Tax risk intelligent identification and training system based on multi-source data fusion

Abstract

The application discloses a tax risk intelligent identification and training system based on multi-source data fusion. A multi-source data fusion and causal graph construction module standardizes multi-modal original data into a feature matrix and constructs a causal network topology structure according to tax rules; a neural symbolic causal decoupling module establishes a hybrid driven neural symbolic structure configuration model according to the structure, and outputs decoupled feature representation through an independence loss function; a core tax risk inference module uses a backdoor adjustment operator to remove statistical bias of interference indicators, generates baseline tax risk dynamic scores and risk causal paths; a game confrontation counterfactual generation engine calculates and optimizes counterfactual perturbations through maximum and minimum game alternation iteration, and generates logically self-consistent counterfactual tax-related characteristic data by using symbolic logic operators to perform algebraic rule calculation; and a dynamic interactive training module performs window rendering on double-track path cases, evaluates decision scores and collects out-of-distribution difficult samples to nourish the basic model.

Description

Technical Field

[0001] This invention relates to the fields of data processing and intelligent computer-aided teaching technology, specifically to a tax risk intelligent identification and training system based on multi-source data fusion. Background Technology

[0002] With the advancement of digital transformation in corporate taxation and finance, utilizing computer information systems to identify risks in multi-source tax-related data and build tax training case libraries has become a common practice in the fields of tax supervision and tax talent training. Currently, the industry typically employs graph analysis techniques or traditional deep neural networks to extract cross-modal data such as business registration information, invoices, and financial statements to establish risk prediction models. Static training materials are then generated based on historical real-world illegal cases to assist in conducting simulations and case reviews of tax audit operations.

[0003] However, existing data analysis and training systems reveal significant technical deficiencies in practical applications. First, because corporate tax filing is frequently influenced by various external environmental factors such as macroeconomic policy fluctuations, regional tax incentives, and industry cycles, traditional machine learning networks cannot effectively eliminate the influence of these external environmental interference signals in the data stream. This leads to statistical pseudo-correlation between features caused by confounding factors, preventing the system from cutting off heterogeneous signals and deeply decoupling underlying features. Second, existing deep learning networks are essentially black-box predictions; their floating-point outputs lack mathematical rigidity control and cannot embed rigid financial and tax laws and regulations and financial identities into the forward propagation process of the causal network. This results in the system's predictions being highly susceptible to numerical non-conservation and mathematical logical imbalances. Finally, existing training case libraries are entirely based on static transmission of historical data, lacking the dynamic adversarial evolution and self-iterative capabilities of cognitive models. Furthermore, the system cannot capture and process the judgments and decisions made by operators in unknown scenarios, resulting in a lack of closed-loop incremental and evolutionary feedback mechanisms. This makes the risk identification model less robust to anomaly identification in unknown scenarios. Summary of the Invention

[0004] To address the problems mentioned in the background, this invention provides a tax risk intelligent identification and training system based on multi-source data fusion. The multi-source data fusion and causal graph construction module acquires multimodal tax-related raw data and performs time-dimensional alignment and standardization on the data to output a standardized underlying feature matrix. Based on preset tax rules and causal discovery algorithms, this module establishes a directed acyclic graph (DAG) among the variables in the standardized underlying feature matrix, dividing the nodes into causal variable sets, internal variable sets, interference variable sets, and outcome variable sets to generate a causal network topology. The neural symbolic causal decoupling module receives the standardized underlying feature matrix and the causal network topology, and uses an internally configured deep neural network layer and symbolic logic engine to project the standardized underlying feature matrix into the latent space, outputting a decoupled feature representation isomorphic to the causal network topology. The core tax risk inference module performs blocking calculations on the interference signal set in the decoupled feature representation and calculates the causal effect distribution of the causal signal set on the outcome signal set to generate a baseline tax risk dynamic score and risk causal path. The game-theoretic counterfactual generation engine receives decoupled feature representations and baseline tax risk dynamic scores. Upon receiving the training start command, it iteratively calculates counterfactual perturbations within the feature space of the causal signal set according to a preset game optimization objective function. These counterfactual perturbations are then input into the symbolic logic engine, which performs forward propagation calculations in conjunction with preset financial audit rules to generate counterfactual tax-related feature data. The dynamic interactive training module receives and presents the baseline tax risk dynamic score, risk causal path, and counterfactual tax-related feature data on its interface. It also receives simulation parameters input by the operator through the interactive interface, which are then passed to the game-theoretic counterfactual generation engine for secondary inference. When preset conditions are met, it collects counterfactual tax-related features and attaches out-of-distribution hard sample labels for feedback.

