Dynamic cognitive diagnosis method based on multivariate learning behavior causal inference

By constructing a learning behavior map and using a causal attention mechanism to distinguish between key behaviors and inefficient behaviors, this study addresses the problem of neglecting the influence of diverse learning behaviors in existing cognitive diagnostic models, thereby achieving accurate assessment of learners' cognitive levels and improving the interpretability of diagnostic results.

CN118940795AActive Publication Date: 2024-11-12WUHAN TEXTILE UNIV

Patent Information

Application Number
CN202410961023.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-17
Publication Date
2024-11-12
Estimated Expiration
2044-07-17

AI Technical Summary

Technical Problem

Existing cognitive diagnostic models neglect the impact of diverse learning behaviors on learners' cognitive states, leading to biased diagnostic results and poor interpretability, and failing to accurately assess learners' overall cognitive level.

Method used

A learning behavior graph is constructed, key behaviors and inefficient behaviors are divided through a causal attention mechanism, and graph neural networks are used for encoding. The causal relationship between multiple behaviors and knowledge states is explored, and a multi-task mechanism is set up to predict the target, thereby realizing dynamic cognitive diagnosis.

Benefits of technology

It improves the accuracy and interpretability of learners' cognitive level diagnosis, and enhances the accuracy and interpretability of diagnostic results through a multi-faceted causal inference method of learning behavior.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118940795A_ABST
    Figure CN118940795A_ABST
Patent Text Reader

Abstract

The invention provides a dynamic cognitive diagnosis method based on multivariate learning behavior causal inference, which is characterized in that learning process behaviors are modeled by means of click streams and activity streams, and a learning behavior graph is constructed through multivariate behavior characteristics. Meanwhile, causal association between learning behaviors and cognition is considered, weight differences are learned through a causal attention mechanism, and automatic division of key behavior sub-graphs and low-efficiency behavior sub-graphs is achieved. Then, learning behaviors are regarded as variables in causal inference, inefficient behaviors are intervened, a combination of key behaviors and inefficient behaviors is established, and an intervention graph is obtained; based on the obtained key behavior sub-graph, the low-efficiency behavior sub-graph and the intervention graph representation, different prediction targets are set through a multi-task mechanism, so that causal association between the learning behavior and implicit cognition is mined, and the interpretability of dynamic diagnosis of the cognition level of the learner is enhanced.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the fields of educational big data mining, graph neural networks and student behavior modeling, and in particular to a dynamic cognitive diagnosis method based on causal inference of multiple learning behaviors. Background Art

[0002] Learning behavior is the process action or result produced by learners when they participate in learning. According to the stage at which learning behavior occurs, it can be divided into result behavior and process behavior.

[0003] Dynamic cognitive diagnosis for outcome-oriented behavior infers learners' potential cognitive level by mining their answer results, where answer results usually refer to binary results of correct or incorrect. Classic cognitive diagnosis models, such as IRT, DINA, and BKT, infer learners' potential cognitive level based on the binary results of their responses. However, in multiple-choice questions, answer results can be expanded to multiple answer options, not just the common binary results. These models all take outcome-oriented behavior as the research target, combine multiple cognitive factors, and jointly mine potential cognitive levels. In fact, most of the current cognitive diagnosis models are based on outcome-oriented learning behavior, ignoring the impact of diversified process-oriented learning behavior on cognitive state, resulting in the separation of learners' cognitive level and learning process. Therefore, further exploration is needed to diagnose learners' true cognitive state by mining process-oriented behavior.

[0004] Dynamic cognitive diagnosis for process behavior, based on the dynamic cognitive diagnosis model for outcome behavior, incorporates learning behavior information in the learning process to infer the learner's potential cognitive level. Since process learning behavior contains rich cognitive state information, cognitive diagnosis integrating process behavior helps to improve the accuracy of diagnosis. Although current studies have introduced process behavior from different perspectives and made certain attempts, they use learning behavior to enrich the representation of answer interaction and do not explore the intrinsic relationship between behavior and knowledge. At the same time, overall, there are relatively few studies on process behavior in current cognitive diagnosis, and existing related studies only use scattered behavioral characteristics, which cannot fully characterize the learning process. Therefore, given the importance of process learning behavior to learning outcomes, it is an important research direction and development trend in this field to explore the learner's potential cognitive state from learning behavior, explore the relationship between behavior and knowledge, and study dynamic cognitive diagnosis modeling for process behavior.

[0005] In the field of massive open online courses (MOOCs), there have been many studies on diversified learning behaviors in the learning process. In particular, some scholars have introduced "click streams" to describe the temporal behavior sequence in the learning process. Click streams are the tracks of learners' continuous access in the learning system, and are often used to predict the dropout rate of learners in online learning platforms. After answering, learners can also participate in some other activities, such as viewing analysis, video learning, and consolidation exercises, which can form "activity streams". In fact, "click streams" and "activity streams" can comprehensively record the sequence of learning behaviors, which is helpful to use neural network technology to simulate the real learning process of learners. However, the application of process behaviors based on click stream and activity stream modeling in the field of cognitive diagnosis has yet to be studied.

