AI personalized tutoring method based on cognitive portrait

Abstract

The invention relates to the technical field of data modeling, in particular to an AI personalized tutoring method based on cognitive portraits. Through implicit feedback dynamic weighting driven by multi-modal emotions, behavior deviation caused by task errors and emotional fluctuation is effectively distinguished, and the input quality and recommendation matching degree of potential factorization are remarkably improved; structured modeling is carried out on multi-dimensional behavior indexes by using a graph convolutional network, and interaction association among the indexes is deeply captured, so that the system can sensitively recognize subtle state changes such as user attention distraction or strategy adjustment; through self-adaptive fusion of a historical portrait and new input information, it is ensured that portrait updating is stable and timely in the process from tension to adaptation or from fatigue to recovery of a user; and finally, reliable dynamic signals are provided for the whole system through a portrait-strategy-feedback-portrait closed-loop mechanism and robustness and interpretability of multi-modal emotion estimation, so that personalized tutoring can be continuously optimized along with data accumulation.

Description

Technical Field

[0001] The present invention relates to the technical field of data modeling, and particularly to an AI personalized tutoring method based on a cognitive portrait. Background Art

[0002] Current personalized tutoring systems often use static confidence settings or weighted methods based on simple statistical values in implicit feedback modeling, making it difficult to distinguish the impact of user mood fluctuations on the authenticity of behavioral signals.

[0003] In addition, multi-dimensional behavioral features are usually treated as independent vectors or coarsely concatenated, ignoring the potential correlation structure between indicators; portrait updates mostly rely on linear mapping or basic RNN models (recurrent neural networks), unable to flexibly balance the retention of historical portraits and the fusion of new information, resulting in slow response to the state evolution of users at different learning stages; emotion estimation mostly relies on single-modal or rough fusion, lacking both robust processing of multi-source heterogeneous signals and the ability to output interpretable fine-grained emotion dimensions. Summary of the Invention

[0004] In view of the above-mentioned shortcomings of the prior art, the present invention provides an AI personalized tutoring method based on a cognitive portrait, which can effectively solve the problem in the prior art that multi-dimensional behavioral indicators are mostly independently processed or coarsely concatenated, ignoring the potential correlations between indicators and making it difficult to capture behavioral causality or structural information, resulting in the inability to generate personalized educational tutoring for users.

[0005] To achieve the above object, the present invention is realized through the following technical solutions: The present invention provides an AI personalized tutoring method based on a cognitive portrait, including at least: Obtain the learning process of the user, introduce a latent cognitive factor vector to capture the user's internal cognitive structure, and combine the explicit scoring error term and the implicit preference confidence term to establish an optimization objective function, and solve for the latent cognitive factor vector; Perform a linear transformation on the latent cognitive factor vector, apply an activation function, and splice the activation results into a cognitive portrait; Model the dynamic transition of the user's cognitive state in the time series, capturing the process from tension to adaptation and from fatigue to recovery, then there is: ; Wherein, represents the cognitive portrait of the user at time , represents the behavioral performance vector of the user at time , represents the user at time The emotional state estimation vector, represents the state transition function, represents the user at the moment The cognitive portrait vector; Define the tutoring strategy set, construct the attribute feature vector for each strategy, input the updated cognitive portrait, output the optimal tutoring strategy and execute it to the user.

[0006] Furthermore, the method for preprocessing the implicit feedback is: Preference indication matrix If , otherwise it is 0; Confidence matrix , , is the implicit feedback value, represents the hyperparameter for adjusting the confidence weight; Introduce an emotion-driven confidence dynamic adjustment mechanism: Establish an adjustment factor , represents the emotion influence intensity hyperparameter, represents the weight vector for projecting the emotional state estimation vector into the scalar space, represents the initial bias, represents the mapping function, represents the transpose; Obtain the updated .

[0007] Furthermore, the method for determining the explicit scoring error term is: According to the user For the content / resource Explicit score; Combine the user's latent cognitive factor vector, latent feature vector, and the bias vectors of both parties; Determine the explicit scoring error term to measure the gap between the model's predicted score and the true cognitive evaluation.

