High-net-value crowd risk intelligent assessment system based on multi-dimensional dynamic interaction

Abstract

The invention discloses a high-net-value crowd risk intelligent assessment system based on multi-dimensional dynamic interaction, and belongs to the technical field of medical health information. Comprising a data acquisition module for acquiring and quantifying multi-dimensional risk factors; the weight distribution module is used for carrying out dynamic weight distribution based on the multi-dimensional risk factors; the data processing module is used for constructing and training a hybrid reinforcement learning model, taking the result of the dynamic weight distribution as the input of the hybrid reinforcement learning model, and outputting a risk scoring result; and the feedback module matches an intervention strategy based on the risk scoring result. According to the method, the problems of static property, low dimension and intervention lag of a traditional method are effectively solved.

Description

Technical Field

[0001] This invention relates to the field of medical and health information technology, and more specifically to a high-net-worth individual risk intelligent assessment system based on multi-dimensional dynamic interaction. Background Technology

[0002] High-net-worth individuals generally refer to those with an annual income of ≥5 million yuan. As important contributors to the core productivity of society, their health status has a significant impact on the social economy.

[0003] Currently, the incidence of cardiovascular and cerebrovascular diseases among high-net-worth individuals is alarming. Relevant statistics show that the incidence rate of these diseases is increasing by an average of 8.5% annually, with young and middle-aged individuals aged 45-55 accounting for over 60%. Furthermore, sudden occurrences of cardiovascular and cerebrovascular diseases such as stroke, acute myocardial infarction, and aortic dissection result in direct economic losses exceeding tens of billions of yuan annually.

[0004] Traditional risk assessment systems for high-net-worth individuals have several shortcomings. First, in terms of weighting, traditional methods use static weights, failing to consider the dynamic changes in risk factors and their interactions. This results in lagging assessments that cannot promptly reflect the true risk situation of high-net-worth individuals. Second, from an assessment perspective, traditional systems rely solely on basic physiological indicators, lacking analysis of deeper risk factors such as behavior, psychology, and imaging, making it difficult to comprehensively and deeply assess the risks faced by high-net-worth individuals. Finally, in terms of intervention strategy development, traditional assessment systems cannot dynamically adjust risk management plans based on real-time assessment results, failing to meet the complex and ever-changing risk prevention and control needs of high-net-worth individuals.

[0005] Therefore, providing a high-net-worth individual risk intelligent assessment system that can comprehensively consider the interaction of multi-dimensional dynamic factors has become a key issue that urgently needs to be addressed in this field. Summary of the Invention

[0006] In view of this, the present invention provides a high-net-worth individual risk intelligent assessment system based on multi-dimensional dynamic interaction, which solves the problems of staticity, low dimensionality, low efficiency and low accuracy of traditional assessment systems.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A high-net-worth individual risk intelligent assessment system based on multi-dimensional dynamic interaction includes: The data acquisition module collects and quantifies multi-dimensional risk factors; The weight allocation module is connected to the data acquisition module and performs dynamic weight allocation based on multi-dimensional risk factors. The data processing module is connected to the weight allocation module to construct and train a hybrid reinforcement learning model, wherein the result of the dynamic weight allocation is used as the input of the hybrid reinforcement learning model and the risk score result is output. The feedback module is connected to the data processing module and matches intervention strategies based on the risk score results.

[0008] Furthermore, it also includes: an adjustment configuration module, connected to the weight allocation module, to obtain the intervention effect value of the current intervention strategy, and to adjust the weight allocation of the weight allocation module based on the intervention effect value.

[0009] Furthermore, the multidimensional risk factors include: basic factors, physiological factors, biochemical factors, behavioral factors, psychological factors, imaging factors, and dynamic monitoring factors.

