above purpose, the present application provides the following technical solution: a personalized learning path recommendation method based on knowledge relationship mining and graph embedding driving, the specific steps are as follows:

[0006] Step 1: Mining knowledge point dependency relationship and constructing a weighted directed graph

[0007] Data analysis was performed on the public datasets ASSISTments and Junyi. The improved Apriori algorithm was used to mine the implicit dependency relationships between knowledge points, fully considering the time dynamics and sequential dependency characteristics in the learning process. Based on the mined dependency relationships and their weights, a weighted directed graph between knowledge points was constructed to accurately represent the knowledge structure of the learning domain.

[0008] Step two: generating knowledge point vector representation based on graph embedding

[0009] Based on the weighted directed graph constructed in step one, the DGI algorithm was used for graph embedding processing to generate dense vector representations for each knowledge point. These vector embeddings effectively capture the structural properties and relationship characteristics of knowledge points in the learning domain, laying the foundation for subsequent processing.

[0010] Step three: knowledge graph enhanced neural network embedding and Transformer processing

[0011] Using the dense vector representations of knowledge points generated in step two, a neural network embedding layer was designed to assign each knowledge point a unique integer index and map it to a low-dimensional vector representation, forming a semantic representation. Subsequently, these vectors were input into the Transformer architecture for processing through an optimized attention mechanism, capturing deep semantic associations and cognitive structures between knowledge points, and exploring complex knowledge dependency patterns and hierarchical relationships.

[0012] Step four: design and generation of personalized learning path recommendation model

[0013] A candidate knowledge point set S = [s1, s2, s3, …, sm] (m < n) was defined and mapped to an embedding vector set X = [x1, x2, x3, …, xm] through an embedding layer to obtain low-dimensional semantic representations of knowledge points. An improved Transformer architecture was used to further mine deep associations between knowledge points, and the attention mechanism was optimized to improve sensitivity to sparse or weakly associated knowledge points. Then, a long short-term memory network (LSTM) was designed to model the student's historical answer sequence, dynamically capturing the temporal dependency of learning behavior. The knowledge point vectors output by the Transformer, the long short-term memory network (LSTM) state vectors, and the predefined learning goals were input into a multilayer perceptron (MLP), and through joint optimization, precise concept selection was achieved. Finally, a personalized learning path was generated that takes into account both knowledge dependency and individual differences.

[0014] As a preferred technical solution of the present application, the Apriori algorithm in step one introduces a minimum confidence threshold as a sequence constraint condition, for a knowledge point pair C1→C2, only when its confidence exceeds the preset threshold, a directed association is established, and through confidence constraint, the directionality of the knowledge point association is ensured to conform to the learning and cognitive rules; in order to capture the time proximity and statistical significance in the knowledge point association, a double weighting method combining a sliding window mechanism and a confidence value is introduced, for a sequence of length n, a fixed window length w (w<n) is defined to analyze the local relationship, and the confidence score of each valid pair is calculated.

[0015] As a preferred technical solution of the present application, the maximum weight of the knowledge point association is determined by two factors:

[0016] Distance-based weight:

[0017]

[0018] Where d represents the distance between points within the window length w, only considering the point pairs within the window length w;

[0019] Confidence-based weight:

[0020] weight c (C1→C2)=confidence(C1→C2)=P(C2|C1)

[0021] =support(C1∪C2) / support(C1)

[0022] Reflects the statistical strength of the association;

[0023] Then the final edge weight calculation formula is:

[0024] weight final (C1→C2)=α*weight d *weight c

[0025] Where α is a normalization factor to ensure that the weight falls within the range [0, 1].

[0026] As a preferred technical solution of the present application, the DGI algorithm in step two maximizes the mutual information between the global graph representation and the local node representation to realize unsupervised knowledge point learning, and the purpose of graph embedding in step two is to ensure that adjacent nodes in the graph have closer representations in the embedding space. In the process of graph embedding, the target embedding vector representation is learned by maximizing the objective function:

[0027]

[0028] where N i represents the neighborhood node set of knowledge point e i , σ(·) is a sigmoid function, e vi and e vk are embedding vectors of knowledge points V i and V k .

