Family education knowledge graph construction and intelligent pushing platform and method

Abstract

The invention relates to the field of artificial intelligence, and discloses a family education knowledge graph construction and intelligent pushing platform and method, and the platform comprises a data collection module, a family state modeling module, an intervention planning module and a pushing module. The family state modeling module is used for collecting multi-modal data, constructing a state vector for each family member and calculating interaction tension among the members so as to generate a family state tensor capable of comprehensively representing a current family system state; and the intervention planning module searches a resonance intervention sequence capable of enabling the family state evolution path cost to be the lowest from a dynamic knowledge graph formed by quantifiable stem budget factors through prospective path simulation based on the family state tensor, and determines an optimal target stem budget factor. According to the method, intelligent intervention which is highly contextualized, personalized and self-evolved can be realized, and the accuracy and effectiveness of family education are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence, in particular to a family education knowledge graph construction and intelligent pushing platform and method. BACKGROUND

[0002] With the development of information technology, digital family education resources are increasingly popular. Currently, the family education application programs or online platforms on the market usually provide a huge content library for users, which contains various types of educational articles, parent-child activity plans, and expert courses. Users can obtain content through keyword search or classification and screening based on preset tags (such as child age, education theme). Some platforms will make preliminary content recommendations based on the initial static profile filled by the user (such as the child's interests, the parent's education difficulties, etc.).

[0003] However, these existing technologies have a deep limitation in application, that is, they lack the ability to perceive the real-time and dynamic state of the complex system of the family. The emotions, energy, cognitive load of family members, and the interaction between members are constantly changing, and the existing recommendation mechanism often ignores these key situational information that determines whether the intervention is effective, resulting in the content pushed at a specific moment may not be applicable, and even may have a counterproductive effect, making it difficult to achieve true personalization.

[0004] More importantly, existing technologies usually adopt a "one-way pushing" mode, which distributes content to users but lacks an effective closed-loop feedback mechanism to evaluate the actual impact of the content on the family state after implementation. The system cannot know whether a certain suggestion is really effective for a particular family, so it cannot learn and evolve from experience, and its knowledge base and recommendation strategy are fixed and cannot be adapted to different families and different situations.

[0005] In addition, the decision logic of these systems is relatively simple, often only recommending a single, isolated educational activity, and failing to plan a series of intervention paths that can guide the evolution of the family state towards the ideal goal from a longer-term and overall perspective. This mode of lacking forward-looking planning ability limits its role in promoting the long-term and positive development of the family. Therefore, the industry urgently needs a new type of intelligent pushing technology that can dynamically model the family state, make forward-looking intervention planning, and achieve self-iterative optimization. SUMMARY

[0006] The technical problem to be solved by the present application is that the existing family education knowledge graph construction method usually organizes knowledge in a static structure, which is difficult to capture the real-time and continuous changes in the state and interaction between family members. Based on such a static knowledge graph, the pushing method often lacks the pertinence of the current specific situation of the family, and the intervention effect is limited.

[0007] To solve the above technical problems, the present application provides a family education intelligent pushing platform and method capable of dynamically modeling the interaction relationship within a family and planning prospective intervention.

[0008] The present application provides a family education knowledge graph construction and intelligent pushing platform, which comprises: a data acquisition module for acquiring multi-modal state data of at least two family member agents; a family state modeling module connected with the data acquisition module, for generating a state vector for each of the agents according to the multi-modal state data, and calculating a tension matrix describing the interaction tension between the agents based on the relationship between the state vectors, and further constructing a family state tensor comprising the state vectors and the tension matrix; an intervention planning module connected with the family state modeling module, for selecting at least one target education intervention operator from a preset operator library comprising a plurality of education intervention operators according to the family state tensor, wherein the operator library and the transformation effect parameters of each education intervention operator therein jointly constitute a dynamic knowledge graph for describing intervention knowledge; a pushing module connected with the intervention planning module, for pushing education content corresponding to the target education intervention operator to a user.

[0009] In an optional embodiment, the family state modeling module processes the multi-modal state data into quantitative indicators including at least emotional valence, cognitive load or physiological energy level, and constructs the state vector based on the quantitative indicators.

