A dynamic knowledge tracking method fusing psychological state gating

Abstract

The application discloses a dynamic knowledge tracking method fusing psychological state gating, and belongs to the technical field of intelligent education. The method firstly extracts features such as test question difficulty and discrimination, and fuses the features with answering results to form an interactive vector, which is input into an LSTM network to capture the short-term interactive state of students; meanwhile, a skill-test dynamic heterogeneous graph is constructed, a neural gating mechanism is used for message transmission, and a global skill memory matrix is maintained through a sliding average strategy to obtain a long-term memory state. On this basis, based on a student answering behavior feature sequence, a psychological state gating coefficient is generated by using a multi-head attention mechanism, the long-term memory is multiplied and modulated, and finally, the mastered probability is output by fusing the modulated long-term state and short-term state. The application effectively realizes the decoupling modeling of knowledge level and test-taking psychology by introducing a psychological gating mechanism, and significantly improves the prediction accuracy and model interpretability of the knowledge tracking task.

Description

Technical Field

[0001] This invention relates to the fields of educational artificial intelligence and smart education technology, specifically to a dynamic knowledge tracking method that integrates psychological state gating to simulate students' future performance in real exam scenarios, thereby improving the accuracy of predictions and the interpretability of the model. Background Technology

[0002] Knowledge tracing (KT) is a core technology in personalized education systems. Its goal is to dynamically predict a student's knowledge acquisition over time based on their historical learning interaction sequences. Existing knowledge tracing models are mainly divided into two categories: traditional deep knowledge tracing models and graph neural network-based models.

[0003] Early knowledge tracing methods were primarily represented by Bayesian Knowledge Tracing (BKT). BKT uses Hidden Markov Models (HMMs) to model knowledge states as binary "mastery" or "lack of mastery," updating state probabilities based on observed responses. However, BKT typically assumes that individual knowledge points are independent, ignoring the complex dependencies between them and failing to capture personalized learning characteristics.

[0004] In recent years, with the rise of deep learning technology, models based on recurrent neural networks (RNNs) and memory networks have gradually become mainstream. For example, the Deep Knowledge Tracing (DKT) model uses LSTM to process long-sequence student answer data, enabling it to capture non-linear dependencies in time series; the Dynamic Key-Value Memory Network (DKVMN) introduces an external memory matrix to store and update the states of different knowledge points, improving the model's interpretability. Furthermore, to utilize the structural information between questions and knowledge points, the Graph-based Knowledge Tracing (GKT) model has been proposed. It treats knowledge points as graph nodes and uses graph convolutional networks to aggregate neighbor information to update node states.

[0005] However, existing technologies still have significant limitations in practical applications: on the one hand, most models only focus on the answer results and ignore the dynamic adjustment of performance by non-cognitive psychological factors (such as guessing, mistakes, and fatigue) and behavioral characteristics (such as answering time and number of attempts), making it difficult to accurately distinguish between students' "true abilities" and "state fluctuations" using multi-feature data; on the other hand, while knowledge structure is modeled by constructing knowledge point relationship graphs, most of them use static graphs or simple dynamic update mechanisms, failing to accurately simulate the cognitive laws of knowledge state transformation during the learning process, especially lacking modeling of the complex interaction mechanism between psychological state and knowledge state, which limits the prediction accuracy and interpretability of the model in complex long-sequence learning scenarios.

[0006] To address the aforementioned issues, a method is needed that can decouple psychological state from knowledge state and quantify the impact of psychological state on future performance under long-term memory knowledge state, in order to improve the accuracy of prediction and the interpretability of the model. Summary of the Invention

[0007] The purpose of this invention is to solve the problem that existing knowledge tracing models are difficult to model simultaneously the effects of temporal interaction, knowledge structure and psychological state, and the dynamic interaction mechanism is complex. The invention provides a dynamic knowledge tracing method with a clear structure and integrated psychological state gating.

[0008] This invention proposes a dynamic knowledge tracking method that integrates mental state gating, which specifically includes the following steps:

[0009] 1. Multidimensional feature encoding module

[0010] This module is responsible for integrating and representing the multidimensional attributes of the test questions themselves in a higher order.

[0011] The input is a sequence of students' historical answers. For each question in the sequence... Extract its inherent attribute feature matrix The attribute features include difficulty, distinguishability, and guessing rate.

