Resource learning path planning method and device based on deep learning

Abstract

The invention discloses a resource learning path planning method and device based on deep learning, and the method comprises the steps: obtaining the knowledge levels and historical record information of a plurality of learners, building a learner model according to the historical record information and the knowledge levels, calculating the learner similarity between any two different learners in the learner model, and obtaining the learning path of the learners. Constructing a learning behavior model corresponding to the learner according to the learner similarity, predicting the cognitive level of the learner according to the learning behavior model and the learner model, constructing a learning path recommendation model according to the cognitive level, and planning a target learning path corresponding to the learner according to the learning path recommendation model and a preset knowledge graph, the learning behavior indexes are quantified according to the learning behaviors of the learner to construct the learner model, the learning path is recommended to the learner, the knowledge graph is combined to further complete accurate learning path planning, and a practical, reasonable and explainable learning path is provided for the learner, so that the learning efficiency and quality of the learner are improved.

Description

Technical Field

[0001] The present invention belongs to the field of education, and in particular relates to a resource learning path planning method and device based on deep learning. Background Art

[0002] The rapid development of information technology is the main driving force promoting the development of the online learning field. Artificial intelligence is deeply changing people's learning methods, and network-based learning has been rapidly developed and widely applied. A learning path refers to an ordered sequence composed of courses and knowledge points that learners learn during the learning process. When learners set learning goals, they need to formulate corresponding learning plans to complete learning tasks in a certain order. Due to certain differences in the characteristics of different courses, corresponding teaching strategies will also be adopted for different course teaching tasks to assist specific teaching tasks to help learners achieve learning goals. Usually, a learning path can be the learning order of multiple related courses formulated by learners to systematically learn a certain aspect of knowledge, or the learning order of multiple related knowledge points formulated by learners to learn a certain course. In the education field, the relationship between courses and courses, and between courses and knowledge points is relatively few, so it can be achieved through manual marking. However, with the popularization of Internet education, the number of questions in online education platforms is increasing continuously, but the relationship between questions and courses and knowledge points is relatively sparse. Many questions only have the question surface, but do not have the knowledge points involved, and cannot accurately and effectively evaluate the learning effect of learners, thus reducing the learning efficiency of learners. Summary of the Invention

[0003] In view of this, the present invention provides a resource learning path planning method and device based on deep learning that can take into account the different learning abilities and needs of different learners, improve the cognitive level of learners and enhance learning efficiency, to solve the above-mentioned existing technical problems, and specifically adopts the following technical solutions to achieve.

[0004] In a first aspect, the present invention provides a resource learning path planning method based on deep learning, including:

[0005] Obtaining the knowledge levels and historical record information of multiple learners, and constructing a learner model according to the historical record information and the knowledge levels, wherein the knowledge levels include knowledge point IDs and mastery degrees, and the historical record information includes learning sequences, learning resources, and homework data;

[0006] Calculate the learner similarity between any two different learners in the learner model, and construct the learning behavior model corresponding to the learner according to the learner similarity. Among them, each learner is expressed as a vector in the learner model, and the cosine similarity is used to evaluate the similarity between two vectors, that is, the cosine value of the included angle between the two learner vectors is used as the similarity score between learners. The expression of the similarity score is where u a represents the vector expression of learner a, and u b represents the vector expression of learner b;

[0007] Predict the cognitive level of the learner according to the learning behavior model and the learner model, and construct a learning path recommendation model according to the cognitive level. Among them, the learning path recommendation model includes constructing two state sets respectively represent whether the learner accepts the recommendation, C represents accepting the recommendation, represents not accepting the recommendation;

[0008] Based on the learning path recommendation model and the preset knowledge graph, plan the corresponding target learning path for the learner.

[0009] As a further improvement of the above technical solution, planning the corresponding target learning path for the learner based on the learning path recommendation model and the preset knowledge graph includes:

[0010] Obtain all knowledge points in the preset knowledge graph as the knowledge point set to be sorted, and preset an initial knowledge point KP0;

[0011] Export the list L from the database corresponding to the preset knowledge graph to determine the knowledge point relationship. Among them, the knowledge point relationship includes the inclusion relationship and parallel relationship between knowledge points. The record information of the list L is (KP m , R, KP n ) indicates that the relationship between knowledge point KP m and KP n is R. The knowledge point relationship includes knowledge point importance, difficulty, centrality, topological hierarchy attribute feature data, and knowledge point sorting index;

[0012] Generate a target learning path using the topological sorting algorithm according to the preset knowledge graph and knowledge point attribute feature data.

[0013] As a further improvement of the above technical solution, generating a target learning path using the topological sorting algorithm according to the preset knowledge graph and knowledge point attribute feature data includes:

[0014] Compare the target learning path with the preset expert path to obtain the path similarity. The expression for calculating the path similarity is where pathep , path at respectively represent the preset expert path and the target learning path, KP mat represents the number of learning paths where the preset expert path matches the target learning path, KP tot represents the total number of learning paths;

[0015] Use the learning path evaluation index fitness to test the quality of the target learning path. The calculation expression of the evaluation index fitness is where pen adj represents the number of learning paths that violate the adjacent principle between knowledge points, pen ord represents the number of learning paths that violate the prerequisite and successor principle between knowledge points. The fewer the number of learning paths that violate the rules, the smaller the fitness value, indicating that the quality of path generation is higher.