[0006] Existing dynamic cognitive diagnosis ignores the impact of diversified behaviors on learners' cognition during long-term learning, which makes it impossible to accurately evaluate learners' comprehensive cognitive level, thus affecting the interpretability of the model. However, there is a complex causal relationship between diversified learning behaviors and learning outcomes, and different behavioral factors have different effects on learners' cognitive states. Mining the key behavioral factors in the learning process will help improve the accuracy and interpretability of diagnostic results. Summary of the invention

[0007] Since the behavioral characteristics in the learning process are complex and diverse, such as learners repeatedly revising their answers during the answering process, a "click stream" will be formed; after finishing the questions, learners will reflect through text analysis or targeted video recommendations to form an "activity stream". There is a causal relationship between learners' learning outcomes and various behavioral elements, and the degree of correlation varies. Therefore, based on the diversified behaviors formed in the learning process, a dynamic cognitive diagnosis method based on causal inference of multiple learning behaviors is proposed. The answering behavior graph is constructed and encoded using a graph neural network. The relationship between multiple behaviors and knowledge states is mined through causal attention, and finally a dynamic diagnosis of learners' cognitive levels is completed.

[0008] The purpose of the present invention is to provide a dynamic cognitive diagnosis method based on causal inference of multiple learning behaviors in order to address the problem that existing cognitive diagnosis ignores the causal analysis of diversified learning behaviors, resulting in diagnostic bias and difficulty in causal interpretation. The method uses click streams and activity streams to model the learning process behavior, and constructs a learning behavior graph through multiple behavioral features. At the same time, the causal relationship between learning behavior and cognition is considered, and the weight difference is learned through the causal attention mechanism to achieve automatic division of key behavior subgraphs and inefficient behavior subgraphs. Subsequently, the learning behavior is regarded as a variable in causal inference, and the inefficient behavior is intervened, a combination of key behavior and inefficient behavior is established, and a generated graph is obtained. Based on the acquired key behavior subgraph, inefficient behavior subgraph and generated graph representation, different prediction targets are set through a multi-task mechanism to explore the causal relationship between learning behavior and implicit cognition, and enhance the interpretability of the dynamic diagnosis of learners' cognitive level.

[0009] The object of the present invention is achieved through the following technical measures: a dynamic cognitive diagnosis method based on causal inference of multivariate learning behavior, comprising the following steps:

[0010] Step 1: Obtain knowledge, behavior, and time series data during the learning process to construct a student answer data set; knowledge data includes the test questions answered by students at a certain moment and the knowledge related to the test questions; behavior data includes 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 the answering process; time series data includes the timestamp information of the current behavior;

[0011] Step 2: Construct a learning behavior graph for the click streams and activity streams of students in the dataset, and map different behavior features to a unified behavior space. Specifically, the graph includes: constructing a click stream subgraph by counting click features using timestamps, constructing an activity stream subgraph by counting activity features using timestamps, and connecting the click stream subgraph and the activity stream subgraph to form a unified learning behavior graph.

[0012] Step 3: divide the behaviors in the learning behavior graph into key behaviors and inefficient behaviors, forming two subgraphs, namely, the key behavior subgraph and the inefficient behavior subgraph, and use the Do operator to generate an intervention graph to explore the causal relationship between the key behaviors and inefficient behaviors and the knowledge state;

[0013] Step 4: Based on the acquired key behavior subgraph, inefficient behavior subgraph, and intervention graph representations, different prediction targets are set through a multi-task mechanism to explore the causal relationship between learning behavior and implicit cognition, and the cross loss entropy is calculated between the key behavior, inefficient behavior, and the actual answer, while the cross loss entropy is calculated between the inefficient behavior and the fixed value. The final loss function is calculated by the weighted sum of the three to achieve key behavior prediction, intervention graph prediction, and inefficient behavior prediction.

[0014] Furthermore, in step 2, click on the flow subgraph The construction method is as follows;

[0015] First, from the click stream middle, represents the click behavior on question n at time i, N c Indicates the total number of click streams; counts the behavioral characteristics of each option of the learner in turn through the timestamp, and extracts the click options The reaction time t of the option n and option repetition times p n Three behavioral characteristics;

[0016] Next, click on the option The timestamp information of the node is set, and the initial node, intermediate node and end node are set. The node is recorded as The initial embedding of nodes is randomly initialized;

[0017] Finally, connect the start and end nodes with the click option nodes in the middle process, connect the reaction time and the number of options with the corresponding click options, complete the edge relationship between the nodes, and construct the edge in the click flow subgraph. By setting nodes and edges, a click flow subgraph is formed To describe the learner's answering process.

