Special crowd identification method and device and computer readable storage medium

Abstract

The invention belongs to the technical field of special crowd recognition, and relates to a special crowd recognition method and device and a computer readable storage medium. Inputting the knowledge graph in the special crowd knowledge graph training set into an encoder, and outputting a feature vector of each individual node in the knowledge graph; calculating a mean value and a covariance of the feature vectors of all the individual nodes, and constructing special population potential feature distribution conforming to normal distribution based on the mean value and the covariance; inputting the feature vector of each individual node and the potential feature distribution of the special crowd into a decoder, obtaining a distribution feature value of the feature vector of each individual in the potential feature distribution of the special crowd, and outputting a prediction probability that each individual node belongs to the special crowd; and constructing a prediction loss function based on the prediction probability that all individual nodes in the knowledge graph belong to the special crowd, carrying out iterative training on an encoder and a decoder, and obtaining a special crowd recognition model based on the trained encoder and decoder.

Description

Technical Field

[0001] The present invention relates to the technical field of special population recognition, and in particular to a special population recognition method, device, and computer-readable storage medium. Background Art

[0002] Special populations refer to those in society who need assistance and support, including the disabled, the elderly, and seriously ill patients, etc. The distribution of such populations in society is characterized by a wide range and diverse types. The accurate identification and precise assistance of special populations have always been an important issue in the field of social services. Traditional special population identification often relies on manual investigation and review, which not only requires a large amount of human resources, has low identification efficiency, but also has a low identification accuracy rate. With the development of artificial intelligence technology, artificial intelligence technology has gradually been applied to special population identification to improve the identification accuracy and efficiency.

[0003] Since the identification of special populations based on artificial intelligence needs to process massive social, age, disease, and other multi-source data to obtain diversified feature information of each individual, so as to identify whether each individual belongs to the special population. And the knowledge graph can classify and sort massive data in different fields according to entities and relationships, and at the same time can build connections between different individuals, thus accurately reflecting the social information and social relationships of individuals, providing strong support for the identification of special populations. Therefore, most of the existing technologies identify special populations by constructing a knowledge graph.

[0004] The special population identification method based on the knowledge graph first collects text data such as social contacts and medical records of multiple individuals through channels in fields such as social networks, public affairs, and medical care; uses a large language model to extract individual names, social contacts, and disease names as entities from the collected large amount of text data, and connects the entities by extracting the relationships between the entities, thereby constructing a knowledge graph; finally, uses a graph convolutional neural network to perform convolution on the knowledge graph, fully mining the social, disease, and other feature information of the individual nodes representing individuals in the knowledge graph, and predicting whether the individuals represented by each individual node belong to the special population based on this feature information. However, this method simply makes a judgment based on the individual's social and disease characteristics, and may misjudge individuals with certain single special characteristics but actually within the normal range of fluctuations as special populations, resulting in a low accuracy rate of the identification result.

[0005] In summary, the existing special population identification method based on the knowledge graph has the problem that it only focuses on the surface characteristics of data such as the individual's social contacts and diseases, resulting in a low accuracy rate of the identification result. Summary of the Invention

[0006] To this end, the technical problem to be solved by the present invention is to overcome the problem in the prior art that the special population recognition method based on the knowledge graph only focuses on the surface features of data such as the social diseases of individuals, resulting in low accuracy of the recognition results.

[0007] To solve the above technical problem, the present invention provides a special population recognition method, including: Input the knowledge graph in the special population knowledge graph training set into the encoder, and output the feature vectors of each individual node in the knowledge graph; wherein, the individual node is the node corresponding to the individual to be recognized; Calculate the mean and covariance of the feature vectors of all individual nodes, and construct a special population latent feature distribution that conforms to the normal distribution based on the mean and covariance; Input the feature vector of each individual node and the special population latent feature distribution into the decoder, obtain the distribution characteristic value of the feature vector of each individual node in the special population latent feature distribution, and output the prediction probability that each individual node belongs to the special population; Construct a prediction loss function based on the prediction probabilities that all individual nodes in the knowledge graph belong to the special population, and perform iterative training on the encoder and the decoder, and obtain a special population recognition model based on the trained encoder and decoder.

