Sitting posture state analysis and recommendation method based on end-to-end spatio-temporal graph attention network

Abstract

The application provides a sitting posture analysis recommendation method based on an end-to-end space-time graph network, belongs to the technical field of computer vision and swarm intelligence, and is used for solving the problem that in the related art, the sitting posture of a student is evaluated in isolation, the mutual influence of a group is ignored, and a recommended strategy is one-sided and short-sighted. The method unifies the modeling of the interaction between students by constructing and dynamically updating a classroom space-time state graph; generates node states by using a network integrated with physical constraints, and fuses group and historical information by using a space-time attention mechanism; then decodes individual scores and group utility from the unified state representation; finally, defines a group potential function and makes an optimization decision based on the group potential function, selects a student who can most improve the overall situation for recommendation, and thus realizes deep understanding and intelligent guidance of the group sitting posture state.

Description

Technical Field

[0001] This application relates to the fields of computer vision and swarm intelligence, and in particular to a sitting posture analysis and recommendation method based on an end-to-end spatiotemporal graph network. Background Technology

[0002] In scenarios such as intelligent teaching and meeting management, automatically analyzing students' posture through video and intelligently recommending display subjects has become an important technological direction for improving interaction quality and participation. Existing technologies mainly rely on computer vision algorithms to independently analyze the posture of a single individual, assessing focus or image by calculating geometric features such as head angle and torso tilt, and often determining display subjects based on simple scoring and ranking.

[0003] However, the aforementioned existing technologies have significant drawbacks. First, they analyze each student in isolation, completely ignoring the mutual influence and group dynamics among students in real-world scenarios, leading to biased assessment results. Second, recommendation strategies based on fixed rules or single score rankings lack a decision-making logic that is disconnected from the deeper goal of "optimizing the overall classroom state through demonstration," failing to achieve intelligent intervention that guides the group towards positive outcomes. Summary of the Invention

[0004] This application provides a sitting posture analysis and recommendation method based on end-to-end spatiotemporal graph networks, which can achieve a deep understanding of the sitting posture state of a group and intelligent guided recommendation in a unified modeling and co-evolutionary manner.

[0005] Firstly, this application provides a posture analysis and recommendation method based on an end-to-end spatiotemporal graph network. The method involves acquiring video stream data from multiple video capture devices, each corresponding to a student; constructing and updating a classroom spatiotemporal state graph based on the video stream data; wherein each student is modeled as a dynamic node in the spatiotemporal state graph, and the state of each dynamic node is generated based on the corresponding video stream data, interacting and iteratively updating with the states of other nodes in the spatiotemporal state graph through a spatiotemporal attention mechanism; generating individual state evaluation data corresponding to each student based on the stable state vector of each node in the updated spatiotemporal state graph; and determining and outputting target student information to be recommended and displayed based on the individual state evaluation data and the spatiotemporal state graph.

[0006] By adopting the above technical solution, this application transforms traditional isolated individual analysis into a holistic modeling of group interaction systems. By constructing and continuously updating a "classroom spatiotemporal state diagram," the state assessment of each student is no longer static and isolated, but rather evolves and is determined through dynamic interactions with other members of the group. This lays the data foundation for making intelligent recommendations that ultimately consider the overall situation of the group, fundamentally solving the problem of existing technologies neglecting group effects.

[0007] Furthermore, for each student, a lightweight perceptron processes the corresponding video stream data to generate the initial state vector of that student node; wherein, the lightweight perceptron integrates a differentiable human rigid body kinematics constraint function and an attitude energy function to guide the network to output a node initial state vector that conforms to physical laws.

[0008] By adopting the above technical solution, prior knowledge of the physical world is embedded into the neural network in the form of differentiable constraints, ensuring the physical rationality and interpretability of the mapping process from the original video to the initial state vector, improving the accuracy and robustness of feature extraction, and avoiding abnormal outputs that may be generated by pure data-driven approaches that violate biomechanics.

[0009] Furthermore, based on the state vectors of each node in the spatiotemporal state graph, the weights of directed edges between any two nodes are dynamically calculated through a meta-relation network. The meta-relation network takes the state vectors of a pair of nodes and an encoding representing the spatial relative relationship between the pair of nodes as input, and outputs attention weights specifically used to model the state propagation relationship between nodes.