[0010] Furthermore, the weight allocation module performs dynamic weight allocation based on multi-dimensional risk factors, including calculating weights and risk interaction effects; The expression for calculating the weights is as follows: ; In the formula, with ( t ) is the first i Each risk factor in t Dynamic weights at time points w i0 For the first i The initial weights of each risk factor; l This is the weight decay coefficient; t For time variables; Δ w i When the event is triggered i The weight adjustment amount for each risk factor; I (Event Trigger) is an event trigger indicator function; it is 1 when an event is triggered, and 0 otherwise. The expression for calculating dynamic interaction effects is: ; In the formula, Risk_Score is the risk score; i Representing different categories of risk factors; w i ( t ) is the first i Each risk factor in t Dynamic weights at time points; X i For the first i The values ​​of each risk factor; The interaction coefficient between risk factors j and k; X j ,X k Let be the values ​​of risk factor j and risk factor k, respectively.

[0011] Furthermore, the structure of the hybrid reinforcement learning model includes: Preprocessing layer: Standardizes the multi-dimensional risk factors to obtain a standardized risk factor vector; Risk factor relationship graph construction layer: Input standardized risk factor vectors, output risk factor graph; The transitive layer takes the risk factor graph as input and constructs a high-order representation of each risk factor. Classification decision layer: Risk classification is performed using the LightGBM classifier.

[0012] Furthermore, the preprocessing layer: standardizes the multi-dimensional risk factors to obtain a standardized risk factor vector, including: Each risk factor is standardized using the following formula: ; In the formula, and Let be the mean and standard deviation of the i-th risk factor; The standardized risk factor vector is: ; In the formula, m is the total number of risk factor vectors.

[0013] Furthermore, the risk factor relationship graph construction layer: inputting standardized risk factor vectors, constructing a risk factor graph, including: Each risk factor is considered as a node in a graph, and the node's characteristics are the standardized risk factor values. Construct edges based on the correlation between risk factors, and let the correlation coefficient between risk factors i and j be: ; when At that time, among them, To set a preset relevance threshold, an edge is created between nodes i and j, with the edge weight being... Obtain and output the risk factor map. , where V is the node set and E is the edge set.

[0014] Furthermore, the transmission layer, taking the risk factor graph as input, constructs a high-order representation of each risk factor, including: For each node The set of neighboring nodes is ; In the Layers, nodes The update formula is: ; In the formula, Is node v at the th l Layer representation, It is a learnable weight matrix. Is node u in the th order? l The representation of layer -1 For the set of neighboring nodes; It is an activation function. It is the weight of the edge between nodes v and u. Is node u in the th order? l Representation of layer -1; By stacking multiple GNN layers, information from neighboring nodes is gradually aggregated to obtain a high-order representation of each risk factor: ; In the formula, L is the number of layers in the GNN.

[0015] Furthermore, the classification decision layer uses the LightGBM classifier for risk grading, including: Let the adjusted risk factor weight vector be... The weighted risk factor is then expressed as: ,in, This indicates element-wise multiplication; The LightGBM classifier is used to classify the weighted risk factor representation and output the risk level.

[0016] Furthermore, the loss function of the hybrid reinforcement learning model is: ; In the formula, L is the cross-entropy loss of the cross-entropy classification task; l Θ represents the L2 regularization coefficient; Θ represents the set of trainable parameters of the model. m The KL divergence penalty coefficient; P 模型 The probability of the risk distribution predicted by the model; P 先验 Let be the prior probability of the risk distribution.

[0017] As can be seen from the above technical solutions, compared with the prior art, the present invention discloses a high-net-worth individual risk intelligent assessment system based on multi-dimensional dynamic interaction. Through multi-dimensional data fusion, dynamic modeling mechanism and intelligent decision-making closed loop, it solves the problems of staticity, low dimensionality, low efficiency and low accuracy, and provides high-net-worth individuals with a more intelligent, accurate and dynamic risk management tool. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only 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 are within the scope of protection of the present invention.

