Population data generation method and device and computer readable storage medium

Abstract

The invention discloses a population data generation method and device and a computer readable storage medium, and the method comprises the steps: carrying out the conversion of an initial population sample from an attribute domain to a copula domain, obtaining a copula domain sample, carrying out the vectorization coding representation of the copula domain sample, and obtaining a matrix sample; performing diffusion model training and sample generation based on the matrix sample to obtain a new matrix sample; and decoding the new matrix sample to obtain a copula domain vector sample, and performing quantile function mapping on the copula domain vector sample to obtain an attribute domain demographic data generation sample. According to the method, the initial population sample is converted to the copula domain, the correlation mode between the attribute variables is reflected through the copula density function, the influence of edge distribution is avoided, the intrinsic dependency constraint relation between the variables can be effectively captured, and the population sample data with high feasibility and diversity is generated.

Description

Technical Field

[0001] This application relates to the field of traffic simulation, specifically to a method, apparatus, and computer-readable storage medium for generating population data. Background Technology

[0002] In traffic simulation, it is necessary to model and represent traffic participants such as pedestrians, motor vehicles, and non-motorized vehicles, as well as traffic elements such as traffic lights, and to simulate the interactions between entities. Reasonable and accurate simulation experiments can provide effective reference and guidance for real-world traffic optimization. Among these, humans, whether pedestrians or drivers, are the core elements of traffic simulation and are direct participants in traffic, requiring effective modeling and representation. However, human behavior is closely related to individual attributes; for example, people with different attributes (gender, age, education, income) have different behavioral patterns, necessitating heterogeneous modeling and representation of people. This means constructing corresponding intelligent agents based on population data and reflecting different individual behavioral patterns. Furthermore, considering factors such as privacy protection and time costs, population data is typically obtained only through surveys, yielding a small sample size, from which larger population data is synthesized.

[0003] Existing population data generation methods include fitting and reweighting methods, which generate data by adjusting the weights of individual sample data and resampling. Simulation-based methods learn the joint probability density of multidimensional attributes of population data and then sample to generate reasonable population sample data. These methods include Bayesian networks, Hidden Markov Models, deep model-based variational autoencoders, and Generative Adversarial Networks (GANs). Existing diffusion model-based population data generation methods reshape individual population data from vectors into square matrices to meet the input format required for two-dimensional convolution operations in diffusion image generation models. They then use the diffusion model to generate and decode new population data.

[0004] However, as a generative model of a stochastic process, the diffusion model performs well in the field of image and video generation. However, in population data generation, the population data generation method based on the diffusion model only changes the form of individual population data when reshaping population data into a square matrix. It does not consider the discrete characteristics of the attribute values ​​of each dimension and effectively represent them. It cannot capture the essential dependency and constraint relationship between variables, resulting in poor feasibility and diversity of generated population data. Summary of the Invention

[0005] This application provides a population data generation method, apparatus, and computer-readable storage medium, which can solve the technical problems existing in the prior art, such as insufficient representation of discrete attribute values ​​of each dimension, inability to capture the essential dependency and constraint relationships between variables, resulting in poor feasibility and diversity of the generated population data.

[0006] In a first aspect, embodiments of this application provide a method for generating population data, including: The initial population sample is transformed from the attribute domain to the copula domain to obtain the copula domain sample, and the copula domain sample is vectorized and encoded to obtain the matrix sample. New matrix samples are obtained by training the diffusion model and generating samples based on matrix samples; The new matrix sample is decoded to obtain the copula domain vector sample, and the copula domain vector sample is mapped by the quantile function to obtain the attribute domain population data generation sample.

[0007] In conjunction with the first aspect, in one implementation, transforming the initial population sample from the attribute domain to the copula domain to obtain a copula domain sample includes: Copula domain samples are obtained by mapping each attribute dimension of the initial population sample data using the empirical cumulative distribution function.

[0008] In conjunction with the first aspect, in one implementation, the initial population sample includes discrete, unordered variables. Before transforming the initial population sample from the attribute domain to the copula domain to obtain the copula domain sample, the method further includes: Map the values ​​of discrete, unordered variables to a continuous sequence of integers.

[0009] In conjunction with the first aspect, in one implementation, the vectorized encoding representation of copula domain samples yields matrix samples, including: The samples in the copula domain are encoded using sine and cosine vectors to obtain matrix samples.

