Intelligent collaborative handling system for protection of minors oriented to cross-department cooperation
By combining privacy computing and graph neural networks, a cross-departmental collaborative response system was constructed, which solved the problems of data privacy and sharing contradictions, inaccurate risk identification, and low efficiency of collaborative scheduling in the protection of minors, and achieved efficient cross-departmental collaborative response and system self-optimization.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GUANGZHOU COLLEGE OF COMMERCE
- Filing Date
- 2026-04-22
- Publication Date
- 2026-07-21
Smart Images
Figure CN122434446A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of data processing, privacy computing, and intelligent collaboration, specifically to an intelligent collaborative handling system for the protection of minors that is designed for cross-departmental collaboration. Background Technology
[0002] The protection of minors involves multiple government departments, including education, public security, civil affairs, health, and justice, and is a complex and systemic project. Currently, each department typically has established independent business systems and data archives based on its own responsibilities, such as the student registration system for the education department, the household registration and public security management system for the public security department, the assistance system for the civil affairs department, and the vaccination system for the health department. However, this fragmented management model leads to the problem of "information silos," resulting in risk information about minors being scattered across various departments, making it difficult to comprehensively and promptly grasp.
[0003] In existing technologies, cross-departmental collaboration mainly faces the following technical challenges: First, the contradiction between data privacy and sharing. Data from various departments involves a large amount of personal privacy, and direct sharing faces legal obstacles and security risks. Second, risk identification methods are limited. Existing risk assessments mostly use rule engines or shallow machine learning models, which are insufficient to capture the complex risk propagation mechanisms in the social networks of minors. Third, cross-departmental task scheduling lacks dynamic optimization. It usually adopts fixed processes or manual coordination, and cannot be intelligently scheduled according to the real-time load and handling capacity of each department. Fourth, there is a lack of closed-loop evolution mechanism. The handling results fail to effectively feed back into model optimization and contingency plan adjustments, making it difficult for the system to continuously improve its efficiency.
[0004] Some existing solutions attempt to improve the above shortcomings by building a unified data platform, applying general machine learning models, and standardizing workflow engines. However, these solutions essentially only limit the application scenarios and cannot fundamentally solve the deficiencies such as cross-departmental data privacy compliance integration, accurate identification of risks to minors, dynamic collaborative scheduling, and closed-loop system evolution. Summary of the Invention
[0005] This invention aims to address the technical shortcomings of existing technologies, such as the contradiction between data privacy and sharing, inaccurate risk identification, low efficiency of cross-departmental collaborative scheduling, and the lack of closed-loop evolution capabilities in the system.
[0006] To address the aforementioned technical problems, the present invention provides a smart collaborative response system for the protection of minors, oriented towards cross-departmental cooperation, comprising: The privacy-preserving computational data fusion module is configured to collaboratively compute and generate encrypted and aligned cross-departmental feature tensors of minors through secure multi-party computation or federated learning protocols, without the original data from multiple heterogeneous data sources leaving the local machine, and to construct a privacy-preserving holographic archive with the minor entity as the core based on the feature tensors. The graph neural network risk identification module is configured to load a pre-trained graph neural network model. The input of the graph neural network model includes the attribute features of minors and their social relationship topology extracted from the privacy-protected holographic file. The output is a multi-dimensional risk probability vector. The risk assessment level is determined based on the multi-dimensional risk probability vector, and the set of key risk factors is located through a graph attention mechanism. The dynamic collaborative scheduling module is configured to receive early warning events containing the set of key risk factors, and based on a multi-objective optimization function, combined with the real-time load index, geographic distance, and handling timeliness requirements of each candidate department, solve for the Pareto optimal solution set to determine the set of target departments participating in collaborative handling and the order of collaboration, and generate a personalized collaborative handling task order for each target department; and The closed-loop evolution module is configured to receive the handling process and result information fed back by each target department, add the handling process and result information as new samples to the training set, update the parameters of the graph neural network model through incremental learning or online learning, and dynamically adjust the plan rules in the collaborative handling plan library according to the handling effect evaluation results.
[0007] In some embodiments of the present invention, the privacy-computing data fusion module further includes: The vertical federated learning unit is used to calculate feature cross statistics using local data from each data source and upload the encrypted statistics to the coordinator through homomorphic encryption or secret sharing technology. The coordinator aggregates and calculates to obtain the global feature tensor without disclosing the original individual data of each party in the process. The entity alignment unit is used to identify the same minor entity across data sources and establish a unified virtual identity based on hidden query or privacy set intersection protocol without exposing the minor identity information of each party.
[0008] In some embodiments of the present invention, the graph neural network model loaded by the graph neural network risk identification module includes: The heterogeneous graph input layer is used to represent the minor entity and its associated guardians, relatives, schools, community nodes, as well as the family relationships, school registration relationships, and social relationships between the entities; The temporal graph convolutional layer is used to capture the patterns of node attributes evolving over time, and calculates the contribution weights of different relationship types to risk propagation through a multi-head attention mechanism. The risk classification output layer is used to output the probability that the minor entity belongs to different preset risk categories, and calculates the contribution score of each node and edge to the risk classification through the gradient weighted class activation mapping method. Nodes and edges with contribution scores exceeding a preset threshold are used as the set of key risk factors.
