A dynamic graph neural cognitive diagnosis method fusing multi-dimensional features

Abstract

The application provides a dynamic graph neural cognitive diagnosis method fusing multi-dimensional features, and the method comprises the following steps: collecting multi-modal data of an online learning platform, and obtaining a multi-dimensional feature splicing vector based on the multi-modal data; based on the multi-dimensional feature splicing vector and by using a graph neural network, obtaining an aggregation matrix from an embedding matrix of a current knowledge point, a hidden state matrix of the current knowledge point, an embedding matrix of a neighboring knowledge point, and a hidden state matrix of the neighboring knowledge point; performing a forgetting process on the embedding matrix of the knowledge point by using a forgetting vector, inputting the aggregation matrix into a memory gate structure model, and capturing a time sequence feature embedding matrix of the knowledge point; based on the time sequence feature embedding matrix and by using the graph neural network, constantly updating the hidden state matrix of the knowledge point, and obtaining an updated knowledge mastery degree of the learner at the next moment. The application comprehensively diagnoses the learning effect of the learner and updates the knowledge mastery state of the learner, and greatly improves the accuracy of the diagnosis result.

Description

Technical Field

[0001] This invention relates to the field of intelligent education technology, and in particular to a dynamic graph neurocognitive diagnostic method that integrates multidimensional features. Background Technology

[0002] Cognitive diagnosis (CD) is a fundamental and crucial task in the field of educational technology. By analyzing learners' behavioral data on online education websites, CD can help them master different knowledge points or skills and predict their performance on learning tasks. It is one of the key technologies for solving applications such as personalized test question recommendation and learning path planning in online education.

[0003] Currently, the main cognitive diagnostic methods include psychometric-based cognitive diagnostic methods and deep learning-based cognitive diagnostic methods. These methods can improve teaching effectiveness and learners' learning outcomes in the education field.

[0004] Psychometric-based cognitive diagnostic methods typically describe learners' characteristics and the features of practice or testing tasks from a specific cognitive perspective, and use empirical response functions to obtain the relationship between learners' test scores and their mastery of knowledge points. Examples of empirical response models include Item Response Theory (IRT), Multidimensional Item Response Theory (MIRT), and Deterministic Inputs, Noisy-And gate model (DINA). While these psychometric-based cognitive diagnostic methods are easy to understand and implement, they primarily target learners' mastery of knowledge points at a single moment. This not only results in low accuracy but also fails to predict learners' mastery at the next moment, thus failing to identify which knowledge points have regressed and require reinforcement, and which knowledge points are already sufficiently mastered and do not require additional training. This leads to limitations in practical applications.

[0005] Deep learning-based cognitive diagnostic methods collect information related to learning resources and learners' learning data. Using deep neural networks, they model the learner's learning process and mastery of knowledge points, thereby extracting hidden information from the collected data and predicting the learner's mastery of knowledge points at the next moment, thus revealing changes in the learner's abilities. Examples include Long Short-Term Memory (LSTM) networks, Neural Cognitive Diagnosis Models (NeuralCDM), and Dynamic Key-Value Memory Networks (DKVMN).

[0006] Cognitive diagnostic methods based on LSTM or NeuralCDM employ attention mechanisms to identify the impact of learners' historical test interaction data on their knowledge state at each moment, and use temporal convolutional networks to extract learners' dynamically changing mastery of knowledge points. DKVMN-based cognitive diagnostic methods predict learners' performance on new test questions by calculating the correlation weights between test questions and knowledge points, thus obtaining learners' dynamic mastery of knowledge points. While these methods can perform cognitive diagnosis and dynamic prediction of learners' answering processes, they fail to fully explore the spatial relationships (parallelism, hierarchy, or sequence) between different knowledge points when diagnosing learners' mastery of knowledge points. This makes them static and unsuitable for dynamic assessment of learners' long-term learning. Furthermore, they do not utilize semantic information, difficulty, and discrimination in a particular test task, and ignore situations where a single test task corresponds to multiple knowledge points in different contexts. This reduces the accuracy of diagnosing and predicting learners' mastery of individual knowledge points during long-term dynamic learning. Summary of the Invention

[0007] This invention provides a dynamic graph neurocognitive diagnostic method that integrates multidimensional features to solve the problem of low accuracy in current cognitive diagnostic methods.