[0005] Preferably, to achieve standardized input of multimodal data, the multi-source data fusion and causal graph construction module uses a word embedding algorithm to convert the business scope text into continuous numerical vectors, and a graph neural network algorithm to aggregate enterprise association graph records into topological node vectors. The multi-source data fusion and causal graph construction module concatenates the numerical vectors, topological node vectors, and the numerical matrix obtained through one-hot encoding to form a time-aligned original feature set; it then constructs a discrete-time sliding window with a natural month as the time step, resamples the original feature set, and uses a max-min scaling algorithm to linearly map the resampled multidimensional intermediate feature tensor to a floating-point range of 0 to 1, generating a standardized underlying feature matrix.

[0006] Preferably, to extract the essential economic causal structure from massive amounts of financial and tax data and eliminate statistical spurious correlations caused by confounding factors, the multi-source data fusion and causal graph construction module uses a conditional independence test algorithm to analyze the feature variables in the standardized underlying feature matrix. Based on the partial correlation coefficient and a preset significance threshold, it removes connection edges that satisfy conditional independence to generate an initial undirected graph. The module then searches for conflict node structures (V-shaped structures) in the initial undirected graph to determine unidirectional connection edges. It reads a preset tax prior rule base and uses the time sequence constraints, business inclusion constraints, and node attribute labels defined in the tax prior rule base to perform forced directional processing on the remaining undirected connection edges to generate a global directed acyclic graph (DAG). Based on the node attribute labels, the module divides the nodes of the global DAG into causal variable sets, internal variable sets, confounding variable sets, and outcome variable sets.

[0007] Preferably, addressing the issue that deep learning models cannot simulate rigid business laws, the neural symbolic causal decoupling module utilizes an internally configured deep neural network layer and a symbolic logic engine to establish a hybrid-driven neural symbolic structure configuration model based on the set of parent nodes for each node variable in the causal network topology. This hybrid-driven neural symbolic structure configuration model establishes node type indicator variables. When the node type indicator variable represents a numerical cognitive attribute for the current node, the system activates a multilayer perceptron neural network model, receiving parent node features and independent exogenous noise signals as input, and calculates and outputs the predicted features of the current node variable through nonlinear mapping. When the node type indicator variable represents a symbolic constraint attribute for the current node, the system stops calling the multilayer perceptron neural network model, calls a non-parametric symbolic logic operator, receives parent node features, performs mathematical logic operations, and outputs verification features that satisfy financial conservation constraints. The value of the node type indicator variable is determined by the neural symbolic causal decoupling module based on the preset attribute labels of each node variable in the causal network topology: if the node variable corresponds to a preset numerical cognitive attribute label, the node type indicator variable is assigned a value of 1; if the node variable corresponds to a preset symbolic constraint attribute label, the node type indicator variable is assigned a value of 0. The symbolic constraint attribute label corresponds to nodes in a financial logic rule base pre-encoded with a cascaded calculation formula for balance sheet balancing identities and VAT payable, and the numerical cognitive attribute label corresponds to nodes that require floating-point estimation through nonlinear mapping. During feature space decoupling, in the model training phase, the hybrid-driven neural symbolic structure configuration model is reorganized based on the node features output by the training sample batches into a causal signal feature matrix, an internal signal feature matrix, and an interference signal feature matrix. The neural symbolic causal decoupling module constructs a total loss function for the decoupling representation, which includes a baseline risk prediction cross-entropy loss term and an independence regularization penalty term. The independence regularization penalty term calculates the statistical correlation between pairs of distinct feature matrices using the Hilbert-Schmidt independence criterion based on radial basis kernel function mapping. The total loss gradient generated by the total loss function initiates the backpropagation algorithm, iteratively updating the weight parameters of the multilayer perceptron neural network model. This aims to reduce the statistical correlation between distinct feature matrices and satisfy the convergence condition, outputting a decoupled feature representation that satisfies statistical independence.

[0008] Preferably, to accurately predict the true tax-related risk status of enterprises after excluding environmental interventions, the core tax risk inference module is internally configured with a risk classification prediction network containing a graph attention layer. The core tax risk inference module inputs the causal variable feature matrix and the interference variable feature matrix into the risk classification prediction network for forward propagation to output a conditional probability distribution. It also statistically analyzes the empirical marginal probability distribution of the interference signal feature matrix in the current batch of input sample data, and uses the empirical marginal probability distribution to perform a weighted summation of the conditional probability distribution to eliminate statistical correlation bias introduced by the interference variable feature matrix, thus calculating the pure causal effect distribution data. The core tax risk inference module uses a preset Sigmoid-type numerical mapping function to perform scaling calculations on the pure causal effect distribution data, nonlinearly mapping it to the 0-100 interval to generate a baseline dynamic tax risk score. Meanwhile, the core tax risk inference module uses the target variable node representing the tax risk status as the starting point for path backtracking. It iterates backward along the directed edges of the causal network topology from the starting point of the path backtracking, iteratively filters out the incoming adjacent nodes in the attention weight matrix of the graph attention layer whose attention weight values ​​are greater than the preset connection threshold and retains the corresponding directed edges, iteratively backtracks to the initial node in the causal variable set that has no incoming causal edges, thereby integrating and generating the risk causal path.

