AI personalized tutoring method based on cognitive portrait

Abstract

The present invention relates to the field of data modeling technology, and specifically to an AI personalized tutoring method based on cognitive profiling. Through the dynamic weighting of implicit feedback driven by multimodal emotions, it effectively distinguishes between behavioral deviations caused by task errors and emotional fluctuations, and significantly improves the input quality and recommendation matching degree of potential factor decomposition; uses graph convolutional networks to perform structured modeling on multidimensional behavioral indicators, deeply captures the interactive correlation between indicators, and enables the system to keenly identify subtle state changes such as user distraction or strategy adjustment; through the adaptive fusion of historical portraits and new input information, it ensures that the portrait update is stable and timely in the process of users going 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 multimodal emotion estimation provide reliable dynamic signals for the entire system, so that personalized tutoring can be continuously optimized as data accumulates.

Description

Technical Field

[0001] The present invention relates to the field of data modeling technology, and in particular to an AI personalized tutoring method based on cognitive profiling. Background Art

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

[0003] In addition, multi-dimensional behavioral features are usually treated as independent vectors or coarse-grained splicing, ignoring the potential correlation structure between various indicators; portrait updates mostly rely on linear mapping or basic RNN models (recurrent neural networks), which cannot flexibly balance the retention of historical portraits and the integration of new information, resulting in slow response to the state evolution of users at different learning stages; emotion estimation mostly relies on a single modality or rough fusion, lacking robust processing of multi-source heterogeneous signals and making it difficult to output interpretable fine-grained emotion dimensions. Summary of the Invention

[0004] In response to the above-mentioned shortcomings of the existing technology, the present invention provides an AI personalized tutoring method based on cognitive profiling, which can effectively solve the problem in the existing technology that multidimensional behavioral indicators are mostly processed independently or coarse-grained, ignoring the potential correlation between indicators, making it difficult to capture behavioral causal or structural information, resulting in the inability to generate personalized educational tutoring for users.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions:

[0006] The present invention provides an AI personalized tutoring method based on cognitive profiling, which at least includes:

[0007] Obtain the user's learning process, introduce the latent cognitive factor vector to capture the user's internal cognitive structure, and combine the explicit rating error term and the implicit preference confidence term to establish an optimization objective function to solve the latent cognitive factor vector;

[0008] Perform a linear transformation on the latent cognitive factor vector and apply an activation function to splice the activation results into a cognitive profile;

[0009] Modeling the dynamic transition of user cognitive states in time series, capturing the process from tension to adaptation, from fatigue to recovery, we have:

[0010] ;

[0011] in, Represents a user At the moment The cognitive portrait below, Represents a user At the moment The behavior performance vector, Represents a user At the moment The estimated emotional state vector of represents the state transition function, Represents a user At the moment The cognitive portrait vector under

[0012] Define a set of coaching strategies, construct an attribute feature vector for each strategy, input the updated cognitive profile, output the optimal coaching strategy and execute it to the user.

[0013] Furthermore, the method for preprocessing the implicit feedback is:

[0014] Preference indicator matrix like , otherwise 0;

[0015] Confidence Matrix , , is the implicit feedback value, represents the hyperparameter for adjusting the confidence weight;

[0016] Introducing a sentiment-driven confidence dynamic adjustment mechanism:

[0017] Establishing adjustment factors , represents the emotional impact intensity hyperparameter, represents the weight vector that projects the emotional state estimation vector into the scalar space, represents the initial bias, represents the mapping function, represents transpose;

[0018] Get updated .

[0019] Furthermore, the method for determining the explicit scoring error term is:

[0020] According to the user About content / resources Explicit rating of

[0021] Combine the user's latent cognitive factor vector, latent feature vector, and the bias vectors of both parties;

[0022] Determine the explicit rating error term to measure the gap between the model's predicted rating and the true cognitive evaluation.

[0023] Furthermore, the implicit preference confidence item is determined by:

[0024] According to the user's implicit behavioral preference indicator matrix for the task;

[0025] Combine the confidence matrix of the interactive behavior, as well as the latent cognitive factor vector and latent feature vector;

[0026] Determine the implicit preference confidence items driven by behavior and simulate real motivations and interest trends.