[0008] This invention provides a dynamic graph neurocognitive diagnostic method that integrates multidimensional features. The method includes:

[0009] Collect multimodal data from an online learning platform, and obtain a multidimensional feature concatenation vector based on the multimodal data;

[0010] Based on the multidimensional feature concatenation vector, and using a graph neural network, the embedding matrix of the current knowledge point, the hidden state matrix of the current knowledge point, the embedding matrix of adjacent knowledge points, and the hidden state matrix of adjacent knowledge points are aggregated to obtain an aggregated matrix;

[0011] The embedding matrix of knowledge points is processed by forgetting vectors, and the aggregation matrix is ​​input into the memory gate structure model to capture the temporal feature embedding matrix of knowledge points.

[0012] Based on the temporal feature embedding matrix and by continuously updating the hidden state matrix of knowledge points using a graph neural network, the learner's mastery of knowledge points after the next time step is obtained.

[0013] Furthermore, obtaining the multidimensional feature concatenation vector based on the multimodal data includes:

[0014] Based on the multimodal data, extract the first type of data and the second type of data;

[0015] The first type of data is decomposed, and test task information and knowledge point information from different decomposition semantic perspectives are fused to obtain test task semantic features.

[0016] Based on the first type of data, the second type of data, and the semantic features of the test task, a multi-dimensional feature concatenation vector is obtained.

[0017] Furthermore,

[0018] The first type of data includes: learner sets, test task sets, knowledge point sets, and learner record datasets of test task responses;

[0019] The second type of data includes: test task difficulty data, learner scores, the time interval between learners learning the same knowledge point and the previous time, the time interval between learners learning the same knowledge point and the number of times learners have repeated the knowledge point.

[0020] Furthermore, the method also includes: based on the learner's mastery of the knowledge points, the relationships between the knowledge points, the difficulty data of the test task, the discrimination data of the test task, and the semantic features of the test task, using an item response theory model or a neurocognitive diagnostic model to predict the probability that the learner will correctly complete the test task in the next moment.

[0021] Furthermore, obtaining the semantic features of the test task includes:

[0022] Using a d-dimensional trainable matrix and Initialize the test tasks and knowledge points separately;

[0023] The interaction information between learners, test tasks, and knowledge points is decomposed from the perspectives of co-occurrence and answer consistency, respectively, to obtain a relationship matrix;

[0024] Under the constraint of the loss function, the relation matrix is ​​input into the test task embedding based on node-level attention aggregation and semantic-level attention aggregation to fuse test task information and knowledge point information from different decomposition semantic perspectives to obtain the semantic features of the test task.

[0025] Furthermore, methods for obtaining multi-dimensional feature concatenation vectors based on the first type of data, the second type of data, and the semantic features of the test task include:

[0026] Based on the second type of data, a test task difficulty matrix and a test task representation matrix of learners' test task completion status are obtained;

[0027] The semantic features of the test task and the difficulty matrix of the test task are combined to obtain the task feature vector;

[0028] The test task is decomposed into multiple knowledge point vectors based on the relationship between the test task and the knowledge points.

[0029] The task feature vector, the test task representation matrix, and multiple knowledge point vectors are concatenated to obtain a multi-dimensional feature concatenation vector for each knowledge point at the current time.

[0030] Furthermore, the aggregation includes:

[0031] The hidden state matrix and the first embedding matrix of the current knowledge point are aggregated to obtain the first matrix of the current knowledge point;

[0032] The hidden state matrix and the second embedding matrix of adjacent knowledge points are aggregated to obtain the second matrix of adjacent knowledge points;

[0033] The aggregation matrix is ​​calculated based on the first matrix and the second matrix.

[0034] Furthermore,

[0035] The aggregation matrix The calculation method is as follows:

[0036] Among them, the The hidden state matrix representing knowledge point k; x t w represents the multidimensional feature concatenation vector. x ∈R N*e , representing the index of knowledge points and the first embedding matrix of the completion status of each test task; w c ∈R N*e The second embedding matrix, w, represents the knowledge points. c (v) represents the second embedding matrix w c In the v-th row, e is the embedding value.

[0037] Furthermore, the aggregation matrix is ​​input into the memory gate structure model to capture the temporal feature embedding matrix of knowledge points, including:

[0038]

[0039]

[0040]

[0041]

[0042]

[0043] Among them, f neighbor This represents a function that defines how information is propagated to neighboring nodes. and All are multilayer perceptrons, [] represents the stitching operation; A i,k A represents the relationship between knowledge point i and knowledge point k; k,i This represents the relationship between knowledge point k and knowledge point i; W1 represents the forgetting vector of knowledge point i; e*4e W2∈R e*4e W3∈R e*5e W4∈R e*2e Both represent weight matrices; b1∈R e b2∈R e b3∈R e b4∈R e , both represent offset; The embedding matrix representing the temporal features of knowledge point i; The aggregation matrix representing knowledge point i at time t; The aggregation matrix represents knowledge point k; The aggregation matrix represents the knowledge point i at time t-1.