[0012] First, three independent linear mapping layers are used to map the three scalar features to a high-dimensional feature space, and the mapped feature vectors are concatenated in the channel dimension to construct a multi-channel feature map.

[0013] Secondly, a two-dimensional convolutional neural network is used to extract and fuse local features from the feature map, with the kernel size set to [value missing]. The step size is 1;

[0014] Finally, the convolutional output is processed by a fully connected layer and a hyperbolic tangent activation function (Tanh) to generate a high-dimensional question feature vector containing the comprehensive attributes of the questions. .

[0015] 2. Short-term interactive status module

[0016] This module aims to capture short-term interaction states and attention fluctuations reflected by recent consecutive question-answering behaviors.

[0017] The feature vectors of the test questions generated by the feature encoding module are compared with the students' actual answers. (Correct answers are 1, incorrect answers are 0) Cross-encoding is performed to generate an interactive embedding vector that integrates question information and answer behavior. The test question features and answer results are encoded into dual-channel vectors. LSTM is used to process the interaction embedding vector sequence to capture the contextual dependence of students in continuous answering and extract short-term interaction states that represent short-term attention and contextual cognition.

[0018] LSTM at time Hidden state This refers to a short-term interactive state, and its update formula is:

[0019] ;

[0020] This state It dynamically reflects students' recent fluctuations in learning performance, learning patterns, and cognitive status.

[0021] 3. Long-term memory state module

[0022] This module is responsible for modeling the long-term stable mastery of structured knowledge.

[0023] First, a heterogeneous graph containing the "knowledge point-test question" relationship is constructed, and a globally persistent skill memory matrix is ​​maintained. It is used to store the status of each knowledge point.

[0024] Knowledge state updates are achieved through feature propagation and state updates via neural gating mechanisms:

[0025] (1) Neural gating mechanism: for any node state in the graph The transmit gate coefficients are calculated using two independent nonlinear mapping networks. With the receiver gate coefficient :

[0026] ;

[0027] ;

[0028] in, , This is a smoothing term.

[0029] (2) Dynamic message passing and aggregation: When node information is propagated, it is first controlled by the sending gate and then passes through the adjacency matrix. After weighted aggregation, the message is then filtered by the receive gate. The calculation is as follows:

[0030] ;

[0031] (3) Residual update and skill state generation: Introduce learnable residual strength parameters Node update status The result is obtained by adding the original state to the weighted message:

[0032] ;

[0033] (4) Global Long-Term Memory Update: The global skill memory matrix is ​​maintained using a momentum update strategy, the updated skill node states are extracted, and the global skill memory matrix is ​​updated using an exponential moving average strategy. :

[0034] ;

[0035] in, Forgetting rate; It is the average skill node status of all students at the last moment.

[0036] 4. Psychological state gating module

[0037] This module, based on behaviorist learning theory in psychology, constructs a mapping mechanism from overt behavioral characteristics to internal psychological states. Given the unobservable nature of psychological states, this module captures and quantifies students' dynamic behavioral characteristics during the answering process (including the number of attempts, response time, etc.), calculating their correlation with the current answering context. This generates a scalar-form psychological state gating coefficient, which serves as an adaptive weight to modulate the student's knowledge state vector in real time during the prediction phase, thereby improving the accuracy of inferring the student's true mastery level.

[0038] (1) Query vector construction: embed the current question into a vector Corresponding long-term cognitive concept state Add them together to form the query vector. :

[0039] ;

[0040] (2) Key / Value Pair Construction: Obtain the student's behavioral feature sequence (including number of attempts and response time), and generate a key vector through linear projection. Sum value vector , behavioral feature sequence Mapped to a high-dimensional representation;

[0041] (3) Multi-head attention calculation: Calculate the association weights between behavioral features and the current cognitive state, and aggregate contextual information:

[0042] ;

[0043] The outputs of multi-head attention are concatenated and fused through a linear layer to obtain a psychological feature vector. .

[0044] (4) Gating coefficient generation: The Sigmoid function is used to map psychological features into scalar coefficients.

[0045] ;

[0046] This coefficient It represents a student's level of focus or tendency towards abnormal states at the current moment. When the value approaches 1, it indicates that the student is in a highly focused and efficient learning state; when... When the level is low, abnormal tendencies such as fatigue or distraction can be identified, reducing the impact of the interaction on ability assessment.

