A dynamic knowledge cognitive level mining method, system, device and terminal

Abstract

The application belongs to the technical field of personalized learning, and discloses a dynamic knowledge cognitive level mining method, system, device and terminal, which collects learning behavior characteristics and learning resource characteristics of learners, generates a learner-question-knowledge interactive heterogeneous graph, extracts subgraphs from the learner-question-knowledge interactive heterogeneous graph, introduces a graph convolutional neural network, performs representation learning on question nodes and knowledge point nodes based on different graph structures, constructs a knowledge state updating layer based on a gate, predicts the mastery level of the learner on questions and knowledge points at the next moment, constructs a dynamic cognitive tracking layer based on a linear neural network, dynamically tracks the cognitive level of the learner according to the updated knowledge point mastery level vector of the learner, fuses the mastery level of the learner on the questions and the mastery level of the learner on the knowledge points examined by the questions, predicts the answering response of the learner at the next moment, and performs training, and applies the trained parameters to predict the learning knowledge state of the learner.

Description

Technical Field

[0001] This invention belongs to the field of personalized learning technology, and in particular relates to a method, system, device and terminal for dynamic knowledge cognition level mining. Background Technology

[0002] With the continuous development of technology, online education is becoming an increasingly popular learning method. While online education brings convenience to students, its learning effectiveness still faces many challenges. One of these challenges is how to achieve personalized teaching to improve learning outcomes. To fully utilize the rich learning resources provided by online learning platforms, and to ensure that each student receives personalized learning guidance and targeted teaching based on their actual situation, thereby improving the effectiveness of online learning, intelligent tutoring systems based on artificial intelligence technology are booming. Intelligent tutoring systems need to mine data from students' learning history to accurately assess their current knowledge level. Learning performance prediction tasks can track students' dynamic knowledge levels over a period of time based on their historical learning records, determine their potential knowledge status, and predict their future knowledge level and answer performance. Furthermore, learning performance prediction can also identify common problems students encounter during the learning process, helping teachers discover and solve students' learning difficulties, thereby better helping students improve their academic performance. Based on the results of learning performance prediction, teachers can develop personalized teaching plans for students, providing targeted reinforcement for students' weak knowledge areas. Therefore, learning performance prediction is a fundamental and key module of personalized teaching, which is of great significance to the development of intelligent tutoring systems and the promotion of personalized learning. It is also one of the research hotspots in the field of educational data mining.

[0003] Current methods for cognitive level mining can be broadly categorized into two types: those based on traditional machine learning and those based on deep learning. Traditional cognitive level mining methods include Item Response Theory (OPT), Bayesian models, and factor analysis models. While these methods offer good interpretability, they rely on theoretical assumptions and require manually constructed input features. Feature construction is often one-sided and limited, failing to fully uncover hidden information within the data, resulting in generally poor predictive performance. In contrast, mainstream deep learning-based learning performance prediction models utilize historical response records and employ different neural network structures to improve experimental results. For example, they use recurrent neural networks (RNNs) such as LSTM or GRU, or convolutional neural networks, to process sequential data and model students' knowledge states. These models further enhance their performance by incorporating attention mechanisms and introducing long short-term memory (LSTM) models. These methods often use test items or the knowledge points contained within them as features, resulting in relatively singular features. While they offer good predictive performance, their results lack interpretability.

[0004] Based on the above analysis, the problems and shortcomings of the existing technology are as follows:

[0005] (1) Current cognitive level mining methods, under the premise of transforming students' answer records into an interaction sequence with knowledge points, ignore the relationship between knowledge points in the test questions and do not consider the connection between test questions and knowledge points. As a result, some high-dimensional information in the answer records cannot be effectively utilized by the model, resulting in poor interpretability.

[0006] (2) Current cognitive level mining methods can only output hidden knowledge states that lack interpretability, usually a decimal between 0 and 1, which cannot be directly applied to educational scenarios. In real educational scenarios, students' knowledge mastery status is not limited to "mastery" and "not mastery", but has different cognitive levels.

[0007] Deep knowledge tracing models are knowledge tracking models that use recurrent neural networks (RNNs). This model predicts a student's mastery of future knowledge points by tracking their learning process. Deep knowledge tracing models input students' answer records as sequential data into the RNN and predict the student's mastery of the next knowledge point at each time step.

[0008] Technical problems with existing technologies:

[0009] 1. Static prediction: Existing deep knowledge tracking models are mainly static. They usually predict students' knowledge mastery status at a fixed point in time, rather than dynamically tracking and predicting changes in students' cognitive levels and knowledge mastery.

[0010] 2. Neglecting the relationships between knowledge points: The DKT model primarily focuses on the learning process over time, but it overlooks the relationships between knowledge points. In actual teaching, there are rich dependencies and connections between knowledge points, which are not fully utilized in deep knowledge tracing models.

[0011] 3. Complexity and Interpretability: As a black-box model, the DKT model's internal workings are complex and difficult to interpret. This limits its application in education to some extent, as teachers and students may need to understand the model's predictions in order to make appropriate teaching or learning decisions.

[0012] 4. Dependence on data: The DKT model is highly dependent on a large amount of student answer data. However, in reality, there may not be enough answer data for each student, which may affect the model's prediction accuracy. Summary of the Invention

[0013] To address the problems existing in the prior art, this invention provides a dynamic knowledge cognition hierarchy mining method.

[0014] This invention is implemented as follows: a dynamic knowledge cognition hierarchy mining method, which includes:

[0015] S1, collect learners’ learning behavior characteristics and learning resource characteristics, and generate a heterogeneous graph based on learner-test question-knowledge interaction;

[0016] S2 extracts subgraphs from the learner-question-knowledge interaction heterogeneous graph, introduces a graph convolutional neural network, and performs representation learning on question nodes and knowledge node based on different graph structures;

[0017] S3, construct a knowledge state update layer based on gating to predict the learner’s mastery level of test questions and knowledge points in the next moment;

[0018] S4, construct a dynamic cognitive tracking layer based on a linear neural network, and dynamically track the learner's cognitive level according to the learner's updated knowledge level mastery horizontal vector;

[0019] S5 integrates the learner's mastery level of the test questions and the mastery level of the knowledge points tested by the test questions to predict the learner's response at the next moment and constructs a loss function for training;

[0020] S6 uses the trained parameters to predict the learner's knowledge status.