[0018] Furthermore, the activity flow subgraph in step 2 The construction method is as follows;

[0019] After the learner has completed the answer, the activity flow consisting of the learner viewing the answer analysis, watching the video, and recommending the practice answer behavior is constructed in chronological order. represents the activity behavior on question n at time i, N a Indicates the total number of active flows;

[0020] First, from the activity stream In the process, the behavioral characteristics of each activity in which the learner participates are counted in sequence by timestamp, and the activity type is extracted. Event Time Number of activities These activities have behavioral characteristics;

[0021] Secondly, according to the three activity behavior characteristics, obtain the activity node The initial embedding of nodes is randomly initialized;

[0022] Thirdly, the edge setting between nodes, according to the information before and after the timestamp, establishes edge connections between activity type nodes, activity time nodes, activity number nodes and corresponding activity type nodes respectively, and constructs the edge in the activity flow subgraph.

[0023] Furthermore, the specific implementation of step 3 is as follows:

[0024] (3-1) Obtaining node representations in the learning behavior graph;

[0025] (3-2) Division of critical behavior subgraphs and inefficient behavior subgraphs

[0026] The causal attention method is used to automatically learn associations and classify behaviors into key learning behaviors that have a significant impact on the knowledge state and inefficient learning behaviors that have a weak impact. Specifically,

[0027] First, two multi-layer perceptrons are set to obtain attention scores from the node level and edge level respectively, namely:

[0028] First, obtain the node-level attention score: For each node in the learning behavior graph, set up a multi-layer perceptron, the node-level attention score of the node The calculation formula is as follows:

[0029]

[0030] Among them, MLP node It is a multilayer perceptron for nodes. The function exp() represents the natural exponential function. h i represents the i-th node representation in the behavior graph, H n Represents all node representations of the behavior graph;

[0031] Second, get the edge-level attention score: For each edge in the learning behavior graph, set up a multi-layer perceptron, and the edge-level attention score of edge e The calculation formula is as follows:

[0032]

[0033] Among them, MLP edge is a multi-layer perceptron for edges, represents the representation of the edge e formed by the i-th and j-th nodes in the behavior graph, Indicates splicing, represents the edges in the clickstream subgraph, represents the edges in the activity flow subgraph;

[0034] Thus, the attention scores of each node and each edge are aggregated to form a node-level attention score for the entire behavior graph. and edge-level attention score

[0035] For inefficient learning behaviors, the scores of the corresponding key learning behaviors are subtracted from 1 to obtain the node attention scores and edge attention scores of the inefficient learning behaviors, that is, The node attention score and edge attention score of the entire behavior graph are expressed as

[0036] (3-3) Use Do operator intervention to obtain intervention graphs to explore the causal relationship between key learning behaviors, inefficient learning behaviors and knowledge states;

[0037] First, the key behavior subgraphs are extracted from the learning behavior graph using the causal attention scores. and the inefficient behavior subgraph Right now ⊙ represents the Hadamard product of the matrix, X n is the representation matrix of the learning behavior graph nodes and the adjacency matrix A of the learning behavior graph n ;

[0038] Then, pooling is used to obtain the representations of the two subgraphs, namely the key behavior representations and inefficient behavior representation And use the queue structure to build a repository, denoted as MB, to record the representations of different inefficient behavior subgraphs of multiple learners;

[0039] Thirdly, in order to explore the behavioral factors that affect learning outcomes, learning behavior is regarded as a variable in causal inference, and inefficient behavior is intervened, that is, the key behavioral variables are fixed and the inefficient behavioral variables are changed. The intervention process formula is:

[0040]

[0041] in, is the graph generated by the intervention operation, Inefficient behavior for fetching from the repository;

[0042] Finally, for each critical behavior, multiple different inefficient behaviors are randomly obtained from the repository Combined with them respectively, several intervention graphs are generated, and the set of intervention graphs is recorded as M is the number of intervention graphs.

[0043] Furthermore, the specific implementation of step (3-1) is as follows:

[0044] First, the graph attention network GAT is used to complete the behavior graph encoding and obtain the representation matrix X of the behavior graph nodes from the learning behavior graph. n and the adjacency matrix A of the behavior graph n ;

[0045] Secondly, according to the correlation between neighbor nodes and themselves, the graph convolution method is used to aggregate the information of the behavior graph nodes to obtain the representation H of all nodes in the entire graph. n , let Hn=GConv(An,Xn⊙Mn), where ⊙ represents the dot product, GConv represents the graph convolution operation, and M n is the attention weight matrix obtained through the self-attention mechanism, that is Among them, the weight W q and W k For the linear transformation in the attention mechanism, d k The embedding dimension of the representation;

[0046] Finally, a linear layer is used to complete the node update and obtain the final node representation.