[0010] In conjunction with the first aspect, in one implementation, new matrix samples are obtained by training a diffusion model and generating samples based on matrix samples, including: Diffusion model trained based on matrix samples; New matrix samples are generated based on the trained diffusion model.

[0011] In conjunction with the first aspect, in one implementation, the diffusion model is constructed based on a denoised probabilistic diffusion model, including a forward Markov chain and a reverse Markov chain.

[0012] In conjunction with the first aspect, in one implementation, decoding the new matrix sample to obtain a copula domain vector sample includes: The new matrix sample is decoded by multi-dimensional weighted optimization to obtain the copula domain vector sample.

[0013] In conjunction with the first aspect, in one implementation, generating population data samples for the attribute domain by performing quantile function mapping on copula domain vector samples includes: The population data of the attribute domain is generated by mapping the copula domain vector samples along each dimension using the empirical quantile function.

[0014] Secondly, embodiments of this application provide a population data generation apparatus, comprising: The sample transformation unit is used to transform the initial population sample from the attribute domain to the copula domain to obtain the copula domain sample, and to vectorize the copula domain sample to obtain the matrix sample. The sample diffusion generation unit is used to train the diffusion model based on matrix samples and generate new matrix samples. The decoding output unit is used to decode the new matrix sample to obtain the copula domain vector sample, and to perform quantile function mapping on the copula domain vector sample to obtain the attribute domain population data generation sample.

[0015] Thirdly, embodiments of this application provide a computer-readable storage medium storing a population data generation program, wherein when the population data generation program is executed by a processor, it implements the steps of the population data generation method as described above.

[0016] The beneficial effects of the technical solutions provided in this application include at least the following: By transforming the initial population sample into the copula domain, the copula density function reflects the correlation pattern between attribute variables, unaffected by marginal distributions. This effectively captures the essential dependency and constraint relationships between variables, generating highly feasible and diverse population sample data. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating an embodiment of the population data generation method of this application; Figure 2 This is a schematic diagram illustrating the diffusion model principle of an embodiment of this application; Figure 3 This is a network structure diagram of the noise prediction network in an embodiment of this application; Figure 4 This is a schematic diagram of the functional modules of an embodiment of the population data generation device of this application. Detailed Implementation

[0018] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0019] The terms "comprising" and "having," and any variations thereof, in the specification, claims, and accompanying drawings of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such process, method, product, or apparatus. The terms "first," "second," and "third," etc., are used to distinguish different objects, etc., and do not indicate a sequence, nor do they limit "first," "second," and "third" to different types.

[0020] In the description of the embodiments of this application, terms such as "exemplary," "for example," or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as "exemplary," "for example," or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary," "for example," or "for instance" is intended to present the relevant concepts in a concrete manner.

[0021] In the description of the embodiments of this application, unless otherwise stated, " / " means "or". For example, A / B can mean A or B. The "and / or" in the text is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist simultaneously, and B exists alone. In addition, in the description of the embodiments of this application, "multiple" means two or more.

[0022] In some processes described in the embodiments of this application, multiple operations or steps are included in a specific order. However, it should be understood that these operations or steps may not be executed in the order they appear in the embodiments of this application, or they may be executed in parallel. The sequence number of the operation is only used to distinguish different operations, and the sequence number itself does not represent any execution order. In addition, these processes may include more or fewer operations, and these operations or steps may be executed sequentially or in parallel, and these operations or steps may be combined.

[0023] First, some of the technical terms used in this application will be explained to help those skilled in the art understand this application.

[0024] Copula domain: refers to the new data space formed after mapping and transforming the original attribute domain data. Its mathematical representation is as follows: ,in The number of attribute dimensions.

[0025] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0026] In a first aspect, embodiments of this application provide a method for generating population data.

[0027] In one embodiment, reference is made to Figure 1 , Figure 1 This is a schematic flowchart of an embodiment of the population data generation method of this application. Figure 1 As shown, population data generation methods include: S101. Transform the initial population sample from the attribute domain to the copula domain to obtain the copula domain sample, and then vectorize and encode the copula domain sample to obtain the matrix sample. S102. Based on matrix samples, perform diffusion model training and sample generation to obtain new matrix samples; S103. Decode the new matrix sample to obtain the copula domain vector sample, and perform quantile function mapping on the copula domain vector sample to obtain the attribute domain population data generation sample.