[0008] Preferably, constructing a special population latent feature distribution that conforms to the normal distribution based on the mean and covariance includes: Use a random generation function to generate multiple Gaussian noise values that follow the standard normal distribution; Perform linear transformation on each Gaussian noise value using the mean and covariance respectively to obtain multiple latent feature distribution values; Input each latent feature distribution value of each individual node into the probability density function of the standard normal distribution, and output the probability density of each latent feature distribution value; Construct a special population latent feature distribution based on all latent feature distribution values and their probability densities.

[0009] Preferably, inputting the feature vector of each individual node and the special population latent feature distribution into the decoder, obtaining the distribution characteristic value of the feature vector of each individual node in the special population latent feature distribution, and outputting the prediction probability that each individual node belongs to the special population includes: Input the special population latent feature distribution into the second graph convolutional neural network for feature extraction, and output the numerical distribution feature vector of the special population latent feature distribution; Construct a probability density function and a cumulative distribution function of the potential feature distribution of the special population based on the numerical distribution feature vector of the potential feature distribution of the special population; and input the feature vector of each individual node into the probability density function and the cumulative distribution function respectively to obtain the probability density function value and the distribution percentage of each individual node. Input the probability density function value and the distribution percentage of each individual node into the linear layer, and output the prediction probability that each individual node belongs to the special population.

[0010] Preferably, input the knowledge graph in the training set of the special population knowledge graph into the encoder, and the output feature vector of each individual node in the knowledge graph includes: Input the knowledge graph into the graph attention neural network, assign weights to each individual node in the knowledge graph, and output the knowledge graph containing the weight coefficients of each individual node. Input the knowledge graph containing the weight coefficients of each individual node into the first graph convolutional neural network for feature extraction, and output the feature vector of each individual node.

[0011] Preferably, input the knowledge graph into the graph attention neural network, and the assignment of weights to each individual node in the knowledge graph includes: Input the knowledge graph into the feature extraction module to extract the initial feature vectors of each individual node and non-individual node in the knowledge graph. Input the initial feature vectors of all individual nodes and non-individual nodes into the linear transformation module, calculate the attention coefficients between each individual node and its connected individual nodes and non-individual nodes, and obtain the weight coefficient of each individual node.

[0012] Preferably, construct a prediction loss function based on the prediction probabilities that all individual nodes in the knowledge graph belong to the special population, including: Based on the loss when the prediction probability that each individual node belongs to the special population approximates the true probability, construct the forward KL divergence loss of each individual node; based on the sum of the forward KL divergence losses of all individual nodes, obtain the forward KL divergence prediction loss function. Based on the loss when the true probability that each individual node belongs to the special population approximates the prediction probability, construct the backward KL divergence loss of each individual node; based on the sum of the backward KL divergence losses of all individual nodes, obtain the backward KL divergence prediction loss function. Perform weighted summation on the forward KL divergence prediction loss function and the backward KL divergence prediction loss function to obtain the prediction loss function.

[0013] Preferably, the forward KL divergence prediction loss function is expressed as: , where, Denote the forward KL divergence prediction loss function; Denote the forward KL divergence loss of the th individual node; Denote the true probability that the th individual node belongs to the special population. When the th individual node belongs to the special population, ; when the th individual node does not belong to the special population, ; Denote the predicted probability that the th individual node belongs to the special population; Denote the number of individual nodes; The backward KL divergence prediction loss function is expressed as: , where Denote the backward KL divergence prediction loss function; Denote the backward KL divergence loss of the th individual node; The prediction loss function is expressed as: , where Denote the prediction loss function; Denote the weight parameter, .

[0014] Preferably, using the special population recognition model for special population recognition includes: Collect the basic information of each individual in the population to be recognized; among them, the basic information includes the identification data, case data, and social data of the individual; the identification data includes the name and age of the individual; Taking the identification data of each individual as the core entity, the disease name in the case data of each individual as the disease entity, and the social object name in the social data of each individual as the social entity; Extract the relationships between the core entity of each individual and the disease entity and social entity respectively, and the relationships between the core entities of every two individuals; Taking the core entity as the individual node, the disease entity and social entity as the nodes, and constructing the knowledge graph to be recognized based on the relationships between the core entities, the relationships between each core entity and the disease entity, and the relationships between each core entity and the social entity; Input the knowledge graph to be recognized into the encoder in the special population recognition model, and output the predicted probability that each individual node in the knowledge graph to be recognized belongs to the special population, so as to determine whether each individual in the population to be recognized is a special population.