[0010] In an optional embodiment, the family state modeling module calculates the tension matrix by the following way: For any two agents, the interaction tension value between them is calculated by a quadratic function based on their respective state vectors and a preset weight matrix, and this value is taken as the corresponding element of the tension matrix.

[0011] Specifically, the interaction tension between agent and agent at time is determined by the following formula: ; wherein, and are the state vectors of the two agents respectively.​ and At time a state vector, is a preset weight matrix that defines the contribution degree of the difference in different state dimensions.

[0012] In an optional embodiment, each educational intervention operator is defined as a structured mathematical transformation that can simultaneously change the state vector and the tension matrix in the family state tensor.

[0013] Specifically, the transformation is implemented by vector addition by applying a predefined change vector to the state vector, and Hadamard product by applying a tension adjustment matrix to the tension matrix.

[0014] In an optional embodiment, the platform further comprises an operator effect calibration module. This module is used to collect the subsequent family state tensor as the actual state after pushing the educational content corresponding to the target educational intervention operator, and update the transformation effect of the target educational intervention operator based on the difference between the actual state and the predicted state predicted by the target educational intervention operator.

[0015] In an optional embodiment, the intervention planning module selects the target educational intervention operator by the following steps: First, select multiple educational intervention operators from the operator library to form at least one time-sequential intervention sequence; Second, simulate at least one of the intervention sequences to generate a corresponding state evolution path, wherein each state evolution path contains a series of family state tensors arranged in chronological order.

[0016] In an optional embodiment, the intervention planning module further completes the selection by the following steps: First, calculate the cost of each state evolution path based on a preset path cost function; Second, select the state evolution path with the lowest cost, and determine the first educational intervention operator in the intervention sequence corresponding to this path as the target educational intervention operator.

[0017] In an optional embodiment, the path cost function at least includes the evaluation of one or more of the following: The process tension cost of the tension matrix contained in all family state tensors in the state evolution path; The final state cost of the tension matrix contained in the terminal family state tensor of the state evolution path; The path smoothness cost of the change intensity between adjacent family state tensors in the state evolution path.

[0018] In one specific embodiment, the path cost function is calculated according to the following formula: ; Wherein: is the calculation result of the path cost function for a state evolution path ; is the total number of intervention steps in the state evolution path; is the serial number of the intervention step; is the family state tensor corresponding to the th step in the state evolution path; is the tension matrix contained in the family state tensor ; is the Frobenius norm of a matrix; is the generalized distance norm between two family state tensors; is the weight coefficient of the process tension cost; is the weight coefficient of the final state cost; is the weight coefficient of the path smoothness cost.

[0019] The second aspect of the present application provides a family education intelligent pushing method, which comprises the following steps: Collecting multi-modal state data of at least two family member agents; Generating a state vector for each of the agents according to the multi-modal state data; Calculating a tension matrix describing the interaction tension between the agents based on the relationship between the state vectors; Constructing the state vector and the tension matrix into a family state tensor; Selecting at least one target education intervention operator from a pre-set operator library containing a plurality of education intervention operators according to the family state tensor; Pushing the education content corresponding to the target education intervention operator to the user.

[0020] The present application provides a family education knowledge graph construction and intelligent pushing platform and method. The present application has the following beneficial effects: 1、The present application constructs a state vector for each family member agent, and further calculates the interaction tension between agents to form a tension matrix, which is finally combined with the state vector to form a family state tensor. This method not only quantifies the individual state of family members, but also quantifies the dynamic interaction relationship between members for the first time. Compared with the prior art which only relies on isolated and static user portraits, the present application provides an accurate description of the overall state of the family system, which is dynamic and multi-dimensional, so as to more deeply identify the root cause of the family problem and provide a reliable data basis for subsequent accurate intervention.

[0021] 2、The present application encapsulates educational knowledge content as an educational intervention operator with a certain mathematical transformation effect, and constructs a dynamic knowledge graph with these operators and their effect parameters. When pushing, the system selects based on the predictable influence of the operator on the family state tensor. This makes each content push become an accurate intervention with clear goals and predictable effects, changing the blind state of the prior art where the effect of content push is unknown, and improving the targeting and effectiveness of the intervention measures.

