A knowledge graph-based adaptive learning resource recommendation method and system

Abstract

The present invention proposes an adaptive learning resource recommendation method and system based on a knowledge graph, comprising: establishing a user cognitive diagnosis model and a test score prediction model to predict the user's mastery of knowledge points and the scores of unfinished test questions, and then selecting a first candidate learning resource; constructing a knowledge graph, using the diagnosis results of the user cognitive diagnosis model to locate the user's better and weaker knowledge points in the knowledge graph, and select a second candidate learning resource; screening the best learning resource from the candidate learning resources and recommending it to the user. The present invention not only considers the user's cognitive level and the user's predicted scores for unfinished test questions, but also considers the semantic relationship between knowledge points, constructs a corresponding knowledge graph, locates the user's better and poorer knowledge points in the knowledge graph in combination with the user's cognitive diagnosis results, selects the most suitable learning resource for the user, and recommends it to the user.

Description

Technical Field

[0001] The present invention relates to the field of online learning technology, and more specifically, to a method and system for adaptively recommending learning resources based on knowledge graphs. Background Art

[0002] With the rapid development of technology, online learning has become a vital learning method. As a crucial component of online learning, online teaching resources are becoming increasingly important in promoting user cognition, enhancing practical skills, and cultivating advanced thinking abilities. However, the explosive growth of online educational resources is gradually exacerbating learners' problems of "information overload" and "knowledge loss." Providing learners with personalized learning services and recommending appropriate educational resources is a challenge that needs to be addressed.

[0003] There is a method for recommending sequential exercises based on cognitive diagnosis. The method constructs a question-answering score matrix for each user based on the question-answering behavior data of all users, and obtains the first knowledge point mastery matrix of each user by combining the question-knowledge point association matrix; obtains the second knowledge point mastery matrix based on the first knowledge point mastery matrix, the error rate and the guessing rate, and obtains the cognitive diagnosis vector of each user by combining the recurrent neural network system; and selects the target user with the highest similarity to the cognitive diagnosis vector of the user to be recommended from the cognitive diagnosis vectors of all users according to the similarity calculation formula, extracts the question-answering behavior data of the target user, and selects the test questions that have not been tested by the user to be recommended, so as to facilitate recommending the selected test questions to the user to be recommended.

[0004] However, the above method recommends exercises to users based on the user's cognitive diagnosis results, ignoring the relationship between knowledge points and the importance of different knowledge points in the examination of test questions. The test questions recommended to users are not the most suitable test questions for users. Summary of the Invention

[0005] In order to solve the problem of how to improve the accuracy of learning resource recommendations, the present invention provides a method and system for adaptive learning resource recommendation based on knowledge graph.

[0006] In order to solve the above technical problems, the technical solutions of the present invention are as follows:

[0007] In a first aspect, the present invention proposes a method for adaptive learning resource recommendation based on knowledge graph, comprising:

[0008] A user cognitive diagnosis model is established, and a learning cognitive diagnosis of the user is performed through the user cognitive diagnosis model, and based on the diagnosis results, the user's mastery of knowledge points is predicted.

[0009] A test question score prediction model is established, and the test question score prediction model is used to predict the user's scores on the unfinished test questions.

[0010] The first candidate learning resource is selected based on the user's mastery of the knowledge points and the user's scores on the unfinished test questions.

[0011] Build a knowledge graph, use the diagnostic results of the user cognitive diagnosis model to locate the knowledge points that the user has mastered well and weakly in the knowledge graph, and select the second candidate learning resources.

[0012] An optimal learning resource is selected from the first candidate learning resource and the second candidate learning resource, and the optimal learning resource is recommended to the user.

[0013] Preferably, performing learning cognitive diagnosis on the user through the user cognitive diagnosis model and predicting the user's mastery of knowledge points based on the diagnosis results specifically includes:

[0014] Define a set P = {p1, p2, ..., p U} is the user set, set T = {t1, t2, ..., t V} is the set of questions, set C = {c1, c2, ..., c K} is a set of knowledge points;

[0015] Define the user's potential learning state as a multi-dimensional knowledge point mastery vector α uk ={α u1 , α u2 ,...,α uK}, α uk Each dimension corresponds to user p u Knowledge point c k The mastery of the situation, when α uk =1 indicates user p u Master knowledge point c k , when α uk =0 indicates that user p u Not mastered knowledge point c k ; Then calculate user p u In the test question v Potential answer situation η uv , whose formula is as follows:

[0016]

[0017] Among them, when η uv =1; when η uv =1, indicating that user p u Able to answer test questions correctly v And mastered the test questions vKnowledge points involved c k When η uv =0, indicating that user p u Unable to answer the test questions correctly v and have not mastered the test questions v Knowledge points involved c k .

[0018] Calculate user p u In the test questions that have not been mastered v Knowledge points involved c k Under the premise of human guessing, the correct answer to the test question is obtained. v The guess rate g v , its formula is as follows:

[0019] g v =P(r uv =1|η uv =0)

[0020] Among them, r uv Indicates user p u In the test question v The score on r uv =1, indicating that user p u Answer the question correctly v .

[0021] Calculate user p u In mastering the test questions v Knowledge points involved c k Under the premise that the test questions cannot be answered correctly due to mistakes, v Error rate v , its formula is as follows:

[0022] s v =P(r uv =0|η uv =1)

[0023] Among them, when r uv =0, indicating that user p u Failed to answer the test questions correctly v .

[0024] According to the guess rate g v , error rate v and potential answer situation η uv , and considering the difficulty of the test questions, the forgetting curve and the number of answers, calculate the user p u In the test question v The probability of a correct answer on , is given by:

[0025]

[0026] Among them, ω represents the difficulty of the test question, Used to fit the forgetting curve; ξ represents user p u In the test question v When the user p u Correctly answer the test questions v ξ takes the value of 1, when user p u Incorrectly answered questions v , ξ takes the value of 0; frequency represents the number of times the question is answered.

[0027] According to user p u In the test question v The probability of correct answer on the ,determine the user's multi-dimensional knowledge point mastery vector α u And calculate the vector α u Estimated value, using vector α u The estimated values ​​construct the user-knowledge point matrix B.

[0028] According to the user-knowledge point matrix B, predict user p u Knowledge point c k The grasp of the situation.

