A multi-level intelligent cognitive tracking method, system, storable medium and terminal

Abstract

The application belongs to the technical field of personalized learning, and discloses a multi-level intelligent cognitive tracking method, system, storable medium and terminal, the method comprising: introducing Bloom cognitive domain education target classification, constructing test question knowledge cognitive tensor TKC, collecting learning resources and answer data of learners, and generating a sequence of learner time sequence answer pairs; introducing a multi-attribute cognitive diagnosis method, combining a deep neural network, and constructing a cognitive level mining model; sorting and encoding the cognitive level mining results of the learners to obtain deep representation features, combining a self-attention mechanism, constructing a multi-level intelligent cognitive tracking model, and further predicting the answer performance of the learners on the test questions. The application is beneficial to accurately and finely modeling the overall knowledge structure and specific level of the learners, thereby promoting personalized learning of the learners and providing a new idea for mining and tracking the cognitive state and level of the learners in an online learning platform.

Description

Technical Field

[0001] This invention belongs to the field of personalized learning technology, and in particular relates to a multi-level intelligent cognitive tracking method, system, storage medium and terminal. Background Technology

[0002] Currently, with the rapid development of internet information technology centered on intelligence, information technology reform in the education sector has also seized the opportunity to ride the wave of this era. Among these advancements, intelligent assessment technologies, represented by learning diagnostics, have become a crucial aspect of educational informatization. By constructing learner profiles and analyzing learning behaviors, these technologies can rationally plan learning paths, efficiently recommend learning resources, and ultimately, through learning intervention, help learners tailor learning plans, identify and address weaknesses, conduct targeted reinforcement exercises, and construct a closed loop of learning, practice, and assessment, truly achieving individualized instruction and personalized learning.

[0003] Intelligent cognitive tracking technology is a key technology for profiling learners' learning behaviors. This technology can extract effective information from learners' answer sequences and the knowledge association matrix of test questions, mark the knowledge tested in the questions, depict the overall test structure, dynamically model and track learners' cognitive levels based on their potential trait differences and test characteristics, predict learners' performance at a future point in time, uncover the cognitive information hidden behind learners' scores, analyze learners' strengths and weaknesses, and finally provide targeted and personalized learning suggestions.

[0004] Currently, intelligent cognitive tracking models proposed by researchers in this field are mainly divided into two categories: static intelligent cognitive tracking based on statistical learning and dynamic intelligent cognitive tracking based on deep learning. Static intelligent cognitive tracking methods based on statistical learning, such as the Deterministic Inputs, Noisy “And” gate (DINA) model, typically model the learner's cognitive level as a binary variable (either 0 or 1), representing whether the learner has "mastered" or "not mastered" a particular piece of knowledge. This variable is then continuously updated using the Q-matrix of test attribute patterns and the EM algorithm, thereby modeling the learner's learning process. The structure and process of this type of method are relatively clear, and the size of the dataset has a relatively low impact on its performance. However, it also has some limitations, such as the fact that this method can only statically diagnose the learner's overall cognitive level based on their responses over a certain period, and cannot track specific knowledge points at specific moments. With the development of artificial intelligence and deep learning technologies, deep learning-based dynamic intelligent cognitive tracking methods have also rapidly emerged. These methods primarily use the learner's answer sequence as input, employing recurrent neural networks to model the learner's answering process. A hidden state vector represents the learner's overall cognitive state, and this hidden state is continuously updated based on the learner's answers. Deep learning-based dynamic intelligent cognitive tracking methods typically offer high accuracy and do not rely heavily on prior knowledge. However, due to the inherently poor interpretability of deep learning technology, these methods often cannot provide convincing explanations for their results, thus limiting their usability.

[0005] How to effectively utilize professional knowledge in the field of education to consider that learners' mastery of each knowledge point is not an extreme "either 0 or 1" when profiling learners' learning behavior, so as to build a more accurate and refined intelligent cognitive tracking model; how to capture the internal influence between different knowledge points and solve the problem of long-distance dependence while reducing the complexity of the model and reducing the computational power required by the model; and how to effectively combine the process of learners' cognitive level changing over time with the spatial change process of mutual influence between knowledge points have become the challenges of intelligent cognitive tracking.

[0006] Based on the above analysis, the problems and shortcomings of the existing technology are as follows:

[0007] (1) Traditional intelligent cognitive tracking methods simplify learners’ mastery of knowledge points and use an extreme “either 0 or 1” variable to represent learners’ mastery of each knowledge point, which does not match the actual learning process and results in coarse-grained model results.

[0008] (2) Traditional intelligent cognitive tracking methods have a relatively singular modeling perspective, ignore the internal influence between knowledge points, cannot "memorize" information over a longer distance, have greater model complexity, and require higher computing power.

[0009] (3) Traditional intelligent cognitive tracking methods usually do not consider combining the process of learners’ cognitive level changing over time with the spatial change process of mutual influence of knowledge points, resulting in poor interpretability of the model and making it difficult to be applied to the actual learning process. Summary of the Invention

[0010] To address the problems existing in the prior art, this invention provides a multi-level intelligent cognitive tracking method, system, storage medium, and terminal.

[0011] This invention is implemented as follows: a multi-level intelligent cognitive tracking method, characterized in that the multi-level intelligent cognitive tracking method includes the following steps:

[0012] Step 1: Introduce Bloom's Taxonomy of Educational Objectives in the Cognitive Domain, construct the modified Test-Knowledge-Cognition tensor (TKC), collect learners' learning resources and answer data, extract features, and generate a temporal sequence of test questions-knowledge points-Bloom hierarchy-answer status for each learner.

[0013] Step 2: Based on the learner's temporal response sequence and the knowledge-based cognitive tensor TKC that integrates Bloom's Taxonomy of Educational Objectives, a multi-attribute cognitive diagnostic method is introduced. Combined with a deep neural network, a cognitive hierarchy mining model is constructed to perform multi-level intelligent cognitive mining on learners.

[0014] Step 3: By sorting and encoding the learner's cognitive level vector obtained from intelligent cognitive mining, its deep representation features are obtained, and a deep self-attention network is introduced to construct a multi-level intelligent cognitive tracking model; based on the tracked learner cognitive level vector and the test question knowledge cognitive tensor TKC, the learner's answer performance on the test questions is predicted.

[0015] Furthermore, step one includes:

[0016] (1.1) Construct a set of learning resources and a set of learners' historical learning data:

[0017] S = {s1, s2, ..., s} M}

[0018] T = {t1, t2, ..., t} N}

[0019] K = {k1,k2,…,k} L}

[0020] Where S is the set of learners, M is the number of learners, T is the set of test questions, N is the number of test questions, K is the set of knowledge points, and L is the number of knowledge points;

[0021] (1.2) Starting from the traditional Q-matrix representing knowledge points in test questions, learners' mastery of knowledge points is divided into six cognitive levels. This yields a specific and unified definition of the examination level for each knowledge point at the cognitive level, thereby constructing a test knowledge cognitive tensor TKC that integrates Bloom's Taxonomy of Educational Objectives. The cognitive level ranges from 0 to 6, where 0: not mastered; 1: know; 2: comprehend; 3: apply; 4: analyze; 5: synthesize; 6: evaluate.