[0010] By adopting the above technical solution, the mutual influence relationships between nodes are no longer preset or fixed, but can dynamically evolve according to the real-time state. This dynamic directed edge weight generated by the meta-relation network can more accurately characterize complex group interaction patterns such as imitation and leadership, providing a high-quality association matrix for subsequent state propagation.

[0011] Furthermore, multiple rounds of spatiotemporal graph attention updates are performed on the spatiotemporal state graph. In each round of updates, for each node, messages from other nodes in the spatial dimension are aggregated after being weighted by edge weights to obtain spatial aggregated messages. Messages from the stable states of the node in the historical time dimension are also aggregated to integrate the temporal characteristics of its personal sitting posture habits to obtain temporal aggregated messages. Then, the spatial aggregated messages and the temporal aggregated messages are merged to update the state vector of the node.

[0012] By adopting the above technical solution, the state information of each node can be propagated and fused in multiple rounds and across time and space on the graph structure. The stable state vector finally obtained by the node not only contains its own posture information, but also encodes comprehensive information about its influence from the group and its own historical habits, thereby realizing a deep state representation of the individual in the group context.

[0013] Furthermore, the step of "merging the spatial aggregation message and the temporal aggregation message to update the state vector of the node" is specifically implemented through the following formula. ;in, Indicates the first In progress, at any time ,node The state vector, For spatial aggregation messages, Aggregate messages by time. This represents vector concatenation. It is a multilayer perceptron. This is a layer normalization operation.

[0014] By adopting the above technical solution, a clear and implementable state update function is provided. This function ensures training stability through residual connections and layer normalization, and achieves nonlinear fusion of spatial and temporal information through MLP, ensuring the efficiency and effectiveness of the state update process.

[0015] Furthermore, the stable state vector of each node is input into a multi-task decoder; the multi-task decoder generates, in parallel, individual state assessment data that includes at least attention score, image suitability score, and expected utility based on the stable state vector.

[0016] By adopting the above technical solution, an efficient mapping from a unified deep state representation to multiple specific evaluation metrics is achieved. This multi-task decoding mechanism ensures that all scores share the same rich set of group context information at the underlying level, making the evaluation results more consistent and holistic, while also improving computational efficiency.

[0017] Furthermore, the multi-task decoder generates the individual state evaluation data in the following manner: ; in, Indicates trainees At any moment Attention score Indicates the image suitability score. This indicates the intended utility being demonstrated. For nodes The stable state vector, , , , For learnable parameters, For a multilayer perceptron, This is the Sigmoid function.

[0018] By adopting the above technical solution, the process of generating evaluation data is defined in a clear mathematical form. Attention level and image rating are normalized using the Sigmoid function, which aligns with the rating intuition; the expected utility is demonstrated through independent MLP learning, which can comprehensively express the potential value of nodes to the group goal, providing direct input for subsequent optimization decisions.

[0019] Furthermore, a group potential energy function is defined to characterize the overall situation of the spatiotemporal state diagram; based on the defined group potential energy function, the expected impact of selecting different trainees as demonstration targets on the value of the group potential energy function is evaluated, and target trainees are selected accordingly.

[0020] By adopting the above technical solution, the recommendation decision problem is formalized into an optimization problem of the overall system state (potential energy). The decision objective changes from "selecting the individual with the highest score" to "selecting the intervention point that maximizes the overall potential energy of the system," thus directly aligning the recommendation strategy with the fundamental goal of "optimizing the overall classroom state."

[0021] Furthermore, the group potential function Defined as the weighted sum of individual utility and interaction energy: ;in, For students The expected utility of the demonstration For the slave node generated by the method according to claim 3 Pointing to node Attention weights A function to characterize the degree of correlation between two utilities, These are the weighting coefficients.

[0022] By employing the above technical solution, the group potential energy is specifically quantified into two parts: the sum of individual utility and the sum of utility correlation based on influence relationships. This function design enables the system not only to encourage outstanding individual performance but also to encourage close and positive correlations between high-influencers and high-utilizers, thus mathematically characterizing the ideal state of a healthy and positive group.

[0023] Furthermore, the "selection of target students accordingly" is achieved through the following calculation: ;in, For the selected target students, , It uses an influence propagation prediction module to "display trainees" based on the current spatiotemporal state diagram and candidate actions. " is the input, and the predicted group potential energy at the next moment is obtained.