[0029] As a preferred technical solution of the present application, the knowledge graph enhanced neural network in step three enhances the attention to sparse knowledge points by adjusting the weight distribution in the attention mechanism, that is, using relative position encoding to fuse the weight between knowledge points in the knowledge graph into the attention mechanism of the Transformer:

[0030]

[0031] To emphasize the importance of knowledge points in the attention mechanism, the weight of the knowledge point graph is introduced into the calculation of the attention score, where R i,j is the relationship embedding between knowledge points i and j in the knowledge point graph:

[0032]

[0033] α i,j =SoftMax(Score i,j )

[0034] Q i =X i W Q , K i =X i W K , V i =X i W V

[0035] W Q , W K , W V are trainable matrices, d is the embedding layer vector dimension, is the vector representation obtained by the knowledge point through the attention mechanism layer, which is used for subsequent tasks of path recommendation.

[0036] As a preferred technical solution of the present application, the personalized learning path recommendation model in step four has three main structures of Transformer, long short-term memory network (LSTM) and a custom embedding layer, Transformer is used to explore the knowledge structure, long short-term memory network (LSTM) is used for the memory of the learning sequence of the student, and the embedding layer is made of all the knowledge point vectors C = [Cv1, Cv2, Cv3, Cv4,..., Cvn] obtained above, when the long short-term memory network (LSTM) performs state memory and update, the initial state is the historical learning concept H = [H1, H2,..., Ht] of the student, and H obtains vector representation through the embedding layer stsates0 represents the initial state:

[0037]

[0038] Since the candidate set S is screened and sorted to obtain a suitable path recommendation for the student, the generated path is π = {π1, π2,..., πm}, it is assumed that the current recommended step i is reached (i <= m), at this time π = {π1, π2,..., πi-1}, then the current long short-term memory network (LSTM) state is:

[0039]

[0040] As a preferred technical solution of the present application, the pre-defined learning goal input multi-layer perceptron (MLP) layer in step four generates the probability of the knowledge point after inputting the learning goal T = {t1, t2,...}, the embedding vector of the candidate set knowledge point and the state of the long short-term memory network (LSTM):

[0041]

[0042] out = softmax(prob)

[0043] P (πi) = argmax (out)

[0044] Wherein, represents the vector representation of the target concept obtained through Embedding, w1, W1, W2 are learnable matrices, softmax is an activation function, and argmax is to pick out the maximum probability from m probabilities.

[0045] As a preferred technical solution of the present application, the personalized learning path recommendation model in step four is used to obtain instant feedback when training the knowledge tracking auxiliary module (KT) as the mastery degree of the student to the concept, and is defined as:

[0046]

[0047] wherein E e and E b represent the mastery level of the student on the target concept before and after the path π, E sup represents the upper limit of the mastery level.

[0048] As a preferred technical solution of the present application, the E T As the reward in reinforcement learning, the policy gradient loss is defined as:

[0049]

[0050] The beneficial effects of the present application are as follows:

[0051] 1. The present application significantly improves the mining and expression ability of complex logical relationships between knowledge points by innovatively mining the implicit dependency relationship between knowledge points and constructing a weighted directed graph, and combining a unique custom embedding layer to integrate graph embedding technology into the recommendation model. Compared with existing recommendation methods based on simple association rules or shallow networks, the present application can effectively reduce the recommendation omission rate of sparse knowledge points, especially in learning scenarios with complex knowledge structures, and exhibits higher applicability and accuracy.

[0052] 2. The present application highly integrates the graph embedding layer with the deep learning network, generates low-dimensional semantic representations of knowledge points through the custom embedding layer, and accurately captures the time dependence of the student learning sequence in combination with the long short-term memory network (LSTM). Compared with existing static recommendation models or path planning relying only on rules, the present application can dynamically optimize the priority and content of the learning path, significantly improving the accuracy and adaptive ability of personalized recommendation, and overcoming the limitations of traditional methods in supporting the dynamic nature of student behavior.

[0053] 3. The present application innovatively integrates knowledge point graph constraints in the Transformer architecture, designs a knowledge graph enhanced attention module, and embeds the knowledge point relationship weight generated by the graph embedding into the attention mechanism. Compared with existing recommendation schemes lacking knowledge structure constraints or dynamic feedback, the present application can ensure the high consistency of the recommended path with the logical dependency of knowledge points in real time, and realize dynamic adjustment of the path through reinforcement learning and knowledge tracking feedback, significantly enhancing the adaptability to changes in student learning behavior. BRIEF DESCRIPTION OF DRAWINGS

[0054] Fig. 1 is a model structure diagram of the present application;

[0055] Fig. 2 is a weighted directed graph of the present application. DETAILED DESCRIPTION

[0056] With reference to the drawings and embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of the present application.