[0009] Preferably, to assess the evolution trend of tax risk under different risk states, the game-theoretic counterfactual generation engine is configured with a Nash game adversarial optimization mechanism. The game-theoretic counterfactual generation engine instantiates an adversarial optimization agent in its internal computing environment, constructing a Nash game adversarial optimization mechanism containing a minimax game optimization objective function. This mechanism drives the counterfactual perturbation towards optimizing to reduce risk exposure by minimizing the risk classification loss function between the updated counterfactual tax risk prediction value and the true label classification boundary. Simultaneously, it limits the range of change in the counterfactual perturbation's impact by calculating the maximum mean difference distance between the original causal signal feature matrix and the updated counterfactual causal feature matrix. In its inner loop, the game-theoretic counterfactual generation engine fixes the weight parameters and maximizes the risk classification loss function along the positive gradient direction of the counterfactual perturbation vector; in its outer loop, it fixes the counterfactual perturbation and minimizes the risk classification loss function along the negative gradient direction of the weight parameters. In each iteration, the game-theoretic counterfactual generation engine uses a projection truncation function to forcibly restrict the updated counterfactual perturbation to the initial range of legal financial values, iteratively calculating until a preset convergence condition is met to output the converged counterfactual perturbation.

[0010] Preferably, to ensure that the features output by the multilayer perceptron neural network model satisfy the financial data logic, the game-theoretic counterfactual generation engine performs matrix addition operations on the optimized counterfactual perturbation and the original causal variable feature matrix to generate a counterfactual causal feature matrix, and solves for the node variables in the causal network topology based on the topological sorting relationship. When the node type indicator variable is 1, the multilayer perceptron neural network model is invoked for node feature mapping; when the node type indicator variable is 0, the multilayer perceptron neural network model is stopped and the symbolic logic operator is activated. The latest counterfactual feature value is used as an input parameter and substituted into the financial logic rule base containing the balance sheet balancing identity and the cascaded calculation formula of VAT payable for algebraic rule calculation. The output is a counterfactual logic feature vector that satisfies the financial conservation constraint. The counterfactual logic features are then used to replace the floating-point estimated output value of the multilayer perceptron neural network model to generate overall logically consistent counterfactual tax-related feature data.

[0011] Preferably, to conduct decision review and achieve incremental data feedback, the dynamic interactive training module converts baseline tax risk dynamic scores and risk causal paths into core case view data, and converts counterfactual tax-related characteristic data into adversarial evolution case view data, which are then rendered in the left and right areas of the training window. The module captures operator-input numerical perturbations or state change commands through a parameter interaction interface that binds causal variable nodes, encapsulates them into intervention command packages containing node identifiers and target values, and transmits them to the game-theoretic adversarial counterfactual generation engine to sever the incoming causal edges of the variable nodes corresponding to the target values ​​and write them as constant terms into the symbolic logic operator for re-execution of the deduction. A reset interface is also configured to restore the initial state. The system records the audit decision sequence and compares its overlap with the risk causal path to generate a decision review analysis report and a comprehensive evaluation score. When operators fail to identify the true violation and the overall assessment score falls below the minimum threshold, the dynamic interactive training module uses a data labeling tool to attach out-of-distribution hard sample labels to the current counterfactual tax-related feature data. This data is then written to the incremental training set partition of the underlying database via a network transmission protocol. During system retraining, this data is mixed with historical training data and incrementally trained using the gradient descent algorithm. The neural symbolic causal decoupling module and the core tax risk inference module read data from the incremental training set partition, mix it with historical training data, and use the gradient descent algorithm to incrementally train the network model to update the weights, thus improving system performance.

[0012] The present invention, by adopting the above technical solution, can bring the following beneficial effects:

[0013] 1. Through the multi-source data fusion and causal graph construction module, a global directed acyclic graph is constructed using the conditional independence test algorithm and the tax prior rule base, and the semantic partitioning of the four major variable sets is completed. The neural symbolic causal decoupling module uses the Hilbert-Schmidt independence criterion function based on radial basis kernel function mapping to calculate the total loss of the decoupling representation to realize the independence constraint of the distinct feature matrices in the latent space, thereby eliminating the statistical pseudo-correlation caused by confounding factors in multimodal fiscal and tax data, and realizing the accurate cutting off of heterogeneous signals and deep decoupling of underlying features.