[0022] 3、The present application simulates an intervention sequence composed of multiple educational intervention operators, and uses a path cost function containing process tension, final state and path smoothness to evaluate the entire state evolution path, thereby realizing forward-looking intervention planning. This method can select the long-term optimal intervention strategy, rather than only focusing on the immediate effect of the next step. This overcomes the short-sightedness of traditional recommendation algorithms, enabling them to perform complex intervention plans that require multiple steps to achieve long-term goals, thereby providing more strategic and fundamental guidance to users. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 The structural block diagram of the family education knowledge graph construction and intelligent push platform of an embodiment of the present application; Figure 2 The data processing and state vectorization flowchart of an embodiment of the present application; Figure 3 The family state tensor construction flowchart of an embodiment of the present application Figure 4 The intervention planning flowchart of an embodiment of the present application.

[0024] Among them, 10, data acquisition module; 20, family state modeling module; 30, intervention planning module; 40, push module. DETAILED DESCRIPTION

[0025] For the purposes of making the objectives, technical solutions, and advantages of the present application clearer, the technical solutions of the present application will be described clearly and completely below. It should be understood that the described embodiments are part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0026] Reference is made to the accompanying drawings Figure 1 , Figure 1 is a structural block diagram of a family education knowledge graph construction and intelligent pushing platform according to an embodiment of the present application. The platform provided by the present application can be a server, a terminal device, or deployed in the form of cloud service on a server cluster, and the hardware architecture thereof includes but is not limited to a processor and a memory. The memory stores a computer program, and the processor implements the functions of the platform of the present application when executing the computer program. The platform includes: a data acquisition module 10, a family state modeling module 20, an intervention planning module 30, and a pushing module 40.

[0027] In an optional embodiment, the platform further includes an operator effect calibration module 50.

[0028] The data acquisition module 10 is configured to acquire multi-modal state data of at least two family member agents by communicating with an application program interface (API) of a user terminal. The data acquisition module 10 transmits the acquired raw data to the family state modeling module 20.

[0029] The family state modeling module 20 is connected with the data acquisition module 10 and is used to receive the multi-modal state data. The module performs data processing and modeling operations, specifically: A state vector is generated for each agent, and a tension matrix describing the interaction tension between agents is calculated based on the relationship between the state vectors, and then a structured family state tensor is constructed. The family state modeling module 20 outputs the generated family state tensor to the intervention planning module 30, and also outputs it to the operator effect calibration module 50 under certain conditions.

[0030] The intervention planning module 30 is connected with the family state modeling module 20 and is used to receive the family state tensor. The module contains or has access to an operator library, which is composed of a plurality of education intervention operators. The operator library and the transformation effect parameters of each education intervention operator therein jointly constitute a dynamic knowledge graph for describing intervention knowledge. The intervention planning module 30 executes a path planning algorithm according to the received family state tensor, selects a target education intervention operator from the operator library, and sends a content identifier associated with the target education intervention operator to the pushing module 40.

[0031] The push module 40 is connected with the intervention planning module 30, and is used for receiving the content identifier. The module retrieves the corresponding educational content from the content database according to the content identifier, and pushes the educational content to the designated user terminal device for presentation through the network interface.

[0032] The operator effect calibration module 50 is connected with the family state modeling module 20 and the intervention planning module 30. The module is used for receiving the predicted state information generated by the intervention planning module 30, and the actual family state tensor generated by the family state modeling module 20 after the intervention is implemented. By comparing the difference between the predicted state and the actual state, the operator effect calibration module 50 executes an optimization algorithm, updates and adjusts the transformation effect parameters of the corresponding educational intervention operator in the operator library used by the intervention planning module 30, so as to realize the iterative update of the dynamic knowledge graph.

[0033] In a complete workflow of the present application, the modules work in the following manner: The data acquisition module 10 continuously acquires data and generates a real-time family state tensor for the family state modeling module 20; The family state modeling module 20 transmits the tensor to the intervention planning module 30; The intervention planning module 30 makes a decision, selects a target intervention operator, and transmits an instruction to the push module 40; The push module 40 performs a push, and after the push, the operator effect calibration module 50 performs effect tracking and parameter calibration, forming a closed-loop control system driven by data and iterated by a model.