[0021] Furthermore, learning behavior characteristics and learning resource characteristics are collected from learners' response sequences, with learning behavior characteristics being response characteristics that include temporal information:

[0022] s={(e1,k1,r1),(e2,k2,r2),…,(e t ,k t ,r t )},s∈S

[0023] Triplet (e) t ,k t ,r t This indicates that at time t, student s answered question e. t This question tests knowledge point k. t r t This refers to the student's answers; if the student answers incorrectly, r t If the student answers correctly, r is 0. t For e t For knowledge point k t The assessment levels are (1: Know, 2: Master, 3: Apply);

[0024] Learning resource characteristics include the knowledge point characteristics of the learning resources, specifically the cognitive level characteristics of the knowledge points tested in the test questions of the learning resources and the relationship characteristics between the knowledge points in the learning resources:

[0025] e = {(k1,l1),(k2,l2),…,(k M ,l M )},e∈E

[0026] Among them, the binary pair (k M ,l M Question e represents the knowledge point k. M The level of investigation is l M M represents the total number of knowledge points tested in this question; the set of knowledge points included in this course is |K|, k∈K; the set of questions included in this course is |E|, e∈E; l∈{1,2,3} represents the level of knowledge point k tested in this question, where 1 represents knowing, 2 represents mastering, and 3 represents applying.

[0027] The learner-question-knowledge point heterogeneous graph is generated based on learners' answer sequences and learning resource characteristics. Specifically, in the learner answer sequence, learners answer questions in a certain order, and each question tests different knowledge points. This process involves a large number of heterogeneous relationships between students, questions, and knowledge points. The learner-question-knowledge point heterogeneous graph SQK = {V, E}, where V represents the node set and E represents the edge set. The node set contains three types of nodes: student nodes n s Question node n q and knowledge point node n k The edge set contains two types of edges: answer edges e. a and the inspection of the border e t .

[0028] Furthermore, the specific process of extracting subgraphs from the learner-question-knowledge interaction heterogeneous graph includes:

[0029] Extract a learner response behavior subgraph (SQ graph) to record learners' answers to test questions; extract a knowledge point testing subgraph (QK graph) to represent the heterogeneous relationship between test questions and knowledge points; extract a relationship subgraph between knowledge points (KK graph) to focus on the connections between knowledge points.

[0030] The learners' responses to the test questions are represented by an SQ graph, where SQ∈R. s_num×q_num SQ ij The number of times learner i answers question j, where s_num represents the number of learners and q_num represents the number of questions;

[0031] The relationship between test questions and knowledge points is represented by a QK graph, where QK∈R.q_num×k_num Where k_num represents the number of knowledge points, if question j tests knowledge point k, then the corresponding SQ jk The positional weight increases. QK(j) represents the j-th row of the matrix, i.e., the knowledge points associated with question j; (QK) T (k) represents the k-th column of the matrix, which contains the test questions associated with knowledge point k;

[0032] The relationships between knowledge points are represented as KK diagrams, and the KK relationships are extracted based on the statistical transition probability matrix. Among them, KK ab n represents the degree of correlation between knowledge points a and b. a,b This represents the number of times knowledge point b is tested after knowledge point a;

[0033] Furthermore, a graph convolutional neural network is introduced. The specific process of learning representations for question nodes and knowledge point nodes based on different graph structures includes:

[0034] Before information propagation in the graph convolutional neural network, the node embedding representations and hidden knowledge states at time t are aggregated:

[0035]

[0036]

[0037] Here, α represents a question node or a knowledge point node, and the method requires two sets of parameters to learn the embedded representation of questions and knowledge points. x represents the hidden state of the question or knowledge point at time t. t ∈{0,1,2,3} 4N E represents the learner's response at time t (0: wrong answer, 1: correct answer and knowledge, 2: correct answer and mastery; 3: correct answer and application); x_α ∈R 4N×e E represents the embedding matrix between test questions or knowledge points and answers. α ∈R N×e It is an embedding matrix of test questions or knowledge points, E α (k) represents E α The k-th row represents the embedding of question k or knowledge point k, where e is the dimension of the embedding. x_α E α All are trainable matrices;

[0038] Based on the graph structure, update the information of the test questions and knowledge points themselves, as well as their neighbors. The method for updating the information of a node itself is as follows:

[0039]

[0040]

[0041] In the formula, and These are the trainable weights and biases used for node aggregation, respectively. The activation function is ReLU.

[0042] Before updating the information of neighboring nodes, it is necessary to first aggregate the updated information of the current node, and then update the information of neighboring nodes:

[0043]

[0044]

[0045] in, This is the updated embedding representation for the current node. and These are the trainable weights and biases used for neighbor node aggregation, respectively. The activation function is ReLU.

[0046] Because the graph structures on which the embedding representation of the question node and the knowledge point node are updated are different, the information update formulas for their neighbor nodes are slightly different. The method for updating the information of the neighbor nodes of the current question node is as follows:

[0047]

[0048] In the formula, SQ(stu_id) represents the learner's answer status in the current row of the sequence number, and c t For the knowledge points being tested at the current moment, QK(c t To update the neighbor node information of the current knowledge point node, which is related to the questions being tested, the method is as follows:

[0049]

[0050] In the formula, q t QK(q) represents the current test questions. t KK(c) represents the knowledge points related to the current test question. t ) and (KK) T These are the c-th graphs of the KK diagram. t line and c t The column represents the knowledge points related to the currently tested knowledge point.

[0051] Furthermore, the specific process of constructing a gating-based knowledge state update layer to predict the learner's mastery level of test questions and knowledge points at the next moment includes:

[0052] The updated test questions and knowledge points are embedded and fed into a gated neural network to model the learner's mastery of the knowledge points and test questions at time t+1.

[0053]

[0054]

[0055] In the formula, G ea The specific calculation process for the erase-addition gate is as follows:

[0056]

[0057]

[0058]

[0059] In the formula, tmp_erase and tmp_add represent the erase and addition amounts of the gate structure, respectively, and W erase W add These are the weight matrices for the erase gate and the addition gate, respectively; b erase b add These are the bias terms corresponding to the erase gate and the addition gate, respectively, and σ is the activation function sigmoid;

[0060] In the formula, G gru The specific calculation process for the gating gate is as follows:

[0061]

[0062]

[0063]

[0064]

[0065] In the formula, To hide the knowledge vector for learners after the erase-addition gate, This is the hidden state from the previous moment. To reset the door, To update the door, Let σ represent the candidate hidden state, σ be the activation function sigmoid, and * represent the Hadamard product.