[0057] As shown in the drawings, the embodiments of the present application provide a personalized learning path recommendation method based on knowledge relationship mining and graph embedding driving, and the specific steps are as follows: Figs. 1-2

[0058] Step one: mining knowledge point dependency relationship and constructing weighted directed graph

[0059] Data analysis is performed on the public data sets ASSISTments and Junyi, the implicit dependency relationship between knowledge points is mined by the improved Apriori algorithm, the time dynamics and sequential dependency characteristics in the learning process are fully considered, the weighted directed graph between knowledge points is constructed according to the dependency relationship and its weight value obtained by mining, so as to accurately represent the knowledge structure of the learning field;

[0060] Step two: generating knowledge point vector representation based on graph embedding

[0061] Based on the constructed weighted directed graph in step, the DGI algorithm is used for graph embedding processing, and dense vector representation is generated for each knowledge point. These vector embeddings effectively capture the structural properties and relationship characteristics of knowledge points in the learning field, laying a foundation for subsequent processing;

[0062] Step three: knowledge graph enhanced neural network embedding and Transformer processing

[0063] Using the knowledge point dense vector generated in step two, a neural network embedding layer is designed to assign a unique integer index to each knowledge point and map it to a low-dimensional vector representation, forming a semantic representation. Then, these vectors are input into the Transformer architecture, which processes them through an optimized attention mechanism to capture deep semantic associations and cognitive structures between knowledge points, explore complex knowledge dependency patterns and hierarchical relationships;

[0064] Step four: design and generate personalized learning path recommendation model

[0065] ​The candidate knowledge point set S is defined as [s1, s2, s3, …, sm] (m<n), which is mapped to the embedding vector set X=[x1, x2, x3, …, xm] through an embedding layer, to obtain the low-dimensional semantic representation of the knowledge points, an improved Transformer architecture is further used to mine the deep correlation between the knowledge points, the attention mechanism is optimized to improve the sensitivity to sparse or weakly correlated knowledge points, then a long short-term memory network (LSTM) is designed to model the student's historical answer sequence, and the time dependence of the learning behavior is dynamically captured, the knowledge point vector output by the Transformer, the long short-term memory network (LSTM) state vector and the predefined learning goal are input into a multilayer perceptron (MLP), and precise concept selection is realized through joint optimization, and finally a personalized learning path is generated, which takes into account the knowledge dependency and individual differences.

[0066] The application fuses knowledge graph embedding knowledge relationship mining, graph enhanced Transformer structure, long short-term memory network (LSTM) and path recommendation of reinforcement learning, and dynamically optimizes the knowledge tracking module, and constructs a precise personalized learning path recommendation model, which not only effectively captures the cognitive association and structural dependence between knowledge points, but also dynamically adjusts the recommended path according to the basic differences and real-time feedback of students, thereby significantly improving the individualization accuracy and adaptability of the learning path, and providing more optimized learning support for students with different bases, Fig. 1 The initial candidate set S forms the vector representation of the knowledge points through a self-defined embedding layer, the Transformer mines the relationship between the knowledge points, and finally generates the final vector representation of the knowledge points, in the decoder module, H represents the historical learning sequence of the student, T represents the target concept to be learned, and the predefined learning goal is input into a multilayer perceptron (MLP) to generate and sort the learning path pi based on the combination of the above information.

[0067] Among them, the Apriori algorithm in step one introduces a minimum confidence threshold as a sequence constraint condition, for a knowledge point pair C1→C2, only when its confidence exceeds the preset threshold, a directed association is established, and the directionality of the knowledge point association is ensured to conform to the learning cognitive law through confidence constraint; in order to capture the time proximity and statistical significance in the knowledge point association, a double weighting method combining a sliding window mechanism and a confidence value is introduced, for a sequence of length n, a fixed window length w (w<n) is defined to analyze the local relationship, and the confidence score of each valid pair is calculated.

