Knowledge-time sequence-behavior ternary second-order coupled dynamic cognitive diagnosis method

Abstract

The invention provides a learning process-oriented ternary second-order coupling dynamic cognitive diagnosis method, which comprises the following steps of: dividing learning units in a learning process into'knowledge, time sequence and behavior 'triples; knowledge-time sequence and behavior-time sequence first-order coupling analysis knowledge state and behavior state changes are obtained through a neural network model and a behavior feature extraction technology, meanwhile, a guess gate, a proficiency gate and a learning gate are set by adopting a GLU gating technology, complex modeling of cognitive state changes of learners in the learning process is achieved, and the learning efficiency is improved. Obtaining a potential ability state of the learner; finally, through multi-task prediction of behavior characteristics and reaction results, diagnosis of the cognitive state of the learner and prediction of future performance of the learner are achieved. According to the invention, the learning condition of the learner can be scientifically and comprehensively predicted, and the purpose of assisting the teacher in precise teaching is achieved.

Description

Technical Field

[0001] The present invention relates to the fields of educational big data mining, attention network and learner modeling, and in particular to a knowledge-time sequence-behavior ternary second-order coupled dynamic cognitive diagnosis method. Background Art

[0002] Cognitive diagnosis is the key to characterizing learning subjects and carrying out personalized learning in an intelligent education environment. Cognitive diagnosis aims to obtain the most intuitive answer data based on the interaction between learners and test questions during the test-taking process, build a cognitive diagnosis model to model the learner's internal cognitive processing process, and then diagnose the learner's knowledge mastery status, and evaluate the effectiveness of the model by predicting the learner's future performance. The results of cognitive diagnosis can help teachers and students quickly locate learners' weak knowledge. Teachers design corresponding intervention strategies based on the group's learning situation to carry out differentiated teaching, and learners carry out targeted remedial exercises based on their personal learning situation to achieve large-scale personalized learning. Therefore, it is urgent to use intelligent technology to implement accurate diagnosis, comprehensively characterize learners' cognitive levels, and reduce students' burdens and increase efficiency by providing accurate education services to achieve high-quality personalized learning.

[0003] The modeling technology of cognitive diagnosis has evolved from the traditional cognitive diagnosis model based on artificially constructed probabilistic graph structure modeling to the intelligent cognitive diagnosis model based on deep learning that automatically learns to improve the fit between learners and project features. According to the different data sources of cognitive diagnosis, current cognitive diagnosis can be roughly divided into two categories: static cognitive diagnosis and dynamic cognitive diagnosis.

[0004] Static cognitive diagnosis aims to infer the knowledge mastery status through diagnostic evaluation of individual cognitive processes, processing skills or knowledge structure, mainly for single learning scenarios, such as final exams. The most classic cognitive diagnosis model, such as the DINA model, uses errors and guessed parameters in the learning process to construct diagnostic functions and infer the learner's knowledge mastery status. With the development of artificial intelligence technology, many studies have introduced deep learning methods into static cognitive diagnosis modeling, namely deep cognitive diagnosis models. For example, Cheng et al. proposed the DIRT model, which uses deep learning algorithms to mine test texts and the relationship between test questions and concepts, thereby enhancing the test question representation to strengthen the diagnostic process; Wang et al. proposed the NeuralCD framework, which uses neural networks to capture the complex interactive relationship between learners and test questions; Gao et al. proposed the DeepCDM model, which uses deep learning to simulate learners' mastery of skills and test questions. However, due to the lack of attention to historical prior data in the learning process, the current results of static cognitive diagnosis are one-sided.

[0005] In order to solve the one-sided problem of static cognitive diagnosis, dynamic cognitive diagnosis came into being. Dynamic cognitive diagnosis accumulates multiple assessment data to form time series data, and obtains the changes in the learner's knowledge mastery status by analyzing the time series data. For example, the classic BKT model abstracts the cognitive state into a set of binary variables, takes the learner's real-time interaction as input, and simulates the changes in the learner's knowledge mastery during the learning process through the hidden Markov model. In fact, due to the good performance of deep learning algorithms, dynamic cognitive diagnosis based on deep learning has become a research hotspot in the field of cognitive diagnosis. For example, the DKT model uses a recurrent neural network to track the learner's cognitive state at different times; the DKVMN model uses a memory network to store the learner's implicit cognitive state, and dynamically updates the network through key-value pairs each time the answer is given; and the EKT model makes full use of natural language processing, domain ontology replacement, and convolutional neural networks to characterize multimodal test features to enhance cognitive diagnosis performance.

[0006] However, most of the current dynamic cognitive diagnosis models based on deep learning take into account the learner's historical answer records, and preliminarily realize the "knowledge-time sequence" binary coupling through time series modeling, which can dynamically track the learner's cognitive level. However, the existing models are still mainly based on result-oriented behavior, and the integration of scattered process behavior characteristics still cannot reflect the relationship between "knowledge, time sequence and behavior". This binary coupling that weakens process behavior has limitations and easily leads to incomplete characterization of learners' performance and inconsistency between their performance and their cognitive state. Summary of the invention

[0007] In order to overcome the above-mentioned deficiencies of the prior art, the present invention provides a knowledge-sequence-behavior three-element second-order coupled dynamic cognitive diagnosis method, comprising the following steps:

[0008] Step 1: Take a certain moment in the learning process as a unit to obtain the learning unit data when students participate in the assessment, which consists of knowledge features, behavior features and time series features. According to the chronological order, several learning units can form the learning process sequence data, so as to construct a time series data set. Knowledge features include the questions answered by students at a certain moment and the concepts related to the questions. Behavior features include the click stream formed by students repeatedly modifying their answers during the answering process, and the activity stream formed by students viewing the analysis and video resources after the answering. Time series features refer to the time interval between each time unit, the time and number of repetitions of the same concept learning interval, etc. Based on the constructed time series data set, a ternary standardized representation of the learning unit is carried out. For knowledge features, each concept related to the current question is first embedded, and then the average pooling method is used to obtain the concept representation associated with the question, which is added to the embedded question to obtain the knowledge representation of the current answer question; for time series features, the embedding method is used to achieve representation.

[0009] Step 2: For knowledge, time series features and behavior, time series features, neural network models and behavior feature extraction techniques are used to implement the first-order coupling analysis of "knowledge-time series" and "behavior-time series" cognitive states. Specifically, for knowledge-time series coupling, knowledge and time series representations are input into the LSTM neural network, and forgetting reduction and learning enhancement functions are set to dynamically adjust; for behavior-time series coupling, five click feature vectors and four activity features are extracted based on the timestamp information in the click stream and activity stream using behavior feature extraction technology, and they are encoded by addition and self-attention networks with causal masks.

[0010] Step 3, based on the behavior analysis and encoding of clickstream and activity stream, combined with knowledge state analysis, realize the second-order coupling analysis of the ability state of knowledge-behavior-time sequence coupling. Specifically, first, based on the output result of the first-order coupling, set the GLU gating function to obtain the student's initial cognitive state; then set the "guessing gate" and "proficiency gate" through clickstream encoding to weaken and enhance the current cognitive state respectively; then, set the "learning gate" through the average pooling of the activity stream to further enhance the current cognitive state; finally, use the time distance attention mechanism to calculate the student's comprehensive ability state representation.