[0027] Furthermore, the method for constructing the behavior performance vector is:

[0028] Constructing behavioral graphs;

[0029] Each behavior dimension node , a total of The number of nodes, the intermediate features are calculated as:

[0030] ;

[0031] in, represents the intermediate eigenvector, Represents a user At the moment No. The original feature vector of the behavioral dimension, represents the first layer weight matrix, represents the activation function, 、 Represents the behavior dimension nodes 、 The degree, Represents a behavioral dimension node With node The adjacency relationship between them;

[0032] Stack all node intermediate features in node order to obtain the intermediate feature matrix;

[0033] Perform a graph convolution propagation on the intermediate feature matrix and output the behavior performance vector.

[0034] Furthermore, the emotional state estimation vector is constructed as follows:

[0035] Collect multiple data modalities of users;

[0036] Each modality extracts feature vectors separately and then performs fusion, including:

[0037] Visual channel feature vector:

[0038] Input the face image tensor and perform the first layer of convolution:

[0039] Extract image features through convolution kernel weight parameters;

[0040] Perform multi-layer convolution: Apply convolution operations layer by layer to extract deep features of the image;

[0041] Reduce the size of the feature map through the pooling layer and output the visual channel vector;

[0042] Input text feature sequence, and extract language channel feature vector based on the neural network architecture of self-attention mechanism;

[0043] Input physiological time series feature sequence and extract physiological channel feature vector through long short-term memory network;

[0044] Map the feature vector of each modality to the same dimension, assign weights to each modality by calculating attention weights, and achieve multimodal fusion through weighted summation to generate an emotion feature representation vector;

[0045] The fused emotion feature representation vector is mapped to a specific emotion space dimension through a multi-layer perceptron, and the final emotion state estimation vector is output.

[0046] Compared with the prior art, the technical solution provided by the present invention has the following beneficial effects:

[0047] Through dynamic weighting of implicit feedback driven by multimodal emotions, we can effectively distinguish between behavioral deviations caused by "task failure" and "emotional fluctuations", significantly improving the input quality and recommendation matching of latent factor decomposition. Using graph convolutional networks to perform structured modeling of multidimensional behavioral indicators, we can deeply capture the interactive correlations between indicators, enabling the system to keenly identify subtle state changes such as user distraction or strategy adjustments.

[0048] Through the adaptive fusion of historical portraits and new input information, we ensure that the portrait updates are both stable and timely as the user transitions 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 multimodal emotion estimation provide reliable dynamic signals for the entire system, allowing personalized tutoring to be continuously optimized as data accumulates. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. Those skilled in the art can also derive other drawings based on these drawings without inventive effort.

[0050] Figure 1 Schematic diagram of the overall method of the present invention. DETAILED DESCRIPTION

[0051] To make the purpose, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0052] The present invention will be further described below with reference to the embodiments.

[0053] Example 1 (see Figure 1 ): The AI ​​personalized tutoring method based on cognitive profiling includes the following steps:

[0054] The user learning process includes explicit feedback, reflecting subjective evaluations (e.g., satisfaction ratings, which refer to users' subjective evaluations of content; recommendation ratings, which refer to users' ratings of recommended resources) and implicit behaviors. This covers every user and depicts behavioral habits, reflecting their actual behavioral trajectories and intrinsic motivations (e.g., dwell time, which refers to how long students stay on a page or video; completion counts, which refer to how often a resource is fully used; and click-through / entry rates, which refer to whether students click on recommended content). Modeling these two aspects in parallel fully captures users' true preferences and cognitive performance, including:

[0055] Collect explicit ratings, such as 1-5 star reviews, satisfaction, etc.

[0056] Collect implicit feedback, such as clicks, completion times, dwell time, etc.

[0057] Data preprocessing is performed on explicit ratings and implicit feedback respectively, and we have:

[0058] Explicit scoring preprocessing:

[0059] Decentralization: , Represents a user About content Explicit rating of represents the average rating. Represents a user About the project The initial score value of

[0060] There are individual preference differences in ratings (for example, some people rate higher, while others rate lower). Centralized processing can eliminate the interference of rating habits, allowing the matrix decomposition results to more accurately reflect the true preferences between users and items.

