A dynamic cognitive diagnostic method based on causal inference of multiple learning behaviors

Abstract

The application provides a dynamic cognitive diagnosis method based on multi-element learning behavior causal inference, which models learning process behavior by means of click flow and activity flow, and constructs learning behavior graph through multi-element behavior characteristics. Meanwhile, the causal correlation between learning behavior and cognition is considered, the weight difference is learned through a causal attention mechanism, and the automatic division of key behavior subgraph and inefficient behavior subgraph is realized. 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 an intervention graph is obtained. Based on the obtained key behavior subgraph, inefficient behavior subgraph and intervention graph representation, different prediction targets are set through a multi-task mechanism, so as to mine the causal correlation between learning behavior and implicit cognition, and enhance the explainability of dynamic diagnosis of learner's cognitive level.

Description

Technical Field

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

[0002] Learning behavior refers to the process actions or results that learners produce when participating in learning. Based on the stage at which learning behavior occurs, it can be divided into outcome-based behavior and process-based behavior.

[0003] Dynamic cognitive diagnosis oriented towards outcome-based behavior infers learners' potential cognitive levels by analyzing their responses, typically referring to binary outcomes of correct or incorrect. Classic cognitive diagnostic models, such as IRT, DINA, and BKT, infer potential cognitive levels based on learners' binary responses. However, in multiple-choice questions, responses can be expanded to include multiple answer options, not just the common binary outcome. These models all target outcome-based behavior, combining multiple cognitive factors to jointly uncover potential cognitive levels. In reality, most current cognitive diagnostic models are based on outcome-based learning behavior, neglecting the impact of diverse process-based learning behaviors on cognitive states, leading to a disconnect between learners' cognitive levels and the learning process. Therefore, further exploration is needed to diagnose learners' true cognitive states by analyzing process-based behaviors.

[0004] Dynamic cognitive diagnosis oriented towards process-based behavior, building upon dynamic cognitive diagnosis models oriented towards outcome-based behavior, incorporates learning behavior information from the learning process to infer learners' potential cognitive levels. Since process-based learning behaviors implicitly contain rich cognitive state information, cognitive diagnosis incorporating process-based behavior helps improve diagnostic accuracy. While current research has introduced process-based behavior from different perspectives and made some attempts, they primarily use learning behaviors to enrich the representation of response interactions, without exploring the intrinsic connection between behavior and knowledge. Furthermore, overall, current research on process-based behavior in cognitive diagnosis is relatively limited, and existing studies only use fragmented behavioral features, which cannot comprehensively characterize the learning process. Therefore, given the importance of process-based learning behavior to learning outcomes, mining learners' potential cognitive states from learning behaviors and exploring the connection between behavior and knowledge, researching dynamic cognitive diagnosis modeling oriented towards process-based behavior is an important research direction and development trend in this field.

[0005] In the field of Massive Open Online Courses (MOOCs), numerous studies have explored the diverse learning behaviors involved in the learning process. In particular, some scholars have introduced the concept of "clickstreams" to describe the temporal sequence of behaviors during learning. A clickstream is the trajectory of a learner's continuous access to a learning system and is commonly used to predict dropout rates on online learning platforms. After completing an answer, learners can participate in other activities, such as viewing explanations, watching videos, and engaging in reinforcement exercises; these activities form an "activitystream." In fact, both clickstreams and activitystreams can comprehensively record the sequence of learning behaviors, which helps in using neural network technology to simulate the learner's actual learning process. However, the application of procedural behaviors modeled based on clickstreams and activitystreams in the field of cognitive diagnostics requires further research.