[0011] Step 4: Based on the acquired comprehensive ability status, the final performance prediction is enhanced through the dual prediction tasks of behavior and reaction. Specifically, three behavior prediction tasks are set: reaction time, click stream length, and activity stream length, and the ability status is converted to the corresponding behavior space, and then the behavior prediction task is completed through the multi-layer perceptron; aggregate learning behavior prediction, and the prediction of the final reaction result in the future moment is realized through the multi-layer perceptron.

[0012] Furthermore, the ternary standardized representation of the learning unit specifically includes:

[0013] (1-1) Knowledge Feature Embedding

[0014] All test questions and concepts together constitute the knowledge space, which is represented by a triple (Q, C, τ), where Q is a non-empty set of test questions, C is a non-empty set of concepts, and τ is a mapping from Q to concept C. For any test question q n ∈Q, τ(q) represents the set of concepts required to solve question q. One question can be associated with multiple concepts. Specifically, it includes three steps: concept embedding, question embedding, and comprehensive knowledge representation.

[0015] First, concept embedding: concept embedding is used to establish the association between test questions and multiple concepts. Specifically, each concept is first embedded, and then the comprehensive knowledge representation of the test question is obtained through average pooling, that is, in, is the overall embedding representation of the concepts involved in the nth test question, is the embedding representation of the concepts involved in the nth test question, d k is the dimension of the embedding, c n,i From q n The knowledge concept set τ(q n );

[0016] Secondly, question embedding: the questions are encoded in an embedded manner, and the embedding of the question index involved in the nth learning unit is represented as

[0017] Finally, the comprehensive knowledge representation of the test question: embedded by the index of the test question and concept embedding Add them together to get k n =q n +c n ,in A comprehensive representation of the knowledge of the test questions;

[0018] (1-2) Temporal feature embedding

[0019] Each learning unit in the learning process occurs at a certain moment and is recorded as the timestamp when the learning event occurs. The timestamps in the learning trajectory are used to extract temporal features, including timestamp distance, time interval, and number of knowledge reviews.

[0020] First, the timestamp distance is used to measure the real distance between learning units. By calculating the timestamp distance between the adjacent answers, the time interval Δs is recorded. n ;

[0021] Secondly, the time interval refers to the time between the current knowledge answer and the last answer of the same knowledge point, which is recorded as the review time interval Δr n ;

[0022] Then, the number of knowledge reviews refers to the cumulative number of answers to the same knowledge point during the learning process, and the number of knowledge reviews is Δc n ; Thus, the timing feature T n =(Δs n ,Δr n ,Δc n );

[0023] (1-3) Behavioral feature extraction

[0024] Use clickstream and activity stream to describe the multi-faceted learning behavior characteristics in the learning process;

[0025] First, extract the clickstream cs from the learning trajectory n ;When answering a test question nWhen learning, learners may modify their answers many times due to uncertainty, thus forming several click answers, which are extracted in chronological order to form a click stream. in, Indicates answering the test question q n The result of the i-th click answer;

[0026] Second, extract the activity flow from the learning trajectory as n ; After completing each test, learners will think and learn further by looking at the answer analysis and watching the video lecture. The learners' behavior after answering the questions is extracted to form an activity flow, recorded as in Indicates that the student answers the test question q n The i-th activity after that includes viewing the explanation or watching the video lecture, i.e.

[0027] Furthermore, the specific implementation process of knowledge-timing coupling is as follows:

[0028] First, the time series feature T n =(Δs n ,Δr n ,Δc n ) are embedded into the same dimension d k After that, the comprehensive representation of the time series features is obtained by adding Then embed it with knowledge features Concatenate to form the final input vector x n , the comprehensive representation of the learning unit is as follows:

[0029]

[0030] The knowledge feature k n and the time series characteristics T n They are input into the LSTM neural network together to capture the dynamic relationship between knowledge points and learning time factors;

[0031] Then, based on the original forget gate, the timestamp distance is used to adjust the forgetting degree. The calculation method is as follows:

[0032] f n =σ(W f ·[h n-1 ,x n ]+b f )·exp(-α·Δs n )

[0033] in, is the adjusted forget gate output, is the hidden state of the previous learning unit; σ is the sigmoid activation function, exp(·) is the exponential function, and is the trainable weight matrix and bias term; Δs n is the timestamp distance between adjacent learning units, and α is an adjustment parameter used to control the impact of timestamp distance on the forget gate;

[0034] Next, in order to enhance the learning process, the review interval Δr is introduced n and knowledge review times Δc n To adjust the output of the input gate, the calculation method is as follows:

[0035]

[0036] in, is the input gate output, and is a trainable parameter; γ is a set hyperparameter used to control the impact of time interval and number of reviews on the input gate;

[0037] Finally, based on the adjusted forget gate f n and input gate i n , update the cell state of LSTM, and update the hidden state. The calculation method is as follows:

[0038]

[0039] in, is the cell state of LSTM, is a candidate cell state, and is a trainable parameter; through the output gate o n Compute the hidden state:

[0040] o n =σ(W o ·[h n-1 ,x n ]+b o ), h n =o n tanh(C n )

[0041] in, is the output of the output gate, Represents the hidden state of the current learning unit.

[0042] Furthermore, behavior-timing coupling includes clickstream behavior-timing coupling and activitystream behavior-timing coupling;

[0043] The specific implementation process of click flow behavior-sequence coupling is as follows:

[0044] First, according to the click stream According to its timestamp information, five behavioral change characteristics are extracted: ① Read whether the learner's answer at time t = 0 matches the standard answer, that is, whether the answer is correct or not, and record the "first answer correct or incorrect" ”; ② Using the same method, read the correctness information of the last modification at time t=i, and record the correctness of the final answer ”; ③ Calculate the difference between the final time and the initial time, and record the “reaction time t consumed” n ”, where t n =t i -t0; ④ Calculate the length of the entire click stream, record "click stream length l n =|cs n ⑤ Analyze the behavioral changes of learners in the clickstream. Learners may eventually modify their answers, which is recorded as "modification" type. They may also change their answers repeatedly and then return to the initial answer, which is recorded as "hesitation". This variable is called "clickstream type p n ”;

[0045] Secondly, these five embedding vectors are combined into one vector and mapped to a unified vector space in is the activity flow cs when the nth answer is given n The embedding representation of k is the embedding dimension, and emb represents the corresponding embedding layer function;

[0046] The specific implementation process of activity flow behavior-timing coupling is as follows:

[0047] First, according to the activity flow According to its timestamp information, four behavior change characteristics are extracted: ① Read the type of the current activity and record the "activity type" ",in, ② Calculate the time from the start to the end of the activity and record the "total activity time" "; ③ Get the activity is to view the analysis The number of times, record "View the number of times parsing "; ④ The acquisition activity is to watch the video The number of times the video is operated is recorded as "Video Operation Count" ”;

[0048] Then, an embedding matrix is ​​used to represent the four activity flow features, where each row of the matrix corresponds to an activity, namely in is the activity flow when answering the nth question nThe embedding representation of the i-th activity of is; the corresponding activities in the activity stream are combined to form an activity stream Among them n is the length of the active stream.