[0029] Preferably, using the vector α u The estimated value constructs the user-knowledge point matrix B specifically including:

[0030] Use EM algorithm to maximize the guessing rate g v and error rate s v The marginal likelihood probability of , respectively get the guess rate g v and error rate s v Estimated value of and

[0031] Determine the user's multi-dimensional knowledge point mastery vector α through the maximum a posteriori probability algorithm u Estimated value of The formula is as follows:

[0032]

[0033] Among them, α represents, R u Represents the score matrix of user u.

[0034] Master the vector α based on the user's multi-dimensional knowledge points u Estimated value of Construct the user-knowledge point matrix B.

[0035] Preferably, using the test question score prediction model to predict the user's score for the unfinished test questions specifically includes:

[0036] Establishing a test question score prediction model, inputting the test question into the test question score prediction model, and obtaining an implicit feature vector corresponding to the test question;

[0037] According to the implicit feature vector corresponding to the test question, the probability matrix is ​​used to identify the test questions of user p. u In the test question v The score on r uv The constructed user test score matrix R is decomposed to obtain the user implicit feature matrix and the test question implicit feature matrix;

[0038] According to the user implicit feature matrix and the test question implicit feature matrix, the KNN algorithm is used to correct the user test question score matrix R after probability matrix decomposition to obtain the final user score for the unfinished test questions.

[0039] Preferably, the test question score prediction model uses a convolutional neural network and an attention mechanism to obtain the implicit feature vector corresponding to the test question. The convolutional neural network includes an embedding layer, an attention layer, a convolution layer, a pooling layer, and an output layer, specifically:

[0040] The relevant text of the test question is input into the embedding layer, and the embedding layer outputs the d-dimensional vector a for each word in the text n , using vector a n Construct a dense digital matrix S, whose formula is as follows:

[0041]

[0042] Input the dense digital matrix S into the attention layer to highlight the features of the key words in the text and obtain the feature expression of the key words The specific calculation is as follows:

[0043]

[0044] s(i)=f(H i *W att +b att )

[0045] Among them, A represents the word feature representation obtained through the attention mechanism; D represents a central word h i The context range, s(i) represents the feature value weight of each word in the context range, W att Indicates the matrix parameters corresponding to the context range, b att represents the bias, f(·) represents the activation function;

[0046] Each word feature w in the context i Multiplying the word's feature value weight s(i) outputs a new feature representation The expression formula is as follows:

[0047]

[0048] Represent the features Input the convolution layer to extract the features C(i) of the key words. The formula is as follows:

[0049]

[0050] Among them, W c represents the shared weight, b c represents bias, f(·) represents activation function;

[0051] The feature C(i) is input into the pooling layer, and the pooling layer outputs the representative feature C out (i)

[0052] C out (i) = MAX(C(i))

[0053] The representative feature C out (i) Input and output layer, obtain the implicit feature vector D corresponding to the relevant text of the test question j :

[0054]

[0055] in, and is the mapping matrix, and and bias vector, f(·) represents the activation function; W represents all weights and biases, H j Represents the word vector corresponding to the relevant text of the test question.

[0056] Preferably, a probability matrix is ​​used to identify the u In the test question v The score on r uv The constructed user test score matrix R is decomposed to obtain the user implicit feature matrix and the test question implicit feature matrix, which specifically include:

[0057] Assume that the user test score matrix R is decomposed into the user implicit feature matrix U, the knowledge point implicit feature matrix K and the test question implicit feature matrix V through the probability matrix;

[0058] The prior probabilities of the user implicit feature matrix U and the knowledge point implicit feature matrix K follow Gaussian distribution and are independent of each other:

[0059]

[0060]

[0061] Among them, m represents the number of users, l represents the number of knowledge points, and represents variance, I represents unit vector;

[0062] The probability distribution of the implicit feature matrix V of the test question is:

[0063]

[0064]

[0065] Among them, W z represents the weights between neurons in a convolutional neural network, represents the variance, n represents the number of test questions, represents variance;

[0066] After the implicit feature vectors corresponding to the user and the test questions are given, the conditional probability distribution of the user test question score matrix R is:

[0067]

[0068] in, represents the variance, represents the implicit features obtained by probability matrix decomposition, Represents the indicator function. If student i has done test j, then otherwise

[0069] The conditional probability distribution of the user knowledge point matrix B is:

[0070]

[0071] Among them, l represents the number of knowledge points, b ic Indicates user i’s mastery of knowledge point c, K c represents the implicit feature vector of the knowledge point, represents the variance, represents the indicator function;

[0072] Combining the probability distribution of the user implicit feature matrix U, the knowledge point implicit feature matrix K, the test question implicit feature matrix V, the user test question score matrix R, and the user knowledge point matrix B, and using the Bayesian criterion and maximization technique, we can finally get the objective function E(K, U, V, B, R) as follows:

[0073]

[0074]

[0075]

[0076]

[0077] Among them, b ij Indicates user i’s mastery of knowledge point j, W i Represents the weight corresponding to the convolutional neural network;

[0078] The optimal solution of the objective function E(K, U, V, B, R) can be solved by the stochastic gradient descent method, thereby obtaining the optimal solutions of the user implicit feature matrix and the test question implicit feature matrix.

[0079] Preferably, based on the user implicit feature matrix and the test question implicit feature matrix, the KNN algorithm is used to correct the user test question score matrix R after the probability matrix decomposition to obtain the final user score for the unfinished test questions, specifically:

[0080] Assume that the KNN algorithm is used to find the N questions that are most similar to question j in the set of questions that user i has completed, to make up for the implicit feature information lost due to probability matrix decomposition, and to correct the user's question score matrix to obtain the final question score R ij :

[0081]

[0082]

[0083] Among them, R ij Indicates the final predicted score of the user for the unfinished test questions, R′ ij is the correction factor, R in represents the user’s original score on the test question, V n The feature vector representing the test question.

[0084] Preferably, selecting the second candidate learning resource specifically includes:

[0085] Build a knowledge graph.

[0086] The formula for calculating the centrality of knowledge meta-nodes in the knowledge graph is as follows:

[0087]

[0088] Among them, G i represents the centrality of knowledge element i, δ represents the contribution of the knowledge element, and λ represents the weight corresponding to the knowledge element; α and β are both weight coefficients, indicating the contribution of the knowledge element itself and its neighboring knowledge elements of each order to the total centrality.