[0022] (1.3) Construct a learner response matrix R to record learners' historical response results:

[0023]

[0024] Where, r mn =0 indicates that learner m answered question n incorrectly, r mn =1 indicates that learner m answered question n correctly;

[0025] (1.4) Based on the knowledge points and their testing levels corresponding to the learner's answers at each moment, the learner's answer records are processed into a temporal answer pair sequence R of learner-question-knowledge point-Bloom level-answer status. s :

[0026] R s ={"user_id":m,

[0027] "test_id":n,

[0028] "score":x,

[0029] "knowledge_code":l,

[0030] "knowledge_level":c}

[0031] Where m is the learner ID, with a value range of 0 ≤ m ≤ M; n is the question ID, with a value range of 0 ≤ n ≤ N; x is the score of learner m on question n, with a value range of 0 / 1; l is the knowledge point ID examined by question n, with a value range of 0 ≤ l ≤ L. When question n examines more than one knowledge point, l can be represented as the set of knowledge point IDs; c is the level of knowledge point l examined by question n, with a value range of 1 ≤ c ≤ 6. When l is the set of knowledge point IDs, c can also be represented as the set of knowledge point levels, corresponding to each dimension in l.

[0032] Furthermore, the knowledge-cognition tensor TKC of the test questions is:

[0033] (t n ,k l ,c i )

[0034] Where the third dimension i = 0 indicates that question n does not test knowledge point l, and 1 ≤ i ≤ 6 indicates that question n tests knowledge points l to c. i level.

[0035] Furthermore, step two, which involves constructing a cognitive hierarchy mining model to perform multi-level intelligent cognitive mining on learners, specifically includes the following steps:

[0036] (2.1) Model learner factors, which include the cognitive level vector F. s :

[0037] h s =x s ×A

[0038] F s =f classifier (h s )

[0039] Where, x s Let A be the learner's one-hot vector, and A be the proficiency matrix of all learners. Multiplying them together yields the learner's proficiency vector h. s f classfier As a classification function, the learner's cognitive level vector F can be obtained. s ;

[0040] (2.2) Modeling test item factors, which include a basic knowledge point relevance vector F. kn Other optional factors F other This includes error parameters and guessing parameters for the test questions:

[0041] F kn =x e ×TKC

[0042]

[0043] Where, x e Here, h is the one-hot vector of the test question, TKC is the knowledge tensor of the test question, and h is the one-hot vector of the test question. g h is the guessing parameter for the test question, representing the probability that the learner answers correctly when they have not reached the cognitive level required to answer the question. sLet B and C be the error parameters of the test question, representing the probability that the learner will answer incorrectly when they have reached the cognitive level of the knowledge points tested in the test question.

[0044] The parameter matrices A, B, and C are obtained through data learning: x s x e x e h obtained by multiplying by parameter matrices A, B, and C respectively s h g h 1-g-s Substitute the item response function described in (2.3) to obtain the prediction results of the cognitive level mining model for the learner's answers. Use the loss function defined in (2.4) to calculate the cross-entropy between the prediction results and the learner's actual score. Train the parameter matrix according to the loss function value and the gradient descent rule.

[0045] (2.3) Construct a cognitive hierarchy mining model. Based on the multi-attribute cognitive diagnosis method, predict the learner's item response function as follows:

[0046]

[0047]

[0048] The obtained prediction results are then used to train a deep neural network:

[0049] f1=φ(W1×x T +b1)

[0050] f2 = φ(W2 × f1 + b2)

[0051] y = φ(W³ × f² + b³)

[0052] Where, k l For knowledge point ID, For the test question on knowledge point k l Relevance on For learners at knowledge point k l The cognitive level is represented by W1 and b1, which are the weight coefficients and bias coefficients in the linear regression function f1, respectively; W2 and b2 are the weight coefficients and bias coefficients in the linear regression function f2, respectively; and W3 and b3 are the weight coefficients and bias coefficients in the linear regression function y, respectively.

[0053] (2.4) Calculate the loss function of the model, which is the cross-entropy between the output y and the true label r:

[0054] loss CLMM =-∑(r i logy i +(1-r ilog(1-y) i ))

[0055] Where y is the predicted value and r is the actual score; after training, the learner's corresponding F s This represents the learning outcome for the learner. Each dimension corresponds to the learner's cognitive level on that knowledge point, with the cognitive level ranging from 0 to 6 (0: not mastered; 1: know; 2: comprehend; 3: apply; 4: analyze; 5: synthesize; 6: evaluate).

[0056] (2.5) Update all weight coefficients (W1, W2, W3) and bias coefficients (b1, b2, b3) in the model based on the loss function value obtained in (2.4) and the gradient descent rule.

[0057] Furthermore, step three includes:

[0058] (3.1) The multi-level cognitive level vectors obtained from the cognitive level mining model are combined with the learners' response time information to construct the Bloom educational goal cognitive level time sequence for each learner;

[0059] (3.2) Encode the learner’s cognitive level time sequence. If the learner’s cognitive level time sequence vector involves M knowledge points, the corresponding encoding length is M dimensions. For a certain knowledge point k, if the learner’s cognitive level is a, 0≤a≤6, the value corresponding to the kth dimension of the encoding is a+a*M. Thus, the deep representation features of each learner’s cognitive level vector can be obtained.

[0060] (3.3) Introduce a deep self-attention network Transformer to construct a multi-level intelligent cognitive tracking model based on Transformer. The multi-level intelligent cognitive tracking model based on Transformer includes an Embedding layer, a Self-attention layer, a Feed forward layer and a Prediction layer.

[0061] (3.4) The tracking model is backpropagated based on a loss function consisting of predicted classification and real label;

[0062] (3.5) Update all weight coefficients (W) in the model based on the loss function value obtained in (3.4) and the gradient descent rule. (1) W (2) ), bias coefficient (b) (1) b (2) ) and some other parameters (W a W p W k W Q WK W V ).

[0063] Furthermore, the cognitive level time sequence includes information such as the learner's cognitive level time sequence vector, the knowledge point ID corresponding to each dimension of the vector, and the total length of the time sequence vector.

[0064] Furthermore, a hyperparameter max step is set in the cognitive level time series to process the cognitive level time series of all learners into integer multiples of max step, and padding with "0" on the left for those that are not long enough to ensure consistent length.

[0065] Furthermore, the embedding layer is an embedding layer that establishes an embedding matrix based on the input knowledge point information and cognitive level information: and The two exist independently. It is the Masteredembedding matrix after adding position encoding. It is the Knowledge embedding matrix:

[0066] M = x a ×W a

[0067] P = x p ×W p

[0068]

[0069]

[0070] Where M is the initial Mastered embedding matrix, and P is the positional encoding matrix. This is the mastered embedding matrix with added positional encoding, where d is the dimension of the embedding layer. x is the knowledge embedding matrix; a It is the encoded temporal sequence of learners' cognitive levels, x p It is the positional information of the knowledge point in the sequence, x k It is the encoded knowledge point ID information, W a W p W k It is a trainable parameter matrix.