[0016] As a further improvement of the above technical solution, predicting the cognitive level of the learner according to the learning behavior model and the learner model includes:

[0017] Divide the cognitive level of the learner into three domains, the positive domain POS, the boundary domain BND, and the negative domain NEG. Determine the learner's cognitive level θ according to the item response theory. Among them, for the item response theory 2PL model, the expression using the item characteristic curve to describe the probability of the learner answering a preset question correctly is where f represents the difficulty coefficient of the question, f ∈ [-2, 2], d represents the constant 1.072, θ represents the ability value, that is, the cognitive level of the learner, p(θ) represents the probability of answering the test question correctly. The POS domain represents learners with a high cognitive level, the NEG domain represents learners with a low cognitive level, and the BND domain represents learners with a medium cognitive level;

[0018] Input the question answering result data in the obtained score matrix and the known question difficulty coefficient into the IRT model, and establish the expression of the maximum likelihood function of the ability parameter as where p i represents the probability of answering correctly obtained by the item response theory model function, y i represents the label of the learner's true answering situation in the score matrix, so as to obtain the expression of the logarithmic maximum likelihood function as Obtain the ability parameter θ by taking the derivative of the logarithmic maximum likelihood function.

[0019] As a further improvement of the above technical solution, the decision information given to each region is A = {α P , α B , α N}, representing the learning path recommendation information for learners with different cognitive levels, α P indicates that the learner has completed the learning task of the current knowledge point and accepts the learning recommendation of other knowledge points, α N indicates that the current cognitive level is low and returns to consolidate basic knowledge, α B indicates accepting exercise recommendations to improve cognitive level;

[0020] The expression of the learning utility function matrix is where λ PP , λ BP and λ NP represent the loss values corresponding to taking actions α P , α B and α N when the object belongs to C, λ PN , λ BN and λ NN represent the loss values corresponding to taking actions α P , α B and α N when the object does not belong to C, s(λ PP ), s(λ BP ) and s(λ NP ) represent the utility values corresponding to taking actions α P , α B and α N when the object belongs to C, s(λ PN ), s(λ BN ) and s(λ NN ) represent the utility values corresponding to taking actions α P , α B and α N when the object does not belong to C; s(λ PP ) ≥ s(λ BP ) ≥ s(λ NP ), indicating that in the state of the object , the utility of determining [x] as the positive domain is greater than the utility of determining it as the boundary domain, and further greater than the utility of determining it as the negative domain; s(λ PN ) ≤ s(λ BN ) ≤ s(λ NN ), indicating that in the state of the object , the utility of determining [x] as the negative domain is less than the utility of determining it as the boundary domain, and further less than the utility of determining it as the positive domain.

[0021] As a further improvement of the above technical solution, export the list L from the database corresponding to the preset knowledge graph to determine the knowledge point relationship, including:

[0022] The text resources are processed using a keyword extraction algorithm to obtain a knowledge point set and the importance of knowledge points, and its expression is where Ws(V i ) represents the weight of node V i , Ws(V j ) represents the weight of node V after the previous iteration j , ω ji represents the similarity between node V j and node V i , and g represents the damping coefficient;

[0023] According to the properties of the knowledge graph, the centrality of knowledge points is calculated using the ratio of the in-degree to the out-degree of knowledge points. The larger the ratio, the higher the centrality of the knowledge point. The expressions for the in-degree and out-degree of knowledge points are The calculation expression for centrality is where represents the ratio of the in-degree to the out-degree of node V i , Pre(V i ) represents the set of first-order predecessor knowledge points of knowledge point V i , and Suc(V i ) represents the set of first-order successor knowledge points of knowledge node V i ;

[0024] The above attribute values are used to calculate the knowledge point sorting index where the importance Imp kpi uses the weight value of node Ws(V i ) in the above expression to obtain the expression where Imp kpi , Diff kpi , Cent kpi and Tp kpi represent the importance, difficulty, centrality, and topological level of the knowledge point kpi respectively, and ω imp , ω diff , ω cent and ω tp represent the assigned attribute weight values.

[0025] As a further improvement of the above technical solution, a learning path recommendation model is constructed according to the cognitive level, including:

[0026] Preset T = [i1, i2... i m , where T represents the knowledge point set and i represents a single knowledge point, then where L represents the learner path sequence and ls represents a single learner, Represents the learning path of learner ls1.

[0027] As a further improvement of the above technical solution, relevant knowledge points are tracked through the learner's questions and the questions are classified. Preset i' = 1, 2... M to represent M question texts in the set, and j′ = 1, 2... N to represent N pre-set knowledge points. The classification matrix is W = w i′j′ , where w i′j′ represents the relationship between the i'-th question text and the j′-th knowledge point.

[0028] The calculation expression for the ratio of the number of nodes completed according to the recommendation for the learner to complete a certain learning goal to the number of nodes where a certain node is located is where n represents the total number of nodes included in the learning path, and i0 represents the total number of actually recommended nodes. When the path value is smaller, the recommended result is more satisfactory and meets the expectations; when the path value is larger, that is, closer to 1, the recommended result does not meet the expectations very much.