[0008] Furthermore, the method for determining the implicit preference confidence term is: According to the user's implicit behavior preference indication matrix for the task; Combine the confidence matrix of the interaction behavior, as well as the latent cognitive factor vector and latent feature vector; Determine the behavior-driven implicit preference confidence term to simulate the true motivation and interest tendency.

[0009] Furthermore, the method for constructing the behavior performance vector is: Construct a behavior graph; Each behavior dimension node , there are number of nodes, and the intermediate feature is calculated as: ; Among them, represents the intermediate feature vector, represents the user at time the th raw feature vector of the represents the first-layer weight matrix, represents the activation function, , respectively represent the degrees of the behavior dimension nodes , ; represents the adjacency relationship between the behavior dimension node and the node ; Stack all the intermediate features of the nodes in the node order to obtain the intermediate feature matrix; Perform a graph convolution propagation on the intermediate feature matrix and output the behavior performance vector.

[0010] Furthermore, the construction method of the emotional state estimation vector is: Collect various data modalities of the user; Extract feature vectors for each modality respectively, and then perform fusion, including: Visual channel feature vector: Input the face image tensor and perform the first-layer convolution: Extract image features through the convolution kernel weight parameters; Perform multi-layer convolution: Apply the convolution operation layer by layer to extract the deep features of the image; Reduce the size of the feature map through the pooling layer and output the visual channel vector; Input the text feature sequence and extract the language channel feature vector based on the neural network architecture of the self-attention mechanism; Input the physiological time series feature sequence and extract the physiological channel feature vector through the long short-term memory network; Map the feature vectors of each modality to the same dimension, calculate the attention weights, assign weights to each modality, and achieve multi-modal fusion to generate the emotional feature representation vector through weighted summation; Map the fused emotional feature representation vector to the specific emotional space dimension through the multi-layer perceptron and output the final emotional state estimation vector.

[0011] The technical solution provided by the present invention has the following beneficial effects compared with the known prior art: By means of implicit feedback dynamic weighting driven by multi-modal emotions, it can effectively distinguish the behavior deviations caused by "task failures" and "emotional fluctuations", and significantly improve the input quality of potential factor decomposition and the recommendation matching degree; using a graph convolutional network to perform structured modeling on multi-dimensional behavior indicators, deeply capturing the interaction correlations between indicators, enabling the system to keenly identify subtle state changes such as user attention dispersion or strategy adjustment; Through the adaptive fusion of historical portraits and new input information, it ensures that the portrait update is both stable and timely during the process of users transitioning from tension to adaptation or from fatigue to recovery; finally, through the "portrait - strategy - feedback - re-portrait" closed-loop mechanism, the robustness and interpretability of multi-modal emotion estimation provide reliable dynamic signals for the entire system, enabling personalized tutoring to be continuously optimized as data accumulates. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings described below are only some embodiments of the present invention, and those of ordinary skill in the art can obtain other drawings based on these drawings without creative efforts.

[0013] Figure 1 It is a schematic diagram of the overall method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0014] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0015] The following further describes the present invention with reference to the embodiments.

[0016] Embodiment 1 (refer to Figure 1 ): The AI personalized tutoring method based on cognitive portraits includes the following steps: The learning process of users includes explicit feedback, which reflects users' subjective evaluations (such as satisfaction scores, referring to users' subjective evaluations of content; recommendation evaluations, referring to the scores users give to recommended resources, etc.) and implicit behaviors, covering each user and depicting behavioral habits, that is, reflecting users' actual behavioral trajectories and internal motivations (such as dwell time, referring to the duration that students stay on a certain page / video; completion times, referring to the frequency of a certain resource being fully used; click / entry rate, referring to whether students click on the recommended content, etc.). Model these two in parallel to comprehensively capture users' true preferences and cognitive performances, including: Collect explicit scores, such as 1-5 star ratings, satisfaction, etc.; Collect implicit feedback, such as clicks, completion times, dwell time, etc. Perform data preprocessing on explicit scores and implicit feedback respectively, then there are: Explicit score preprocessing: Decentralization: , denotes the user 's explicit score for the content , denotes the average score, denotes the user 's initial score value for the item ; There are individual preference differences in scores (such as "some people score higher, some people score lower"). Centralization processing can eliminate the interference of scoring habits and make the matrix factorization result more accurately reflect the true preferences between users and items.