[0047] Furthermore, the specific implementation method of realizing the key behavior prediction in step 4 is as follows;

[0048] For key behaviors, there is a causal relationship between key learning behaviors and correct or incorrect responses. represents the key behavior, where qn+1 is the question to be answered, σ is the sigmoid activation function, and MLP k is a multi-layer perceptron, is the learner's cognitive state, and the calculation formula is as follows:

[0049]

[0050] Among them, MHSA stands for Multi-Head Self-Attention Network, Represents knowledge characteristics, Represents the time series characteristics and represents the key behavioral characteristics, obtained through step 3;

[0051] Will Calculate the cross loss entropy with the actual answer result, and the loss function is recorded as where r n+1 Is the test question waiting to be answered? n+1 The actual reaction result, r n+1 ∈{0, 1}, 0 indicates an incorrect answer, 1 indicates a correct answer, and N indicates the total number of answers given by the learner.

[0052] Furthermore, the specific implementation method of realizing the inefficient behavior prediction in step 4 is as follows;

[0053] For inefficient behavior, inefficient learning behavior has no causal relationship with answer performance, so it will pass through the linear layer, that is, Get the prediction results The cross loss entropy is calculated by comparing it with a fixed value of 0.5. The fixed value of 0.5 is the middle value of the two classifications, which means that inefficient behavior is unlikely to affect the answer result. The loss function is recorded as MLP l It is a multi-layer perceptron.

[0054] Furthermore, the specific implementation method of implementing the intervention graph prediction in step 4 is as follows;

[0055] Finally, the intervention map It is generated by combining intervention operations and key behaviors for inefficient behaviors. Therefore, for each intervention graph, it is passed through a linear layer, i.e. q n+1 It is a test question to be answered, MLP i It is a multi-layer perceptron, and the prediction result is obtained With the collection The intervention graph corresponds to , and then the cross loss entropy is calculated with the corresponding real answer result, that is

[0056]

[0057] Furthermore, the final loss function is obtained by weighted summation, that is, where λ 1 and λ 2 is a hyperparameter weight used to adjust the impact of the last two loss functions.

[0058] Furthermore, the knowledge feature represents the knowledge information of the test questions answered by the learners, using the test question number or the knowledge points associated with the test question, and is embedded using Embedding;

[0059] The time series feature calculates the time interval between each student's answer and the last answer based on the timestamp information in the time series data of the learning process, and embeds it using Embedding.

[0060] Compared with the prior art, the beneficial effect of the present invention is that the existing dynamic cognitive diagnosis ignores the impact of diversified behaviors on learners' cognition during long-term learning, resulting in the inability to accurately assess the learners' comprehensive cognitive level, which in turn affects the interpretability of the model. The method of the present invention considers the complex causal relationship between diversified learning behaviors and learning outcomes, and the differences in the impact of different behavioral elements on learners' cognitive states, thereby exploring the key behavioral elements in the learning process, which helps to improve the accuracy and interpretability of the diagnostic results. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 It is a flowchart of the method of the present invention.

[0062] Figure 2Schematic diagram of constructing a learning behavior graph in an embodiment of the present invention. DETAILED DESCRIPTION

[0063] 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.

[0064] like Figure 1 As shown, an embodiment of the present invention provides a dynamic cognitive diagnosis method based on causal inference of multivariate learning behaviors, comprising the following steps:

[0065] Step 1: Obtain knowledge, behavior, and time series data during the learning process to construct a student answer data set; knowledge data includes the test questions answered by students at a certain moment and the knowledge related to the test questions; behavior data includes 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 the answering process; time series data includes the timestamp information of the current behavior;

[0066] Step 2: Construct a learning behavior graph for the click streams and activity streams of students in the dataset, and map different behavior features to a unified behavior space. Specifically, the graph includes: constructing a click stream subgraph by counting click features using timestamps, constructing an activity stream subgraph by counting activity features using timestamps, and connecting the click stream subgraph and the activity stream subgraph to form a unified learning behavior graph.

[0067] like Figure 1 , behavior graph construction part; in the continuous learning process, the learner's learning behavior consists of click stream cs and activity stream as. In order to fully characterize the learning behavior in the learning process, this method constructs a behavior graph for each answer of the learner. Clickstream subgraph and the activity flow subgraph A represents an edge and X represents a node.

[0068] (2-1) Click flow subgraph The construction, such as Figure 2 shown.

[0069] First, from the click stream middle, represents the click behavior on question n at time i, N c Represents the total number of click streams. The behavioral characteristics of each option of the learner are counted in turn through the timestamp, and the click options are extracted The reaction time t of the option n and option repetition times p n Three behavioral characteristics;

[0070] Next, click on the option The timestamp information of the answer is set, and the initial node, intermediate node and end node are set to highlight the initial and final answer options. The nodes are recorded as The initial embedding of nodes is initialized randomly.

[0071] Finally, connect the start and end nodes with the click option nodes in the middle process, connect the reaction time and the number of options with the corresponding click options, complete the edge relationship between the nodes, and construct the edge in the click flow subgraph. By setting nodes and edges, a click flow subgraph is formed To describe the learner's answering process.