[0029] As a further improvement of the above technical solution, the knowledge levels and historical record information of multiple learners are obtained, and a learner model is constructed according to the historical record information and the knowledge levels, including:

[0030] Determine the value of the similar learner clustering cluster k according to the obtained learner behavior data. The expression for calculating the distance between all learner coordinates and k seed points is where Y represents the learner model;

[0031] Compare the distance between each learner coordinate and each seed point. If the distance between the learner coordinate G and the seed point Kn is the smallest, then the learner coordinate G belongs to the learner cluster Kn;

[0032] Move the seed points so that each seed point moves to the center of its cluster, that is, make the sum of the distances between the seed point Kn and the learner coordinates belonging to this learner cluster the smallest, and repeat the above process until the seed points no longer move.

[0033] In the second aspect, the present invention also provides a resource learning path planning device based on deep learning, including:

[0034] A data acquisition module for obtaining the knowledge levels and historical record information of multiple learners, and constructing a learner model according to the historical record information and the knowledge levels, where the knowledge levels include knowledge point IDs and mastery levels, and the historical record information includes learning sequences, learning resources, and homework data;

[0035] The first construction module is used to calculate the learner similarity between any two different learners of the learner model, and construct the learning behavior model corresponding to the learner according to the learner similarity. Among them, each learner is represented by a vector in the learner model, and the cosine similarity is used to evaluate the similarity between two vectors, that is, the cosine value of the included angle between two learner vectors is used as the similarity score between learners. The expression of the similarity score is where u a represents the vector expression of learner a, and u b represents the vector expression of learner b;

[0036] The second construction module is used to predict the cognitive level of the learner according to the learning behavior model and the learner model, and construct a learning path recommendation model according to the cognitive level. Among them, the learning path recommendation model includes constructing two state sets respectively represent whether the learner accepts the recommendation. C represents accepting the recommendation, represents not accepting the recommendation;

[0037] The path planning module is used to plan the target learning path corresponding to the learner based on the learning path recommendation model and the preset knowledge graph.

[0038] The present invention provides a resource learning path planning method and device based on deep learning. By obtaining the knowledge levels and historical record information of multiple learners, constructing a learner model according to the historical record information and the knowledge level, calculating the learner similarity between any two different learners of the learner model, constructing the learning behavior model corresponding to the learner according to the learner similarity, predicting the cognitive level of the learner according to the learning behavior model and the learner model, constructing a learning path recommendation model according to the cognitive level, planning the target learning path corresponding to the learner according to the learning path recommendation model and the preset knowledge graph, quantifying the learner's own learning behavior into learning behavior indicators to construct a learner model, predicting the cognitive level of the learner according to the learner behavior model to accurately improve the cognitive level of the learner, making personalized learning path recommendations to the learner, and further completing accurate learning path planning in combination with the knowledge graph, providing a practical, reasonable and interpretable learning path for the learner, thereby improving the learning efficiency and quality of the learner. Brief Description of the Drawings

[0039] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required to be used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention, and therefore should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can be obtained according to these drawings without creative efforts.

[0040] Figure 1 Flow chart of the resource learning path planning method based on deep learning provided by the present invention;

[0041] Figure 2 Block diagram of the structure of the resource learning path planning device based on deep learning provided by the present invention. Detailed implementation manners

[0042] The embodiments of the present invention will be described in detail below. The examples of the embodiments are shown in the accompanying drawings, where the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described by referring to the accompanying drawings are exemplary only for explaining the present invention and should not be construed as a limitation to the present invention.

[0043] Refer to Figure 1 , the present invention provides a resource learning path planning method based on deep learning, including:

[0044] S1: Obtain the knowledge levels and historical record information of multiple learners, and construct a learner model according to the historical record information and the knowledge levels. Among them, the knowledge levels include knowledge point IDs and mastery levels, and the historical record information includes learning sequences, learning resources, and assignment data;

[0045] S2: Calculate the learner similarity between any two different learners of the learner model, and construct a learning behavior model corresponding to the learner according to the learner similarity. Among them, each learner is expressed by a vector in the learner model, and the cosine similarity is used to evaluate the similarity between two vectors, that is, the cosine value of the included angle between two learner vectors is used as the similarity score between learners. The expression of the similarity score is where u a represents the vector expression of learner a, and u b represents the vector expression of learner b;

[0046] S3: Predict the cognitive level of the learner according to the learning behavior model and the learner model, and construct a learning path recommendation model according to the cognitive level. Among them, the learning path recommendation model includes constructing two state sets respectively represent whether the learner accepts the recommendation, C represents accepting the recommendation, represents not accepting the recommendation;

[0047] S4: Plan the target learning path corresponding to the learner based on the learning path recommendation model and a preset knowledge graph.