[0006] Current dynamic cognitive diagnostics neglect the impact of diverse behaviors on learners' cognition during long-term learning processes, leading to an inability to accurately assess learners' overall cognitive level and consequently affecting the interpretability of the model. Furthermore, there are complex causal relationships between diverse learning behaviors and learning outcomes, and different behavioral elements have varying impacts on learners' cognitive states. Identifying key behavioral elements in the learning process will help improve the accuracy and interpretability of diagnostic results. Summary of the Invention

[0007] Because the behavioral characteristics during the learning process are complex and diverse—for example, learners repeatedly modifying their answers, forming a "clickstream"; and learners reflecting after completing exercises through text analysis or targeted video recommendations, forming an "activitystream"—learners' learning outcomes are causally correlated with various behavioral elements, and the degree of correlation varies. Therefore, based on the diversified behaviors formed during the learning process, this paper proposes a dynamic cognitive diagnostic method based on causal inference of multiple learning behaviors. It constructs an answer behavior graph and encodes it using a graph neural network, then uses causal attention to explore the correlation between multiple behaviors and knowledge states, ultimately achieving a dynamic diagnosis of learners' cognitive levels.

[0008] The purpose of this invention is to address the diagnostic bias and causal incomprehension problems caused by the neglect of causal analysis of diverse learning behaviors in existing cognitive diagnostic methods. This invention provides a dynamic cognitive diagnostic method based on causal inference of diverse learning behaviors. It models learning process behaviors using clickstreams and activity streams, constructing a learning behavior graph through diverse behavioral features. Simultaneously, it considers the causal relationship between learning behaviors and cognition, automatically dividing key behavior subgraphs and inefficient behavior subgraphs through causal attention mechanisms and learning weight differences. Then, learning behaviors are treated as variables in causal inference, and inefficient behaviors are intervened to establish combinations of key and inefficient behaviors, obtaining a generated graph. Based on the obtained key behavior subgraphs, inefficient behavior subgraphs, and generated graph representations, a multi-task mechanism is used to set different prediction targets, thereby uncovering the causal relationship between learning behaviors and implicit cognition, enhancing the interpretability of dynamic diagnosis of learners' cognitive levels.

[0009] The objective of this invention is achieved through the following technical measures: a dynamic cognitive diagnostic method based on causal inference of multiple learning behaviors, comprising the following steps:

[0010] Step 1: Acquire knowledge, behavioral, and temporal data during the learning process to construct a student answer dataset; knowledge data includes the questions answered by students at a certain moment and the knowledge related to the questions; behavioral data includes the click stream formed by students repeatedly modifying their answers during the answering process, and the activity stream formed by students viewing explanations and video resources after answering; temporal data includes the timestamp information of the current behavior.

[0011] Step 2: For the click flow and activity flow of students in the dataset, construct a learning behavior graph to map different behavioral features to a unified behavior space. Specifically, this includes: constructing a click flow subgraph by statistically analyzing click features using timestamps, constructing an activity flow subgraph by statistically analyzing activity features using timestamps, and connecting the click flow subgraph and the activity flow subgraph to form a unified learning behavior graph.

[0012] Step 3: Divide the behaviors in the learning behavior map into key behaviors and inefficient behaviors, forming two sub-maps: a key behavior sub-map and an inefficient behavior sub-map. Then, use the Do operator to generate an intervention map to explore the causal relationship between key behaviors and inefficient behaviors and knowledge state.

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

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

[0015] First, from the clickstream middle, N represents the click action on question n at time i, and N represents the click action on question n at time i. c This represents the total number of clickstreams; by using timestamps, the behavioral characteristics of each option chosen by the learner are analyzed sequentially to extract the clicked options. The reaction time t consumed by the option n and the number of times the option is repeated p n Three behavioral characteristics;

[0016] Secondly, based on the clicked options The timestamp information is used to set the initial node, intermediate nodes, and end node, and the node is denoted as . The initial embedding of nodes uses a random initialization method;

[0017] Finally, the start and end nodes are connected to the click option nodes in the intermediate process, and the reaction time and the number of times an option is selected are connected to the corresponding click option. This completes the edge relationships between the nodes and constructs the edges in the click flow subgraph. By setting the nodes and edges, a click flow subgraph is formed. To describe the learner's response process.

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

[0019] After learners complete their answers, an activity flow is constructed in chronological order, consisting of learners viewing answer explanations, watching videos, and requesting recommended practice answers. N represents the action of the user on question n at time i. a Indicates the total number of activity flows;

[0020] First, from the activity flow In this process, behavioral characteristics of learners participating in each activity are statistically analyzed using timestamps, and activity types are extracted. Event Time Number of activities These activities and behavioral characteristics;

[0021] Secondly, activity nodes are obtained based on the three types of activity behavior characteristics. The initial embedding of nodes uses a random initialization method;

[0022] Next, regarding the edge settings between nodes, edges are established between activity type nodes based on the timestamp information. Edges are also established between activity time nodes, activity count nodes, and their corresponding activity type nodes to construct the edges in the activity flow subgraph.