[0072] (2-2) Activity Flow Subgraph Construction

[0073] After the learner has completed the answer, the activity flow is constructed in chronological order, including the learner's behavior of viewing the answer analysis, watching the video, and answering the recommended exercises. represents the activity behavior on question n at time i, N a Indicates the total number of active flows;

[0074] First, from the activity stream In the process, the behavioral characteristics of each activity in which the learner participates are counted in sequence by timestamp, and the activity type is extracted. Event Time Number of activities vt expi Other activity behavior characteristics;

[0075] Secondly, according to the three activity behavior characteristics, obtain the activity node The initial embedding of nodes is initialized randomly.

[0076] Thirdly, the edge setting between nodes, according to the information before and after the timestamp, establishes edge connections between activity type nodes, activity time nodes, activity number nodes and corresponding activity type nodes respectively, and constructs the edge in the activity flow subgraph.

[0077] (2-3) Connect the click flow subgraph and the activity flow subgraph

[0078] Connect the end node of the click flow subgraph to the activity type node of the activity flow subgraph, thereby connecting the click flow subgraph and the activity flow subgraph into a complete behavior graph To describe the entire process of learners' answers.

[0079] Step 3: divide the behaviors in the learning behavior graph into key behaviors and inefficient behaviors, forming two subgraphs, namely, the key behavior subgraph and the inefficient behavior subgraph, and use the Do operator to generate an intervention graph to explore the causal relationship between the key behaviors and inefficient behaviors and the knowledge state;

[0080] like Figure 1 , the graph processing part based on causal attention; at the same time, considering the causal relationship between learning behavior and cognition, the weight difference is learned through the causal attention mechanism to achieve automatic division of key behavior subgraphs and inefficient behavior subgraphs. Then, the learning behavior is regarded as a variable in causal inference, the inefficient behavior is intervened, the combination of key behavior and inefficient behavior is established, and the intervention graph is obtained.

[0081] (3-1) Behavior Graph Coding

[0082] First, the graph attention network GAT is used to complete the behavior graph encoding and obtain the representation matrix X of the behavior graph nodes from the learning behavior graph. n and the adjacency matrix A of the behavior graph n .

[0083] Secondly, according to the correlation between neighbor nodes and themselves, the graph convolution method is used to aggregate the information of the behavior graph nodes to obtain the representation H of all nodes in the entire graph. n , let Hn=GConv(An,Xn⊙Mn), where ⊙ represents the dot product, GConv represents the graph convolution operation, and M n is the attention weight matrix obtained through the self-attention mechanism, that is Among them, the weight W q and W k For the linear transformation in the attention mechanism, d k The embedding dimension of the representation.

[0084] Finally, a linear layer is used to complete the node update and obtain the final node representation.

[0085] (3-2) Division of critical behavior subgraphs and inefficient behavior subgraphs

[0086] The causal attention method is used to automatically learn associations and classify behaviors into key learning behaviors that have a significant impact on the knowledge state and inefficient learning behaviors that have a weak impact. Specifically,

[0087] First, two multi-layer perceptrons are set up to obtain the attention scores of key learning behaviors from the node level and edge level, namely

[0088] First, obtain the node-level attention score. For each node in the behavior graph, set up a multi-layer perceptron, and the node-level attention score α ki The calculation formula is as follows:

[0089]

[0090] Among them, MLP n ode is a multilayer perceptron for nodes, the function exp() represents the natural exponential function, hi Represents the i-th node representation in the behavior graph.

[0091] Second, get the edge-level attention score. For each edge in the behavior graph, set up a multi-layer perceptron, and the edge-level attention score of edge e The calculation formula is as follows:

[0092]

[0093] Among them, MLP edge is a multi-layer perceptron for edges, represents the representation of the edge e formed by the i-th and j-th nodes in the behavior graph, Indicates splicing.

[0094] From this, we aggregate the attention scores of each node and each edge to form a node-level attention score for the entire behavior graph and edge-level attention score

[0095] For inefficient learning behaviors, we subtract the score of the corresponding key learning behavior from 1 to obtain the node attention score and edge attention score of the inefficient learning behavior, that is, The node attention score and edge attention score of the entire behavior graph are expressed as

[0096] (3-3) Obtaining the intervention map

[0097] Use Do operator intervention to obtain intervention graphs to explore the causal relationship between key learning behaviors and inefficient learning behaviors and knowledge states; specifically,

[0098] First, the key behavior subgraphs are extracted from the learning behavior graph using the causal attention scores. and the inefficient behavior subgraph Right now ⊙ represents the Hadamard product of the matrix.

[0099] Then, pooling is used to obtain the representations of the two subgraphs, namely the key behavior representations and inefficient behavior representation A queue structure is used to build a memory bank, denoted as MB, to record the representations of different inefficient behavior subgraphs of multiple learners.