[0048] In this embodiment, planning the target learning path corresponding to the learner based on the learning path recommendation model and the preset knowledge graph includes: obtaining all knowledge points in the preset knowledge graph as the knowledge point set to be sorted, and presetting an initial knowledge point KP0; exporting a list L from the database corresponding to the preset knowledge graph to determine the knowledge point relationships, where the knowledge point relationships include the inclusion relationship and parallel relationship between knowledge points, and the record information of the list L is (KP m ,R,KP n ) indicating that the relationship between the knowledge point KP m and KP n is R, and the knowledge point relationships include knowledge point importance, difficulty, centrality, topological hierarchy attribute feature data, and knowledge point sorting indicators; using a topological sorting algorithm to generate the target learning path according to the preset knowledge graph and knowledge point attribute feature data. Using a topological sorting algorithm to generate the target learning path according to the preset knowledge graph and knowledge point attribute feature data includes: comparing the target learning path with the preset expert path to obtain the path similarity, and the expression for calculating the path similarity is where path ep , path at respectively represent the preset expert path and the target learning path, KP mat represents the number of learning paths that match between the preset expert path and the target learning path, and KP tot represents the total number of learning paths; using the learning path evaluation index fitness to test the quality of the target learning path, and the calculation expression of the evaluation index fitness is where pen adj represents the number of times the learning path violates the adjacent principle between knowledge points, and pen ord represents the number of times the learning path violates the prerequisite and successor principle between knowledge points. The fewer the number of times the learning path violates the rules, the smaller the fitness value, indicating that the quality of the path generation is higher. Obtaining the knowledge levels and historical record information of multiple learners, and constructing a learner model according to the historical record information and the knowledge level includes: determining the value of the similar learner clustering cluster k according to the obtained learner behavior data, and the expression for calculating the distance between all learner coordinates and k seed points is where Y represents the learner model; comparing the distance between each learner coordinate and each seed point. If the distance between the learner coordinate G and the seed point Kn is the smallest, then the learner coordinate G belongs to the learner cluster Kn; moving the seed points so that each seed point moves to the center of its cluster, that is, making the sum of the distances between the seed point Kn and the learner coordinates belonging to this learner cluster the smallest, and repeating the above process until the seed points no longer move.

[0049] It should be noted that the basic information can be name, student ID, username, password, gender, class, etc. The mastery of a single knowledge point is calculated based on the learner's mastery ability after learning that knowledge point. First, the mastery ability value of a single knowledge point is obtained through calculation. The calculation expression of the one-parameter model of the corresponding theory of the project is where P ij (θ k ) represents that the mastery ability of learner k for knowledge point i is θ k , and it is the probability of being able to answer the j-th question of the i-th knowledge point. Q ij (θ k ) represents that the mastery ability of learner k for knowledge point i is θ k when the probability of answering the j-th question of the i-th knowledge point wrongly. θ k is obtained by performing maximum likelihood estimation on P ij (θ k ). When LL(μ1, μ2... μ n ) is the largest, the θ k value is used for estimation, and its calculation expression is where μ j takes values of 0 or 1. When μ j = 1, it means that learner k has answered the exercise j under knowledge point i correctly. When μ j = 0, it means that learner k has answered the exercise j under knowledge point i wrongly. n represents the number of exercises examining this knowledge point. Then, the Newton-Raphson iteration method is used to iterate different estimates of θ k until LL(μ1, μ2... μ n ) reaches the maximum to obtain the specific value of θ k . The iterative expression of θ is where θ t+1 and θ t are the mastery ability values of a certain knowledge point obtained in the (t + 1)-th and t-th iterations. h t represents the knowledge point mastery ability correction factor, is the first derivative of the log-likelihood function lnLL with respect to θ, the second derivative of the log-likelihood function lnLL with respect to θ. When h t is small enough or the number of iterations is large enough, the θ value at this time is the learner's mastery ability value for the knowledge point. θ k takes values in the preset interval [-m, m]. m represents that learner k has the best mastery ability for knowledge point j, 0 represents medium, and -m represents the worst mastery ability. The expression for not being able to master knowledge point i is γ i represents the weighted score of the knowledge point. When it is greater than the threshold c iWhen it is the case, it is considered that the learner has mastered the knowledge point i; otherwise, the learner has not mastered it.

[0050] It should be understood that in order to recommend an appropriate learning path to the learner, it is very necessary to construct a learner model. The learning path refers to the sequence followed when learning each knowledge point to master the target knowledge point. According to the learner's mastery of each knowledge point, combining the learning goal with the current learning state, a series of knowledge points are orderly listed under the guidance of the corresponding learning strategy, and finally a learning path is formed. The learning path reflects the sequence of learning activities, which should be the same as the sequence of the association relationship between knowledge points and should conform to the original precursor relationship and successor relationship between knowledge points. The learning path is a sequence formed by a learning activity. The knowledge points in the sequence are coherent, that is, there cannot be missing knowledge nodes in the path. Each knowledge point forms a knowledge graph through complex association relationships. The knowledge points are connected by directed graph line segments. Countless directed line segments are interconnected to form a knowledge graph. The learning path is mainly the orderly arrangement of knowledge point objects, and the path to reach the target object is not unique. In the process of automatically planning from the knowledge graph to the learning path, it lays a solid foundation for the generation of personalized learning paths, the recommendation of learning resources, and even the construction of personalized learning models. The feasibility of learning path planning is relatively high and the quality is good. During the online learning process, it can replace the path formulated by experts and provide a general, reasonable and interpretable learning path for online learners, thereby improving the learning efficiency and quality of online learners.