[0014] 2. A hybrid-driven neural symbolic structure configuration model is established through a neural symbolic decoupling module. When the node type indicator variable is 0, a symbolic logic operator embedded with the balance sheet balancing identity and the VAT payable calculation formula is invoked to directly replace and cover the floating-point estimated output value of the multilayer perceptron neural network model. The game-theoretic counterfactual generation engine performs symbolic logic self-consistency verification and neural algebra rule calculation along the topological sorting iterative relationship when reconstructing counterfactual data. Thus, while ensuring the high-dimensionality capability of the network, rigid fiscal and tax legal rules are embedded, and the data reconstruction meets the financial numerical conservation constraints.

[0015] 3. By instantiating an adversarial optimization agent through a game-theoretic counterfactual generation engine and using a minimax game to iteratively optimize the objective function and projection truncation function to optimize the output counterfactual perturbation, the dynamic interactive training module converts the updated counterfactual tax-related feature data into adversarial evolution case view data and uses a data labeling tool to backflow difficult samples that fail to identify the true violation status to the incremental training set partition to perform incremental training of the network model. This establishes an adaptive training environment and completes data closed-loop feedback, thereby realizing the cyclical improvement of the risk identification system's robustness in anomaly identification under unknown scenarios outside the distribution. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the tax risk intelligent identification and training system based on multi-source data fusion according to the present invention.

[0017] Figure 2 This is a flowchart illustrating the overall workflow of the system of the present invention;

[0018] Figure 3 This is a flowchart of the multi-source data fusion and causal graph construction process of the present invention;

[0019] Figure 4 This is a flowchart of the neural symbolic causal decoupling process of the present invention;

[0020] Figure 5 This is a flowchart illustrating the core tax risk assessment process of this invention.

[0021] Figure 6This is a flowchart of the game-theoretic counterfactual generation process of the present invention;

[0022] Figure 7 This is a flowchart of the dual-track parallel rendering, hypothetical intervention, and feedback process of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Example 1

[0025] Please refer to the appendix. Figure 1 The present invention provides a tax risk intelligent identification and training system based on multi-source data fusion, including: a multi-source data fusion and causal graph construction module 110, a neural symbolic causal decoupling module 120, a core tax risk inference module 130, a game-theoretic counterfactual generation engine 140, and a dynamic interactive training module 150.

[0026] The multi-source data fusion and causal graph construction module 110 is used to receive tax-related multi-source raw data. The multi-source data fusion and causal graph construction module 110 performs data calibration and standardization processing, mapping the data into a feature matrix with a unified time dimension. The multi-source data fusion and causal graph construction module 110 is configured with a constraint-based causal graph generation algorithm, which is used to establish a directed acyclic graph among feature variables and output a causal network topology structure containing a set of cause variables, a set of internal variables, a set of interference variables, and a set of result variables.

[0027] The neural symbolic causal decoupling module 120 is communicatively connected to the multi-source data fusion and causal graph construction module 110. The neural symbolic causal decoupling module 120 receives the standardized low-level feature matrix and the causal network topology. The neural symbolic causal decoupling module 120 is internally configured with a deep neural network layer and a symbolic logic engine, which is used to project the standardized low-level feature matrix to the latent space and output the decoupling feature representation isomorphic to the causal network topology.

[0028] The core tax risk inference module 130 is connected to the neural symbolic causal decoupling module 120. The core tax risk inference module 130 receives the decoupling feature representation. The core tax risk inference module 130 uses the backdoor adjustment operator to perform blocking calculation on the set of interference variables in the decoupling feature representation, cuts off the gradient propagation path of the interference signal feature matrix to the target risk node, and outputs the enterprise's baseline tax risk dynamic score and risk causal path.

[0029] The game-theoretic counterfactual generation engine 140 is connected to the neural symbolic causal decoupling module 120 and the core tax risk inference module 130, respectively. The game-theoretic counterfactual generation engine 140 receives the decoupling feature representation and the baseline tax risk dynamic score. The game-theoretic counterfactual generation engine 140 is configured with an adversarial optimization algorithm to calculate counterfactual perturbations within the set causal graph interference boundary. Based on the calculated counterfactual perturbations, the game-theoretic counterfactual generation engine 140 re-executes the forward propagation calculation through the symbolic logic engine and outputs counterfactual tax-related feature data that meets the constraints.

[0030] The dynamic interactive training module 150 is connected to the core tax risk inference module 130 and the game-theoretic counterfactual generation engine 140. The dynamic interactive training module 150 is used to receive and render baseline tax risk dynamic scores, risk causal paths and counterfactual tax-related feature data. The dynamic interactive training module 150 provides a parameter interaction interface. The dynamic interactive training module 150 sends the received interaction parameters back to the game-theoretic counterfactual generation engine 140 to trigger regeneration, records the exercise operation sequence, and feeds back the adversarial sample data outside the preset interaction conditions to the system storage.