[0068] The common Apriori algorithm cannot fully explore the logical relationship of deep knowledge points, but the introduction of the minimum confidence threshold as a sequence constraint condition and the introduction of the double weighting method combining the sliding window mechanism and the confidence value can realize the full exploration of the logical relationship of deep knowledge points. At the same time, for the generated knowledge points, whether they meet the logical relationship of knowledge points is observed in detail through manual inspection, so as to ensure the effectiveness of the knowledge points.

[0069] Among them, the maximum weight of the knowledge point association is determined by two factors:

[0070] Distance-based weight:

[0071]

[0072] Among them, d represents the distance between points within the window length w, only considering the point pair within the window length w;

[0073] Confidence-based weight:

[0074] weight c (C1→C2)=confidence(C1→C2)=P(C2|C1)

[0075] =support(C1∪C2) / support(C1)

[0076] Reflects the statistical strength of the association;

[0077] Then the final edge weight calculation formula is:

[0078] weight final (C1→C2)=α*weight d *weight c

[0079] Among them, α is a normalization factor to ensure that the weight falls within the range [0, 1].

[0080] Through the above method, the logical relationship between knowledge points can be determined, and with this logical relationship, the construction of the weighted directed graph can be carried out, such as Fig. 2 The knowledge point relationship graph G(V, E) in the figure, where the node set V represents all knowledge points, and the edge set E represents the association between these knowledge points. The weight of the edge reflects the relationship strength between them.

[0081] Among them, the DGI algorithm in step two realizes unsupervised knowledge point learning by maximizing the mutual information between global graph representation and local node representation. The purpose of graph embedding in step two is to ensure that adjacent nodes in the graph have closer representations in the embedding space. In the process of graph embedding, the target embedding vector representation is learned by maximizing the objective function:

[0082]

[0083] where N i represents the neighborhood node set of knowledge point e i , σ(·) is the sigmoid function, e vi and e vk are the embedding vectors of knowledge points V i and V k .

[0084] The DGI algorithm can effectively capture the topological structure and weight information between knowledge points, while maintaining good interpretability, and these vectors provide rich representative information for subsequent learning and reasoning tasks; and the method of learning the target embedding vector representation by maximizing the objective function ensures that knowledge points that are related to each other have close representations in the embedding space, so that the model can better capture the implicit relationships between knowledge points. After obtaining the vector representation of the knowledge points by the above method, the vectors are embedded into the personalized learning path recommendation model.

[0085] wherein the knowledge graph enhanced neural network in step three enhances the attention to sparse knowledge points by adjusting the weight distribution in the attention mechanism, i.e. using relative position encoding to fuse the weight between knowledge points in the knowledge graph into the attention mechanism of the Transformer:

[0086]

[0087] To emphasize the importance of knowledge points in the attention mechanism, the weight of the knowledge graph is introduced into the calculation of the attention score, where R i,j is the relationship embedding between knowledge points i and j in the knowledge graph:

[0088]

[0089] α i,j =SoftMax(Score i,j )

[0090] Q i =X i W Q , K i =X i W K , V i =X i W V

[0091] W Q , W K , W Vis trainable matrix, d is embedding layer vector dimension, is the vector representation of knowledge points obtained through the attention mechanism layer, which is used for subsequent tasks of path recommendation.

[0092] Fig. 2 As can be seen from the above table, Through the T-SNE dimension reduction method, it can be clearly seen that the knowledge points with close logical relationships are well aggregated together in the two-dimensional coordinate system.

[0093] Among them, the personalized learning path recommendation model in step four has three main structures: Transformer, long short-term memory network (LSTM) and custom embedding layer. Transformer is used to explore the knowledge structure, long short-term memory network (LSTM) is used for the memory of the student's learning sequence, and the embedding layer is made of all the knowledge point vectors C = [Cv1, Cv2, Cv3, Cv4, …, Cvn] obtained above. When the long short-term memory network (LSTM) performs state memory and update, the initial state is the student's historical learning concept H = [H1, H2, …Ht], and H obtains vector representation after embedding layer states0 represents the initial state:

[0094]

[0095] Since the candidate set S is filtered and sorted to get the appropriate path recommendation to the student, the generated path is π = {π1, π2, …, πm}, assuming that the current recommendation is the i-th step (i <= m), at this time π = {π1, π2, …, πi-1}, then the current long short-term memory network (LSTM) state is:

[0096]

[0097] Each step of the recommended concept needs to be added to the long short-term memory network (LSTM) network for state update to select the next appropriate concept.