[0009] In some embodiments of the present invention, in the dynamic cooperative scheduling module, the multi-objective optimization function comprehensively considers the following objectives: Minimize the total cross-departmental response time, minimize the workload imbalance between departments, and maximize the match between the tasks handled and the departmental functions; The solution set for Pareto optimality is obtained using a decomposition-based multi-objective evolutionary algorithm, and knowledge graphs are used to perform semantic constraint pruning on departmental functions and contingency plans during the solution process.
[0010] In some embodiments of the present invention, the closed-loop evolution module further includes: The effectiveness evaluation unit is used to calculate the risk reduction and recurrence rate indicators based on the feedback of treatment results, and compare them with similar historical cases to generate a treatment effectiveness score. The contingency plan optimization unit is used to extract high-frequency handling paths from successful cases through association rule mining when the handling effect score is lower than a preset threshold, and automatically update the collaborative process templates of the corresponding risk type in the contingency plan library.
[0011] According to another aspect of the present invention, a smart collaborative handling method for the protection of minors oriented towards cross-departmental collaboration is provided, applied to the above-mentioned system, comprising the following steps: Step A: Through secure multi-party computation or federated learning protocols, without leaving the local source of raw data from multiple heterogeneous data sources, collaborative computation generates encrypted and aligned cross-departmental feature tensors of minors, and constructs privacy-protected holographic profiles with minor entities at their core based on the feature tensors. Step B: Input the minor's attribute features and social relationship topology extracted from the privacy-protected holographic file into a pre-trained graph neural network model, output a multi-dimensional risk probability vector, and determine the risk assessment level and locate the set of key risk factors through the graph attention mechanism based on the multi-dimensional risk probability vector; Step C: When the risk assessment level exceeds the preset threshold, based on the early warning event containing the set of key risk factors, a multi-objective optimization function is used to solve the Pareto optimal solution set by combining the real-time load index, geographic distance and handling time requirements of each candidate department to determine the set of target departments participating in collaborative handling and the order of collaboration, and to generate a personalized collaborative handling task sheet for each target department. Step D: Receive the handling process and result information from each target department, add the handling process and result information as new samples to the training set, update the parameters of the graph neural network model through incremental learning or online learning, and dynamically adjust the contingency plan rules in the collaborative handling contingency plan library based on the handling effect evaluation results.
[0012] In some embodiments of the present invention, step A further includes: Through vertical federated learning, feature cross statistics are calculated using local data from each data source, and the encrypted statistics are uploaded to the coordinator through homomorphic encryption. The coordinator then aggregates and calculates the global feature tensor. By using a privacy set intersection protocol, the same minor entity across data sources can be identified and a unified virtual identity can be established without exposing the identity information of the minors of all parties.
[0013] In some embodiments of the present invention, in step B, the graph neural network model uses a heterogeneous graph input layer to represent minor entities and their associated nodes and relationships, uses a temporal graph convolutional layer to capture the evolution pattern of node attributes over time, and calculates the contribution weight of different relationship types to risk propagation through a multi-head attention mechanism. Finally, the risk classification output layer outputs the probability that the minor entity belongs to different preset risk categories, and uses a gradient-weighted class activation mapping method to locate the nodes and edges that contribute the most to risk classification as the set of key risk factors.
[0014] In some embodiments of the present invention, in step C, the multi-objective optimization function comprehensively considers minimizing the total cross-departmental response time, minimizing the workload imbalance of each department, and maximizing the matching degree between the handling task and the departmental function; a decomposition-based multi-objective evolutionary algorithm is used to solve the Pareto optimal solution set, and during the solution process, a knowledge graph is used to perform semantic constraint pruning on the departmental functions and handling plans.
[0015] In some embodiments of the present invention, step D further includes: Based on the feedback on the treatment results, calculate the risk reduction and recurrence rate indicators, compare them with similar historical cases, and generate a treatment effectiveness score. When the handling effect score is lower than a preset threshold, high-frequency handling paths are extracted from successful cases through association rule mining, and the collaborative process templates of the corresponding risk type in the contingency plan library are automatically updated.
[0016] The advantages of this invention compared to the prior art are: This invention, through a privacy-preserving computational data fusion module, employs secure multi-party computation or federated learning protocols to achieve cross-departmental collaborative data computation without disclosing the original individual data of each department, thereby constructing a privacy-preserving holographic archive. This not only breaks down "data silos" but also meets data privacy compliance requirements, resolving legal obstacles and security risks associated with the cross-departmental sharing of sensitive data of minors. This invention uses a graph neural network risk identification module to model the risks of minors within their social networks. It captures complex social relationship topologies and risk propagation mechanisms using heterogeneous graphs and temporal graph convolutional layers. Compared to traditional rule engines or shallow machine learning models, it can better identify potential risks and locate key risk factors through graph attention mechanisms. This invention achieves dynamic and intelligent optimization of cross-departmental task scheduling by using a dynamic collaborative scheduling module, employing multi-objective optimization functions and Pareto optimal solution sets, and comprehensively considering multi-dimensional factors such as real-time load of each department, geographical distance, and timeliness requirements. This invention uses a closed-loop evolution module to feed back the handling results to the incremental updates of the graph neural network model and the dynamic adjustment of the contingency plan library, forming a complete technical closed loop of "identification-scheduling-handling-feedback-optimization". This enables the system to have the ability to continuously learn and self-evolve, and the accuracy of risk identification and the efficiency of collaborative handling can be continuously improved as the running time increases. Attached Figure Description
[0017] Figure 1 This is a structural block diagram of an intelligent collaborative response system for the protection of minors, designed for cross-departmental cooperation, provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the multi-objective optimization process of the dynamic collaborative scheduling module provided in an embodiment of the present invention; Figure 3 The overall flowchart of the intelligent collaborative handling method for the protection of minors with cross-departmental cooperation provided in the embodiments of the present invention is shown.