[0047] 5. Dual-state modulation prediction module

[0048] This module integrates short-term and long-term states for final prediction and uses psychological state gating coefficients to adaptively calibrate the prediction results.

[0049] First, based on short-term interaction states respectively and long-term knowledge status Two preliminary accuracy predictions were obtained through independent prediction networks. and .

[0050] Then, using the mental state gating coefficient For short-term interaction states respectively and long-term knowledge status The prediction results (after Q-matrix mapping) are dynamically modulated. The final student answer accuracy prediction is then calculated. The calculation formula is:

[0051] ;

[0052] ;

[0053] ;

[0054] in, As a balancing factor, its value range is , Values This means adopting an average fusion strategy to simultaneously take into account students' recent fluctuations in performance and long-term stability of ability. and pass The gating function allows the model to automatically lower the confidence level of the prediction probability when it detects abnormal psychological behaviors such as guessing or making mistakes by students, thereby correcting the prediction results.

[0055] This invention proposes a dynamic knowledge tracking method that integrates mental state gating. Addressing the shortcomings of existing models in distinguishing between knowledge acquisition and cognitive mental states, and the lack of dynamic modeling of knowledge point associations, a three-tiered architecture of "short-term interaction state - long-term memory state - mental gating modulation" is designed. By modeling the evolution of knowledge structure through a neural-gated dynamic graph convolutional network, and utilizing attention mechanisms to generate mental state gating coefficients for adaptive modulation of predictions, fine-grained and interpretable modeling of learners' cognitive states is achieved, significantly improving prediction accuracy and robustness, and providing reliable technical support for personalized education. Attached Figure Description

[0056] Figure 1 The flowchart illustrates the overall process of the dynamic knowledge tracking method that integrates psychological state gating, as provided in this embodiment of the invention.

[0057] Figure 2 This is an overall architecture diagram of the dynamic knowledge tracking method that integrates mental state gating provided in an embodiment of the present invention.

[0058] Figure 3 This is a flowchart of a neural gated dynamic graph convolutional network in an embodiment of the present invention.

[0059] Figure 4 This is a flowchart of the psychological state gating generation module in an embodiment of the present invention. Detailed Implementation

[0060] The following detailed description, with reference to the accompanying drawings, illustrates the specific implementation of the present invention, "A Dynamic Knowledge Tracking Method Integrating Mental State Gating." This embodiment uses a student answering question scenario in an online education platform as a background to elaborate on the processing flow and data calculation logic of the present invention in practical applications.

[0061] A dynamic knowledge tracking method integrating psychological state gating is proposed, and the specific implementation steps are as follows:

[0062] 1. Data Preprocessing and Feature Engineering

[0063] Basic Interaction Sequence: Extract the sequence of answer triples for each student. These represent the question ID, the answer result (0 / 1), and the timestamp, respectively.

[0064] Multidimensional attribute association: The Q-matrix is ​​used to establish the mapping relationship between test questions and knowledge points; at the same time, the inherent attributes of the test questions (such as difficulty and discrimination) are extracted. If the dataset is not provided, it is initialized based on global interaction statistics.

[0065] Behavioral Feature Engineering: Input Feature Dimensions of the Mental State Module Set to 2, including:

[0066] Attempts: The number of attempts for this problem after normalization, adjusted to the minimum value after Min-Max normalization. ;

[0067] Response Time: Calculates the time taken to answer the question and performs a logarithmic transformation. To eliminate the influence of long-tail distribution;

[0068] To meet the input requirements of the psychological state module, a behavioral feature vector is constructed. ;

[0069] Sequence truncation and padding: Setting the maximum sequence length For sequences that are too short, zero padding is performed; for sequences that are too long, sliding window segmentation is used.

[0070] 2. Short-term interactive status module

[0071] The feature vectors of the test questions generated by the feature encoding module are compared with the students' actual answers. (1 for correct, 0 for incorrect) Perform cross-coding to generate interactive embedding vectors. .

[0072] The `encode_inputs` function encodes the question features and the answer results into a two-channel vector. If the answer is correct, the input is... If incorrect, enter: By using LSTM to process the sequence of interactive embedding vectors, we can capture the contextual dependence of students in continuous question answering and extract short-term interactive states that represent short-term attention and contextual cognition.