[0089] Classify the learning resources contained in the knowledge meta-node.

[0090] Using the diagnostic results of the user cognitive diagnosis model, the knowledge points that the user has mastered well and weakly are located in the knowledge graph, and the second candidate learning resources are selected.

[0091] Preferably, constructing a knowledge graph specifically includes:

[0092] Collect text datasets;

[0093] Perform entity extraction on the text dataset to obtain the entities in the text dataset;

[0094] Extract relationships between entities to obtain the relationships between entities;

[0095] Artificial knowledge fusion is performed on entities and the relationships between them, and the fusion results are used to construct a knowledge graph, which is then stored in a database.

[0096] In a second aspect, the present invention proposes an adaptive learning resource recommendation system based on a knowledge graph, comprising:

[0097] The user cognitive diagnosis module is used to perform learning cognitive diagnosis on users and predict their mastery of knowledge points based on the diagnosis results.

[0098] The test score prediction module is used to predict the user's score for unfinished test questions.

[0099] The knowledge graph construction module is used to construct the knowledge graph and use the diagnosis results of the user cognitive diagnosis module to locate the knowledge points that the user has better and weaker grasp of in the knowledge graph.

[0100] The learning resource acquisition module is used to obtain the first candidate learning resource based on the output results of the user cognitive diagnosis module and the test score prediction module, and the second candidate learning resource based on the output results of the knowledge graph construction module.

[0101] The recommendation module is used to select the best learning resource from the first candidate learning resource and the second candidate learning resource, and recommend the best learning resource to the user.

[0102] Compared with the existing technology, the beneficial effect of the technical solution of the present invention is: the present invention not only takes into account the user's cognitive level and the user's predicted score for unfinished test questions, but also takes into account the semantic relationship between knowledge points, constructs a corresponding knowledge graph, and combines the user's cognitive diagnosis results to locate the knowledge points that the user has mastered better and worse in the knowledge graph, selects the learning resources that are most suitable for the user, and recommends them to the user, thereby realizing accurate recommendation of learning resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0103] Figure 1 This is a flowchart of the adaptive learning resource recommendation method based on knowledge graph in Example 1.

[0104] Figure 2 This is a flowchart of the adaptive learning resource recommendation method based on knowledge graph in Example 2.

[0105] Figure 3 This is an architectural diagram of the adaptive learning resource recommendation system based on knowledge graph in Example 3. DETAILED DESCRIPTION

[0106] The accompanying drawings are for illustrative purposes only and are not to be construed as limiting this patent;

[0107] The technical solution of the present invention is further described below with reference to the accompanying drawings and embodiments.

[0108] Example 1

[0109] See also Figure 1 This embodiment proposes an adaptive learning resource recommendation method based on knowledge graph, including:

[0110] A user cognitive diagnosis model is established, and a learning cognitive diagnosis of the user is performed through the user cognitive diagnosis model, and based on the diagnosis results, the user's mastery of knowledge points is predicted.

[0111] Based on the existing I-DINA (I-Deterministic Inputs, Noisy "And" gate model) model, this embodiment constructs a cognitive diagnosis model called AMI-DINA (AMI-Deterministic Inputs, Noisy "And" gate model). The construction of the AMI-DINA cognitive diagnosis model includes calculating the user's potential answers to test questions; calculating the guess rate of the user's correct answers to test questions through manual guessing without mastering the knowledge points involved in the test questions; calculating the error rate of the user's failure to answer test questions correctly due to mistakes under the premise that the user has mastered the knowledge points involved in the test questions; introducing the reference factors of test question difficulty, forgetting curve and number of answers to calculate the probability of the user's correct answer to the test questions; and constructing a user-knowledge point matrix.

[0112] The AMI-DINA cognitive diagnosis model not only considers the impact of error rate and guessing rate on the results of careful diagnosis, but also considers the impact of test question difficulty, forgetting curve and number of answers on the correct answer to the test question.

[0113] A test question score prediction model is established, and the test question score prediction model is used to predict the user's scores on the unfinished test questions.

[0114] This embodiment proposes a test score prediction model called KCAMF (KNN&CNN-Attention&Probabilistic MatrixFactorization). This model deeply mines the implicit features of the test text by utilizing a convolutional neural network with an integrated attention mechanism, integrates a cognitive diagnosis model to obtain implicit features of knowledge points and users, and combines PMF (Probabilistic Matrix Factorization) and the KNN (K-Nearest Neighbor) nearest neighbor algorithm to achieve test score prediction.

[0115] The first candidate learning resource is selected based on the user's mastery of the knowledge points and the user's scores on the unfinished test questions.

[0116] Build a knowledge graph, use the diagnostic results of the user cognitive diagnosis model to locate the knowledge points that the user has mastered well and weakly in the knowledge graph, and select the second candidate learning resources.

[0117] In this embodiment, the semantic relationship between knowledge points is considered. For knowledge points that the user has a good grasp of, the subsequent knowledge points are identified. For example, if the user has a good grasp of knowledge point a, the user can directly learn knowledge point b, and knowledge point b is the successor knowledge point of knowledge point a. For knowledge points that the user has a weak grasp of, the predecessor knowledge points are identified. For example, if the user has a weak grasp of knowledge point a, the user needs to learn knowledge point b, and knowledge point b is the predecessor knowledge point of knowledge point a.

[0118] An optimal learning resource is selected from the first candidate learning resource and the second candidate learning resource, and the optimal learning resource is recommended to the user.

[0119] In terms of the user's cognitive level, by constructing the AMI-DINA cognitive diagnosis model, considering the impact of test question difficulty, forgetting curve and number of answers on the correct answer to the test questions, the user's learning cognitive diagnosis is conducted. According to the diagnosis results, the user's mastery of knowledge points is predicted, and then the KCAMF test question score prediction model is constructed to predict the user's score for the unfinished test questions. The user's mastery of knowledge points and the user's score for the unfinished test questions are used to select the first candidate learning resource; and considering the semantic relationship between knowledge points, the corresponding knowledge graph is constructed. Combined with the user's cognitive diagnosis results, the knowledge points that the user has mastered well and poorly are located in the knowledge graph, and the successor knowledge points of the knowledge points that the user has mastered well are found. The predecessor knowledge points of the knowledge points that the user has mastered weakly are found, and the second candidate learning resource is selected; finally, the most suitable learning resource for the user is selected from the first candidate learning resource and the second candidate learning resource and recommended to the user.