[0024] By adopting the above technical solution, the decision-making criterion is clearly defined as "maximizing the expected potential energy increment". This requires the system to not only consider the current static value of the candidate trainees, but also to proactively simulate the group state evolution chain that may be triggered by the recommendation action through the prediction module, and finally select the trainee that can bring about the most positive global change, thus realizing a truly strategic intelligent recommendation.

[0025] In summary, this application has at least the following beneficial effects: 1. A sitting posture analysis method is provided that can uniformly model group interaction and make intelligent recommendations based on system optimization, overcoming the shortcomings of isolated evaluation and simple ranking in existing technologies; 2. By using a dynamic spatiotemporal graph attention mechanism, a deep state representation of individuals within a group context is achieved, making the evaluation results more comprehensive and accurate; 3. By defining and optimizing the group potential function for decision-making, the recommendation strategy can proactively guide the group state to evolve toward the expected goal, thus realizing a closed loop from perception to intervention.

[0026] It should be understood that the description in the Summary Section is not intended to limit the key or essential features of the embodiments of this application, nor is it intended to restrict the scope of this application. Other features of this application will become readily apparent from the following description. Attached Figure Description

[0027] The above and other features, advantages, and aspects of the embodiments of this application will become more apparent from the accompanying drawings and the following detailed description. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein: Figure 1 A schematic diagram of an exemplary operating environment in which embodiments of this application can be implemented is shown; Figure 2 The flowchart illustrates a method for analyzing and recommending sitting postures based on an end-to-end spatiotemporal graph attention network, as described in an embodiment of this application. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0029] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0030] This application provides a sitting posture analysis and recommendation method based on an end-to-end spatiotemporal graph network. It constructs a dynamic classroom spatiotemporal state graph to uniformly model the interaction between students and makes decisions based on group potential energy optimization. This realizes an intelligent method that transforms passive assessment of individual sitting posture into active guidance of the overall classroom state to evolve in a positive direction.

[0031] Figure 1 A schematic diagram of an exemplary operating environment in which embodiments of this application can be implemented is shown.

[0032] Reference Figure 1 The operating environment includes a distributed heterogeneous hardware system for implementing the posture analysis and recommendation method based on an end-to-end spatiotemporal graph network. This system consists of multiple intelligent sensing terminals deployed at student seats, one or more collaborative computing hubs located in a remote data center, a high-speed communication network connecting all the aforementioned devices, and interactive display devices for presenting the recommendation results. It aims to provide the necessary physical computing and data exchange foundation for constructing and updating spatiotemporal state graphs from video stream data and making intelligent recommendations.

[0033] Specifically, the intelligent sensing terminal mainly consists of a high-definition camera module with a built-in lightweight neural network acceleration unit. Its function is to acquire raw video streams and process them locally using a lightweight sensing network that integrates physical constraints, generating initial state vectors for each learner's nodes. This enables preliminary data perception and feature extraction while protecting the privacy of the original video. The collaborative computing hub consists of a server cluster equipped with a graphics processor and an independent federated learning coordination server. Its function is to receive feature vector streams from all intelligent sensing terminals, run a complex spatiotemporal graph attention network to iteratively update, decode, and make decisions on the state graph, and securely coordinate the federated learning process among multiple such systems to achieve collaborative model evolution.

[0034] Connecting the intelligent sensing terminals, collaborative computing hub, and display devices is a high-speed, low-latency network infrastructure. This network aggregates data from various terminals locally via a local area network (LAN), and remotely establishes secure connections between the hub and terminals, as well as between different hubs, via a wide area network (WAN) or dedicated network. Its function is to ensure reliable, real-time, and secure transmission of feature data, model parameters, and control commands throughout the system, serving as the link for collaborative operation across the entire system. The display and interactive devices include a main display screen and a management terminal. Their function is to receive and visualize recommended target learner images, while also providing administrators with an interactive interface to monitor the overall system status.

[0035] In summary, this operating environment, through its hierarchical design of edge perception, cloud cognition, network collaboration, and terminal interaction, forms a complete "collection-processing-decision-display" closed loop, providing the necessary hardware support for the complete technical chain of the posture analysis and recommendation method, from individual feature extraction to group potential energy optimization calculation.

[0036] Based on the above operating environment, the specific steps of the method embodiments of this application will be described in detail below. Figure 2 A flowchart illustrating a sitting posture state analysis and recommendation method based on an end-to-end spatiotemporal graph attention network according to an embodiment of this application is shown. (Refer to...) Figure 2 The method is executed by the collaborative computing center and specifically includes the following steps: S1: Acquire video stream data from multiple video capture devices, with each video capture device corresponding to one student.