[0100] Thirdly, in order to explore the behavioral factors that affect learning outcomes, learning behaviors are considered as variables in causal inference, and inefficient behaviors are intervened, that is, the key behavioral variables are fixed and the inefficient behavioral variables are changed. The intervention process formula is:

[0101]

[0102] in, is the graph generated by the intervention operation, is the inefficient behavior obtained from the repository. Note that for each key behavior, we randomly obtain multiple different inefficient behaviors from the repository. Combined with them respectively, several intervention graphs are generated, and the set of intervention graphs is recorded as M is the number of intervention graphs.

[0103] Step 4: Based on the acquired key behavior subgraph, the masked attention mechanism is used to fuse it with knowledge and temporal features to dynamically diagnose the student's knowledge status. At the same time, different prediction targets are set for the acquired key behavior subgraph, inefficient behavior subgraph and generated graph representation. The cross loss entropy is calculated between the key behavior, inefficient behavior and the actual answer, while the cross loss entropy is calculated between the inefficient behavior and the fixed value. The final loss function is calculated by the weighted sum of the three to achieve behavioral performance prediction.

[0104] like Figure 1 , performance prediction part; based on the acquired key behavior subgraph, inefficient behavior subgraph and intervention graph representation, different prediction goals are set through a multi-task mechanism to explore the causal relationship between learning behavior and implicit cognition.

[0105] (4-1) Temporal Attention for Key Behavior-Knowledge Fusion

[0106] First, feature acquisition. This step requires three features in the data set: knowledge features, time series features, and behavior features. ① Knowledge features, the knowledge information of the test questions answered by learners, generally using the test question number or the knowledge points associated with the test question, and embedding is used to embed and record ② Time series features: According to the timestamp information in the learning process time series data, the time interval between each student’s answer and the last answer is calculated and recorded as the time series feature Embedding is used for embedding; ③Key behavioral features, i.e., the key behavioral features processed in step (3-3)

[0107] Secondly, cognitive tracking that integrates knowledge, timing, and key behavioral features. A multi-head self-attention mechanism with masking is set up to mask the question sequence after the current moment, so as to transfer the temporal influence of the previous answering behavior to the current answer, obtain the potential cognitive state, and record

[0108]

[0109] Among them, MHSA stands for Multi-Head Self-Attention Network, The knowledge representation of the corresponding test questions, is the corresponding temporal feature embedding, It is the learner's cognitive state.

[0110] The intervention graph and the inefficient learning behavior subgraph do not pass through the self-attention network.

[0111] (4-2) Behavioral performance prediction

[0112] First, for key behaviors, there is a causal relationship between key learning behaviors and correct or incorrect responses. represents the key behavior, where qn+1 is the question to be answered, σ is the sigmoid activation function, and MLP k is a multi-layer perceptron. Calculate the cross loss entropy with the actual answer result, and the loss function is recorded as where r n+1 Is the test question waiting to be answered? n+1 The actual reaction result, r n+1 ∈{0, 1}, 0 indicates an incorrect answer, 1 indicates a correct answer, and N indicates the total number of answers given by the learner.

[0113] Secondly, for inefficient behavior, inefficient learning behavior has no causal relationship with answer performance, so we will pass the linear layer, that is, Get inefficient behavior prediction results The cross loss entropy is calculated with a fixed value of 0.5, which is the middle value of the two-class classification, meaning that inefficient behavior is unlikely to affect the answer result. The loss function is recorded as N represents the total number of answers given by learners.

[0114] Finally, the intervention map The inefficient behaviors are generated by the combination of intervention operations and key behaviors. Therefore, for each intervention graph, we pass it through a linear layer, i.e. Get the intervention map prediction results With the collection Then, we calculate the cross loss entropy with the corresponding real answer result, that is, N represents the total number of learners’ responses, and M represents the number of intervention diagrams.

[0115] The final loss function is obtained by weighted summation, that is, where λ 1 and λ 2 is a hyperparameter weight used to adjust the impact of the last two loss functions.

[0116] The contents not described in detail in this specification belong to the prior art known to those skilled in the art.

[0117] 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 dynamic cognitive diagnosis method based on causal inference of multivariate learning behaviors, characterized in that: The steps include: Step 1: Obtain knowledge, behavior, and time series data during the learning process to construct a student answer data set; knowledge data includes the test questions answered by students at a certain moment and the knowledge related to the test questions; behavior data includes 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 the answering process; time series data includes the timestamp information of the current behavior; Step 2: Construct a learning behavior graph for the click streams and activity streams of students in the dataset, and map different behavior features to a unified behavior space. Specifically, the graph includes: constructing a click stream subgraph by counting click features using timestamps, constructing an activity stream subgraph by counting activity features using timestamps, and connecting the click stream subgraph and the activity stream subgraph to form a unified learning behavior graph. Step 3: divide the behaviors in the learning behavior graph into key behaviors and inefficient behaviors, forming two subgraphs, namely, the key behavior subgraph and the inefficient behavior subgraph, and use the Do operator to generate an intervention graph to explore the causal relationship between the key behaviors and inefficient behaviors and the knowledge state; Step 4: Based on the acquired key behavior subgraph, inefficient behavior subgraph, and intervention graph representations, different prediction targets are set through a multi-task mechanism to explore the causal relationship between learning behavior and implicit cognition, and the cross loss entropy is calculated between the key behavior, inefficient behavior, and the actual answer, while the cross loss entropy is calculated between the inefficient behavior and the fixed value. The final loss function is calculated by the weighted sum of the three to achieve key behavior prediction, intervention graph prediction, and inefficient behavior prediction.