[0049] Furthermore, the specific implementation of step 3 includes:

[0050] (3-1) Based on the output results of the first-order coupling, the GLU gating function is set to obtain the student's initial cognitive state, and the guessing gate and the proficiency gate are set through click stream coding to weaken and enhance the current cognitive state respectively;

[0051] (3-2) Setting GLU-based learning gates based on activity flow to enhance the current cognitive state;

[0052] (3-3) Through the time-distance attention mechanism, the comprehensive ability status representation of students is obtained.

[0053] Furthermore, the specific implementation of (3-1) is as follows:

[0054] First, the learner's knowledge state is controlled by the gate function GLU(), which can be expressed as:

[0055]

[0056] in, is the initial cognitive state, is the output result of the first-order coupling of the cognitive state in the previous step, and Represents the corresponding weight and bias parameters;

[0057] Second, based on the students’ modification operations in the clickstream, the degree of guessing during the learner’s answering process is calculated; the cognitive state is adjusted by setting a guessing gate, and the function is expressed as:

[0058] K′ n =K n -GLU g (CS n ·W g +b g ),

[0059] in, is the state of knowledge after passing through the guessing gate; represents the embedding representation of the click stream, and To guess the weight and bias parameters of the gate;

[0060] Third, based on the time attributes in the clickstream, the learner’s proficiency in knowledge is inferred; the knowledge state is enhanced by setting a proficiency gate, and the function is expressed as:

[0061] K″ n =K′ n +GLU p (CS n ·W p +b p ),

[0062] in, It is the state of knowledge after passing the proficiency gate; and are the weight and bias parameters of the proficiency gate.

[0063] Furthermore, the specific implementation of (3-2) is as follows:

[0064] In order to unify the representation learning activities, the average pooling method is used to obtain the average value of each activity embedding in the activity stream, thereby aggregating all activity information. The calculation method is as follows:

[0065]

[0066] in, is the aggregate representation of the activity flow of the nth answer, For activity streams n The length of , that is, the number of activities, represents the embedding representation of the i-th activity in the n-th answer;

[0067] Secondly, the learning activities after answering the questions can enhance the learner's knowledge mastery, and the learning gate is set to simulate the improvement of their high-order cognitive state; the function is expressed as:

[0068] K″′ n =K″ n +GLU l (AS n ),

[0069] in, represents the knowledge state after being enhanced by the learning gate, GLU is the knowledge state before entering the learning gate. l It is a learning gate function used to enhance the influence of activity flow on knowledge state.

[0070] Furthermore, the specific implementation of (3-3) is as follows:

[0071] Based on the current cognitive state h n And (3-2) get the knowledge state change K″′ caused by the behavior n , and further infer the potential comprehensive ability state of students θ through the time distance attention mechanism n+1; In order to introduce the influence of time factors, the query vector is set to the question representation k at time n+1 n+1 , and use an exponential decay function to weight the historical information at different times, and introduce the time interval between the n+1 answer unit and each historical unit in the attention calculation;

[0072] First, define the query, key, and value vectors: n+1 Use the test question embedding representation k at time n+1 n+1 Generate, and the key and value vectors come from the current and all historical learning units; the calculation is as follows:

[0073] Q n+1 =W Q ·k n+1

[0074] K 1:n =[W K h1,W K ·h2,...,W K ·h n ]

[0075] V 1:n =[W V h1,W V ·h2,...,W V ·h n ]

[0076] in, represents the test embedding at the n+1th moment, is the mapping matrix, used to map questions and hidden states to query, key and value spaces; is the query vector, is the key vector matrix, is a matrix of value vectors;

[0077] Then, through the query vector Q n+1 and the key vector K 1:n Calculate the similarity and combine it with the time decay function exp(-βΔt n+1,i ) adjusts the attention weight; the calculation formula of attention weight is:

[0078]

[0079] Among them, α n+1,i represents the attention weight of the n+1th unit to the i-th historical unit, Δt n+1,i The time interval between the n+1th unit and the i-th historical unit. β is a hyperparameter of time decay, which is used to control the effect of the time interval on the attention weight.

[0080] Finally, the time-weighted attention weight α is usedn+1,i With the value vector V 1:n Perform weighted summation to generate the potential comprehensive ability state θ of the n+1th learning unit n+1 :

[0081]

[0082] in, Represents the potential comprehensive ability state of the inferred n+1th learning unit.

[0083] Furthermore, in step 4, the reaction time prediction formula is expressed as: The loss function is denoted by L t , where t n+1 is the actual reaction time at time n+1;

[0084] The click stream length prediction formula is expressed as: The loss function is denoted by L l , Among them l n+1 is the actual click length at time n+1;

[0085] The activity flow length prediction formula is expressed as: The loss function is denoted by L a , where a n+1 is the actual activity length at time n+1;

[0086] in is the predicted reaction time, is the predicted clickstream length, is the predicted activity flow length, θ n+1 It is the potential ability state of the n+1th one. In each behavior prediction, a feedforward network layer FNN is set to convert the ability state into the corresponding behavior space, and then the prediction is completed through the respective multi-layer perceptron MLP.

[0087] Furthermore, the prediction of the final reaction result in step 4 includes:

[0088] First, we obtain the behavior aggregation representation by aggregating the three vectors used to predict the multivariate learning behavior and the current potential ability state θ n+1 , forming a comprehensive behavioral representation:

[0089]

[0090] in, is the behavior representation of the n+1th unit, combining the average information of each behavior feature; FNN(·) represents the n+1Feed-forward network layers that map to their respective behavior spaces;

[0091] Secondly, based on the aggregated behavior representation R n+1 , set up a multilayer perceptron to predict the learner's final response result, and the calculation process is expressed as:

[0092]

[0093] in, is the predicted reaction result, MLP is a multi-layer perceptron; the loss function is recorded as L r , where r n+1 is the true response at time n+1;

[0094] The final loss function is obtained by weighted summation, that is, L = L r +λ1L t +λ2L l +λ3L a , where λ1, λ2, and λ3 are hyperparameter weights used to adjust the impact of the latter two loss functions.

[0095] Compared with the prior art, the beneficial effects of the present invention are as follows: the existing dynamic cognitive diagnosis is essentially a binary coupling of "sequence-knowledge", and the binary coupling of weakened procedural behavior has limitations, which will eventually lead to the problem of incomplete learning characterization. The method of the present invention proposes a dynamic cognitive diagnosis method of ternary second-order coupling for the learning process, which divides the learning units in the learning process into "knowledge, sequence, behavior" triplets, and obtains the first-order coupling of "knowledge-sequence" and "behavior-sequence" through neural network models and behavioral feature extraction technology to analyze the changes in knowledge state and behavioral state. At the same time, the GLU gating technology is used to set the guessing gate, proficiency gate and learning gate to achieve complex modeling of the learner's cognitive state changes during the learning process, and obtain the learner's potential ability state; finally, through the multi-task prediction of behavioral characteristics and reaction results, the diagnosis of the learner's cognitive state and the prediction of the learner's future performance are realized. The present invention can scientifically and comprehensively predict the learner's learning situation, so as to achieve the purpose of assisting teachers in precise teaching. BRIEF DESCRIPTION OF THE DRAWINGS

[0096] Figure 1 This is the framework diagram of this example method.

[0097] Figure 2 is a flowchart of the method of this example.

[0098] Figure 3 This is an example diagram of the structure of a learning unit.