[0037] The video acquisition device is an embedded AI camera module deployed in front of each student's seat in the operating environment, which operates at a fixed sampling frequency. (e.g., 25 frames per second) Synchronously acquire RGB format video stream data superscript Unique identifier Each student and their corresponding camera Total number of students, subscript A unified timestamp index ensures temporal alignment of multiple video streams. Each camera module's built-in lightweight computing unit provides a low-latency data interface for the lightweight sensing network in subsequent steps. (Video stream data) The data is transmitted in real time to the central processing server, forming the raw data source for continuous processing in subsequent steps. ,in The processing time is specified. The acquired video stream data serves as the raw perceptual input for constructing the entire classroom dynamic model.

[0038] S2: Based on the video stream data, construct and update a classroom spatiotemporal state diagram.

[0039] In this step, the system constructs a graph model with students as nodes and their mutual influence relationships as edges, and iteratively updates it. First, for each student, a lightweight perceptron processes the corresponding video stream data to generate the initial state vector of that student node. Its dimensions are , These are the preset hyperparameters for the network output dimension. This network integrates differentiable human rigid body kinematics constraint functions and attitude energy functions. Specifically, the attitude energy function... Defined as the predicted torso direction vector (From the predicted key point in the neck to the midpoint of the hip) and the direction of physical gravity. The negative logarithm of the cosine of the angle between them, i.e. This term is incorporated into network training as a regularization loss to penalize physically impossible severe tilting postures. (Kinematic constraint loss) The limb length ratios used to constrain predictions are within a reasonable range determined by prior knowledge of human anatomy, and their specific forms are, for example, as follows: ,in The limb length ratio predicted by the network. and These are preset upper and lower limits of a reasonable range. These constraints, used as regularization terms, are trained together with the network's main regression loss to guide the network's output to conform to physical laws. dimensional initial state vector Next, based on the state vectors of each node, the weights of directed edges (attention weights) between any two nodes are dynamically calculated through a meta-relational network. The network is in a state of pairs of nodes. , and spatial relative relationship coding For input, where It is a fixed, low-dimensional vector, obtained by transforming the normalized coordinate difference between the two students' seats in the classroom plane coordinate system through a sinusoidal position encoding function. The meta-relational network calculates... The original attention score is obtained, where and It is a learnable linear transformation matrix. It is a learnable attention vector. These are the key / query dimension hyperparameters of the attention mechanism. This represents a vector concatenation operation. Then, the softmax function is applied to the nodes... Normalize the values ​​on all outgoing neighbors to obtain the final directed attention weights. This models dynamic, asymmetric inter-node influence relationships. Subsequently, multiple rounds (e.g.,) are performed on the spatiotemporal state graph. (Round) Spacetime graph attention update. In the first... wheel( (Updating, targeting nodes) Aggregates messages from other nodes in the spatial dimension: ,in Represents all pointing nodes in the graph The set of neighboring nodes, It is the first The learnable value transformation matrix updated in each round. This is based on the message dimension. Simultaneously, messages from the time dimension are aggregated to incorporate personal preferences. ,in It is the first A gated loop unit that is updated in cycles. It was the previous step The final stable state vector of this node. Then, the two types of messages are fused to update the node state: ,in A small multilayer perceptron is used for nonlinear fusion of spatial and temporal messages. This is a layer normalization operation. After... After the round of updates, we obtain the time of each node at time [time value missing]. The final stable state vector It deeply encodes the physically corrected postures of individual learners, the context of group interactions, and their personal historical habits.

[0040] S3: Based on the stable state vector of each node in the updated spatiotemporal state graph, generate individual state evaluation data corresponding to each student.

[0041] In this step, the stable state vector of each node is... The data is input into a multi-task decoder to generate individual state assessment data in parallel. The multi-task decoder generates the data in the following manner: ; Among them, focus score Image suitability score Through independent linear transformations and the Sigmoid function The range is obtained. A higher value indicates a better state. , and , All parameters are learnable. Demonstrates expected utility. Through an independent multilayer perceptron Calculate, its range is It can flexibly represent the potential comprehensive value of the trainee in improving the overall state of the group; The internal weights and biases are all learnable parameters. All learnable parameters are learned from labeled data during the model training phase using gradient descent.