[0017] Implicit feedback preprocessing: Preference indication matrix If , otherwise 0, which describes that if the user has an interaction with the content (such as clicking, completing, etc.), then it is 1; otherwise it is 0, indicating whether the user likes the content; Confidence matrix , , is the implicit feedback value (such as click count / dwell time), denotes the hyperparameter that adjusts the confidence weight; Furthermore, by considering the influence of users' mental states on the meaning of behavioral signals, enhancing the model's sensitivity to the meaning of behaviors in different emotional states, introduce an emotion-driven confidence dynamic adjustment mechanism: Obtain the state estimation vector through multimodal fusion, see below. Therefore, design the adjustment factor , denotes the emotion influence intensity hyperparameter, which is used to control the contribution degree of emotion to confidence adjustment, It represents the weight vector that projects the emotional state estimation vector into the scalar space and is used to measure the importance of each dimension of emotion for confidence adjustment. It represents the initial bias, which is used for translational adjustment of the linear transformation to help the model give a suitable baseline output when there is no specific emotional tendency. It represents the mapping function. It represents the transpose, from which the updated is obtained. This method enables different weights to be assigned to behavioral signals when the user is in a state such as anxiety or high fatigue, so as to more accurately interpret implicit feedback in matrix factorization optimization and subsequent recommendations, improving the personalization effect and robustness.

[0018] Implicit behaviors are difficult to be directly interpreted as "like" or "dislike". By constructing a preference indication matrix and a confidence function, "interaction presence" and "interaction intensity" can be distinguished.

[0019] Furthermore, the cognitive characteristics, knowledge structure, and preference state of users are complex and not directly observable. Therefore: Introduce a latent cognitive factor vector to capture the user's internal cognitive structure, combine explicit ratings and implicit feedback, and establish an optimization objective function : ; Among them, represents the explicit rating error term, represents the implicit preference confidence term, , represent the weights of the explicit rating loss term (the degree of importance the model attaches to rating accuracy) and the implicit preference confidence term (regulating the impact of implicit behaviors on training), represents the user 's latent cognitive factor vector, which describes the distribution of this user in the cognitive space, represents the latent feature vector of the content / resource and describes the position of this content in the cognitive space, represents the regularization weight, which controls the complexity of the latent factors, represents the user latent factor matrix, which describes the latent cognitive factor vector of the th row user , represents the content latent factor matrix, which describes the latent feature vector of the th behavioral content , represents the user bias vector, represents the content bias vector; For the explicit rating error term: ; Among them, represents the set of explicit rating pairs, describing all known rating items , that is, the user has a clear rating sample for the content ; represents the rating bias term of the user , describing its overall preference rating level, represents the rating bias term of the content , describing the overall rating deviation of the content; represents transpose; By introducing the centered rating and bias terms, the individual tendency in the rating (such as a high score) is modeled to improve the accuracy of rating prediction.

[0020] For the implicit preference confidence term: ; Among them, and represent the total number of users and the total number of contents (resources) respectively. A large amount of user behavior data is not rated and labeled, but contains clues such as cognitive interest, attention flow, and ability performance. By weighted fitting the preference expression in the implicit behavior, the model can learn the potential motivation without relying on ratings, enhancing the model's adaptability to low-rating users.

[0021] Furthermore, the fusion of explicit ratings and implicit feedback leads to a complex optimization objective function. Therefore, the alternating least squares (ALS) method is suitable for large-scale matrix factorization and supports the joint optimization of explicit and implicit feedback. Then, we have: By fixing and , optimize all potential feature vectors and the rating bias term of the content , and minimize ; By fixing and , optimize all potential cognitive factor vectors and the rating bias term of the user , and minimize ; Update alternately until the objective function converges; Through the ALS alternating update strategy, solve the user and item vectors step by step to ensure the optimization of each step's closed-form solution; update the bias terms synchronously to further improve the fitting effect.