[0066] Furthermore, the specific process of constructing a dynamic cognitive tracking layer based on a linear neural network to dynamically track the learner's cognitive levels includes:

[0067] Cognitive level prediction is performed using the hidden knowledge state obtained at time t+1 from the knowledge state update layer.

[0068]

[0069] In the formula, W l It is a cognitive state weight matrix for all knowledge point nodes, b k σ is the offset of knowledge point node k, and σ is the softmax function, which means predicting whether the student belongs to one of the four cognitive levels at this moment.

[0070] C = Argmax(l t )

[0071] In the formula l t The cognitive hierarchy matrix of students for all concepts at the next moment is obtained by using the Argmax function to take the index of the item with the highest predictability, which is the cognitive hierarchy C of students for all knowledge points.

[0072] The learner's level of mastery of the test questions and their level of mastery of the knowledge points tested by the test questions are represented as follows:

[0073]

[0074]

[0075]

[0076] In the formula, and It is a prediction weight matrix for answering knowledge point nodes and question nodes, b k and b q σ is the offset between knowledge point k and question q, and σ is the sigmoid function. and These represent the learner's level of mastery of the test questions and knowledge points at time t;

[0077] The model uses a linear layer to fit students' understanding of the questions and knowledge points tested at time t, thus obtaining the students' answer performance y. t The model is trained by minimizing the negative log-likelihood (NLL) loss to obtain the best model performance.

[0078] The prediction of the learner's response at the next moment is expressed as follows:

[0079] p t =y t ·q t+1

[0080] In the formula, q t+1 For the questions learners need to answer in the next moment, y tLet p be the student's response vector to the test question at this moment. t That is, the probability that the learner can answer the question correctly;

[0081] Training is performed by minimizing the negative log-likelihood loss, based on the predicted learner responses and the actual learner responses.

[0082] L = logSoftmax(p t ,y true )

[0083] Finally, the Adam optimizer is used to update the network parameters of the above model.

[0084] Another objective of this invention is to provide a dynamic knowledge cognition hierarchy mining system, which includes:

[0085] The node embedding representation learning module is used to generate corresponding heterogeneous graphs based on learner-question-knowledge interaction according to learner learning behavior characteristics and learning resource characteristics; it uses the learner-question-knowledge interaction heterogeneous graph to generate subgraphs, and introduces graph convolutional neural networks to perform embedding representation learning on question nodes and knowledge point nodes respectively based on different graph structures.

[0086] The knowledge status update module is used to construct a gate structure to perform temporal prediction on the updated question nodes and knowledge point nodes, and calculate the learner's mastery status of all questions and knowledge points at the next moment.

[0087] The dynamic cognitive tracking module is used to dynamically track the learner's cognitive level of all knowledge points by utilizing the learner's mastery state vector of all knowledge points based on a linear neural network.

[0088] The response prediction module is used to predict the learner's response at the next moment based on the learner's learning interaction sequence, the characteristics of the question the learner will answer at the next moment, and the characteristics of the learner's response to the next question and the knowledge points it tests.

[0089] Another object of the present invention is to provide a computer device, the computer device including a memory and a processor, the memory storing a computer program, and when the computer program is executed by the processor, causing the processor to perform the steps of the dynamic knowledge cognition hierarchy mining method described above.

[0090] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the dynamic knowledge cognition hierarchy mining method.

[0091] Another objective of this invention is to provide an information data processing terminal for implementing the aforementioned dynamic knowledge cognition hierarchy mining system.

[0092] Based on the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by this invention are as follows:

[0093] First, existing deep learning-based cognitive hierarchy mining methods neglect the relationships between knowledge points within test questions and fail to consider the connections between test questions and knowledge points. This results in some high-dimensional information in the answer records being unable to be effectively utilized by the model, leading to poor interpretability. The challenge in solving this problem lies in how to effectively extract and model the heterogeneous relationships between test questions and knowledge points, and between knowledge points themselves, within learning resources. This invention fully utilizes learners' answer sequences and learning resources, constructing a SQK heterogeneous graph from students' answer interaction sequences, and deconstructing this graph into SQ, QK, and KK graphs. These three graphs encompass the relationships between learners and test questions in learner answer interactions, the relationships between test questions and knowledge points in the course structure, and the connections between knowledge points, better modeling students' answer interaction behavior for more accurate subsequent predictions by the model. Experimental results show that it outperforms other deep learning-based cognitive hierarchy mining methods in terms of AUC, ACC, and RMSE.

[0094] Existing deep learning-based cognitive level mining methods can only output hidden knowledge states that lack interpretability, typically a decimal between 0 and 1, making them unsuitable for direct application in educational scenarios. In real-world educational settings, students' knowledge mastery states are not limited to "mastery" and "lack of mastery," but rather involve different cognitive levels. The challenge lies in extending the model's output from a purely non-zero or single-valued state to multi-class cognitive levels. This invention innovatively proposes a cognitive tracking problem and designs a dynamic cognitive tracking layer to dynamically acquire the student's current cognitive level.

[0095] Secondly, the dynamic knowledge cognition hierarchy mining method based on heterogeneous graph representation of learning interaction proposed in this invention represents learner interaction behavior and learning resource features as a learner-question-knowledge interaction heterogeneous graph, extracts subgraphs from it, introduces them into graph convolutional neural networks, and applies them to the embedding representation learning of different nodes. The updated node representations are used to model the student's mastery level of the questions and knowledge points in the next time step using a gate-based neural network, predicting the learner's response and dynamically tracking the learner's cognitive level. Compared with traditional methods, representing learner interaction behavior and learning resources as heterogeneous graphs and introducing cognitive tracking tasks has greatly improved the accuracy of prediction and has better interpretability, thus improving the shortcomings of traditional methods in predicting learning performance.

[0096] Third, as supplementary evidence of the inventive step of the claims of this invention, it is also reflected in the following important aspects:

[0097] (1) The expected benefits and commercial value of the technical solution of this invention after transformation are as follows:

[0098] The dynamic knowledge cognition hierarchy mining method based on heterogeneous graph representation of learning interaction proposed in this invention represents learner interaction behavior as a heterogeneous graph and introduces a cognition tracking module. It realizes the dynamic output of learner's current cognition level of all knowledge points based on learner's historical answer data, making the model more practically educational and applicable to real educational scenarios. It provides reliable support for personalized teaching and has great commercial value.