[0042] S4: Based on the individual state assessment data and the spatiotemporal state diagram, determine and output the target student information to be recommended for display.

[0043] In this step, the system performs forward-looking decision-making based on population potential optimization. First, a population potential function is defined to quantify the overall situation: This function is composed of the weighted sum of individual utilities and interaction energies. Wherein, To demonstrate the expected utility from S3, The directed attention weights generated from the S2 meta-relation network, It is a node The set of outgoing neighbors. Let cosine similarity be a function that characterizes the degree of correlation between two utilities. One specific implementation of cosine similarity is... ,in For a very small positive number (such as This is to prevent division by zero errors. A pre-defined weighting coefficient hyperparameter greater than 0 is used to balance the relative importance of the sum of individual utility and the sum of interaction energy in the potential energy. The decision objective is to select the student who can maximize the increase in potential energy at the next moment, calculated by the following formula: in . Generated by an influence propagation prediction module, which uses the current spatiotemporal state graph (including all nodes) as an example. Edge weight (etc.) and candidate actions "demonstrating student" Using the learned dynamic laws as input, the system quickly simulates the state evolution after intervention and calculates the new population potential energy after the evolution. .expect The randomness that may exist in the simulation can be considered. This can be achieved by evaluating all candidate trainees. Corresponding expected potential energy increment Select the optimal target .

[0044] S5: Output the target student information.

[0045] The system will identify the target students. and its associated assessment data ( , , The video stream is encapsulated and pushed to the display control unit via a low-latency interface. This unit schedules students from the video stream server. The system displays the corresponding real-time video feed, overlaying the student's identifier and assessment score graphically, and then outputs it to a physical display device for presentation, thus completing the intelligent guidance process. Simultaneously, the system updates the student's "last presentation time" stamp internally. This variable will be stored in a historical database to incorporate fairness considerations in long-term decision-making.

[0046] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, because according to the embodiments of this application, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this application.

[0047] By synchronously acquiring video stream data from multiple learners (S1), the system provides accurate and consistent perceptual input for constructing a time-aligned group dynamic model, a fundamental prerequisite for any group collaborative analysis. Based on this, the system creatively constructs and continuously updates a classroom spatiotemporal state graph (S2). Here, a lightweight perceptual network with injected physical constraints generates an initial state for each learner node that conforms to biomechanical principles, ensuring the rationality and authenticity of the model input. Through a meta-relational network, directed edge weights are dynamically calculated, and asymmetric mutual influence relationships among learners are modeled in real time, thus embedding isolated individuals into a dynamic social interaction structure. Furthermore, through a multi-round spatiotemporal graph attention update mechanism, the state of each node aggregates the influence of its neighbors in the spatial dimension and integrates its own historical habits in the temporal dimension. The resulting stable state vector becomes a unified representation encoding the multidimensional context of "individual-group-history." This series of methods works together to fundamentally eliminate the isolation in the system's assessment of learner states, achieving a deep understanding of the true state of individuals within the group.

[0048] Subsequently, a multi-task decoder performs parallel decoding (S3) on the aforementioned unified stable state vector to directly generate attention scores, image suitability scores, and expected performance utility. Since all scores share the same state representation enhanced by group and historical information, they naturally possess consistency and context awareness. Attention and image assessments are no longer contradictory, and expected performance utility implies its potential value within the group. This resolves the conflicts and biases that may arise from the isolated calculation of various assessment indicators in traditional methods.

[0049] Furthermore, the system introduces a group potential energy function to quantify the overall situation of the entire spatiotemporal state graph, and makes recommendation decisions based on this function (S4). The design of this function combines individual utility with interaction energy based on influence networks, making the system's optimization objective explicitly defined as increasing overall potential energy, rather than simply selecting the currently optimal individual. During decision-making, the influence propagation prediction module simulates the global impact of different recommendation actions on the future, and selects the student who brings the largest expected increase in potential energy. This mechanism transforms short-sighted, statically ranked recommendations into long-term, systematic interventions aimed at guiding the group towards a better state.

[0050] Ultimately, the system outputs recommended targets and updates relevant states (S5), forming a complete closed loop of "perception-modeling-evaluation-decision-execution". Each step is interconnected, from ensuring data synchronization to building a group interaction model, and then to achieving collaborative evaluation and global optimization decisions. Ultimately, it logically and consistently achieves a posture analysis and recommendation effect that can understand group dynamics and proactively provide intelligent guidance, overcoming the core shortcomings of existing technologies such as isolated evaluation and short-sighted decision-making.