[0071] Furthermore, the self-attention layer is a self-attention layer, according to and We obtain the query (sequence), key, and value of the input information, and then use the scaled dot product attention mechanism to obtain the output Attention(Q,K,V) of the self-attention layer, as shown in the formula:

[0072]

[0073]

[0074]

[0075]

[0076] Among them, W Q W K W V It is a trainable parameter matrix, and d is a matrix QK. T Dimensions.

[0077] Furthermore, the feed forward layer is a forward propagation network that propagates the output of self-attention forward:

[0078] y = SW (1) +b (1)

[0079] y = ReLU(y)

[0080] y = yW (2) +b (2)

[0081] F = dropout(y)

[0082] Where S is the output of the self-attention layer, and W... (1) W (2) Let b be a trainable weight parameter matrix. (1) b (2) Given a trainable bias parameter matrix, ReLU is an activation function expressed as f(x) = max(0,x). When the input is negative, it outputs 0, meaning the neurons will not be activated. This implies that only a portion of the neurons are activated at any given time, resulting in a sparse network that is computationally efficient. The dropout function, during the training of a deep network, randomly and temporarily discards a portion of the neurons with a certain probability. Specifically, dropout is applied to each mini-batch of training data. Due to its mechanism of randomly discarding some neurons, it is equivalent to training a neural network with a different structure in each iteration.

[0083] Furthermore, the prediction layer is a fully connected layer. The output of self-attention is propagated forward to obtain matrix F, which is then passed through the prediction layer and activated by the softmax function to output the probability that each knowledge point is mastered by the learner at different levels.

[0084] p i =softmax(f classifier (F)).

[0085] Furthermore, the tracking model performs backpropagation based on a loss function composed of the predicted classification and the true label, wherein the loss function is:

[0086]

[0087] Where K is the category label, and in the model K=7, representing Level 0: Not mastered; Level 1: Known; Level 2: Comprehended; Level 3: Applied; Level 4: Analyzed; Level 5: Synthesized; Level 6: Evaluated; y is the label, where the sample category is i, then y i =1, otherwise y i =0; the output p of the prediction model i Let i be the probability that the sample belongs to class i.

[0088] Another object of the present invention is to provide a multi-level intelligent cognitive tracking system applying the aforementioned multi-level intelligent cognitive tracking method, characterized in that the multi-level intelligent cognitive tracking system comprises:

[0089] The Test Knowledge Cognition Tensor (TKC) construction module is used to construct the Test Knowledge Cognition Tensor (TKC) based on learners' learning resource information and by introducing Bloom's Taxonomy of Educational Objectives in the Cognitive Domain. Starting from the traditional Q-matrix representing test knowledge points, it divides learners' mastery of knowledge points into six ability levels, thereby constructing the Test Knowledge Cognition Tensor (TKC).

[0090] The answer sequence generation module is used to collect learners’ learning resources and answer data and extract features to generate a temporal answer sequence of questions-knowledge points-Bloom level-answer status for each learner.

[0091] The cognitive level mining module is used to construct a cognitive level mining model based on learners' historical answer records and the knowledge-based cognitive tensor TKC that integrates Bloom's Taxonomy of Educational Objectives. It introduces a multi-attribute cognitive diagnosis method and combines it with a deep neural network to perform multi-level intelligent cognitive mining on learners.

[0092] The multi-level intelligent cognitive tracking module is used to sort and encode the learner's cognitive level vector obtained by intelligent cognitive mining, obtain its deep representation features, and introduce a deep self-attention (Transformer) network to build a multi-level intelligent cognitive tracking model. Based on the learner's encoded deep representation features, the model is trained and the learner's cognitive level at the next moment is tracked.

[0093] The future performance prediction module is used to predict learners' performance on test questions based on the learner's cognitive level vector and the test question knowledge cognitive tensor TKC obtained from tracking.

[0094] The learning ability analysis module is used to analyze learners' learning abilities based on the learner's cognitive level vector obtained through tracking.

[0095] The personalized resource recommendation module is used to recommend learning resources to learners based on the learner's cognitive level vector and learning resource information obtained through tracking.

[0096] 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 following steps:

[0097] This study introduces Bloom's Taxonomy of Educational Objectives in the cognitive domain, constructs a Test Knowledge Cognition Tensor (TKC), collects learners' learning resources and response data, and generates learners' temporal response pairs. It also introduces a multi-attribute cognitive diagnostic method, combined with deep neural networks, to construct a cognitive hierarchy mining model. After sorting and encoding the learners' cognitive hierarchy mining results, its deep representation features are obtained. Combined with a self-attention mechanism, a multi-level intelligent cognitive tracking model is constructed to further predict learners' performance on test questions.

[0098] Another objective of this invention is to provide an information data processing terminal for implementing the aforementioned multi-level intelligent cognitive tracking system.

[0099] Based on the above technical solutions and the technical problems solved, please analyze the advantages and positive effects of the technical solution to be protected by this invention from the following aspects:

[0100] This invention provides a multi-level intelligent cognitive tracking method and system. First, it introduces Bloom's Classification of Educational Objectives (CCA) for cognitive domains, constructing a modified test knowledge cognitive tensor (TKC) based on the traditional Q-matrix representing test knowledge points. It collects learners' learning resources and answer data, extracts features, and generates a temporal sequence of test questions, knowledge points, Bloom's hierarchy, and answer responses for each learner. Then, based on the learners' temporal answer sequences and the TKC incorporating Bloom's Classification of Educational Objectives, it introduces a multi-attribute cognitive diagnostic method, combined with a deep neural network, to construct a cognitive hierarchy mining model for multi-level intelligent cognitive mining of learners. Finally, by sorting and encoding the learners' cognitive level vectors obtained from intelligent cognitive mining, its deep representation features are obtained, and a deep self-attention (Transformer) network is introduced to construct a multi-level intelligent cognitive tracking model. Based on the tracked learners' cognitive level vectors and the TKC, the model predicts the learners' answer performance on the test questions. This invention facilitates the accurate and effective fine-grained modeling and mining of learners' overall knowledge structure and specific cognitive level, thereby promoting personalized learning and providing new ideas for tracking and predicting learners' knowledge structure and cognitive level in online learning platforms.

[0101] This invention provides a more detailed model of learners' cognitive states, associating learners' cognitive levels at each knowledge point with the Test Knowledge Cognition Tensor (TKC) that incorporates Bloom's Taxonomy of Educational Objectives. Furthermore, it uses a custom cognitive level mining function to perform fine-grained mining of learners' cognitive levels, enabling accurate estimation of learners' overall knowledge mastery and specific cognitive levels based on TKC tensors at different levels.