[0044] Furthermore, the hidden state matrix of knowledge points, based on the aforementioned temporal feature embedding matrix and continuously updated using a graph neural network, is as follows:

[0045]

[0046] in, f represents the updated hidden state matrix; self This represents a function that defines how information is propagated to its own node. It is a multilayer perceptron. Represents the weight matrix; b5∈R e , represents the offset; f neighbor This represents a function that defines how information is propagated to neighboring nodes. This represents the temporal feature embedding matrix of knowledge point i.

[0047] In summary, this invention provides a dynamic graph neurocognitive diagnostic method that integrates multidimensional features. Compared with existing technologies, the technical solution conceived in this invention can achieve the following beneficial effects:

[0048] Firstly, by comprehensively utilizing big data technology, deep learning, and natural language processing technology, this approach performs semantically enhanced multidimensional interactive feature learning across temporal and spatial dimensions. It also leverages the advantages of graph neural networks and forgetting memory gate structures to dynamically update learners' knowledge mastery status, thereby acquiring information about learners' knowledge mastery and predicting their future performance. Furthermore, by using graph neural networks to analyze learners' performance on different knowledge points, it effectively integrates multidimensional interactive data. The graph neural network's ability to process graph structures demonstrates significant performance in tasks such as node classification, edge information propagation, and graph clustering. This greatly improves the accuracy of predicting knowledge mastery levels, enabling personalized learning suggestions and resource recommendations for learners with low learning effectiveness, as well as targeted teaching for those with poor future performance.

[0049] Secondly, by collecting and processing data from the entire process of learners' interaction with the learning platform, it not only considers the spatial relationships between knowledge points and the characteristics of a single knowledge point corresponding to multiple test tasks, but also integrates various behavioral characteristics such as the contextual semantic features, difficulty, discrimination, completion time, spatiotemporal features, and forgetting factors of the test tasks. This greatly enhances the importance of diagnosing learners' learning outcomes, improves the shortcomings of traditional evaluation methods in terms of content and tools, and addresses the deficiencies of current intelligent diagnosis. It strengthens the accompanying, automated, and intelligent nature of learners' knowledge mastery status, promotes timely feedback and application of mastery status results, and has significant educational practical significance and application prospects.

[0050] Thirdly, by using item response theory models or neurocognitive diagnostic models to predict the probability of learners correctly completing test tasks in the next moment, and thus predicting learners' future performance, we can provide teachers and learners with comprehensive and personalized learning diagnosis and tutoring services. This not only improves teaching efficiency but also promotes learners' personalized learning and growth. It has great potential and applications in the field of dynamic diagnosis and prediction that integrates learners' longitudinal multidimensional characteristics. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0052] Figure 1 This is a schematic diagram illustrating the principle of a dynamic graph neurocognitive diagnostic method that integrates multidimensional features, as provided by the present invention.

[0053] Figure 2 This is a schematic diagram of the semantic features of a test task for a dynamic graph neurocognitive diagnostic method that integrates multi-dimensional features, as provided by the present invention.

[0054] Figure 3 This is a schematic diagram of the graph neural network principle of a dynamic graph neural cognitive diagnostic method that integrates multi-dimensional features, provided by the present invention.

[0055] Figure 4 This is a schematic diagram of the memory gate structure of a dynamic graph neurocognitive diagnostic method that integrates multidimensional features, provided by the present invention. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings and embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0057] It should be noted that, in the description of the embodiments of the present invention, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a method, step, or system that includes a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to the method, step, or system. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the method, step, or system that includes said element.

[0058] This invention proposes a cognitive diagnostic model based on Graph Neural Networks (GNNs). It should be noted that GNNs have strong compatibility with multidimensional interactive data, and are particularly adept at processing graph-structured data.

[0059] like Figure 1 The diagram shown illustrates a method for dynamic graph neurocognitive diagnosis that integrates multidimensional features, as provided by this invention. The method includes:

[0060] S101: Collect multimodal data from online learning platforms and obtain multidimensional feature concatenation vectors based on the multimodal data.

[0061] It should be noted that multimodal data refers to valid data generated in real time on online learning platforms when learners complete test tasks (i.e., test questions or problems). This includes, for example, multidimensional data such as test questions, learners' accuracy rates, and answer times.

[0062] As an embodiment of the present invention, multimodal data is obtained by collecting data in real time from the database of an online learning platform using an API interface and performing data preprocessing.

[0063] Preferably, data preprocessing includes removing invalid data and data standardization.