[0015] The present invention also provides a special population recognition device, including: A feature extraction module, configured to input a knowledge graph in a special population knowledge graph training set into an encoder, and output a feature vector of each individual node in the knowledge graph; wherein, the individual node is the node corresponding to the individual to be recognized; A potential feature distribution construction module, configured to calculate the mean and covariance of the feature vectors of all individual nodes, and construct a special population potential feature distribution conforming to a normal distribution based on the mean and covariance; A probability prediction module, configured to input the feature vector of each individual node and the special population potential feature distribution into a decoder, obtain a distribution characteristic value of the feature vector of each individual node in the special population potential feature distribution, and output a prediction probability that each individual node belongs to the special population; A model training and acquisition module, configured to construct a prediction loss function based on the prediction probabilities that all individual nodes in the knowledge graph belong to the special population, perform iterative training on the encoder and the decoder, and obtain a special population recognition model based on the trained encoder and decoder.

[0016] The present invention also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the above special population recognition method are implemented.

[0017] The special population recognition method provided by this application has the following beneficial effects: Since there are obvious differences in data such as age, pathology, and social interaction between special populations and non-special populations, taking the elderly as an example, in a group, the ages of non-special populations will be concentrated near the mean, thus forming a peak, while the elderly will be distributed near a relatively low level due to their advanced age, forming an approximately normal distribution. Similarly, when disabled people, seriously ill people, or people with social difficulties are in a group, they will also be distributed near a low level far from the peak. Therefore, the distribution characteristics of special populations in a group conform to a normal distribution; based on this, after extracting the feature vectors of individual nodes in the knowledge graph, the present application constructs a special population potential feature distribution conforming to a normal distribution based on the mean and covariance of the feature vectors of all individual nodes, thereby mapping the features of individual nodes to a latent space with a normal distribution, using the decoder to analyze the distribution characteristics of the features of each individual in this latent space with a normal distribution, and outputting the prediction probability that each individual node belongs to the special population by judging whether the feature distribution of the individual node is at the center or the edge of this latent feature distribution. A loss function is constructed using the prediction probabilities of all individual nodes in the knowledge graph and the encoder and decoder are trained. Based on feature information such as social pathology, the distribution characteristics of special populations are comprehensively considered, enabling the model to comprehensively learn the data feature information and distribution characteristic information of special populations, thereby improving the accuracy of special population recognition. Brief Description of the Drawings

[0018] To make the content of the present invention easier to be clearly understood, the following further describes the present invention in detail according to specific embodiments of the present invention in conjunction with the accompanying drawings, where: Figure 1 is a flowchart of the special population recognition method provided by this application; Figure 2 is a schematic diagram of the knowledge graph construction process provided by this application; Figure 3 is a schematic diagram of the structure of the special population recognition device provided by this application. Detailed Embodiments

[0019] The following further describes the present invention in conjunction with the accompanying drawings and specific embodiments, so that those skilled in the art can better understand the present invention and be able to implement it, but the embodiments cited are not used as a limitation to the present invention.

[0020] Please refer to Figure 1 , Figure 1 shown is a flowchart of the special population recognition method provided by this application, and the method specifically includes: S10: Input the knowledge graph in the special population knowledge graph training set into the encoder, and output the feature vectors of each individual node in the knowledge graph; wherein, the individual node is the node corresponding to the individual to be recognized.

[0021] S20: Calculate the mean and covariance of the feature vectors of all individual nodes, and construct a special population latent feature distribution that conforms to the normal distribution based on the mean and covariance.

[0022] Specifically, the construction steps of the special population latent feature distribution include: Step 1: Use a random generation function to generate multiple Gaussian noise values that follow the standard normal distribution.

[0023] Step 2: Perform linear transformation on each Gaussian noise value using the mean and covariance respectively to obtain multiple latent feature distribution values.

[0024] , wherein, represents the th latent feature distribution value; represents the mean; represents the covariance; represents the th Gaussian noise value.

[0025] Step 3: Input each latent feature distribution value of each individual node into the probability density function of the standard normal distribution, and output the probability density of each latent feature distribution value.

[0026] Step 4: Based on all potential feature distribution values and their probability densities, construct the potential feature distribution of the special population.

[0027] S30: Input the feature vector of each individual node and the potential feature distribution of the special population into the decoder, obtain the distribution characteristic value of the feature vector of each individual node in the potential feature distribution of the special population, and output the prediction probability that each individual node belongs to the special population.