[0099] Figure 4 This is a comparison diagram of the coupling mechanism between this example method and the traditional cognitive diagnosis method. DETAILED DESCRIPTION

[0100] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0101] A knowledge-time-behavior triple second-order coupling dynamic cognitive diagnosis method, such as Figure 1 As shown in the figure, it is a ternary second-order cognitive diagnosis architecture. The bottom layer uses "learning units" to describe educational situations. Learning units include three types of characteristics: knowledge, timing, and behavior. For knowledge and behavior characteristics, first-order coupling analysis is performed with timing characteristics to obtain cognitive states, that is, a first-order coupling analysis of knowledge-timing is performed to obtain the evolution of learners' knowledge states, and a first-order coupling analysis of behavior-timing is performed to analyze learners' behavioral changes. Then, a second-order coupling of knowledge states and behavioral states is performed at the time granularity to obtain the comprehensive ability states of learners. According to the ability states, behavior and reaction dual tasks are set to enhance learners' performance prediction.

[0102] The embodiment of the present invention provides a knowledge-sequence-behavior three-element second-order coupled dynamic cognitive diagnosis method, comprising the following steps:

[0103] Step 1: Take a certain moment in the learning process as a unit to obtain the learning unit data when students participate in the assessment, which consists of knowledge features, behavior features and time series features. According to the chronological order, several learning units can form the learning process sequence data, so as to construct a time series data set. Knowledge features include the questions answered by students at a certain moment and the concepts related to the questions. Behavior features include the click stream formed by students repeatedly modifying their answers during the answering process, and the activity stream formed by students viewing the analysis and video resources after the answering. Time series features refer to the time interval between each time unit, the time and number of repetitions of the same concept learning interval, etc. Based on the constructed time series data set, a ternary standardized representation of the learning unit is carried out. For knowledge features, each concept related to the current question is first embedded, and then the average pooling method is used to obtain the concept representation associated with the question, which is added to the embedded question to obtain the knowledge representation of the current answer question; for time series features, the embedding method is used to achieve representation.

[0104] Step 2: For knowledge, time series features and behavior, time series features, neural network models and behavior feature extraction techniques are used to implement the first-order coupling analysis of "knowledge-time series" and "behavior-time series" cognitive states. Specifically, for knowledge-time series coupling, knowledge and time series representations are input into the LSTM neural network, and forgetting reduction and learning enhancement functions are set to dynamically adjust; for behavior-time series coupling, five click feature vectors and four activity features are extracted based on the timestamp information in the click stream and activity stream using behavior feature extraction technology, and they are encoded by addition and self-attention networks with causal masks.

[0105] Step 3, based on the behavior analysis and encoding of clickstream and activity stream, combined with knowledge state analysis, realize the second-order coupling analysis of the ability state of knowledge-behavior-time sequence coupling. Specifically, first, based on the output of the first-order coupling, set the GLU gating function to obtain the student's initial cognitive state; then, based on the clickstream encoding, set the guessing gate and the proficiency gate to weaken and enhance the current cognitive state respectively; then, set the GLU-based learning gate based on the activity stream to enhance the current cognitive state; finally, through the time distance attention mechanism, obtain the student's comprehensive ability state representation.

[0106] Step 4: Based on the acquired comprehensive ability status, the final performance prediction is enhanced through the dual prediction tasks of behavior and reaction. Specifically, three behavior prediction tasks are set: reaction time, click stream length, and activity stream length, and the ability status is converted to the corresponding behavior space, and then the behavior prediction task is completed through the multi-layer perceptron; aggregate learning behavior prediction, and the prediction of the final reaction result in the future moment is realized through the multi-layer perceptron.

[0107] The specific implementation of each step is as follows:

[0108] (1) Ternary Normalized Representation of Learning Units

[0109] like Figure 2 The "learning unit" layer in , describes the three-element standardized representation of the learning unit. First, define a learning unit consisting of three elements: knowledge, timing, and behavior, denoted by LC i =(K i ,T i ,B i ), which is used to represent the learning events that occur within a certain time unit i. i Indicates the knowledge characteristics of the test, such as the test questions and the concepts they consider, T i represents the time series features in time series data, B i It represents the behavioral characteristics of the learner in the learning process. The learner's learning process X consists of N+1 learning units LC i Composition, let X = {LC1, LC2, ..., LCn ,...,LC N+1}, where LC n Represents the nth learning unit.

[0110] Example: The structure of the learning process is as follows: Figure 3 As shown in Figure 2, the learning process consists of multiple learning units, LC i =(K i ,T i ,B i ), where K i Refers to the knowledge formed by the test questions and related concepts done by students at this moment. A test question q includes multiple concepts c; T i Refers to the temporal characteristics at the i-th moment, such as the sequence time interval, review time interval and number of knowledge reviews; B i Refers to the temporal features at the i-th moment, including click stream and activity stream.

[0111] The calculation process of the standardized representation of the learning unit includes: knowledge feature embedding, time series feature embedding, and behavior feature embedding.

[0112] (1-1) Knowledge Feature Embedding

[0113] All test questions and concepts together constitute the knowledge space, which is represented by a triple (Q, C, τ). Among them, Q is a non-empty set of test questions, C is a non-empty set of concepts, and τ is a mapping from Q to concept C. For any test question q n ∈Q, τ(q) represents the set of concepts required to solve question q, and one question can be associated with multiple concepts. Specifically, it includes three steps: concept embedding, question embedding, and comprehensive knowledge representation.

[0114] First, concept embedding. Concept embedding is used to establish the association between test questions and multiple concepts. Specifically, each concept is first embedded, and then the comprehensive knowledge representation of the test question is obtained through average pooling, that is, in, is the overall embedding representation of the concepts involved in the nth test question, is the embedding representation of the concepts involved in the nth test question, d k is the dimension of the embedding, c n,i From q n The knowledge concept set τ(q n ).

[0115] Secondly, the test questions are embedded. The test questions are encoded in an embedded manner, and the embedding of the test question index involved in the nth learning unit is represented as

[0116] Finally, the knowledge of the test question is comprehensively represented. It is embedded by the index of the test question. and concept embedding Add them together to get k n =q n +c n ,in Represents the comprehensive knowledge representation of the test question.

[0117] (1-2) Temporal feature embedding

[0118] Each learning unit in the learning process occurs at a certain moment and is recorded as the timestamp when the learning event occurs. We use the timestamps in the learning trajectory to extract temporal features, including timestamp distance, time interval, and number of knowledge reviews.

[0119] First, we use timestamp distance to measure the true distance between learning units. By calculating the timestamp distance between the adjacent answers, we record the time interval Δs n ;

[0120] Secondly, the time interval refers to the time interval between the current knowledge answer and the last answer of the same knowledge point, which can be used to reflect the learner's familiarity with a certain knowledge point. The review time interval Δr n ;

[0121] Then, the number of knowledge reviews refers to the cumulative number of answers to the same knowledge point during the learning process. Similarly, to a certain extent, it can reflect the learner's familiarity with the knowledge point. The number of knowledge reviews is Δc n ; Thus, the timing feature T n =(Δs n ,Δr n ,Δc n ).

[0122] (1-3) Behavioral feature extraction

[0123] In order to achieve comprehensiveness and diversity in the characterization of learning behavior, we propose to use click streams and activity streams to describe the multiple learning behavior characteristics in the learning process.

