Financial data controllable generation method and system based on stochastic control generative adversarial

Abstract

This disclosure provides a method and system for controllable generation of financial data based on stochastic control generative adversarial techniques, relating to the fields of anomaly detection and data generation. The method includes: acquiring financial voucher data from an enterprise ERP system and encoding it as a feature vector as the target distribution of the terminal state in a stochastic control evolution process; defining interpretable control signals; constructing an adversarial network model for financial data generation; and modeling the generation process from latent space vectors to terminal financial data vectors as a generation path trajectory of a dynamic system described by stochastic differential equations, ultimately generating output financial table data. This disclosure introduces stochastic optimal control techniques, modeling the generation process as an evolutionary system of stochastic differential equations, achieving refined, continuous, and interpretable semantic control of the generation path. This method, to a certain extent, solves the problems of instability and pattern collapse in traditional GAN ​​training, significantly improving the diversity of generated samples and the reliability of training convergence.

Description

Technical Field

[0001] This disclosure relates to the field of anomaly detection and data generation technology, specifically to a method and system for controllable generation of financial data based on random control generative adversarial mechanisms. Background Technology

[0002] The statements in this section are merely background information relating to this disclosure and do not necessarily constitute prior art.

[0003] With the deepening of enterprise digital transformation, Business AI has become a core driving force for enterprise-level application software such as Enterprise Resource Planning (ERP), Customer Relationship Management (CRM), and Business Intelligence (BI). In typical application scenarios of Business AI, such as intelligent auditing, risk control, sales forecasting, and supply chain optimization, deep generative models play a crucial role. For example, when training an enterprise's abnormal transaction detection model, a serious imbalance between positive and negative samples is encountered; real abnormal samples (such as financial fraud records and logs of violations) are extremely scarce and difficult to obtain. Simultaneously, due to data security and privacy compliance considerations, raw data containing sensitive business information cannot be directly shared between different departments within an enterprise or between the enterprise and external institutions. Against this backdrop, there is an urgent need for a data augmentation and privacy protection technology that can precisely control the generated attributes according to business semantics while ensuring the statistical authenticity of the generated data and consistency with business logic.

[0004] Generative adversarial networks (GANs), as powerful deep generative models, have achieved significant success in areas such as image, speech, and video synthesis. However, existing GANs and their variants have obvious limitations in addressing the needs of commercial AI: (i) Traditional GAN ​​training is unstable and prone to pattern collapse, resulting in insufficient diversity of generated enterprise operation data (such as financial vouchers and inventory records), making it difficult to cover rare business scenarios with long-tail distribution.

[0005] (ii) Controllable generation schemes such as conditional GAN ​​or styleGAN, although they allow the generation results to be affected by category labels or style codes, their control mechanisms are mostly discrete and black box-like. Business analysts cannot intervene in the specific business attributes of the generated samples in an intuitive, continuous and interpretable way. For example, they cannot continuously adjust the degree of abnormality of a transaction from fully compliant to highly suspicious, or control the amount and industry characteristics of the generated financial records. Summary of the Invention

[0006] To address the aforementioned issues, this disclosure proposes a controllable generation method and system for financial data based on stochastic control generative adversarial systems. By introducing stochastic optimal control technology, the data point generation process of generative adversarial networks is modeled as the evolution trajectory of a dynamic system described by stochastic differential equations. By solving the stochastic optimal control problem, the generation path is achieved with fine, continuous, and interpretable semantic control, significantly improving the diversity of generated samples and the reliability of training convergence.

[0007] According to some embodiments, the present disclosure adopts the following technical solutions: A method for controllable generation of financial data based on stochastic control generative adversarial approaches includes: Acquire financial voucher data from the enterprise's ERP system and encode it into a feature vector as the target distribution of the terminal state in the stochastic control evolution process; Define interpretable control signals, construct an adversarial network model for generating financial data, and model the process of generating financial data vectors from latent space vectors to terminal financial data vectors as a generation path trajectory of a dynamic system described by stochastic differential equations, ultimately generating financial table data in adversarial form. The adversarial network model includes a controllable generator and a path space discriminator. The controllable generator of the adversarial network model is designed as a stochastic controller. The controllable generator receives the current state, time and interpretable control signals, and drives the initial random noise state to the target data distribution area by solving the stochastic optimal control problem. The path space discriminator is used to evaluate the authenticity of the finally generated data points.

[0008] Furthermore, the step of acquiring financial voucher data from the enterprise ERP system and encoding it into a feature vector as the target distribution of the terminal state in the stochastic control evolution process includes: Extract financial voucher data from the enterprise ERP system. The mixed-type fields for each record include: debit amount, credit amount, accounting code, voucher date, supplier ID, and department code. All fields are fused into a fixed-dimensional feature vector by an encoder, which serves as the target distribution of the terminal state in the stochastic control evolution process.