[0034] Referring to the accompanying drawings Figure 2 , Figure 2 is a data processing and state vectorization flowchart according to an embodiment of the present application. The flowchart is first executed by the data acquisition module 10, and the subsequent processing is completed by the family state modeling module 20.

[0035] The specific function of the data acquisition module 10 is to acquire state data for each family member agent in the system.

[0036] In an embodiment, the system creates and binds a dedicated family member agent for each family member. The data acquisition module 10 acquires two types of data generated by the user through the application interface on the user terminal device bound to the member: The first type is data actively submitted by the user, for example, the user selects from a preset emotion vocabulary (such as containing the words “calm”, “tired”, “concentrated”, “irritated”, etc.), or marks the current physiological energy level by dragging a slider with a value range of 0 to 100; The second kind is passive sensing data collected after obtaining explicit authorization from the user, for example, local real-time analysis of the audio stream when the user interacts with the platform through voice, extracting non-content acoustic features such as fundamental frequency, pitch variation range, speech rate, signal-to-noise ratio, etc.

[0037] In a specific embodiment, the acoustic features include but are not limited to: Mel Frequency Cepstral Coefficients (MFCCs), fundamental frequency (F0), formants, and jitter and shimmer parameters that describe the stability of the sound.

[0038] The data acquisition module 10 can perform frame processing on the audio stream locally on the user terminal, and extract the above-mentioned features on each frame, and then input the feature sequence into the subsequent emotion recognition model.

[0039] The home state modeling module 20 receives the raw data transmitted by the data acquisition module 10 and performs a series of processing steps to generate a quantized state vector. The process includes: First, the raw data is quantized and converted into a numerical value in the preset state dimension. For example, the user-selected emotional vocabulary is converted into a continuous value between -1 and +1 in the "emotional valence" dimension through a preset mapping table (Look-up Table). The user input physiological energy level slider value is mapped to a value between 0 and 1 in the "physiological energy" dimension through linear normalization processing. The extracted multiple acoustic features are input into a pre-trained acoustic emotion recognition model (such as a support vector machine or a small neural network model), and the output is a comprehensive "emotional arousal" dimension value. The user's application usage frequency and switching times in a specific time period are input into a preset function to calculate the "cognitive load" dimension value.

[0040] In a specific embodiment, the preset function for calculating cognitive load can be a linear combination function. For example, the cognitive load dimension value can be determined by the following formula: ; wherein, is the average switching frequency of the application program on the user terminal within the preset time window, is the number of application programs in the active state within the time window. and are preset non-negative weight coefficients for adjusting the contribution of different behavior indicators to the calculation of cognitive load. The calculation result is further normalized to a preset numerical value range.

[0041] Secondly, all the quantized state dimension values ​​generated in the previous step are combined into a structured state vector. For the first state in a family system... A smart agent At any moment state vector Represented as: ; in, It belongs to 3D real space Column vectors. It is the first of the vectors Each component represents a specific state dimension.

[0042] In one embodiment, It can represent emotional valence. It can represent physiological energy. It can represent cognitive load. This can represent parenting patience, etc. Vector dimensions. It is a configurable integer, set according to the specific application scenario. All components The values ​​are all constrained to a uniform numerical range, such as [-1, 1], to ensure the validity of subsequent calculations.

[0043] Through the above steps, the family state modeling module 20 transforms the discrete, multimodal raw input data into a continuous, standardized state vector that can be used for mathematical modeling, providing input for subsequent calculation of interaction tension.

[0044] After generating the state vectors of each family member agent, the family state modeling module 20 continues to perform subsequent modeling steps to construct a family state tensor that can comprehensively represent the dynamics of the family system.

[0045] See attached document Figure 3 , Figure 3 This is a flowchart of a family state tensor construction process according to an embodiment of the present invention. The core of this process lies in not only processing the states of individuals but also quantitatively modeling the interactions between individuals. The family state modeling module 20 receives the state vector set containing N agents generated in the previous stage. As input.