[0124] First, extract the clickstream cs from the learning trajectory n . When answering a test question n When learning, learners may modify their answers multiple times due to uncertainty, resulting in several clicks to answer. We extract them in chronological order to form a click stream in, Indicates answering the test question q n For example, when answering a multiple-choice question, a learner may first choose A, then choose B, and finally choose C. Then the click stream of this learner is cs = {'A', 'B', 'C'}.

[0125] Second, extract the activity flow from the learning trajectory as n After answering each test question, learners will think and learn further by looking at the answer analysis, watching video lectures, and other activities. We extract the learners' behavior activities after answering the questions to form an activity flow, which is recorded as in Indicates that the student answers the test question q n The i-th activity after that includes viewing the explanation or watching the video lecture, i.e.

[0126] (2) First-order coupling analysis of cognitive states

[0127] like Figure 2 The "first-order coupling" layer in describes the first-order coupling analysis of cognitive states. For knowledge, time series features and behavior, time series features, neural network models and behavior feature extraction techniques are used to implement the first-order coupling analysis of "knowledge-time series" and "behavior-time series" cognitive states. Specifically, for knowledge-time series coupling, the knowledge and time series representations are input into the LSTM neural network, and forgetting reduction and learning enhancement functions are set to dynamically adjust; for behavior-time series coupling, according to the timestamp information in the click stream and activity stream, the behavior feature extraction technology is used to extract five click feature vectors and four activity features, which are encoded by addition and self-attention network with causal mask.

[0128] (2-1) Analysis of changes in cognitive states (knowledge-time coupling)

[0129] First, we transform the time series feature T n =(Δs n ,Δr n ,Δc n ) are embedded into the same dimension d k After that, the comprehensive representation of the time series features is obtained by adding Then embed it with knowledge features Concatenate to form the final input vector x n , the comprehensive representation of the learning unit is as follows:

[0130]

[0131] We will define the knowledge feature k n and the time series characteristics T n They are input into the LSTM neural network together to capture the dynamic relationship between knowledge points and learning time factors.

[0132] Subsequently, the basic update process of the traditional LSTM unit involves the forget gate, input gate, and output gate. Based on the original forget gate, we use the timestamp distance to adjust the forget degree, and the calculation method is as follows:

[0133] f n =σ(W f ·[h n-1 ,x n ]+b f )·exp(-α·Δs n )

[0134] in, is the adjusted forget gate output, is the hidden state of the previous learning unit. σ is the sigmoid activation function, exp(·) is the exponential function, and is the trainable weight matrix and bias term. Δs n is the timestamp distance between adjacent learning units, and α is an adjustment parameter used to control the impact of the timestamp distance on the forget gate. Through such a forget gate design, the longer the time interval between learning units, the greater the possibility of forgetting, and the deeper the model forgets previous knowledge.

[0135] Next, to enhance the learning process, we introduce the review interval Δr n and knowledge review times Δc n To adjust the output of the input gate, the calculation method is as follows:

[0136]

[0137] in, is the input gate output, and is a trainable parameter. γ is a set hyperparameter that controls the effect of the time interval and the number of reviews on the input gate. In the input gate, if the learner reviews frequently (i.e., Δc n larger), and recently reviewed (i.e., Δr n Smaller), then the learning effect is better and the weight of memory input is increased.

[0138] Finally, we use the adjusted forget gate f n and input gate i n , update the cell state of LSTM, and update the hidden state. The calculation method is as follows:

[0139]

[0140] in, is the cell state of LSTM, is a candidate cell state, and is a trainable parameter. We pass the output gate o n Compute the hidden state:

[0141] o n =σ(W o ·[h n-1 ,x n ]+b o ), h n =o n tanh(C n )

[0142] in, is the output of the output gate, In summary, we use the time series features to adjust the forget gate and input gate, obtain the hidden state of the learning unit based on the LSTM architecture, and preliminarily realize the knowledge-time series coupling.

[0143] (2-2) Analysis of click behavior changes (click flow behavior-time coupling)

[0144] First, according to the click stream According to its timestamp information, five behavioral change characteristics are extracted: ① Read whether the learner's answer at time t = 0 matches the standard answer, that is, whether the answer is correct or not, and record the "first answer correct or incorrect" ”; ② Using the same method, read the correctness information of the last modification at time t=i, and record the correctness of the final answer ”; ③ Calculate the difference between the final time and the initial time, and record the “reaction time t consumed” n ”, where t n =t i -t0; ④ Calculate the length of the entire click stream, record "click stream length l n =|cs n ⑤ Analyze the behavioral changes of learners in the clickstream. Learners may eventually modify their answers, which is recorded as "modification" type. They may also change their answers repeatedly and then return to the initial answer, which is recorded as "hesitation". This variable is called "clickstream type p n ”.

[0145] Secondly, these five embedding vectors are combined into one vector and mapped to a unified vector space in is the activity flow cs when the nth answer is given n The embedding representation of k is the embedding dimension, and emb represents the corresponding embedding layer function.

[0146] (2-3) Analysis of activity behavior changes (activity flow behavior-time sequence coupling)

[0147] First, according to the activity flow According to its timestamp information, four behavior change characteristics are extracted: ① Read the type of the current activity and record the "activity type" ",in, ② Calculate the time from the start to the end of the activity and record the "total activity time" "; ③ Get the activity is to view the analysis (ie ) times, record the number of times you have viewed the analysis "; ④ The acquisition activity is to watch the video (i.e. ) times, record the number of video operations ”.

[0148] Then, an embedding matrix is ​​used to represent the four activity flow features, where each row of the matrix corresponds to an activity, namely in is the activity flow when answering the nth question n The corresponding activities in the activity stream are combined to form an activity stream Among them n is the length of the active stream.

[0149] (3) Second-order coupling analysis of capability states

[0150] Figure 2 The "second-order coupling" layer in describes the second-order coupling analysis of the ability state. Based on the behavior analysis and encoding of clickstream and activity stream, combined with the knowledge state analysis, the second-order coupling analysis of the ability state of knowledge-behavior-time sequence coupling is realized. Specifically, based on the output result of the first-order coupling, the GLU gating function is set to obtain the student's initial cognitive state; then, the guessing gate and the proficiency gate are set through the clickstream encoding to weaken and enhance the current cognitive state respectively; then, the GLU-based learning gate is set based on the activity stream to enhance the current cognitive state; finally, the comprehensive ability state representation of the student is obtained through the time distance attention mechanism.

[0151] (3-1) Cognitive state update based on clickstream

[0152] First, the learner's knowledge state can be controlled by the gate function GLU(), which can be expressed as:

[0153]

[0154] in, is the initial cognitive state, is the output result of the first-order coupling of the cognitive state in the previous step, and Denotes the corresponding weight and bias parameters.

[0155] Second, based on the students’ modification operations in the clickstream, the degree of guessing in the learner’s answering process is calculated. If guessing exists, the level of knowledge mastery should be reduced by the error caused by guessing. By setting a “guessing gate” to adjust the cognitive state, the function is expressed as:

[0156] K′ n =K n -GLU g (CS n ·W g +b g ),

[0157] in, is the knowledge state after the guess gate, GLU g represents the gating function of the guess gate; represents the embedding representation of the click stream, and To guess the weight and bias parameters of the gate.