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

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

[0025] (3-2) Division of Key Behavior Subgraphs and Inefficient Behavior Subgraphs

[0026] This approach uses causal attention to automatically learn associations and categorizes behaviors into key learning behaviors that significantly impact knowledge state and inefficient learning behaviors that have a weak impact. Specifically...

[0027] First, two multilayer perceptrons are set up to obtain attention scores from the node level and the edge level, respectively:

[0028] First, obtain node-level attention scores: For each node in the learning behavior graph, set up a multilayer perceptron to obtain the node-level attention score. 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 H represents the representation of the i-th node in the behavioral graph. n This represents the representation of all nodes in the behavior graph;

[0031] Second, obtain the edge-level attention score: For each edge in the learning behavior graph, set up a multilayer perceptron, and obtain the edge-level attention score for edge e. The calculation formula is as follows:

[0032]

[0033] Among them, MLP edge It is a multilayer perceptron for edges. Let e ​​represent the edge e formed by the i-th and j-th nodes in the behavioral graph. Indicates splicing, This indicates clicking on an edge in the flow subgraph. Represents the edges in the active flow subgraph;

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

[0035] For inefficient learning behaviors, subtracting the score of the corresponding key learning behavior from 1 yields the node attention score and edge attention score for the inefficient learning behavior. The node attention score and edge attention score of the entire behavior graph are represented as follows:

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

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

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

[0039] Furthermore, in order to uncover the behavioral factors influencing learning effectiveness, learning behaviors are treated as variables in causal inference. Interventions are implemented for inefficient behaviors, i.e., key behavioral variables are fixed while inefficient behavioral variables are changed. The intervention process formula is as follows:

[0040]

[0041] in, The graph is generated through intervention operations. Inefficient behavior of fetching from the repository;

[0042] Finally, for each key behavior, several different inefficient behaviors are randomly selected from the repository. Each of these can be combined to generate several intervention diagrams, and the set of intervention diagrams is denoted as . M represents the number of intervention maps.

[0043] Furthermore, step (3-1) is implemented in the following specific way;

[0044] First, a graph attention network (GAT) is used to encode the behavior graph, obtaining the representation matrix X of the behavior graph nodes from the learned behavior graph. n Adjacency matrix A of the behavior graph n ;

[0045] Secondly, based on the relevance of neighboring nodes to itself, a graph convolution method is used to aggregate the information of behavioral graph nodes, obtaining the representation H of all nodes in the entire graph. n Let Hn = GConv(An, Xn⊙Mn), where ⊙ denotes dot product, GConv denotes graph convolution operation, and Mn represents the graph convolution operation. n It is the attention weight matrix obtained through the self-attention mechanism, i.e. Wherein, weight W q and W k Linear transformation used in attention mechanisms, d k The embedding dimension of the representation;

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

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

[0048] For key behaviors, a causal relationship exists between key learning behaviors and correct or incorrect responses. This represents the key behavior, where qn+1 is the question to be answered, σ is the sigmoid activation function, and MLP is used. k It is a multilayer perceptron. It refers to the learner's cognitive state, calculated using the following formula:

[0049]

[0050] MHSA stands for Multi-Head Self-Attention Network. Representing knowledge characteristics, Representing time series characteristics and Key behavioral characteristics are identified and obtained through step 3;

[0051] Will The cross-loss entropy is calculated by comparing the actual responses with the actual responses, and the loss function is denoted as . Where r n+1 It is the unanswered question q n+1 The actual reaction results on, r n+1 ∈{0, 1}, where 0 indicates an incorrect answer, 1 indicates a correct answer, and N represents the total number of answers given by the learner.