[0028] Specifically, considering that existing technologies directly reshape the initial population sample vector into a square matrix, they fail to take into account the discrete characteristics of the attribute values ​​of each dimension and cannot effectively represent them, thus failing to capture the essential dependency and constraint relationships between variables, this embodiment captures the essential dependency and constraint relationships between variables by transforming the population sample from the attribute domain to the copula domain and representing it effectively.

[0029] First, the implementation example maps the initial population samples to the copula domain using the empirical cumulative distribution function. Then, sine and cosine encoding is used to vectorize each dimension of the copula domain samples, resulting in matrix samples. The processed matrix samples are then used to train a diffusion model, which generates new matrix samples. In the decoding stage, considering the noise inherent in the generated matrix data, the implementation example uses a multi-dimensional weighted optimization method for robust decoding, decoding the generated matrix samples back to the copula domain to obtain copula domain vector samples. Then, considering the discrepancy between the empirical cumulative distribution function and the true cumulative distribution function, and the noise inherent in the samples generated by the diffusion model, the implementation example finally maps the generated copula domain samples to the attribute domain using the empirical quantile function (empirical inverse cumulative distribution function), resulting in population data generated samples.

[0030] In this embodiment, by transforming the training to the copula domain, the copula density function can effectively reflect the correlation patterns between attribute variables, such as tail dependence and nonlinear relationships, without being affected by marginal distributions, thus characterizing the essential dependency constraints between variables. Consequently, population data with higher diversity and feasibility can be generated. This solves the technical problem in existing technologies that fail to consider the discrete characteristics of attribute values ​​in each dimension, thus failing to capture the essential dependency constraints between variables, resulting in poor feasibility and diversity of the generated population data.

[0031] Furthermore, in one embodiment, transforming the initial population sample from the attribute domain to the copula domain to obtain a copula domain sample includes: Copula domain samples are obtained by mapping each attribute dimension of the initial population sample data using the empirical cumulative distribution function.

[0032] Furthermore, in one embodiment, the initial population sample includes discrete, unordered variables. Before transforming the initial population sample from the attribute domain to the copula domain to obtain the copula domain sample, the method further includes: Map the values ​​of discrete, unordered variables to a continuous sequence of integers.

[0033] Specifically, in order to capture the correlation patterns between attribute variables, the embodiment maps the initial population sample to the copula domain using an empirical cumulative distribution function.

[0034] In this process, the initial individual population sample is denoted as... , That is, the total number of samples is... Individuals, where each person's characteristics are represented by vectors This represents the multiple attributes of an individual, recorded as a total of Attributes, such as gender, age, and educational background, can be used for modeling heterogeneous agents. Let each population feature be a random vector. It is a random vector One implementation of this is sampling.

[0035] In the training set The above embodiment calculates the empirical marginal distribution of each dimension. , That is, to count the frequency of each attribute value in each dimension, and thus obtain the cumulative distribution of experience in each dimension. , .

[0036] Furthermore, for discrete and unordered variables present in the initial data, such as gender and commuting method, these discrete values ​​cannot be sorted, so the empirical cumulative distribution cannot be directly calculated. To address this, the embodiment maps their values ​​to ordered integers starting from 0 before calculating the cumulative distribution function.

[0037] It should be noted that for discrete, unordered variables, the performance of training and generation is independent of the mapping order of variable values. For example, for the gender attribute, the training and generation results are stable under both mapping methods: mapping "male" to 0 and "female" to 1 and mapping "male" to 1 and "female" to 0. Therefore, the implementation examples do not limit the mapping order.

[0038] Then, based on the calculated empirical cumulative distribution functions for each dimension, the attribute domain samples can be transformed into the copula domain. Let the individual population data be... The data representation of its corresponding copula field is as follows ,in , .

[0039] In this embodiment, by transforming to the copula domain, the correlation patterns between attribute variables, such as tail dependence and nonlinear correlation, can be effectively captured without being affected by marginal distributions, thus characterizing the essential dependency constraints between variables. This solves the technical problem in existing technologies that fail to consider the discrete characteristics of attribute values ​​in each dimension, making it impossible to capture the essential dependency constraints between variables and resulting in poor feasibility and diversity of the generated population data.