[0120] Example 2

[0121] See also Figure 2 This embodiment proposes an adaptive learning resource recommendation method based on knowledge graph, including:

[0122] Establish a user cognitive diagnosis model, use the user cognitive diagnosis model to perform learning cognitive diagnosis on the user, and predict the user's mastery of knowledge points based on the diagnosis results. The user cognitive diagnosis model in this embodiment uses the AMI-DINA cognitive diagnosis model. Specifically, it includes the following steps:

[0123] Define a set P = {p1, p2, ..., p U} is the user set, set T = {t1, t2, ..., t V} is the set of questions, set C = {c1, c2, ..., c K} is a set of knowledge points; the matrix Q is the knowledge point association matrix of the test question, and each element q in the matrix Q vk Indicates the title t v Knowledge point c k The matrix R is the user score matrix, and each element r in the matrix R uv Indicates user p u In the test question v The score on the matrix η is the user's potential answer matrix, and each element η in the matrix η uv Indicates user p u In the test question v Potential answer situation.

[0124] Define the user's potential learning state as a multi-dimensional knowledge point mastery vector α uk ={α u1 , α u2 ,...,α uK}, α uk Each dimension corresponds to user p u Knowledge point c k The mastery of the situation, when α uk =1 indicates user p u Master knowledge point c k , when α uk =0 indicates that user p u Not mastered knowledge point c k ; Then calculate user p u In the test question v Potential answer situation η uv , whose formula is as follows:

[0125]

[0126] Among them, when ηuv =1; when η uv =1, indicating that user p u Able to answer test questions correctly v And mastered the test questions v Knowledge points involved c k When η uv =0, indicating that user p u Unable to answer the test questions correctly v and have not mastered the test questions v Knowledge points involved c k .

[0127] Calculate user p u In the test questions that have not been mastered v Knowledge points involved c k Under the premise of human guessing, the correct answer to the test question is obtained. v The guess rate g v , its formula is as follows:

[0128] g v =P(r uv =1|η uv =0)

[0129] Among them, r uv Indicates user p u In the test question v The score on r uv =1, indicating that user p u Answer the question correctly v .

[0130] Calculate user p u In mastering the test questions v Knowledge points involved c k Under the premise that the test questions cannot be answered correctly due to mistakes, v Error rate v , its formula is as follows:

[0131] s v =P(r uv =0|η uv =1)

[0132] Among them, when r uv =0, indicating that user p u Failed to answer the test questions correctly v .

[0133] According to the guess rate g v , error rate v and potential answer situation η uv , and considering the difficulty of the test questions, the forgetting curve and the number of answers, calculate the user pu In the test question v The probability of a correct answer on , is given by:

[0134]

[0135] Among them, ω represents the difficulty of the test question, Used to fit the forgetting curve; ξ represents user p u In the test question v When the user p u Correctly answer the test questions v ξ takes the value of 1, when user p u Incorrectly answered questions v , ξ takes the value of 0; frequency represents the number of times the question is answered.

[0136] According to user p u In the test question v The probability of correct answer on the ,determine the user's multi-dimensional knowledge point mastery vector α u And calculate the vector α u Estimated value, using vector α u The estimated values ​​construct the user-knowledge point matrix B. The rows of the matrix B represent users and the columns represent knowledge points, that is, the row vector represents each user's mastery of the knowledge point.

[0137] In this embodiment, since the constructed formula contains hidden variables that cannot be directly observed, that is, the user's knowledge level, it is necessary to use the EM algorithm to maximize the marginal likelihood probability of the formula to obtain the estimated values ​​of the two test parameters corresponding to each question. and At the same time, the user's multi-dimensional knowledge point mastery vector α is determined by the maximum a posteriori probability algorithm u Estimated value of The formula is as follows:

[0138]

[0139] Among them, α represents the knowledge point mastery vector, R u Represents the score matrix of user u.

[0140] The EM algorithm can solve the parameter estimation problem for incomplete data. When the amount of missing information is large or the data volume is very large, the AMI-DINA cognitive diagnosis model achieves acceleration by dividing the user data into multiple user blocks, accessing one user block at each iteration, and using multi-threaded training techniques.

[0141] According to the user-knowledge point matrix B, predict user p u Knowledge point c k The grasp of the situation.

[0142] When the multi-dimensional knowledge point mastery vector α corresponding to all users is obtained u After this, the final user-knowledge point matrix B is obtained. This matrix has two main functions: First, it combines the user-knowledge point matrix B with the test question knowledge point association matrix Q to predict the performance of all users. Second, it uses the user-knowledge point matrix B to locate the user's weak and strong knowledge points in the knowledge graph. Based on the order of the knowledge points on the knowledge graph, it then provides targeted, high-precision learning resource recommendations for the user.

[0143] The AMI-DINA cognitive diagnostic model introduces parameters such as guessing rate, error rate, test question difficulty parameter, fitted forgetting curve and number of answers to more accurately diagnose the user's cognitive learning status at the multi-dimensional knowledge level.

[0144] Establishing a KCAMF test question score prediction model, and using the KCAMF test question score prediction model to predict the user's score on the unfinished test questions, specifically including the following steps:

[0145] A KCAMF test question score prediction model is established, and the test questions are input into the KCAMF test question score prediction model to obtain the implicit feature vector corresponding to the test questions.

[0146] In this embodiment, the attention mechanism is integrated into the convolutional neural network in the KCAMF test score prediction model to obtain a CNN-Attention model, which includes an embedding layer, an attention layer, a convolution layer, a pooling layer, and an output layer. The process of obtaining the implicit feature vector corresponding to the test question is as follows:

[0147] The relevant text of the test question (including questions, answers, analysis and comments) is input into the embedding layer, and the embedding layer outputs the d-dimensional vector a for each word in the text n , using vector a n Construct a dense digital matrix S, whose formula is as follows:

[0148]

[0149] Input the dense digital matrix S into the attention layer to highlight the features of the key words in the text and obtain the feature expression of the key words The specific calculation is as follows:

[0150]

[0151] s(i)=f(H i *W att +b att )

[0152] Among them, A represents the word feature representation obtained through the attention mechanism; D represents a central word h i The context range, s(i) represents the feature value weight of each word in the context range, W att Indicates the matrix parameters corresponding to the context range, b att represents the bias, and f(·) represents the activation function.