[0022] Furthermore, finally, the cognitive portrait of the user can be output based on the potential cognitive factor vector ; , specifically, a linear transformation is performed on the potential cognitive factor vector, the intermediate vector after the linear transformation is calculated, and an activation function is applied element-wise to the intermediate vector to obtain the cognitive portrait values. The cognitive portrait values are concatenated into a cognitive portrait, that is, a cognitive portrait vector, which will not be elaborated here.

[0023] For each dimension, it may include: knowledge mastery, confidence, learning motivation, attention stability, fatigue estimation, strategy usage frequency, emotional stability index, etc.

[0024] Furthermore, the cognitive portrait is non-static and usually evolves dynamically with the user's behavior and emotions. By modeling the dynamic transition of the user's cognitive state in the time series, capturing the process from tension to adaptation and from fatigue to recovery, and supporting accurate state estimation based on historical information, we have: ; Among them, represents the cognitive portrait of the user at time , represents the behavior performance vector of the user at time , represents the emotional state estimation vector of the user at time (including continuous or categorical emotion labels such as anxiety level, positivity, fatigue index, etc.), represents the state transition function, a non-linear mapping function.

[0025] In the cognitive portrait modeling of the user's learning process, behavior performance is an important information carrier that directly reflects the user's learning state and preferences. Traditional methods usually simply process the user's multi-dimensional behavior indicators (such as click count, task completion rate, switching frequency, dwell time, etc.) as independent features, ignoring the potential correlations and structures between behaviors. However, actual learning behaviors are highly correlated systematic activities, and there are causal or associative relationships between some behaviors. For example, an increase in dwell time may be related to task difficulty adjustment, and task switching frequency may reflect the user's attention dispersion or strategy adjustment, etc. If the associative information between multi-dimensional behaviors is ignored, it is easy to lead to redundant behavior information, confounding of behavior dimension effects, and failure to identify potential interactions, thereby affecting the accurate modeling of the cognitive portrait and personalized tutoring. Therefore: The graph convolutional network (GCN) jointly models the user's multi-dimensional behavior data based on the associative graph structure between behavior dimensions, so we have: Construct a behavior graph , represents the node set, describing that each node corresponds to a behavior dimension (such as task completion rate, page dwell time, switching frequency, etc.), Represents the edge set, describing the potential correlation relationships between behaviors. Represents the original adjacency matrix. For sub-elements Describes the behavior dimension nodes. And the node Whether there is an edge between them. If there is an associated edge, it is 1; otherwise, it is 0 (Pearson correlation coefficient, mutual information, or other similarity metrics can be calculated from the original feature sequences of the behavior dimension node And the node . Whether an associated edge is considered to exist is determined by comparing with a threshold). For each behavior dimension node , there are a total of number of nodes. The intermediate feature is calculated as: ; Among them, Represents the intermediate feature vector (representing the intermediate feature vector of the th behavior dimension node after the first layer propagation of the GCN). Represents the user At time The th original feature vector of the behavior dimension. Represents the weight matrix of the first layer. Represents the activation function (such as ReLU). , Respectively represent the degrees of the behavior dimension nodes , (i.e., how many connected nodes, including itself). Represents the adjacency relationship between the behavior dimension node And the node (the original adjacency matrix plus the identity matrix). By weighted aggregation of the features of each node (behavior dimension) in its adjacent nodes, the modeling of the mutual influence relationship between user behavior dimensions is realized, and the nonlinear relationship is captured through the activation function; Stack all the intermediate features of the nodes in the node order: ; Among them, Represents the intermediate feature matrix (the intermediate feature matrix of all behavior dimension nodes after the first layer propagation of the GCN). Represents the hidden layer dimension; Perform one GCN propagation on the intermediate feature matrix and output the final behavior performance vector: ; Among them, Represents a probability distribution function, which normalizes the weights for each behavior dimension to make the result a probability distribution vector, reflecting the relative importance of different behavior characteristics. Represents a diagonal matrix. Represents the learnable weight matrix of the second layer, which is used to transform the hidden features into the final behavior representation space. Represents a symmetrically normalized graph propagation matrix, which is used to prevent gradient explosion or vanishing. Represents an adjacency matrix (the original adjacency matrix plus the identity matrix (diagonal is 1 and others are 0), and the original adjacency matrix refers to the connection relationship between nodes).