[0040] Furthermore, in one embodiment, the vectorized encoding representation of the copula domain samples to obtain matrix samples includes: The samples in the copula domain are encoded using sine and cosine vectors to obtain matrix samples.

[0041] Specifically, to ensure that the input data meets the input requirements of the diffusion model, the embodiment uses sine and cosine coding to vectorize the copula domain samples. After vector coding, each individual can be represented as a matrix. In the neural network, a multi-head attention module can be used to fully explore the correlations between the attribute values ​​represented as matrix rows, thereby learning the relationships between different dimensions.

[0042] Specifically, for attribute values , Its sine and cosine vector encoding is:

[0043]

[0044] in, Indicates the first The attribute values ​​of each dimension, The length of the encoded vector. , For the first A frequency basis in multiple dimensions. In this way, attribute values... Encoded as a length of The vector. In the interval The value is taken from the middle. In sine and cosine vector encoding... Multiplying by 100 to expand the range of values ​​allows for the dispersion of different values ​​within the same dimension through mapping, making the vector representation more dissimilar. However, it should be understood that the multiplier of 100 for expanding the range of values ​​is only one setting in this embodiment; other values ​​may be used in other embodiments. In this embodiment, the frequency base is a prime number near a whole number of 100, such as {101, 199, 307, 401, 503, 601, 701, 809, 907} for a population sample containing 9 population attributes. Setting different frequency bases and different multipliers is within the scope of protection of this application.

[0045] After the above sine and cosine encoding, each attribute value Represented as a row in a matrix, individual Represented as a The matrix is ​​used as input for training the diffusion model. .

[0046] In this embodiment, by using sine and cosine encoding, the copula domain samples that reflect the essential dependency and constraint relationships between variables can be transformed into a sample format suitable for the diffusion model. In this process, a multiplier is set to map and disperse attribute values ​​under the same dimension, making the representation vectors more dissimilar and improving the quality of the generated samples of the diffusion model.

[0047] Furthermore, in one embodiment, new matrix samples are obtained by training the diffusion model and generating samples based on matrix samples, including: Diffusion model trained based on matrix samples; New matrix samples are generated based on the trained diffusion model.

[0048] Furthermore, in one embodiment, the diffusion model is constructed based on a denoised probabilistic diffusion model, including a forward Markov chain and a reverse Markov chain.

[0049] Specifically, the diffusion model in this embodiment employs a denoising probabilistic diffusion model, comprising two Markov chains. The forward chain progressively adds noise to the data, while the backward chain learns how to recover the original data from the noise. The forward chain is typically designed to transform complex data distributions into simpler prior distributions, usually a standard Gaussian distribution. Conversely, the backward chain uses a deep neural network to learn a transfer kernel, thereby reversing the effects of the forward chain to capture complex dependencies and patterns in the original data.

[0050] Figure 2 This is a schematic diagram illustrating the diffusion model principle of an embodiment of this application. (Combined with...) Figure 2 Let's look at the given data distribution. The forward Markov chain process generates a series of random vectors. Its transfer nucleus is The most common choice for the transfer kernel is a Gaussian distribution, represented as:

[0051] Among them, hyperparameters It was determined before training. along with It increases and grows. With... growth, data sample Gradually losing its unique characteristics. When large enough, It approximates a Gaussian distribution.

[0052] Furthermore, through iterative equations and reparameterization techniques, it is possible to derive... Specific form:

[0053] in , .

[0054] The reverse process is also a Markov chain, which can be represented as a prior distribution. and a series of learnable transfer kernels Choose a prior distribution It is because the forward construction process makes Learnable transfer kernels have the following form:

[0055] Among them, the mean and variance Represented by a parameterized deep neural network, Represents the learnable model parameters. Generate data. The process can be achieved by first sampling the noise vector. , and then from arrive Gradually apply the learned transfer kernel This is achieved by generating a series of samples, specifically through ancestor sampling. This process is repeated until a result is obtained. .

[0056] The training process of the diffusion model aims to reverse the time of the forward Markov chain in the reverse Markov chain. During training, the learnable transfer kernel is adjusted. parameters To minimize the cross-entropy loss, it is defined as:

[0057] Using the reparameterization technique, a simplified form of the cross-entropy loss is:

[0058] in, Indicates in set A uniform distribution on the surface. With parameters. Deep neural networks based on and Predicted noise vector .pass The mean of the reverse transfer kernel can be calculated. ,as follows:

[0059] For the covariance of learnable transfer kernels It can be set up directly without learning. .

