Individual risk dynamic assessment method and system based on spatiotemporal causal diagram network

Abstract

This invention discloses a method and system for dynamic individual risk assessment based on spatiotemporal causal graph networks, relating to the field of data processing technology. The method includes the following steps: acquiring risk-related data of the individual to be assessed; converting the aforementioned data into structured data to obtain risk-related structured data of the individual to be assessed, wherein the risk-related structured data is divided into static features, dynamic features, and causal features; calculating the static feature risk score and dynamic feature risk score of the individual to be assessed based on the risk-related structured data to determine the final risk score; and acquiring risk classification information, and predicting the dynamic risk trend information with the highest confidence based on multiple classified dynamic risk trends, combined with dynamic spatiotemporal fusion prediction results and a causal feature matrix. This invention effectively improves the reliability, accuracy, and timeliness of dynamic individual risk assessment, enabling more efficient and dynamic prediction and assessment of the individual's risk status.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for dynamic assessment of individual risk based on spatiotemporal causal graph networks. Background Technology

[0002] Risk assessment refers to the process of predicting the impact or losses of a risk event on society, the economy, and public order before it occurs, and quantifying the risk status of the assessed object. Risk assessment is commonly used in various fields such as finance, public safety, transportation, healthcare, and power systems. The assessed object can be any entity within the corresponding field, such as equipment, facilities, management systems, processes, working environments, individuals, or external factors. Through risk assessment, the risk information of the assessed object can be quantified, facilitating the development of response strategies in advance. Furthermore, it can assist relevant institutions and researchers in making more accurate risk management (providing a scientific basis for risk management) and in making proactive decisions to prevent the immeasurable impact of sudden events.

[0003] Taking individual risk as an example, there are currently solutions for automatically assessing user risk levels in the financial, public safety, and healthcare sectors. For instance, risk assessment models can be built to predict and assess user bank credit risk, pedestrian traffic safety risk, and the risk of a patient's condition worsening. When building these risk assessment models, methods such as strong logical rules, matrix calculations, and machine learning are typically used. Taking the assessment of user bank credit risk as an example, the assessment process is usually as follows: After obtaining information on the user's historical financial situation and whether they have a negative credit history through a questionnaire survey, the aforementioned information is analyzed and processed by the constructed credit risk assessment model to predict related risks. Risk level assessments in other scenarios are largely similar.

[0004] The commonly used risk assessment schemes mentioned above have the following shortcomings: 1) When acquiring individual information, they mainly focus on historical or static characteristics. While this can prevent the recurrence of historical risks to some extent, it cannot fully reflect the individual's current state and is not conducive to comprehensive individual assessment and real-time updates. 2) When assessing individual risk, traditional statistical and survival analysis methods have certain limitations and are difficult to model high-dimensional features and nonlinear relationship data. While machine learning and matrix calculation methods can handle high-dimensional data, they still cannot learn and train on sequential, spatiotemporally correlated, and multi-intersecting data.

[0005] On the other hand, with the rapid development of computers, trajectory prediction algorithms have gradually transitioned from machine learning to deep learning. Currently, prediction models based on deep learning methods mainly include Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTMs), and Graph Convolutional Networks (GCNs). The appropriate method can be selected for trajectory prediction based on the specific application scenario and data characteristics. Among them, Graph Convolutional Networks (GCNs) are a very popular GNN architecture, using convolutional layers to create graph embeddings. Interconnected graph data is ubiquitous in all fields around us, from molecular structures to social networks to urban design structures. As a powerful method, Graph Neural Networks (GNNs) are being used to model and learn the spatial and graph structures of this type of data, and they have been applied in drug discovery, simulation systems, and social networks. GNNs can also combine ideas from other machine learning models, such as combining GNNs with sequence models to form Spatial-Temporal Graph Neural Networks (STGNNs). Spatiotemporal Graph Neural Networks (STGNNs) can capture the temporal and spatial dependencies of data. Each time step in an STGNN is a graph, which can be passed through a Generative Neural Network (GNN) to obtain an encoded graph that embeds the spatial dependencies of the data. These encoded graphs can then be modeled like time-series data. Currently, STGNNs have been used in traffic flow prediction and vehicle trajectory prediction. For example, Chinese patent application CN202410614595.3 discloses a vehicle trajectory prediction method based on STGNN-Net. By constructing a vehicle trajectory prediction network based on STGNN-Net, it fully utilizes the temporal features in the trajectory data and combines them with the spatial features at different times to achieve more accurate and robust trajectory prediction. Through graph neural networks and attention mechanisms, the network can consider interactive information, thereby optimizing features and achieving higher trajectory prediction accuracy.

[0006] In summary, given the advantages of the Spatiotemporal Graph Neural Network (STGNN) model, how to combine it with the assessment needs of individual risk status to provide a solution that can efficiently and dynamically predict and assess individual risk status is a technical problem that urgently needs to be solved. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a method and system for dynamic individual risk assessment based on spatiotemporal causal graph networks. The method for dynamic individual risk assessment based on spatiotemporal causal graph networks provided by this invention considers the dynamic changes and diverse characteristics of individuals, simultaneously taking into account their dynamic and static features and the causal relationships between these features. Furthermore, by combining spatiotemporal fusion and causal reasoning, it predicts the potential risk value and future risk trends of the individual to be assessed, effectively improving the reliability, accuracy, and timeliness of dynamic individual risk assessment. This allows for more efficient and dynamic prediction and assessment of the individual's risk status, achieving the goals of risk prevention and management.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] A dynamic individual risk assessment method based on spatiotemporal causal graph networks includes the following steps:

[0010] Obtain risk-related data for the individuals to be assessed, and classify the risk-related data into unstructured data and semi-structured data according to the complexity of the data;

[0011] The aforementioned unstructured and semi-structured data are transformed into structured data to obtain the risk-related structured data of the individuals to be assessed. The risk-related structured data is divided into three dimensions: static features, dynamic features, and causal features. The static features are features that do not interfere with each other. The dynamic features are features that are related to time and space, including time and space-related dimensions. The causal features are used to represent the feature orientation relationship and are represented by a causal feature matrix.

[0012] Based on the data content of the three dimensions in the aforementioned risk-related structured data, the static characteristic risk score of the aforementioned individual to be assessed is calculated respectively. Xgb and dynamic feature risk score Stgnn The final risk score is determined based on the static and dynamic feature risk scores. final Furthermore, it obtains risk classification information, and based on the multiple dynamic risk trends classified, combines the dynamic spatiotemporal fusion prediction results and the aforementioned causal feature matrix to predict the dynamic risk trend information with the highest confidence.