[0009] Furthermore, the definition of interpretable control signals is defined as structured service parameter vectors, including: Global controls include anomaly level, company size, and industry category; local controls include monetary scale factor and time distribution offset. Among them, the anomaly level is divided into continuous values ​​from 0 to 1, where 0 represents completely normal and 1 represents highly abnormal. Intermediate values ​​can generate suspicious records with blurred edges. The amount scale factor controls the order of magnitude of the generated transaction amount, and the time distribution offset controls the seasonal offset of the voucher date.

[0010] Furthermore, the adversarial network model for generating financial data is constructed, comprising a controllable generator and a path space discriminator. The adversarial network model is trained using a two-stage training method, specifically: First, in the pre-training phase, the stochastic differential equation is pre-trained using a large number of normal financial records reconstruction loss, so that it learns the basic topological structure of the financial feature space. In the adversarial training phase, a path space discriminator and interpretable control signals are introduced to conduct adversarial training using real anomaly samples and generated samples. A batch of data is sampled from the real data distribution, and control signals are sampled from the prior distribution of control signals. A structured initial state is generated through an invertible neural network. Using the parameters of the current controllable generator, the generated path and its terminal state are obtained by numerically solving stochastic differential equations. The generated path is discretely sampled to obtain the generated path sequence and its statistical characteristics are calculated. The real data, terminal state, generated path sequence and its statistical characteristics are input into the path space discriminator to calculate the adversarial loss. The regularization loss is calculated according to the regularization method. Based on the training objective, the gradients of the generator and discriminator parameters are calculated using the adjoint sensitivity method of stochastic differential equations. The parameters are then updated using an optimization algorithm to complete the training.

[0011] Furthermore, the training objective is constructed as a stochastic optimal control problem, specifically:

[0012] in, and These represent the path distributions of real data and generated data, respectively. It is a proposed regularization term, whose function is to make the control strategy learned by the generator physically or semantically reasonable; It is the coefficient of the regularization term, which is used to adjust the influence of the regularization term during the training process; For real data, Terminal status. and For model parameters, The target distribution of terminal states, To control the prior distribution of the signal, For interpretable control signals, This is the initial distribution.

[0013] Furthermore, the process of generating the latent space vector into the terminal financial data vector is modeled as a generation path trajectory of a dynamic system described by stochastic differential equations, including: Let the distribution of real data be Learning a stochastic process This makes its terminal distribution close to The data point generation path is defined as a stochastic differential equation:

[0014] in, yes The state at time t is a high-dimensional feature vector; It is standard Brownian motion, representing random noise; the controllable generator of generative adversarial networks. parameters Used to parameterize drift coefficients and diffusion coefficient ; For deterministic thrust coefficients, The coefficients of the random disturbance; Indicates time encoding; This is the initial distribution.

[0015] Furthermore, the drift coefficient is parameterized by a deep neural network, with its inputs being state, time encoding, and control signals. The time encoding employs a multi-frequency sine and cosine function. The diffusion coefficient is designed as a state- and time-dependent diagonal matrix.

[0016] Furthermore, the controllable generator is designed as a stochastic controller that receives the current state, time, and an interpretable control signal vector, and outputs a drift term for dynamically adjusting the stochastic control. By changing the control signal, the controllable generator guides the dynamic system to different attractor regions, thereby corresponding to different patterns of data distribution, and thus achieving continuous, smooth, and interpretable sample control. The control signals are designed with a hierarchical structure:

[0017] in, It is a global control signal that controls the overall properties of the sample; It serves as a local control signal, controlling the properties of different time scales or spatial regions.

[0018] According to some embodiments, the present disclosure adopts the following technical solutions: A controllable financial data generation system based on stochastic control generative adversarial mechanisms includes: The data acquisition and encoding module is used to acquire financial voucher data from the enterprise's ERP system and encode it into feature vectors as the target distribution of the terminal state in the stochastic control evolution process. The stochastic control adversarial generation module is used to define interpretable control signals, construct an adversarial network model for generating financial data, and model the process of generating latent space vectors into terminal financial data vectors as a generation path trajectory of a dynamic system described by stochastic differential equations. Finally, the adversarial generation outputs financial table data. The adversarial network model includes a controllable generator and a path space discriminator. The controllable generator of the adversarial network model is designed as a stochastic controller. The controllable generator receives the current state, time and interpretable control signals, and drives the initial random noise state to the target data distribution area by solving the stochastic optimal control problem. The path space discriminator is used to evaluate the authenticity of the finally generated data points.