[0051] Optionally, predicting the cognitive level of the learner according to the learning behavior model and the learner model includes:

[0052] The cognitive level of the learner is divided into three domains: the positive domain POS, the boundary domain BND, and the negative domain NEG. According to the item response theory, the learner's cognitive level θ is determined. Among them, for the 2PL model of the item response theory, the expression for using the item characteristic curve to describe the probability that the learner answers a preset question is where f represents the difficulty coefficient of the question, f ∈ [-2, 2], d represents the constant 1.072, θ represents the ability value, that is, the cognitive level of the learner, p(θ) represents the probability of answering the test question correctly. The POS domain represents learners with a high cognitive level, the NEG domain represents learners with a low cognitive level, and the BND domain represents learners with a medium cognitive level;

[0053] The data of the test results in the obtained score matrix and the known question difficulty coefficient are input into the IRT model, and the expression for establishing the maximum likelihood function of the ability parameter is where p i represents the probability of answering correctly obtained by the item response theory model function, y iIndicates the label of the learner's true answering situation in the score matrix, and thus the expression of the logarithmic maximum likelihood function is The ability parameter θ is obtained by taking the derivative of the logarithmic maximum likelihood function.

[0054] In this embodiment, the decision information given to each region is A = {α P , β B , α N}, which respectively represent the learning path recommendation information for learners with different cognitive levels. α P indicates accepting the learning recommendation of other knowledge points after completing the learning task of the current knowledge point. α N indicates returning to consolidate basic knowledge when the current cognitive level is low. α B indicates accepting exercise recommendations to improve the cognitive level; the expression of the learning utility function matrix is where λ PP , λ BP and λ NP represent the loss values corresponding to taking actions α P , α B and α N when the object belongs to C. λ PN , λ BN and λ NN represent the loss values corresponding to taking actions α P , α B and α N when the object does not belong to C. s(λ PP ), s(λ BP ) and s(λ NP ) represent the utility values corresponding to taking actions α P , α B and α N when the object belongs to C. s(λ PN ), s(λ BN ) and s(λ NN ) represent the utility values corresponding to taking actions α P , α B and α N when the object does not belong to C; s(λ PP ) ≥ s(λ BP ) ≥ s(λ NP ), indicating that in the state of the object , the utility of judging [x] as the positive domain is greater than the utility of judging it as the boundary domain, and further greater than the utility of judging it as the negative domain; s(λ PN ) ≤ s(λ BN ) ≤ s(λ NN ), indicating that in the state of the object In the [x] state, the utility of determining [x] as the negative domain is less than the utility of determining it as the boundary domain, and further less than the utility of determining it as the positive domain.

[0055] It should be noted that the cognitive level refers to the learner's domain knowledge level, which reflects the learner's learning performance information. The model corresponding to the cognitive level consists of two parts: the target course and the initial level, and its expression is Cog = (Les, <t1, l1>, <t2, l2>... <t m , l m >), where Cog represents the learner's cognitive level, Les represents the learner's target course, that is, the course the learner wants to learn, t m represents the m-th learning unit, l m represents the mastery level of knowledge, <t m , l m > represents a learner's cognitive level of the knowledge of the m-th learning unit, which is mainly obtained through pre-course tests. Using the item response theory to establish the functional relationship between the learner's cognitive level and item parameters can be described by the function of item characteristics. In fact, it is the regression curve of the probability of correct answering of the learner on the item to the latent trait score, and corresponding models can be selected according to different response data to estimate the parameters. Although the cognitive levels of different learners are different, each learner hopes to improve their cognitive level through a certain period of learning based on their current cognitive level. Learners assigned to the BND domain need to adopt a delayed decision-making processing method, that is, reprocess the BND domain through the three-way decision boundary domain processing model based on learning utility to obtain a personalized learning path. The personalized learning path recommends different learning paths for different learners according to the differences between the learner's behavior habits and their respective cognitive levels. The personalized learning path needs to conform to the learner's personalized learning habits and can obtain a higher cognitive level through learning the recommended path, so as to improve the accuracy of learning path planning.

[0056] Optionally, export the list L from the database corresponding to the preset knowledge graph to determine the knowledge point relationship, including:

[0057] Use the keyword extraction algorithm to process the text resources to obtain the knowledge point set and knowledge point importance, and its expression is where Ws(V i ) represents the weight of node V i , Ws(V j ) represents the weight of node V j after the previous iteration, ω ji represents the similarity between node V j and node V i , and g represents the damping coefficient;

[0058] Calculate the centrality of knowledge points according to the properties of the knowledge graph by using the ratio of the in-degree to the out-degree of the knowledge points. The larger the ratio, the higher the centrality of the knowledge point. The expressions for the in-degree and out-degree of the knowledge points are The calculation expression for centrality is Where Represents the node V i The ratio of the in-degree to the out-degree, Pre(V i ) represents the set of first-order predecessor knowledge points of the knowledge point V i , and Suc(V i ) represents the set of first-order successor knowledge points of the knowledge node V i ;

[0059] Use the above attribute values to calculate the knowledge point sorting index Where the importance Imp kpi Use the weight value of the node Ws(V i ) in the above expression to obtain the expression Where Imp kpi , Diff kpi , Cent kpi And Tp kpi Respectively represent the importance, difficulty, centrality and topological level of the knowledge point kpi, ω imp , ω diff , ω cent And ω tp Represent the assigned attribute weight values.