[0153] Each word feature w in the context i Multiplying the word's feature value weight s(i) outputs a new feature representation The expression formula is as follows:

[0154]

[0155] Represent the features Input the convolution layer to extract the features C(i) of the key words. The formula is as follows:

[0156]

[0157] Among them, W c represents the shared weight, b c represents the bias, and f(·) represents the activation function.

[0158] The feature C(i) is input into the pooling layer, and the pooling layer outputs the representative feature C out (i)

[0159] C out (i) = MAX(C(i))

[0160] The representative feature C out (i) Input and output layer, obtain the implicit feature vector D corresponding to the relevant text of the test question j :

[0161]

[0162] in, and is the mapping matrix, and and bias vector, f(·) represents the activation function; W represents all weights and biases, H j Represents the word vector corresponding to the relevant text of the test question.

[0163] According to the implicit feature vector corresponding to the test question, the probability matrix is ​​used to identify the test questions of user p. u In the test question v The score on r uvThe constructed user test score matrix R is decomposed to obtain the user implicit feature matrix and the test question implicit feature matrix, which specifically includes the following steps:

[0164] Assume that the user test score matrix R is decomposed into the user implicit feature matrix U, the knowledge point implicit feature matrix K and the test question implicit feature matrix V through the probability matrix;

[0165] The prior probabilities of the user implicit feature matrix U and the knowledge point implicit feature matrix K follow Gaussian distribution and are independent of each other:

[0166]

[0167]

[0168] Among them, m represents the number of users, l represents the number of knowledge points, and represents variance, I represents unit vector;

[0169] The probability distribution of the implicit feature matrix V of the test question is:

[0170]

[0171]

[0172] Among them, W z represents the weights between neurons in a convolutional neural network, represents the variance, n represents the number of test questions, represents variance;

[0173] After the implicit feature vectors corresponding to the user and the test questions are given, the conditional probability distribution of the user test question score matrix R is:

[0174]

[0175] in, represents the variance, represents the implicit features obtained by probability matrix decomposition, Represents the indicator function. If student i has done test j, then otherwise

[0176] The conditional probability distribution of the user knowledge point matrix B is:

[0177]

[0178] Among them, l represents the number of knowledge points, b ic Indicates user i’s mastery of knowledge point c, K c represents the implicit feature vector of the knowledge point, represents the variance, represents the indicator function;

[0179] Combining the probability distribution of the user implicit feature matrix U, the knowledge point implicit feature matrix K, the test question implicit feature matrix V, the user test question score matrix R, and the user knowledge point matrix B, and using the Bayesian criterion and maximization technique, we can finally get the objective function E(K, U, V, B, R) as follows:

[0180]

[0181]

[0182] Among them, b ij Indicates user i’s mastery of knowledge point j, W i Represents the weight corresponding to the convolutional neural network;

[0183] The objective function E(K, U, V, B, R) is solved by the stochastic gradient descent method to obtain the optimal solutions of the user implicit feature matrix and the test question implicit feature matrix.

[0184] S2.3: Based on the obtained user implicit feature matrix and test question implicit feature matrix, use the KNN algorithm to correct the user test question score matrix R after probability matrix decomposition to obtain the final user score for the unfinished test questions.

[0185] Although PMF can overcome the problem of excessive sparsity of the test score matrix R, the test score matrix R is prone to lose some hidden feature information during the matrix decomposition process. Therefore, this embodiment introduces the KNN algorithm to correct the score of each test question. Specifically, it is assumed that the KNN algorithm is used to find the N test questions most similar to test question j in the test question set that user i has completed, to make up for the implicit feature information lost due to PMF, and to correct the user test score matrix to obtain the final test question score R ij :

[0186]

[0187]

[0188] Among them, R ij Indicates the final predicted score of the user for the unfinished test questions, R′ ij is a correction factor used to compensate for the accuracy problem caused by partial information loss due to matrix decomposition; R in represents the user’s original score on the test question, V n The feature vector representing the test question.

[0189] The first candidate learning resource is selected based on the user's mastery of the knowledge points and the user's scores on the unfinished test questions.

[0190] In this embodiment, the user's mastery of knowledge points and the user's scores on unfinished test questions are ranked, and combined with the test question difficulty coefficient λ, candidate learning resources are selected within the user's expected difficulty range. The calculation formula for λ is as follows:

[0191]

[0192] The user's potential answer situation η uv The difficulty coefficient of the test question is obtained by normalizing it to the range of [0, 1]. Finally, the learning resources whose difficulty coefficient meets the expected (λ1, λ2) are selected from the learning resources as candidate learning resources.

[0193] Build a knowledge graph, use the diagnostic results of the user cognitive diagnosis model to locate the knowledge points that the user has mastered well and weakly in the knowledge graph, and select the second candidate learning resources.

[0194] Knowledge graphs can quickly describe the concepts of entities in the physical world and the relationships between them, while also expressing a wider range of knowledge content and semantic relationships. The application of knowledge graphs in learning resource recommendation is still in its infancy, with issues such as over-reliance on subject matter experts in the construction process, a low level of automation, and fuzzy knowledge granularity, with the use of graphs primarily to express concepts without fully representing knowledge elements.

[0195] Therefore, based on the traditional knowledge graph, this paper proposes a method for constructing a knowledge graph suitable for learning resource recommendation. This method uses an additional binary pair (centrality and included learning resources) for each knowledge meta-node to provide a more in-depth and detailed description. Specifically, it includes the following steps:

[0196] Building a knowledge graph:

[0197] The text dataset is collected using crawler technology; the text dataset includes structured data, semi-structured data and unstructured data.

[0198] Perform entity extraction on text datasets to obtain entities in the text datasets.

[0199] Entity extraction is also known as named entity recognition, which is the automatic identification of named entities from a text dataset. This embodiment uses the Bi-LSTM+CRF model to complete test question recognition. The Bi-LSTM+CRF model consists of three layers. The first layer is the input layer, which is used to convert text data into word vectors and character vectors; the second layer is a double-layer LSTM, which is used to input the word vectors and character vectors obtained in the first layer into the Bi-LSTM to obtain the scores of all labels for each word; the third layer is the sequence annotation layer, which is used to take the output of the second layer as input to obtain the probability value of the label sequence.