[0018] Figure 4 The flowchart of step S1 of the intelligent collaborative handling method for the protection of minors with cross-departmental cooperation provided in the embodiments of the present invention.
[0019] Figure 5 The flowchart of step S2 of the intelligent collaborative handling method for the protection of minors with cross-departmental cooperation provided in the embodiments of the present invention.
[0020] Figure 6 The flowchart of step S3 of the intelligent collaborative handling method for the protection of minors with cross-departmental cooperation provided in the embodiments of the present invention.
[0021] Figure 7The flowchart of step S4 of the intelligent collaborative handling method for the protection of minors with cross-departmental cooperation provided in the embodiments of the present invention. Detailed Implementation
[0022] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.
[0023] It should be noted that the terms "comprising" and "having" and any variations thereof in the specification, claims and accompanying drawings of this invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such processes, methods, products or devices.
[0024] Example 1 like Figure 1 As shown, this embodiment provides an intelligent collaborative handling system 100 for the protection of minors in a cross-departmental manner, including a privacy computing data fusion module 110, a graph neural network risk identification module 120, a dynamic collaborative scheduling module 130, and a closed-loop evolution module 140.
[0025] The privacy-preserving computational data fusion module 110 is configured to generate encrypted and aligned cross-departmental feature tensors for minors through secure multi-party computation or federated learning protocols, without the original data from multiple heterogeneous data sources (such as the student registration system 210 of the education department, the household registration system 220 of the public security department, the assistance system 230 of the civil affairs department, the medical system 240 of the health department, and the case system 250 of the judicial department) leaving the local area. Based on the feature tensors, a privacy-preserving holographic archive 111 with the minor entity as the core is constructed.
[0026] The privacy-preserving computational data fusion module 110 further includes a vertical federated learning unit 112 and an entity alignment unit 113.
[0027] The vertical federated learning unit 112 is used to calculate feature cross statistics using local data from each data source and upload the encrypted statistics to the coordinator through homomorphic encryption or secret sharing technology. The coordinator aggregates and calculates the global feature tensor without disclosing the original individual data of each party. Specifically, assuming there are K data source parties, each data source party i holds different feature dimensions X_i of the same set of underage samples. Each party first standardizes and encodes the features locally, and then encrypts the local feature cross terms based on a homomorphic encryption algorithm (such as Paillier encryption or CKKS scheme) to generate Enc(X_i^TX_i) and Enc(X_i^TY) (where Y is a virtual label, which can be provided by the coordinator or generated through negotiation among the parties). The coordinator aggregates the encrypted statistics uploaded by each party to obtain Enc(∑X_i^TX_i) and Enc(∑X_i^TY), and obtains the global feature covariance matrix after decryption, thereby training the global model parameters. Throughout the process, the coordinator can only access the encrypted statistics and cannot know the original feature data of any party.
[0028] The entity alignment unit 113 is used to identify the same minor entity across data sources and establish a unified virtual identity based on a hidden query or privacy set intersection protocol, without exposing the minor identity information of each party. Specifically, each party first hashes its own minor identity (such as the hash value of a combination of name, date of birth, and guardian information), and then uses a privacy set intersection protocol based on unintentional transmission to jointly calculate the virtual identity in the intersection. For example, a PSI protocol based on Diffie-Hellman key exchange can be used. Each party blinds its own identity set and sends it to the coordinator. The coordinator determines the shared minor entity by calculating the intersection of the blinded values and assigns a globally unique virtual ID to each shared entity. This virtual ID is only used for internal system association calculations and does not map back to any party's original identity.
[0029] Through the aforementioned privacy computing architecture, the privacy computing data fusion module 110 can construct a holographic profile 111 of minors containing multi-dimensional features while protecting the data privacy of various departments. This profile is indexed by a virtual identity and includes the minor's basic attribute features (age, gender, etc.), family features (guardian information, economic status, etc.), educational features (school status, attendance records, etc.), health features (vaccination, medical records, etc.), legal features (case-related information, etc.), and cross-statistics calculated from these features.
[0030] The graph neural network risk identification module 120 is configured to load a pre-trained graph neural network model 121. The input of the graph neural network model 121 includes the minor attribute features and social relationship topology extracted from the privacy-protected holographic file 111. The output is a multi-dimensional risk probability vector. The risk assessment level is determined based on the multi-dimensional risk probability vector, and the key risk factor set 122 is located through the graph attention mechanism.
[0031] The graph neural network model 121 includes a heterogeneous graph input layer 1211, a temporal graph convolutional layer 1212, and a risk classification output layer 1213.