[0031] Example 2

[0032] Please refer to the appendix. Figure 2 Based on Example 1 and the aforementioned system, this example provides a workflow for a tax risk intelligent identification and training system based on multi-source data fusion, including the following steps:

[0033] Step S100: The multi-source data fusion and causal graph construction module 110 reads the enterprise's multi-source tax-related raw data, including enterprise business registration information, tax return data, and invoice circulation records; the multi-source data fusion and causal graph construction module 110 performs time-dimensional alignment and standardization processing on the multi-source tax-related raw data, and outputs a standardized underlying feature matrix; the multi-source data fusion and causal graph construction module 110 establishes a directed acyclic graph among the variables of the standardized underlying feature matrix according to preset tax rules and causal discovery algorithms; the multi-source data fusion and causal graph construction module 110 divides the nodes of the directed acyclic graph into a set of causal variables, a set of internal variables, a set of interference variables, and a set of result variables, generates a causal network topology, and sends the standardized underlying feature matrix and the causal network topology to the neural symbolic causal decoupling module 120;

[0034] Step S200: The neural symbolic causal decoupling module 120 receives the standardized low-level feature matrix and causal network topology structure sent by the multi-source data fusion and causal graph construction module 110. The neural symbolic causal decoupling module 120 schedules the corresponding computation logic according to the node classification attributes of the causal network topology structure. The neural symbolic causal decoupling module 120 processes the standardized low-level feature matrix through the internally configured deep neural network layer and symbolic logic engine, splits the standardized low-level feature matrix and projects it to the latent space. The neural symbolic causal decoupling module 120 outputs a decoupling feature representation isomorphic to the causal network topology structure. The neural symbolic causal decoupling module 120 sends the decoupling feature representation to the core tax risk inference module 130 and the game adversarial counterfactual generation engine 140.

[0035] Step S300: The core tax risk inference module 130 receives the decoupled feature representation sent by the neural symbolic causal decoupling module 120. The core tax risk inference module 130 uses a backdoor to adjust the calculation operator to perform blocking calculation on the set of interference variables in the decoupled feature representation. The core tax risk inference module 130 calculates the causal impact of the set of cause variables on the set of result variables. The core tax risk inference module 130 generates a baseline tax risk dynamic score and the corresponding risk causal path. The core tax risk inference module 130 sends the baseline tax risk dynamic score and the risk causal path to the dynamic interactive training module 150 and the game-theoretic counterfactual generation engine 140.

[0036] Step S400: The game-theoretic counterfactual generation engine 140 receives the decoupled feature representation and the baseline tax risk dynamic score. When the system receives the training start instruction, the game-theoretic counterfactual generation engine 140 calculates the counterfactual perturbation in the feature space corresponding to the causal variable set according to the preset game optimization objective function. The game-theoretic counterfactual generation engine 140 inputs the counterfactual perturbation into the symbolic logic engine, performs forward propagation calculation in combination with the preset verification rules, generates counterfactual tax-related feature data, and sends the counterfactual tax-related feature data to the dynamic interactive training module 150.

[0037] Step S500: The dynamic interactive training module 150 receives the baseline tax risk dynamic score, risk causal path, and counterfactual tax-related feature data. The dynamic interactive training module 150 presents the baseline tax risk dynamic score, risk causal path, and counterfactual tax-related feature data on the interface. The dynamic interactive training module 150 receives the hypothetical interaction parameters input by the operator through the interactive interface and sends the hypothetical interaction parameters to the game adversarial counterfactual generation engine 140. The game adversarial counterfactual generation engine 140 re-executes causal inference and forward propagation calculations based on the hypothetical interaction parameters to generate updated counterfactual tax-related feature data and updated risk causal paths. The game adversarial counterfactual generation engine 140 returns the updated counterfactual tax-related feature data and updated risk causal paths to the dynamic interactive training module 150 for rendering. The dynamic interactive training module 150 records the operation sequence during the interaction process and stores adversarial sample data outside the preset range into the basic database.

[0038] Example 3

[0039] To further illustrate the specific implementation process of each step in Embodiment 2, the following is in conjunction with the appendix. Figure 3 To be continued Figure 7 The specific methods involved in each module are explained in detail.

[0040] In step S100, the multi-source data fusion and causal graph construction module 110 may employ the following methods when performing time dimension and standardization processing:

[0041] Step S111: Multi-source data fusion and causal graph construction module 110 acquires multimodal tax-related raw data. The multimodal tax-related raw data includes formatted features and unformatted features. Formatted features include registered capital, tax declaration amount and invoice issuance time. Unformatted features include business scope text and enterprise association graph records.