[0064] As an embodiment of the present invention, obtaining a multidimensional feature concatenation vector based on multimodal data includes:

[0065] S201: Extracting the first and second types of data based on multimodal data.

[0066] Preferably, the first type of data includes: learner set, test task set, knowledge point set, and learner's record dataset of answering test tasks.

[0067] Preferably, the second type of data includes: test task difficulty data, learner scores, the time interval between learners learning the same knowledge point and the previous time, the time interval between learners learning the same knowledge point and the previous time, and the number of times learners repeatedly learn the knowledge point.

[0068] It should be noted that the second type of data was extracted from the multimodal data using methods such as the BERT model (Bidirectional Encoder Representations from Transformer, a pre-trained language representation model), the IRT model, and statistical analysis. This data reflects learners' performance in different aspects, providing rich information for subsequent diagnosis of knowledge mastery.

[0069] The BERT model is used to match learners' test task response text data with the key points of the answers, calculate the similarity between the two, and obtain the learners' scores. These scores include objective task scores and subjective task scores. Objective tasks include multiple-choice questions, true / false questions, and matching questions; subjective tasks include short-answer questions and application questions.

[0070] The IRT model was used to analyze learners' objective task scores and subjective task scores, and the distribution function between the test task and learners' scores was calculated to obtain test task difficulty data.

[0071] More specifically, a task-knowledge graph structure is constructed based on multimodal data to obtain the relationships between learners, test tasks, and knowledge points. That is, a task-knowledge graph structure and a trainable matrix Q are constructed based on the relationships between test tasks, knowledge points, competencies, and learners' task response results.

[0072] Using statistical analysis methods, based on the timestamps at t-1 (task start) and t (task end) and the number of times the task was repeated (including the number of times the learner consulted help and the number of times the learner watched videos), the time interval between the learner and the previous time point when learning the same knowledge point, the time interval between the learner and the previous time point when learning, and the number of times the learner repeated the knowledge point were calculated.

[0073] S202: Decompose the first type of data and integrate the test task information and knowledge point information from different decomposition semantic perspectives to obtain the test task semantic features;

[0074] Specifically, by analyzing the learner set, test task set, knowledge point set, and the record dataset of learners' answers to test tasks, the underlying relational information in the interaction is decomposed, and under the constraint of the loss function, test task information and knowledge point information from different decomposed semantic perspectives are fused to obtain the semantic features of the test task.

[0075] Preferred, such as Figure 2 As shown, the semantic features of the test task are obtained including:

[0076] S301: Using a d-dimensional trainable matrix and Initialize the test tasks and knowledge points separately.

[0077] Preferably, the one-hot representation vector x is... i The transpose of the matrix is ​​multiplied by the trainable matrix Q to initialize the test task, resulting in the initial test task q. i Knowledge point c k The initial embedding is the transpose of the k-th row of the trainable matrix C.

[0078] Specifically, the input layer is modeled as follows: Where, q i ,c k ∈R d ,x i ∈{0,1} M ,x k ∈{0,1} K .

[0079] S302: Decompose the interaction information between learners, test tasks, and knowledge points from the perspectives of co-occurrence and answer consistency to obtain the relationship matrix.

[0080] The co-occurrence perspective mainly decomposes the interaction between test tasks and knowledge points to find homogeneous neighbors of test tasks and knowledge points.

[0081] The test task transaction matrix can be represented as follows: Furthermore, each row of this matrix can be viewed as a test task transaction. Therefore, the relationship matrix of test tasks from a co-occurrence perspective is represented as follows: Among them, relationship strength Then it means q i and q j The relationship between them.

[0082] It should be noted that the strength of the relationship between any two test tasks It can be calculated as follows: in, Represents matrix T q The i-th and j-th columns, |||1 represents the 1-norm of the vector, and K represents the number of knowledge points.

[0083] In obtaining the homogeneous neighborhood q of the target test task i Subsequently, the normalized test task relationship matrix from the co-occurrence perspective is defined as follows: Specifically: Where, λ q Indicates used to control Hyperparameters.

[0084] The answer consistency perspective is mainly used to uncover relational information in the interaction between learners and test tasks and the interaction between learners and knowledge points.

[0085] As an embodiment of the present invention, the learner's interaction sequence (q) with the test task is extracted. i ,q j The strength of the relationships in the interaction of variables is calculated, thus obtaining the normalized test task relationship matrix from the perspective of answer consistency.

[0086] Preferably, the strength of the relationship in the variable interaction is calculated as follows: Wherein, p(a i =m,a j =n) represents the joint probability distribution, α i and α j represents the i-th and j-th columns of the answer, respectively; m and n represent the different values ​​of the student's answer, respectively.