[0032] The heterogeneous graph input layer 1211 is used to represent minor entities and their associated entity nodes, as well as the relationship edges between entities. Specifically, the constructed heterogeneous graph G=(V,E,R) contains multiple types of nodes V and multiple types of relationship edges E. The node types include minor nodes, guardian nodes, relative nodes, school nodes, community nodes, etc., and the relationship edge types include family relationships (minor-guardian), school registration relationships (minor-school), residence relationships (minor-community), kinship relationships (minor-relative), etc. Each node v∈V is associated with an initial feature vector h_v^(0), which is extracted from the privacy-protected holographic archive 111. For example, the features of minor nodes include age, gender, health status, etc., and the features of school nodes include school type, geographical location, student size, etc.
[0033] The temporal graph convolutional layer 1212 is used to capture patterns in the evolution of node attributes over time and calculates the contribution weights of different relationship types to risk propagation through a multi-head attention mechanism. The temporal graph convolutional layer 1212 employs the following update rules: h_v^(l+1)=σ(∑_{r∈R}∑_{u∈N_v^r}α_{vu}^rW_r^(l)h_u^(l)+β_v^(l)h_v^(l)) Where h_v^(l) represents the feature representation of node v at layer l, N_v^r represents the set of neighbor nodes connected to node v through relation r, W_r^(l) is the trainable weight matrix of relation r at layer l, α_{vu}^r is the attention weight calculated by the multi-head attention mechanism, representing the importance of neighbor node u to node v through relation r, β_v^(l) is the weight of the self-loop connection, and σ is the nonlinear activation function.
[0034] Meanwhile, the temporal graph convolutional layer 1212 introduces a temporal encoding mechanism. For each node v, its historical time series features are encoded through a gated recurrent unit or a long short-term memory network to obtain a temporal feature representation h_v^(time), which is then fused with the feature representation of the current layer. h_v^(l+1,combined)=γ·h_v^(l+1)+(1-γ)·h_v^(time) Here, γ is the equilibrium parameter, which can be learned through training.
[0035] The risk classification output layer 1213 outputs the probability that the minor entity belongs to different preset risk categories, and uses a gradient-weighted class activation mapping method to locate the nodes and edges that contribute the most to the risk classification as the key risk factor set 122. Specifically, assuming there are C risk categories (such as dropout risk, domestic violence risk, left-behind risk, psychological crisis risk, and crime risk), the risk classification output layer outputs a probability vector p∈R^C using a softmax function. For a given target minor node v, its risk level is determined by argmax(p). To locate the key risk factors, a gradient-weighted class activation mapping method is used to calculate the gradient of risk category c with respect to each node and edge in the graph structure, thereby obtaining the contribution score of each node and edge to the classification result c. Score_node(u)=ReLU(∑_kα_k^c·A_u^k) Score_edge(u,r,v)=ReLU(∑_kβ_k^c·E_{u,r,v}^k) Where α_k^c and β_k^c represent the weights of the k-th feature channel to category c, A_u^k is the activation value of node u in the k-th channel, and E_{u,r,v}^k is the activation value of edge (u,r,v) in the k-th channel. Nodes and edges whose contribution scores exceed a preset threshold are extracted and constitute a key risk factor set 122, such as {"Father node contribution 0.85", "Family relationship edge contribution 0.76", "School absence record node contribution 0.92"}.
[0036] The dynamic collaborative scheduling module 130 is configured to receive an early warning event 123 containing the set of key risk factors 122, and based on a multi-objective optimization function, combine the real-time load index, geographical distance and handling time requirements of each candidate department to solve the Pareto optimal solution set to determine the set of target departments participating in collaborative handling and the order of collaboration, and generate a personalized collaborative handling task sheet 134 for each target department.
[0037] like Figure 2 As shown, the dynamic collaborative scheduling module 130 includes a multi-objective optimization unit 131 and a task order generation unit 132.
[0038] The multi-objective optimization unit 131 constructs the following multi-objective optimization function: minF(x)=(f_1(x),f_2(x),-f_3(x)) Here, x is a decision variable, representing the selection of the target department set and the task allocation scheme.
[0039] f_1(x)=∑_{j∈S}t_j(x)+T_transfer(x) represents the total cross-department response time, where t_j(x) is the estimated processing time for department j to execute the assigned task, and T_transfer(x) is the task handover time between departments.
[0040] f_2(x)=√((1 / |S|)∑_{j∈S}(L_j-L_avg)^2) represents the workload imbalance of each department, where L_j is the current load index of department j (which can be calculated based on the number of pending tasks, the urgency of tasks, etc.), and L_avg is the average load of participating departments.
[0041] f_3(x)=∑_{j∈S}match_score(j,task_j) represents the matching degree between the task and the department's function. The match_score is calculated by the semantic similarity of the functions in the knowledge graph.
[0042] The multi-objective optimization unit 131 employs a decomposition-based multi-objective evolutionary algorithm to solve for the Pareto optimal solution set. Specifically, the MOEA / D algorithm is used to decompose the multi-objective optimization problem into several scalar subproblems, each subproblem corresponding to a set of weight vectors. These subproblems are then collaboratively optimized through an evolutionary algorithm. During the solution process, a pre-constructed departmental function knowledge graph is used for semantic constraint pruning: for each candidate department j, it is checked whether it possesses the functional authority to handle the corresponding task_j (inferred through the "possesses functional authority" relationship in the knowledge graph). If it does not possess the authority, it is directly excluded from the candidate set, significantly reducing the search space.