[0073] The calculation formula is as follows: Calculate LSTM at time 10:00 The hidden state, this state It dynamically reflects students' recent fluctuations in learning performance, learning patterns, and cognitive status.

[0074] 3. Long-term memory state module

[0075] Graph Construction: Construct a heterogeneous graph containing "knowledge point-question" relationships. Edges are derived from the Q-matrix and the co-occurrence relationships of questions. For example, if two questions share the same knowledge point, then an edge is established between the knowledge point nodes associated with them.

[0076] Initialization: Initialize a globally persistent skill memory matrix. , representing the initial state of each knowledge point.

[0077] The neural gating mechanism is used for feature propagation and state updating. The specific implementation steps are as follows:

[0078] (1) Neural gating mechanism: For any node in the graph, the sending coefficient is calculated using the send gating network (Gate Send) and the receive gating network (Gate Recv) respectively. and acceptance coefficient :

[0079] ;

[0080] ;

[0081] (2) Dynamic message passing and aggregation: When node information is propagated, it is first controlled by the sending gate and then passes through the adjacency matrix. After weighted aggregation, the message is then filtered by the receive gate. The calculation is as follows:

[0082] ;

[0083] (3) Residual update and skill state generation: Introduce learnable residual strength parameters The strength of message passing is controlled after activation via softplus. This controls the node's update status. It is obtained by adding the original state and the weighted message, as shown in the following formula:

[0084] ;

[0085] (4) Global Long-Term Memory Update: The global skill memory matrix is ​​maintained using a momentum update strategy, the updated skill node states are extracted, and the global skill memory matrix is ​​updated using an exponential moving average strategy. :

[0086] ;

[0087] Among them, the update coefficient: Setting it to 0.05 ensures that the model is robust to long-term capability assessments and is not severely affected by abnormal knowledge state performance.

[0088] 4. Psychological state gating module

[0089] (1) Query vector construction: embed the current question into a vector using a linear layer. Corresponding long-term cognitive concept state Add them together to form the query vector. : The current test question is embedded and integrated with the short-term cognitive context provided by the short-term interaction module, with a dimension of 128.

[0090] (2) Key / Value Pair Construction: Extract the student's behavioral feature sequence (including number of attempts and answer time), and generate a key vector through a linear projection layer. Sum value vector , behavioral feature sequence Mapped to a high-dimensional representation of dimension 128;

[0091] (3) Multi-head attention calculation: Number of heads in this implementation scheme: Dimensions of each head Calculate the association weights between behavioral features and the current cognitive state, and aggregate contextual information:

[0092] ;

[0093] The outputs of multi-head attention are concatenated and fused through a linear layer to obtain a psychological feature vector. .

[0094] (4) Gating coefficient generation: The Sigmoid function is used to map psychological features into scalar coefficients.

[0095] ;

[0096] This coefficient It represents the student's level of focus or tendency to be in an abnormal state at the current moment.

[0097] 5. Dual-state modulation prediction module

[0098] Using psychological state gating coefficient For short-term interaction states respectively and long-term knowledge status The prediction results (after Q-matrix mapping) are dynamically modulated. The final student answer accuracy prediction is then calculated. The calculation formula is:

[0099] ;

[0100] ;

[0101] ;

[0102] in, As a balancing factor, a value of 0.5 is used. By employing an average fusion strategy, the prediction accuracy is improved by simultaneously considering the students' recent state fluctuations and long-term ability stability.

Claims

1. A dynamic knowledge tracking method integrating psychological state gating, characterized in that, Includes the following steps: S1. Multidimensional Feature Encoding: Extract the difficulty, discrimination, and guessing rate features of the test questions, perform linear mapping on each feature, and concatenate them into a multi-channel feature matrix. Use a two-dimensional convolutional layer to extract local correlations between features and generate test question feature vectors. ; S2. Short-term interaction state modeling: Based on the aforementioned question feature vectors and the students' answers Constructing the interaction vector of positive and negative dual-channel encoding Input into the Long Short-Term Memory network to extract short-term interaction states ; S3. Long-term memory state modeling: Construct a dynamic graph structure containing skill nodes and question nodes, perform message passing between nodes based on neural gating mechanisms to update node states, and update the global skill memory matrix through a moving average strategy. Output long-term knowledge state ; S4. Mental State Gating Generation: Based on the student's answer behavior feature sequence, a multi-head attention mechanism is used to calculate the association weight between the current question and the behavioral features, generating mental state gating coefficients. ; S5, Dual-state modulation prediction: This involves predicting the long-term knowledge state... Mapped to predicted probability Using the psychological state gating coefficient The predicted probability is multiplied and modulated, and then combined with the short-term interaction state. Weighted fusion is used to output the student's final probability of mastering the current test question. .