[0019] According to some embodiments, the present disclosure adopts the following technical solutions: An electronic device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to perform the aforementioned method for controllable generation of financial data based on random control generative adversarial mechanisms.

[0020] Compared with the prior art, the beneficial effects of this disclosure are as follows: This disclosed method for controllable generation of financial data based on stochastic control generative adversarial methods introduces stochastic optimal control techniques to model the generation process of data points as the evolutionary trajectory of a dynamic system described by stochastic differential equations, achieving fine, continuous, and interpretable semantic control of the generation path. In this process, the generator is no longer a simple feedforward neural network but acts as a "stochastic controller," driving the initial random noise state to the target data distribution region by solving a stochastic optimal control problem. The discriminator then evaluates the authenticity of the final state (i.e., the final generated data points). This method, to a certain extent, solves the problems of training instability and pattern collapse in traditional GANs, significantly improving the diversity of generated samples and the reliability of training convergence.

[0021] This disclosed method for controllable generation of financial data based on stochastic control generative adversarial approaches addresses the pain point of scarce abnormal samples in enterprises (typically less than 1% of the total sample). By setting the anomaly level of the control signal to a high value range (e.g., 0.7-1.0), it can generate high-quality, diverse simulated abnormal financial voucher data in batches. By mixing the generated data with real data and retraining anomaly detection models such as graph neural networks, the model recall rate can be improved, enhancing the enterprise's ability to identify hidden financial risks. Furthermore, by adjusting the control signal through an interactive interface, the characteristics of financial data under different fraudulent methods (such as inflating revenue or concealing expenses) can be simulated, and the reaction of the current risk control model can be observed, thereby identifying rule loopholes in advance and optimizing audit strategies.

[0022] The disclosed method for controllable generation of financial data based on randomized generative adversarial mechanisms poses compliance risks when directly sharing real financial data during cross-terminal interactions. Using the controlled generator of this disclosure to generate synthetic financial data to replace real data for shared analysis preserves key statistical characteristics and business patterns while avoiding the leakage of original sensitive information, thus meeting compliance requirements. Attached Figure Description

[0023] The accompanying drawings, which form part of this disclosure, are used to provide a further understanding of this disclosure. The illustrative embodiments of this disclosure and their descriptions are used to explain this disclosure and do not constitute an undue limitation of this disclosure.

[0024] Figure 1 This is a flowchart illustrating the training process of a controllable financial data generation method based on stochastic control generative adversarial methods according to an embodiment of this disclosure. Figure 2 A flowchart illustrating the training process of an adversarial network model generated from financial data in an embodiment of this disclosure; Figure 3 This is a system architecture diagram of a controllable financial data generation system based on random control generative adversarial methods, according to an embodiment of this disclosure. Detailed Implementation

[0025] The present disclosure will be further described below with reference to the accompanying drawings and embodiments.

[0026] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of this disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure pertains.

[0027] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this disclosure. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms “comprising” and / or “including” are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0028] Example 1 One embodiment of this disclosure provides a method for controllable generation of financial data based on stochastic control generative adversarial methods, the method steps of which include: Step 1: Obtain the financial voucher data from the enterprise's ERP system and encode it into a feature vector as the target distribution of the terminal state in the stochastic control evolution process; Step 2: Define interpretable control signals, construct an adversarial network model for generating financial data, and model the process of generating financial data vectors from latent space vectors to terminal financial data vectors as a generation path trajectory of a dynamic system described by stochastic differential equations, ultimately generating output financial table data. The adversarial network model includes a controllable generator and a path space discriminator. The controllable generator of the adversarial network model is designed as a stochastic controller. The controllable generator receives the current state, time and interpretable control signals, and drives the initial random noise state to the target data distribution area by solving the stochastic optimal control problem. The path space discriminator is used to evaluate the authenticity of the finally generated data points.

[0029] As one example, training financial anomaly detection models (such as those identifying fake invoices, abnormal expense reports, and misappropriation of funds) in Enterprise Resource Planning (ERP) and Business Intelligence (BI) systems faces a severe imbalance between positive and negative samples. Real anomaly samples (such as records of financial fraud) are extremely scarce and difficult to obtain. Furthermore, due to data security and privacy compliance considerations, companies cannot share original sensitive financial data across departments or enterprises. Traditional tabular data generation methods based on statistical distribution fitting or variational autoencoders often suffer from schema collapse or inconsistencies between the generated data and actual business logic (such as accounting identity constraints) when generating high-dimensional mixed-type data (including continuous values ​​such as amounts, discrete categories such as account codes, and timestamps).