[0099] (2) The technical solution of this invention fills a technical gap in the industry both domestically and internationally:

[0100] Existing deep learning-based cognitive level mining methods neglect the learner's multiple cognitive states regarding knowledge points. This invention constructs a cognitive tracking layer to achieve richer predictive analysis of the learner's knowledge state and mastery.

[0101] (3) The technical solution of the present invention solves a technical problem that people have long wanted to solve but have never been able to solve successfully:

[0102] Existing deep learning-based cognitive hierarchy mining methods generally neglect the relationships between questions and knowledge points in learning resources, as well as between different knowledge points, and ignore the modeling of students' answering behavior. This invention fully models the relationships between learners' answering interactions and the learning resources, and updates learners' states regarding questions and knowledge points based on a graph structure. This enables the prediction and analysis of learners' knowledge states and mastery levels, improves the model's interpretability, and helps learners and educators conduct more targeted learning and teaching.

[0103] (4) The technical solution of the present invention overcomes technical bias:

[0104] Existing deep learning-based cognitive level mining methods generally ignore the premise that learners have multiple cognitive levels of knowledge points, and only output the hidden knowledge state with poor interpretability. This invention makes full use of the features in learning resources, combined with the different levels of examination of different knowledge points in the test questions, to dynamically track and evaluate the learner's cognitive level of knowledge points, making the model output more interpretable.

[0105] Fourth, the acquisition of the main steps of this invention represents a significant technological advancement.

[0106] (a) Collect learners' learning behavior characteristics and learning resource characteristics to generate a heterogeneous graph of learner-question-knowledge interaction:

[0107] The significant technological advancement of this step lies in its ability to generate a heterogeneous graph by collecting learners' learning behavior characteristics and learning resource characteristics, thus providing a data foundation for subsequent analysis and data mining. This heterogeneous graph can better reflect the interaction relationships between learners, test questions, and knowledge points, providing rich information for subsequent analysis.

[0108] (b) Extract subgraphs from the learner-question-knowledge interaction heterogeneous graph, introduce graph convolutional neural networks, and learn representations for question nodes and knowledge node based on different graph structures:

[0109] The significant technological advancement in this step lies in its introduction of graph convolutional neural networks, which are capable of extracting useful information from complex graph structures and learning representations for question nodes and knowledge point nodes. This method is able to better capture and represent the complex relationships between questions and knowledge points.

[0110] (c) Construct a gating-based knowledge state update layer to predict the learner's mastery level of test questions and knowledge points in the next time step:

[0111] The significant technological advancement of this step lies in its construction of a gated knowledge state update layer, capable of dynamically predicting learners' mastery levels of test questions and knowledge points. This method can more accurately reflect learners' actual knowledge mastery, providing crucial information for subsequent teaching strategies.

[0112] (d) Construct a dynamic cognitive tracking layer based on a linear neural network to dynamically track the learner's cognitive level according to the learner's updated knowledge mastery level vector:

[0113] The significant technological advancement of this step lies in its introduction of linear neural networks, which provide a deeper analysis of learner performance by dynamically tracking learners' cognitive levels. This approach allows for a better understanding of learners' learning processes and provides more effective instructional suggestions.

[0114] (e) By integrating learners' mastery of the test questions and their mastery of the knowledge points tested, the learners' responses at the next moment are predicted, and a loss function is constructed for training:

[0115] The significant technological advancement of this step lies in its ability to predict learners' responses at the next moment by integrating their level of mastery of the test questions and their mastery of the knowledge points tested in the questions, thereby achieving accurate prediction of learners' learning progress and knowledge mastery.

[0116] (f) Predicting learners' knowledge status using trained parameters:

[0117] The significant technological advancement of this step lies in its ability to use trained models to predict learners' knowledge status, thereby providing educators with accurate and personalized educational recommendations and improving the personalization and precision of educational outcomes. Attached Figure Description

[0118] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0119] Figure 1 This is a flowchart of the dynamic knowledge cognition hierarchy mining method provided in the embodiments of the present invention;

[0120] Figure 2 This is a schematic diagram of the dynamic knowledge cognition hierarchy mining method provided in the embodiments of the present invention;

[0121] Figure 3 This is a structural diagram of the dynamic knowledge cognition hierarchy mining system provided in the embodiments of the present invention;

[0122] Figure 4 This is a schematic diagram comparing the experimental results of the TIMSS2019 dataset provided in this embodiment of the invention;

[0123] Figure 5 This is a schematic diagram of the training process of the TIMSS2019 dataset provided in this embodiment of the invention on the AUC metric;

[0124] Figure 6 This is a schematic diagram of the training process of the TIMSS2019 dataset on the ACC metric provided in this embodiment of the invention;

[0125] Figure 7 This is a schematic diagram of the training process of the TIMSS2019 dataset provided in this embodiment of the invention on the RMSE metric. Detailed Implementation

[0126] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0127] To address the problems existing in the prior art, this invention provides a method, system, device, and terminal for dynamic knowledge cognition level mining. The invention will be described in detail below with reference to the accompanying drawings.

[0128] like Figure 1As shown, the dynamic knowledge cognition hierarchy mining method provided in this embodiment of the invention includes the following steps:

[0129] S101, extract the content features of the questions in the student question interaction sequence, generate a heterogeneous graph based on learner-question-knowledge point, extract subgraphs from the learner-question-knowledge interaction heterogeneous graph, introduce graph convolutional neural network, and perform representation learning on question nodes and knowledge point nodes based on different graph structures;

[0130] S102, Construct a knowledge state update layer based on gating, and predict the learner's mastery level of each question and knowledge point in the next moment based on the updated question and knowledge point embedding representation;

[0131] S103 combines a matrix of learners' mastery levels of knowledge points with a linear neural network to dynamically track learners' cognitive levels.

[0132] S104 uses the learner's mastery level of each knowledge point and test question at the next moment to predict the student's answer score, and uses training parameters to analyze the learner's mastery of each knowledge point over time.

[0133] The principle diagram of the dynamic knowledge cognition hierarchy mining method provided in this embodiment of the invention is as follows: Figure 2 As shown.

[0134] like Figure 3 As shown, the dynamic knowledge cognition hierarchy mining system provided in this embodiment of the invention includes:

[0135] The node embedding representation learning module is used to generate corresponding heterogeneous graphs based on learner-question-knowledge interaction according to learner learning behavior characteristics and learning resource characteristics; it uses the learner-question-knowledge interaction heterogeneous graph to generate subgraphs, and introduces graph convolutional neural networks to perform embedding representation learning on question nodes and knowledge point nodes respectively based on different graph structures.