[0021] Example 1: See Figure 1 This invention discloses a high-net-worth individual risk intelligent assessment system based on multi-dimensional dynamic interaction, comprising: The data acquisition module collects and quantifies multi-dimensional risk factors; The weight allocation module is connected to the data acquisition module and performs dynamic weight allocation based on multi-dimensional risk factors. The data processing module is connected to the weight allocation module to construct and train a hybrid reinforcement learning model, wherein the result of the dynamic weight allocation is used as the input of the hybrid reinforcement learning model and the risk score result is output. The feedback module is connected to the data processing module and matches intervention strategies based on the risk score results; The configuration adjustment module is connected to the weight allocation module to obtain the intervention effect value of the current intervention strategy, and the weight allocation of the weight allocation module is adjusted based on the intervention effect value.

[0022] In one specific embodiment, the multidimensional risk factors include: basic factors, physiological factors, biochemical factors, behavioral factors, psychological factors, imaging factors, and dynamic monitoring factors.

[0023] Specifically, basic factors (age, gender, etc.), physiological factors (heart rate, blood pressure, etc.), biochemical factors (blood glucose, blood lipids, etc.), behavioral factors (exercise frequency, smoking history, etc.), psychological factors (anxiety scale scores, etc.), imaging factors (CT image features, etc.), and dynamic monitoring factors (real-time data from wearable devices) are collected and standardized into numerical vectors.

[0024] a. Basic factors: age, sex, family history of cardiovascular and cerebrovascular diseases in immediate relatives; b. Physiological factors: resting heart rate, systolic blood pressure / diastolic blood pressure, BMI index; c. Biochemical factors: lipid profile (TC, TG, LDL-C, HDL-C), fasting blood glucose, and serum homocysteine ​​levels; d. Behavioral factors: Smoking history: daily number of cigarettes smoked (0, 1-10, >10), years of smoking (<5 years, 5-20 years, >20 years); Drinking history: daily alcohol intake (0g, <25g, 25-60g, >60g); e. Psychological factors: The Hamilton Anxiety Scale (HAMA) was used to assess anxiety status (total score <7 is normal, 7-14 is mild, and >14 is moderate to severe). f. Dynamic monitoring factors: 12-lead electrocardiogram to obtain QTc interval, ST segment deviation amplitude, and frequency of atrial fibrillation / premature ventricular contractions; g. Imaging factors: Color Doppler ultrasound of the carotid and vertebral arteries to detect intima-media thickness (IMT), plaque location, and degree of stenosis.

[0025] In one specific embodiment, the weight allocation module performs dynamic weight allocation based on multi-dimensional risk factors, including calculating weights and risk interaction effects; Specifically, dynamic weight calculation: the weights of each factor are updated based on time decay and event triggering mechanisms to reflect the timeliness of factors (such as weight mutations triggered by acute physiological changes). Interaction effect calculation: Identify key risk combinations by quantifying the synergistic effect between factors through the correlation coefficient and variance product.

[0026] The expression for calculating the weights is as follows: ; In the formula, with ( t ) is the first i Each risk factor in t Dynamic weights at time points w i0 For the first i The initial weights of each risk factor; l This is the weight decay coefficient; t For time variables; Δ w i When the event is triggered i The weight adjustment amount for each risk factor; I (Event Trigger) is the event trigger indicator function, which is 1 when an event is triggered and 0 otherwise.

[0027] The expression for calculating dynamic interaction effects is: ; In the formula, Risk_Score is the risk score; i Representing different categories of risk factors; w i ( t ) is the first i Each risk factor in t Dynamic weights at time points; X i For the first i The values ​​of each risk factor; The interaction coefficient between risk factors j and k; X j , X k Let be the values ​​of risk factor j and risk factor k, respectively.

[0028] In one specific embodiment, the structure of the hybrid reinforcement learning model includes: Preprocessing layer: Standardizes the multi-dimensional risk factors to obtain a standardized risk factor vector; Risk factor relationship graph construction layer: Input standardized risk factor vectors, output risk factor graph; The transitive layer takes the risk factor graph as input and constructs a high-order representation of each risk factor. Classification decision layer: Risk classification is performed using the LightGBM classifier.