[0060] In the process of training the diffusion model based on matrix samples in the embodiment, for the constructed learnable noise prediction network The network input is and , Output For the corresponding Cumulative noise vector at time step An approximation. See also Figure 3 , Figure 3 This is a network structure diagram of the noise prediction network according to an embodiment of this application. The part enclosed by the dashed box represents the structure of the noise prediction network. Samples from the copula domain... The matrix is ​​obtained after sine and cosine encoding. Determine the noise-adding step Add noise Later obtained The output after inputting into the prediction network ,calculate and The distance between them is considered as the loss. Figure 3 middle This indicates the batch size; population data is... Dimensional, in the previous steps correspond, This represents the length of the sine and cosine vector encoding, compared to the previous steps. The corresponding [correspondence]. The example embodiment sets... .

[0061] The example uses a trained diffusion model to generate new matrix samples. It should be noted that the data generated at this point is... The matrix corresponds to the training input of the diffusion model, and further decoding algorithms are needed to decode it into the copula domain. The sample is then transformed from the copula domain back to the attribute domain to obtain the final generated population data.

[0062] In this embodiment, a diffusion model is trained based on initial population data, and new population data is generated using the trained diffusion model, which effectively expands the population size and allows for the acquisition of diverse population data based on a limited initial sample.

[0063] Furthermore, in one embodiment, decoding the new matrix sample to obtain a copula domain vector sample includes: The new matrix sample is decoded by multi-dimensional weighted optimization to obtain the copula domain vector sample.

[0064] Specifically, during the decoding process of the new matrix sample, considering that the generation process is a random process and each step of the generation algorithm contains random noise, the generated matrix data is noisy. Denote the generated vector corresponding to the attribute value as . Then will not be strictly equal to a certain . Therefore, the embodiment adopts a weighted optimization method for robust decoding. The formula is expressed as:

[0065] where , , , .

[0066] Dividing by 100 in the formula is corresponding to the being amplified by 100 times during sine-cosine encoding. is the loss term corresponding to the th frequency component, and is the corresponding weight (higher weight for low-frequency dimensions).

[0067] In this embodiment, by decoding the matrix sample to the copula domain through a multi-dimensional weighted optimization method, the robustness of the decoding process can be ensured, and the decoding deviation caused by noise can be reduced.

[0068] Furthermore, in one embodiment, mapping the copula domain vector sample through the quantile function to obtain the attribute domain population data generation sample includes: Mapping the copula domain vector sample on each dimension using the empirical quantile function to obtain the attribute domain population data generation sample. [[ID=P41]]

[0069] Specifically, after decoding to obtain the sample in the copula domain, theoretically, it can be mapped to the data in the attribute domain based on the empirical quantile function (i.e., the empirical inverse cumulative distribution function) of each attribute dimension. However, on the one hand, there is a deviation between the empirical cumulative distribution function and the true cumulative distribution function, and on the other hand, the samples generated by the diffusion model are noisy, and errors will also be introduced during the decoding process. Strictly calculating the inverse according to the empirical inverse cumulative distribution function will lead to mapping errors, and a small deviation will cause mapping errors. Taking the gender dimension as an example, assume the empirical probability distribution of the gender variable is , where 0 represents female and 1 represents male. The empirical cumulative distribution function of the gender variable is then , . If the value in the copula domain of the gender dimension obtained during decoding, then direct mapping will have In fact, the gender dimension decoding yields... It will cluster at 0.4921 and 1.0 ( The area near the point in the value range may contain noise and errors due to the sample generation and decoding process, and may also be a real point. The value is 0.49211. Therefore, for The mapping should yield 0s, not 1s, in the attribute domain. Since the population attributes being considered are discrete variables (continuous variables such as age have been discretized into intervals), the decoded values... The empirical cumulative distribution function for the corresponding dimension is denoted as ,turn up In the range (finite set) The closest point, as the actual decoded value For example, for the gender variable in the previous example, we can use 0.4921 instead. As the actual decoded value, it is then mapped to the sample of the attribute domain. .

[0070] This completes the generation of population data based on the copula domain diffusion model.