2. A dynamic cognitive diagnosis method based on causal inference of multiple learning behaviors as claimed in claim 1, characterized in that: Click on the flow subgraph in step 2 The construction method is as follows; First, from the click stream middle, represents the click behavior on question n at time i, N c Indicates the total number of click streams; Count the behavioral characteristics of each option of the learner in turn through the timestamp and extract the clicked options The reaction time t of the option n and option repetition times p n Three behavioral characteristics; Next, click on the option The timestamp information of the node is set, and the initial node, intermediate node and end node are set. The node is recorded as The initial embedding of nodes is randomly initialized; Finally, connect the start and end nodes with the click option nodes in the middle process, connect the reaction time and the number of options with the corresponding click options, complete the edge relationship between the nodes, and construct the edge in the click flow subgraph. By setting nodes and edges, a click flow subgraph is formed To describe the learner's answering process.

3. A dynamic cognitive diagnosis method based on causal inference of multiple learning behaviors as claimed in claim 1, characterized in that: Activity flow subgraph in step 2 The construction method is as follows; After the learner has completed the answer, the activity flow consisting of the learner viewing the answer analysis, watching the video, and recommending the practice answer behavior is constructed in chronological order. represents the activity behavior on question n at time i, N a Indicates the total number of active flows; First, from the activity stream In the process, the behavioral characteristics of each activity in which the learner participates are counted in sequence by timestamp, and the activity type is extracted. Event Time Number of activities These activities have behavioral characteristics; Secondly, according to the three activity behavior characteristics, obtain the activity node The initial embedding of nodes is randomly initialized; Thirdly, the edge setting between nodes, according to the information before and after the timestamp, establishes edge connections between activity type nodes, activity time nodes, activity number nodes and corresponding activity type nodes respectively, and constructs the edge in the activity flow subgraph.

4. A dynamic cognitive diagnosis method based on causal inference of multiple learning behaviors as claimed in claim 1, characterized in that: The specific implementation of step 3 is as follows: (3-1) Obtaining node representations in the learning behavior graph; (3-2) Division of critical behavior subgraphs and inefficient behavior subgraphs The causal attention method is used to automatically learn associations and classify behaviors into key learning behaviors that have a significant impact on the knowledge state and inefficient learning behaviors that have a weak impact. Specifically, First, two multi-layer perceptrons are set to obtain attention scores from the node level and edge level respectively, namely: First, obtain the node-level attention score: For each node in the learning behavior graph, set up a multi-layer perceptron, the node-level attention score of the node The calculation formula is as follows: Among them, MLP node It is a multilayer perceptron for nodes. The function exp() represents the natural exponential function. h i represents the i-th node representation in the behavior graph, H n Represents all node representations of the behavior graph; Second, get the edge-level attention score: For each edge in the learning behavior graph, set up a multi-layer perceptron, and the edge-level attention score of edge e The calculation formula is as follows: Among them, MLP edge is a multi-layer perceptron for edges, represents the representation of the edge e formed by the i-th and j-th nodes in the behavior graph, Indicates splicing, represents the edges in the clickstream subgraph, represents the edges in the activity flow subgraph; Thus, the attention scores of each node and each edge are aggregated to form a node-level attention score for the entire behavior graph. and edge-level attention score For inefficient learning behaviors, the scores of the corresponding key learning behaviors are subtracted from 1 to obtain the node attention scores and edge attention scores of the inefficient learning behaviors, that is, The node attention score and edge attention score of the entire behavior graph are expressed as and (3-3) Use Do operator intervention to obtain intervention graphs to explore the causal relationship between key learning behaviors, inefficient learning behaviors and knowledge states; First, the key behavior subgraphs are extracted from the learning behavior graph using the causal attention scores. and the inefficient behavior subgraph Right now and ⊙ represents the Hadamard product of the matrix, X n is the representation matrix of the learning behavior graph nodes and the adjacency matrix A of the learning behavior graph n ; Then, pooling is used to obtain the representations of the two subgraphs, namely the key behavior representations and inefficient behavior representation And use the queue structure to build a repository, denoted as MB, to record the representations of different inefficient behavior subgraphs of multiple learners; Thirdly, in order to explore the behavioral factors that affect learning outcomes, learning behavior is regarded as a variable in causal inference, and inefficient behavior is intervened, that is, the key behavioral variables are fixed and the inefficient behavioral variables are changed. The intervention process formula is: in, is the graph generated by the intervention operation, Inefficient behavior for fetching from the repository; Finally, for each critical behavior, multiple different inefficient behaviors are randomly obtained from the repository Combined with them respectively, several intervention graphs are generated, and the set of intervention graphs is recorded as M is the number of intervention graphs.