[0061] Implicit feedback preprocessing:

[0062] Preference indicator matrix like , otherwise 0, describing if the user and content If there is interaction (such as click, completion, etc.), it is 1; otherwise it is 0, indicating whether the content is liked;

[0063] Confidence Matrix , , is an implicit feedback value (such as click count / stay time), represents the hyperparameter for adjusting the confidence weight;

[0064] Furthermore, by considering the impact of the user's psychological state on the meaning of behavioral signals, the model's sensitivity to the meaning of behaviors under different emotional states is enhanced, and an emotion-driven confidence dynamic adjustment mechanism is introduced:

[0065] The state estimation vector is obtained by multimodal fusion, see below, so the design adjustment factor , It represents the emotional impact intensity hyperparameter, which is used to control the contribution of emotions to confidence adjustment. Represents the weight vector that projects the emotional state estimation vector into the scalar space, which is used to measure the importance of each dimension of emotion to the confidence adjustment. Represents the initial bias, which is used to perform translation adjustment on the linear transformation to help the model give a suitable baseline output when there is no specific emotional tendency. represents the mapping function, Represents transposition, thus obtaining the updated This method assigns different weights to behavioral signals when the user is in a state of anxiety, high fatigue, etc., thereby more accurately interpreting implicit feedback in matrix decomposition optimization and subsequent recommendations, and improving personalization effects and robustness.

[0066] Implicit behavior is difficult to directly interpret as "like" or "dislike". By constructing a preference indicator matrix and a confidence function, we can distinguish between "interaction existence" and "interaction intensity".

[0067] Furthermore, users’ cognitive characteristics, knowledge structure, and preference status are complex and cannot be directly observed. Therefore:

[0068] The latent cognitive factor vector is introduced to capture the user's internal cognitive structure, and the explicit rating and implicit feedback are combined to establish the optimization objective function. :

[0069] ;

[0070] in, represents the explicit scoring error term, represents the implicit preference confidence term, 、 represents the weight of the explicit rating loss term (how much the model values ​​the rating accuracy), the weight of the implicit preference confidence term (regulating the impact of implicit behavior on training), Represents a user The latent cognitive factor vector of describes the user’s distribution in the cognitive space. Represents content / resources The latent feature vector of , describing the position of the content in the cognitive space, represents the regularization weight, controlling the complexity of the latent factor, Represents the user's latent factor matrix, describing the Line User The latent cognitive factor vector of Represents the content latent factor matrix, describing the Behavior content The latent feature vector of represents the user bias vector, represents the content bias vector;

[0071] For an explicit scoring error term:

[0072] ;

[0073] in, Represents a set of explicit rating pairs, describing all known rated items , that is, the user About content There are samples with clear scores, Represents a user The rating bias term describes its overall preference rating level, Display content The rating bias term describes the overall rating deviation of the content. represents transpose;

[0074] By introducing centralized scoring and bias terms, we can model individual tendencies in scoring (such as over-scoring) and improve the accuracy of score predictions.

[0075] For implicit preference confidence items:

[0076] ;

[0077] in, 、 They represent the total number of users and the total number of contents (resources) respectively. A large amount of user behavioral data is not labeled with ratings, but contains clues such as cognitive interests, attention flow, and ability performance. By weighted fitting the preference expression in implicit behavior, the model can learn potential motivations without relying on ratings, thereby enhancing the model's adaptability to low-rated users.

[0078] Furthermore, the fusion of explicit scoring 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 joint optimization of explicit and implicit feedback.

[0079] By fixing 、 , optimize all potential feature vectors and content The scoring bias term is minimized ;

[0080] By fixing 、 , optimize all potential cognitive factor vectors and users The scoring bias term is minimized ;

[0081] Alternate iterative updates until the objective function converges;

[0082] Through the ALS alternating update strategy, the user and item vectors are solved step by step to ensure the closed-form solution optimization at each step; the bias term is updated synchronously to further improve the fitting effect.

[0083] Then, finally, we can use the latent cognitive factor vector Output user cognitive portrait Specifically, the latent cognitive factor vector is linearly transformed, the intermediate vector after the linear transformation is calculated, the activation function is applied to the intermediate vector element by element to obtain the cognitive portrait value, and the cognitive portrait values ​​are spliced ​​into a cognitive portrait, that is, a cognitive portrait vector, which will not be repeated here.