[0052] Furthermore, the specific implementation method for predicting inefficient behavior in step 4 is as follows;

[0053] For inefficient behaviors, there is no causal relationship between inefficient learning behaviors and response performance; therefore, a linear layer will be used, i.e. Obtain 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 median for binary classification, meaning that inefficient behavior is unlikely to affect the answer result; the loss function is denoted as... MLP l It is a multilayer perceptron.

[0054] Furthermore, the specific implementation method for intervention map prediction in step 4 is as follows;

[0055] Finally, the intervention diagram It is generated by intervening inefficient behaviors and combining key behaviors. Therefore, for each intervention map, it is passed through a linear layer, i.e. q n+1 These are unanswered questions, MLP i It is a multilayer perceptron, which obtains the prediction results. With set The intervention map corresponds to the actual response result, and then the cross-loss entropy is calculated by comparing it with the actual response result.

[0056]

[0057] Furthermore, the final loss function is obtained by weighted summation, i.e. λ1 and λ2 are hyperparameter weights used to adjust the influence of the latter two loss functions.

[0058] Furthermore, knowledge features represent the knowledge information of the questions answered by learners, using question numbers or related knowledge points, and are embedded using embedding.

[0059] The temporal features are calculated based on the timestamp information in the temporal data of the learning process, and the time interval between each student's answer and the previous answer is embedded using Embedding.

[0060] Compared with existing technologies, the beneficial effects of this invention are as follows: Existing dynamic cognitive diagnosis ignores the impact of diverse behaviors on learners' cognition during long-term learning processes, resulting in the inability to accurately assess learners' comprehensive cognitive level, thereby affecting the interpretability of the model. In contrast, the method of this invention considers the complex causal relationship between diverse learning behaviors and learning outcomes, and that different behavioral elements have different impacts on learners' cognitive states. This approach aims to uncover key behavioral elements in the learning process, thereby improving the accuracy and interpretability of diagnostic results. Attached Figure Description

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

[0062] Figure 2 This is a schematic diagram illustrating the construction of the learning behavior graph in an embodiment of the present invention. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

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

[0065] Step 1: Acquire knowledge, behavioral, and temporal data during the learning process to construct a student answer dataset; knowledge data includes the questions answered by students at a certain moment and the knowledge related to the questions; behavioral data includes the click stream formed by students repeatedly modifying their answers during the answering process, and the activity stream formed by students viewing explanations and video resources after answering; temporal data includes the timestamp information of the current behavior.

[0066] Step 2: For the click flow and activity flow of students in the dataset, construct a learning behavior graph to map different behavioral features to a unified behavior space. Specifically, this includes: constructing a click flow subgraph by statistically analyzing click features using timestamps, constructing an activity flow subgraph by statistically analyzing activity features using timestamps, and connecting the click flow subgraph and the activity flow subgraph to form a unified learning behavior graph.

[0067] like Figure 1 The behavior graph construction part involves constructing a behavior graph for each learner's response during continuous learning. In this process, the learner's learning behavior consists of a click flow (CS) and an activity flow (AS). From click flow subgraph and active flow subgraph The structure consists of A representing an edge and X representing a node.

[0068] (2-1) Click on the flow graph The construction, such as Figure 2 As shown.

[0069] First, from the clickstream middle, N represents the click action on question n at time i, and N represents the click action on question n at time i. c This represents the total number of clickstreams. The learner's behavioral characteristics for each option are analyzed sequentially using timestamps to extract the clicked options. The reaction time t consumed by the option n and the number of times the option is repeated p n Three behavioral characteristics;

[0070] Secondly, based on the clicked options The timestamp information is used to set initial, intermediate, and ending nodes to highlight the initial and final answer options. The nodes are denoted as... The initial embedding of nodes uses a random initialization method.

[0071] Finally, the start and end nodes are connected to the click option nodes in the intermediate process, and the reaction time and the number of times an option is selected are connected to the corresponding click option. This completes the edge relationships between the nodes and constructs the edges in the click flow subgraph. By setting the nodes and edges, a click flow subgraph is formed. To describe the learner's response process.