[0026] By establishing the propagation path between behavior dimensions through the adjacency matrix, it can explicitly model the mutual influence relationships among behaviors such as task completion frequency, switching times, click density, etc., and overcome the problem of fragmented behavior information in traditional feature independent modeling methods.

[0027] The representation of each behavior dimension node not only depends on its own features but also integrates the feature propagation of its structural neighbors, so that the embedding of this dimension has semantic information with behavior context, enhancing the model's perception ability of the user state driven by multi-behavior intersection.

[0028] Since each layer of propagation operation implies the effects of information smoothing, noise suppression, and local feature aggregation, the final behavior representation vector can better reflect the comprehensive behavior state of the user at the current moment, enhancing the robustness and prediction accuracy of downstream cognitive state estimation and policy selection tasks.

[0029] Furthermore, considering that in an online learning or tutoring system, relying solely on a single channel (such as a self-assessment questionnaire or a single physiological signal) often cannot accurately and comprehensively capture the user's emotional state. A single modality may be affected by noise, low user initiative, device limitations, etc., and cannot fully reflect real emotional fluctuations. In addition, in different behavior scenarios, users may not have obvious facial expressions or may be inconvenient to fill out questionnaires, while only behavior signals or physiological signals can still provide clues. Moreover, emotions have complex multi-dimensional attributes (such as anxiety, fatigue, pleasure, activation level, etc.), and a single channel is difficult to cover simultaneously. Therefore, through the collaborative judgment of multi-modal and multi-source information, the accuracy and robustness of emotion estimation are improved, misjudgment or omission is reduced, and the determination of the emotion state estimation vector is achieved, including: Collecting multiple data modalities such as the user's facial images, speech / text, and physiological or behavior signals, and each modality has different acquisition methods, noise characteristics, and dimensions. Perform preprocessing operations such as face detection and alignment, image normalization, text cleaning and tokenization, speech recognition or acoustic feature extraction, sensor signal filtering and standardization. Through preprocessing, eliminate interference, unify the scale, and enhance the robustness and generalization ability of the emotion estimation model for subsequent feature input; Extract high-quality feature vectors for each modality respectively, and then perform fusion, then there is: For the visual channel feature vector : Input user At the moment The face image tensor collected , perform the first layer of convolution: ; Among them, Represents the value of the first layer output feature map at the spatial position , the th output channel value, Represents the number of input channels, , Respectively represent the radius parameters of the convolution kernel in the height and width directions, Represents the convolution kernel weight parameter in the first layer convolution kernel, Represents the bias vector of the first layer and the th output channel, , Represents the corresponding spatial position on , subsequent multi-layer convolution: ; Among them, Represents the th layer output feature map, Represents the th layer convolution kernel weight tensor, Represents the th layer bias vector, Represents the convolution operation; Pooling / global average pooling to obtain the visual channel vector , Represents capturing facial expressions, Represents the th layer output feature map; Input the text feature sequence, that is, a piece of text of the user at the current moment or during a certain interaction, such as the answer content, chat record, comment, etc., and extract the language channel feature vector based on the neural network architecture (Transformer) of the self-attention mechanism to capture the language emotion; Input the physiological time series feature sequence of the user at consecutive time steps, including the physiological feature vector of the user at the time step. Extract the physiological channel feature vector through the Long Short-Term Memory Network (LSTM) to capture the physiological emotion. These two methods of extracting feature vectors are well-known and will not be elaborated here.

[0030] Therefore, map the high-quality feature vectors of each modality to a unified dimension. By calculating the attention weights, since different modalities contribute differently to emotions, weights are assigned to each modality. The three modalities are fused through weighted summation to generate an emotion feature representation vector. The contribution of different modalities to the user's emotion varies with the scenario and the quality of real-time acquisition. For example, when the light is insufficient, the quality of the visual channel may decline, and the language channel or the physiological channel may be more reliable. By calculating the attention weights to dynamically adjust the proportion of each modality in the final vector, both the respective advantages can be retained, and the noise impact of weak modalities can be suppressed. In addition, multi-modal fusion also allows capturing the interaction features between modalities, such as the linkage between facial expressions and tone of voice, and the temporal coupling between physiological signals and emotional expressions, thereby obtaining more comprehensive and robust emotion features.