[0102] This invention introduces a deep self-attention network, the Transformer, and integrates a custom knowledge-mastery self-attention mechanism. Compared to data-driven short-term neural networks, it considers long-term dependencies between sequences. While updating the learner's cognitive level at each moment, it not only considers explicit factors of the learner's actual mastery level but also implicit information about the mutual influence between knowledge points under the self-attention mechanism, effectively simulating changes in the learner's cognitive level in real-world learning scenarios. This invention achieves higher accuracy in mining and tracking the learner's overall knowledge structure and specific cognitive level, surpassing traditional intelligent cognitive tracking methods. It provides students with more accurate self-assessment information and teachers with more effective guidance.

[0103] This invention effectively mines learners' cognitive levels on knowledge points, and takes into account that learners' cognitive levels on different knowledge points may influence each other. It also uses a self-attention mechanism to model the relationship between knowledge points, which improves the accuracy of the model and makes it more realistic.

[0104] This invention uses the Test Knowledge Cognition Tensor (TKC) that integrates Bloom's Taxonomy of Educational Objectives to define cognitive level information, avoiding the use of the traditional Q matrix for a rough definition. It makes full use of professional knowledge in the field of education and improves the effectiveness and interpretability of the model.

[0105] The multi-level intelligent cognitive tracking method provided by this invention can mine learners' cognitive level on knowledge points, fully consider the long-term dependencies between answer sequences, and thus track and predict learners' overall knowledge structure and specific cognitive level. It provides learners with effective mining and tracking information, helps learners adjust their subsequent learning plans, fill in gaps in their knowledge, and improve learning efficiency.

[0106] The expected benefits and commercial value of the technical solution of this invention after transformation are as follows: This invention supports personalized learning with big data and artificial intelligence, realizes the educational concept of teaching according to aptitude, and can be widely used in fields such as smart education systems, intelligent learning guidance systems, intelligent teaching assistants, and adaptive learning systems, and has great commercial value.

[0107] The technical solution of this invention fills a technological gap in the domestic and international industry: By constructing a multi-level intelligent cognitive tracking model, this invention integrates professional knowledge and technology in the field of education, and designs a long-term dependency tracking model that integrates a custom knowledge mastery self-attention mechanism. This model can obtain more granular cognitive level information, filling the gap in the industry's technology that only constructs learner cognitive level vectors from the perspectives of "0" and "1", and cannot conduct multi-dimensional, comprehensive, and granular mining and tracking of learners' cognitive levels. Thus, it effectively models the learner's overall knowledge structure and specific cognitive level, providing comprehensive and accurate guidance information for learners' personalized learning.

[0108] The technical solution of this invention solves a long-standing but unresolved technical problem: in the mining and tracking of learners' cognitive levels, there has been a persistent desire to achieve a fine-grained, interpretable method for mining knowledge structures and cognitive levels. Learners' mastery of knowledge is not limited to extreme "0"s and "1s," but rather this cognitive level constantly changes as learning deepens or forgetting occurs. This invention, based on the mining of dynamic knowledge structures and cognitive levels, overcomes the technical challenge of real-time diagnosis and tracking of cognitive levels.

[0109] The technical solution of this invention overcomes technological bias: traditional intelligent cognitive tracking methods simply assume that learners' learning status for knowledge points falls into only two categories: "mastery" and "lack of mastery." This bias greatly hinders the application of intelligent cognitive tracking. This invention, starting from professional educational knowledge, constructs a multi-level intelligent cognitive tracking method, breaking through traditional technological bias and greatly enhancing the application value of this invention. Attached Figure Description

[0110] Figure 1 This is a flowchart of the multi-level intelligent cognitive tracking method provided in the embodiments of the present invention;

[0111] Figure 2 This is a tensor graph of the multi-level intelligent cognitive tracking method provided in the embodiments of the present invention;

[0112] Figure 3 This is an example diagram of the Test Knowledge Cognition TKC tensor provided in an embodiment of the present invention;

[0113] Figure 4 This is an example radar chart of learner cognitive level provided in an embodiment of the present invention;

[0114] Figure 5 Here are schematic diagrams of the multi-level intelligent cognitive tracking method provided in this embodiment of the invention: (a) is a schematic diagram of the method, and (b) is a schematic diagram of the cognitive level mining model.

[0115] Figure 6 This is a structural block diagram of the multi-level intelligent cognitive tracking system provided in the embodiments of the present invention;

[0116] The diagram shows: 1. Test Knowledge Cognition Tensor (TKC) Construction Module; 2. Answer Sequence Generation Module; 3. Cognitive Level Mining Module; 4. Multi-level Intelligent Cognitive Tracking Module; 5. Future Performance Prediction Module; 6. Learning Ability Analysis Module; 7. Personalized Resource Recommendation Module. Detailed Implementation

[0117] 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.

[0118] To enable those skilled in the art to fully understand how the present invention is specifically implemented, this section provides an explanatory description of the embodiments that expand upon the technical solutions of the claims.

[0119] like Figure 1 As shown, the multi-level intelligent cognitive tracking method provided in this embodiment of the invention includes the following steps:

[0120] S101 introduces Bloom's Taxonomy of Cognitive Domains and constructs a modified test knowledge cognitive tensor (TKC) based on the traditional Q matrix representing test knowledge points. It collects learners' learning resources and answer data and extracts features to generate a temporal sequence of test questions, knowledge points, Bloom's hierarchy, and answer status for each learner.

[0121] S102, based on learners' temporal response pairs and the knowledge-cognition tensor TKC of test questions that integrates Bloom's Taxonomy of Educational Objectives, introduces a multi-attribute cognitive diagnosis method, combines it with a deep neural network, and constructs a cognitive hierarchy mining model to perform multi-level intelligent cognitive mining of learners.

[0122] S103: By sorting and encoding the learner's cognitive level vector obtained from intelligent cognitive mining, its deep representation features are obtained, and a deep self-attention (Transformer) network is introduced to construct a multi-level intelligent cognitive tracking model; based on the tracked learner cognitive level vector and the test question knowledge cognitive tensor TKC, the learner's answer performance on the test questions is predicted.

[0123] Some symbols appearing in the embodiments of this invention are shown in Table 1.

[0124] Table 1 Symbol Description

[0125]

[0126]

[0127] Furthermore, S101 includes:

[0128] (1.1) Construct a set of learning resources and a set of learners' historical learning data:

[0129] S = {s1, s2, ..., s} M}

[0130] T = {t1, t2, ..., t} N}

[0131] K = {k1,k2,…,k} L}

[0132] Where S is the set of learners, M is the number of learners, T is the set of test questions, N is the number of test questions, K is the set of knowledge points, and L is the number of knowledge points;

[0133] (1.1.1) The data collected from the "C Programming Language" class of the 2021 undergraduate students of the School of Artificial Intelligence Education of Central China Normal University is shown in Table 2;

[0134] Table 2 Dataset Information

[0135] Number of students 51 Number of knowledge points 17 Number of questions 124 highest level of knowledge points 3 Total number of response records 6165 Longest number of student responses 124 Shortest number of student response records 10

[0136] (1.1.2) Perform dataset preprocessing, delete the answer records with missing values, and sort the answer records in chronological order;

[0137] (1.2) Starting from the traditional Q-matrix representing knowledge points in test questions, learners' mastery of knowledge points is divided into six cognitive levels. This yields a specific and unified definition of the examination level for each knowledge point at the cognitive level, thereby constructing a test knowledge cognitive tensor TKC that integrates Bloom's Taxonomy of Educational Objectives. The cognitive level ranges from 0 to 6, where 0: not mastered; 1: know; 2: comprehend; 3: apply; 4: analyze; 5: synthesize; 6: evaluate.