[0013] Furthermore, the aforementioned final risk score and / or dynamic risk trend information will be output on the data visualization interface; or,

[0014] The aforementioned final risk score and / or dynamic risk trend information are sent to the associated user terminal so that the associated personnel can obtain the risk information of the aforementioned individual to be assessed, and the risk response measures information fed back by the aforementioned associated personnel's terminal is output.

[0015] Furthermore, the static feature risk scores and dynamic feature risk scores are weighted and calculated to obtain the final risk score. The calculation formula is as follows:

[0016] Score final =αScore Xgb +βScore Stgnn ,

[0017] Wherein, α represents the risk weight of static features, and β represents the risk weight of dynamic features. α and β are set by system default, user-defined, or received from the weight settings of the associated expert experience module.

[0018] Furthermore, the risk-related data of the individual to be evaluated is encrypted data, which is mapped into ciphertext through a data encryption algorithm; after obtaining the aforementioned encrypted data of the individual to be evaluated, the encrypted data is first decrypted, and then the decrypted data is divided into unstructured data and semi-structured data.

[0019] Furthermore, unstructured and semi-structured data are transformed into structured data through a hybrid logic model, which is configured with keyword matching rules. Temporal and spatial location regular expression matching rules and rules for extracting logical thinking chains

[0020] The hybrid logic model is configured as follows:

[0021] Obtain unstructured data of the individuals to be evaluated: UD = {ud1, ud2, ..., ud} n} and semi-structured data SSD={ssd1,ssd2,...,ssd n In fact, n represents the dimensions of the individual risk-related data. The risk-related data of the individual to be evaluated consists of n different dimensions, where n is an integer greater than or equal to 1. According to the aforementioned configured rules, the unstructured data UD and semi-structured data SSD are converted into structured data. The structured data SD contains three dimensions: static feature data X, dynamic feature data Y, and causal feature matrix Z. X is configured with a feature dimensions, Y is configured with b feature dimensions, and Z is configured with c feature dimensions. a, b, and c are all integers greater than or equal to 1, and a + b + c ≥ n, c ≤ n.

[0022] At this time, the static feature data is represented as:

[0023] X = {x1, x2, ..., x} a},

[0024] X aThis represents the static feature data of the a-th dimension;

[0025] The dynamic feature data is represented as follows:

[0026]

[0027] Among them, y t ={(dx,dy)} t ,(x a+1 ,x a+2 ,...,x a+b ) t Let} represent the state of the aforementioned individual to be evaluated at observation time t, where (dx, dy) t This represents the latitude and longitude coordinates of the individual to be evaluated at observation time t, (x a+1 ,x a+2 ,...,x a+b ) t Let represent the value vector of b dynamic feature dimensions of the individual to be evaluated at observation time t, where T represents the entire observation time series, and t represents the final observation time. This represents each observation time unit, where tT represents the initial observation time. This indicates the time of the next prediction after the observation time t. Indicates the time of the next prediction from the observation time tT;

[0028] The causal feature matrix Z is generated based on the strength of the causal relationship between features. The strength level of the causal relationship between multiple features can be determined through the causal feature matrix Z.

[0029] Furthermore, a causal feature matrix generation module is provided, which is configured to perform the following steps:

[0030] The causal strength levels between features are initialized, and different causal strength levels map to different causal feature values, which can reflect the strength level of causal association between features.

[0031] Collect information on the directional relationships between user-configured features and the causal strength levels between features;

[0032] Based on the causal strength level information between features, and the mapping relationship between the aforementioned causal strength level and causal feature values, the causal feature values ​​between multiple features are obtained. The causal feature values ​​between multiple features are then converted into the correlation matrix of the graph, which is the causal feature matrix Z.

[0033] Furthermore, a sample size of m individuals to be evaluated was obtained, and a static characteristic risk score was calculated for each individual. Xgb The steps are as follows:

[0034] S310, perform extremum-removal adjustment on the continuous variables in the static feature X, as follows: Calculate the mean μ and standard deviation σ of each feature. Based on the preset reasonable range of the features [μ-m*σ, μ+m*σ], adjust the static feature X to obtain the extremum-removed static feature X'. For any feature X in the static feature X... i For i = 1, 2, ..., a, the formula for calculating the de-extreme adjustment is as follows:

[0035]

[0036] S320, input the de-extremeized static feature X' into the XGBoost model; where the maximum tree depth of the XGBoost model is set according to the sensitivity of individual risk to feature dimensions;

[0037] S330, by training the XGBoost model, calculate and output the confidence level of whether an individual is at risk, and use this confidence level as the static feature risk score. Xgb Score Xgb It is an m-dimensional vector containing the static feature risk score of each of the m samples, where each score satisfies the Score Xgb ∈[0-100].

[0038] Furthermore, when calculating the dynamic characteristic risk score of the individual to be evaluated, a dynamic spatiotemporal fusion prediction model is used to predict the spatial location coordinates of the individual to be evaluated at future time points and the characteristic changes of b dynamic characteristic dimensions.

[0039] The dynamic spatiotemporal fusion prediction model employs a dynamic spatial-dual-track temporal prediction model, which consists of a graph convolutional neural network and a temporal prediction network. The future time point is the time following the final observation time t. From the start to the future time T, t+T, the formula is expressed as:

[0040]

[0041] In the formula, This represents the characteristic data of the aforementioned individual to be evaluated at the aforementioned future time points. Represents spatial features G and historical T time series data. Related functions;

[0042] The final output of the dynamic spatial-dual-track temporal prediction model is the value of the individual to be evaluated at time [time value missing]. The predicted state is expressed by the following formula:

[0043]

[0044] At this point, the steps for calculating the dynamic characteristic risk score of the individual to be assessed are as follows:

[0045] Obtain the output results of the dynamic spatial domain-dual parallel time domain prediction model

[0046] Based on the aforementioned output results, the dynamic characteristic risk score is calculated using the risk integration assessment module. Stgnn The risk integration assessment module is configured to perform the following operations:

[0047]

[0048] Part One express The average increase or decrease in risk characteristics at any given time; an increase in value indicates a higher level of risk, and a decrease in value indicates a lower level of risk; the middle part express The rate of change of the trajectory at any given time; a larger value indicates a faster rate of change and a higher risk. (Part Three) Indicates the nearest risk location to the individual. The greater the distance, the lower the score; the smaller the distance, the higher the score. v represents the highest risk value, which is greater than 0.

[0049] Furthermore, the dynamic spatial-dual-track temporal prediction model includes a graph convolutional neural network (GCN) model, a long short-term memory (LSTM) network model, and an autoregressive moving average (Arima) model. The prediction steps are as follows:

[0050] Obtain historical T time series data Will The data is input into the GCN model, which is used to detect individual trajectory information to obtain the trajectory topology and thus the spatial characteristics of the individual.