[0084] Each dimension may include: knowledge mastery, self-confidence, learning motivation, attention stability, fatigue estimation, strategy use frequency, emotional stability index, etc.

[0085] Furthermore, cognitive profiles are not static and typically evolve dynamically with user behavior and emotions. Modeling the dynamic transition of user cognitive states in a time series captures the process from tension to adaptation, from fatigue to recovery, and supports accurate state estimation based on historical information. This allows for:

[0086] ;

[0087] in, Represents a user At the moment The cognitive portrait below, Represents a user At the moment The behavior performance vector, Represents a user At the moment The emotional state estimation vector (including continuous or categorical emotional labels such as anxiety level, positivity, fatigue index, etc.), Represents state transfer function, nonlinear mapping function.

[0088] In cognitive profiling of the user learning process, behavioral performance is an important information carrier that directly reflects the user's learning status and preferences. Traditional methods often simply treat users' multidimensional behavioral indicators (such as number of clicks, task completion rate, switching frequency, dwell time, etc.) as independent features, ignoring the potential correlation and structure between behaviors. However, actual learning behavior is a highly correlated and systematic activity, and certain behaviors have causal or correlational relationships. For example, an increase in dwell time may be related to an adjustment in task difficulty, and the frequency of task switching may reflect the user's distraction or strategy adjustment. If the correlation information between multidimensional behaviors is ignored, it can easily lead to behavioral information redundancy, overlapping of behavioral dimension influences, and the inability to identify potential interactions, which in turn affects the accurate modeling of cognitive profiles and personalized tutoring. Therefore:

[0089] The Graph Convolutional Network (GCN) jointly models user multi-dimensional behavior data based on the association graph structure between behavior dimensions, which gives:

[0090] Building a Behavior Diagram , Represents a node set, describing each node corresponding to a behavioral dimension (such as task completion rate, page dwell time, switching frequency, etc.), Represents an edge set, describing the potential correlation between behaviors, represents the original adjacency matrix, , for child elements , describing the behavior dimension node With node Is there an edge between them? If there is an associated edge, it is 1, otherwise it is 0 (can be obtained from large-scale user behavior logs, for behavior dimension nodes With node The original feature sequence is used to calculate the Pearson correlation coefficient, mutual information or other similarity metrics, and the comparison threshold is used to determine whether there is an associated edge.

[0091] For each behavior dimension node , a total of The number of nodes, the intermediate features are calculated as:

[0092] ;

[0093] in, represents the intermediate eigenvector (representing the The intermediate feature vector of the behavior dimension node after propagation in the first layer of GCN), Represents a user At the moment No. The original feature vector of the behavioral dimension, represents the first layer weight matrix, represents the activation function (e.g. ReLU), 、 Represents the behavior dimension nodes 、 The degree of (i.e., how many connected nodes it has, including itself), Represents a behavioral dimension node With node The adjacency relationship between them (the original adjacency matrix plus the identity matrix) is calculated by weighting the features of each node (behavior dimension) on its adjacent nodes to model the mutual influence relationship between user behavior dimensions and capture nonlinear relationships through activation functions.

[0094] Stack all node intermediate features in node order:

[0095] ;

[0096] in, represents the intermediate feature matrix (the intermediate feature matrix of all behavioral dimension nodes after the first layer of GCN propagation), represents the hidden layer dimension;

[0097] Perform a GCN propagation on the intermediate feature matrix and output the final behavior performance vector:

[0098] ;

[0099] in, Represents the probability distribution function, normalizes the weight of each behavior dimension, and makes the result a probability distribution vector, which reflects 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 the symmetric normalized graph propagation matrix, which is used to prevent gradient explosion or disappearance. Represents the adjacency matrix (the original adjacency matrix plus the identity matrix (diagonal is 1, the rest is 0). The original adjacency matrix refers to the connection relationship between nodes).

[0100] By establishing the propagation path between behavioral dimensions through the adjacency matrix, we can explicitly model the mutual influence relationship between behaviors such as task completion frequency, number of switches, and click density, overcoming the problem of behavioral information fragmentation in traditional feature-independent modeling methods.

[0101] The representation of each behavior dimension node not only relies on its own features, but also integrates the feature propagation of its structural neighbors, so that the dimension is embedded with semantic information with behavioral context, improving the model's ability to perceive user status driven by multiple behaviors.