[0138] (1.3) Construct a learner response matrix R to record learners' historical response results:

[0139]

[0140] Where, r mn =0 indicates that learner m answered question n incorrectly, r mn =1 indicates that learner m answered question n correctly;

[0141] (1.4) Based on the knowledge points and their testing levels corresponding to the learner's answers at each moment, the learner's answer records are processed into a temporal answer pair sequence R of learner-question-knowledge point-Bloom level-answer status. s :

[0142] R s ={"user_id":m,

[0143] "test_id":n,

[0144] "score":x,

[0145] "knowledge_code":l,

[0146] "knowledge_level":c}

[0147] Where m is the learner ID, with a value range of 0 ≤ m ≤ M; n is the question ID, with a value range of 0 ≤ n ≤ N; x is the score of learner m on question n, with a value range of 0 / 1; l is the knowledge point ID examined by question n, with a value range of 0 ≤ l ≤ L. When question n examines more than one knowledge point, l can be represented as the set of knowledge point IDs; c is the level of knowledge point l examined by question n, with a value range of 1 ≤ c ≤ 6. When l is the set of knowledge point IDs, c can also be represented as the set of knowledge point levels, corresponding to each dimension in l.

[0148] In this C language dataset, an example of a learner's temporal response pair sequence can be represented as follows:

[0149] R s ={"user_id":23,

[0150] "test_id":11,

[0151] "score":1,

[0152] "knowledge_code":2,5,

[0153] "knowledge_level":1,3}

[0154] like Figure 2 , Figure 3 As shown, the knowledge cognition tensor TKC of the test questions is:

[0155] (t n ,k l ,c i )

[0156] Where the third dimension i = 0 indicates that question n does not test knowledge point l, and 1 ≤ i ≤ 6 indicates that question n tests knowledge points l to c. i Level. Based on Bloom's Taxonomy of Cognitive Objectives, educational experts labeled the knowledge points and levels tested in each test question, constructing different knowledge and ability tensors (TKC) for each question. In this C language dataset, the highest level of knowledge points tested in the test questions is Level 3 Application.

[0157] Furthermore, the construction of the cognitive hierarchy mining model in S102, which performs multi-level intelligent cognitive mining on learners, specifically includes the following steps:

[0158] (2.1) Model learner factors, which include the cognitive level vector F. s :

[0159] h s =x s ×A

[0160] F s =f classifier (h s )

[0161] Where, x s Let A be the learner's one-hot vector, and A be the proficiency matrix of all learners. Multiplying them together yields the learner's proficiency vector h. s f classfier As a classification function, the learner's cognitive level vector F can be obtained. s Example radar chart of learner cognitive level Figure 4 As shown;

[0162] (2.2) Modeling test item factors, which include a basic knowledge point relevance vector F. kn Other optional factors F other This includes error parameters and guessing parameters for the test questions:

[0163] F kn =x e ×TKC

[0164]

[0165] Where, x e Here, h is the one-hot vector of the test question, TKC is the knowledge tensor of the test question, and h is the one-hot vector of the test question. g h is the guessing parameter for the test question, representing the probability that the learner answers correctly when they have not reached the cognitive level required to answer the question. s Let B and C be the error parameters of the test question, representing the probability that the learner will answer incorrectly when they have reached the cognitive level of the knowledge points tested in the test question.

[0166] The parameter matrices A, B, and C are obtained through data learning: x s x e x e h obtained by multiplying by parameter matrices A, B, and C respectively s h g h 1-g-s Substitute the item response function described in (2.3) to obtain the prediction results of the cognitive level mining model for the learner's answers. Use the loss function defined in (2.4) to calculate the cross-entropy between the prediction results and the learner's actual score. Train the parameter matrix according to the loss function value and the gradient descent rule.

[0167] (2.3) Construct a cognitive hierarchy mining model. Based on the multi-attribute cognitive diagnosis method, predict the learner's item response function as follows:

[0168]

[0169]

[0170] The obtained prediction results are then used to train a deep neural network:

[0171] f1=φ(W1×x T +b1)

[0172] f2 = φ(W2 × f1 + b2)

[0173] y = φ(W³ × f² + b³)

[0174] Where, k l For knowledge point ID, For the test question on knowledge point k l Relevance on For learners at knowledge point k l The cognitive level is represented by W1 and b1, which are the weight coefficients and bias coefficients in the linear regression function f1, respectively; W2 and b2 are the weight coefficients and bias coefficients in the linear regression function f2, respectively; and W3 and b3 are the weight coefficients and bias coefficients in the linear regression function y, respectively.

[0175] (2.4) Calculate the loss function of the model, which is the cross-entropy between the output y and the true label r:

[0176] loss CLMM =-∑(r i logy i +(1-r i log(1-y) i ))

[0177] Where y is the predicted value and r is the actual score; after training, the learner's corresponding F s This represents the learning results for the learner. Each dimension corresponds to the learner's cognitive level on that knowledge point. The cognitive level ranges from 0 to 6, and for the C language dataset, the cognitive level range is 0 to 3 (0: not mastered; 1: know; 2: comprehend; 3: apply).

[0178] (2.5) Update all weight coefficients (W1, W2, W3) and bias coefficients (b1, b2, b3) in the model based on the loss function value obtained in (2.4) and the gradient descent rule.

[0179] The experimental parameter settings for this embodiment of the invention are shown in Table 3.

[0180] The experimental development platform used the PyTorch framework. All learners' responses were divided by week, and cognitive level mining was performed on each week to obtain cognitive level vectors for learners at different time points. In this step, AUC and ACCURACY were used as evaluation metrics. The experimental results are shown in Table 4.