[0029] Specifically, the preprocessing layer standardizes risk factors to eliminate the influence of dimensions; Risk factor relationship graph construction layer: Construct a factor graph based on correlation, where nodes are factors and edges represent correlation strength; Transmission layer: Utilizes graph neural networks (GNNs) to extract higher-order correlation features of factors; Classification decision layer: Outputs risk scores (levels 1-5, with higher values ​​indicating higher risk) through the LightGBM classifier. Specifically, the preprocessing layer: standardizes the multi-dimensional risk factors to obtain a standardized risk factor vector, including: Each risk factor is standardized using the following formula: ; In the formula, and Let be the mean and standard deviation of the i-th risk factor; The standardized risk factor vector is: ; In the formula, m is the total number of risk factor vectors.

[0030] Specifically, the risk factor relationship graph construction layer: inputting standardized risk factor vectors, constructing a risk factor graph, including: Each risk factor is considered as a node in a graph, and the node's characteristics are the standardized risk factor values. Construct edges based on the correlation between risk factors, and let the correlation coefficient between risk factors i and j be: ; when At that time, among them, To set a preset relevance threshold, an edge is created between nodes i and j, with the edge weight being... Obtain and output the risk factor map. , where V is the node set and E is the edge set.

[0031] Specifically, the transitive layer takes the risk factor graph as input and constructs a high-order representation of each risk factor, including: For each node The set of neighboring nodes is ; In the Layers, nodes The update formula is: ; In the formula, Is node v at the th l Layer representation, It is a learnable weight matrix. Is node u in the th order? l The representation of layer -1 For the set of neighboring nodes; It is an activation function. It is the weight of the edge between nodes v and u. Is node u in the th order? l Representation of layer -1; By stacking multiple GNN layers, information from neighboring nodes is gradually aggregated to obtain a high-order representation of each risk factor: ; In the formula, L is the number of layers in the GNN.

[0032] Specifically, in the graph neural network layers, information from neighboring nodes is gradually aggregated through multi-layer stacking, utilizing learnable weight matrices, activation functions, vector concatenation operations, and edge weights to ultimately obtain a high-order representation of each risk factor. These high-order representations can better capture the complex relationships between risk factors, providing richer feature information for subsequent risk assessment and intervention strategy matching.

[0033] Specifically, the classification decision layer uses the LightGBM classifier for risk grading, including: Let the adjusted risk factor weight vector be... The weighted risk factor is then expressed as: ,in, This indicates element-wise multiplication; The LightGBM classifier is used to classify the weighted risk factor representation and output the risk level.

[0034] Specifically, the loss function of the hybrid reinforcement learning model is: ; In the formula, L is the cross-entropy loss of the cross-entropy classification task; l Θ represents the L2 regularization coefficient; Θ represents the set of trainable parameters of the model. m The KL divergence penalty coefficient; P 模型 The probability of the risk distribution predicted by the model; P 先验 Let be the prior probability of the risk distribution.

[0035] Specifically, it also includes a reinforcement learning policy adjustment layer, the input of which is the higher-order representation of the risk factor H. GNN The specific processing procedure is as follows: The higher-order representation of risk factors H GNN As the current state s; Define a set of actions In this embodiment, the weights of different risk factors in the assessment are adjusted and different assessment model parameters are selected. Use a neural network To represent the strategy, where, These are the parameters of the policy network, which outputs the probability distribution of each action based on the current state. Define reward function In this embodiment, the reward is based on the accuracy of the risk assessment (the difference between the assessment result and the actual risk level). The closer the assessment result is to the actual risk level, the higher the reward. The REINFORCE policy gradient algorithm is used to update the policy network parameters. The policy gradient expression is: ; Specifically, pre-defined strategies (such as health management and psychological counseling) are matched based on risk scores. The weight allocation is dynamically adjusted based on the feedback of the intervention effect (such as the improvement rate of physiological indicators) to form a closed-loop optimization.