[0042] Step S112: The multi-source data fusion and causal graph construction module 110 performs feature mapping on the multimodal tax-related raw data. The multi-source data fusion and causal graph construction module 110 uses a word embedding algorithm to convert the business scope text into a continuous numerical vector. The multi-source data fusion and causal graph construction module 110 uses a graph neural network algorithm to aggregate the enterprise association graph records into a topological node vector.

[0043] Step S113: The multi-source data fusion and causal graph construction module 110 converts discrete business registration categories and industry classification codes into numerical matrices using a one-hot coding algorithm. The multi-source data fusion and causal graph construction module 110 then concatenates the numerical vectors, topological node vectors, and the numerical matrices obtained through the one-hot coding algorithm to form a time-aligned original feature set.

[0044] Step S114: The multi-source data fusion and causal graph construction module 110 performs time step resampling processing on the original feature set, constructs a discrete time sliding window, and performs resampling calculation on the original feature set with natural month as the time step.

[0045] Step S115: For data sources with inconsistent frequencies, the multi-source data fusion and causal graph construction module 110 applies differentiated aggregation rules. For single-item flow features with sampling frequencies higher than the time step, the multi-source data fusion and causal graph construction module 110 performs monthly summation and averaging calculations. For business changes and annual financial report features with sampling frequencies lower than the time step, the multi-source data fusion and causal graph construction module 110 performs forward filling calculations. The multi-source data fusion and causal graph construction module 110 assigns the data values ​​of missing time nodes to the data values ​​of the most recent preceding time nodes. The forward filling calculation is suitable for low-frequency static features where attribute values ​​remain stable within adjacent time steps to avoid introducing trend deviations that do not conform to the actual business due to linear interpolation extrapolation. For low-frequency features with obvious time-series evolution trends, periodic extension calculation based on the business cycle is used, or the most recent valid value is maintained and a missing value is marked to suppress statistical pseudo-correlation caused by repeated filling.

[0046] Step S116: After time dimension analysis and processing, the multi-source data fusion and causal graph construction module 110 outputs a multi-dimensional intermediate feature tensor. The multi-source data fusion and causal graph construction module 110 performs dimensional normalization processing on the multi-dimensional intermediate feature tensor. The multi-source data fusion and causal graph construction module 110 calculates the extreme values ​​of each feature vector in the multi-dimensional intermediate feature tensor. The multi-source data fusion and causal graph construction module 110 uses the maximum-minimum scaling algorithm to linearly map the values ​​of different dimensions to the floating-point range of 0 to 1.

[0047] Step S117: After the multi-source data fusion and causal graph construction module 110 completes the dimensional normalization process, it generates a standardized low-level feature matrix. The multi-source data fusion and causal graph construction module 110 then transmits the standardized low-level feature matrix to the neural symbolic causal decoupling module 120.

[0048] Based on the aforementioned time dimension and standardization processing, in order to further construct the causal dependencies between variables, in step S100, the multi-source data fusion and causal graph construction module 110 can adopt the following methods when performing causal graph construction, including:

[0049] Step S121: The multi-source data fusion and causal graph construction module 110 receives the standardized low-level feature matrix and extracts all feature variables from the standardized low-level feature matrix.

[0050] Step S122: The multi-source data fusion and causal graph construction module 110 uses the conditional independence test algorithm to perform traversal calculations on the feature variables. The multi-source data fusion and causal graph construction module 110 judges the statistical independence between the feature variables based on the partial correlation coefficient of the calculation results and the preset significance threshold.

[0051] Step S123: Multi-source data fusion and causal graph construction module 110 removes the connection edges of feature variables that satisfy conditional independence and generates an initial undirected graph that retains the dependencies;

[0052] Step S124: The multi-source data fusion and causal graph construction module 110 searches for conflict node structure (V-shaped structure) in the initial undirected graph, and the multi-source data fusion and causal graph construction module 110 determines the unidirectional connection edge between adjacent nodes based on the conflict node structure;

[0053] Step S125: The multi-source data fusion and causal graph construction module 110 reads the preset tax prior rule library and extracts the time order constraints, business inclusion constraints and node attribute labels defined in the tax prior rule library.

[0054] Step S126: The multi-source data fusion and causal graph construction module 110 uses time order constraints to perform forced directionification on the remaining undirected connection edges to generate a global directed acyclic graph;

[0055] Step S127: The multi-source data fusion and causal graph construction module 110 performs a set partitioning operation on the nodes of the global directed acyclic graph. Based on the node attribute labels, the multi-source data fusion and causal graph construction module 110 divides the nodes that represent the basic operating attributes of the enterprise and have inbound causal edges into a set of causal variables. The set of causal variables includes the enterprise registration time node and the legal person change frequency node.