[0051] By inputting time series with spatial characteristics into LSTM and Arima models, the dynamic changes of the time series can be obtained through information transfer between units, thus capturing the time features.

[0052] The values ​​output by the LSTM model and the Arima model are concatenated to obtain the next time step. spatiotemporal characteristics

[0053] The GCN model uses a two-layer structure to detect human trajectory information. Let A represent the adjacency matrix of the observation points, and then use the formula... The transformation yields a self-connected structure, and the preprocessing is represented as follows: In the formula, The degree matrix representing the trajectory of all individuals at the observation point. Obtain the weight matrices for the first and second layers, W0 and W1 respectively. Use Ω() and ReLU() as activation functions. The formula for the two-layer GCN model is as follows:

[0054]

[0055] Furthermore, the gating mechanism of the LSTM model is used to acquire and record temporal correlations, denoted as h. t-1 Let x be the hidden layer state at time t-1. a+1 ,x a+2 ,...,x a+b ) t Let f(A,Y) be the feature content at time t. t The process of GCN graph convolution is shown below, where W and k are the weights and biases during training, respectively. The training and update process of the LSTM model is as follows:

[0056] u t =Ω(W u [f(A,Y t ),h t-1 ]+k u ),

[0057] r t =Ω(W r [f(A,Y t ),h t-1 ]+k r ),

[0058] s t =tanh(W c [f(A,Y t ),(r t *h t-1 )]+k c ),

[0059] h t =u t *h t-1 +(1-u t )*s t ,

[0060] Where, r t This represents the reset gate state at time t, used to control the degree to which information from the previous state is ignored. The corresponding weight and bias are w. r and k r ;u tLet w represent the update gate at time t, used to control the degree to which the state information from the previous time step is incorporated into the current state. The corresponding weight and bias are w. u and k u ;s t The information stored at time t is represented by w, with corresponding weights and biases. c and k c h t Indicates the output state at time t;

[0061] And, predicting future moments using the Arima model. The risk profile is represented by Arima(p,d,q), where p represents the autoregressive order of the AR term, q represents the moving average order of the MR term, and d represents the difference order for the sequence to become stable. The feature prediction results output by the LSTM are denoted as... Let the feature prediction result output by Arima be denoted as final The risk assessment score output at any time is in, This indicates the weight of the prediction result, which is set by the system or the user.

[0062] Furthermore, during the training of the dynamic spatial domain-dual-track temporal domain prediction model, the loss function is configured as follows:

[0063] loss = ||Y t -Y t ′||+λL reg ;

[0064] Among them, Y t and Y t ′ represent the true value and predicted value of the feature label at time t, respectively; L reg λ is the L2 regularization term used to avoid overfitting, and it is a preset value; λ is a preset hyperparameter.

[0065] Furthermore, the task prediction module predicts the dynamic risk trend information with the highest confidence level. This task prediction module is configured to: acquire multiple dynamic risk trend information segments; and, for each dynamic risk trend, generate the results output by the dynamic spatial-dual-track temporal prediction model. Combined with the causal feature matrix Z, the dynamic risk trend with the highest confidence is predicted through a configured fully connected neural network;

[0066] For the causal feature matrix Z, we obtain c feature dimensions of causal association, and sum Z by rows to obtain a vector with c rows and 1 column, i.e., zc. c*c =sum(z) c*c c)=z c*1In the formula, sum(·) represents the summation function; then, the feature dimensions without causal relationships are completed, set to 0, and transformed into z. n*1 ; matrix z n*1 Compared with the prediction results Concatenate to obtain a vector The vector dimension is (n+2+b)*1; the aforementioned vector is used to predict the optimal dynamic risk trend category and its confidence level through a configured multi-layer fully connected neural network.

[0067] Another embodiment of the present invention also provides an individual risk dynamic assessment system, the system comprising:

[0068] The data acquisition module is used to acquire risk-related data for the individuals to be assessed.

[0069] The data processing module is used to divide the risk-related data into unstructured data and semi-structured data according to the complexity of the data; and to transform the aforementioned unstructured data and semi-structured data into structured data to obtain the risk-related structured data of the aforementioned individual to be evaluated. The risk-related structured data is divided into three dimensions according to static features, dynamic features, and causal features. The static features are features that do not interfere with each other; the dynamic features are features related to time and space, including time and space related dimensions; the causal features are used to represent the feature orientation relationship, represented by a causal feature matrix.

[0070] The risk dynamic assessment module is used to calculate the static characteristic risk score of the individual to be assessed based on the three dimensions of data content in the aforementioned risk-related structured data. Xgb and dynamic feature risk score Stgnn The final risk score is determined based on the static and dynamic feature risk scores. final Furthermore, it obtains risk classification information, and based on the multiple dynamic risk trends classified, combines the dynamic spatiotemporal fusion prediction results and the aforementioned causal feature matrix to predict the dynamic risk trend information with the highest confidence.

[0071] Compared with existing technologies, this invention, by adopting the above technical solutions, has the following advantages and positive effects: The individual risk dynamic assessment method based on spatiotemporal causal graph networks considers the dynamic and static characteristics of individuals and the causal relationships between these characteristics from the perspective of individual dynamic changes and diverse features. Furthermore, by combining spatiotemporal fusion and causal reasoning, it predicts the potential risk value and future risk trends of the individual to be assessed, effectively improving the reliability, accuracy, and timeliness of individual risk dynamic assessment. This allows for more efficient and dynamic prediction and assessment of individual risk status, achieving the goal of risk prevention and management. Attached Figure Description

[0072] Figure 1 A flowchart of an individual risk dynamic assessment method based on spatiotemporal causal graph networks provided in an embodiment of the present invention.

[0073] Figure 2 This is a schematic diagram illustrating the information processing for risk assessment provided in an embodiment of the present invention.

[0074] Figure 3 This is a schematic diagram illustrating the feature orientation relationship of causal features provided in an embodiment of the present invention.

[0075] Figure 4 This is a schematic diagram of the logical structure of the improved dynamic spatial-dual parallel time-domain prediction model provided in an embodiment of the present invention. Detailed Implementation

[0076] The present invention discloses a method and system for dynamic individual risk assessment based on spatiotemporal causal graph networks, which will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that the technical features or combinations of technical features described in the following embodiments should not be considered isolated; they can be combined to achieve better technical effects. In the accompanying drawings of the following embodiments, the same reference numerals in each drawing represent the same features or components, which can be applied to different embodiments. Therefore, once an item is defined in one drawing, it does not need to be further discussed in subsequent drawings.