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

[0073] After learners complete their answers, an activity flow is constructed in chronological order, consisting of learners viewing answer explanations, watching videos, and requesting practice questions. N represents the action of the user on question n at time i. a Indicates the total number of activity flows;

[0074] First, from the activity flow In this process, behavioral characteristics of learners participating in each activity are statistically analyzed using timestamps, and activity types are extracted. Event Time Number of activities vt expi Characteristics of activities and behaviors;

[0075] Secondly, activity nodes are obtained based on the three types of activity behavior characteristics. The initial embedding of nodes uses a random initialization method.

[0076] Next, regarding the edge settings between nodes, edges are established between activity type nodes based on the timestamp information. Edges are also established between activity time nodes, activity count nodes, and their corresponding activity type nodes to construct the edges 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. It enables the description of the entire process of learners responding to questions.

[0079] Step 3: Divide the behaviors in the learning behavior map into key behaviors and inefficient behaviors, forming two sub-maps: a key behavior sub-map and an inefficient behavior sub-map. Then, use the Do operator to generate an intervention map to explore the causal relationship between key behaviors and inefficient behaviors and knowledge state.

[0080] like Figure 1The graph processing component is based on causal attention. It considers the causal relationship between learning behavior and cognition, and automatically divides key behavior subgraphs and inefficient behavior subgraphs by learning weight differences through a causal attention mechanism. Then, learning behavior is treated as a variable in causal inference, and interventions are applied to inefficient behaviors to establish combinations of key and inefficient behaviors, resulting in an intervention graph.

[0081] (3-1) Behavioral graph coding

[0082] First, a graph attention network (GAT) is used to encode the behavior graph, obtaining the representation matrix X of the behavior graph nodes from the learned behavior graph. n Adjacency matrix A of the behavior graph n .

[0083] Secondly, based on the relevance of neighboring nodes to itself, a graph convolution method is used to aggregate the information of behavioral graph nodes, obtaining the representation H of all nodes in the entire graph. n Let Hn = GConv(An, Xn⊙Mn), where ⊙ denotes dot product, GConv denotes graph convolution operation, and Mn represents the graph convolution operation. n It is the attention weight matrix obtained through the self-attention mechanism, i.e. Wherein, weight W q and W k Linear transformation used in attention mechanisms, d k The embedded dimension is represented.

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

[0085] (3-2) Division of Key Behavior Subgraphs and Inefficient Behavior Subgraphs

[0086] This approach uses causal attention to automatically learn associations and categorizes behaviors into key learning behaviors that significantly impact knowledge state and inefficient learning behaviors that have a weak impact. Specifically,

[0087] First, two multilayer perceptrons are set up to obtain attention scores for key learning behaviors from the node level and edge level, respectively.

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

[0089]

[0090] Among them, MLP n An ode is a multilayer perceptron for nodes, and the function exp() represents the natural exponential function. i This represents the i-th node in the behavioral graph.

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

[0092]

[0093] Among them, MLP edge It is a multilayer perceptron for edges. Let e ​​represent the edge e formed by the i-th and j-th nodes in the behavioral graph. Indicates splicing.

[0094] Therefore, we aggregate the attention scores of each point and each edge to form a node-level attention score for the entire behavior graph. and marginal 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. The node attention score and edge attention score of the entire behavior graph are represented as follows:

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

[0097] The Do operator is used to obtain an intervention map to uncover the causal relationship between key learning behaviors and inefficient learning behaviors and knowledge state; specifically,

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

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

[0100] Furthermore, to uncover the behavioral factors influencing learning effectiveness, learning behavior is treated as a variable in causal inference. Intervention is then implemented for inefficient behaviors, meaning the key behavioral variables are fixed while the inefficient behavioral variables are modified. The intervention process formula is:

[0101]

[0102] in, The graph is generated through intervention operations. These are inefficient behaviors retrieved from the repository. Note that for each key behavior, we randomly retrieve several different inefficient behaviors from the repository. Each of these can be combined to generate several intervention diagrams, and the set of intervention diagrams is denoted as . M represents the number of intervention maps.

[0103] Step 4: Based on the acquired key behavior subgraphs, a masked attention mechanism is used to fuse them with knowledge and temporal features to dynamically diagnose students' knowledge status. At the same time, different prediction targets are set for the acquired key behavior subgraphs, inefficient behavior subgraphs, and generated graph representations. Cross-loss entropy is calculated between key behaviors, inefficient behaviors, and actual answers, while cross-loss entropy is calculated between inefficient behaviors and fixed values. Finally, the loss function is calculated by weighted summation of the three to achieve behavioral performance prediction.