[0028] S40: Construct a prediction loss function based on the prediction probabilities that all individual nodes in the knowledge graph belong to the special population, and perform iterative training on the encoder and the decoder. Obtain the special population recognition model based on the trained encoder and decoder.

[0029] Furthermore, in some embodiments of the present application, the specific implementation manner of step S10 is as follows: S100: Input the knowledge graph into the graph attention neural network, assign weights to each individual node in the knowledge graph, and output the knowledge graph containing the weight coefficients of each individual node.

[0030] S101: Input the knowledge graph containing the weight coefficients of each individual node into the first graph convolutional neural network for feature extraction, and output the feature vector of each individual node.

[0031] Furthermore, inputting the knowledge graph into the graph attention neural network in step S100 and assigning weights to each individual node in the knowledge graph specifically includes: Input the knowledge graph into the feature extraction module to extract the initial feature vectors of each individual node and non-individual node in the knowledge graph; Input the initial feature vectors of all individual nodes and non-individual nodes into the linear transformation module, calculate the attention coefficients between each individual node and the individual nodes and non-individual nodes connected to it, and obtain the weight coefficient of each individual node.

[0032] Specifically, first calculate the attention coefficient between each individual node and each node (individual node or non-individual node) connected to it, and then perform a weighted sum of the attention coefficients of each individual node and all connected nodes to obtain the weight coefficient calculated for each individual.

[0033] Specifically, the calculation formula for the attention coefficient between each individual node and each node connected to it is: , where, represents the th individual node and the th node connected to it; represents the activation function; Represents the feature vector of the th individual node; Represents the th node connected to the th individual node; Represents the set of nodes connected to the th individual node; Represents the weight matrix.

[0034] Furthermore, in some embodiments of the present application, the specific implementation manner of step S30 includes: S300: Input the potential feature distribution of the special population into the second graph convolutional neural network for feature extraction, and output the numerical distribution feature vector of the potential feature distribution of the special population.

[0035] S301: Based on the numerical distribution feature vector of the potential feature distribution of the special population, construct the probability density function and cumulative distribution function of the potential feature distribution of the special population; and input the feature vector of each individual node into the probability density function and cumulative distribution function respectively to obtain the probability density function value and distribution percentage of each individual node.

[0036] S302: Input the probability density function value and distribution percentage of each individual node into the linear layer, and output the predicted probability that each individual node belongs to the special population.

[0037] Since the distribution characteristics of the special population in the group tend to be marginal, therefore, the present application reflects whether the individual node belongs to the special population by obtaining the position and probability of the feature vector of the individual node in the potential feature distribution of the special population. Specifically, if the probability density function value of the feature vector of a certain individual node in the potential feature distribution of the special population that conforms to the normal distribution is low, it means that the probability of the feature of the individual node appearing in the feature space is very small, that is, it is in a very uncommon position in the overall feature distribution. At the same time, if the distribution percentage of the feature vector of the individual node is extremely high or extremely low, it can also be considered that the feature of the individual node belongs to a few special cases. Therefore, by combining the distribution probability density value and distribution percentage of the individual node in the potential feature distribution of the special population, it can be determined whether the individual node is in a position far from the peak of the normal distribution, so as to determine whether the individual node belongs to the special population.

[0038] Furthermore, this application uses the KL divergence to construct a loss function to prevent the model from overfitting and improve the robustness of the model against noise and outliers. Specifically, the forward KL divergence prefers that the main part of the predicted data distribution result covers the main part of the true data distribution as much as possible, so that the model can learn the average features in the true data distribution. However, this will ignore the main pattern features in the true data distribution; while the backward KL divergence loss focuses more on making the maximum probability in the predicted data distribution result fit the maximum probability in the true data distribution as much as possible, so that the model learns the main patterns in the true data distribution result.

[0039] Since this application identifies special populations based on the distribution characteristics of special populations in the group, the distribution of individual nodes belonging to special populations and the distribution of individual nodes belonging to non-special populations among multiple individual nodes in each knowledge graph conform to the normal distribution. Therefore, the predicted probability distribution of the model for multiple individual nodes in each knowledge graph should also conform to the normal distribution. For this reason, this application uses both the forward KL divergence loss and the backward KL divergence loss when constructing the loss function, so that the predicted probability distribution of all individual nodes in a knowledge graph can cover the true probability distribution as much as possible, and at the same time, the maximum value in the predicted probability distribution can also be close to the maximum value of the true probability distribution.