[0158] Third, based on the time attributes in the clickstream, the learner’s proficiency in knowledge is inferred. The learner’s reasonable time allocation for doing questions can enhance the student’s proficiency, and their knowledge mastery level can be added with the knowledge enhancement brought by proficiency. By setting a “proficiency gate” to enhance the knowledge state, the function is expressed as:

[0159] K″ n =K′ n +GLU p (CS n ·W p +b p ),

[0160] in, GLU is the knowledge status after passing the proficiency gate. p represents the gating function of the skilled gate; and are the weight and bias parameters of the proficiency gate.

[0161] (3-2) Enhancement of cognitive state based on activity flow

[0162] First, learners have a variety of learning activities, such as viewing analysis and watching videos. To uniformly represent learning activities, we use the average pooling method to obtain the average value of each activity embedding in the activity stream, thereby aggregating all activity information. The calculation method is as follows:

[0163]

[0164] in, is the aggregate representation of the activity flow of the nth answer, For activity streams n The length of , that is, the number of activities, Represents the embedding representation of the i-th activity in the n-th answer.

[0165] Secondly, the learning activities after answering the questions can enhance the learner's knowledge. That is, the learner's knowledge level needs to be enhanced after the learning activities, and the "learning gate" is set to simulate the degree of improvement of their high-level cognitive state. The function is expressed as:

[0166] K″′ n =K″ n +GLU l (AS n ),

[0167] in, represents the knowledge state after being enhanced by the learning gate, GLU is the knowledge state before entering the learning gate. l It is a learning gate function used to enhance the influence of activity flow on knowledge state.

[0168] (3-3) Potential Ability Reasoning Based on Temporal Distance Attention Mechanism

[0169] Based on (2-1) the current cognitive state h n And (3-2) get the knowledge state change K″′ caused by the behavior n , we further infer the potential comprehensive ability state of students θ through the temporal distance attention mechanism n+1 In order to introduce the influence of time factors, we set the query vector to the question representation k at time n+1. n+1 , and use an exponential decay function to weight the historical information at different times, and introduce the time interval between the n+1 answer unit and each historical unit in the attention calculation.

[0170] First, we define the query, key, and value vectors. The query vector Q n+1 Use the test question embedding representation k at time n+1 n+1 Generate, and the key and value vectors come from the current and all historical learning units. The calculation is as follows:

[0171] Q n+1 =W Q ·k n+1

[0172] K 1:n =[W K h1,W K h2,...,W K ·hn ]

[0173] V 1:n =[W V h1,W V ·h2,...,W V ·h n ]

[0174] in, represents the test embedding at the n+1th moment, is the mapping matrix used to map questions and hidden states to query, key, and value spaces. is the query vector, is the key vector matrix, is a matrix of value vectors.

[0175] Then, through the query vector Q n+1 and the key vector K 1:n Calculate the similarity and combine it with the time decay function exp(-βΔt n+1,i ) adjusts the attention weight. The calculation formula of attention weight is:

[0176]

[0177] Among them, α n+1,i represents the attention weight of the n+1th unit to the i-th historical unit, Δt n+1,i The time interval between the n+1th unit and the i-th historical unit, β is a hyperparameter of time decay, which is used to control the impact of the time interval on the attention weight.

[0178] Finally, we use the time-weighted attention weight α n+1,i With the value vector V 1:n Perform weighted summation to generate the potential comprehensive ability state θ of the n+1th learning unit n+1 :

[0179]

[0180] in, Represents the potential comprehensive ability state of the inferred n+1th learning unit.

[0181] (4) Dual-task prediction of behavior and response

[0182] like Figure 2In the performance prediction part, based on the acquired comprehensive ability status, the final performance prediction is enhanced through the dual prediction tasks of behavior and reaction. Specifically, three behavior prediction tasks, namely reaction time, click stream length and activity stream length, are set first, the ability status is converted to the corresponding behavior space, and then the behavior prediction task is completed through the multi-layer perceptron; finally, the behavior prediction is aggregated and the prediction of the final reaction result in the future is realized through the multi-layer perceptron.

[0183] (4-1) Identification and prediction of multi-faceted learning behaviors

[0184] According to the learner's ability status, predict the learner's behavioral performance, including reaction time, click stream length and activity stream length.

[0185] The reaction time prediction formula is expressed as: The loss function is denoted by L t , where t n+1 is the actual reaction time at moment n+1.

[0186] The click stream length prediction formula is expressed as: The loss function is denoted by L l , Among them l n+1 It is the actual click length at time n+1.

[0187] The activity flow length prediction formula is expressed as: The loss function is denoted by L a , where a n+1 It is the actual activity length at time n+1.

[0188] in is the predicted reaction time, is the predicted clickstream length, is the predicted activity flow length, θ n+1 is the potential ability state of the n+1th. In each behavior prediction, a feedforward network layer FNN is set to convert the ability state into the corresponding behavior space, and then the prediction is completed through the respective multi-layer perceptron MLP.

[0189] (4-2) Behavior analysis enhanced response outcome prediction

[0190] First, we obtain the aggregated representation of the behavior by aggregating the three vectors used to predict the multivariate learning behavior in the previous step and the current potential ability state θ n+1 , forming a comprehensive behavioral representation:

[0191]

[0192] in, is the behavior representation of the n+1th unit, combining the average information of each behavior feature. FNN(·) represents the n+1 Feed-forward network layers that map to their respective behavior spaces.

[0193] Secondly, based on the aggregated behavior representation R n+1 , set up a multilayer perceptron to predict the learner's final response result, and the calculation process is expressed as:

[0194]

[0195] in, is the predicted response result, MLP is a multi-layer perceptron. The loss function is denoted by L r , where r n+1 This is the true reaction at time n+1.

[0196] The final loss function is obtained by weighted summation, that is, L = L r +λ1L t +λ2L l +λ3L a , where λ1, λ2, and λ3 are hyperparameter weights used to adjust the impact of the latter two loss functions.

[0197] The present invention proposes a dynamic cognitive diagnosis method with three-element second-order coupling for the learning process, which divides the learning units in the learning process into three-element groups of "knowledge, timing, and behavior", and introduces a variety of deep learning technologies to realize the second-order modeling of cognition and ability. The comparison between the static cognitive diagnosis method and the dynamic cognitive diagnosis method is shown in Table 1. As shown in Table 1, the static cognitive diagnosis method only uses the knowledge features in the educational context, does not consider the timing and behavior features in the real educational context, is applicable to a single learning context, and is essentially a one-element coupling of knowledge; the dynamic cognitive diagnosis method considers the knowledge features and the sequence features of the learning process, is applicable to multiple learning contexts, ignores the time interval attributes and behavior process features in the learning process, and is essentially a weak coupling of knowledge and timing features, such as Figure 4 As shown in (a) of , the binary coupling of weakened procedural behavior has limitations, which will eventually lead to the problem of incomplete learning characterization; the three-element second-order coupling cognitive diagnosis method proposed in this invention considers the second-order coupling of three features, such as Figure 4 As shown in (b) in the figure, it is possible to scientifically and comprehensively model the learner’s cognitive process and achieve accurate prediction of the learner’s future performance.

[0198] Table 1 Comparison of coupling mechanisms between this example method and traditional cognitive diagnosis methods

[0199]

[0200] The contents not described in detail in this specification belong to the prior art known to those skilled in the art.