[0030] To address the aforementioned issues with corporate financial data, this paper utilizes a controllable financial data generation method based on stochastic control generative adversarial network (GAN) to construct a controllable financial table data generator. A novel GAN ​​framework based on stochastic control technology—Stochastic Control Generative Adversarial Network (GAN)—is proposed. Its core idea is to model the data point generation process as the evolutionary trajectory of a dynamic system described by stochastic differential equations. In this process, the generator is no longer a simple feedforward neural network but acts as a "stochastic controller," driving the initial random noise state to the target data distribution region by solving a stochastic optimal control problem. The discriminator then evaluates the authenticity of the final state (i.e., the finally generated data points). The specific implementation process is as follows: Step 1: Obtain the financial voucher data from the enterprise's ERP system and encode it into a feature vector as the target distribution of the terminal state in the stochastic control evolution process.

[0031] Extract real financial voucher data from the enterprise ERP system. Each record contains the following mixed-type fields: debit amount (continuous value), credit amount (continuous value), accounting subject code (discrete category), voucher date (time), supplier ID (high base category), department code (discrete), etc.

[0032] The encoder fuses all fields of the financial voucher data into a fixed-dimensional feature vector. This vector represents the final state of the evolution process of the stochastic differential equation (SDE). Target distribution .

[0033] As one example, a stochastic dynamic system is first designed, and the data generation process is reformulated as a terminal-time state distribution matching problem of a controlled stochastic differential equation. Let the real data target distribution be... This disclosure aims to learn a stochastic process. This makes its terminal distribution close to .

[0034] Specifically, this disclosure defines the data point generation path as a stochastic differential equation (SDE):

[0035] in, yes The state at time t can be viewed as a high-dimensional feature vector; It is standard Brownian motion, representing random noise; the controllable generator of generative adversarial networks. parameters Used to parameterize drift coefficients and diffusion coefficient ; It can be considered as a deterministic thrust coefficient. It can be regarded as the coefficient of random perturbation; Indicates time encoding; This represents the initial state distribution.

[0036] Furthermore, the drift coefficient Parameterized by a deep neural network, whose input is the current state. Time coding and interpretable control signals Time coding employs multi-frequency sine and cosine functions: , Among them, frequency Distributed in a geometric progression to enable time coding The control signals cover different time scales, from low to high frequencies. By incorporating hierarchical encoding and attention mechanisms, the core architecture includes a feature fusion module and residual block stacking, and finally generates a drift vector through the output projection layer.

[0037] Furthermore, the diffusion coefficient Designed as a diagonal matrix form that depends on both state and time: , Among them, each diagonal element To ensure positive definiteness, the neural network performs independent computation, employing a U-Net encoder-decoder architecture with skip connections. Specifically, the U-Net encoder consists of four downsampling stages, each composed of two 3×3 convolutional layers, a group normalization layer, and a SiLU activation function; the bottleneck layer captures contextual features at the lowest resolution. The decoder consists of four upsampling stages, restoring dimensionality through transposed convolutions and performing skip connections with the feature maps of the corresponding stages in the encoder via channel concatenation to achieve multi-scale information fusion. The network's final output layer uses a 1×1 convolutional layer to map the number of channels to... d The output is then activated by the Softplus function to ensure that it is positive, thus forming a diagonal element vector.

[0038] It should be noted that the above architecture based on convolution and U-Net is particularly suitable for images or data with spatial structure. When the input data is tabular data, sequence data, or graph structure data, the above convolutional layers can be replaced with fully connected layers, self-attention mechanism layers, or graph convolutional layers.

[0039] Furthermore, to avoid mode collapse, the initial state distribution... Instead of using the traditional standard Gaussian method, a reversible neural network is used. generate: , in, Using a normalized flow architecture (such as RealNVP or Glow), it will standardize Gaussian random variables. According to the control signal The initial state is transformed into a structured state. This design ensures that different control signals correspond to different initial distributions and makes the mapping from the initial state to the data more direct, thereby shortening the evolution time of the SDE. T .

[0040] Step 2: Define interpretable control signals, construct an adversarial network model for generating financial data, and model the process of generating financial data vectors from latent space vectors to terminal financial data vectors as a generation path trajectory of a dynamic system described by stochastic differential equations, ultimately generating financial table data in adversarial form.

[0041] (1) First, define interpretable control signals.

[0042] To achieve precise and controllable data generation, control signals are defined based on the actual needs of financial analysts. The following is a structured vector of business parameters.

[0043] control signals It's not a simple vector concatenation, but rather designed as a hierarchical structure:

[0044] Specifically, global control signals The overall attributes of the control samples include anomaly level (a continuous value from 0 to 1, where 0 represents complete normal and 1 represents high anomaly; intermediate values ​​can generate suspicious records with blurred edges), enterprise size (micro / small / medium / large, one-hot encoded), and industry category (manufacturing / retail / service, etc.); local control signals. It controls attributes at different time scales or in different spatial regions, including monetary scale factor (controlling the order of magnitude of the generated transaction amount) and time distribution offset (controlling the seasonal offset of the voucher date).