[0102] Since each layer of propagation operation implicitly contains the effects of information smoothing, noise suppression, and local feature aggregation, the final behavior performance vector can better reflect the user's comprehensive behavioral state at the current moment, enhancing the robustness and prediction accuracy of downstream cognitive state estimation and strategy selection tasks.

[0103] Furthermore, considering that in online learning or tutoring systems, relying solely on a single channel (such as a self-assessment questionnaire or a single physiological signal) often fails to accurately and comprehensively capture the user's emotional state. A single modality may be affected by noise, low user cooperation, equipment limitations, etc., and cannot fully reflect real emotional fluctuations. In addition, in different behavioral scenarios, users may not have obvious facial expressions or be inconvenient to fill out questionnaires, and only behavioral signals or physiological signals can still provide clues. Furthermore, emotions have complex multidimensional attributes (such as anxiety, fatigue, pleasure, activation, etc.), which are difficult to cover simultaneously with a single channel. Therefore, through the collaborative judgment of multimodal and multi-source information, the accuracy and robustness of emotion estimation can be improved, misjudgment or omission can be reduced, and the emotional state estimation vector can be realized. Determination includes:

[0104] Collect multiple data modalities such as user facial images, voice / text, and physiological or behavioral signals. Each modality has different acquisition methods, noise characteristics, and dimensions.

[0105] Perform preprocessing operations such as face detection and alignment, image normalization, text cleaning and word segmentation, speech recognition or acoustic feature extraction, sensor signal filtering and normalization. Preprocessing eliminates interference, unifies scales, and serves as subsequent feature input, enhancing the robustness and generalization ability of the emotion estimation model.

[0106] Extract high-quality feature vectors from each modality and then fuse them, and we have:

[0107] For the visual channel feature vector :

[0108] Enter User At the moment The collected face image tensor , perform the first convolution:

[0109] ;

[0110] in, Indicates the spatial location of the first layer output feature map , No. The value of the output channel, Indicates the number of input channels, 、 Represent the radius parameters of the convolution kernel in the height and width directions respectively, Represents the convolution kernel weight parameter in the first layer of convolution kernel, Indicates the first layer The bias vector of the output channels, , Indicates The corresponding spatial position, subsequent multi-layer convolution:

[0111] ;

[0112] in, Indicates the Layer output feature map, Indicates the Layer convolution kernel weight tensor, Indicates the layer bias vector, Represents the convolution operation;

[0113] Pooling / global average pooling to obtain the visual channel vector , Capturing facial expressions, Indicates the The layer outputs a feature map;

[0114] The input text feature sequence is a segment of text from the user at the current moment or in a certain interaction, such as answers, chat logs, comments, etc. The self-attention-based neural network architecture (Transformer) extracts language channel feature vectors to capture language sentiment.

[0115] The user's physiological temporal feature sequence at consecutive time steps is input, including the user's physiological feature vector at the time step, and the physiological channel feature vector is extracted through the long short-term memory network (LSTM) to capture physiological emotions. The extraction of feature vectors by these two methods is well known and will not be described in detail here.

[0116] Therefore, the high-quality feature vectors of each modality are mapped to a unified dimension. By calculating the attention weight, each modality is assigned a weight because different modalities contribute differently to emotions. The three modalities are then weighted and summed to achieve multimodal fusion to generate an emotion feature representation vector.

[0117] The contribution of different modalities to user emotions varies depending on the scenario and the quality of real-time acquisition. For example, in low-light conditions, the visual channel may degrade, while the language or physiological channels may be more reliable. By calculating attention weights and dynamically adjusting the weight of each modality in the final vector, we can preserve the strengths of each modality while suppressing the noise impact of weaker modalities. Furthermore, multimodal fusion allows for the capture of interactive features between modalities, such as the linkage between facial expressions and tone of voice, and the temporal coupling between physiological signals and emotional expressions, resulting in more comprehensive and robust emotional features.