[0031] Specifically, since the fused emotion feature representation vector is a comprehensive representation of multi-modal features and has a high dimension and does not have directly interpretable emotion meanings, map the fused emotion feature representation vector to the specific emotion space dimension (such as anxiety level, fatigue level, pleasure level, etc.) through the Multi-Layer Perceptron (MPL) to output the final emotion state estimation vector. In order to accurately estimate the cognitive portrait (reflecting the mental state), the system can perform precise intervention and guidance based on the current emotion state. The mapping method of the Multi-Layer Perceptron (MPL) is well-known and will not be elaborated here.

[0032] Finally, define the set of all tutoring strategies that the system can execute, and each strategy in it represents a specific tutoring measure, for example: Push low-difficulty exercises, push high-challenge tasks; Give detailed video explanations of problem-solving, provide demonstration cases; Give positive encouragement or feedback prompts; Arrange a quiz to detect mastery; Suggest rest or relaxing thinking activities; Organize peer discussions or collaboration groups; Adjust the order of the learning path, etc.

[0033] Construct an attribute feature vector (describing the attributes of the strategy) for each strategy ; Input the updated cognitive portrait and call the supervised learning model to predict the expected effects of each strategy under this portrait. Multiple-objective comprehensive scores can be calculated by combining other indicators (such as acceptance, cognitive cost, emotional conflict, long-term value, etc.). The optimal tutoring strategy is output according to the score ranking and executed for the user. After the user executes it, new behavioral / emotional feedback is collected for the next round of update and fine-tuning of the mapping model, thus forming a closed loop of portrait update → strategy prediction → execution feedback → portrait update to achieve personalized auxiliary guidance for education.

[0034] A personalized tutoring system, applied to the AI personalized tutoring method based on cognitive portraits, includes: An objective function establishment module that introduces a latent cognitive factor vector to capture the user's internal cognitive structure, combines an explicit scoring error term and an implicit preference confidence term, establishes an optimization objective function, and solves for the latent cognitive factor vector; A cognitive portrait construction module that performs a linear transformation on the latent cognitive factor vector, applies an activation function, and splices the activation results into a cognitive portrait; A cognitive portrait update module that models the dynamic transition of the user's cognitive state in the time series, captures the process from tension to adaptation and from fatigue to recovery, and realizes the update of the cognitive portrait; A tutoring strategy output module that defines a set of tutoring strategies, constructs an attribute feature vector for each strategy, inputs the updated cognitive portrait, outputs the optimal tutoring strategy, and executes it for the user.

[0035] A computer device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of the method are implemented.

[0036] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the method are implemented.

[0037] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements will not cause the essence of the corresponding technical solutions to deviate from the protection scope of the technical solutions of the embodiments of the present invention.

Claims

1. An AI personalized tutoring method based on a cognitive portrait, characterized in that It includes the following steps: Obtain the user's learning process, introduce a latent cognitive factor vector to capture the user's internal cognitive structure, combine the explicit rating error term and the implicit preference confidence term, establish an optimization objective function, and solve for the latent cognitive factor vector; Perform a linear transformation on the latent cognitive factor vector, apply an activation function, and concatenate the activation results into a cognitive portrait; Model the dynamic transition of the user's cognitive state in the time series, capture the process from tension to adaptation and from fatigue to recovery, then there is: ； Among them, represents the cognitive portrait of the user at the moment ; represents the behavior performance vector of the user at the moment ; represents the estimated vector of the emotional state of the user at the moment ; represents the state transition function, represents the cognitive portrait vector of the user at the moment ; Define a set of tutoring strategies, construct an attribute feature vector for each strategy, input the updated cognitive portrait, output the optimal tutoring strategy and execute it for the user.