[0040] Specifically, constructing a prediction loss function based on the predicted probabilities of all individual nodes in the knowledge graph belonging to special populations includes: 1. Construct the forward KL divergence loss of each individual node based on the loss when the predicted probability of each individual node belonging to the special population approximates the true probability; obtain the forward KL divergence prediction loss function based on the sum of the forward KL divergence losses of all individual nodes.

[0041] Specifically, the forward KL divergence prediction loss function is expressed as: , where, represents the forward KL divergence prediction loss function; represents the forward KL divergence loss of the th individual node; represents the true probability that the th individual node belongs to the special population. When the th individual node belongs to the special population, ; when the th individual node does not belong to the special population, ; represents the predicted probability that the th individual node belongs to the special population; represents the number of individual nodes.

[0042] 2. Construct the backward KL divergence loss for each individual node based on the loss when approximately predicting the probability that each individual node belongs to a special population; obtain the backward KL divergence prediction loss function based on the sum of the backward KL divergence losses of all individual nodes.

[0043] Specifically, the backward KL divergence prediction loss function is expressed as: , where, represents the backward KL divergence prediction loss function; represents the th backward KL divergence loss of an individual node.

[0044] 3. Perform weighted summation on the forward KL divergence prediction loss function and the backward KL divergence prediction loss function to obtain the prediction loss function.

[0045] Specifically, the prediction loss function is expressed as: , where, represents the prediction loss function; represents the weight parameter, .

[0046] Furthermore, when using the special population recognition model for special population recognition, it is necessary to first construct a knowledge graph to be recognized based on the information of the population to be recognized, so as to extract the feature information of each individual to be recognized based on this knowledge graph to be recognized, and determine whether each individual to be recognized belongs to a special population based on the distribution characteristics of the feature information of each individual to be recognized in the population to be recognized.

[0047] Specifically, the steps of using the special population recognition model for special population recognition include: Collect the basic information of each individual in the population to be recognized; among them, the basic information includes the identification data, case data, and social data of the individual; the identification data includes the name and age of the individual. Take the identification data of each individual as the core entity, the disease name in the case data of each individual as the disease entity, and the social object name in the social data of each individual as the social entity. Extract the relationships between the core entity of each individual and the disease entity and social entity respectively, as well as the relationships between the core entities of every two individuals. Take the core entity as the individual node, the disease entity and social entity as nodes, and construct the knowledge graph to be recognized based on the relationships between the core entities, the relationships between each core entity and the disease entity, and the relationships between each core entity and the social entity. Input the knowledge graph to be recognized into the encoder in the special population recognition model, and output the prediction probability that each individual node in the knowledge graph to be recognized belongs to the special population, so as to determine whether each individual in the population to be recognized is a special population.

[0048] Exemplarily, as Figure 2 shown in the schematic diagram of the knowledge graph construction process provided by this application. When collecting the basic information of an individual, data such as the name, age, medical records, and social interactions of the individual can be collected from fields such as social networks, public affairs, medical care, and education. Among them, social data can include the individual's comments, likes, and follow data, and public affairs data can include data such as low-income household data and unemployed person data in government public information.

[0049] Furthermore, in order to improve the data quality, the collected data can be preprocessed such as cleaning and deduplication first. If the preprocessed data is structured data, that is, entity-relationship-entity triple data, then directly construct a knowledge graph based on the preprocessed data. If the preprocessed data is unstructured data, then a large language model based on Bert needs to be used to extract entities and relationships in the data to obtain triple data. In addition, experts can also analyze and evaluate the quality of the triple data to further improve the data quality. Finally, a knowledge graph is constructed according to the entities and relationships in all triple data.