[0046] Based on this input, the family state modeling module 20 can model any pair of intelligent agents in the system. Calculate its time at time Interactive tension Interaction tension is a scalar whose magnitude quantifies the degree of inconsistency or disharmony between the state vectors of two agents. Its calculation is determined by the following formula: ; wherein, is the difference vector of the two agent state vectors, representing the size and direction of the difference in all state dimensions. is a weight matrix, which is a symmetric positive semi-definite matrix, specially designed to define the interaction and importance of the difference in different state dimensions in a specific relationship pair.

[0047] In one embodiment, the diagonal elements of the weight matrix represent the independent contribution weight of the difference in the th state dimension to the total tension. For example, when defining the weight matrix of the “mother-child” specific relationship pair, the weight of the difference in the “parenting patience” dimension and the “physical energy” dimension can be set to be higher. The non-diagonal elements of the matrix represent the coupling effect between the difference in the th state dimension and the difference in the th state dimension. For example, when the difference in the “cognitive load” dimension of the parent and the difference in the “social willingness” dimension of the child are both large, a positive non-diagonal weight value can make the calculated tension value higher. The weight matrix is bound to a specific agent pair , which means that the weight matrix of the “spouse” relationship pair can be different from the weight matrix of the “parent-child” relationship pair, so as to be able to model the differences in different properties of the internal relationship of the family. The initial value of the matrix can be pre-set by experts in the field of family education according to professional knowledge, or learned and optimized in a data-driven manner during the running of the system.

[0048] In a specific embodiment, the learning and optimization process of the weight matrix is coupled with the workflow of the operator effect calibration module 50. When the operator effect calibration module 50 calculates the error term between the predicted state and the actual state, not only the effect parameters of the intervention operator are optimized, but also the gradient of each element in the weight matrix with respect to the error term is calculated. Subsequently, the system updates the weight matrix using the gradient descent method or its variant algorithm: ; wherein, is a small learning rate used to control the update rate of the weight matrix. In this way, if the difference in a certain state dimension is repeatedly proven to be related to a higher prediction error, the weight value corresponding to this dimension in the weight matrix will be automatically adjusted. This iterative process enables the platform to autonomously learn the true importance of each state dimension in different family relationships from the data, thereby continuously refining the interaction force field model.

[0049] After calculating the interaction force values between all pairs of agents, the family state modeling module 20 assembles these values into a force matrix . This matrix is a symmetric matrix, and the value of the element located at the th row and the th column is . The diagonal elements of the matrix are always 0 by definition. This force matrix fully presents the relationship force spectrum between all pairs of members in the family system at this moment.

[0050] Finally, the family state modeling module 20 structurally combines the individual states and the relationship states to construct the family state tensor . This tensor is a data container, and its structure is defined as: ; It contains two components: the first component is the set of state vectors of all agents , which describes the state of each person; the second component is the force matrix , which describes the relationship state between all people.

[0051] After completion of the construction, the family state tensor is used as the input basis for the subsequent intervention planning module 30 to make decisions, providing a comprehensive and quantitative snapshot of the current state of the family system.

[0052] The intervention planning module 30 is the core decision-making unit of the platform. After receiving the family state tensor generated by the family state modeling module 20, the module performs a series of planning and decision-making operations to determine the optimal intervention measures.

[0053] Referring to the accompanying Figure 4 , Figure 4 is an intervention planning flowchart according to an embodiment of the present application. First, the intervention planning module 30 abstracts each piece of educational content or activity suggestion stored in the knowledge base into a mathematical entity, i.e., an educational intervention operator All these operators and their associated effect parameters are stored together in an operator library, which constitutes the dynamic knowledge graph defined in the present application. Each operator is defined as a function that transforms an input home state tensor into another home state tensor. The transformation effect is decomposed into specific actions on two components of the state tensor: 1. Action on state vector: Operator applies a predefined vector addition to the state vector of the agent , i.e. where is a change vector bound to the operator and the agent , whose components represent the expected influence amount of the intervention content on the corresponding state dimension of the agent.