[0077] It should be noted that the structures, proportions, sizes, etc., illustrated in the accompanying drawings are merely for illustrative purposes and to aid those skilled in the art in understanding and reading the invention. They are not intended to limit the conditions under which the invention can be implemented. Any modifications to the structure, changes in proportions, or adjustments to size, provided they do not affect the effectiveness or purpose of the invention, should fall within the scope of the technical content disclosed in the invention. The scope of the preferred embodiments of the present invention includes other implementations, wherein functions may be performed not in the order stated or discussed, including substantially simultaneously or in reverse order, depending on the functions involved. This should be understood by those skilled in the art to which the embodiments of the present invention pertain.

[0078] Techniques, methods, and apparatus known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and apparatus should be considered part of the specification. In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0079] In the description of the embodiments of this application, " / " means "or", and "and / or" is used to describe the relationship between related objects, indicating that there can be three relationships. For example, "A and / or B" means: A and B exist alone, B exists alone, and A and B exist simultaneously. In the description of the embodiments of this application, "multiple" refers to two or more.

[0080] Example

[0081] See Figure 1 As shown, this invention provides a method for dynamic assessment of individual risk based on spatiotemporal causal graph networks.

[0082] The method includes the following steps:

[0083] S100, Obtain risk-related data for the individuals to be assessed, and classify the risk-related data into unstructured data and semi-structured data based on the complexity of the data. Structured data refers to data that follows predefined data formats and length specifications. It has a clear meaning, a strict and consistent order, and a clear data type. It is stored and managed in highly organized tables or databases. The most common type of structured data is data in relational databases, primarily in the form of two-dimensional tables. In contrast to structured data is unstructured data, which refers to unfiltered information without fixed organizational principles. It does not have a predefined data model. Unstructured data takes many forms, typically including data in formats such as images, videos, audio files, and text information. This type of data is difficult to store and manage using traditional relational databases, and its processing complexity is high. Semi-structured data is data that lies between structured and unstructured data, possessing certain structured characteristics but not fully conforming to them (e.g., data that does not fully conform to the format requirements of tabular data models or relational databases). It contains some easily analyzable structured elements, such as tags, making data processing and utilization more convenient. Common examples of semi-structured data include log files, XML documents, JSON documents, and HTML documents. Due to its certain structured characteristics, semi-structured data is less complex than unstructured data.

[0084] S200, the aforementioned unstructured and semi-structured data are transformed into structured data to obtain the risk-related structured data of the individuals to be assessed. This risk-related structured data is divided into three dimensions: static features, dynamic features, and causal features. The static features are non-interfering features; the dynamic features are spatiotemporally related features, including time and space-related dimensions; the causal features are used to represent feature orientation relationships and are represented by a causal feature matrix. The data format and specifications of the structured data can be set by system defaults or customized by the user according to data processing needs; no restrictions are imposed here.

[0085] S300, based on the data content of the three dimensions in the aforementioned risk-related structured data, calculate the static characteristic risk score for the aforementioned individual to be assessed. Xgb and dynamic feature risk score Stgnn The final risk score is determined based on the static and dynamic feature risk scores. final Furthermore, it obtains risk classification information, and based on the multiple dynamic risk trends classified, combines the dynamic spatiotemporal fusion prediction results and the aforementioned causal feature matrix to predict the dynamic risk trend information with the highest confidence.

[0086] In practice, the final risk score can be obtained by weighting the static feature risk score and the dynamic feature risk score. The formula for calculating the final risk score is as follows:

[0087] Score final =αScore Xgb +βScore Stgnn ,

[0088] Where α represents the risk weight of static features, and β represents the risk weight of dynamic features. α and β can be set by system default, user-defined, or receive weight settings from the associated expert experience module. Preferably, see [reference needed]. Figure 2 As shown, the weight settings of the associated expert experience module are received, that is, the risk weights α and β are set by experts in the relevant application domain.

[0089] In one implementation, after step S300, step S410 may be included: outputting the aforementioned final risk score and / or dynamic risk trend information on a data visualization interface. This allows the user to obtain the potential risk value and future risk trend information of the individual to be assessed.

[0090] In another embodiment, after step S300, step S420 may be included: sending the aforementioned final risk score and / or dynamic risk trend information to the associated user terminal so that associated personnel can obtain the risk information of the aforementioned individual to be assessed; and receiving and outputting risk response measures information fed back by the associated personnel's terminal. Furthermore, the associated personnel can take different positive guidance, prevention and early warning, or rescue measures based on the different risk scores of the individual to be assessed.

[0091] The relevant personnel include, but are not limited to, the subject of the assessment, risk assessment researchers, safety professionals, risk managers, and rescue personnel. This allows these personnel to better understand individual risks in different scenarios and take proactive response measures.

[0092] In this embodiment, the risk-related data of the individual to be evaluated in step S100 is preferably encrypted data, which is mapped into ciphertext through a data encryption algorithm. After obtaining the aforementioned encrypted data of the individual to be evaluated, the encrypted data is first decrypted, and then the decrypted data is divided into unstructured data and semi-structured data.

[0093] The risk-related data may include data from one or more dimensions. For example, the preset individual risk-related data may include n different dimensions, where n is an integer greater than or equal to 1. As an example and not a limitation, the n dimensions may be configured as personal basic information, time and space trajectory information, work information, life information, etc. Considering data privacy and security, the individual risk-related data is encrypted and mapped into ciphertext.

[0094] Preferably, considering that some risk-related data may involve privacy information, the following steps may be included before obtaining the risk-related data of the individual to be evaluated: sending an inquiry to the individual to be evaluated regarding whether the use of the risk-related data is permitted, collecting feedback from the individual to be evaluated regarding the aforementioned inquiry, and when the feedback indicates that the data is permitted, reading and processing the aforementioned risk-related data.

[0095] The following is combined with Figure 2 The individual risk dynamic assessment process of this embodiment is described in detail.

[0096] Before model analysis, unstructured and semi-structured data that do not meet the model processing requirements need to be transformed into structured data that meets the model processing requirements.

[0097] In practice, a hybrid logic model is used to transform unstructured and semi-structured data into structured data. This hybrid logic model is configured with at least keyword matching rules. Temporal and spatial location regular expression matching rules and rules for extracting logical thinking chains

[0098] The hybrid logic model is configured to: acquire unstructured data UD = {ud1, ud2, ..., ud} of the individual to be evaluated. n} and semi-structured data SSD={ssd1,ssd2,...,ssd n In fact, n represents the number of dimensions for the individual risk-related data. The risk-related data of the individual to be evaluated consists of n different dimensions, where n is an integer greater than or equal to 1. According to the aforementioned configured rules, the unstructured data UD and semi-structured data SSD are converted into structured data SD.