[0043] The task order generation unit 132 generates a personalized collaborative handling task order 134 for each target department selected from the Pareto optimal solution set. The task order 134 includes at least the following fields: event ID, warning event description, list of key risk factors, list of pending items (automatically generated based on departmental functions and risk factors), handling time limit, associated virtual identity identifier and data access token for minors (generated based on attribute-based encryption, allowing access only to data fields necessary to complete the task), list of collaborating partner departments and their contact information.
[0044] The closed-loop evolution module 140 is configured to receive the handling process and result information 141 fed back by each target department, add the handling process and result information 141 as new samples to the training set, update the parameters of the graph neural network model 121 through incremental learning or online learning, and dynamically adjust the plan rules in the collaborative handling plan library according to the handling effect evaluation results.
[0045] The closed-loop evolution module 140 further includes an information update unit 142, an effect evaluation unit 143, and a contingency plan optimization unit 144.
[0046] The information update unit 142 is used to structure the received handling process and result information 141 and append it to the minor's holographic file 111. The handling process and result information 141 includes the execution records of each department (such as home visit time, home visit record, measures taken, difficulties encountered, etc.), the final handling result (such as "student has returned to school", "domestic violence risk has been eliminated", etc.), and follow-up suggestions. This information is converted into time series features and updated to the corresponding minor's time series feature vector, providing historical basis for subsequent risk assessment.
[0047] The effectiveness evaluation unit 143 is used to calculate the risk reduction magnitude and recurrence rate indicators based on the feedback treatment result information, and compare them with similar historical cases to generate a treatment effectiveness score. Specifically, the risk reduction magnitude ΔR is defined as R_before - R_after, where R_before is the highest risk value in the risk probability vector before treatment, and R_after is the highest risk value in the risk probability vector reassessed after treatment. The recurrence rate indicator calculates whether the warning is triggered again within a certain period of time after treatment (e.g., 3 months, 6 months). By comparing with similar historical cases, a relative effectiveness score is calculated. score=w_1·(ΔR / ΔR_avg)+w_2·(1-recidivism_flag)+w_3·(T_response / T_response_avg) Where ΔR_avg is the average risk reduction of similar historical cases, recidivism_flag is the recurrence marker, T_response is the response time of this case, T_response_avg is the average response time of similar historical cases, and w_1, w_2, and w_3 are weighting coefficients.
[0048] The contingency plan optimization unit 144 is used to extract high-frequency handling paths from successful cases (cases with handling effect scores higher than 0.8) through association rule mining when the handling effect score is lower than a preset threshold (e.g., 0.6), and automatically update the collaborative process templates of the corresponding risk type in the contingency plan library. Specifically, the FP-Growth algorithm is used to mine frequent itemsets from successful cases, and each itemset is represented as a set of (risk type, department combination, task sequence, handling duration). The mined high-frequency itemsets are used to update the collaborative process templates in the contingency plan library. For example, for the "domestic violence risk" case, if the high-frequency itemset shows that the combination of "public security-women's federation-civil affairs" and the task sequence of "public security issue warning-women's federation psychological intervention-civil affairs temporary assistance" has the best handling effect, then the corresponding contingency plan template is updated and the process is set as the default recommended process.
[0049] Furthermore, the closed-loop evolution module 140 updates the parameters of the graph neural network model 121 through incremental learning. After accumulating a certain number of new samples (e.g., 100), the system employs an elastic weight consolidation algorithm to fine-tune the model using new samples while retaining historical knowledge. Specifically, for model parameters θ, the importance metric Ω_i of each parameter to historical tasks is calculated, and a regularization term is added during incremental learning. L(θ)=L_new(θ)+λ∑_iΩ_i(θ_i-θ_i^old)^2 Where L_new(θ) is the loss function on new samples, θ_i^old is the historical optimal parameter, and λ is the regularization coefficient. In this way, the model can adapt to new samples without drastically changing the parameters that are important for historical samples, thus avoiding catastrophic forgetting.
[0050] Example 2 like Figures 3 to 7 As shown, this embodiment provides an intelligent collaborative handling method for the protection of minors through cross-departmental cooperation. This method can be applied to the system 100 in Embodiment 1, and includes the following steps S1 to S4. It should be noted that the technical details of the method described in this embodiment correspond to those of the system described in Embodiment 1, and the same content will not be repeated.
[0051] Step S1: Privacy-preserving computational data fusion to construct a holographic profile of minors. Through secure multi-party computation or federated learning protocols, without the original data from multiple heterogeneous data sources leaving the local machine, a cryptographically aligned cross-departmental feature tensor for minors is generated through collaborative computation. Based on the feature tensor, a privacy-protected holographic profile centered on the minor entity is constructed.
[0052] Specifically, this step includes: Sub-step S11: Through vertical federated learning, feature cross statistics are calculated at each data source using local data, and the encrypted statistics are uploaded to the coordinator via homomorphic encryption. The coordinator aggregates and calculates the global feature tensor. Each data source uses the Paillier homomorphic encryption algorithm to encrypt its local feature covariance matrix. The coordinator performs additive homomorphic aggregation on the encrypted result and then decrypts it to obtain the global feature covariance matrix. Finally, matrix factorization is used to obtain the global feature representation of each sample.