[0181] Table 3 Experimental Parameter Settings

[0182] parameter value batch_size 32 epoch 200 / 500 / 800 dropout 0.5 learning rate 0.001 linear_num 1024 / 512 / 256 / 128

[0183] Table 4. Experimental Results of Step 1

[0184] time AUC ACC week2 0.7650 0.8000 week3 0.7533 0.8900 week4 0.8663 0.8000 week5 0.6756 0.6600 week6 0.8181 0.7450 week9 0.9077 0.8500 week11 0.7478 0.7200 week14 0.8555 0.7867 week15 0.8022 0.7667 week16 0.7337 0.7000

[0185] Furthermore, S103 includes:

[0186] (3.1) The multi-level cognitive level vectors obtained from the cognitive level mining model are combined with the learners' response time information to construct the Bloom educational goal cognitive level time sequence for each learner;

[0187] (3.2) Encode the learner’s cognitive level time sequence. If the learner’s cognitive level time sequence vector involves M knowledge points, the corresponding encoding length is M dimensions. For a certain knowledge point k, if the learner’s cognitive level is a (0≤a≤6), the value corresponding to the kth dimension of the encoding is a+a*M. Thus, the deep representation features of each learner’s cognitive level vector can be obtained.

[0188] (3.3) Introduce a deep self-attention network Transformer to construct a multi-level intelligent cognitive tracking model based on Transformer. The multi-level intelligent cognitive tracking model based on Transformer includes an Embedding layer, a Self-attention layer, a Feed forward layer and a Prediction layer.

[0189] (3.4) The tracking model is backpropagated based on a loss function consisting of predicted classification and real label;

[0190] (3.5) Update all weight coefficients (W) in the model based on the loss function value obtained in (3.4) and the gradient descent rule. (1) W (2) ), bias coefficient (b) (1) b (2) ) and some other parameters (W a W p W k W Q W K WV ).

[0191] Furthermore, the cognitive level time sequence includes information such as the learner's cognitive level time sequence vector, the knowledge point ID corresponding to each dimension of the vector, and the total length of the time sequence vector.

[0192] Furthermore, a hyperparameter max step is set in the cognitive level time series to process the cognitive level time series of all learners into integer multiples of max step, and padding with "0" on the left for those that are not long enough to ensure consistent length.

[0193] Furthermore, the embedding layer is an embedding layer that establishes an embedding matrix based on the input knowledge point information and cognitive level information: and The two exist independently. It is the Masteredembedding matrix after adding position encoding. It is the Knowledge embedding matrix:

[0194] M = x a ×W a

[0195] P = x p ×W p

[0196]

[0197]

[0198] Where M is the initial Mastered embedding matrix, and P is the positional encoding matrix. This is the mastered embedding matrix with added positional encoding, where d is the dimension of the embedding layer. x is the knowledge embedding matrix; a It is the encoded temporal sequence of learners' cognitive levels, x p It is the positional information of the knowledge point in the sequence, x k It is the encoded knowledge point ID information, W a W p W k It is a trainable parameter matrix.

[0199] Furthermore, the self-attention layer is a self-attention layer, according to and We obtain the query, key, and value of the input information, and then use the scaled dot product attention mechanism to obtain the output Attention(Q,K,V) of the self-attention layer, as shown in the formula:

[0200]

[0201]

[0202]

[0203]

[0204] Among them, W Q W K W V It is a trainable parameter matrix, and d is a matrix QK. T Dimensions.

[0205] Furthermore, the feed forward layer is a forward propagation network that propagates the output of self-attention forward:

[0206] y = SW (1) +b (1)

[0207] y = ReLU(y)

[0208] y = yW (2) +b (2)

[0209] F = dropout(y)

[0210] Where S is the output of the self-attention layer, and W... (1) W (2) Let b be a trainable weight parameter matrix. (1) b (2) Given a trainable bias parameter matrix, ReLU is an activation function expressed as f(x) = max(0,x). When the input is negative, it outputs 0, meaning the neurons will not be activated. This implies that only a portion of the neurons are activated at any given time, resulting in a sparse network that is computationally efficient. The dropout function, during the training of a deep network, randomly and temporarily discards a portion of the neurons with a certain probability. Specifically, dropout is applied to each mini-batch of training data. Due to its mechanism of randomly discarding some neurons, it is equivalent to training a neural network with a different structure in each iteration.

[0211] Furthermore, the prediction layer is a fully connected layer. The output of self-attention is propagated forward to obtain matrix F, which is then passed through the prediction layer and activated by the softmax function to output the probability that each knowledge point is mastered by the learner at different levels.

[0212] p i =softmax(f classifier (F)).

[0213] Furthermore, the tracking model performs backpropagation based on a loss function composed of the predicted classification and the true label, wherein the loss function is:

[0214]

[0215] Where K is the category label, K=7 in the model and K=4 in the model applied to the C language dataset, representing level 0: not mastered; level 1: known; level 2: comprehended; level 3: applied; y is the label, if the sample category is i, then y i =1, otherwise y i =0; the output p of the prediction model i Let i be the probability that the sample belongs to class i.

[0216] The experimental parameter settings for this embodiment of the invention are shown in Table 5.

[0217] Table 5 Experimental Parameter Settings II

[0218] parameter value batch_size 32 epoch 200 / 500 / 800 dropout 0.2 learning rate 0.01 max step 10 hidden_num 128

[0219] The principle diagram of the multi-level intelligent cognitive tracking method provided in this embodiment of the invention is as follows: Figure 5 As shown.

[0220] like Figure 6 As shown, the multi-level intelligent cognitive tracking system provided in this embodiment of the invention includes:

[0221] The Test Knowledge Cognition Tensor TKC Construction Module 1 is used to construct the Test Knowledge Cognition Tensor TKC based on learners' learning resource information and by introducing Bloom's Taxonomy of Educational Objectives in the Cognitive Domain. Starting from the traditional Q-matrix representing test knowledge points, it divides learners' mastery of knowledge points into six ability levels, thereby constructing the Test Knowledge Cognition Tensor TKC.

[0222] The answer sequence generation module 2 is used to collect learners' learning resources and answer data and extract features to generate a time-series answer sequence for each learner, consisting of questions, knowledge points, Bloom level, and answer status.

[0223] The cognitive level mining module 3 is used to construct a cognitive level mining model based on learners' historical answer records and the knowledge cognitive tensor TKC of test questions that integrates Bloom's Taxonomy of Educational Objectives in the cognitive domain. It introduces a multi-attribute cognitive diagnosis method and combines it with a deep neural network to perform multi-level intelligent cognitive mining on learners.

[0224] The multi-level intelligent cognitive tracking module 4 is used to sort and encode the learner's cognitive level vector obtained by intelligent cognitive mining, obtain its deep representation features, and introduce a deep self-attention (Transformer) network to build a multi-level intelligent cognitive tracking model. Based on the learner's encoded deep representation features, the model is trained and the learner's cognitive level at the next moment is tracked.

[0225] Future performance prediction module 5 is used to predict learners' performance on test questions based on the learner's cognitive level vector and test question knowledge cognitive tensor TKC obtained from tracking.

[0226] Learning ability analysis module 6 is used to analyze learners’ learning abilities based on the learners’ cognitive level vectors obtained through tracking.

[0227] The personalized resource recommendation module 7 is used to make personalized learning resource recommendations for learners based on the learner's cognitive level vector and learning resource information obtained from tracking.

[0228] To demonstrate the inventiveness and technical value of the technical solution of this invention, this section provides specific product or related technology application examples of the technical solution claimed.