[0036] In one specific embodiment, the configuration adjustment module is connected to the weight allocation module to obtain the intervention effect value of the current intervention strategy, and adjusts the weight allocation of the weight allocation module based on the intervention effect value. The intervention strategy matching process is as follows: Define intervention strategy rules: Low risk (R=1): Regular risk monitoring is conducted periodically, without the need for special intervention.

[0037] Low to medium risk (R=2): Provide risk warnings (e.g., maintain + regular monitoring).

[0038] Medium risk (R=3): Develop a detailed risk management plan (such as a low-salt diet + 3 aerobic exercises per week).

[0039] Medium to high risk (R=4): Take emergency measures (such as immediately arranging an electrocardiogram and a cardiology consultation).

[0040] High risk (R=5): Conduct a comprehensive review of the situation and develop a comprehensive risk mitigation plan.

[0041] Matching intervention strategies refers to directly selecting the corresponding intervention strategy based on the risk score result R.

[0042] Specifically, starting with the input of multi-dimensional risk factor data, the process involves data preprocessing, construction of risk factor relationship graphs, graph neural network processing, reinforcement learning strategy adjustment, risk assessment and classification, and finally, intervention strategy matching. Higher-order representations of risk factors are calculated and output at the graph neural network layer, subsequently participating in the reinforcement learning strategy adjustment and risk assessment and classification processes to jointly complete the task of intelligent risk assessment and intervention strategy matching for high-net-worth individuals.

[0043] Specifically, this invention fundamentally improves the accuracy and interpretability of risk assessment through multiple dimensions, particularly by introducing quantum computing and causal reasoning. The quantum computing acceleration layer utilizes the properties of quantum mechanics to process high-dimensional data; the causal structure discovery layer constructs causal relationships between risk factors; the federated learning layer addresses data privacy issues; the dynamic knowledge graph layer realizes structured representation and reasoning of knowledge; and the meta-learning optimization layer enhances the model's adaptability. These features enable this invention to more accurately assess the risk profile of high-net-worth individuals.

[0044] Example 2: This embodiment discloses the basic risk assessment process of a high-net-worth individual risk intelligent assessment system based on multi-dimensional dynamic interaction provided by the present invention. The application scenario is the collection of annual health checkup data for a high-net-worth client, specifically including: The data acquisition module collects user A's baseline factors (age 50, male), physiological factors (blood pressure 140 / 90 mmHg, heart rate 85 beats / min), and biochemical factors (fasting blood glucose 6.5 mmol / L). The weight allocation module initially sets the weights to a uniform distribution. When no event is triggered, the weights are allocated according to... l =0.01 Exponential decay The data processing module, the factor graph construction layer, uses correlation thresholds. i =0.5 Connecting the blood pressure and blood glucose nodes (correlation coefficient 0.6), after extracting second-order features in the transfer layer, LightGBM outputs a risk score of level 3 (medium risk). Feedback module, matching intervention strategy: recommend a low-salt diet + 3 aerobic exercises per week.

[0045] Example 3: This example is based on Client A from Example 2, who suddenly experienced chest pain one month later; The data acquisition module detected a sudden increase in heart rate to 120 beats per minute (event triggered).

[0046] In the weight allocation module, when triggered (I(event trigger)=1, the physiological factor weight increases by 0.3, and the dynamic weight jumps to 1.2 times the initial value; In the data processing module, the updated factor graph strengthens the correlation between physiological factors, and the score is improved to level 4 (medium-high risk); in the feedback module, the strategy is adjusted: immediately arrange an electrocardiogram examination + cardiology specialist consultation.

[0047] Example 4: This example is based on user B having both hyperlipidemia (a biochemical factor with a correlation coefficient of 0.7 with coronary heart disease) and a history of smoking (a behavioral factor with a correlation coefficient of 0.6).

[0048] The weight allocation module calculates the interaction effect. i , j =0.5×0.7×0.6=0.21, which is significantly higher than the contribution of a single factor; In the data processing module, the graph neural network captures the higher-order interaction between the two, and the score is improved from level 2, which is evaluated separately, to level 3; Feedback module, customized strategy: combined implementation of lipid-lowering drug intervention and smoking cessation intervention.