[0136] The knowledge status update module is used to construct a gate structure to perform temporal prediction on the updated question nodes and knowledge point nodes, and calculate the learner's mastery status of all questions and knowledge points at the next moment.

[0137] The dynamic cognitive tracking module is used to dynamically track the learner's cognitive level of all knowledge points by utilizing the learner's mastery state vector of all knowledge points based on a linear neural network.

[0138] The response prediction module is used to predict the learner's response at the next moment based on the learner's learning interaction sequence, the characteristics of the question the learner will answer at the next moment, and the learner's characteristics of the next question and the knowledge points it tests. The technical solution of the present invention will be further described below with reference to the explanation of symbols.

[0139] The symbols involved in the embodiments of the present invention are shown in Table 1.

[0140] Table 1. Symbols involved in the embodiments of the present invention.

[0141]

[0142]

[0143] S101 provided in the embodiments of the invention specifically includes:

[0144] (1.1) Collect learners' learning behavior characteristics and learning resource characteristics from learners' response sequences:

[0145] The TIMSS2019 dataset was compiled from interaction data collected in the seventh evaluation period of the IEA (International Energy Agency) TIMSS dataset across 64 countries and 8 benchmark systems in grades 4 and 8. Specifically, this embodiment used mathematics data from New Zealand grade 4 students for the experiment, removing some redundant data. The processed dataset contains information such as the longest number of interactions, the number of knowledge points, and the number of test questions, as shown in Table 2 below.

[0146] Table 2. Relevant information about the TIMSS2019 dataset.

[0147] The longest number of interactions between learners and practice 34 Number of knowledge points 12 Number of questions 213 Number of learners 5003 The highest cognitive level of knowledge points Level 3 Average number of knowledge points tested in the test 1.76

[0148] Learner learning behavior features and learning resource features were selected from the TIMSS2019 dataset. The selected external learning behavior features and resource features are shown in Table 3 below.

[0149] Table 3 shows the features selected from the TIMSS2019 dataset.

[0150]

[0151] (1.2) Generate an interaction heterogeneous graph based on learners-questions-knowledge points based on learners' answer sequences and learning resource features.

[0152] In the learner-response sequence, learners answer questions in a certain order, each question testing different knowledge points at different levels. This process involves numerous heterogeneous relationships between students, questions, and knowledge points. A heterogeneous graph SQK = {V, E} based on learner-question-knowledge-point relationships is constructed, where V represents the node set and E represents the edge set. The node set contains three types of nodes: student nodes n... s Question node n q and knowledge point node n kThe edge set contains two types of edges: answer edges e. a and the inspection of the border e t .

[0153] (1.3) The subgraphs extracted from the learner-question-knowledge interaction heterogeneous graph are as follows:

[0154] Extract a learner response behavior subgraph (SQ graph) to record learners' answers to test questions; extract a knowledge point testing subgraph (QK graph) to represent the heterogeneous relationship between test questions and knowledge points; extract a relationship subgraph between knowledge points (KK graph) to focus on the connections between knowledge points.

[0155] The learner's answers to the test questions are represented as an SQ graph, where SQ∈R. s_num×q_num SQ ij The number of times learner i answers question j, where s_num represents the number of learners and q_num represents the number of questions.

[0156] The relationship between test questions and knowledge points is represented as a QK graph, where QK∈R. q_num×k_num Where k_num represents the number of knowledge points, if question j tests knowledge point k, then the corresponding SQ jk The positional weight increases. QK(j) represents the j-th row of the matrix, i.e., the knowledge points associated with question j; (QK) T (k) represents the k-th column of the matrix, which represents the test questions associated with knowledge point k.

[0157] The relationship between knowledge points is represented as a KK graph, and the KK relationship is extracted based on the statistical transition probability matrix. Among them, KK ab n represents the degree of correlation between knowledge points a and b. a,b ∑ represents the number of times knowledge point b is tested after knowledge point a. k a,k represent the total number of times knowledge point a was tested;

[0158] (1.4) The specific process of introducing a graph convolutional neural network to learn representations of question nodes and knowledge point nodes based on different graph structures includes:

[0159] Before information propagation in the graph convolutional neural network, the node embedding representations and hidden knowledge states at time t are aggregated. The method is as follows:

[0160]

[0161]

[0162] Here, α represents a question node or a knowledge point node, and the method requires two sets of parameters to learn the embedded representation of questions and knowledge points. x represents the hidden state of the question or knowledge point at time t. t ∈{0,1,2,3} 4N α represents the learner's response at time t (0: wrong answer, 1: correct answer and knowledge, 2: correct answer and mastery; 3: correct answer and application), t This represents the initialization representation of the question node or knowledge point node at time t. x_α ∈R 4N×e E represents the embedding matrix between test questions or knowledge points and answers. α ∈R N×e It is an embedding matrix of test questions or knowledge points, E α (k) represents E α The k-th row represents the embedding of question k or knowledge point k, where e is the dimension of the embedding. x_α E α All of them are trainable matrices.

[0163] Update the information of the test questions and knowledge points themselves and their neighbors according to the graph structure.

[0164] The node's own information update method F self as follows:

[0165]

[0166]

[0167] In the formula, and These are the trainable weights and biases used for node aggregation, respectively. The activation function is ReLU.

[0168] Before updating the information of neighboring nodes, it is necessary to first aggregate the updated information of the current node, and then update the information of neighboring nodes. The method is as follows:

[0169]

[0170]

[0171] in, This is the updated embedding representation for the current node. and These are the trainable weights and biases used for neighbor node aggregation, respectively. The activation function is ReLU.

[0172] Specifically, since the graph structures on which the embedding representation of test question nodes and knowledge point nodes are updated are different, the information update formulas for their neighbor nodes are slightly different.

[0173] The methods for updating the neighbor node information of the current question node are as follows:

[0174]

[0175] In the formula, SQ(stu_id) represents the learner's answer status in the current row of the sequence number, and c t For the knowledge points being tested at the current moment, QK(c t This refers to the test questions related to the currently tested knowledge point;

[0176] The method for updating the neighbor node information of the current knowledge point node is as follows:

[0177]

[0178] In the formula, q t QK(q) represents the current test questions. t KK(c) represents the knowledge points related to the current test question. t ) and (KK) T These are the c-th graphs of the KK diagram. t line and c t The column represents the knowledge points related to the currently tested knowledge point.