[0229] The multi-level intelligent cognitive tracking method provided in the application embodiment of the present invention is applied to 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, the processor performs the steps of the multi-level intelligent cognitive tracking method.

[0230] 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.

[0231] The embodiments of the present invention have achieved some positive results during the research and development or use process, and have indeed great advantages compared with the prior art. The following content describes them in conjunction with the data, charts and other information of the experimental process.

[0232] To illustrate the positive effects of the embodiments of the present invention, the multi-level intelligent cognitive tracking method proposed in this invention is used to track and model the learner's cognitive level, thereby predicting the learner's subsequent answer performance. In addition, the experimental results of this method are compared with those of traditional intelligent cognitive tracking methods, using AUC and ACCURACY as evaluation indicators. The comparison results are shown in Table 6.

[0233] Table 6 Experimental Results

[0234] method AUC ACC Traditional intelligent cognitive tracking methods 61.82％ 57.36％ This method 67.91％ 60.79％

[0235] The experimental results show that, compared with traditional intelligent cognitive tracking methods, the multi-level intelligent cognitive tracking method proposed in this invention improves the AUC and ACC by 6.09% and 3.43% respectively. This indicates that the invention can better explore learners' mastery of different knowledge points, track learners' cognitive levels at different times, and has higher accuracy in predicting learners' answer performance, making it more effective than traditional intelligent cognitive tracking methods.

[0236] Experimental results demonstrate that the multi-level intelligent cognitive tracking method provided in this invention, which integrates the Test Knowledge Cognition Tensor (TKC) based on Bloom's Taxonomy of Cognitive Objectives and the cognitive level mining model, is effective in fine-grained mining of learners' cognitive levels. It can accurately estimate learners' overall knowledge mastery and specific cognitive levels based on TKC tensors at different levels. Furthermore, by introducing a deep self-attention network (Transformer) and integrating a custom knowledge mastery self-attention mechanism, a prediction model that considers long-term dependencies between sequences is designed. This modeling of students' cognitive levels helps to model the changes in learners' cognitive levels in both temporal and spatial senses. While updating learners' cognitive levels at each moment, it not only considers the explicit factors of learners' actual responses but also the implicit information of mutual influence between knowledge points under the self-attention mechanism. This effectively simulates the changes in learners' knowledge states in actual learning scenarios, enabling more accurate mining and tracking of learners' overall knowledge structure and specific cognitive levels.

[0237] 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 multi-level intelligent cognitive tracking method, characterized in that, The multi-level intelligent cognitive tracking method comprises the following steps: Step one, introducing the Bloom cognitive domain educational objective classification, constructing the modified test knowledge cognitive tensor TKC, collecting the learning resources and answer data of the learners and performing feature extraction, and generating the time-series test-knowledge point-Bloom level-answer situation answer pair sequence for each learner; Step two, based on the time-series answer pair sequence of the learners and the test knowledge cognitive tensor TKC fused with the Bloom cognitive domain educational objective classification, introducing the multi-attribute cognitive diagnosis method, combining the deep neural network, and constructing the cognitive level mining model to perform multi-level intelligent cognitive mining on the learners; Step three, sorting and encoding the cognitive level vector of the learners obtained through intelligent cognitive mining to obtain the deep representation features, introducing the deep self-attention network, and constructing the multi-level intelligent cognitive tracking model; and predicting the answer performance of the learners on the test according to the cognitive level vector of the learners obtained through tracking and the test knowledge cognitive tensor TKC; The step one comprises: (1.1) constructing a learning resource set and a learner historical learning data set: Wherein, S is a learner set, M is the number of learners, T is a test question set, N is the number of test questions, K is a knowledge point set, and L is the number of knowledge points; (1.2) starting from the traditional test knowledge point representation Q matrix, dividing the mastery of knowledge points by learners into six cognitive levels to obtain specific and unified definitions of the examination levels of each knowledge point at the cognitive level, thereby constructing the test knowledge cognitive tensor TKC fused with the Bloom cognitive domain educational objective classification, wherein the cognitive level ranges from 0 to 6, wherein 0: not mastered; 1: know; 2: understand; 3: apply; 4: analyze; 5: synthesize; and 6: evaluate; S = {s1, s2,..., S M} T = {t1, t2,..., t N} K = {k1, k2,..., k L} (1.3) constructing a learner answer matrix R to record the historical answer results of the learners: "test_id":n, "score":x, where r mn = 0 indicates that the learner m answered the question n incorrectly, and r mn = 1 indicates that the learner m answered the question n correctly. (1.4) According to the knowledge point and the examination level corresponding to the learner's answer to the question at each moment, the learner's answer record is processed into a time sequence answer pair sequence R of learner-question-knowledge point-Bloom level-answer condition s : R s = {"user_id": m, "knowledge_code":l, "knowledge_level":c} Wherein, m is a learner ID, the value range is 0≤m≤M; n is a test ID, the value range is 0≤n≤N; x is the score of the learner m on the test n, the value range is 0 / 1; l is the knowledge point ID of the test n, the value range is 0≤l≤L, when the test n examines more than one knowledge point, l can represent the set of examined knowledge point IDs; c is the level of the test n examined knowledge point l, the value range is 1≤c≤6, when l is a set of examined knowledge point IDs, c can also represent a set of examined knowledge point levels, each dimension in l corresponds to each dimension in c; The test knowledge cognitive tensor TKC is: ​ ​ (t n ,k l ,c i ) wherein the third dimension i = 0 represents that the test question n does not examine the knowledge point l, and 1≤i≤6 represents that the test question n examines the knowledge point l to c i horizontal.

2. The multi-tiered intelligent cognitive tracking method of claim 1, wherein, The step two of constructing the cognitive level mining model includes the following steps. (2.1) modelling learner factors, said learner factors being a cognitive level vector F s : h s = x s x A F s = f classifier (h s ) Wherein, x s is the one-hot vector of the learner, A is the proficiency parameter matrix of all learners, and the multiplication can obtain the proficiency vector h s of the learner classfier , f s is the classification function, and the cognitive level vector F of the learner can be obtained. (2.2) modelling test item factors, including a basic knowledge point relevance vector F kn and other optional factors F other , including a test item failure parameter and a guessing parameter: F kn = x e x TKC wherein x e is the one-hot vector of the test question, TKC is the knowledge cognition tensor of the test question, h g is the guessing parameter of the test question, representing the probability of the learner answering correctly when the learner has not reached the cognitive level of the knowledge point examined by the test question, h s is the error parameter of the test question, representing the probability of the learner answering incorrectly when the learner has reached the cognitive level of the knowledge point examined by the test question, and B and C are parameter matrices. (2.3) Constructing the cognitive level mining model, and predicting the item response function of the learner based on the multi-part attribute cognitive diagnosis method: The obtained prediction result is trained by using a deep neural network: f1 = φ(W1 x x + b1) T + b1) f2 = φ (W2 x f1 + b2) y = φ (W3 x f2 + b3) Where, k l For knowledge point ID, For the test question on knowledge point k l Relevance on For learners at knowledge point k l The cognitive level is represented by W1 and b1, which are the weight coefficients and bias coefficients in the linear regression function f1, respectively; W2 and b2 are the weight coefficients and bias coefficients in the linear regression function f2, respectively; and W3 and b3 are the weight coefficients and bias coefficients in the linear regression function y, respectively. (2.4) Calculating the loss function of the model, and the loss function is the cross entropy between the output prediction value y and the real label r: After the training, the F s That is the mining result of the learner, and each dimension corresponds to the cognitive level of the learner on the knowledge point. (2.5) Updating all weight coefficients W1, W2, W3 and bias coefficients b1, b2, b3 in the model according to the loss function value obtained in (2.4) and the gradient descent rule. The parameter matrix A, B and C are obtained by training through data learning, and the process is as follows: First, x s , x e , x e are multiplied by parameter matrices A, B, and C respectively to obtain h s , h g , h 1-g-s , which are substituted into the item response function to obtain the predicted results of the cognitive level mining model for the answers of the learners. Secondly, the cross entropy between the prediction result and the real score of the learner is calculated by using the loss function, and the parameter matrix is trained according to the loss function value and the gradient descent rule.