[0099]

[0100] The structured data SD contains three dimensions: static feature data X, dynamic feature data Y, and a causal feature matrix Z. X has 'a' feature dimensions, Y has 'b' feature dimensions, and Z has 'c' feature dimensions, where a, b, and c are all integers greater than or equal to 1, and a + b + c ≥ n, c ≤ n.

[0101] The static feature data is represented as follows:

[0102] X = {x1, x2, ..., x} a},

[0103] X a Let X1, X2, ..., X be the static feature data of the a-th dimension. a These are characteristic indicators or values ​​that do not interfere with each other.

[0104] The dynamic feature data is represented as follows:

[0105]

[0106] Among them, y t ={(dx,dy)} t ,(x a+1 ,x a+2 ,...,x a+b ) t Let} represent the state of the aforementioned individual to be evaluated at observation time t, where (dx, dy) t This represents the latitude and longitude coordinates of the individual to be evaluated at observation time t, (x a+1 ,x a+2 ,...,x a+b ) t Let represent the value vector of b dynamic feature dimensions of the individual to be evaluated at observation time t, where T represents the entire observation time series, and t represents the final observation time. This represents each observation time unit, where tT represents the initial observation time. This indicates the time of the next prediction after the observation time t. This indicates the time of the next prediction after the observation time tT.

[0107] The causal feature matrix Z is generated based on the strength of the causal relationship between features. The strength level of the causal relationship between multiple features can be determined through the causal feature matrix Z.

[0108] Specifically, a causal feature matrix generation module is included.

[0109] The causal feature matrix generation module is configured to perform the following steps: initialize the causal strength level between features, with different causal strength levels mapping to different causal feature values, whereby the causal feature values ​​reflect the strength level of the causal relationship between features; collect the directional relationship between features and the causal strength level information between features configured by the user; based on the causal strength level information between features and the aforementioned mapping relationship between causal strength levels and causal feature values, obtain the causal feature values ​​between multiple features, and convert the causal feature values ​​between multiple features into an association matrix of the graph, which is the causal feature matrix Z.

[0110] As an example, and not a limitation, the strength of the causal relationship between the features may include strong causal relationship, medium causal relationship, weak causal relationship, and no causal relationship, with corresponding causal feature values ​​of 3, 2, 1, and 0, respectively.

[0111] The following is combined with Figure 3 Describe in detail how the causal feature matrix Z is generated.

[0112] See Figure 3 As shown, there are 6 features X1, X3, X... a-2 X a X b X c There are causal inferences and associations between them, and these causal associations are divided into four categories: strong, medium, weak, and none, with corresponding causal characteristic values ​​of [3, 2, 1, 0]. Specifically, the causal association between X1 and itself is non-causal, the causal association between X1 and X3 is non-causal, and the causal association between X1 and X... a-2 The causal relationship between X1 and X is a weak causal relationship. a The causal relationship between X1 and X is not causal. b The causal relationship is a medium causal relationship, X1 and X c The causal relationship is a strong causal relationship, meaning that feature X1 can be related to itself, X3, and X... a-2 X a X b X c The causal feature values ​​are 0, 0, 1, 0, 2, 3 in sequence; and so on. Based on the causal relationship established between any two features, other features X3 and X can be obtained. a-2 X a X b X c The causal feature values ​​between any two features of the aforementioned six features, along with their own and other features, are converted into a graph correlation matrix, as follows:

[0113]

[0114] The matrix described above is the causal feature matrix Z for the six features mentioned above. The first row represents the causal feature value row for feature X1, the second row represents the causal feature value row for feature X3, and the third row represents the causal feature value row for feature X4. a-2 The causal feature value row, the 4th row represents feature X b The causal feature value row, the 5th row represents feature X b The causal feature value row, the 6th row represents feature X c The causal characteristic value row.

[0115] In this embodiment, let the sample size of the individuals to be evaluated be m. Then, calculate the static characteristic risk score of each individual to be evaluated. Xgb The specific steps may include steps S311-S313.

[0116] S311, perform extremum-removal adjustment on the continuous variables in the static feature X, as follows: Calculate the mean μ and standard deviation σ of each feature. Based on the preset reasonable range of the features [μ-m*σ, μ+m*σ], adjust the static feature X to obtain the extremum-removed static feature X'. For any feature X in the static feature X... i For i = 1, 2, ..., a, the formula for calculating the de-extreme adjustment is as follows:

[0117]

[0118] By performing the above extremum removal process, we can avoid the situation where the feature values ​​of the input model are too high or too low, thus preventing the serious interference caused by excessively high or low label feature values ​​to the model.

[0119] In step S312, the de-extremeized static feature X' is input into the XGBoost model. The maximum tree depth of the XGBoost model is set based on the sensitivity of individual risk to feature dimensions.

[0120] Empirical tests have shown that when the ratio of sample size to feature size is too small, the model is prone to overfitting. The greater the tree depth, the higher the risk of overfitting. Therefore, it is necessary to set the maximum tree depth of the XGBoost model appropriately.

[0121] Preferably, in this embodiment, the maximum tree depth is set to 10. In practical applications, it has been found that this can effectively reduce the risk of model overfitting. In this case, the calculation formula for the maximum tree depth is configured as follows:

[0122]

[0123] S313, By training the XGBoost model, calculate and output the confidence level of whether an individual is at risk, and use this confidence level as the static feature risk score.Xgb For m samples, Score Xgb It is an m-dimensional vector containing the static feature risk score of each of the m samples, where each score satisfies the Score Xgb ∈[0-100].

[0124] Dynamic risk is spatiotemporally related data. When calculating the dynamic characteristic risk score of an individual to be evaluated, this invention uses a dynamic spatiotemporal fusion prediction model to predict the spatial location coordinates of the individual to be evaluated at future time points and the characteristic changes of b dynamic characteristic dimensions.

[0125] Specifically, the dynamic spatiotemporal fusion prediction model of this invention adopts a dynamic spatial-dual parallel temporal prediction model based on GCN-LSTM-Arima. This dynamic spatial-dual parallel temporal prediction model consists of two parts: a graph convolutional network and a temporal prediction network. The model aims to predict future time points (e.g., the next time step). The spatial coordinates and b-dimensional feature changes up to time T (t+T) are expressed by the formula:

[0126]

[0127] In the formula, This represents the characteristic data of the aforementioned individual to be evaluated at the aforementioned future time points. Represents spatial features G and historical T time series data. Related functions.