[0054] 2. Action on tension matrix: Operator applies an element-wise product operation to the tension matrix , i.e. where is a tension adjustment matrix associated with the operator , whose element values are within the interval [0, 1], used to directly adjust the interaction tension between specific agents.

[0055] In an embodiment, for the initial deployment of the system or the introduction of an entirely new field of educational intervention operators, their initial effect parameters can be estimated and set by domain experts based on their professional knowledge. For example, an expert can set a vector for an "engage in 15-minute parent-child reading" operator that slightly increases the "emotional valence" dimension and slightly reduces the "cognitive load" dimension, and set an adjustment matrix that slightly reduces the interaction tension between parents and children. These initial values based on expert experience will serve as the starting point for subsequent data-driven optimization through the operator effect calibration module.

[0056] The planning process of the intervention planning module 30 is not continuously running, but is triggered by specific conditions. In an embodiment, the trigger condition is set as: when the maximum interaction tension value in the home state tensor exceeds the preset tension threshold , or the absolute value of any state component of any agent exceeds the preset state deviation threshold , the planning process starts.

[0057] After the process is initiated, the intervention planning module 30 performs a forward-looking path simulation. The module first selects a subset of operators from the operator library and permutes these operators into a plurality of candidate intervention sequences of length . For each intervention sequence , the module starts from the current time's family state tensor and recursively applies each operator in the sequence to generate a state evolution path , where any point in the path is obtained from the previous point by applying the operator, i.e. .

[0058] After all candidate state evolution paths are generated, the intervention planning module 30 evaluates each path using a path cost function . The function is calculated according to the following formula: ; The formula contains three weighted parts: The first part, the process tension cost, quantifies the overall level of disharmony in the entire intervention process by accumulating the Frobenius norm of the tension matrix of each state in the path.

[0059] The second part, the final state cost, quantifies the level of harmony in the family system at the end of the intervention by calculating the Frobenius norm of the tension matrix of the path's end state.

[0060] The third part, the path smoothness cost, quantifies the degree of state change by accumulating the generalized distance norm between adjacent family state tensors in the path to penalize intervention paths that cause drastic jumps in the system state. , , , are configurable non-negative weight coefficients corresponding to the three parts of the cost, respectively, used to balance different optimization objectives.

[0061] To address the issue of huge computational overhead caused by evaluating all possible paths, in a preferred embodiment, the intervention planning module 30 employs a beam search algorithm to find the state evolution path with the lowest cost. At each step of the simulation process (i.e., after applying an intervention operator), the module only keeps the lowest-cost candidate paths and discards all other higher-cost paths. The parameter referred to as "beam width", is a configurable positive integer. By limiting the search width to , the method is able to reduce the computational complexity from exponential level to linear level with respect to the path length and the beam width , ensuring the engineering feasibility of the planning process.

[0062] Finally, the intervention planning module 30 computes and compares the cost values of all candidate paths, selects the state evolution path that makes the cost value minimum, and determines the corresponding intervention sequence as the optimal intervention sequence. The first educational intervention operator in the sequence is selected as the target educational intervention operator for this time. The intervention planning module 30 outputs the identifier of the target educational intervention operator to the push module 40.

[0063] Upon receiving the identifier of the target educational intervention operator sent by the intervention planning module 30, the push module 40 executes the content pushing process. The identifier uniquely corresponds to a specific educational intervention content. The push module 40 retrieves the specific data of the intervention content from a pre-set content database according to the identifier. The content data can be a combination of one or more formats, such as a text and image article, a uniform resource locator (URL) of a video file, or a structured activity suggestion scheme (including fields such as activity name, suggested duration, participating members, step-by-step instructions, etc.).

[0064] While obtaining the content data, the push module 40 also receives the target user information for this time of pushing from the intervention planning module 30, which specifies which family member agent the content should be sent to. The push module 40 encapsulates the content data into one or more push messages according to the target user information. Subsequently, the module calls the push service interface provided by the operating system or a third-party service, such as the Apple Push Notification Service (APNS) or a unified push service, to send the encapsulated messages to the application program on the target user's terminal device, thereby completing the presentation of the content.