[0045] Furthermore, each control component is first mapped to a key vector through an independent linear projection layer. Sum value vector Meanwhile, the current state With time encoding The fused features are mapped to a query vector through another linear layer. Subsequently, query vector With all key vectors Perform scaled dot product attention calculations to obtain attention weights, and apply these weights to all value vectors. The weighted summation is performed to output the fused control context vector, which is then added to the fused feature and processed through subsequent residual blocks.

[0046] Furthermore, by continuously adjusting the anomaly level component, financial records in any transitional state between full compliance and obvious violation can be generated for training or evaluating the sensitivity and robustness of the anomaly detection model. This continuous control capability is not available in traditional conditional GANs based on discrete labels (abnormal / normal).

[0047] Furthermore, to bridge the gap between static financial data records and continuous-time stochastic differential equation models, this disclosure metaphorically represents the generation process from latent space vectors to final financial data vectors as a portfolio evolution process under a risk-neutral measure. Specifically, although the final generated financial record is itself a static snapshot at a time cross-section, the numerical recursive steps of the SDE solver in the latent space can be viewed as a "virtual evolution time" of the firm's financial situation on the characteristic manifold. From this perspective, the drift term of the SDE... This corresponds to the company's operational inertia and industry benchmark trends (e.g., pushing account balances from a noisy state to a compliant debt-to-equity balance range); while the diffusion coefficient This corresponds to the unavoidable random fluctuations and abnormal fraud noise that need to be captured during the operation of an enterprise. In this way, the SDE framework, originally used for physical process modeling, is naturally migrated to the task of generating tabular data, so that the generation path not only satisfies the statistical similarity of data distribution, but also has interpretable financial semantics at the level of evolutionary trajectory.

[0048] (2) Construct an adversarial network model for generating financial data.

[0049] Specifically, the adversarial network model includes a controllable generator and a pathspace discriminator; Among them, the controllable generator of generative adversarial networks Designed as a random controller, it receives the current state. ,time And an interpretable control signal Its output is used to dynamically adjust the drift term of the SDE, that is:

[0050] By changing the control signal c This controllable generator can guide a dynamic system toward different attractor regions, thereby corresponding to different patterns of data distribution, and thus achieving continuous, smooth, and interpretable sample control.

[0051] Path space discriminator Not only assess the final state It also evaluates the entire state path. It is part of a larger system. It is trained to distinguish between "likelihood paths derived from real data" and "paths simulated forward by a controllable generator." This draws on path space metric theory in stochastic control, which can provide richer training signals.

[0052] Furthermore, the path space discriminator Compared to only assessing terminal status Unlike traditional methods, this disclosed pathspace discriminator A multi-scale evaluation mechanism is adopted, receiving four types of inputs: 1) Original terminal status ; 2) Discretized path sequence ,in ; 3) First-order difference features of the path ; 4) Second-order statistical characteristics of multiple generation paths within a batch, including each discrete time step. The trace of the batch variance or covariance matrix of the state vector.

[0053] The above four types of input information together constitute a complete characterization of the generation path of stochastic differential equations. Among them, the first-order difference features are used to characterize the local velocity and smoothness of the latent space evolution, reflecting the temporal consistency within a single path; the second-order statistical features are used to monitor the changes in pattern diversity during the generation process, revealing whether the generator has experienced pattern collapse or premature convergence at the batch level.

[0054] Specifically, during the model inference or deployment phase, when the batch size is 1, the second-order statistical features can be replaced by the moving average variance of all generated samples at each discrete time point pre-calculated during the training phase, or by generating virtual neighbor samples to calculate local statistical features by applying small perturbations to the latent variables within the current batch.

[0055] Furthermore, the path space discriminator It adopts a multi-branch coding and layered fusion architecture, the core of which contains four parallel branches: 1) Terminal status encoder processing Generate features ; 2) Path timing encoder processes discrete path sequences Generate features ; 3) First-order difference sequence of the differential feature encoder processing path Generate features ; 4) The statistical feature encoder processes the second-order statistics of the state at each time step within a batch to generate features. .

[0056] The feature vectors of the four parallel branches are interactively fused through a cross-attention mechanism to obtain the fused features. .

[0057] Subsequently, the fusion features The original features of each branch , , , By splicing or through a gating mechanism, each independent discriminant output head (e.g., a single-layer fully connected network with Sigmoid activation) is mapped to a corresponding scalar truth probability. , , , .

[0058] The final discriminator outputs a weighted sum:

[0059] in, , , , For learnable or fixed weight coefficients, satisfying .