2. The dynamic knowledge tracking method integrating psychological state gating according to claim 1, characterized in that, The specific steps of multidimensional feature encoding in step S1 also include: The fused comprehensive feature vector of test questions The input is fed into the impact factor prediction layer, where it is mapped to a scalar through a linear transformation and normalized to the (-1, 1) interval using a hyperbolic tangent activation function. This is used to quantify the potential direction and intensity of the test questions' impact on knowledge status updates. ； The output value is compressed to the range of (-1, 1). A value close to 1 indicates that the question has a positive and high influence, a value close to 0 indicates that the question has little impact on the state update, and a value close to -1 indicates that it has a negative impact.

3. The dynamic knowledge tracking method integrating psychological state gating according to claim 1, characterized in that, The specific steps for short-term interaction state modeling in step S2 include: The difficulty, discrimination, and guessing rate features of the test questions are mapped separately and then concatenated. A two-dimensional convolutional layer is used to capture the local correlations between the features to generate the test question feature vector. Then construct the input vector for the LSTM. Its construction method involves encoding the question feature vector and the answer result using positive and negative dual-channel encoding: ； in This represents vector concatenation. ∈{0,1} represents the student's answer result.

4. The dynamic knowledge tracking method integrating psychological state gating according to claim 1, characterized in that, The specific steps for modeling long-term memory states in step S3 include: (1) Neural gating mechanism: For any node in the graph, the transmission coefficient is calculated using the transmit gating network and the receive gating network respectively. and acceptance coefficient : ； ； in: It is a node The current state; , These are the learnable parameters of the MLP; For smoothing terms; (2) Dynamic message passing and aggregation: based on adjacency matrix The information propagation between nodes is calculated, and a learnable residual strength parameter γ is introduced to update the node state: ； ； in: These are the weights of the static adjacency matrix; This represents the Hadamard product, which is formed by multiplying elements one by one. γ is the magnitude of the change in the original state by the adaptive control graph convolutional layer; (3) Global long-term memory update: Extract the updated skill node state and update the global skill memory matrix using an exponential moving average strategy. : ； ； in: It is a global skill memory matrix; It is the average skill node status of all students at the last moment; λ is the momentum coefficient, which determines the forgetting rate.

5. The dynamic knowledge tracking method integrating psychological state gating according to claim 1, characterized in that, The specific steps of the mental state gating described in step S4 include: Construct a query vector based on the current question embedding vector and short-term interaction state. ; Constructing key directions based on behavioral feature sequences AND value vector , behavioral feature sequence Mapped to a high-dimensional representation; Attention weights are calculated using a multi-head attention mechanism, and the behavioral representation is weighted and aggregated. ; Will After fusing with the current cognitive context vector, the input is fed into the Sigmoid function to generate ; Gating generation: ； ； in, For the cognitive context of the current moment, The output of multi-head attention, This is the Sigmoid function.

6. The dynamic knowledge tracking method integrating psychological state gating according to claim 1, characterized in that, The specific steps of the dual-state modulation prediction in step S5 include: ； in, For the Sigmoid function, ∈[0,1] represents the hyperparameters that balance the short-term and long-term states; As a balancing factor, its value range is , Values This means adopting an average integration strategy to simultaneously consider students' recent fluctuations in performance and their long-term ability stability.

7. A dynamic knowledge tracking method integrating psychological state gating as described in any one of claims 1-6, characterized in that, include: The feature encoding module is used to extract multi-dimensional features of test questions and construct interaction vectors; The dynamic graph neural network module is used to update the state of skill nodes based on neural gating mechanisms. The mental state calculation module is used to generate mental gating coefficients based on attention mechanisms; The dual-state prediction modulation module is used to fuse short-term states, long-term states, and psychological coefficients for probability prediction.

Images