[0179] S102 provided in this embodiment of the invention specifically includes:

[0180] (2.1) The updated test questions and knowledge points are embedded and fed into the erase-addition gate to construct the learner's forgetting and learning process in time:

[0181]

[0182] In the formula, G ea The specific calculation process for the erase-addition gate is as follows:

[0183]

[0184]

[0185]

[0186] In the formula, tmp_erase and tmp_add represent the erase and addition amounts of the gate structure, respectively, and W erase W add These are the weight matrices for the erase gate and the addition gate, respectively; b erase b add σ represents the bias terms corresponding to the erase gate and the addition gate, respectively, and σ is the activation function sigmoid.

[0187] (2.2) The mastery vector obtained after the erase-addition gate is fed into the gating gate to predict the learner's mastery level of the test questions and knowledge points at the next time step:

[0188]

[0189] In the formula, G gru The specific calculation process for the gating gate is as follows:

[0190]

[0191]

[0192]

[0193]

[0194] In the formula, To hide the knowledge vector for learners after the erase-addition gate, This is the hidden state from the previous moment. To reset the door, To update the door, Let σ represent the candidate hidden state, σ be the activation function sigmoid, and * represent the Hadamard product.

[0195] S103 provided in this embodiment of the invention specifically includes:

[0196] (3.1) Cognitive level prediction is performed using the knowledge hidden state obtained at time t+1 from the knowledge state update layer:

[0197]

[0198] In the formula, W l It is a cognitive state weight matrix for all knowledge point nodes, b k This is the offset of knowledge node k. Here, σ is the softmax function, which predicts whether the learner belongs to one of the four cognitive levels at that moment.

[0199] (3.2) Obtain the learner's current cognitive level of all knowledge points through the Argmax function.

[0200] C = Argmax(l t )

[0201] In the formula l t The cognitive hierarchy matrix of students for all concepts at the next moment is obtained by using the Argmax function to take the index of the item with the highest predictability, which is the cognitive hierarchy C of students for all knowledge points.

[0202] S104 provided in this embodiment of the invention specifically includes:

[0203] (4.1) The learner's level of mastery of the test questions and their level of mastery of the knowledge points tested by the test questions are represented as follows:

[0204]

[0205]

[0206]

[0207] In the formula, and It is a prediction weight matrix for answering knowledge point nodes and question nodes, b k and b q σ is the offset between knowledge point k and question q, and σ is the sigmoid function. and These represent the learner's level of mastery of the test questions and knowledge points at time t.

[0208] By fitting linear layers to students' understanding of the questions and knowledge points at time t, we can obtain the students' answer performance y. t .

[0209] (4.2) Predicting the learner's response at the next moment:

[0210] p t =y t ·q t+1

[0211] In the formula, q t+1 For the questions learners need to answer in the next moment, y t Let p be the student's response vector to the test question at this moment. t This is the probability that the learner can answer the question correctly.

[0212] (4.3) Based on the predicted learner responses and the actual learner responses, training is performed using the method of minimizing the negative log-likelihood loss:

[0213] L = logSoftmax(p t ,y true )

[0214]

[0215] Among them, y true p represents the actual learner responses. t The result is the learner's response predicted by the method.

[0216] (4.4) Use the Adam optimizer to update the network parameters of the above model.

[0217] An application embodiment of the present invention provides a computer device, which includes a memory and a processor. The memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the dynamic knowledge cognition hierarchy mining method.

[0218] An application embodiment of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the dynamic knowledge cognition hierarchy mining method.

[0219] An application embodiment of the present invention provides an information data processing terminal, which is used to implement the dynamic knowledge cognition hierarchy mining system.

[0220] The main development environment used in this invention experiment includes: Windows 10, Intel(R) Core(TM) i7-8565U CPU@1.80GHz, RAM 8.00GB, GTX 3090Ti graphics, PyTorch 1.2.0, Python 3.6, and the specific hyperparameter settings of the model are shown in Table 4 below.

[0221] Table 4. Hyperparameter settings of the experimental model

[0222] batch_size 32 epoch 50 dropout 0.5 learning_rate 0.001 embedding_dim 32 hidden_dim 32

[0223] This invention compares a dynamic knowledge cognitive hierarchy mining method based on learning-interactive heterogeneous graph representation with a cognitive hierarchy mining method based on graph deep learning, using the area under the curve (AUC), prediction accuracy (ACC), and root mean square error (RMSE) as comparison metrics. AUC and ACC are evaluation metrics for binary classification models, accurately describing the overall performance of the model even with imbalanced positive and negative samples. An AUC or ACC value of 0.5 represents a randomly obtainable score, with scores closer to 1 indicating more accurate predictions. RMSE quantifies the difference between predicted and true values; a lower RMSE indicates better model performance.

[0224] This invention was compared with a cognitive hierarchy mining method based on graph deep learning. To ensure a fair comparison, the hyperparameters of the same modules in both models were set to the same values. Table 5 shows the AUC, ACC, and RMSE comparison results of a dynamic knowledge cognitive hierarchy mining method based on learning interactive heterogeneous graph representations and a cognitive hierarchy mining method based on graph deep learning on the TIMSS2019 dataset. A schematic diagram of the model training process on the TIMSS2019 dataset is shown below. Figure 5 , Figure 6 , Figure 7 As shown.

[0225] Table 5 Comparison of experimental results from different methods Experimental results show that the dynamic knowledge cognitive hierarchy mining method based on heterogeneous graph representation of learning interaction proposed in this invention improves AUC by 2.8%, ACC by 6.2%, and reduces RMSE by 5.4% on the TIMSS2019 dataset. This indicates that this invention represents student response interactions as a heterogeneous graph, extracts and models high-dimensional information and relationships between learners, questions, and knowledge points, updates the embedding representations of question nodes and knowledge point nodes based on different graph structures, and uses the node representations updated by the graph neural network as input to the recurrent neural network. This effectively models learners' response behavior and utilizes the rich features contained in learning resources, resulting in better performance on student response prediction tasks. Its results outperform cognitive hierarchy mining methods based on graph deep learning. Experiments show that the dynamic knowledge cognitive hierarchy mining method based on heterogeneous graph representation of learning interaction proposed in this invention is more effective than the cognitive hierarchy mining method based on graph deep learning in terms of AUC, ACC, and RMSE. In conclusion, this invention has the best experimental results.