[0118] Specifically, because the fused emotion feature representation vector is a comprehensive representation of multimodal features, it has a high dimension and does not have a directly interpretable emotional meaning. The fused emotion feature representation vector is mapped to a specific emotion space dimension (anxiety, fatigue, pleasure, etc.) through the multi-layer perceptron MPL, and the final emotion state estimation vector is output. In order to accurately estimate the cognitive portrait (reflecting the psychological state), the system can accurately intervene and guide based on the current emotional state; the multi-layer perceptron MPL mapping method is well known and will not be described here.

[0119] Finally, define the set of all coaching strategies that the system can execute, and each strategy within it Represents a specific counseling measure, such as:

[0120] Push low-difficulty exercises and push high-challenge tasks;

[0121] Provide detailed problem-solving video explanations and demonstration cases;

[0122] Provide positive encouragement or feedback;

[0123] Arrange quizzes to test mastery;

[0124] Suggesting a break or light-hearted thinking activity;

[0125] Organize peer discussion or collaboration groups;

[0126] Adjust the order of learning paths, etc.

[0127] Construct an attribute feature vector for each strategy (describing the attributes of the strategy) ;

[0128] Enter an updated cognitive profile , and call the supervised learning model to predict the expected effect of each strategy under this portrait;

[0129] It can be combined with other indicators (such as acceptance, cognitive cost, emotional conflict, long-term value, etc.) to make a multi-target comprehensive score, output the optimal counseling strategy according to the score sorting and execute it to the user. After the user executes it, new behavioral / emotional feedback is collected for the next round of updates and mapping model fine-tuning, thus forming a closed loop of portrait update → strategy prediction → execution feedback → portrait update, realizing personalized auxiliary guidance in education.

[0130] The personalized tutoring system, applied to the aforementioned AI personalized tutoring method based on cognitive profiling, includes:

[0131] The objective function establishment module introduces the latent cognitive factor vector to capture the user's internal cognitive structure, and combines the explicit rating error term and the implicit preference confidence term to establish an optimization objective function and solve the latent cognitive factor vector;

[0132] The cognitive profile construction module linearly transforms the latent cognitive factor vector, applies an activation function, and splices the activation results into a cognitive profile;

[0133] The cognitive profile update module models the dynamic transition of the user's cognitive state in a time series, capturing the process from tension to adaptation and from fatigue to recovery, and thus updating the cognitive profile.

[0134] The coaching strategy output module defines a coaching strategy set, constructs an attribute feature vector for each strategy, inputs an updated cognitive profile, outputs the optimal coaching strategy, and executes it to the user.

[0135] A computer device includes a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the method when executing the computer program.

[0136] A computer-readable storage medium stores a computer program, which implements the steps of the method when executed by a processor.

[0137] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, 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 various embodiments of the present invention.

Claims

1. The AI ​​personalized tutoring method based on cognitive profiling is characterized by: The steps include: The user's learning process is obtained, and the latent cognitive factor vector is introduced to capture the user's internal cognitive structure. The explicit rating error term and the implicit preference confidence term are combined to establish an optimization objective function to solve the latent cognitive factor vector. Specifically, The latent cognitive factor vector is introduced to capture the user's internal cognitive structure, and the explicit rating and implicit feedback are combined to establish the optimization objective function. ; ； in, represents the explicit scoring error term, represents the implicit preference confidence term, 、 represents the weight of the explicit rating loss term and the weight of the implicit preference confidence term, Represents a user The latent cognitive factor vector of describes the user’s distribution in the cognitive space. Represents content / resources The latent feature vector of , describing the position of the content in the cognitive space, represents the regularization weight, controlling the complexity of the latent factor, Represents the user's latent factor matrix, describing the Line User The latent cognitive factor vector of Represents the content latent factor matrix, describing the Behavior content The latent feature vector of represents the user bias vector, represents the content bias vector; Perform a linear transformation on the latent cognitive factor vector and apply an activation function to splice the activation results into a cognitive profile; Modeling the dynamic transition of user cognitive states in time series, capturing the process from tension to adaptation, from fatigue to recovery, we have: ； in, Represents a user At the moment The cognitive portrait below, Represents a user At the moment The behavior performance vector, Represents a user At the moment The estimated emotional state vector of represents the state transition function, Represents a user At the moment The cognitive portrait vector under Define a set of coaching strategies, construct an attribute feature vector for each strategy, input the updated cognitive profile, output the optimal coaching strategy, and execute it on the user; Preprocess the implicit feedback, specifically: Preference indicator matrix ; Confidence Matrix , is the implicit feedback value, represents the hyperparameter for adjusting the confidence weight; Introducing a sentiment-driven confidence dynamic adjustment mechanism: Establishing adjustment factors , represents the emotional impact intensity hyperparameter, represents the weight vector that projects the emotional state estimation vector into the scalar space, represents the initial bias, represents the mapping function, represents transpose; Get updated .