[0050] Based on the special population recognition method provided in the above embodiments, the embodiments of this application also provide a special population recognition device, as Figure 3 shown. The device specifically includes: A feature extraction module 10, configured to input the knowledge graph in the special population knowledge graph training set into the encoder, and output the feature vector of each individual node in the knowledge graph; wherein, the individual node is the node corresponding to the individual to be recognized; A latent feature distribution construction module 20, configured to calculate the mean and covariance of the feature vectors of all individual nodes, and construct a special population latent feature distribution that conforms to a normal distribution based on the mean and covariance; A probability prediction module 30, configured to input the feature vector of each individual node and the special population latent feature distribution into the decoder, obtain the distribution characteristic value of the feature vector of each individual node in the special population latent feature distribution, and output the prediction probability that each individual node belongs to the special population; A model training and acquisition module 40, configured to construct a prediction loss function based on the prediction probabilities that all individual nodes in the knowledge graph belong to the special population, and perform iterative training on the encoder and the decoder, and obtain a special population recognition model based on the trained encoder and decoder.

[0051] The embodiment of the present application also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the above-mentioned special population recognition method are implemented.

[0052] The special population recognition method provided by the present application extracts multi-dimensional feature information of each individual to be recognized by using a knowledge graph. At the same time, based on the characteristic that the distribution of special populations in the population conforms to the normal distribution, a potential feature distribution space that conforms to the normal distribution is constructed by using the feature vectors of all individual nodes in the knowledge graph, and all individual nodes are mapped into the potential space of the normal distribution. Thus, by judging the positions of each individual node in this normal distribution, it is determined whether each individual node belongs to a special population. Through experiments, it is proved that the present application can improve the accuracy of special population recognition based on the distribution characteristics of special populations, and thus can provide data support for rescue strategy decision-makers based on the recognition results.

[0053] Those skilled in the art should understand that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0054] The present application is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each flow and / or block in the flowchart and / or block diagram, as well as the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate means for implementing the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0055] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured article including instruction means, and the instruction means implements the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0056] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus, so that a series of operation steps are executed on the computer or other programmable apparatus to produce a computer-implemented process, and thus the instructions executed on the computer or other programmable apparatus provide steps for realizing the functions specified in one process or a plurality of processes and / or blocks Figure 1 one process or a plurality of processes and / or blocks Figure 1 steps for realizing the functions specified in one block or a plurality of blocks.

[0057] Obviously, the above embodiments are only examples for clear illustration, and are not limitations on the implementation manners. For those of ordinary skill in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation manners here. And the obvious changes or modifications derived therefrom are still within the protection scope of the present invention.

Claims

1. A method for identifying a special group of people, characterized in that: include: Input the knowledge graph in the special population knowledge graph training set into the encoder, and output the feature vector of each individual node in the knowledge graph; where the individual node is the node corresponding to the individual to be identified; Calculate the mean and covariance of the feature vectors of all individual nodes, and construct the potential feature distribution of special populations that conforms to the normal distribution based on the mean and covariance; Input the feature vector of each individual node and the potential feature distribution of the special population into the decoder, obtain the distribution characteristic value of the feature vector of each individual node in the potential feature distribution of the special population, and output the predicted probability that each individual node belongs to the special population; A prediction loss function is constructed based on the predicted probability that all individual nodes in the knowledge graph belong to special populations, and the encoder and decoder are iteratively trained to obtain a special population recognition model based on the trained encoder and decoder.

2. The special population identification method according to claim 1 is characterized in that: The potential characteristic distribution of special populations that conform to normal distribution based on mean and covariance includes: Generate multiple Gaussian noise values ​​that obey the standard normal distribution using a random generation function; Each Gaussian noise value is linearly transformed using the mean and covariance to obtain multiple potential feature distribution values; Input each potential feature distribution value of each individual node into the probability density function of the standard normal distribution, and output the probability density of each potential feature distribution value; Based on all potential feature distribution values ​​and their probability density, the potential feature distribution of special populations is constructed.

3. The special population identification method according to claim 1, characterized in that: The feature vector of each individual node and the potential feature distribution of the special population are input into the decoder to obtain the distribution characteristic value of the feature vector of each individual node in the potential feature distribution of the special population, and the predicted probability of each individual node belonging to the special population is output, including: Input the potential feature distribution of the special population into the second graph convolutional neural network for feature extraction, and output the numerical distribution feature vector of the potential feature distribution of the special population; Based on the numerical distribution feature vector of the potential feature distribution of the special population, the probability density function and cumulative distribution function of the potential feature distribution of the special population are constructed; and the feature vector of each individual node is input into the probability density function and the cumulative distribution function respectively to obtain the probability density function value and distribution percentage of each individual node; The probability density function value and distribution percentage of each individual node are input into the linear layer, and the predicted probability that each individual node belongs to a special population is output.