[0201] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A knowledge-sequence-behavior three-element second-order coupling dynamic cognitive diagnosis method, characterized by: The steps include: Step 1: Obtain the learning unit data of students when they take the assessment, including the three elements of knowledge, behavior, and time sequence. According to the chronological order, several learning units are combined into learning process sequence data to construct a time series data set. Based on the constructed time series data set, obtain the ternary standardized representation of the learning unit; Among them, knowledge features include the test questions answered by students at a certain moment and the concepts related to the test questions; behavioral features include the click stream formed by students repeatedly modifying their answers during the answering process, and the activity stream formed by students viewing analysis and video resources after answering; temporal features refer to the time interval between each time unit, the time and number of repetitions between learning the same concept; Step 2: For the represented knowledge and behavior characteristics, conduct first-order coupling analysis with the time series characteristics to obtain the cognitive state, that is, conduct first-order coupling analysis on knowledge-time series to obtain the evolution of the learner's knowledge state, and conduct first-order coupling analysis on behavior-time series to analyze the learner's behavior changes; Specifically, for knowledge-time series coupling, knowledge and time series representations are input into the LSTM neural network, and forgetting reduction and learning enhancement functions are set to dynamically adjust; for behavior-time series coupling, according to the timestamp information in the click stream and activity stream, the behavior feature extraction technology is used to extract the click feature vector and activity feature, and they are encoded by addition and self-attention network with causal mask respectively; Step 3: perform second-order coupling of knowledge state and behavior state at the time granularity to obtain the learner's comprehensive ability state; Step 4: Based on the acquired comprehensive ability state representation, the final performance prediction is enhanced through the dual prediction task of behavior and reaction; Specifically, it includes setting three behavior prediction tasks: reaction time, click stream length, and activity stream length, converting the capability state to the corresponding behavior space, and then completing the behavior prediction task through a multi-layer perceptron; aggregate learning behavior prediction, and realizing the prediction of the final reaction result at future moments through a multi-layer perceptron.

2. The knowledge-sequence-behavior three-element second-order coupled dynamic cognitive diagnosis method as claimed in claim 1, characterized in that: The three-element standardized representation of the learning unit specifically includes: (1-1) Knowledge Feature Embedding All test questions and concepts together constitute the knowledge space, which is represented by a triple (Q, C, τ), where Q is a non-empty set of test questions, C is a non-empty set of concepts, and τ is a mapping from Q to concept C. For any test question q n ∈Q, τ(q) represents the set of concepts required to solve question q. One question can be associated with multiple concepts. Specifically, it includes three steps: concept embedding, question embedding, and comprehensive knowledge representation. First, concept embedding: concept embedding is used to establish the association between test questions and multiple concepts. Specifically, each concept is first embedded, and then the comprehensive knowledge representation of the test question is obtained through average pooling, that is, in, is the overall embedding representation of the concepts involved in the nth test question, is the embedding representation of the concepts involved in the nth test question, d k is the dimension of the embedding, c n,i From q n The knowledge concept set τ(q n ); Secondly, question embedding: the questions are encoded in an embedded manner, and the embedding of the question index involved in the nth learning unit is represented as Finally, the comprehensive knowledge representation of the test question: embedded by the index of the test question and concept embedding Add them together to get k n =q n +c n ,in A comprehensive representation of the knowledge of the test questions; (1-2) Temporal feature embedding Each learning unit in the learning process occurs at a certain moment and is recorded as the timestamp when the learning event occurs. The timestamps in the learning trajectory are used to extract temporal features, including timestamp distance, time interval, and number of knowledge reviews. First, the timestamp distance is used to measure the real distance between learning units. By calculating the timestamp distance between the adjacent answers, the time interval Δs is recorded. n ; Secondly, the time interval refers to the time between the current knowledge answer and the last answer of the same knowledge point, which is recorded as the review time interval Δr n ; Then, the number of knowledge reviews refers to the cumulative number of answers to the same knowledge point during the learning process, and the number of knowledge reviews is Δc n ; Thus, the timing feature T n =(Δs n ,Δr n ,Δc n ); (1-3) Behavioral feature extraction Use clickstream and activity stream to describe the multi-faceted learning behavior characteristics in the learning process; First, extract the clickstream cs from the learning trajectory n ;When answering a test question n When learning, learners may modify their answers many times due to uncertainty, thus forming several click answers, which are extracted in chronological order to form a click stream. in, Indicates answering the test question q n The result of the i-th click answer; Second, extract the activity flow from the learning trajectory as n ; After completing each test, learners will think and learn further by looking at the answer analysis and watching the video lecture. The learners' behavior after answering the questions is extracted to form an activity flow, recorded as in Indicates that the student answers the test question q n The i-th activity after that includes viewing the explanation or watching the video lecture, i.e.

3. The knowledge-sequence-behavior three-element second-order coupled dynamic cognitive diagnosis method as claimed in claim 2, characterized in that: The specific implementation process of knowledge-timing coupling is as follows: First, the time series feature T n =(Δs n ,Δr n ,Δc n ) are embedded into the same dimension d k After that, the comprehensive representation of the time series features is obtained by adding Then embed it with knowledge features Concatenate to form the final input vector x n , the comprehensive representation of the learning unit is as follows: The knowledge feature k n and the timing characteristics T n They are input into the LSTM neural network together to capture the dynamic relationship between knowledge points and learning time factors; Then, based on the original forget gate, the timestamp distance is used to adjust the forgetting degree. The calculation method is as follows: f n =σ(W f ·[h n-1 ,x n ]+b f )·exp(-α·Δs n ) in, is the adjusted forget gate output, is the hidden state of the previous learning unit; σ is the sigmoid activation function, exp(·) is the exponential function, and is the trainable weight matrix and bias term; Δs n is the timestamp distance between adjacent learning units, and α is an adjustment parameter used to control the impact of timestamp distance on the forget gate; Next, in order to enhance the learning process, the review interval Δr is introduced n and knowledge review times Δc n To adjust the output of the input gate, the calculation method is as follows: in, is the input gate output, and is a trainable parameter; γ is a set hyperparameter used to control the impact of time interval and number of reviews on the input gate; Finally, based on the adjusted forget gate f n and input gate i n , update the cell state of LSTM, and update the hidden state. The calculation method is as follows: in, is the cell state of LSTM, is a candidate cell state, and is a trainable parameter; through the output gate o n Compute the hidden state: o n =σ(W o ·[h n-1 ,x n ]+b o ),h n =o n ·tanh(C n ) in, is the output of the output gate, Represents the hidden state of the current learning unit.