[0051] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the foregoing disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A method for analyzing and recommending sitting postures based on an end-to-end spatiotemporal graph attention network, characterized in that, include: Acquire video stream data from multiple video capture devices, with each video capture device corresponding to one student; Based on the video stream data, a classroom spatiotemporal state diagram is constructed and updated; wherein, each student is modeled as a dynamic node in the spatiotemporal state diagram, and the state of each dynamic node is generated based on the corresponding video stream data, and interacts and iteratively updates with the state of other nodes in the spatiotemporal state diagram through a spatiotemporal attention mechanism; based on the stable state vector of each node in the updated spatiotemporal state diagram, individual state evaluation data corresponding to each student is generated. Based on the individual state assessment data and the spatiotemporal state diagram, the target student information to be recommended and displayed is determined and output.

2. The method according to claim 1, characterized in that, The "constructing and updating a classroom spatiotemporal state diagram" includes: For each student, a lightweight perceptron processes the corresponding video stream data to generate the initial state vector of that student node. The lightweight perceptron integrates differentiable human rigid body kinematics constraint functions and attitude energy functions to guide the network to output a node initial state vector that conforms to physical laws.

3. The method according to claim 1, characterized in that, The "constructing and updating a classroom spatiotemporal state diagram" includes: Based on the state vectors of each node in the spatiotemporal state diagram, the weights of directed edges between any two nodes are dynamically calculated through a meta-relation network. The meta-relation network takes the state vectors of a pair of nodes and an encoding representing the spatial relative relationship between the pair of nodes as input, and outputs attention weights specifically used to model the state propagation relationship between nodes.

4. The method according to claim 1, characterized in that, The "constructing and updating a classroom spatiotemporal state diagram" includes: Multiple rounds of spatiotemporal graph attention updates are performed on the spatiotemporal state graph. In each round of updates, for each node, messages from other nodes in the spatial dimension are aggregated after being weighted by edge weights to obtain spatial aggregated messages. Messages from the stable states of the node in the historical time dimension are also aggregated to integrate the temporal characteristics of its personal sitting posture habits to obtain temporal aggregated messages. The spatial aggregated messages and the temporal aggregated messages are then merged to update the state vector of the node.

5. The method according to claim 4, characterized in that, The step of "merging the spatial aggregation message and the temporal aggregation message to update the node's state vector" is specifically implemented through the following formula: ;in, Indicates the first In progress, at any time ,node The state vector, For spatial aggregation messages, Aggregate messages by time. This represents vector concatenation. It is a multilayer perceptron. This is a layer normalization operation.

6. The method according to claim 1, characterized in that, The phrase "generating individual state evaluation data corresponding to each trainee based on the stable state vector of each node in the updated spatiotemporal state graph" includes: The stable state vector of each node is input into a multi-task decoder; the multi-task decoder generates, in parallel, individual state assessment data, including at least attention score, image suitability score, and expected utility, based on the stable state vector.

7. The method according to claim 6, characterized in that, The multi-task decoder generates the individual state evaluation data in the following manner: ； in, Indicates trainees At any moment Attention score Indicates the image suitability score. This indicates the intended utility being demonstrated. For nodes The stable state vector, , , , For learnable parameters, For a multilayer perceptron, This is the Sigmoid function.

8. The method according to claim 1, characterized in that, The phrase "based on the individual state assessment data and the spatiotemporal state diagram, determine and output the target student information to be recommended and displayed" includes: Define a group potential energy function that characterizes the overall situation of the spatiotemporal state diagram; based on the defined group potential energy function, evaluate the expected impact of selecting different trainees as demonstration targets on the value of the group potential energy function, and select target trainees accordingly.

9. The method according to claim 8, characterized in that, The group potential function Defined as the weighted sum of individual utility and interaction energy: ;in, For students The expected utility of the demonstration For the slave node generated by the method according to claim 3 Pointing to node Attention weights A function to characterize the degree of correlation between two utilities, These are the weighting coefficients.

10. The method according to claim 8 or 9, characterized in that, The "selection of target students based on this" is achieved through the following calculation: ;in, For the selected target students, , It uses an influence propagation prediction module to display students' current spatiotemporal state and candidate actions. " is the input, and the predicted group potential energy at the next moment is obtained.

Images