3. The multi-tiered intelligent cognitive tracking method of claim 1, wherein, The step three includes: (3.1) Constructing the Bloom educational goal cognitive level time sequence of each learner by combining the multi-level cognitive level vector obtained from the cognitive level mining model and the answer time information of the learner, wherein the cognitive level time sequence includes the cognitive level time vector of the learner, the knowledge point ID corresponding to each dimension in the vector and the total length of the time vector; (3.2) Encoding the cognitive level time sequence of the learner, if the cognitive level time vector of the learner involves M knowledge points, the corresponding encoding length is M dimensions; for the knowledge point k, if the cognitive level of the learner to it is a, 0≤a≤6, the value corresponding to the kth dimension of the encoding is a+a*M, thereby obtaining the deep representation feature of the cognitive level vector of each learner; (3.3) Introducing a deep self-attention network Transformer, and constructing a multi-level intelligent cognitive tracking model based on the Transformer, wherein the multi-level intelligent cognitive tracking model based on the Transformer includes an Embedding layer, a Self-attention layer, a Feed forward layer and a Prediction layer; (3.4) The tracking model is reversely propagated based on the loss function composed of the prediction classification and the real label; (3.5) updating the weight coefficients W in the model according to the loss function value and gradient descent rule obtained in (3.4) (1) , W (2) , the bias coefficient b (1) , b (2) , and the parameter W a , W p , W k , W Q , W K , W V .

4. The multi-tiered intelligent cognitive tracking method of claim 3, wherein, The super parameter max step is set in the cognitive level time sequence, and the cognitive level time sequences of all learners are processed into an integer multiple of max step, and "0" is used for left padding if the length is not enough, so as to ensure the consistency of the length; The Embedding layer is an embedding layer, and an Embedding matrix is established according to input knowledge point information and cognitive level information coding: and Both are independent of each other; The is the Mastered embedding matrix with position encoding added, and is the Knowledge embedding matrix: where M is the initial Mastered embedding matrix, P is the position encoding matrix, d is the dimension of the embedding layer; x a is the encoded learner cognitive level time series, x p is the position information of the knowledge point in the sequence, x k is the encoded knowledge point ID information, W a , W P , W k is a trainable parameter matrix.

5. The multi-tiered intelligent cognitive tracking method of claim 3, wherein, The Self-attention layer is a self-attention layer, according to the and the Obtain the sequence of input information query, key and value value, expressed as: The scaled dotproduct attention mechanism is used to obtain the output Attention (Q, K, V) of the self-attention layer, which is expressed as: where W Q , W K , and W V are trainable parameter matrices, and d is the dimension of the matrix QK T .

6. The multi-tiered intelligent cognitive tracking method of claim 3, wherein, The Feed forward layer is a forward propagation network, which performs forward propagation on the output of the self-attention, and then iteratively trains the matrix F after randomly discarding a part of the neural nodes by using the dropout function, and the process is as follows: y = SW (1) + b (1) y = ReLU(y) y = yw (2) + b (2) F = dropout(y) where S is the output of the self-attention layer, W (1) , W (2) are trainable weight parameter matrices, b (1) , b (2) are trainable bias parameter matrices, and ReLU is an activation function with the expression: f(x) = max(0, x). The Prediction layer is a full connection layer, and the matrix F is output through the Prediction layer and a softmax activation function, outputting the probability of each knowledge point being mastered by the learner at different levels: p i = softmax(f classifier (F)) ; The tracking model performs back propagation based on a loss function composed of the predicted classification and the true label, and the loss function is: wherein K is a category label, K = 7, including level 0: no grasp; level 1: know; level 2: understand; level 3: apply; level 4: analyze; level 5: synthesize; level 6: evaluate; the sample category of label y is i, then y i = 1, otherwise y i = 0.

7. A multi-level intelligent cognitive tracking system for implementing the multi-level intelligent cognitive tracking method according to any one of claims 1-6, characterized by, The multi-level intelligent cognitive tracking system comprises: A test question knowledge cognitive tensor TKC construction module, which is configured to introduce Bloom cognitive domain educational objective classification based on learning resource information of a learner, divide mastery of knowledge points by the learner into six ability levels from a traditional test question knowledge point representation Q matrix, and thereby construct a test question knowledge cognitive tensor TKC; An answer pair sequence generation module, which is configured to collect learning resources and answer data of the learner and perform feature extraction, and generate a time-series test question-knowledge point-Bloom level-answer condition answer pair sequence for each learner; A cognitive level mining module, which is configured to introduce a multi-attribute cognitive diagnosis method based on historical answer records of the learner and the test question knowledge cognitive tensor TKC fused with Bloom cognitive domain educational objective classification, combine a deep neural network, construct a cognitive level mining model, and perform multi-level intelligent cognitive mining on the learner; A multi-level intelligent cognitive tracking module, which is configured to sort and encode a cognitive level vector of the learner obtained through intelligent cognitive mining, obtain deep representation features of the cognitive level vector, introduce a deep self-attention network, construct a multi-level intelligent cognitive tracking model, train the model according to the deep representation features of the learner after encoding, and track a cognitive level of the learner at a next time; A future performance prediction module, which is configured to predict answer performance of the learner on a test question based on the cognitive level vector of the learner obtained through tracking and the test question knowledge cognitive tensor TKC; A learning ability analysis module, which is configured to perform learning ability analysis on the learner based on the cognitive level vector of the learner obtained through tracking; An individualized resource recommendation module, which is configured to perform individualized learning resource recommendation for the learner based on the cognitive level vector of the learner obtained through tracking and learning resource information. 8.A computer readable storage medium storing a computer program, the computer program being executed by a processor to cause the processor to perform steps of the multi-level intelligent cognitive tracking method according to any one of claims 1-6.

Images