[0060] In this embodiment, a learning path recommendation model is constructed according to the cognitive level, including: presetting T = [i 1, i2...i m , where T represents the set of knowledge points, and i represents a single knowledge point, then Where L represents the learner path sequence, and ls represents a single learner, Represents the learning path of the learner ls1. Track relevant knowledge points through the learner's questions and classify the questions. Preset i' = 1, 2... M to represent M question texts in the set, and j′ = 1, 2... N to represent N pre-set knowledge points. The classification matrix is W = w i′j′ , where w i′j′ Represents the relationship between the i'th question text and the j′th knowledge point, According to the calculation expression for the ratio of the number of nodes recommended for the learner to complete a certain learning goal to the number of nodes where a certain node is located Where n represents the total number of nodes included in the learning path, and i represents the total number of nodes actually recommended. When the value of path is smaller, the recommended result is more satisfactory and meets the expectations; when the value of path is larger, that is, closer to 1, the recommended result does not meet the expectations well.

[0061] It should be noted that learners usually confuse the target knowledge points involved in the questions and do not know which knowledge point content they should actually learn. In order to quickly track relevant knowledge points through the learners' questions, it is necessary to classify the questions. For example, first collect text corpora, store the text corpora classified according to instructions, clean, sort, and segment the data, calculate the term frequency-inverse document frequency, that is, TF-IDF, as the weight factor, and introduce the weight factor into the conditional probability of the Naive Bayes function. Then, determine the rank factor from the weight factor and integrate the rank factor into the Bayes function to enhance the importance of key words, thereby reducing the impact brought by the feature independence assumption of Naive Bayes and improving the processing efficiency of system data to a certain extent.

[0062] See Figure 2 , the present invention also provides a resource learning path planning device based on deep learning, including:

[0063] A data acquisition module, configured to acquire the knowledge levels and historical record information of multiple learners, and construct a learner model according to the historical record information and the knowledge levels. Among them, the knowledge levels include knowledge point IDs and mastery levels, and the historical record information includes learning sequences, learning resources, and assignment data;

[0064] A first construction module, configured to calculate the learner similarity between any two different learners of the learner model, and construct a learning behavior model corresponding to the learner according to the learner similarity. Among them, each learner is expressed as a vector in the learner model, and the cosine similarity is used to evaluate the similarity between two vectors, that is, the cosine value of the included angle between two learner vectors is used as the similarity score between learners. The expression of the similarity score is where u a represents the vector expression of learner a, and u b represents the vector expression of learner b;

[0065] A second construction module, configured to predict the cognitive level of the learner according to the learning behavior model and the learner model, and construct a learning path recommendation model according to the cognitive level. Among them, the learning path recommendation model includes constructing two state sets respectively represent whether the learner accepts the recommendation, C represents accepting the recommendation, represents not accepting the recommendation;

[0066] A path planning module for planning a target learning path corresponding to the learner based on the learning path recommendation model and a preset knowledge graph.

[0067] In this embodiment, a feature-weighted Naive Bayes classifier is used to obtain target knowledge points. The main process of obtaining knowledge points includes: feature item processing: performing jieba word segmentation on the text collection of each knowledge point category, sequentially reading each word in each question text under this category, counting the number of texts h in which the word appears in this category of text, counting and calculating the frequency r of the occurrence of this word in this category of text, multiplying h by r to obtain the TF-IDF value of this feature word, recording this feature word and its weight in the candidate set of the corresponding classification category, and then sequentially running the next feature word after completion; forming a vector matrix: listing the candidate sets of text feature values under each knowledge point category in descending order according to the weight, selecting feature items in each question text to form a set representing the features of this question text, constructing a dictionary for the key words of each question text extracted under all knowledge point categories, arranging them in the order of the dictionary position, representing the involved text using a Boolean model, and finally forming two matrices, one is the TF-IDF weight factor space vector matrix of all knowledge point categories, and the other is the vector matrix represented by the question texts under each knowledge point classification category; constructing a classifier: obtaining the above two vector matrices, calculating the TF and TF-IDF under the corresponding knowledge point categories, obtaining the Naive Bayes conditional probability vector of each feature item under the relevant knowledge point categories, determining the rank factor vector from the word frequency-inverse document frequency (TF-IDF) vectors under each knowledge point category, and then effectively calculating the posterior probability under each knowledge point category.

[0068] It should be noted that by obtaining the knowledge levels and historical record information of multiple learners, constructing a learner model based on the historical record information and the knowledge level, calculating the learner similarity between any two different learners of the learner model, constructing a learning behavior model corresponding to the learner according to the learner similarity, predicting the cognitive level of the learner according to the learning behavior model and the learner model, constructing a learning path recommendation model according to the cognitive level, planning a target learning path corresponding to the learner according to the learning path recommendation model and a preset knowledge graph, quantifying the learner's own learning behavior into learning behavior indicators to construct a learner model, predicting the cognitive level of the learner according to the learner behavior model to accurately improve the learner's cognitive level, making personalized learning path recommendations to the learner, and further completing accurate learning path planning in combination with the knowledge graph, providing a practical, reasonable and interpretable learning path for the learner, thereby improving the learning efficiency and quality of the learner.

[0069] In all examples shown and described herein, any specific values should be construed as merely exemplary and not as a limitation. Thus, other examples of the exemplary embodiments may have different values.