[0056] Step S128: Multi-source data fusion and causal graph construction module 110 matches node attribute labels and divides the nodes that represent the enterprise's transaction behavior directly driven by the causal variable set into internal variable sets. The internal variable sets include invoice issuance time difference nodes and fund turnover frequency nodes.

[0057] Step S129: Multi-source data fusion and causal graph construction module 110 matches node attribute labels and divides nodes that represent the external macro environment and simultaneously have a causal impact on the causal variable set and the final state into interference variable set. The interference variable set includes industry prosperity index nodes and regional hot policy nodes.

[0058] Step S1210: The multi-source data fusion and causal graph construction module 110 sets the nodes representing the final tax violation status as the result variable set. The multi-source data fusion and causal graph construction module 110 integrates the cause variable set, internal variable set, interference variable set and result variable set to generate a causal network topology. The multi-source data fusion and causal graph construction module 110 sends the causal network topology to the neural symbolic causal decoupling module 120.

[0059] After constructing the causal network topology through the above steps, the system enters the neural symbolic causal decoupling stage. In step S200, the neural symbolic causal decoupling module 120 can use the following methods when outputting the decoupling feature representation:

[0060] Step S211: The neural symbolic causal decoupling module 120 receives the standardized low-level feature matrix and causal network topology structure sent by the multi-source data fusion and causal graph construction module 110. The neural symbolic causal decoupling module 120 extracts all node variables in the causal network topology structure.

[0061] Step S212: The neural symbolic causal decoupling module 120 parses the attribute labels of each node variable in the causal network topology. The neural symbolic causal decoupling module 120 configures a node type indicator variable for each node variable according to the attribute label by iterating through the node attribute mapping dictionary of the parent node. For node variables representing numerical cognitive attributes, the neural symbolic causal decoupling module 120 sets the value of the node type indicator variable to 1; for node variables representing symbolic constraint attributes, the neural symbolic causal decoupling module 120 sets the value of the node type indicator variable to 0.

[0062] Step S213: The neural symbolic causal decoupling module 120 establishes a structural configuration based on the set of parent nodes of each node variable. The mathematical expression of the hybrid-driven neural symbolic structural configuration model is shown in formula (1):

[0063] (1)

[0064] In formula (1), Represents the first in the causal network topology Individual node metrics; Indicates the first The node type indicator variable corresponding to each node indicator, when representing numerical cognitive attributes... When representing symbolic constraint attributes ; Indicates the first The set of parent nodes for each node variable; Represents exogenous noise variables; Indicated by A multilayer perceptron neural network model with weight parameters; Represents non-parametric symbolic logic operators;

[0065] Step S214: When the value of the node type indicator variable is 1, the neural symbolic causal decoupling module 120 activates the multilayer perceptron neural network model. The multilayer perceptron neural network model receives the variable features in the parent node set and the exogenous noise variable as input, and outputs the first... Predictive features of node variables;

[0066] Step S215: When the value of the node type indicator variable is 0, the neural symbolic causal decoupling module 120 activates the symbolic logic operator. The symbolic logic operator is pre-encoded with the tax law calculation formula and the accounting balance identity. The symbolic logic operator receives the numerical features in the parent node set, performs mathematical logic operations, and outputs the first node that satisfies the financial conservation constraint. Verification characteristics of individual node variables;

[0067] Step S216: The neural symbolic causal decoupling module 120 traverses each node variable in the causal network topology, and based on the value of the node type indicator variable, takes the corresponding prediction feature or verification feature as the first... The final output representation of each node indicator completes the construction of the hybrid-driven neural symbolic structure configuration model. The neural symbolic causal decoupling module 120 sends the hybrid-driven neural symbolic structure configuration model to the core tax risk inference module 130.

[0068] After obtaining the above-mentioned structural configuration model, in order to further train and obtain the decoupled feature representation, the core tax risk inference module 130 may use the following methods in step S200 when training the decoupled feature representation:

[0069] Step S221: During the model training phase, the neural symbolic causal decoupling module 120 obtains the node features of the hybrid-driven neural symbolic structure configuration model based on the batch output of training samples. The neural symbolic causal decoupling module 120 reorganizes the node features into a cause signal feature matrix, an internal signal feature matrix, and an interference signal feature matrix according to the node partitioning relationship of the causal network topology.

[0070] Step S222: The neural symbolic causal decoupling module 120 constructs the target loss function for network training optimization. The target loss function includes a prediction loss term for risk state classification and an independence regularization penalty term for feature distribution constraints.