[0104] like Figure 1 The performance prediction component, based on the acquired key behavior subgraphs, inefficient behavior subgraphs, and intervention graph representations, uses a multi-task mechanism to set different prediction targets, thereby exploring the causal relationship between learning behavior and implicit cognition.

[0105] (4-1) Temporal attention in key behavior-knowledge fusion

[0106] First, feature acquisition. This step requires three types of features from the dataset: knowledge features, temporal features, and behavioral features. ① Knowledge features: These are the knowledge information from the learners' answers to the questions, typically using question numbers or related knowledge points, embedded using embedding. ② Temporal characteristics: Based on the timestamp information in the temporal data of the learning process, the time interval between each student's answer and the previous answer is calculated and denoted as the temporal characteristics. Embedding is used for embedding; ③ Key behavioral features, namely the key behavioral features processed in step (3-3).

[0107] Secondly, cognitive tracking integrates knowledge, temporal sequence, and key behavioral features. A masked multi-head self-attention mechanism is implemented to mask the sequence of questions answered after the current moment, thereby transferring the temporal influence of previous responses to the current response, acquiring the potential cognitive state, and recording it.

[0108]

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

[0110] Intervention maps and inefficient learning behavior sub-maps 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. This represents the key behavior, where qn+1 is the question to be answered, σ is the sigmoid activation function, and MLP is used. k It is a multilayer perceptron. The cross-loss entropy is calculated by comparing the actual responses with the actual responses, and the loss function is denoted as . Where r n+1 It is the unanswered question q n+1 The actual reaction results on, r n+1 ∈{0, 1}, where 0 indicates an incorrect answer, 1 indicates a correct answer, and N represents the total number of answers given by the learner.

[0113] Secondly, regarding inefficient behaviors, there is no causal relationship between inefficient learning behaviors and response performance. Therefore, we will use a linear layer, i.e. Obtaining inefficient behavior prediction results We calculate the cross-loss entropy by comparing it with a fixed value of 0.5. The fixed value of 0.5 is the median for binary classification, meaning that inefficient behavior is unlikely to affect the answer result. The loss function is denoted as... N represents the total number of responses from learners.

[0114] Finally, the intervention diagram It is generated by intervening inefficient behaviors and combining key behaviors. Therefore, for each intervention map, we pass it through a linear layer, i.e. Intervention map prediction results With set The intervention diagram corresponds to this. Furthermore, we calculate the cross-loss entropy by comparing it with the corresponding actual response results, i.e. N represents the total number of responses from learners, and M represents the number of intervention maps.

[0115] The final loss function is obtained by weighted summation, i.e. λ1 and λ2 are hyperparameter weights used to adjust the influence of the latter two loss functions.