[0070] It should be noted that like reference numerals and letters refer to like items in the following figures. Thus, once an item is defined in one figure, it need not be further defined and explained in subsequent figures.

[0071] The above-described embodiments merely represent several embodiments of the present invention. The description thereof is relatively specific and detailed, but it should not be construed as limiting the scope of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all fall within the protection scope of the present invention.

Claims

1. A resource learning path planning method based on deep learning, characterized in that Including the following steps: Obtain the knowledge levels and historical record information of multiple learners, and construct a learner model based on the historical record information and the knowledge levels. Among them, the knowledge levels include knowledge point IDs and mastery levels, and the historical record information includes learning sequences, learning resources, and assignment data; Calculate the learner similarity between any two different learners of the learner model, and construct a learning behavior model corresponding to the learner according to the learner similarity. Among them, each learner is expressed as a vector in the learner model, and the cosine similarity is used to evaluate the similarity between two vectors, that is, the cosine value of the included angle between two learner vectors is used as the similarity score between learners. The expression of the similarity score is where u a represents the vector expression of learner a, and u b represents the vector expression of learner b; Predict the cognitive level of the learner according to the learning behavior model and the learner model, and construct a learning path recommendation model according to the cognitive level, wherein the learning path recommendation model includes constructing two state sets respectively represent whether the learner accepts the recommendation, C represents accepting the recommendation, represents not accepting the recommendation; Plan the corresponding target learning path for the learner based on the learning path recommendation model and a preset knowledge graph.

2. The resource learning path planning method based on deep learning according to claim 1, characterized in that Planning the corresponding target learning path for the learner based on the learning path recommendation model and a preset knowledge graph includes: Obtain all knowledge points in the preset knowledge graph as a set of knowledge points to be sorted, and preset an initial knowledge point KP0; Export list L from the database corresponding to the preset knowledge graph to determine the knowledge point relationships, where the knowledge point relationships include the inclusion relationship and the parallel relationship between knowledge points, and the record information of list L is (KP m ,R,KP n ) indicating that the relationship between knowledge point KP m and KP n is R. The knowledge point relationships include the importance, difficulty, centrality, and topological hierarchy attribute feature data of knowledge points, as well as the knowledge point sorting index; Generate a target learning path using a topological sorting algorithm according to the preset knowledge graph and knowledge point attribute feature data.

3. The resource learning path planning method based on deep learning according to claim 2, characterized in that, Generating a target learning path using a topological sorting algorithm according to the preset knowledge graph and knowledge point attribute feature data includes: Compare the target learning path with the preset expert path to obtain the path similarity. The expression for calculating the path similarity is where path ep and path at represent the preset expert path and the target learning path respectively, and KP mat represents the number of learning paths that match between the preset expert path and the target learning path, and KP tot represents the total number of learning paths; The quality of the target learning path is tested using the learning path evaluation metric fitness. The calculation expression of the evaluation metric fitness is where pen adj represents the number of violations of the adjacent principle between knowledge points by the learning path, and pen ord represents the number of violations of the prerequisite and successor principle between knowledge points by the learning path. The fewer the number of violations of the rules by the learning path, the smaller the fitness value, indicating the higher the quality of the path generation.

4. The resource learning path planning method based on deep learning according to claim 1, wherein Predict the cognitive level of the learner according to the learning behavior model and the learner model, including: The cognitive level of learners is divided into three domains, the positive domain POS, the boundary domain BND, and the negative domain NEG. According to the item response theory, the cognitive level θ of learners is determined. Among them, for the 2PL model of the item response theory, the expression using the item characteristic curve to describe the probability that learners answer the preset questions correctly is where f represents the difficulty coefficient of the question, f ∈ [-2, 2], d represents the constant 1.072, θ represents the ability value, that is, the cognitive level of the learner, p(θ) represents the probability of answering the test question correctly. The POS domain represents learners with a high cognitive level, the NEG domain represents learners with a low cognitive level, and the BND domain represents learners with a medium cognitive level; Input the test result data in the obtained score matrix and the known item difficulty coefficient into the IRT model. The expression for the maximum likelihood function of the ability parameter is where p i represents the probability of answering correctly obtained from the item response theory model function, and y i represents the true answer situation label of the learner in the score matrix. Thus, the expression for the log-likelihood function is The ability parameter θ is obtained by taking the derivative of this log-likelihood function.

5. The resource learning path planning method based on deep learning according to claim 4, wherein Also included: The given decision information for each region is A = {α P , α B , α N}, which respectively represent the learning path recommendation information for learners with different cognitive levels. α P indicates that the learner has completed the learning task of the current knowledge point and accepts the learning recommendation of other knowledge points. α N indicates that the current cognitive level is low and the learner needs to return to consolidate basic knowledge. α B indicates that the learner accepts the exercise recommendation to improve the cognitive level; The expression of the learning utility function matrix is where λ PP , λ BP and λ NP represent the loss values corresponding to taking actions α P , α B and α N when the object belongs to C, and λ PN , λ BN and λ NN represent the loss values corresponding to taking actions α P , α B and α N when the object does not belong to C. s(λ PP ), s(λ BP ) and s(λ NP ) represent the utility values corresponding to taking actions α P , α B and α N when the object belongs to C, and s(λ PN ), s(λ BN ) and s(λ NN ) represent the utility values corresponding to taking actions α P , α B and α N when the object does not belong to C; s(λ PP ) ≥ s(λ BP ) ≥ s(λ NP ), indicating that in the state of the object , the utility of judging [x] as the positive region is greater than the utility of judging it as the boundary region, and further greater than the utility of judging it as the negative region; s(λ PN ) ≤ s(λ BN ) ≤ s(λ NN ), indicating that in the state of the object , the utility of judging [x] as the negative region is less than the utility of judging it as the boundary region, and further less than the utility of judging it as the positive region.