2. The AI personalized tutoring method based on cognitive portraits according to claim 1, wherein, The method for preprocessing the implicit feedback is: Preference indication matrix If , otherwise 0; Confidence matrix , , are implicit feedback values, denote hyperparameters for adjusting confidence weights; Introduce an emotion-driven confidence dynamic adjustment mechanism: Establish adjustment factor , denotes the hyperparameter of emotion influence intensity, denotes the weight vector that projects the emotion state estimation vector into the scalar space, denotes the initial bias, denotes the mapping function, denotes the transpose; Get updated .

3. The AI personalized tutoring method based on cognitive portraits according to claim 1, wherein The method for determining the explicit rating error term is: According to the user for content / resources explicit rating Combine the latent cognitive factor vector, latent feature vector of this user, and the bias vectors of both sides; Determine the explicit rating error term to measure the gap between the model-predicted rating and the true cognitive evaluation.

4. The AI personalized tutoring method based on a cognitive profile according to claim 1, wherein The method for determining the implicit preference confidence term is: According to the user's implicit behavior preference indication matrix for tasks; Combine the confidence matrix of interaction behaviors, as well as the latent cognitive factor vector and latent feature vector; Determine the behavior-driven implicit preference confidence term to simulate the true motivation and interest tendency.

5. The AI personalized tutoring method based on cognitive portraits according to claim 1, characterized in that The method for constructing the behavior performance vector is: Construct a behavior graph; Each behavioral dimension node , with a total of number of nodes, and the intermediate feature calculation is as follows: ； Among them, represents the intermediate feature vector, represents the user at the moment the original feature vector of the represents the first-layer weight matrix, represents the activation function, and respectively represent the degrees of the behavior dimension nodes and ; represents the adjacency relationship between the behavior dimension node and the node ; Stack the intermediate features of all nodes in the node order to obtain an intermediate feature matrix; Perform one graph convolution propagation on the intermediate feature matrix and output the behavior performance vector.

6. The AI personalized tutoring method based on cognitive portraits according to claim 1, wherein, The method for constructing the emotion state estimation vector is: Collect multiple data modalities of the user; Extract feature vectors for each modality respectively, and then perform fusion, including: Visual channel feature vector : Input face image tensor , perform the first layer of convolution: Extract image features through convolutional kernel weight parameters; Perform multi-layer convolution: apply convolution operations layer by layer to extract the deep features of the image; Reduce the size of the feature map through a pooling layer and output the visual channel vector; Input the text feature sequence, and extract the language channel feature vector based on the neural network architecture of the self-attention mechanism; Input the physiological time series feature sequence and extract the physiological channel feature vector through a long short-term memory network; Map the feature vectors of each modality to the same dimension, calculate the attention weights, assign weights to each modality, and achieve multi-modal fusion through weighted summation to generate an emotion feature representation vector; Map the fused emotion feature representation vector to the specific emotion space dimension through a multi-layer perceptron and output the final emotion state estimation vector.

7. The AI personalized tutoring method based on cognitive portraits according to claim 6, wherein The expression for extracting image features through convolutional kernel weight parameters is: ； Among them, represents the value of the output feature map of the first layer at the spatial position , the -th value of the output channel, represents the number of input channels, , respectively represent the radius parameters of the convolutional kernel in the height and width directions, represents the convolutional kernel weight parameter in the first layer of convolutional kernels, represents the bias vector of the -th output channel of the first layer, , represents the corresponding spatial position at .

8. A personalized tutoring system applied to the AI personalized tutoring method based on a cognitive portrait according to any one of claims 1-7, characterized in that, It includes: An objective function establishment module, which introduces a latent cognitive factor vector to capture the user's internal cognitive structure, combines the explicit rating error term and the implicit preference confidence term, establishes an optimization objective function, and solves for the latent cognitive factor vector; A cognitive portrait construction module, which performs a linear transformation on the latent cognitive factor vector, applies an activation function, and concatenates the activation results into a cognitive portrait; A cognitive portrait update module, which models the dynamic transition of the user's cognitive state in the time series, captures the process from tension to adaptation and from fatigue to recovery, and realizes the update of the cognitive portrait; The tutoring strategy output module defines a set of tutoring strategies, constructs an attribute feature vector for each strategy, inputs the updated cognitive profile, outputs the optimal tutoring strategy, and executes it for the user.

9. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method described in any one of claims 1 to 7.

Images