[0116] The contents not described in detail in this specification are existing technologies 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 principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A dynamic cognitive diagnosis method based on multi-element learning behavior causal inference, characterized in that, Comprise the following steps: Step 1, obtain the knowledge, behavior and time sequence data in the learning process, build a student answer data set; knowledge data includes the knowledge related to the test questions answered by the student at a certain time, behavior data includes the click stream formed by the student repeatedly modifying the answer in the answering process, and the activity stream formed by the student viewing the analysis and video resources after answering; the time sequence data includes the time stamp information of the current behavior; Step 2, for the click stream and activity stream of the student in the data set, a learning behavior graph is constructed, and different behavior characteristics are mapped to a unified behavior space, which specifically includes: constructing a click stream subgraph by timestamping click features, constructing an activity stream subgraph by timestamping activity features, 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 to form a key behavior subgraph and an inefficient behavior subgraph, and use Do operator intervention to generate an intervention graph to mine the causal relationship between key behaviors and inefficient behaviors and knowledge state; Among them, using Do operator intervention to obtain the intervention graph includes: First, the key behavior sub-graphs and the inefficient behavior sub-graphs are extracted from the learning behavior graph using the scores of causal attention and low efficient behavior sub-graphs Subsequently, the representations of the two sub-graphs, i.e., the key behavior representation and the inefficient behavior representation, are obtained by pooling and low efficient behavior representation and a memory bank, denoted as MB, is constructed using a queue structure to record the representations of the different inefficient behavior sub-graphs of multiple learners; Thirdly, in order to mine the behavior elements affecting learning effectiveness, the learning behavior is regarded as a variable in causal inference, and the inefficient behavior is intervened, that is, the key behavior variable is fixed and the inefficient behavior variable is changed, and the intervention process formula is: wherein, is the graph resulting from the intervention operation, is the inefficient behavior retrieved from the repository; for each key behavior, a number of different inefficient behaviors is randomly retrieved from the repository are combined with, respectively, so as to generate a number of intervention graphs, the set of intervention graphs being denoted M is the number of intervention graphs; Step 4, based on the obtained key behavior subgraph, inefficient behavior subgraph and intervention graph representation, different prediction targets are set through a multi-task mechanism to mine the causal relationship between learning behavior and implicit cognition, and the cross-entropy loss of key behavior, inefficient behavior and real answer calculation is calculated, and the cross-entropy loss of inefficient behavior and fixed value is calculated, and the final loss function is calculated by weighted summation of the three, to realize key behavior prediction, intervention graph prediction and inefficient behavior prediction.

2. The dynamic cognitive diagnosis method based on multi-element learning behavior causal inference according to claim 1, wherein: Clickstream sub-diagram in Step 2 is constructed as follows; First, from the click stream represents the click behavior of the test question n at time i, N c represents the total number of click streams;​ The behavior characteristics of each option of the learner are counted in sequence by time stamp, and the clicked option is extracted Reaction time t of the option n And the number of option repetitions p n Three behavior characteristics Second, according to the timestamp information of the click option , set the initial node, intermediate node and end node, and the node is recorded as The initial embedding of the node adopts the method of random initialization; Finally, the start and end nodes are connected with the intermediate process of the click option node, the reaction time and the number of options are connected with the corresponding click options, the edge relationship between nodes is completed, and the edge of the click flow subgraph is constructed Through the setting of nodes and edges, the click flow subgraph is formed to describe the learner's answering process.

3. The dynamic cognitive diagnosis method based on multi-element learning behavior causal inference according to claim 1, wherein: The active sub-graph in step 2 is constructed as follows; After the learner completes the answer, the time sequence is constructed to build the learner to view the answer analysis, watch the video, and recommend the practice answer behavior to form the activity flow, and record N represents the activity behavior of the test question n at time i a N represents the total number of activity flows First, from the activity stream , the behavior characteristics of each activity participated by the learner are sequentially counted by timestamp, and the activity type activity time activity times These activity behavior characteristics; Secondly, according to the three kinds of activity behavior characteristics, the activity nodes are obtained The initial embedding of the node adopts the method of random initialization; Again, the edges between nodes are set up, according to the before and after information of the timestamp, to establish edge connection between the activity type nodes, and the activity time nodes, the activity frequency nodes and the corresponding activity type nodes respectively establish edge connection, to construct the edges in the activity flow subgraph 4. The dynamic cognitive diagnosis method based on multi-element learning behavior causal inference of claim 1, wherein: The specific implementation of step 3 is as follows: (3-1) Obtain the node representation in the learning behavior graph; (3-2) Division of key behavior subgraph and inefficient behavior subgraph The correlation is automatically learned by the causal attention method, and the behavior is divided into key learning behavior which has significant influence on the knowledge state and inefficient learning behavior which has weak influence, specifically, Firstly, set two multi-layer perceptrons to obtain attention scores from node level and edge level respectively, that is: First, obtain the node-level attention score: for each node in the learned behavior graph, set a multi-layer perceptron, and the node-level attention score of the node The calculation formula is as follows: where MLP node is a multi-layer perceptron for the node, the function exp() represents the natural exponential function, h i denotes the representation of the i-th node in the behavior graph, H n denotes the representation of all nodes in the behavior graph; Second, obtain the edge-level attention score: for each edge in the learned behavior graph, set a multi-layer perceptron, and the edge-level attention score of edge e is The calculation formula is as follows: where MLP edge is a multi-layer perceptron for edges, represents a representation of an edge e formed by i and j nodes in a behavior graph, represents concatenation, represents an edge in a clickstream subgraph, represents an edge in an activity stream subgraph; Thereby, the attention scores for each point and each edge are aggregated, forming node-level attention scores for the entire behavior graph and edge-level attention scores For the inefficient learning behavior, subtract 1 from the score of the corresponding key learning behavior to obtain the node attention score and the edge attention score of the inefficient learning behavior, i.e. The node attention score and the edge attention score of the whole behavior graph are respectively represented as and (3-3) Use Do operator intervention to obtain intervention graph to mine the causal relationship between key learning behavior and inefficient learning behavior and knowledge state.