[0200] Perform relationship extraction on entities to obtain the relationships between entities.

[0201] Relationship extraction is used to determine whether there is a certain defined relationship type between the extracted entities. Currently, the present invention uses a bidirectional GRU combined with an Attention model to perform relationship extraction on the entities extracted in step 2.

[0202] Artificial knowledge fusion is performed on entities and the relationships between them, and the fusion results are used to construct a knowledge graph, which is then stored in a database.

[0203] There may be a large amount of duplicate or erroneous information in the triples extracted through entities and relationships. Since the knowledge graph constructed in this embodiment is ultimately used as supporting data for personalized learning resource recommendations, it is necessary to perform manual knowledge fusion processing on the entities and relationships to ensure the accuracy of the professional terms of the nodes in the knowledge graph and the relationships between the nodes.

[0204] There are two main storage methods for knowledge graphs: RDF and graph databases. Compared to RDF, graph databases prioritize efficient graph queries and searches. Graph databases generally use property graphs as their primary representation, allowing entities and relationships to include attributes, making it easier to express real-world business scenarios. Therefore, in this example, knowledge point entities and their relationships are stored in the graph database Neo4j.

[0205] Calculate the centrality of knowledge meta-nodes in the knowledge graph.

[0206] Currently, there is no unified definition of centrality in academia. It is generally considered to be a measure of the importance of a certain attribute in a network node. This embodiment combines the characteristics of knowledge points and the learning patterns of users to define knowledge meta-centrality as the influence of a knowledge meta-node in the knowledge graph.

[0207] The centrality of a knowledge element is the result of the interaction and contribution of all knowledge points in the knowledge graph. The closer the distance between two knowledge element nodes, the stronger their mutual dependence. The dependency between knowledge points shows an exponential decay trend as the distance increases. At the same time, it is necessary to consider the different requirements for different knowledge points in the field of learning resource recommendation. According to the education syllabus, the importance of knowledge points can be divided into understanding, comprehension, mastery, and application, and assigned weights of 0.1, 0.2, 0.3, and 0.4, respectively. The formula for calculating centrality is as follows:

[0208]

[0209] Among them, G i represents the centrality of knowledge element i, δ represents the contribution of the knowledge element, and λ represents the weight corresponding to the knowledge element; α and β are both weight coefficients, indicating the contribution of the knowledge element itself and its neighboring knowledge elements of each order to the total centrality.

[0210] Classify the learning resources contained in the knowledge meta-node.

[0211] This embodiment uses the Bi-LSTM+Attention model to automatically label the knowledge point labels of the questions and classify the learning resources contained in the knowledge meta-nodes.

[0212] The Bi-LSTM+Attention model integrates the Attention layer into the Bi-LSTM model. The Attention layer first calculates the weight of each time series, then takes the weighted sum of all time series vectors as the feature vector, and finally performs Softmax classification.

[0213] Using the diagnostic results of the user cognitive diagnosis model, the knowledge points that the user has mastered well and weakly are located in the knowledge graph, and the second candidate learning resources are selected.

[0214] In this embodiment, the semantic relationships between knowledge points are considered. For knowledge points that the user has a good grasp of, their successor knowledge points are identified. For example, if the user has a good grasp of knowledge point a and can directly learn knowledge point b, then knowledge point b is the successor knowledge point of knowledge point a. For knowledge points that the user has a weaker grasp of, their predecessor knowledge points are identified. For example, if the user has a weaker grasp of knowledge point a and needs to learn knowledge point b, then knowledge point b is the predecessor knowledge point of knowledge point a. After identifying the successor and predecessor knowledge points, the knowledge points are sorted according to their centrality and, combined with the difficulty coefficient λ, the second candidate learning resource is selected.

[0215] An optimal learning resource is selected from the first candidate learning resource and the second candidate learning resource, and the optimal learning resource is recommended to the user.

[0216] In terms of the user's cognitive level, by constructing the AMI-DINA cognitive diagnosis model, considering the impact of test question difficulty, forgetting curve and number of answers on the correct answer to the test questions, the user's learning cognitive diagnosis is conducted. According to the diagnosis results, the user's mastery of knowledge points is predicted, and then the KCAMF test question score prediction model is constructed to predict the user's score for the unfinished test questions. The user's mastery of knowledge points and the user's score for the unfinished test questions are used to select the first candidate learning resource; and considering the semantic relationship between knowledge points, the corresponding knowledge graph is constructed. Combined with the user's cognitive diagnosis results, the knowledge points that the user has mastered well and poorly are located in the knowledge graph, and the successor knowledge points of the knowledge points that the user has mastered well are found. The predecessor knowledge points of the knowledge points that the user has mastered weakly are found, and the second candidate learning resource is selected; finally, the most suitable learning resource for the user is selected from the first candidate learning resource and the second candidate learning resource and recommended to the user.

[0217] The pseudo code of this embodiment is as follows:

[0218] enter:

[0219] Learner set P, test set T, learner-test score matrix R, test-knowledge point association matrix Q, domain knowledge graph KG

[0220] Output:

[0221] Recommended learning resource set T rec

[0222] 1.for each learner p u :

[0223] 2. According to formula (2-5) and combined with the test question-knowledge point association matrix Q, calculate the learner's knowledge point mastery vector α u

[0224] 3.end for

[0225] 4. According to formula (2-20), continuously update the user feature matrix and the test feature matrix until the termination condition is met

[0226] 5.For each test question T v ∈T:

[0227] 6. Calculate learner p according to formula (2-22) u The predicted score on this question μ v

[0228] 7. Calculate the difficulty coefficient λ corresponding to the test question according to formula (2-23)

[0229] 8.endfor

[0230] 9. Select the test questions with difficulty coefficient λ in the difficulty range (λ1, λ2) from the test question set as T1

[0231] 10. Master the vector α according to the learner’s knowledge points u , calculate the knowledge points C1 and C2 that the learners have mastered firmly and weakly, and sort C1 and C2 in descending order according to the calculation results of formula (2-6).