[0128] The final output of the dynamic spatial-dual-track temporal prediction model is the value of the individual to be evaluated at time [time value missing]. The predicted state is expressed by the following formula:

[0129]

[0130] At this point, the steps for calculating the dynamic characteristic risk score of the individual to be assessed can be as follows:

[0131] S321, Obtain the output of the dynamic spatial-dual-track temporal prediction model.

[0132] S322, Based on the aforementioned output results, the dynamic characteristic risk score is calculated through the risk integration assessment module. Stgnn .

[0133] The risk integration assessment module is configured to perform the following calculations:

[0134]

[0135] Part One express The average increase or decrease of risk characteristics at any given time; an increase in the value indicates a higher level of risk, while a decrease in the value indicates a lower level of risk.

[0136] Middle part express The rate of change of the trajectory at any given time; the larger the value, the faster the trajectory changes, and the higher the risk.

[0137] Part Three Indicates the nearest risk location to the individual. The greater the distance, the lower the score; the smaller the distance, the higher the score. v represents the highest risk value, which is greater than 0.

[0138] In specific implementation, the dynamic spatial-dual-track temporal prediction model includes a graph convolutional neural network (GCN) model, a long short-term memory (LSTM) network model, and an autoregressive moving average (Arima) model. (See [link to relevant documentation]). Figure 4 As shown, the dynamic spatial-dual-track temporal prediction model is configured to perform the following steps:

[0139] S3201, Obtain historical T time series data. Will The data is input into the GCN model, which uses the GCN model to detect individual trajectory information to obtain the trajectory topology, thereby obtaining the spatial characteristics of the individual.

[0140] S3202 inputs spatially characteristic time series into an LSTM gated recurrent unit model and an Arima time series prediction model, and obtains the dynamic changes of the time series through information transmission between units, thus capturing time features.

[0141] S3203 concatenates the output values ​​of the LSTM model and the Arima model to obtain the next time step. spatiotemporal characteristics

[0142] Among them, the GCN model is a multi-layered structure model for detecting human trajectory information.

[0143] Preferably, this embodiment employs a two-layer GCN model, where A represents the adjacency matrix of the observation points, and the formula is used to... The transformation yields a self-connected structure, and the preprocessing is represented as follows:

[0144]

[0145] In the formula, The degree matrix representing the trajectory of all individuals at the observation point.

[0146] Obtain the weight matrices for the first and second layers, W0 and W1 respectively. Use the functions Ω() and ReLU() as activation functions (an activation function is a function that runs on the neurons of an artificial neural network, responsible for mapping the neuron's input to its output). The formula for the two-layer GCN model is as follows:

[0147]

[0148] Obtaining temporal correlation is crucial for risk prediction. This embodiment uses the gating mechanism of an LSTM model to acquire and record temporal correlation. Let h be the time correlation. t-1 Let x be the hidden layer state at time t-1. a+1 ,x a+2 ,...,x a+b ) t Let f(A,Y) be the feature content at time t. t The process of GCN graph convolution is shown, where w and k are the weights and biases during training, respectively, and r is the weight. t This represents the reset gate state at time t, used to control the degree to which information from the previous state is ignored, and uses the activation function Ω(); t This represents the update gate at time t, used to control the degree to which the state information from the previous time step is incorporated into the current state, using the activation function Ω(); s t The information stored at time t is represented using the activation function tanh(); h t This represents the output state at time t; corresponding weights are set for r, u, and s, which are w in order. r w u w c , and the corresponding deviation, k in sequence r k u k c .

[0149] At this point, the training and update process of the LSTM model is as follows:

[0150] u t =Ω(W u [f(A,Y t ),h t-1 ]+k u ),

[0151] r t =Ω(W r [f(A,Y t ),h t-1 ]+k r),

[0152] s t =tanh(W c [f(A,Y t ),(r t *h t-1 )]+k c ),

[0153] h t =u t *h t-1 +(1-u t )*s t .

[0154] Compared to the most widely used sequence processing neural network model RNN (Recurrent Neural Network) in existing technologies, this invention solves the problems of gradient vanishing and gradient explosion that exist when using the RNN model by using the gating mechanism of the LSTM model to remember more effective information over time.

[0155] This embodiment uses the Arima model to predict future moments. The risk profile is represented by Arima(p,d,q), where p represents the autoregressive order of the AR term, q represents the moving average order of the MR term, and d represents the difference order for the sequence to become stable. The feature prediction results output by the LSTM are denoted as... Let the feature prediction result output by Arima be denoted as final The risk assessment score output at any time is in, This indicates the weight of the prediction result, which is set by the system or the user.

[0156] In this embodiment, during the training of the dynamic spatial domain-dual parallel time domain prediction model, the loss function is...

[0157] The loss is configured as follows:

[0158] loss = ||Y t -Y t ′||+λL reg ;

[0159] Among them, Y t and Y t ′ represent the true value and predicted value of the feature label at time t, respectively; L reg λ is the L2 regularization term used to avoid overfitting, and it is a preset value; λ is a preset hyperparameter.

[0160] In this embodiment, the task prediction module is used to predict the dynamic risk trend information with the highest confidence level.

[0161] Specifically, the task prediction module is configured to: acquire multiple dynamic risk trend information segments; and, for each dynamic risk trend segment, generate the results output by the dynamic spatial-dual-track temporal prediction model. Combined with the causal feature matrix Z, the dynamic risk trend with the highest confidence is predicted through a configured fully connected neural network.

[0162] For the causal feature matrix Z, we obtain c feature dimensions of causal association, and sum Z by rows to obtain a vector with c rows and 1 column, i.e., zc. c*c =sum(z) c*c c)=z c*1 In the formula, sum(·) represents the summation function; then, the feature dimensions without causal relationships are completed, set to 0, and transformed into z. n*1 ; matrix z n*1 Compared with the prediction results Concatenate to obtain a vector The vector dimension is (n+2+b)*1. The aforementioned vector is used to predict the optimal dynamic risk trend category and its confidence level through a configured multi-layer fully connected neural network. The loss function can be the sigmoid function.

[0163] The Fully Connected Neural Network (FCNN), also known as a feedforward neural network, is the most basic type of neural network structure. Each neuron in each layer is connected to every neuron in the layer below, hence the name "fully connected." Preferably, this embodiment uses a three-layer fully connected neural network, including an input layer, a hidden layer, and an output layer.

[0164] The solution provided by this invention employs an improved Dynamic Spatial-Dual-Track Temporal Prediction (STGNN) model, enabling more accurate and rapid prediction of individual risks. This invention categorizes features into static, dynamic, and causal features, and predicts the potential risk value and future risk trends of the individual being assessed through a combination of spatiotemporal fusion and causal reasoning. It can efficiently and dynamically calculate the risk status of the assessed object, solving the problems of one-sided, static, and incomplete risk assessment. This achieves a comprehensive consideration and understanding of risk factors, allowing researchers, rescue personnel, and other relevant personnel to quickly and accurately grasp individual risks and take proactive response measures. Compared with existing pure machine learning clustering and regression algorithms, the model of this invention significantly improves prediction accuracy and demonstrates remarkable effectiveness.