[0053] Sub-step S12: Using a privacy set intersection protocol, identify the same minor entity across data sources and establish a unified virtual identity without exposing the minor's identity information of each party. Each party uses the PSI protocol based on Diffie-Hellman key exchange to blind their own identity set before sending it to the coordinator. The coordinator calculates the intersection of the blinded values and assigns a globally unique virtual ID to each intersection element.
[0054] Sub-step S13: Associate the global feature tensor with the virtual ID to construct a privacy-protected holographic profile indexed by the virtual ID. The profile contains multi-dimensional information such as the minor's attribute characteristics, family characteristics, educational characteristics, health characteristics, and legal characteristics.
[0055] Step S2: Graph neural network risk identification to locate key risk factors. The attribute features of minors and their social relationship topology extracted from the privacy-protected holographic archives are input into a pre-trained graph neural network model, which outputs a multidimensional risk probability vector. The risk assessment level is determined based on the multidimensional risk probability vector, and the set of key risk factors is located through a graph attention mechanism.
[0056] Specifically, this step includes: Sub-step S21: Construct a heterogeneous graph, with minor entities and their associated guardians, relatives, schools, and communities as nodes, and family relationships, school registration relationships, residential relationships, and social relationships as edges, and extract the initial feature vectors of each node from the privacy-protected holographic archive.
[0057] Sub-step S22: Input the heterogeneous graph into a pre-trained graph neural network model, pass it through multiple layers of temporal graph convolution, and capture the patterns of node attribute evolution over time and the risk propagation mechanism in the graph structure. The model uses a multi-head attention mechanism to calculate the contribution weights of different relationship types to risk propagation.
[0058] Sub-step S23: Through the risk classification output layer, obtain the probability vectors of minor entities belonging to different preset risk categories, and take the category corresponding to the highest probability as the risk assessment level.
[0059] Sub-step S24: Using the gradient-weighted class activation mapping method, calculate the contribution score of each node and edge to the risk classification result, and extract the nodes and edges whose contribution scores exceed the preset threshold to form a set of key risk factors. For example, output {"Father node domestic violence record contribution score 0.85", "School absence record node contribution score 0.92", "Family relationship edge contribution score 0.76"}.
[0060] Sub-step S25: When the risk assessment level exceeds a preset threshold (such as "medium risk" and above), an early warning event is generated. This event includes the minor's virtual ID, risk assessment level, set of key risk factors, and associated access token for the minor's holographic profile.
[0061] Step S3: Multi-objective dynamic collaborative scheduling to generate personalized task lists. When the risk assessment level exceeds a preset threshold, based on the early warning event containing the set of key risk factors, a multi-objective optimization function is used to solve the Pareto optimal solution set by combining the real-time load index, geographical distance and handling time requirements of each candidate department to determine the set of target departments participating in collaborative handling and the order of collaboration, and to generate a personalized collaborative handling task sheet for each target department.
[0062] Specifically, this step includes: Sub-step S31: Analyze the early warning event, obtain the risk type and key risk factor set, and determine the core functions required for handling.
[0063] Sub-step S32: Based on the geographic information system, select a list of candidate departments within the jurisdiction where the target minor resides.
[0064] Sub-step S33: Construct a multi-objective optimization function that comprehensively considers minimizing the total cross-departmental response time, minimizing the workload imbalance of each department, and maximizing the matching degree between the handling tasks and departmental functions.
[0065] Sub-step S34: Use a decomposition-based multi-objective evolutionary algorithm to solve for the Pareto optimal solution set. During the solution process, semantic constraint pruning is performed using a pre-constructed departmental function knowledge graph to exclude candidate departments that do not possess the corresponding functional authority.
[0066] Sub-step S35: Select the final target department set and collaboration order from the Pareto optimal solution set. The selection strategy can be set according to actual needs, such as selecting the solution most similar to historical successful cases, or selecting the solution with the optimal load balancing among departments.
[0067] Sub-step S36: For each department in the target department set, generate a personalized collaborative task list based on its statutory responsibilities, key risk factors, and current workload. The task list includes information such as event description, key risk factors, to-do list, processing time limit, data access permissions, and a list of collaborating departments.
[0068] Sub-step S37: Dispatch the task order to the business systems or terminals of each target department through the security interface.
[0069] Step S4: Handle feedback and close-loop evolution, update the model and contingency plan. The system receives information on the handling process and results from various target departments, adds this information as new samples to the training set, updates the parameters of the graph neural network model through incremental learning or online learning, and dynamically adjusts the rules of the collaborative handling plan library based on the evaluation results of the handling effect.
[0070] Specifically, this step includes: Sub-step S41: Receive feedback from each target department regarding the handling process and results, including execution records, handling results, and follow-up suggestions. Structure the information and extract key fields.
[0071] Sub-step S42: Add the structured disposal information to the minor's holographic file and update its historical time-series feature vector.
[0072] Sub-step S43: Based on the treatment results information, calculate the risk reduction rate and recurrence rate indicators, compare them with similar historical cases, and generate a treatment effect score.