4. The special population identification method according to claim 1, characterized in that: The knowledge graph in the special population knowledge graph training set is input into the encoder, and the feature vector of each individual node in the output knowledge graph includes: Input the knowledge graph into the graph attention neural network, assign weights to each individual node in the knowledge graph, and output a knowledge graph containing the weight coefficients of each individual node; The knowledge graph containing the weight coefficients of each individual node is input into the first graph convolutional neural network for feature extraction, and the feature vector of each individual node is output.

5. The special population identification method according to claim 4, characterized in that: Input the knowledge graph into the graph attention neural network and assign weights to each individual node in the knowledge graph: Input the knowledge graph into the feature extraction module to extract the initial feature vector of each individual node and non-individual node in the knowledge graph; The initial feature vectors of all individual nodes and non-individual nodes are input into the linear transformation module, and the attention coefficient between each individual node and the individual nodes and non-individual nodes connected to it is calculated to obtain the weight coefficient of each individual node.

6. The special population identification method according to claim 1, characterized in that: The prediction loss function is constructed based on the predicted probability that all individual nodes in the knowledge graph belong to a special population, including: Based on the loss when the predicted probability of each individual node belonging to a special population approximates the true probability, the forward KL divergence loss of each individual node is constructed; based on the sum of the forward KL divergence losses of all individual nodes, the forward KL divergence prediction loss function is obtained; Based on the loss when the true probability of each individual node belonging to a special population approximates the predicted probability, the backward KL divergence loss of each individual node is constructed; based on the sum of the backward KL divergence losses of all individual nodes, the backward KL divergence prediction loss function is obtained; The prediction loss function is obtained by weighted summing the forward KL divergence prediction loss function and the backward KL divergence prediction loss function.

7. The special population identification method according to claim 6, characterized in that: The forward KL divergence prediction loss function is expressed as: ， in, represents the forward KL divergence prediction loss function; Indicates Forward KL divergence loss of individual nodes; Indicates The true probability that an individual node belongs to a special population is When an individual node belongs to a special population, ; When the When an individual node does not belong to a special population, ; Indicates The predicted probability that each individual node belongs to a special population; represents the number of individual nodes; The backward KL divergence prediction loss function is expressed as: ， in, represents the backward KL divergence prediction loss function; Indicates Backward KL divergence loss of individual nodes; The prediction loss function is expressed as: ， in, represents the prediction loss function; represents the weight parameter, .

8. The special population identification method according to claim 1, characterized in that: Using the special crowd identification model to identify special crowds includes: Collect basic information of each individual in the population to be identified; basic information includes individual identification data, case data, and social data; identification data includes individual name and age; The identification data of each individual is taken as the core entity, the disease name in the case data of each individual is taken as the disease entity, and the social object name in the social data of each individual is taken as the social entity; Extract the relationship between each individual's core entity and disease entity, social entity, and the relationship between each two individuals' core entities; The knowledge graph to be identified is constructed by taking core entities as individual nodes, disease entities and social entities as nodes, and based on the relationship between each core entity, the relationship between each core entity and the disease entity, and the relationship between each core entity and the social entity; The knowledge graph to be identified is input into the encoder in the special population identification model, and the predicted probability that each individual node in the knowledge graph to be identified belongs to the special population is output, so as to determine whether each individual in the population to be identified is a special population.

9. A device for identifying special groups of people, characterized in that: include: The feature extraction module is used to input the knowledge graph in the special population knowledge graph training set into the encoder and output the feature vector of each individual node in the knowledge graph; wherein the individual node is the node corresponding to the individual to be identified; The latent feature distribution construction module is used to calculate the mean and covariance of the feature vectors of all individual nodes, and to construct the latent feature distribution of special populations that conforms to the normal distribution based on the mean and covariance; The probability prediction module is used to input the feature vector of each individual node and the potential feature distribution of the special population into the decoder, obtain the distribution characteristic value of the feature vector of each individual node in the potential feature distribution of the special population, and output the predicted probability that each individual node belongs to the special population; The model training and acquisition module is used to construct a prediction loss function based on the predicted probability that all individual nodes in the knowledge graph belong to special populations, and iteratively train the encoder and decoder to obtain a special population recognition model based on the trained encoder and decoder.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by the processor, the steps of the special population identification method described in any one of claims 1 to 8 are implemented.

Images