[0065] In an optional embodiment, the platform further includes an operator effect calibration module 50, which is used to implement adaptive updating of the dynamic knowledge graph. After the push module 40 completes a content pushing, the operator effect calibration module 50 is activated after a pre-set time window . The calibration process of the module is as follows: First, the module obtains the state predicted for this intervention from the intervention planning module 30. Specifically, when the intervention planning module 30 selects the target operator At this time, it has calculated the predicted family state tensor after applying the operator in the simulation process . The operator effect calibration module 50 takes this as the prediction benchmark.

[0066] Secondly, the module obtains the actual family state tensor after the end of the time window from the family state modeling module 20 . The tensor is generated in real time by the data acquisition module 10 and the family state modeling module 20 after the implementation of the intervention according to the new data collected.

[0067] Then, the operator effect calibration module 50 calculates the error between the predicted state and the actual state. The error is defined as an error term , which can be the mean square error of all corresponding components between the two, for example: ; wherein, represents the square of the L2 norm of the vector, represents the square of the Frobenius norm of the matrix.

[0068] Finally, the operator effect calibration module 50 adjusts the effect parameters of the operator used in this intervention based on the calculated error term . In one embodiment, the adjustment process is realized by a gradient descent algorithm. Let the effect parameter set of the operator be , then the update rule of the parameter is: ; wherein, is a preset learning rate, is the gradient of the error term with respect to the parameter set . Through this step, the effect parameters of the operator are fine-tuned in the direction that can reduce the future prediction error. The updated parameter is written back to the operator library of the intervention planning module 30, thereby completing one iteration optimization of the dynamic knowledge graph.

[0069] The present application also provides a family education intelligent pushing method, which can be executed by the platform described above, and specifically includes the following steps: S100, acquiring multi-modal state data of at least two family member agents through a user terminal device. The multi-modal state data includes data actively submitted by the user through an application interface, and passive perception data related to terminal device use behavior collected after the user authorizes.

[0070] S200, execute family state modeling. This step is based on the multi-modal state data collected in S100, first generate a quantitative state vector for each family member agent; then, calculate an interaction tension value between any two agents based on their state vectors, and combine all interaction tension values into a tension matrix; finally, all state vectors and tension matrix are jointly constructed into a family state tensor.

[0071] S300, execute intervention planning. In a preferred embodiment, this step is triggered when a preset indicator (e.g. the maximum interaction tension value) in the family state tensor exceeds a preset threshold. This step specifically includes: S310, from an operator library consisting of multiple educational intervention operators, heuristically filter out a candidate operator subset highly relevant to the currently over-limit indicator.

[0072] S320, combine operators in the candidate operator subset into multiple candidate intervention sequences with a preset length.

[0073] S330, perform forward-looking simulation on each candidate intervention sequence, i.e. start from the current family state tensor, and gradually apply operators in the sequence to generate a corresponding state evolution path.

[0074] S340, use a preset path cost function to calculate a total cost value for each state evolution path. The path cost function includes a comprehensive quantification of path process tension, final state tension, and path smoothness.

[0075] S350, compare the total cost values of all candidate paths, determine the state evolution path with the lowest total cost value as the optimal path, and select the first educational intervention operator in the intervention sequence corresponding to the optimal path as the target intervention operator.

[0076] S400, execute content pushing. According to the identifier of the target intervention operator determined in S350, retrieve the corresponding educational content from a content database, and push the educational content to the specified user terminal device(s) through the network for presentation.

[0077] S500, in a preferred embodiment, the method further includes an operator effect calibration step. This step is executed after a preset time window after content pushing is completed, specifically: obtain a predicted family state tensor generated by the intervention planning step, and an actual family state tensor generated by the state modeling step at the same time; calculate the error between the predicted state tensor and the actual state tensor; and according to the error, use an optimization algorithm (e.g. gradient descent method) to update the transformation effect parameters of the target intervention operator used in this push in the operator library.

[0078] While embodiments of the application have been shown and described, it is to be understood that the application is not limited to the details of the embodiments described, since numerous changes, modifications, substitutions and variations can be made thereto without departing from the spirit and scope of the application as defined by the appended claims and their equivalents.