[0165] Another embodiment of the present invention also provides an individual risk dynamic assessment system.

[0166] The system includes a data acquisition module, a data processing module, and a dynamic risk assessment module.

[0167] The data acquisition module is used to acquire risk-related data of the individual to be assessed.

[0168] The data processing module is used to divide the risk-related data into unstructured data and semi-structured data according to the complexity of the data; and to transform the aforementioned unstructured and semi-structured data into structured data to obtain the risk-related structured data of the aforementioned individual to be assessed. The risk-related structured data is divided into three dimensions: static features, dynamic features, and causal features. The static features are features that do not interfere with each other. The dynamic features are spatiotemporally related features, including time and space-related dimensions. The causal features are used to represent the feature orientation relationship and are represented by a causal feature matrix.

[0169] The aforementioned dynamic risk assessment module is used to calculate the static characteristic risk score of the individual to be assessed based on the three dimensions of data content in the aforementioned risk-related structured data. Xgb and dynamic feature risk score Stgnn The final risk score is determined based on the static and dynamic feature risk scores. final Furthermore, it obtains risk classification information, and based on the multiple dynamic risk trends classified, combines the dynamic spatiotemporal fusion prediction results and the aforementioned causal feature matrix to predict the dynamic risk trend information with the highest confidence.

[0170] Other technical features are described in the preceding embodiments and will not be repeated here.

[0171] In the above description, the disclosure of this invention is not intended to limit itself to these aspects. Rather, within the scope of the objectives of this disclosure, components can be selectively and operationally combined in any number. Furthermore, terms such as “comprising,” “encompassing,” and “having” should be interpreted by default as inclusive or open-ended, rather than exclusive or closed, unless explicitly defined as such. All technical, scientific, or other terms are to be understood by those skilled in the art, unless defined as such. Public terms found in dictionaries should not be interpreted in the context of the relevant technical documents in an overly idealistic or impractical manner, unless explicitly defined as such in this disclosure. Any modifications or alterations made by those skilled in the art based on the foregoing disclosure are within the scope of the claims.

Claims

1. A method for dynamic assessment of individual risk based on spatiotemporal causal graph network, characterized in that The method comprises the steps of: obtaining risk-related data of an individual to be evaluated, and dividing the risk-related data into unstructured data and semi-structured data according to the complexity of the data; The foregoing unstructured data and semi-structured data are converted into structured data by a hybrid logic model to obtain risk-related structured data of the foregoing individual to be evaluated, which is divided into three dimensions according to static features, dynamic features and causal features; the static features are features that do not interfere with each other; the dynamic features are features related to space-time, including time and space related dimensions; the causal features are used to represent the directional relationship of the features, which are represented by a causal feature matrix; the hybrid logic model is configured with keyword matching rules , time and space position regular matching rules and logic thinking chain extraction rules ; According to the data content of the three dimensions in the aforementioned risk-related structured data, a static characteristic risk score of the aforementioned individual to be evaluated is calculated and a dynamic characteristic risk score , and a final risk score is determined according to the static characteristic risk score and the dynamic characteristic risk score ; and the division information of the risk is obtained, the dynamic risk trend information with the highest confidence is predicted according to the divided multiple dynamic risk trends, combined with the dynamic space-time fusion prediction result and the aforementioned causal characteristic matrix. The mixed logic model is configured to obtain unstructured data of an individual to be evaluated and semi-structured data , where n represents the dimension of the set individual risk-related data, the risk-related data of the individual to be evaluated is composed of n different dimensions, and n is an integer greater than or equal to 1; and the unstructured data UD and semi-structured data SSD are converted into structured data according to the rules configured in the foregoing , the structured data SD contains the contents of three dimensions, namely static feature data X, dynamic feature data Y, and causal feature matrix Z, the X is configured with a feature dimension, the Y is configured with b feature dimensions, and the Z is configured with c feature dimensions, a, b, and c are all integers greater than or equal to 1, and , ; At this time, the static feature data is represented as: ， X a Xa represents static feature data of the a-th dimension; The dynamic feature data is represented as: ， wherein, represents the state of the individual to be evaluated at the observation time t, (dx, dy) t represents the latitude and longitude coordinates of the individual to be evaluated at the observation time t, t represents the value vector of the b dynamic feature dimensions of the individual to be evaluated at the observation time t, T represents the time sequence of the entire observation, and t represents the final observation time, represents each observation time unit, where t-T represents the initial observation time, represents the time of the next step prediction at the observation time t, represents the time of the next step prediction at the observation time t-T. The causal feature matrix Z is generated according to the strength of the causal correlation between the features, and the strength of the causal correlation between the features can be determined through the causal feature matrix Z.

2. The method of claim 1, wherein: The final risk score and / or dynamic risk trend information are output on a data visualization interface; or, The final risk score and / or dynamic risk trend information are sent to an associated user terminal for an associated person to obtain risk information of the individual to be evaluated, and risk response measure information fed back by the associated person terminal is output after receiving the risk response measure information.

3. The method of claim 1, wherein: The static feature risk score and the dynamic feature risk score are weighted to obtain a final risk score, and the calculation formula is as follows, ， wherein, represents a risk weight of a static feature, represents a risk weight of a dynamic feature, and The weight setting value is set by a system default or by a user self-defined or received associated expert experience module.

4. The method according to any one of claims 1-3, characterized in that: The risk-related data of the individual to be evaluated is encrypted data, which is mapped into ciphertext through a data encryption algorithm; after obtaining the encrypted data of the individual to be evaluated, the encrypted data is decrypted, and then the decrypted data is divided into unstructured data and semi-structured data.

5. The method of claim 1, wherein: A causal feature matrix generation module is provided, which is configured to perform the following steps: Initialize the causal strength level between the features, and different causal strength levels are mapped to different causal feature values, which can reflect the strength level of the causal correlation between the features; Collect the directional relationship between the features and the causal strength level information between the features configured by the user; According to the causal strength level information between the features, based on the mapping relationship between the causal strength level and the causal feature value, the causal feature values between the features are obtained, and the causal feature values between the features are converted into a graph association matrix, that is, a causal feature matrix Z.