5. The dynamic cognitive diagnosis method based on multi-element learning behavior causal inference according to claim 4, characterized in that: The specific implementation of step (3-1) is as follows: First, the behavior graph encoding is completed using the graph attention network GAT to obtain the representation matrix X of the behavior graph nodes from the learned behavior graph n and the adjacency matrix A of the behavior graph n ; Secondly, according to the relevance degree of neighbor nodes and itself, the information aggregation of behavior graph nodes is carried out by using graph convolution method, and the representation H of all nodes of the whole graph is obtained n , record H n =GConv(A n ,X n ⊙M n ), wherein, ⊙ represents dot product, GConv represents graph convolution operation, M n is the attention weight matrix obtained by self-attention mechanism, that is wherein, the weight W q and W k are used for linear conversion in attention mechanism, d k represents embedding dimension; Finally, a linear layer is used to complete node update to obtain the final node representation.

6. The dynamic cognitive diagnosis method based on multivariate learning behavior causal inference of claim 5, wherein: The specific implementation of realizing key behavior prediction in step 4 is as follows: For key behaviors, there is a causal relationship between key learning behaviors and the correctness of the response, and the correctness of the response is represented as , where q n+1 is the test 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: wherein MHSA represents a multi-head self-attention network, representing knowledge features, representing temporal features and representing key behavior features, obtained by step 3; will be calculated The cross-entropy loss function is denoted as where r n+1 is the true response result on the test question q n+1 , r n+1 ∈{0,1}, 0 represents an incorrect answer, 1 represents a correct answer, and N represents the total number of answers by the learner.

7. The dynamic cognitive diagnosis method based on multivariate learning behavior causal inference of claim 6, wherein: The specific implementation of realizing inefficient behavior prediction in step 4 is as follows: For inefficient behavior, inefficient learning behavior has no causal connection with answering performance, so it will pass through a linear layer, i.e. Get the prediction result Calculate the cross-entropy loss with a fixed value of 0.5, which is the middle value of binary classification, which means that inefficient behavior is difficult to affect the answering result; the loss function is denoted as MLP l is a multi-layer perceptron.

8. The dynamic cognitive diagnosis method based on multi-element learning behavior causal inference of claim 7, wherein: The specific implementation of realizing intervention graph prediction in step 4 is as follows: The specific implementation of realizing intervention graph prediction in step 4 is as follows: Finally, intervention graph is generated by intervention operation on inefficient behaviors and key behavior combination, thus, for each intervention graph, it is passed through a linear layer, i.e. q n+1 is the test question to be answered, MLP i is a multi-layer perceptron to get the predicted result corresponds to the intervention graph in the set , and then the cross-entropy loss is calculated with the corresponding real answer result, i.e.

9. The dynamic cognitive diagnosis method based on multi-element learning behavior causal inference according to claim 8, characterized in that: The final loss function is then obtained by a weighted sum, i.e. where λ1and λ2are hyperparameter weights to adjust the influence of the last two loss functions.

10. The dynamic cognitive diagnosis method based on multi-element learning behavior causal inference of claim 6, Characterized in that: Knowledge feature represents the knowledge information of the answered test questions of the learners, adopts the test question number or the knowledge points associated with the test questions, and adopts Embedding for embedding; The time sequence feature calculates the time interval between each answer and the last answer of the student according to the timestamp information in the time sequence data of the learning process, and adopts Embedding for embedding.

Images