Claims

1. A home education knowledge graph construction and intelligent pushing platform, characterized in that, The method comprises the following steps: a data collection module is configured to collect multi-modal state data of at least two family member agents; a family state modeling module is connected to the data collection module and configured to generate a state vector for each of the agents based on the multi-modal state data, and calculate a tension matrix describing the interaction tension between the agents based on the relationship between the state vectors, and further construct a family state tensor comprising the state vectors and the tension matrix; an intervention planning module is connected to the family state modeling module and configured to select at least one target educational intervention operator from a preset operator library comprising a plurality of educational intervention operators based on the family state tensor; a pushing module is connected to the intervention planning module and configured to push educational content corresponding to the target educational intervention operator to a user. 2.The home education knowledge graph construction and intelligent pushing platform according to claim 1, characterized in that, The family state modeling module is specifically configured to: process the multi-modal state data into quantitative indicators including valence of emotion, cognitive load, or physiological energy level; and construct the state vector based on the quantitative indicators. 3.The home education knowledge graph construction and intelligent pushing platform according to claim 1, characterized in that, When calculating the tension matrix, the family state modeling module is specifically configured to: for any two agents, calculate the interaction tension value between them through a quadratic function based on their respective state vectors and a preset weight matrix, and take the interaction tension value as the corresponding element of the tension matrix. 4.The home education knowledge graph construction and intelligent pushing platform according to claim 3, characterized in that, Each of the educational intervention operators is defined as a structured mathematical transformation, which specifically includes: a change vector bound to the agent, used to update the state vector of the agent through vector addition; a tension adjustment matrix, used to update the tension matrix through Hadamard product. 5.The home education knowledge graph construction and intelligent pushing platform according to claim 1, characterized in that, The platform further comprises: an operator effect calibration module configured to collect a subsequent family state tensor as an actual state after pushing the educational content corresponding to the target educational intervention operator, and update the transformation effect of the target educational intervention operator based on the difference between the actual state and a predicted state predicted by the target educational intervention operator. 6.The home education knowledge graph construction and intelligent pushing platform according to claim 1, characterized in that, When selecting the target educational intervention operator, the intervention planning module is specifically configured to: select a plurality of educational intervention operators from the operator library to form at least one time-sequential intervention sequence; simulate at least one of the intervention sequences to generate a corresponding state evolution path, wherein each state evolution path comprises a series of family state tensors arranged in chronological order.

7. The family education knowledge graph construction and intelligent pushing platform according to claim 6, characterized in that, The intervention planning module is further specifically configured to: calculate the cost of each state evolution path based on a preset path cost function; select the state evolution path with the lowest cost, and determine the first educational intervention operator in the intervention sequence corresponding to the path as the target educational intervention operator. 8.The home education knowledge graph construction and intelligent pushing platform according to claim 7, characterized in that, The path cost function at least includes an evaluation of one or more of the following: a process tension cost of the tension matrix contained in all family state tensors in the state evolution path; a final state cost of the tension matrix contained in the terminal family state tensor of the state evolution path. a path smoothness cost of a degree of change between adjacent family state tensors on the state evolution path. 9.The home education knowledge graph construction and intelligent pushing platform according to claim 8, characterized in that, a calculation manner of the path cost function is determined by the following formula: ； wherein, is the result of the computation of the path cost function for a state evolution path ; is the total number of intervention steps in the state evolution path; is the sequence number of an intervention step; is the family state tensor corresponding to the th step in the state evolution path; is the tension matrix contained in the family state tensor ; is the Frobenius norm of a matrix; is the generalized distance norm between two family state tensors; is the weight coefficient of the process tension cost; is the weight coefficient of the final state cost; is the weight coefficient of the path smoothness cost.

10. The method of smart push of home education based on the platform of any one of claims 1-9, characterized in that, comprising the following steps: collecting multi-modal state data of at least two family member agents; generating a state vector for each of the agents according to the multi-modal state data; calculating a tension matrix describing interaction tension between the agents based on a relationship between the state vectors; constructing the state vectors and the tension matrix into a family state tensor; selecting at least one target educational intervention operator from a preset operator library containing multiple educational intervention operators according to the family state tensor; pushing educational content corresponding to the target educational intervention operator to a user.

Images