4. The knowledge-sequence-behavior three-element second-order coupled dynamic cognitive diagnosis method as claimed in claim 2, characterized in that: Behavior-timing coupling includes clickstream behavior-timing coupling and activitystream behavior-timing coupling; The specific implementation process of click flow behavior-sequence coupling is as follows: First, according to the click stream According to its timestamp information, five behavioral change characteristics are extracted: ① Read whether the learner's answer at time t=0 matches the standard answer, that is, whether the answer is correct or not, and record "the correct or incorrect first answer" ”; ② Using the same method, read the correctness information of the last modification at time t=i, and record the correctness of the final answer ”; ③ Calculate the difference between the final time and the initial time, and record the reaction time t n ”, where t n =t i -t0; ④ Calculate the length of the entire click stream, record "click stream length l n =|cs n ⑤ Analyze the behavioral changes of learners in the clickstream. The learner may eventually modify the answer, which is recorded as "modification" type, or may change back to the initial answer after repeated changes, which is recorded as "hesitation". This variable is called "clickstream type p n ”; Secondly, these five embedding vectors are combined into one vector and mapped to a unified vector space in is the activity flow cs when the nth answer is given n The embedding representation of k is the embedding dimension, emb represents the corresponding embedding layer function; The specific implementation process of activity flow behavior-timing coupling is as follows: First, according to the activity flow According to its timestamp information, four behavior change characteristics are extracted: ① Read the type of the current activity and record the "activity type ",in, ② Calculate the time from the start to the end of the activity and record the total activity time "; ③ Get the activity is to view the analysis The number of times, record "View the number of parsing "; ④ The activity of obtaining is to watch the video The number of times the video is operated is recorded as " ”; Then, an embedding matrix is ​​used to represent the four activity flow features, where each row of the matrix corresponds to an activity, namely in is the activity flow when answering the nth question n The embedding representation of the i-th activity of is; the corresponding activities in the activity stream are combined to form an activity stream Among them n is the length of the active stream.

5. The knowledge-sequence-behavior three-element second-order coupled dynamic cognitive diagnosis method as claimed in claim 1, characterized in that: The specific implementation of step 3 includes: (3-1) Based on the output results of the first-order coupling, the GLU gating function is set to obtain the student's initial cognitive state, and the guessing gate and the proficiency gate are set through click stream coding to weaken and enhance the current cognitive state respectively; (3-2) Setting GLU-based learning gates based on activity flow to enhance the current cognitive state; (3-3) Through the time-distance attention mechanism, the comprehensive ability status representation of students is obtained.

6. The knowledge-sequence-behavior three-element second-order coupled dynamic cognitive diagnosis method as claimed in claim 5, characterized in that: The specific implementation of (3-1) is as follows: First, the learner's knowledge state is controlled by the gate function GLU(), which can be expressed as: in, is the initial cognitive state, is the output result of the first-order coupling of the cognitive state in the previous step, and Represents the corresponding weight and bias parameters; Second, based on the students’ modification operations in the clickstream, the degree of guessing during the learner’s answering process is calculated; the cognitive state is adjusted by setting a guessing gate, and the function is expressed as: K′ n =K n -GLU g (CS n ·W g +b g ), in, is the state of knowledge after passing through the guessing gate; represents the embedding representation of the click stream, and To guess the weight and bias parameters of the gate; Third, based on the time attributes in the clickstream, the learner’s proficiency in knowledge is inferred; the knowledge state is enhanced by setting a proficiency gate, and the function is expressed as: K″ n =K′ n +GLU p (CS n ·W p +b p ), in, It is the state of knowledge after passing the proficiency gate; and are the weight and bias parameters of the proficiency gate.

7. The knowledge-sequence-behavior three-element second-order coupled dynamic cognitive diagnosis method according to claim 6, characterized in that: The specific implementation of (3-2) is as follows: In order to unify the representation learning activities, the average pooling method is used to obtain the average value of each activity embedding in the activity stream, thereby aggregating all activity information. The calculation method is as follows: in, is the aggregate representation of the activity flow of the nth answer, For activity streams n The length of , that is, the number of activities, represents the embedding representation of the i-th activity in the n-th answer; Secondly, the learning activities after answering the questions can enhance the learner's knowledge mastery, and the learning gate is set to simulate the improvement of their high-order cognitive state; the function is expressed as: K″′ n =K″ n +GLU l (AS n ), in, represents the knowledge state after being enhanced by the learning gate, GLU is the knowledge state before entering the learning gate. l It is a learning gate function used to enhance the influence of activity flow on knowledge state.

8. The knowledge-sequence-behavior three-element second-order coupled dynamic cognitive diagnosis method according to claim 7, characterized in that: The specific implementation of (3-3) is as follows: Based on the current cognitive state h n And (3-2) get the knowledge state change K″′ caused by the behavior n , and further infer the potential comprehensive ability state of students θ through the time distance attention mechanism n+1 ; In order to introduce the influence of time factors, the query vector is set to the question representation k at time n+1 n+1 , and use an exponential decay function to weight the historical information at different times, and introduce the time interval between the n+1 answer unit and each historical unit in the attention calculation; First, define the query, key, and value vectors: n+1 Use the test question embedding representation k at time n+1 n+1 Generate, and the key and value vectors come from the current and all historical learning units; the calculation is as follows: Q n+1 =W Q ·k n+1 K 1:n =[W K ·h1,W K ·h2,...,W K ·h n ] V 1:n =[W V ·h1,W V ·h2,...,W V ·h n ] in, represents the test embedding at the n+1th moment, is the mapping matrix, used to map questions and hidden states to query, key and value spaces; is the query vector, is the key vector matrix, is a matrix of value vectors; Then, through the query vector Q n+1 and the key vector K 1:n Calculate the similarity and combine it with the time decay function exp(-βΔt n+1,i ) adjusts the attention weight; the calculation formula of attention weight is: Among them, α n+1,i represents the attention weight of the n+1th unit to the i-th historical unit, Δt n+1,i The time interval between the n+1th unit and the i-th historical unit. β is a hyperparameter of time decay, which is used to control the effect of the time interval on the attention weight. Finally, the time-weighted attention weight α is used n+1,i With the value vector V 1:n Perform weighted summation to generate the potential comprehensive ability state θ of the n+1th learning unit n+1 : in, Represents the potential comprehensive ability state of the inferred n+1th learning unit.

9. The knowledge-sequence-behavior three-element second-order coupled dynamic cognitive diagnosis method as claimed in claim 1, characterized in that: In step 4, the reaction time prediction formula is expressed as: The loss function is denoted by L t , where t n+1 is the actual reaction time at time n+1; The click stream length prediction formula is expressed as: The loss function is denoted by L l , Among them l n+1 is the actual click length at time n+1; The activity flow length prediction formula is expressed as: The loss function is denoted by L a , where a n+1 is the actual activity length at time n+1; in is the predicted reaction time, is the predicted clickstream length, is the predicted activity flow length, θ n+1 It is the potential ability state of the n+1th one. In each behavior prediction, a feedforward network layer FNN is set to convert the ability state into the corresponding behavior space, and then the prediction is completed through the respective multi-layer perceptron MLP.

10. The knowledge-sequence-behavior three-element second-order coupled dynamic cognitive diagnosis method according to claim 9, characterized in that: The prediction of the final reaction results in step 4 includes: First, we obtain the behavior aggregation representation by aggregating the three vectors used to predict the multivariate learning behavior and the current potential ability state θ n+1 , forming a comprehensive behavioral representation: in, is the behavior representation of the n+1th unit, combining the average information of each behavior feature; FNN(·) represents the n+1 Feed-forward network layers that map to their respective behavior spaces; Secondly, based on the aggregated behavior representation R n+1 , set up a multilayer perceptron to predict the learner's final response result, and the calculation process is expressed as: in, is the predicted reaction result, MLP is a multi-layer perceptron; the loss function is recorded as L r , where r n+1 is the true response at time n+1; The final loss function is obtained by weighted summation, that is, L = L r +λ1L t +λ2L l +λ3L a , where λ1, λ2, and λ3 are hyperparameter weights used to adjust the impact of the latter two loss functions.

Images