[0087] The calculation method for the normalized test task relation matrix from the perspective of answer consistency is as follows:

[0088] S303: Under the constraint of the loss function, the relation matrix is ​​input into the test task embedding based on node-level attention aggregation and semantic-level attention aggregation to fuse test task information and knowledge point information from different decomposition semantic perspectives to obtain the semantic features of the test task.

[0089] That is, the above-mentioned relational information is jointly input into the test task embedding through node-level attention aggregation and semantic-level attention aggregation.

[0090] Node-level attention aggregation is used to learn the target test task q i The weights of node pairs (i,j). As an embodiment of the present invention, for the target test task node i, the weights of homogeneous node pairs (i,j) can be calculated as follows:

[0091]

[0092]

[0093]

[0094] in, This represents the weight of the homogeneous node pair (i,j) from the perspectives of answer consistency and co-occurrence. This represents the weight of the homogeneous node pair (i,j) obtained from the attention mechanism; γ represents the score of homogeneous node pair (i,j) calculated from priors; γ represents the hyperparameter used to control the balance of the two weights. This represents the node-level attention vector for a specific viewpoint ψ; This represents a measure of the correlation or importance between node i and node j from the perspective ψ; q j and q j’ Let represent the embedding vectors of node j and node j', respectively.

[0095] Preferably, the weights between node pairs are determined by considering node characteristics and statistical information, thereby effectively integrating information from homogeneous and heterogeneous neighbors in the aggregation layer to improve the quality and comprehensiveness of problem and concept embedding.

[0096] As an embodiment of the present invention, a specific angle ψ, test task q i Node-level embedding Aggregation is performed using the mapping features of homogeneous neighbors and the corresponding attention weights.

[0097] As a preferred embodiment, node-level embedding The aggregation formula is: Where σ() represents the sigmoid activation function; The representation of the learning task q is related to the target learning task. i Related neighborhood.

[0098] Preferably, node features are injected into node-level attention, aggregating two sets of node embeddings for specific test tasks. and knowledge point node embedding This effectively integrates information from specific problem nodes and concept nodes, thereby improving the embedding quality and comprehensiveness of problem nodes and concept nodes.

[0099] Semantic-level attention aggregation is used to compute the weights of the two semantic perspectives of the target test task node i, and learn a more comprehensive node embedding.

[0100] Specifically, the embedding of a specific semantic is first transformed through a nonlinear transformation, and then the average of all contributions of each neighbor of node i is taken.

[0101] As an embodiment of the present invention, the normalized importance of each semantic pair for node i for:

[0102]

[0103] in, This represents the feature vector of node i in a given context φ; ρ represents the feature vector of node i in another context φ'; T ρ represents the transpose of vector ρ; W represents a learnable weight matrix used to learn the linear transformation of the node representation. represents the query vector for node i; b represents the bias term, used to introduce the offset.

[0104] As an embodiment of the present invention, for the target test task node i, an updated test task embedding is obtained through fusion embedding.

[0105] Specifically, in, These represent the learnable weight coefficients from different perspectives, and the weighted query vectors for different context types. These represent the query vectors of node i from different perspectives; c represents the set of nodes that are directly related to node i in its context; k This indicates a knowledge point.

[0106] For a specific knowledge point, the same method is used to obtain the updated knowledge point embedding. I will not go into detail here.

[0107] To improve the quality of feature learning, this invention utilizes feature constraints and structural constraints.

[0108] In the feature constraint section, the features of the test task include the text features of the test task, the test task type, etc. Preferably, the test task q i The eigenvectors are represented as: Among them, W f Let b represent a weight matrix. f Let F represent the bias vector and F represent the number of features. q i eigenvectors.

[0109] Furthermore, preferably, for test task q i Define the set of related knowledge points as For example, Use The average representation of all knowledge point embeddings in q is taken as q i Embedded.

[0110] Specifically, it can be expressed as a formula:

[0111] As an embodiment of the present invention, a three-layer multilayer perceptron (MLP) is used to... Merged into the final feature And using nonlinear layers to Perform mapping to obtain in and b d These are trainable parameters.

[0112] Preferably, the loss function for feature constraints is:

[0113] In the structural constraints section, the explicit relationship between the problem and the concept is modeled by estimating the local proximity between the test task and the knowledge point vertices in the embedding space. The test task q... i And knowledge point c k The modeling of the local proximity between them is as follows:

[0114] To ensure that the local proximity between test tasks and knowledge points in the embedding space remains consistent with their relationship in the knowledge graph, as a preferred embodiment of the present invention, the loss function formula for structural constraints is:

[0115] Preferably, the loss function (joint optimization function) includes a combination of the loss function of the feature constraint part and the loss function of the structural constraint part.