[0049] Example 5: In this example, after user C implemented the intervention strategy for 3 months, their blood pressure dropped to 130 / 80 mmHg (an improvement of 20%), as detailed below: Adjust the configuration module to reduce the physiological factor weight decay coefficient λ to 0.005 based on the effect value (preset threshold 15%), thereby extending its effectiveness. In the weighting module, the rate of weight decay of the blood pressure factor slows down in the next round of evaluation, and the impact of historical improvement data on the current score is enhanced. In the data processing module, the score was downgraded from level 3 to level 2 (low risk), and the strategy was adjusted to maintain the existing interventions plus regular monitoring.

[0050] Example 6: This embodiment trains the hybrid reinforcement learning model provided by this invention based on a sample data of 100,000 high-net-worth individuals. Specifically, it includes the following: Data preprocessing: Standardize factor vectors, and fill missing values ​​using KNN interpolation; Model structure: The risk factor relationship graph contains 50 nodes (7 types of factor sub-items), the transit layer uses a 2-layer GCN, and the classification layer LightGBM tree depth is set to 6; Training parameters: In the loss function, λ=0.001, μ=0.01, and KL divergence constrains the model to align with the prior risk distribution (e.g., the high-risk proportion of the middle-aged and elderly population is 20%). Evaluation metrics: Test set accuracy 92%, AUC-ROC 0.95, outperforming the traditional logistic regression model (accuracy 85%).

[0051] This invention captures high-order nonlinear relationships between factors through a risk factor relationship graph construction layer, correlation threshold filtering, and a graph neural network transmission layer. For example, it can identify a three-level transmission path of "anxiety (psychological factor) → decreased sleep quality (behavioral factor) → blood pressure fluctuation (physiological factor)," a complex causal chain that traditional linear models cannot resolve. By employing the LightGBM algorithm in the classification decision layer, combined with weighted risk factor representation, computational complexity is significantly reduced while maintaining accuracy. In actual testing, single-sample evaluation takes only 80ms, meeting the requirements for real-time risk warning.

[0052] This invention collects and quantifies multi-dimensional risk factors, employs a dynamic weight allocation method based on entropy weighting, and constructs a hybrid reinforcement learning model comprising a preprocessing layer, a temporal modeling layer, an attention mechanism layer, and a classification decision layer. Finally, it matches corresponding intervention strategies based on the risk score results. The collaborative work among these layers enables a more accurate assessment of the risk of high-net-worth individuals.

[0053] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0054] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A high-net-worth individual risk intelligent assessment system based on multi-dimensional dynamic interaction, characterized in that, include: The data acquisition module collects and quantifies multi-dimensional risk factors; The weight allocation module is connected to the data acquisition module and performs dynamic weight allocation based on multi-dimensional risk factors. The data processing module is connected to the weight allocation module to construct and train a hybrid reinforcement learning model, wherein the result of the dynamic weight allocation is used as the input of the hybrid reinforcement learning model and the risk score result is output. The feedback module is connected to the data processing module and matches intervention strategies based on the risk score results.

2. The intelligent risk assessment system for high-net-worth individuals based on multi-dimensional dynamic interaction according to claim 1, characterized in that, It also includes: an adjustment configuration module, which is connected to the weight allocation module to obtain the intervention effect value of the current intervention strategy and adjust the weight allocation of the weight allocation module based on the intervention effect value.

3. The intelligent risk assessment system for high-net-worth individuals based on multi-dimensional dynamic interaction according to claim 1, characterized in that, The multidimensional risk factors include: basic factors, physiological factors, biochemical factors, behavioral factors, psychological factors, imaging factors, and dynamic monitoring factors.