[0060] (3) Training objective design based on random control.

[0061] Specifically, the training objective is constructed as a stochastic optimal control problem:

[0062] in, and These represent the path distributions of real data and generated data, respectively. It is a proposed regularization term, whose function is to make the control strategy learned by the generator physically or semantically reasonable; It is the coefficient of the regularization term, which can be used to adjust the influence of the regularization term during the training process. For real data, Terminal status. and For model parameters, The target distribution of terminal states, To control the prior distribution of the signal, For interpretable control signals, This is the initial distribution.

[0063] Furthermore, regularization terms You can choose one or a combination of the following definition methods depending on the specific application scenario: 1) Based on the concept of optimal control, a "least effort" control strategy can be used, that is:

[0064] Its function is to encourage control strategies to minimize the above energy function, making the generation process more efficient and smoother.

[0065] 2) The probability density evolution of the generation process can be directly constrained. Let... yes If the probability density is such that it satisfies the Fokker-Planck equation: .

[0066] Therefore, regularization terms can be constructed to encourage Covering all real data, that is: , in, represent The divergence function has two terms: the first term ensures that the terminal distribution matches the true distribution, and the second integral term maximizes the entropy along the path to prevent premature distribution collapse.

[0067] 3) Furthermore, for applications such as scientific computing, domain-specific physical constraints can be introduced. For example, if generating fluid dynamics data, residual terms from the Navier-Stokes equations can be added, i.e.: , This regularization ensures that the generated data is not only statistically accurate but also physically plausible. Correspondingly, when generating financial, textual, or other non-physical data, this regularization term can be replaced with logical consistency constraints specific to the domain.

[0068] As one embodiment, this disclosure pertains to the generation process of enterprise financial data. Therefore, this disclosure incorporates regularization design based on accounting constraints. To overcome the problem of traditional table generation models generating data that violates basic accounting principles (such as the requirement that debits and credits must equal each other), this embodiment incorporates regularization items... The document specifically introduces domain-specific accounting identity constraints, including: Let the terminal generate vector After decoding and restoring the original fields, the sum of the debit amounts is The sum of the credit amounts is Then add the following constraint loss: , in, For indicator functions, This is the penalty weight. This regularization term works in conjunction with the minimum effort control regularization term or the Fokker-Planck equation regularization term to ensure that the generated data both statistically approximates the real data and strictly satisfies the basic accounting logic.

[0069] Furthermore, to improve discrimination ability and training stability, it is also possible to optionally... An additional slice Wasserstein distance based on optimal transmission theory is introduced as an auxiliary adversarial loss term. This criterion calculates the difference between real path samples and generated path samples in random directions. θp-Wasserstein distance of the upper projection distribution:

[0070] in, It is the p-Wasserstein distance. and Represents distribution and The forward distribution under projection. This term, together with the aforementioned adversarial loss, optimizes the controllable generator, enabling it to provide meaningful gradients even when the distribution support sets do not overlap, further mitigating the mode collapse problem.

[0071] (4) Achieve overall training of the adversarial network model.

[0072] Specifically, such as Figure 2 As shown, this disclosure employs a structured two-stage training strategy and an efficient numerical solution scheme to ensure that the generative adversarial network based on stochastic differential equations can converge stably and efficiently, specifically including: The first stage is the SDE dynamics pre-training stage. The main purpose of this stage is to enable the SDE to learn a basic, deterministic mapping, thereby avoiding the controllable generator in the early stages of adversarial games. It collapses due to insufficient capability. During this phase, the control signal is fixed at a neutral value (e.g., the zero vector). And temporarily disable the pathspace discriminator. This phase focuses on initializing the parameters of the SDE to capture the basic geometric structure of the data distribution. The training objective of this phase is to minimize a deterministic reconstruction loss: , in, It is to use real data samples Through reversible neural networks Mapping yields the initial state Then, it evolves to the terminal time via the currently parameterized SDE. T The result.

[0073] The second stage is the adversarial training and control learning stage, which aims to introduce game theory and gradually apply all the constraints defined in the theoretical framework. This stage introduces a pathspace discriminator. and variable control signals Perform full adversarial optimization and alternately update the pathspace discriminator. and controllable generator The parameters. During training, each iteration includes the following sub-steps: (1) From the distribution of real data A batch of data was sampled. ,in Batch size; (2) From the prior distribution of control signals Medium sampling control signal And through a reversible neural network Generate structured initial state ; (3) Use the current generator parameters The generation path is obtained by numerically solving the aforementioned SDE. and its terminal status The numerical solution can be obtained using numerical methods for solving stochastic differential equations, such as the Euler-Maruyama method. (4) Discretely sample the generated path to obtain the sequence. And calculate its statistical characteristics.