[0116] Specifically, the loss function is minL = min(L f +L s It should be noted that the optimal semantic feature q for the test task can be obtained when the function optimization converges. text .

[0117] S203: Based on the first type of data, the second type of data, and the semantic features of the test task, a multi-dimensional feature concatenation vector is obtained.

[0118] As one embodiment of the present invention, such as Figure 3 As shown, methods for obtaining multi-dimensional feature concatenation vectors based on the first type of data, the second type of data, and the semantic features of the test task include:

[0119] The test task difficulty matrix q is obtained based on the second type of data. difficulty Test task representation matrix a, showing learners' performance on the test task. t ; where, the test task representation matrix a t Represented as a decimal between 0 and 1.

[0120] The semantic features q of the test task text Combined with the test task difficulty matrix q difficulty To obtain the task feature vector

[0121] The test task is decomposed into multiple knowledge point vectors according to the relationship between the test task and the knowledge points; that is, according to the relationship between the test task and the knowledge points in the task-knowledge graph structure, a test task is decomposed into several knowledge point vectors (i,j,k).

[0122] Task feature vector Test task representation matrix a t The vectors of multiple knowledge points (i,j,k) are concatenated to obtain the multidimensional feature concatenation vector of each knowledge point at the current time.

[0123] The preferred formula for obtaining the multidimensional feature concatenation vector is:

[0124]

[0125]

[0126]

[0127] Where ω1 and b1 are the learning parameters, q rsThis is the representation of the s-th knowledge point in test task r after feature enhancement. The dimension of vector O = (O, O, ..., O) is related to q. rs Consistent.

[0128] S102: Based on multi-dimensional feature concatenation vectors, and using graph neural networks, the embedding matrix of the current knowledge point, the hidden state matrix of the current knowledge point, the embedding matrix of adjacent knowledge points, and the hidden state matrix of adjacent knowledge points are aggregated to obtain an aggregated matrix.

[0129] As one embodiment of the present invention, the aggregation includes:

[0130] The hidden state matrix and the first embedding matrix of the current knowledge point are aggregated to obtain the first matrix of the current knowledge point;

[0131] The hidden state matrix and the second embedding matrix of adjacent knowledge points are aggregated to obtain the second matrix of adjacent knowledge points;

[0132] The aggregation matrix is ​​calculated based on the first and second matrices.

[0133] It should be noted that each knowledge point is represented by a string of length d. h The vector represents the knowledge state; N knowledge points will have N vectors, forming an N×d vector. h The matrix represents the knowledge state H of each learner; the values ​​of the elements in the matrix range from 0 to 1, with the closer to 1 indicating a higher level of mastery.

[0134] Preferred, aggregation matrix The calculation method is as follows:

[0135] in, The hidden state matrix representing knowledge point k; X t W represents a multidimensional feature concatenation vector. x ∈R N*e W represents the first embedding matrix, which shows the index of knowledge points and the completion status of each test task; c ∈R N*e W represents the second embedding matrix of knowledge points. c (V) represents the second embedding matrix w c In the v-th row, e is the embedding value.

[0136] S103: The embedding matrix of knowledge points is processed by forgetting vectors, and the aggregation matrix is ​​input into the memory gate structure model to capture the temporal feature embedding matrix of knowledge points.

[0137] Preferably, when knowledge point k transmits a message to knowledge point i, the forgetting vector is used. Forgetting is performed on the knowledge point embedding matrix of knowledge point i, and the aggregation matrix is ​​then processed. Input into the memory gate structure model.

[0138] As one embodiment of the present invention, such as Figure 4 As shown, the aggregation matrix is ​​input into the memory gate structure model to capture the temporal feature embedding matrix of knowledge points, including:

[0139]

[0140]

[0141]

[0142]

[0143]

[0144] Among them, f neighbor This represents a function that defines how information is propagated to neighboring nodes. and All are multilayer perceptrons, [] represents the stitching operation; A i,k A represents the relationship between knowledge point i and knowledge point k; k,i This represents the relationship between knowledge point k and knowledge point i; W1 represents the forgetting vector of knowledge point i; e*4e W2∈R e*4e W3∈R e*5e W4∈R e*2e Both represent weight matrices; b1∈R e b2∈R e b3∈R e b4∈R e , both represent offset; The embedding matrix representing the temporal features of knowledge point i; The aggregation matrix representing knowledge point i at time t; The aggregation matrix represents knowledge point k; The aggregation matrix represents the knowledge point i at time t-1.