[0232] 11.For each learner p u :

[0233] 12.For each knowledge point c∈C1:

[0234] 13. Locate knowledge point c in KG and find all target knowledge points that are successor to c

[0235] 14. Select questions with a difficulty coefficient within the difficulty range (λ1, λ2) from the learning resources corresponding to the knowledge point

[0236] 15.As T2

[0237] 16.end for

[0238] 17.endfor

[0239] 18.For each learner p u :

[0240] 19.For each knowledge point c∈C2:

[0241] 20. Locate knowledge point c in KG and find all target knowledge points that have a predecessor relationship with c

[0242] 21. Select questions with a difficulty coefficient within the difficulty range (λ1, λ2) from the learning resources corresponding to the knowledge point

[0243] 22.As T3

[0244] 23.end for

[0245] 24.endfor

[0246] 25. Final test question recommendation set T rec ←T1∪T2∪T3

[0247] Example 3

[0248] See also Figure 3,This embodiment proposes an adaptive learning resource recommendation system based on knowledge graph,,including: a user cognitive diagnosis module, a test score prediction module, a knowledge graph construction module, a learning resource acquisition module and a recommendation module.

[0249] During the specific implementation process, the user conscientiousness diagnosis module obtains the user's test data, the test knowledge point correlation matrix, and the user score matrix. It uses the AMI-DINA cognitive diagnosis model to calculate the user's guess rate, error rate, and potential answer situation. It also takes into account the test difficulty, forgetting curve, and the number of answers to calculate the user's correct answer probability on the test question to obtain the user's mastery of the knowledge points. The user's mastery of the knowledge points is output to the learning resource acquisition module.

[0250] The test question score prediction module obtains the implicit feature vector corresponding to the test question from the test question, and uses the probability matrix to decompose the user test question score matrix based on the implicit feature vector corresponding to the test question. The KNN algorithm is used to correct the user test question score matrix R after the probability matrix decomposition to obtain the final test question prediction score, and the final test question prediction score is output to the learning resource acquisition module.

[0251] The knowledge graph construction module constructs the basic knowledge graph, calculates the centrality of the knowledge meta-nodes in the knowledge graph, classifies the learning resources contained in the knowledge meta-nodes, and uses the diagnosis results of the user cognitive diagnosis model to locate the knowledge points that the user has mastered well and weakly in the knowledge graph, and outputs the successor knowledge points and predecessor knowledge points.

[0252] The learning resource acquisition module selects the first candidate learning resource based on the output results of the user cognitive diagnosis module and the test score prediction module, combined with the user's expected difficulty range; selects the second candidate learning resource based on the output results of the knowledge graph construction module, combined with the user's expected difficulty range, and outputs the first and second candidate learning resources to the recommendation module.

[0253] The recommendation module selects the best learning resource from the first candidate learning resource and the second candidate learning resource, and recommends the best learning resource to the user.

[0254] The terms used in the drawings to describe positional relationships are for illustrative purposes only and should not be construed as limiting this patent;

[0255] Obviously, the above embodiments of the present invention are merely examples for the purpose of clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. Those skilled in the art will appreciate that other variations or modifications can be made based on the above description. It is not necessary and impossible to enumerate all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the claims of the present invention.

Claims

1. A method for adaptive learning resource recommendation based on knowledge graph, characterized in that: include: Establishing a user cognitive diagnosis model, conducting a learning cognitive diagnosis on the user through the user cognitive diagnosis model, and predicting the user's mastery of knowledge points based on the diagnosis results; Establish a test score prediction model, and use the test score prediction model to predict the user's score for the unfinished test questions, including: Establishing a test question score prediction model, inputting the test question into the test question score prediction model, and obtaining an implicit feature vector corresponding to the test question; According to the implicit feature vector corresponding to the test question, the probability matrix is ​​used to In the test questions Scores on Constructed user test score matrix R Decompose to obtain the user implicit feature matrix and the test question implicit feature matrix; According to the user implicit feature matrix and the test question implicit feature matrix, the user test question score matrix after probability matrix decomposition is performed using the KNN algorithm R Make corrections to obtain the final user's scores for the unfinished test questions; Select the first candidate learning resource based on the user's mastery of knowledge points and the user's scores on unfinished test questions; Build a knowledge graph and use the diagnostic results of the user cognitive diagnosis model to locate the knowledge points that the user has good and weak grasp of in the knowledge graph, and select the second candidate learning resources; An optimal learning resource is selected from the first candidate learning resource and the second candidate learning resource, and the optimal learning resource is recommended to the user.

2. The adaptive learning resource recommendation method based on knowledge graph according to claim 1 is characterized in that: The user's learning cognitive diagnosis is performed on the user through the user cognitive diagnosis model, and based on the diagnosis results, the user's mastery of the knowledge points is predicted, specifically including: Defining a Collection For user collection, collection Collection, Collection is a collection of knowledge points; Define the user's potential learning state as a multi-dimensional knowledge point mastery vector , Each dimension corresponds to a user Knowledge points The grasp of the situation, when =1 indicates the user Master the knowledge points ,when =0 means user Not mastered the knowledge points ; Then calculate the user In the test questions Potential answers on , whose formula is as follows: Among them, when When the user Able to answer test questions correctly and mastered the test questions Knowledge points involved ;when When the user Unable to answer the test questions correctly and have not mastered the test questions Knowledge points involved ; Calculating users Not mastering the test questions Knowledge points involved Under the premise of human guessing, the correct answer is obtained Guess rate , its formula is as follows: in, Represents a user In the test questions The score on =1, indicating that the user Answer the test questions correctly ; Calculating users Mastering the test questions Knowledge points involved Under the premise that the test questions are not answered correctly due to mistakes Error rate , its formula is as follows: Among them, when =0, indicating that the user Failure to answer the test questions correctly ; According to the guess rate , error rate and potential answers , and taking into account the difficulty of the test questions, the forgetting curve and the number of answers, calculate the user In the test questions The probability of a correct answer on , is given by: in, Indicates the difficulty of the test question. Used to fit the forgetting curve; Represents a user In the test questions When the user Answer the test questions correctly , The value is 1, when the user Incorrectly answered questions , The value is 0; frequency Indicates the number of times the question was answered; According to the user In the test questions The probability of correct answer on the ,determine the user's multi-dimensional knowledge point mastery vector And calculate the vector Estimates, using vector The estimated value constructs the user-knowledge point matrix B; According to the user-knowledge point matrix B, predict the user Knowledge points The grasp of the situation.