4. The intelligent risk assessment system for high-net-worth individuals based on multi-dimensional dynamic interaction according to claim 1, characterized in that, The weight allocation module performs dynamic weight allocation based on multi-dimensional risk factors, including calculating weights and risk interaction effects. The expression for calculating the weights is as follows: ； In the formula, wi ( t ) is the first i Each risk factor in t Dynamic weights at time points w i0 For the first i The initial weights of each risk factor; λ This is the weight decay coefficient; t For time variables; Δ w i When the event is triggered i The weight adjustment amount for each risk factor; I (Event Trigger) is an event trigger indicator function; it is 1 when an event is triggered, and 0 otherwise. The expression for calculating dynamic interaction effects is: ； In the formula, Risk_Score is the risk score; i Representing different categories of risk factors; w i ( t ) is the first i Each risk factor in t Dynamic weights at time points; X i For the first i The values ​​of each risk factor; The interaction coefficient between risk factors j and k; X j , X k Let be the values ​​of risk factor j and risk factor k, respectively.

5. The intelligent risk assessment system for high-net-worth individuals based on multi-dimensional dynamic interaction according to claim 1, characterized in that, The structure of the hybrid reinforcement learning model includes: Preprocessing layer: Standardizes the multi-dimensional risk factors to obtain a standardized risk factor vector; Risk factor relationship graph construction layer: Input standardized risk factor vectors, output risk factor graph; The transitive layer takes the risk factor graph as input and constructs a high-order representation of each risk factor. Classification decision layer: Risk classification is performed using the LightGBM classifier.

6. The intelligent risk assessment system for high-net-worth individuals based on multi-dimensional dynamic interaction according to claim 5, characterized in that, The preprocessing layer: standardizes the multi-dimensional risk factors to obtain a standardized risk factor vector, including: Each risk factor is standardized using the following formula: ； In the formula, and Let be the mean and standard deviation of the i-th risk factor; The standardized risk factor vector is: ； In the formula, m is the total number of risk factor vectors.

7. The intelligent risk assessment system for high-net-worth individuals based on multi-dimensional dynamic interaction according to claim 5, characterized in that, The risk factor relationship graph construction layer: Inputs standardized risk factor vectors and constructs a risk factor graph, including: Each risk factor is considered as a node in a graph, and the node's characteristics are the standardized risk factor values. Construct edges based on the correlation between risk factors, and let the correlation coefficient between risk factors i and j be: ; when At that time, among them, To set a preset relevance threshold, an edge is created between nodes i and j, with the edge weight being... Obtain and output the risk factor map. , where V is the node set and E is the edge set.

8. The intelligent risk assessment system for high-net-worth individuals based on multi-dimensional dynamic interaction according to claim 5, characterized in that, The transmission layer takes the risk factor graph as input and constructs a high-order representation of each risk factor, including: For each node The set of neighboring nodes is ; In the Layers, nodes The update formula is: ； In the formula, Is node v at the th l Layer representation, It is a learnable weight matrix. Is node u in the th order? l The representation of layer -1 For the set of neighboring nodes; It is an activation function. It is the weight of the edge between nodes v and u. Is node u in the th order? l Representation of layer -1; By stacking multiple GNN layers, information from neighboring nodes is gradually aggregated to obtain a high-order representation of each risk factor: ； In the formula, L is the number of layers in the GNN.

9. A high-net-worth individual risk intelligent assessment system based on multi-dimensional dynamic interaction as described in claim 5, characterized in that, The classification decision layer uses the LightGBM classifier for risk grading, including: Let the adjusted risk factor weight vector be... The weighted risk factor is then expressed as: ,in, This indicates element-wise multiplication; The LightGBM classifier is used to classify the weighted risk factor representation and output the risk level.

10. A high-net-worth individual risk intelligent assessment system based on multi-dimensional dynamic interaction according to claim 5, characterized in that, The loss function of the hybrid reinforcement learning model is: ； In the formula, L is the cross-entropy loss of the cross-entropy classification task; λ Θ represents the L2 regularization coefficient; Θ represents the set of trainable parameters of the model. μ The KL divergence penalty coefficient; P 模型 The probability of the risk distribution predicted by the model; P 先验 Let be the prior probability of the risk distribution.

Images