[0226] In summary, the dynamic knowledge cognition level mining method and system based on heterogeneous graph representation of learning interaction provided by this invention not only achieves more accurate prediction of learner response, but also predicts the learner's cognitive level of knowledge points, thereby assisting students to carry out more targeted learning work efficiently. This method collects learner learning behavior features and learning resource features from learner response sequences to generate a learner-question-knowledge interaction heterogeneous graph; extracts subgraphs from the learner-question-knowledge interaction heterogeneous graph, introduces graph convolutional neural networks, and performs representation learning on question nodes and knowledge point nodes based on different graph structures; constructs a gating-based knowledge state update layer to predict the learner's mastery level of questions and knowledge points at the next moment; constructs a dynamic cognitive tracking layer based on a linear neural network to dynamically track the learner's cognitive level according to the updated knowledge point mastery level vector; integrates the learner's mastery level of questions and the mastery level of knowledge points tested by the questions to predict the learner's response at the next moment, and constructs a loss function for training; and applies the trained parameters to predict the learner's learning knowledge state.

[0227] Example 1: Online Education Platform

[0228] Online education platforms can leverage this dynamic knowledge hierarchy mining method to improve the accuracy of their personalized recommendations and the efficiency of educational resource utilization.

[0229] 1. Online education platforms can collect students' learning behavior data, such as learning time, frequency, and learning content, and collect the characteristics of learning resources, such as test difficulty and knowledge points, to generate a learner-test-knowledge interaction heterogeneous graph.

[0230] 2. Use graph convolutional neural networks to extract subgraphs from heterogeneous graphs and learn representations for question nodes and knowledge point nodes.

[0231] 3. A gating-based knowledge state update layer is used to predict the student's mastery level of the test questions and knowledge points at the next moment, and a dynamic cognitive tracking layer based on a linear neural network is used to track the student's cognitive level.

[0232] 4. After predicting students' responses, the model is trained and optimized using a loss function.

[0233] 5. Finally, the trained model is used to predict students' knowledge status, and based on the prediction results, appropriate learning resources are pushed to students to achieve personalized learning.

[0234] Example 2: Enterprise Internal Training System

[0235] Companies can apply this approach to their internal training systems to more effectively assess employee learning progress and deliver personalized training content.

[0236] 1. Enterprises can record employees' learning behaviors on the training system, such as the courses they learn and the exercises they complete, and collect the characteristics of learning resources, such as the difficulty of the questions and the knowledge points, to generate a heterogeneous graph of learner-question-knowledge interaction.

[0237] 2. Use graph convolutional neural networks to extract subgraphs from heterogeneous graphs and learn representations for question nodes and knowledge point nodes.

[0238] 3. A knowledge state update layer based on gating is used to predict the employee's mastery level of test questions and knowledge points in the next moment, and a dynamic cognitive tracking layer based on a linear neural network tracks the employee's cognitive level.

[0239] 4. After predicting the employees' responses, the model is trained and optimized using a loss function.

[0240] 5. Finally, the trained model is applied to predict employees' learning knowledge status, and based on the prediction results, more suitable training content is provided to employees to achieve personalized training.

[0241] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.

[0242] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for dynamic knowledge cognitive hierarchy mining, characterized in that, This method employs a heterogeneous graph approach based on learner-question-knowledge interaction to understand learners' learning behaviors and learning resource characteristics. A graph convolutional neural network is used to extract subgraphs from the heterogeneous graph for representation learning. Furthermore, this method constructs a gated knowledge state update layer and a dynamic cognitive tracking layer based on a linear neural network, which can dynamically predict and track learners' mastery levels of questions and knowledge points, as well as their cognitive levels. By fusing learners' mastery levels of questions and the knowledge points tested in the questions, the method predicts learners' responses at the next moment and uses a loss function for training to predict learners' knowledge states. include: S1, collect learners’ learning behavior characteristics and learning resource characteristics, and generate a heterogeneous graph based on learner-test question-knowledge interaction; S2 extracts subgraphs from the learner-question-knowledge interaction heterogeneous graph, introduces a graph convolutional neural network, and performs representation learning on question nodes and knowledge node based on different graph structures; S3, construct a knowledge state update layer based on gating to predict the learner’s mastery level of test questions and knowledge points in the next moment; S4, construct a dynamic cognitive tracking layer based on a linear neural network, and dynamically track the learner's cognitive level according to the learner's updated knowledge level mastery horizontal vector; S5 integrates the learner's mastery level of the test questions and the mastery level of the knowledge points tested by the test questions to predict the learner's response at the next moment and constructs a loss function for training; S6 uses the trained parameters to predict the learner's knowledge status.

2. The dynamic knowledge cognition hierarchy mining method as described in claim 1, characterized in that, Learning behavior characteristics and learning resource characteristics are collected from learners' response sequences. Learning behavior characteristics are response characteristics that include temporal information. ； Triple group This indicates that at time t, student s answered the test question. This question tested knowledge points. , It reflects the students' answers; if a student answers incorrectly, If the student answers correctly, the value is 0. for Knowledge points The level of investigation; Learning resource characteristics include the knowledge point characteristics of the learning resources, specifically the cognitive level characteristics of the knowledge points tested in the test questions of the learning resources and the relationship characteristics between the knowledge points in the learning resources: ； Among them, the binary pair Question e, representing the test question, examines the following knowledge points. The level of investigation is M represents the number of knowledge points tested in this question; the course includes the following set of knowledge points. , The course includes a question bank of the following: , , This represents the level of knowledge point k being tested in the question, with 1 representing knowledge, 2 representing mastery, and 3 representing application. The learner-question-knowledge point heterogeneity graph is generated based on learners' answer sequences and learning resource characteristics. In the learner answer sequence, learners answer questions in a certain order, and each question tests different knowledge points, containing a large number of heterogeneous relationships between students, questions, and knowledge points. Where V represents the node set and E represents the edge set; the node set contains three types of nodes, namely student nodes. Test Question Nodes and knowledge point nodes ; The edge set contains two types of edges: answer edges and response edges. and the border .