[0074] (5) Use real data Generate terminal status Generate path sequences and their statistical features as input to the discriminator. Calculate the countermeasure loss ; (6) Calculate the regularization loss based on the selected regularization method (such as least effort control, Fokker-Planck constraint, or physical constraint). ; (7) Based on the training objective of stochastic control, the gradients of the parameters of the controllable generator and the path space discriminator are calculated using the adjoint sensitivity method of stochastic differential equations, and the parameters are updated using optimization algorithms (such as Adam). and .

[0075] Furthermore, in the initial stage of the second phase of training, the adversarial loss can be primarily optimized to learn the data distribution; in the middle stage, the regularization term can be gradually enhanced. The weights are assigned to introduce constraints such as distribution coverage or energy efficiency; fine-tuning is then performed later.

[0076] As one example, the training process of the adversarial network model for generating financial data includes: First, during the pre-training phase, the reconstructed loss is obtained using a large number of normal financial records. The SDE is pre-trained to learn the basic topological structure of the financial feature space; then, during the adversarial training phase, a path space discriminator is introduced. Adversarial training was conducted using real anomaly samples (collected from historical audit cases) and generated samples. The discriminator not only judges the terminal vector... The authenticity of the data is further evaluated by comprehensively assessing: the first-order difference features of the generation path (reflecting whether the instantaneous rate of change of key fields such as the debit amount in the latent space evolution is smooth and natural), and the degree of dispersion of multiple generation paths within the same batch at the same evolution stage (i.e., second-order statistics, used to detect whether the diversity of the generated samples is sufficient). Through the above multi-dimensional path features, the discriminator can more accurately judge the rationality of the generation process. Finally, after training, the model is embedded as the core data generation engine into the intelligent audit or risk control module of the existing platform.

[0077] Example 2 One embodiment of this disclosure provides a controllable financial data generation system based on stochastic control generative adversarial mechanisms, comprising: The data acquisition and encoding module is used to acquire financial voucher data from the enterprise's ERP system and encode it into feature vectors as the target distribution of the terminal state in the stochastic control evolution process. The stochastic control adversarial generation module is used to define interpretable control signals, construct an adversarial network model for generating financial data, and model the process of generating latent space vectors into terminal financial data vectors as a generation path trajectory of a dynamic system described by stochastic differential equations, ultimately generating output financial table data. The adversarial network model includes a controllable generator and a path space discriminator. The controllable generator of the adversarial network model is designed as a stochastic controller. The controllable generator receives the current state, time and interpretable control signals, and drives the initial random noise state to the target data distribution area by solving the stochastic optimal control problem. The path space discriminator is used to evaluate the authenticity of the finally generated data points.

[0078] Example 3 One embodiment of this disclosure provides an electronic device, including: a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to perform the aforementioned method for controllable generation of financial data based on random control generative adversarial mechanisms.

[0079] This disclosure is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0080] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0081] While the specific embodiments of this disclosure have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of this disclosure. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of this disclosure are still within the scope of protection of this disclosure.

Claims

1. A method for controllable generation of financial data based on stochastic control generative adversarial, characterized in that, include: Acquire financial voucher data from the enterprise's ERP system and encode it into a feature vector as the target distribution of the terminal state in the stochastic control evolution process; Define interpretable control signals, construct an adversarial network model for generating financial data, and model the process of generating financial data vectors from latent space vectors to terminal financial data vectors as a generation path trajectory of a dynamic system described by stochastic differential equations, ultimately generating financial table data in adversarial form. The adversarial network model includes a controllable generator and a path space discriminator. The controllable generator of the adversarial network model is designed as a stochastic controller. The controllable generator receives the current state, time and interpretable control signals, and drives the initial random noise state to the target data distribution area by solving a stochastic optimal control problem. The path space discriminator is used to evaluate the authenticity of the finally generated data points.

2. The stochastic control generated adversarial financial data controllable generation method of claim 1, wherein, The step of acquiring financial voucher data from the enterprise ERP system and encoding it into a feature vector as the target distribution of the terminal state in the stochastic control evolution process includes: Extract financial voucher data from the enterprise ERP system. The mixed-type fields for each record include: debit amount, credit amount, accounting code, voucher date, supplier ID, and department code. All fields are fused into a fixed-dimensional feature vector by an encoder, which serves as the target distribution of the terminal state in the stochastic control evolution process.

3. The stochastic control generated adversarial financial data controllable generation method of claim 1, wherein, The definition of interpretable control signals is defined as structured service parameter vectors, including: Global controls include anomaly level, company size, and industry category; local controls include monetary scale factor and time distribution offset. Among them, the anomaly level is divided into continuous values ​​from 0 to 1, where 0 represents completely normal and 1 represents highly abnormal. Intermediate values ​​can generate suspicious records with blurred edges. The amount scale factor controls the order of magnitude of the generated transaction amount, and the time distribution offset controls the seasonal offset of the voucher date.