[0071] In this embodiment, by mapping copula domain samples to the attribute domain using nearest neighbor approximation and based on the empirical inverse cumulative distribution function, mapping errors caused by noise interference during sample generation and decoding process errors can be avoided.

[0072] Secondly, embodiments of this application also provide a population data generation apparatus.

[0073] In one embodiment, reference is made to Figure 4 , Figure 4 This is a schematic diagram of the functional modules of an embodiment of the population data generation device of this application. Figure 4 As shown, the population data generation device includes: The sample conversion unit 401 is used to convert the initial population sample from the attribute domain to the copula domain to obtain the copula domain sample, and to vectorize the copula domain sample to obtain the matrix sample. The sample diffusion generation unit 402 is used to train the diffusion model based on matrix samples and generate new matrix samples. The decoding output unit 403 is used to decode the new matrix sample to obtain the copula domain vector sample, and to perform quantile function mapping on the copula domain vector sample to obtain the attribute domain population data generation sample.

[0074] The functions of each module in the aforementioned population data generation device correspond to the steps in the aforementioned population data generation method embodiment, and their functions and implementation processes will not be described in detail here.

[0075] Thirdly, embodiments of this application also provide a readable storage medium.

[0076] The present application has a population data generation program stored on a readable storage medium, wherein when the population data generation program is executed by a processor, it implements the steps of the population data generation method described above.

[0077] The method implemented when the population data generation program is executed can be referred to in various embodiments of the population data generation method of this application, and will not be repeated here.

[0078] It should be noted that the sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0079] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device to execute the methods of the various embodiments of this application.

[0080] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for generating population data, characterized in that, include: The initial population sample is transformed from the attribute domain to the copula domain to obtain the copula domain sample, and the copula domain sample is vectorized and encoded to obtain the matrix sample. New matrix samples are obtained by training a diffusion model and generating samples based on the matrix samples. The new matrix sample is decoded to obtain a copula domain vector sample, and the copula domain vector sample is mapped by a quantile function to obtain an attribute domain population data generation sample.

2. The population data generation method according to claim 1, characterized in that, The process of transforming the initial population sample from the attribute domain to the copula domain to obtain the copula domain sample includes: Copula domain samples are obtained by mapping each attribute dimension of the initial population sample data using the empirical cumulative distribution function.

3. The population data generation method according to claim 1, characterized in that, The initial population sample includes discrete, unordered variables. Before transforming the initial population sample from the attribute domain to the copula domain to obtain the copula domain sample, the process further includes: The discrete, unordered variable values ​​are mapped to a continuous integer sequence.

4. The population data generation method according to claim 1, characterized in that, The process of vectorizing and encoding the copula domain samples to obtain matrix samples includes: The copula domain samples are encoded using sine and cosine vectors to obtain matrix samples.

5. The population data generation method according to claim 1, characterized in that, The process of training the diffusion model and generating new matrix samples based on the matrix samples includes: The diffusion model is trained based on the matrix samples; New matrix samples are generated based on the trained diffusion model.

6. The population data generation method according to claim 1, characterized in that, The diffusion model is constructed based on a denoised probabilistic diffusion model, which includes a forward Markov chain and a reverse Markov chain.

7. The population data generation method according to claim 1, characterized in that, Decoding the new matrix sample to obtain the copula domain vector sample includes: The new matrix sample is then subjected to multi-dimensional weighted optimization decoding to obtain the copula domain vector sample.

8. The population data generation method according to claim 1, characterized in that, The step of generating population data samples by performing quantile function mapping on the copula domain vector samples includes: The copula domain vector samples are mapped along each dimension using an empirical quantile function to generate population data samples for the attribute domain.

9. A population data generation device, characterized in that, include: The sample transformation unit is used to transform the initial population sample from the attribute domain to the copula domain to obtain a copula domain sample, and to vectorize and encode the copula domain sample to obtain a matrix sample. The sample diffusion generation unit is used to train the diffusion model and generate new matrix samples based on the matrix samples; The decoding output unit is used to decode the new matrix sample to obtain a copula domain vector sample, and to perform quantile function mapping on the copula domain vector sample to obtain an attribute domain population data generation sample.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a population data generation program, wherein when the population data generation program is executed by a processor, it implements the steps of the population data generation method as described in any one of claims 1 to 8.

Images