[0073] Sub-step S44: When the handling effect score is lower than the preset threshold, the contingency plan optimization mechanism is triggered. The FP-Growth algorithm is used to mine high-frequency handling paths from successful cases (effect scores higher than 0.8), extract frequent itemsets (risk type, department combination, task sequence, handling duration), and automatically update the corresponding risk type collaborative process template in the contingency plan library.
[0074] Sub-step S45: Trigger incremental model updates periodically (e.g., weekly or every 100 new samples). Employ an elastic weight consolidation algorithm to fine-tune the graph neural network model with new samples while preserving historical knowledge.
[0075] Sub-step S46: Blockchain-based evidence storage is performed on the entire lifecycle of the warning event from its generation to its closure to ensure the immutability and traceability of the handling process.
[0076] Through the above steps S1 to S4, the present invention realizes a complete closed loop of cross-departmental collaborative handling of minor protection. From data fusion for privacy protection, accurate risk identification, intelligent dynamic scheduling, to feedback-driven closed-loop evolution, each link is closely coupled to form an intelligent system that can continuously learn and self-optimize.
[0077] To verify the technical effectiveness of this invention, the system was deployed in the protection of minors in a certain city and compared with existing solutions based on general big data platforms and rule engines. The testing period was 6 months, covering five departments: education, public security, civil affairs, health, and justice, involving approximately 500,000 data entries related to minors. The test results are as follows: 1. Data Fusion Efficiency: Employing a privacy-computing data fusion module, and fully complying with data privacy regulations, the system successfully fused 23 types of features from 5 departments to construct holographic profiles of over 450,000 minors. Compared to traditional plaintext aggregation solutions, the additional overhead of privacy computation was only 23%, while avoiding the risk of data leakage.
[0078] 2. Risk Identification Accuracy: The graph neural network risk identification module achieved an early warning accuracy of 91.5% (precision 89.2%, recall 93.7%), a significant improvement compared to rule engine solutions (precision 67.3%, recall 58.6%) and logistic regression solutions (precision 76.8%, recall 71.4%). The advantage of graph neural networks is particularly evident in identifying complex risks (such as school dropout + domestic violence + poverty), with an accuracy 28 percentage points higher than the baseline solution.
[0079] 3. Collaborative scheduling efficiency: The dynamic collaborative scheduling module reduced the average cross-departmental response time from 3.8 days to 1.2 days, a reduction of 68%; the standard deviation of workload in each department decreased from 0.42 to 0.18, and the load balance improved by 57%; the matching degree between tasks and departmental functions increased from 72% in manual scheduling to 94%.
[0080] 4. Closed-loop evolution effect: After 6 months of incremental learning and contingency plan optimization, the model's early warning accuracy increased from the initial 86.3% to 91.5%, an increase of 5.2 percentage points; the success rate of the contingency plan (no recurrence within 6 months after treatment) increased from 73% to 85%, an increase of 12 percentage points.
[0081] The above verification results show that the present invention has achieved good technical results in the field of cross-departmental collaboration in the protection of minors by organically integrating non-general technical means such as privacy computing, graph neural networks, multi-objective optimization, and closed-loop evolution.
[0082] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.
Claims
1. A smart collaborative handling system for the protection of minors, characterized in that, include: The privacy-preserving computational data fusion module is configured to collaboratively compute and generate encrypted and aligned cross-departmental feature tensors of minors through secure multi-party computation or federated learning protocols, without the original data from multiple heterogeneous data sources leaving the local machine, and to construct a privacy-preserving holographic archive with the minor entity as the core based on the feature tensors. The graph neural network risk identification module is configured to load a pre-trained graph neural network model. The input of the graph neural network model includes the attribute features of minors and their social relationship topology extracted from the privacy-protected holographic file. The output is a multi-dimensional risk probability vector. The risk assessment level is determined based on the multi-dimensional risk probability vector, and the set of key risk factors is located through a graph attention mechanism. The dynamic collaborative scheduling module is configured to receive early warning events containing the set of key risk factors, and based on a multi-objective optimization function, combine the real-time load index, geographic distance and handling time requirements of each candidate department to solve the Pareto optimal solution set to determine the set of target departments participating in collaborative handling and the order of collaboration, and generate a personalized collaborative handling task sheet for each target department. as well as The closed-loop evolution module is configured to receive the handling process and result information fed back by each target department, add the handling process and result information as new samples to the training set, update the parameters of the graph neural network model through incremental learning or online learning, and dynamically adjust the plan rules in the collaborative handling plan library according to the handling effect evaluation results.
2. The intelligent collaborative handling system for the protection of minors oriented towards cross-departmental cooperation as described in claim 1, characterized in that, The privacy-preserving computational data fusion module further includes: The vertical federated learning unit is used to calculate feature cross statistics using local data from each data source and upload the encrypted statistics to the coordinator through homomorphic encryption or secret sharing technology. The coordinator aggregates and calculates to obtain the global feature tensor without disclosing the original individual data of each party in the process. The entity alignment unit is used to identify the same minor entity across data sources and establish a unified virtual identity based on hidden query or privacy set intersection protocol without exposing the minor identity information of each party.