3. The dynamic knowledge cognition hierarchy mining method as described in claim 1, characterized in that, The specific process of extracting subgraphs from the learner-question-knowledge interaction heterogeneous graph includes: Extract a subgraph of learner response behavior to record learners' answers to test questions; extract a subgraph of knowledge points tested by test questions to represent the heterogeneous relationship between test questions and knowledge points; extract a subgraph of relationships between knowledge points to focus on the connections between knowledge points. The learners' responses to the test questions are represented by an SQ graph. ,in The number of times learner i answers question j. Represents the number of learners. Represents the number of test questions; The relationship between test questions and knowledge points is represented by a QK graph. ,in This represents the number of knowledge points. If question j tests knowledge point k, then the corresponding... Position weight increases; This represents the j-th row of the matrix, which contains the knowledge points associated with question j. This represents the k-th column of the matrix, which contains the test questions associated with knowledge point k. The relationships between knowledge points are represented as KK diagrams, and the KK relationships are extracted based on the statistical transition probability matrix. ,in, This represents the degree of correlation between knowledge points a and b. This represents the number of times knowledge point b is tested after knowledge point a.

4. The dynamic knowledge cognition hierarchy mining method as described in claim 1, characterized in that, The specific process of introducing graph convolutional neural networks and learning representations for question nodes and knowledge point nodes based on different graph structures includes: Before information propagation in the graph convolutional neural network, the node embedding representations and hidden knowledge states at time t are aggregated: ； ； Where α represents a question node or a knowledge point node, the method requires two sets of parameters to learn the embedded representation of questions and knowledge points; This represents the hidden state of the question or knowledge point at time t. This represents the learner's response at time t; An embedding matrix representing test questions or knowledge points and answers. It is an embedding matrix of test questions or knowledge points. represent The k-th row represents the embedding of question k or knowledge point k, where e is the dimension of the embedding; All are trainable matrices; Based on the graph structure, update the information of the test questions and knowledge points themselves, as well as their neighbors. The method for updating the information of a node itself is as follows: ； ； In the formula, and These are the trainable weights and biases used for node self-aggregation; the activation function is ReLU; Before updating the information of neighboring nodes, it is necessary to first aggregate the updated information of the current node, and then update the information of neighboring nodes: ； ； in, This is the updated embedding representation for the current node. and These are the trainable weights and biases for neighbor node aggregation; the activation function is ReLU. Because the graph structures on which the embedding representation of the question node and the knowledge point node are updated are different, the information update formulas for their neighbor nodes are slightly different. The method for updating the information of the neighbor nodes of the current question node is as follows: ； In the formula, This shows the learner's answer status at the current moment, in the row numbered sequence. To test the knowledge points at the current moment, To update the neighbor node information of the current knowledge point node in relation to the questions being tested, the method is as follows: ； In the formula, This refers to the current test questions. Information on knowledge points related to the current test questions. and These are the KK diagrams. Action and The column represents the knowledge points related to the currently tested knowledge point.

5. The dynamic knowledge cognition hierarchy mining method as described in claim 1, characterized in that, The specific process of constructing a gating-based knowledge state update layer to predict learners' mastery level of test questions and knowledge points at the next moment includes: The updated test questions and knowledge points are embedded and fed into a gated neural network to model the learner's mastery of the knowledge points and test questions at time t+1. ； ； In the formula, The specific calculation process for the erase-addition gate is as follows: ； ； ； In the formula, and These represent the erase and addition amounts of the door structure, respectively. , These are the weight matrices for the erase gate and the addition gate, respectively; , These are the bias terms corresponding to the erase gate and the addition gate, respectively. The activation function is sigmoid; In the formula, The specific calculation process for the gating gate is as follows: ； ； ； ； In the formula, To hide the knowledge vector for learners after the erase-addition gate, This is the hidden state from the previous moment. To reset the door, To update the door, In the candidate hidden state, The activation function is sigmoid, and * represents the Hadamard product.

6. The dynamic knowledge cognition hierarchy mining method as described in claim 1, characterized in that, The specific process of constructing a dynamic cognitive tracking layer based on a linear neural network to dynamically track the learner's cognitive levels includes: Cognitive level prediction is performed using the hidden knowledge state obtained at time t+1 from the knowledge state update layer. ； In the formula, It is a cognitive state weight matrix for all knowledge point nodes. It is the offset of knowledge point node k. It is the softmax function, which predicts whether a student belongs to one of the four cognitive levels at a given moment. ； In the formula This refers to the student's cognitive hierarchy matrix for all concepts at the next moment. Using the Argmax function, the index of the highest predictive value is used to represent the student's cognitive hierarchy across all knowledge points. ; The learner's level of mastery of the test questions and their level of mastery of the knowledge points tested by the test questions are represented as follows: ； ； ； In the formula, and It is a weighted matrix for answer prediction based on knowledge point nodes and question nodes. and It is the offset between knowledge point k and question q. It is the sigmoid function; and These represent the learner's level of mastery of the test questions and knowledge points at time t; The model obtains students' answers by fitting the students' understanding of the questions and knowledge points at time t using a linear layer. And it is trained by minimizing the negative log-likelihood loss to obtain the best model performance; The prediction of the learner's response at the next moment is expressed as follows: ； In the formula, These are the test questions that learners will need to answer in the next moment. The result is the student's response vector to the test question at this moment. That is, the probability that the learner can answer the question correctly; Training is performed by minimizing the negative log-likelihood loss, based on the predicted learner responses and the actual learner responses. ； Finally, the Adam optimizer is used to update the network parameters of the above model.

7. A dynamic knowledge recognition hierarchy mining system, characterized in that, The dynamic knowledge cognition hierarchy mining system includes: The node embedding representation learning module is used to generate corresponding heterogeneous graphs based on learner-question-knowledge interaction according to learner learning behavior characteristics and learning resource characteristics; it uses the learner-question-knowledge interaction heterogeneous graph to generate subgraphs, and introduces graph convolutional neural networks to perform embedding representation learning on question nodes and knowledge point nodes respectively based on different graph structures. The knowledge status update module is used to construct a gate structure to perform temporal prediction on the updated question nodes and knowledge point nodes, and calculate the learner's mastery status of all questions and knowledge points at the next moment. The dynamic cognitive tracking module is used to dynamically track the learner's cognitive level of all knowledge points by utilizing the learner's mastery state vector of all knowledge points based on a linear neural network. The response prediction module is used to predict the learner's response at the next moment based on the learner's learning interaction sequence, the characteristics of the question the learner will answer at the next moment, and the characteristics of the learner's response to the next question and the knowledge points it tests.

8. A computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the dynamic knowledge cognitive hierarchy mining method as described in any one of claims 1 to 6.

9. An information data processing terminal, characterized in that, The information data processing terminal is used to implement the dynamic knowledge cognition hierarchy mining system as described in claim 7.

Images