3. The adaptive learning resource recommendation method based on knowledge graph according to claim 2 is characterized in that: Using vectors The estimated value constructs the user-knowledge point matrix B specifically including: Use EM algorithm to maximize the guessing rate and error rate The marginal likelihood probability of and error rate Estimated value of and ; Determine the user's multi-dimensional knowledge point mastery vector through the maximum a posteriori probability algorithm Estimated value of , whose formula is as follows: in, Represents the knowledge point mastery vector, Represents a user u The scoring matrix of Master vectors based on the user's multi-dimensional knowledge points Estimated value of , construct the user-knowledge point matrix B.

4. The adaptive learning resource recommendation method based on knowledge graph according to claim 2 is characterized in that: The test score prediction model uses a convolutional neural network and an attention mechanism to obtain the implicit feature vector corresponding to the test question. The convolutional neural network includes an embedding layer, an attention layer, a convolution layer, a pooling layer, and an output layer. Specifically: The relevant text of the test question is input into the embedding layer, and the embedding layer outputs the d-dimensional vector of each word in the text , using vector Constructing dense digital matrices S , whose formula is as follows: Dense digital matrix S Input the attention layer to highlight the features of key words in the text and obtain the feature expression of key words , the specific calculation is as follows: in, A Represents the word feature representation obtained through the attention mechanism; D Indicates a central word The context scope, Represents the feature value weight of each word in the context, Represents the matrix parameters corresponding to the context range, Indicates the offset, represents the activation function; Each word feature in the context The feature value weight of the word Multiply and output new feature representation , and its expression formula is as follows: ； Represent the features Input convolution layer to extract the features of key words , whose formula is as follows: in, represents the shared weight, Indicates bias, represents the activation function; The features Input pooling layer, pooling layer outputs representative features : The representative features Input and output layer, get the implicit feature vector corresponding to the relevant text of the test question : in, and is the mapping matrix, and With the bias vector, represents the activation function; represents all weights and biases, Represents the word vector corresponding to the relevant text of the test question.

5. The adaptive learning resource recommendation method based on knowledge graph according to claim 4 is characterized in that: Use the probability matrix to In the test questions Scores on Constructed user test score matrix R Decomposition is performed to obtain the user implicit feature matrix and the test question implicit feature matrix, which specifically include: Assume that the user test score matrix R Decompose the user implicit feature matrix through probability matrix , knowledge point implicit feature matrix and the implicit feature matrix of the test question V ; User implicit feature matrix And the knowledge point implicit feature matrix The prior probabilities of follow a Gaussian distribution and are independent of each other: in, Indicates the number of users, Indicates the number of knowledge points, and represents the variance, represents a unit vector; Implicit feature matrix of test questions V The probability distribution of is: in, represents the weights between neurons in a convolutional neural network, represents the variance, n Indicates the number of test questions, represents variance; Given the implicit feature vectors corresponding to the user and the test question, the user-test-question score matrix R The conditional probability distribution of is: in, represents the variance, represents the implicit features obtained by probability matrix decomposition, Represents the indicator function. If student i has done test j, then ,otherwise ; User knowledge matrix B The conditional probability distribution of is: in, Indicates the number of knowledge points, Indicates user i’s mastery of knowledge point c, represents the implicit feature vector of the knowledge point, represents the variance, represents the indicator function; Combined with user implicit feature matrix , knowledge point implicit feature matrix , Implicit feature matrix of test questions V , user test score matrix R and user knowledge matrix B The probability distribution of , and the use of Bayesian criteria and maximization techniques, can finally get the objective function as follows: in, Represents a user i Knowledge points j The grasp of Represents the weight corresponding to the convolutional neural network; The objective function is obtained by stochastic gradient descent method. Solve it to obtain the optimal solution of the user implicit feature matrix and the test question implicit feature matrix.

6. The adaptive learning resource recommendation method based on knowledge graph according to claim 5 is characterized in that: According to the user implicit feature matrix and the test question implicit feature matrix, the user test question score matrix after probability matrix decomposition is performed using the KNN algorithm R Correction is performed to obtain the final predicted scores of users for the unfinished test questions, specifically: Assume that the user i Use KNN algorithm to find the questions that match the completed questions in the set of questions. j The N most similar test questions are used to make up for the implicit feature information lost due to probability matrix decomposition, and the user's test score matrix is ​​corrected to obtain the final test score. : in, Indicates the final predicted score of the user for the unfinished test questions, is the correction factor, Indicates the user's original score on the test question, The feature vector representing the test question.

7. The method for adaptive learning resource recommendation based on knowledge graph according to any one of claims 1 to 6, characterized in that: The second candidate learning resources include: Build a knowledge graph; The formula for calculating the centrality of knowledge meta-nodes in the knowledge graph is as follows: in, Representing knowledge elements i The centrality of represents the contribution of knowledge element, Indicates the weight corresponding to the knowledge element; and are weight coefficients, indicating the contribution of the knowledge element itself and its neighboring knowledge elements of each order to the total centrality; Classify the learning resources contained in the knowledge meta-node; Using the diagnostic results of the user cognitive diagnosis model, the knowledge points that the user has mastered well and weakly are located in the knowledge graph, and the second candidate learning resources are selected.

8. The method for adaptive learning resource recommendation based on knowledge graph according to claim 7, characterized in that: Building a knowledge graph specifically includes: Collect text datasets; Perform entity extraction on the text dataset to obtain the entities in the text dataset; Extract relationships between entities to obtain the relationships between entities; Artificial knowledge fusion is performed on entities and the relationships between them, and the fusion results are used to construct a knowledge graph, which is then stored in a database.

9. A knowledge graph-based adaptive learning resource recommendation system, applied to the knowledge graph-based adaptive learning resource recommendation method according to any one of claims 1 to 8, characterized in that: include: The user cognitive diagnosis module is used to conduct learning cognitive diagnosis on users and predict their mastery of knowledge points based on the diagnosis results; The test score prediction module is used to predict the user's score for the unfinished test questions; The knowledge graph construction module is used to construct the knowledge graph and use the diagnosis results of the user cognitive diagnosis module to locate the knowledge points that the user has a good grasp of and the knowledge points that the user has a weak grasp of in the knowledge graph; The learning resource acquisition module is used to obtain the first candidate learning resource based on the output results of the user cognitive diagnosis module and the test score prediction module; and the second candidate learning resource based on the output results of the knowledge graph construction module; The recommendation module is used to select the best learning resource from the first candidate learning resource and the second candidate learning resource, and recommend the best learning resource to the user.

Images