6. The resource learning path planning method based on deep learning according to claim 1, wherein Export a list L from the database corresponding to the preset knowledge graph to determine knowledge point relationships, including: The text resources are processed by a keyword extraction algorithm to obtain a knowledge point set and the importance of knowledge points, and its expression is where Ws(V i ) represents the weight of node V i , Ws(V j ) represents the weight of node V j after the previous iteration, ω ji represents the similarity between node V j and node V i , and g represents the damping coefficient; Calculate the centrality of knowledge points according to the properties of the knowledge graph by using the ratio of the in-degree to the out-degree of knowledge points. The larger the ratio, the higher the centrality of the knowledge point. The expressions for the in-degree and out-degree of knowledge points are The calculation expression for centrality is where represents the node V i is the ratio of the in-degree to the out-degree, Pre(V i ) represents the set of first-order predecessor knowledge points of the knowledge point V i , and Suc(V i ) represents the set of first-order successor knowledge points of the knowledge node V i ; Calculate the knowledge point sorting index using the above attribute values Among them, the importance Imp kpi Use the weight value of the Ws(V i ) node in the above expression to obtain the expression as Among them, Imp kpi , Diff kpi , Cent kpi and Tp kpi respectively represent the importance, difficulty, centrality, and topological level of the knowledge point kpi, and ω imp , ω diff , ω cent and ω tp represent the assigned attribute weight values.

7. The resource learning path planning method based on deep learning according to claim 1, wherein Construct a learning path recommendation model according to the cognitive level, including: Preset \(T = [i 1, i_2...i m \), where \(T\) represents the set of knowledge points and \(i\) represents a single knowledge point. Then where \(L\) represents the learner path sequence and \(ls\) represents a single learner. represents the learning path of learner \(ls_1\).

8. The resource learning path planning method based on deep learning according to claim 7, characterized in that Also included: Track relevant knowledge points through learners' questions and classify the questions. Presuppose \(i' = 1, 2,\cdots,M\) to represent \(M\) question texts in the set, and \(j\) ′ \(= 1, 2,\cdots,N\) to represent \(N\) pre - set knowledge points. The classification matrix is \(W = w\) i′j′ , where \(w\) i′j′ represents the relationship between the \(i'\)-th question text and the \(j\) ′ -th knowledge point. The calculation expression for the ratio of the number of nodes completed according to the recommendation to achieve a certain learning goal to the number of nodes where a certain node is located is as follows where n represents the total number of nodes included in the learning path, and i0 represents the total number of nodes actually recommended. When the value of path is smaller, the recommended result is more satisfactory and meets the expectations; when the value of path is larger, that is, closer to 1, the recommended result does not meet the expectations very well.

9. The resource learning path planning method based on deep learning according to claim 1, wherein Obtain the knowledge levels and historical record information of multiple learners, and construct a learner model based on the historical record information and the knowledge levels, including: Determine the value of k for the similar learner clustering cluster based on the obtained learner behavior data. The expression for calculating the distances between all learner coordinates and the k seed points is where Y represents the learner model; Compare the distance between each learner coordinate and each seed point. If the distance between the learner coordinate G and the seed point Kn is the smallest, then the learner coordinate G belongs to the learner cluster Kn; Move the seed points so that each seed point moves to the center of its cluster, that is, make the sum of the distances between the seed point Kn and the learner coordinates belonging to this learner cluster the smallest, and repeat the above process until the seed points no longer move.

10. A deep learning-based resource learning path planning device for the deep learning-based resource learning path planning method according to any one of claims 1-9, characterized in that, Including: A data acquisition module for obtaining the knowledge levels and historical record information of multiple learners, and constructing a learner model based on the historical record information and the knowledge levels. Among them, the knowledge levels include knowledge point IDs and mastery levels, and the historical record information includes learning sequences, learning resources, and assignment data; The first construction module is used to calculate the learner similarity between any two different learners of the learner model, and construct the learning behavior model corresponding to the learner according to the learner similarity. Among them, each learner is expressed by a vector in the learner model, and the cosine similarity is used to evaluate the similarity between two vectors, that is, the cosine value of the included angle between two learner vectors is used as the similarity score between learners. The expression of the similarity score is where u a represents the vector expression of learner a, and u b represents the vector expression of learner b; A second construction module, configured to predict the cognitive level of the learner according to the learning behavior model and the learner model, and construct a learning path recommendation model according to the cognitive level, wherein the learning path recommendation model includes constructing two state sets respectively indicating whether the learner accepts the recommendation, where C indicates acceptance of the recommendation, indicating non-acceptance of the recommendation; A path planning module for planning the corresponding target learning path for the learner based on the learning path recommendation model and a preset knowledge graph.

Images