[0145] It should be noted that, Among them, P t (i)=[CR t (i),FA t [i] represents the learner's mastery of existing knowledge at time t, CR t (i) represents the number of times the learner viewed help at time t, FA t (i) is the number of times the learner viewed the video; RT t(i) represents the time interval between the learner's last learning of the same knowledge point at time t; LT t (i) represents the number of times the learner repeats the knowledge point at time t; ST t (i) represents the time interval between the learner's last learning session and time t; OM t-1 (i) represents the learner's initial mastery of the knowledge points at time t-1.

[0146] S104: Based on the temporal feature embedding matrix and using a graph neural network to continuously update the hidden state matrix of knowledge points, the learner's mastery of knowledge points after the next time step is obtained.

[0147] As an embodiment of the present invention, the hidden state matrix of knowledge points, which is based on the temporal feature embedding matrix and continuously updated using a graph neural network, is as follows:

[0148]

[0149] in, f represents the updated hidden state matrix; self This represents a function that defines how information is propagated to its own node. It is a multilayer perceptron. Represents the weight matrix; b5∈R e , represents the offset; f neighbor This represents a function that defines how information is propagated to neighboring nodes. This represents the temporal feature embedding matrix of knowledge point i.

[0150] As an embodiment of the present invention, the level of knowledge mastery updated at the next time step t+1. for: Among them, W hθ Weight matrix (learnable parameters); b represents the hidden state matrix of knowledge point i after the update; hθ This represents the offset (a learnable parameter).

[0151] As an embodiment of the present invention, the method further includes:

[0152] S105: Based on the level of knowledge mastery, the relationship between knowledge points, the difficulty data of the test task, the discrimination data of the test task, and the semantic features of the test task, use the item response theory model or the neurocognitive diagnostic model to predict the probability that the learner will correctly complete the test task in the next moment.

[0153] As an example, the levels of knowledge mastery, relationships between knowledge points, test task difficulty data, test task discrimination data, and test task semantic features are input into the Project Response Theory model, specifically including: in, and These represent the test task discrimination and test task difficulty at time t+1, respectively.

[0154] As another embodiment, the knowledge point mastery level, knowledge point relationships, test task difficulty data, test task discrimination data, and test task semantic features are input into the neurocognitive diagnostic model, specifically including:

[0155]

[0156]

[0157]

[0158]

[0159] in, It is the m-th row of the Q matrix, which is also the input variable, and It is the activation function tanh. This represents the degree to which student n has mastered the knowledge points, and is estimated based on the student's task response vector.

[0160] It should be noted that during the training phase, the model parameters are learned by minimizing the standard cross-entropy loss between the predicted probability of a student correctly completing the task at time t+1 and the actual probability of a student correctly completing the task at time t+1.

[0161] Preferably, the standard cross-entropy loss is: in, It represents the students' true probability of mastering knowledge, as determined by the test task q. t The representation vector α of the correct completion status t It is calculated.

[0162] In summary, through the design and implementation of its components, this invention fully utilizes multi-dimensional interactive feature extraction and fusion technology, combined with a diagnostic method based on knowledge point graph structures, to provide teachers and learners with comprehensive and personalized learning diagnosis and tutoring services. This technical solution not only improves teaching efficiency but also promotes personalized learning and growth for learners.

[0163] It should be noted that, for the sake of simplicity, the foregoing embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0164] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0165] In the several embodiments provided in this application, it should be understood that the disclosed methods or systems can be implemented in other ways. For example, the embodiments described above are merely illustrative. For instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0166] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0167] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0168] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage device. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application.

[0169] Those skilled in the art will understand that all or part of the circuits in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, which may include: a flash drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, etc.

[0170] The foregoing description is merely an exemplary embodiment of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure. Those skilled in the art will readily conceive of embodiments of this disclosure upon considering the specification and practicing the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not described herein. The specification and embodiments are to be considered exemplary only, and the scope and spirit of this disclosure are defined by the claims.