3. The intelligent collaborative handling system for the protection of minors oriented towards cross-departmental cooperation as described in claim 1, characterized in that, The graph neural network model loaded by the graph neural network risk identification module includes: The heterogeneous graph input layer is used to represent the minor entity and its associated guardians, relatives, schools, community nodes, as well as the family relationships, school registration relationships, and social relationships between the entities; The temporal graph convolutional layer is used to capture the patterns of node attributes evolving over time, and calculates the contribution weights of different relationship types to risk propagation through a multi-head attention mechanism. The risk classification output layer is used to output the probability that the minor entity belongs to different preset risk categories, and calculates the contribution score of each node and edge to the risk classification through the gradient weighted class activation mapping method. Nodes and edges with contribution scores exceeding a preset threshold are used as the set of key risk factors.
4. The intelligent collaborative handling system for the protection of minors oriented towards cross-departmental cooperation as described in claim 1, characterized in that, In the dynamic collaborative scheduling module, the multi-objective optimization function comprehensively considers the following objectives: minimize Total cross-departmental response time, minimizing workload imbalance among departments, and maximizing the match between tasks and departmental functions; The solution set for Pareto optimality is obtained using a decomposition-based multi-objective evolutionary algorithm, and knowledge graphs are used to perform semantic constraint pruning on departmental functions and contingency plans during the solution process.
5. The intelligent collaborative handling system for the protection of minors oriented towards cross-departmental cooperation as described in claim 1, characterized in that, The closed-loop evolution module also includes: The effectiveness evaluation unit is used to calculate the risk reduction and recurrence rate indicators based on the feedback of treatment results, and compare them with similar historical cases to generate a treatment effectiveness score. The contingency plan optimization unit is used to extract high-frequency handling paths from successful cases through association rule mining when the handling effect score is lower than a preset threshold, and automatically update the collaborative process templates of the corresponding risk type in the contingency plan library.
6. A smart collaborative handling method for the protection of minors oriented towards cross-departmental collaboration, applied to the system described in any one of claims 1 to 5, characterized in that, Includes the following steps: Step A: Through secure multi-party computation or federated learning protocols, without leaving the local source of raw data from multiple heterogeneous data sources, collaborative computation generates encrypted and aligned cross-departmental feature tensors of minors, and constructs privacy-protected holographic profiles with minor entities at their core based on the feature tensors. Step B: Input the minor's attribute features and social relationship topology extracted from the privacy-protected holographic file into a pre-trained graph neural network model, output a multi-dimensional risk probability vector, and determine the risk assessment level and locate the set of key risk factors through the graph attention mechanism based on the multi-dimensional risk probability vector; Step C: When the risk assessment level exceeds the preset threshold, based on the early warning event containing the set of key risk factors, a multi-objective optimization function is used to solve the Pareto optimal solution set by combining the real-time load index, geographic distance and handling time requirements of each candidate department to determine the set of target departments participating in collaborative handling and the order of collaboration, and to generate a personalized collaborative handling task sheet for each target department. Step D: Receive the handling process and result information from each target department, add the handling process and result information as new samples to the training set, update the parameters of the graph neural network model through incremental learning or online learning, and dynamically adjust the contingency plan rules in the collaborative handling contingency plan library based on the handling effect evaluation results.
7. The intelligent collaborative handling method for the protection of minors oriented towards cross-departmental cooperation as described in claim 6, characterized in that, Step A further includes: Through vertical federated learning, feature cross statistics are calculated using local data from each data source, and the encrypted statistics are uploaded to the coordinator through homomorphic encryption. The coordinator then aggregates and calculates the global feature tensor. By using a privacy set intersection protocol, the same minor entity across data sources can be identified and a unified virtual identity can be established without exposing the identity information of the minors of all parties.
8. The intelligent collaborative handling method for the protection of minors oriented towards cross-departmental cooperation as described in claim 6, characterized in that, In step B, the graph neural network model uses a heterogeneous graph input layer to represent minor entities and their associated nodes and relationships, uses a temporal graph convolutional layer to capture the evolution pattern of node attributes over time, and uses a multi-head attention mechanism to calculate the contribution weight of different relationship types to risk propagation. Finally, the risk classification output layer outputs the probability that the minor entity belongs to different preset risk categories, and uses a gradient-weighted class activation mapping method to locate the nodes and edges that contribute the most to risk classification as the set of key risk factors.
9. The intelligent collaborative handling method for the protection of minors oriented towards cross-departmental cooperation as described in claim 6, characterized in that, In step C, the multi-objective optimization function comprehensively considers minimizing the total cross-departmental response time, minimizing the workload imbalance of each department, and maximizing the matching degree between the handling tasks and departmental functions; a decomposition-based multi-objective evolutionary algorithm is used to solve the Pareto optimal solution set, and a knowledge graph is used to perform semantic constraint pruning on departmental functions and handling plans during the solution process.
10. The intelligent collaborative handling method for the protection of minors oriented towards cross-departmental cooperation as described in claim 6, characterized in that, Step D further includes: Based on the feedback on the treatment results, calculate the risk reduction and recurrence rate indicators, compare them with similar historical cases, and generate a treatment effectiveness score. When the handling effect score is lower than a preset threshold, high-frequency handling paths are extracted from successful cases through association rule mining, and the collaborative process templates of the corresponding risk type in the contingency plan library are automatically updated.