4. The stochastic control generated adversarial financial data controllable generation method of claim 1, wherein, The adversarial network model for generating financial data is constructed, comprising a controllable generator and a path space discriminator. The adversarial network model is trained using a two-stage training method, specifically: First, in the pre-training phase, the stochastic differential equation is pre-trained using a large number of normal financial records reconstruction loss, so that it learns the basic topological structure of the financial feature space. In the adversarial training phase, a path space discriminator and interpretable control signals are introduced to conduct adversarial training using real anomaly samples and generated samples. A batch of data is sampled from the real data distribution, and control signals are sampled from the prior distribution of control signals. A structured initial state is generated through an invertible neural network. Using the parameters of the current controllable generator, the generated path and its terminal state are obtained by numerically solving stochastic differential equations. The generated path is discretely sampled to obtain the generated path sequence and its statistical characteristics are calculated. The real data, terminal state, generated path sequence and its statistical characteristics are input into the path space discriminator to calculate the adversarial loss. The regularization loss is calculated according to the regularization method. Based on the training objective, the gradients of the generator and discriminator parameters are calculated using the adjoint sensitivity method of stochastic differential equations. The parameters are then updated using an optimization algorithm to complete the training.

5. The stochastic control generated adversarial financial data controllable generation method of claim 4, wherein, The training objective is constructed as a stochastic optimal control problem, specifically: in, and These represent the path distributions of real data and generated data, respectively. It is a proposed regularization term, whose function is to make the control strategy learned by the generator physically or semantically reasonable; It is the coefficient of the regularization term, which is used to adjust the influence of the regularization term during the training process; For real data, Terminal status. and For model parameters, The target distribution of terminal states, To control the prior distribution of the signal, For interpretable control signals, This is the initial distribution.

6. The method for controllable generation of financial data based on stochastic control generative adversarial methods as described in claim 1, characterized in that, The process of generating latent space vectors into terminal financial data vectors is modeled as a generation path trajectory of a dynamic system described by stochastic differential equations, including: Let the distribution of real data be Learning a stochastic process This makes its terminal distribution close to The data point generation path is defined as a stochastic differential equation: in, yes The state at time t is a high-dimensional feature vector; It is standard Brownian motion, representing random noise; the controllable generator of generative adversarial networks. parameters Used to parameterize drift coefficients and diffusion coefficient ; For deterministic thrust coefficients, The coefficients of the random disturbance; Indicates time encoding; This is the initial distribution.

7. The method for controllable generation of financial data based on stochastic control generative adversarial methods as described in claim 6, characterized in that, The drift coefficient is parameterized by a deep neural network, with inputs including state, time encoding, and control signals. The time encoding uses multi-frequency sine and cosine functions. The diffusion coefficient is designed as a state- and time-dependent diagonal matrix.

8. The method for controllable generation of financial data based on stochastic control generative adversarial methods as described in claim 1, characterized in that, The controllable generator is designed as a stochastic controller that receives the current state, time, and an interpretable control signal vector, and outputs a drift term for dynamically adjusting the stochastic control. By changing the control signal, the controllable generator guides the dynamic system to different attractor regions, thereby corresponding to different patterns of data distribution, and thus achieving continuous, smooth, and interpretable sample control. The control signals are designed with a hierarchical structure: in, It is a global control signal that controls the overall properties of the sample; It serves as a local control signal, controlling the properties of different time scales or spatial regions.

9. A controllable financial data generation system based on stochastic control generative adversarial mechanisms, characterized in that: include: The data acquisition and encoding module is used to acquire financial voucher data from the enterprise's ERP system and encode it into feature vectors as the target distribution of the terminal state in the stochastic control evolution process. The stochastic control adversarial generation module is used to define interpretable control signals, construct an adversarial network model for generating financial data, and model the process of generating latent space vectors into terminal financial data vectors as a generation path trajectory of a dynamic system described by stochastic differential equations. Finally, the adversarial generation outputs financial table data. The adversarial network model includes a controllable generator and a path space discriminator. The controllable generator of the adversarial network model is designed as a stochastic controller. The controllable generator receives the current state, time and interpretable control signals, and drives the initial random noise state to the target data distribution area by solving a stochastic optimal control problem. The path space discriminator is used to evaluate the authenticity of the finally generated data points.

10. An electronic device, characterized in that, include: The electronic device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to perform the controllable generation method for financial data based on random control generative adversarial as described in any one of claims 1-8.

Images