6. The method of claim 1, wherein: The step of obtaining a sample of m individuals to be evaluated, calculating the static feature risk score of the individual to be evaluated is: S310, the continuity variable in the static feature X is adjusted by de-extreme value, as follows: calculate the average value of each feature and standard deviation , based on the pre-set reasonable range of the feature , adjust the static feature X to obtain the de-extreme value of the static feature X', for any feature X in the static feature X i , i=1, 2, …, a, the calculation formula of de-extreme value adjustment is as follows, ； S320, input the de-extreme value static feature X' into the XGBoost model; wherein the maximum tree depth of the XGBoost model is set according to the sensitivity of the individual risk to the feature dimension; S330, by training the XGBoost model, calculating and outputting the confidence of whether the individual is at risk, and taking the confidence as a static feature risk score , is an m-dimensional vector containing the static feature risk score of each sample in the m samples, wherein each score satisfies .

7. The method of claim 1, wherein: When calculating the dynamic feature risk score of the individual to be evaluated, a dynamic space-time fusion prediction model is used to predict the spatial position coordinates of the individual to be evaluated at a future time point and the feature change of the b dynamic feature dimensions; The dynamic space-time fusion prediction model adopts a dynamic space-dual parallel time prediction model, which is composed of a graph convolutional neural network and a time series prediction network. From the beginning to the future T moment The formula is represented as: ， wherein represents the feature data of the individual to be evaluated at the future point in time, represents a function related to the spatial feature G and the T time series data of the past T. The final output of the dynamic spatial-dual parallel temporal prediction model is the predicted state of the aforementioned individual to be evaluated at time , which is expressed by the following formula: ， At this time, the steps of calculating the dynamic feature risk score of the individual to be evaluated are as follows: Obtaining output results of a dynamic spatial-dual rail temporal prediction model , According to the foregoing output result, a dynamic characteristic risk score is calculated by a risk integration evaluation module configured to perform the following operation: ， wherein the first part represents the average amplitude of increase or decrease of the risk feature at the moment, the value increase indicates the risk level increases, and the value decrease indicates the risk level decreases; the middle part represents the change rate of the trajectory at the moment, the greater the value, the faster the change rate of the trajectory, and the higher the risk; the third part represents the distance from the individual's nearest risk position , the greater the distance, the lower the score, and the smaller the distance, the higher the score, represents a preset highest risk value and is greater than 0.

8. The method of claim 7, wherein: The dynamic space-time domain double-track prediction model includes a graph convolutional neural network GCN model, a long short-term memory network LSTM model and an autoregressive moving average model Arima model, and the prediction steps are as follows: Obtain historical T time series data ,Will The data is input into the GCN model, which is used to detect individual trajectory information to obtain the trajectory topology and thus the spatial characteristics of the individual. The time series with spatial features are input into the LSTM model and the Arima model, the dynamic changes of the time series are obtained through information transmission between units, and the time features are captured; The LSTM model is spliced with the values output by the Arima model to obtain the spatiotemporal features of the next moment . ; Wherein, the GCN model adopts a two-layer structure model for detecting the character trajectory information, and A represents an adjacent matrix of the observation point, and the formula is The transformation obtains a self-connection structure, and the pretreatment process is represented as , wherein, A represents a degree matrix of all individual trajectories at the observation point, ; the weight matrix of the first layer and the second layer is obtained, which is W0 and W1 respectively, and the function 、 is used as an activation function, and the formula of the two-layer structure GCN model is represented as follows: ； and the time correlation is obtained and recorded through the gating mechanism of the LSTM model is the hidden layer state at time t-1, is the feature content at time t, and t is denoted as the process of GCN graph convolution, and w and k are the weights and biases in the training process, respectively. The training and updating process of the LSTM model is as follows: ， ， ， ， wherein, r t denotes the reset gate state at time t, which is used to control the degree of ignoring the state information of the previous moment, and the corresponding weight and bias are w r and k r ; u t denotes the update gate at time t, which is used to control the degree of bringing the state information of the previous moment into the current state, and the corresponding weight and bias are w u and k u ; s t denotes the information content stored at time t, and the corresponding weight and bias are w c and k c ; h t denotes the output state at time t; and the risk condition at future time is predicted by Arima model denotes Arima model, where p denotes the autoregressive order of AR term, q denotes the moving average order of MR term, d denotes the difference order to become a stable sequence; the feature prediction result output by LSTM is recorded as the feature prediction result output by Arima is recorded as the final risk evaluation score output at time is wherein, denotes prediction result weight, which is set by system or user;​​ And, in the dynamic air space-double parallel track time domain prediction model training, the loss function loss is configured as: ​ wherein, and respectively represent the true value and the predicted value of the feature label at time t; is an L2 regularization term to avoid overfitting, and is a preset value; is a preset hyperparameter.

9. The method of claim 7, wherein: The task prediction module is configured to: obtain the divided plurality of dynamic risk trend information; for the plurality of dynamic risk trends, output a result of a dynamic airspace-double rail time domain prediction model and the causal feature matrix Z, the highest confidence dynamic risk trend is predicted by the configured fully connected neural network; Wherein, for the causal feature matrix Z, c feature dimensions of causal correlation are obtained, Z is summed by rows to obtain a vector of c rows and 1 column, that is , wherein sum(·) represents a summation function; then, the feature dimensions without causal correlation are completed and set to 0, and transformed into ; the matrix is spliced with the prediction result to obtain a vector , and the vector dimension is ; the foregoing vector is predicted through a configured multi-layer fully connected neural network to obtain the best dynamic risk trend category and the confidence thereof.

10. A system for dynamic assessment of individual risk according to the method of claim 1, characterized by The method comprises the steps of: a data acquisition module for obtaining risk-related data of an individual to be evaluated; a data processing module for dividing the risk-related data into unstructured data and semi-structured data according to the complexity of the data; And, the aforementioned unstructured data and semi-structured data are converted into structured data to obtain risk-related structured data of the aforementioned individual to be evaluated, which is divided into three dimensions according to static characteristics, dynamic characteristics and causal characteristics; the static characteristics are characteristics that do not interfere with each other; the dynamic characteristics are characteristics related to space-time, including time and space related dimensions; the causal characteristics are used to represent the directional relationship of characteristics, which is represented by a causal characteristic matrix; a risk dynamic assessment module, configured to calculate a static characteristic risk score of the aforementioned individual to be assessed according to the data content of the three dimensions of the aforementioned risk-related structured data and a dynamic characteristic risk score , and determine a final risk score according to the static characteristic risk score and the dynamic characteristic risk score ; And, the division information of the risk is obtained, and the dynamic risk trend information with the highest confidence is predicted according to the divided multiple dynamic risk trends, combined with the dynamic space-time fusion prediction result and the aforementioned causal characteristic matrix.

Images