5. A dynamic cognitive diagnosis method based on causal inference of multiple learning behaviors as claimed in claim 4, characterized in that: The specific implementation method of step (3-1) is as follows: First, the graph attention network GAT is used to complete the behavior graph encoding and obtain the representation matrix X of the behavior graph nodes from the learning behavior graph. n and the adjacency matrix A of the behavior graph n ; Secondly, according to the correlation between neighbor nodes and themselves, the graph convolution method is used to aggregate the information of the behavior graph nodes to obtain the representation H of all nodes in the entire graph. n , note H n =GConv(A n ,X n ⊙M n ), where ⊙ represents point multiplication, GConv represents graph convolution operation, and M n is the attention weight matrix obtained through the self-attention mechanism, that is Among them, the weight W q and W k For the linear transformation in the attention mechanism, d k The embedding dimension of the representation; Finally, a linear layer is used to complete the node update and obtain the final node representation.

6. A dynamic cognitive diagnosis method based on causal inference of multiple learning behaviors as claimed in claim 5, characterized in that: The specific implementation method of realizing key behavior prediction in step 4 is as follows; For key behaviors, there is a causal relationship between key learning behaviors and correct or incorrect responses. represents the key behavior, where q n+1 is the test question to be answered, σ is the sigmoid activation function, MLP k is a multi-layer perceptron, is the learner's cognitive state, and the calculation formula is as follows: Among them, MHSA stands for Multi-Head Self-Attention Network, Represents knowledge characteristics, Represents the time series characteristics and represents the key behavioral characteristics, obtained through step 3; Will Calculate the cross loss entropy with the actual answer result, and the loss function is recorded as where r n+1 Is the test question waiting to be answered? n+1 The actual reaction result, r n+1 ∈{0, 1}, 0 indicates an incorrect answer, 1 indicates a correct answer, and N indicates the total number of answers given by the learner.

7. A dynamic cognitive diagnosis method based on causal inference of multiple learning behaviors as claimed in claim 6, characterized in that: The specific implementation method of realizing the inefficient behavior prediction in step 4 is as follows; For inefficient behavior, inefficient learning behavior has no causal relationship with answer performance, so it will pass through the linear layer, that is, Get the prediction results The cross loss entropy is calculated by comparing it with a fixed value of 0.

5. The fixed value of 0.5 is the middle value of the two classifications, which means that inefficient behavior is unlikely to affect the answer result. The loss function is recorded as MLP l It is a multi-layer perceptron.

8. A dynamic cognitive diagnosis method based on causal inference of multiple learning behaviors as claimed in claim 7, characterized in that: The specific implementation method of implementing the intervention map prediction in step 4 is as follows; Finally, the intervention map It is generated by combining intervention operations and key behaviors for inefficient behaviors. Therefore, for each intervention graph, it is passed through a linear layer, i.e. q n+1 It is a test question to be answered, MLP i It is a multi-layer perceptron, and the prediction result is obtained With the collection The intervention graph corresponds to , and then the cross loss entropy is calculated with the corresponding real answer result, that is 9. A dynamic cognitive diagnosis method based on causal inference of multiple learning behaviors as claimed in claim 8, characterized in that: The final loss function is obtained by weighted summation, that is, Among them, λ1 and λ2 are hyperparameter weights used to adjust the impact of the latter two loss functions.

10. A dynamic cognitive diagnosis method based on causal inference of multiple learning behaviors as claimed in claim 6, Features: The knowledge feature represents the knowledge information of the test questions answered by the learners, using the test question number or the knowledge points associated with the test question, and is embedded using Embedding; The time series feature calculates the time interval between each student's answer and the last answer based on the timestamp information in the time series data of the learning process, and embeds it using Embedding.

Citation Information

Patent Citations

  • Multi-level intelligent cognitive tracking method and system, storable medium and terminal

    CN115545155A

  • Industrial situation security basic framework and network attack behavior feature analysis method

    CN116545679A

  • Deep knowledge tracing with transformers

    US20210390873A1

Cited By

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

    CN119514590A

  • Hierarchical attention cognition diagnosis method based on knowledge association and abnormal behavior mixing

    CN119514591A

  • E-commerce scene cold start recommendation method combining meta-learning and causal inference

    CN119741096A

  • Task scene spatio-temporal data prediction method, system and device and medium

    CN120180073A

  • Learning behavior analysis method and system based on smart learning

    CN121169201A