[0171] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0172] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A dynamic graph neurocognitive diagnostic method integrating multidimensional features, characterized in that, The method includes: Collect multimodal data from an online learning platform, and obtain a multidimensional feature concatenation vector based on the multimodal data; Based on the multidimensional feature concatenation vector and using a graph neural network, an aggregation matrix is ​​obtained by combining the embedding matrix of the current knowledge point, the hidden state matrix of the current knowledge point, the embedding matrix of adjacent knowledge points, and the hidden state matrix of adjacent knowledge points. The embedding matrix of knowledge points is processed by forgetting vectors, and the aggregation matrix is ​​input into the memory gate structure model to capture the temporal feature embedding matrix of knowledge points. Based on the temporal feature embedding matrix and by continuously updating the hidden state matrix of knowledge points using a graph neural network, the learner's mastery of knowledge points after the next time step is obtained. The multidimensional feature concatenation vector obtained based on the multimodal data includes: Based on the multimodal data, extract the first type of data and the second type of data; The first type of data is decomposed, and test task information and knowledge point information from different decomposition semantic perspectives are fused to obtain test task semantic features. Based on the semantic features of the first type of data, the second type of data, and the test task, a multi-dimensional feature concatenation vector is obtained; The acquisition of the semantic features of the test task includes: Using a d-dimensional trainable matrix and Initialize the test tasks and knowledge points separately; The interaction information between learners, test tasks, and knowledge points is decomposed from the perspectives of co-occurrence and answer consistency, respectively, to obtain a relationship matrix; Under the constraint of the loss function, the relation matrix is ​​input into the test task embedding based on node-level attention aggregation and semantic-level attention aggregation to fuse test task information and knowledge point information from different decomposition semantic perspectives to obtain the semantic features of the test task.

2. The dynamic graph neurocognitive diagnostic method integrating multidimensional features as described in claim 1, characterized in that, The first type of data includes: learner sets, test task sets, knowledge point sets, and learner record datasets of test task responses; The second type of data includes: test task difficulty data, learner scores, the time interval between learners learning the same knowledge point and the previous time, the time interval between learners learning the same knowledge point and the number of times learners have repeated the knowledge point.

3. The dynamic graph neurocognitive diagnostic method integrating multidimensional features as described in claim 2, characterized in that, The method further includes: based on the learner's mastery of the knowledge points, the relationship between the knowledge points, the difficulty data of the test task, the discrimination data of the test task, and the semantic features of the test task, using the item response theory model or the neurocognitive diagnostic model to predict the probability that the learner will correctly complete the test task in the next moment.

4. The dynamic graph neurocognitive diagnostic method integrating multidimensional features as described in claim 3, characterized in that, Methods for obtaining multi-dimensional feature concatenation vectors based on the first type of data, the second type of data, and the semantic features of the test task include: Based on the second type of data, a test task difficulty matrix and a test task representation matrix of learners' test task completion status are obtained; The semantic features of the test task and the difficulty matrix of the test task are combined to obtain the task feature vector; The test task is decomposed into multiple knowledge point vectors based on the relationship between the test task and the knowledge points. The task feature vector, the test task representation matrix, and multiple knowledge point vectors are concatenated to obtain a multi-dimensional feature concatenation vector for each knowledge point at the current time.

5. The dynamic graph neurocognitive diagnostic method integrating multidimensional features as described in claim 4, characterized in that, The aggregation includes: The hidden state matrix and the first embedding matrix of the current knowledge point are aggregated to obtain the first matrix of the current knowledge point; The hidden state matrix and the second embedding matrix of adjacent knowledge points are aggregated to obtain the second matrix of adjacent knowledge points; The aggregation matrix is ​​calculated based on the first matrix and the second matrix; the aggregation matrix The calculation method is as follows: ; wherein, the The hidden state matrix representing knowledge point k; This represents the concatenated vector of the multidimensional features; , representing the index of knowledge points and the first embedding matrix of the completion status of each test task; The second embedding matrix represents the knowledge points. Represents the second embedding matrix The OK, It is an embedded value.

6. The dynamic graph neurocognitive diagnostic method integrating multidimensional features as described in claim 5, characterized in that, The aggregation matrix is ​​input into the memory gate structure model to capture the temporal feature embedding matrix of knowledge points, including: ； ； ； ； ； in, This represents a function that defines how information is propagated to neighboring nodes. , and All are multilayer perceptrons, and [ ] represents the stitching operation; This represents the relationship between knowledge point i and knowledge point k; This represents the relationship between knowledge point k and knowledge point i; Represents the forgetting vector of knowledge point i; , , , , where , represent the weight matrix; , , , , both represent offset; The embedding matrix representing the temporal features of knowledge point i; The aggregation matrix representing knowledge point i at time t; The aggregation matrix representing knowledge point k; The aggregation matrix represents the knowledge point i at time t-1.

7. The dynamic graph neurocognitive diagnostic method integrating multidimensional features as described in claim 6, characterized in that, The hidden state matrix of knowledge points, based on the aforementioned temporal feature embedding matrix and continuously updated using a graph neural network, is as follows: ； in, This represents the updated hidden state matrix; This represents a function that defines how information is propagated to the node itself. , It is a multilayer perceptron. ; , represents the weight matrix; , indicating the offset; This represents a function that defines how information is propagated to neighboring nodes. This represents the temporal feature embedding matrix of knowledge point i.

Images