[0098] Among them, the pre-defined learning goal input multi-layer perception (MLP) layer in step four inputs the learning goal T = {t1, t2, …} and the embedding vector of the candidate set knowledge point And the state of the long short-term memory network (LSTM) generates the probability of the knowledge point as:

[0099]

[0100] out = softmax(prob)

[0101] P(πi) = argmax(out)

[0102] wherein, denotes the vector representation of the target concept through Embedding, w1, W1, W2 are learnable matrices, softmax is the activation function, and argmax is to pick out the maximum probability from m probabilities.

[0103] The previously recommended concepts will be set to 0 at the corresponding position to avoid path duplication.

[0104] wherein, the personalized learning path recommendation model in step four is used to obtain real-time feedback, and the knowledge tracking auxiliary module (KT) is trained as the mastery of the concept by the student, and is defined as:

[0105]

[0106] wherein, E e and E b represent the mastery of the target concept before and after the path π by the student, and E sup represents the upper limit of the mastery.

[0107] E e and E b are the scores obtained through the knowledge tracking auxiliary module (KT) as the examination environment.

[0108] wherein, E T as the reward in reinforcement learning, the policy gradient loss is defined as:

[0109]

[0110] Through the knowledge tracking auxiliary module (KT), the feedback of the student on the concept is introduced in real time, and the adaptability and accuracy of the recommended path are dynamically optimized, so as to provide more accurate personalized learning path recommendation for students with different foundations.

[0111] It should be noted that, in this article, relational terms such as first and second are used merely to distinguish one entity or action from another entity or action, without necessarily requiring or implying any such actual relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variant thereof are intended to cover non-exclusive inclusions, so that a process, method, article, or apparatus that includes a list of elements does not only include those elements, but also includes other elements not explicitly listed, or inherent to such a process, method, article, or apparatus.

[0112] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary of the principles and application of the present application. Numerous modifications and adaptions can be effected without departing from the spirit and scope of the present application, which is not limited to the exact construction and arrangement described. It is intended, therefore, to cover all modifications and adaptions that fall within the scope of the claims and their equivalents.