2. The AI ​​personalized tutoring method based on cognitive profiling according to claim 1 is characterized in that: The method for determining the explicit scoring error term is: According to the user About content / resources Explicit rating of Combine the user's latent cognitive factor vector, latent feature vector, and the bias vectors of both parties; Determine the explicit rating error term to measure the gap between the model's predicted rating and the true cognitive evaluation.

3. The AI ​​personalized tutoring method based on cognitive profiling according to claim 1 is characterized in that: The method for determining the implicit preference confidence item is: According to the user's implicit behavioral preference indicator matrix for the task; Combine the confidence matrix of the interactive behavior, as well as the latent cognitive factor vector and latent feature vector; Determine the implicit preference confidence items driven by behavior and simulate real motivations and interest trends.

4. The AI ​​personalized tutoring method based on cognitive profiling according to claim 1 is characterized in that: The method for constructing the behavior performance vector is: Constructing behavioral graphs; Each behavior dimension node , a total of The number of nodes, the intermediate features are calculated as: ； in, represents the intermediate eigenvector, Represents a user At the moment No. The original feature vector of the behavioral dimension, represents the first layer weight matrix, represents the activation function, 、 Represents the behavior dimension nodes 、 The degree, Represents a behavioral dimension node With node The adjacency relationship between them; Stack all node intermediate features in node order to obtain the intermediate feature matrix; Perform a graph convolution propagation on the intermediate feature matrix and output the behavior performance vector.

5. The AI ​​personalized tutoring method based on cognitive profiling according to claim 1 is characterized in that: The method for constructing the emotional state estimation vector is: Collect multiple data modalities of users; Each modality extracts feature vectors separately and then performs fusion, including: Visual channel feature vector : Input face image tensor , perform the first convolution: Extract image features through convolution kernel weight parameters; Perform multi-layer convolution: Apply convolution operations layer by layer to extract deep features of the image; Reduce the size of the feature map through the pooling layer and output the visual channel vector; Input text feature sequence, and extract language channel feature vector based on the neural network architecture of self-attention mechanism; Input physiological time series feature sequence and extract physiological channel feature vector through long short-term memory network; Map the feature vector of each modality to the same dimension, assign weights to each modality by calculating attention weights, and achieve multimodal fusion through weighted summation to generate an emotion feature representation vector; The fused emotion feature representation vector is mapped to a specific emotion space dimension through a multi-layer perceptron, and the final emotion state estimation vector is output.

6. The AI ​​personalized tutoring method based on cognitive profiling according to claim 5 is characterized in that: The expression for extracting image features through convolution kernel weight parameters is: ； in, Indicates the spatial location of the first layer output feature map , No. The value of the output channel, Indicates the number of input channels, 、 Represent the radius parameters of the convolution kernel in the height and width directions respectively, Represents the convolution kernel weight parameter in the first layer of convolution kernel, Indicates the first layer The bias vector of the output channels, Indicates The corresponding spatial position.

7. A personalized tutoring system, applied to the AI ​​personalized tutoring method based on cognitive profiling according to any one of claims 1 to 6, characterized in that: include: The objective function establishment module introduces the latent cognitive factor vector to capture the user's internal cognitive structure, and combines the explicit rating error term and the implicit preference confidence term to establish an optimization objective function and solve the latent cognitive factor vector; The cognitive profile construction module linearly transforms the latent cognitive factor vector, applies an activation function, and splices the activation results into a cognitive profile; The cognitive profile update module models the dynamic transition of the user's cognitive state in a time series, capturing the process from tension to adaptation and from fatigue to recovery, and thus updating the cognitive profile. The coaching strategy output module defines a coaching strategy set, constructs an attribute feature vector for each strategy, inputs an updated cognitive profile, outputs the optimal coaching strategy, and executes it to the user.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images