Claims

1. A personalized learning path recommendation method based on knowledge relationship mining and graph embedding driving, characterized in that, The specific steps are as follows: Step one: excavate the dependency relationship of knowledge points and construct a weighted directed graph Data analysis is performed on the public data sets ASSISTments and Junyi. The improved Apriori algorithm is used to mine the implicit dependency relationship between knowledge points, fully considering the time dynamics and sequential dependency characteristics in the learning process. According to the dependency relationship and its weight value obtained by mining, a weighted directed graph between knowledge points is constructed to accurately represent the knowledge structure of the learning field. Step two: generate knowledge point vector representation based on graph embedding Based on the weighted directed graph constructed in step one, the DGI algorithm is used for graph embedding processing to generate dense vector representation for each knowledge point. These vector embeddings effectively capture the structural properties and relationship characteristics of knowledge points in the learning field, laying the foundation for subsequent processing. Step three: knowledge graph enhanced neural network embedding and Transformer processing Using the dense vector of knowledge points generated in step two, a neural network embedding layer is designed to assign a unique integer index to each knowledge point and map it to a low-dimensional vector representation, forming a semantic representation. Then, these vectors are input into the Transformer architecture for processing through the optimized attention mechanism, capturing deep semantic associations and cognitive structures between knowledge points, and exploring complex knowledge dependency patterns and hierarchical relationships. Step four: design and generate personalized learning path recommendation model Define the candidate knowledge point set S = [s1, s2, s3, …, sm], map it to the embedding vector set X = [x1, x2, x3, …, xm] through the embedding layer to obtain the low-dimensional semantic representation of the knowledge points. Use the improved Transformer architecture to further mine the deep association between knowledge points, optimize the attention mechanism to improve the sensitivity to sparse or weakly associated knowledge points. Then, design a long short-term memory network to model the student's historical answer sequence, dynamically capturing the temporal dependency of learning behavior. Input the knowledge point vector output by the Transformer, the long short-term memory network state vector, and the predefined learning goal into a multilayer perceptron to achieve precise concept selection through joint optimization. Finally, generate a personalized learning path that takes into account knowledge dependency and individual differences. 2.The knowledge relationship based mining and graph embedding driven personalized learning path recommendation method according to claim 1, characterized in that: The Apriori algorithm described in step one introduces a minimum confidence threshold as a sequence constraint condition. For a knowledge point pair C1→C2, only when its confidence exceeds the preset threshold, a directed association is established. By confidence constraint, the directionality of knowledge point association is ensured to comply with the learning and cognitive rules. To capture the temporal proximity and statistical significance in knowledge point association, a double weighting method combining sliding window mechanism and confidence value is introduced. For a sequence of length n, a fixed window length w is defined to analyze local relationships, and the confidence score of each valid pair is calculated. 3.The knowledge relationship based mining and graph embedding driven personalized learning path recommendation method according to claim 2, characterized in that: The maximum weight of the knowledge point association is determined by two factors: where d represents the distance between points within the window length w, only considering point pairs within the window length w. Confidence-based weights: where d denotes the distance between points within the window length w, only considering pairs of points within the window length w. Confidence-based weights: Reflects statistical strength of association; The final edge weight calculation formula is: wherein, is a normalization factor that ensures the weight falls in the range [0, 1]. 4.The knowledge relationship based mining and graph embedding driven personalized learning path recommendation method according to claim 1, characterized in that: The DGI algorithm described in step two achieves unsupervised knowledge point learning by maximizing the mutual information between global graph representation and local node representation, and the purpose of the graph embedding described in step two is to ensure that adjacent nodes in the graph have closer representations in the embedding space. In the process of graph embedding, the target embedding vector representation is learned by maximizing the objective function: wherein, represents a knowledge point a set of neighborhood nodes of the knowledge point, is a function, and is an embedding vector of the knowledge point and . 5.The knowledge relationship based mining and graph embedding driven personalized learning path recommendation method according to claim 1, characterized in that: The knowledge graph enhanced neural network described in step three enhances the focus on sparse knowledge points by adjusting the weight distribution in the attention mechanism, i.e. using relative position encoding to fuse the weights between knowledge points in the knowledge graph into the attention mechanism of the Transformer: To emphasize the importance of knowledge points in the attention mechanism, the weights of the knowledge point graph are introduced into the calculation of the attention score, where, is the relationship embedding between knowledge points i and j in the knowledge point graph: : : : is a trainable matrix, d is the embedding layer vector dimension, is the vector representation obtained by the knowledge point through the attention mechanism layer, which is used for subsequent tasks of path recommendation. 6.The knowledge relationship based mining and graph embedding driven personalized learning path recommendation method according to claim 1, characterized in that: The personalized learning path recommendation model described in step four has three main structures: Transformer, long short-term memory network and custom embedding layer. Transformer is used to explore the knowledge structure, long short-term memory network is used for the memory of the learning sequence of students, and embedding layer is made of all knowledge point vectors C = [Cv1, Cv2, Cv3, Cv4,...,Cvn] obtained above. When the long short-term memory network performs state memory and update, the initial state is the historical learning concept of the student H = [H1, H2,...,Ht], and H obtains vector representation after passing through the embedding layer , The initial state is represented as: Since the candidate set S is screened and sorted to recommend a suitable path to the student, the generated path is π = {π1, π2,..., πm}. Assuming that the current recommendation is to the i-th step, there is π = {π1, π2,..., πi-1} at this time, and the current long short-term memory network state is: . 7.The knowledge relationship based mining and graph embedding driven personalized learning path recommendation method according to claim 1, characterized in that: The pre-defined learning target input multi-layer perception layer in step four inputs the embedding vector of the candidate set knowledge point , and the state of the long short-term memory network : : : , wherein represents the vector representation of the target concept obtained through Embedding, W1, W1, and W2 are learnable matrices, is an activation function, is the probability with the largest probability selected from the m probabilities. 8.The knowledge relationship based mining and graph embedding driven personalized learning path recommendation method according to claim 1, characterized in that: The personalized learning path recommendation model described in step four, when used, to obtain immediate feedback, train a knowledge tracking assistance module as the mastery level of the student on the concept, then define: Where E e和 represents the mastery level of the student on the path of the target concept, represents the upper limit of the mastery level. 9.The knowledge relationship based mining and graph embedding driven personalized learning path recommendation method according to claim 8, characterized in that: The As a reward in reinforcement learning, the policy gradient loss is defined as: .

Images