[0071] Step S223: The mathematical expression of the target loss function for the latent space decoupling and correlation constraints is shown in Equation (2):

[0072] (2)

[0073] In formula (2), The total loss function represents the decoupling process. This represents the baseline risk prediction cross-entropy loss term; This represents the predicted output value of the network's forward propagation. This represents the true label distribution corresponding to the multi-source raw data; This represents the hyperparameter of the independence constraint weights; and These represent the sets of feature vector samples extracted from the three types of matrices: the causal signal feature matrix, the internal signal feature matrix, and the interference signal feature matrix, respectively. This represents the Hilbert-Schmidt independence criterion function based on radial basis kernel function mapping, used to measure the statistical correlation between two sets of feature vector sample sets;

[0074] Step S224: The neural symbolic causal decoupling module 120 calculates the error gradient of the prediction loss term using the predicted output value and the true label distribution. The neural symbolic causal decoupling module 120 calculates the statistical dependency penalty gradient between distinct feature matrices through the Hilbert-Schmidt independence criterion function.

[0075] Step S225: The neural symbolic causal decoupling module 120 sums the prediction error gradient and the mutual information penalty gradient to obtain the total loss gradient. The neural symbolic causal decoupling module 120 executes the backpropagation algorithm based on the total loss gradient.

[0076] Step S226: The neural symbolic causal decoupling module 120 updates the weight parameters of the multilayer perceptron neural network model through the backpropagation algorithm. The neural symbolic causal decoupling module 120 reduces the statistical correlation between distinct feature matrices through parameter iterative updates.

[0077] Step S227: After the neural symbolic causal decoupling module 120 completes training, it outputs a decoupling feature representation that satisfies statistical independence. The neural symbolic causal decoupling module 120 sends the decoupling feature representation to the core tax risk inference module 130 and the game adversarial counterfactual generation engine 140.

[0078] After obtaining the decoupled feature representation, the system enters the core tax risk inference stage. In step S300, the core tax risk inference module 130 may use the following methods when calculating the risk score:

[0079] Step S311: The core tax risk inference module 130 receives the decoupling feature representation sent by the neural symbolic causal decoupling module 120, and the core tax risk inference module 130 extracts the causal signal feature matrix, internal signal feature matrix and interference signal feature matrix from the decoupling feature representation;

[0080] Step S312: The core tax risk inference module 130 is internally configured with a risk classification prediction network containing a graph attention layer. The core tax risk inference module 130 transmits the cause signal feature matrix and the interference signal feature matrix as input data to the risk classification prediction network. The core tax risk inference module 130 performs forward propagation calculation and outputs the conditional probability distribution under the model parameters.

[0081] Step S313: The core tax risk inference module 130 starts the backdoor adjustment calculation program. After truncating the gradient propagation path of the interference variable feature matrix to the target risk node in the calculation graph, the core tax risk inference module 130 calculates the empirical marginal probability distribution of the interference variable feature matrix in the current batch of input sample data.

[0082] Step S314: The core tax risk inference module 130 uses the empirical marginal probability distribution of the interference signal feature matrix to perform a weighted summation calculation on the conditional probability distribution in order to eliminate the statistical correlation bias introduced by the interference signal feature matrix and calculate the pure causal effect distribution data.

[0083] Step S315: The core tax risk inference module 130 maps the causal effect value to a continuously output numerical result as the enterprise's baseline tax risk dynamic score. The core tax risk inference module 130 reads the attention weight values ​​between each node in the corresponding causal network topology in the graph attention layer.

[0084] Step S316: The core tax risk inference module 130 exports the causal network topology structure, extracts the incoming neighbor nodes whose attention weight values ​​exceed the preset connection threshold, and generates risk causal paths by extracting the incoming neighbor nodes according to the causal network topology structure.

[0085] Step S317: The core tax risk inference module 130 sends the baseline tax risk dynamic score and risk causal path to the dynamic interactive training module 150. The core tax risk inference module 130 synchronously transmits the baseline tax risk dynamic score to the game adversarial counterfactual generation engine 140 as the optimization benchmark for the generation process.

[0086] After obtaining the risk score, in order to transform the causal effect into an explainable risk score and propagation path, in step S300, the core tax risk inference module 130 may use the following methods when outputting the risk score and path:

[0087] Step S321: The core tax risk inference module 130 obtains the calculated pure causal effect distribution data. The core tax risk inference module 130 is configured with a Sigmoid type numerical mapping function for nonlinearly mapping effect values ​​to the interval of 0 to 100.

[0088] Step S322: The core tax risk inference module 130 uses a Sigmoid-type numerical mapping function to perform scaling transformation calculations on the pure causal effect distribution data, and the core tax risk inference module 130 generates a baseline tax risk dynamic score;

[0089] Step S323: The core tax risk inference module 130 extracts the node features and attention weight matrix of the graph attention layer in the risk classification prediction network. The core tax risk inference module 130 uses the target variable node representing the tax risk status as the starting point for path backtracking.

[0090] Step S324: The core tax risk inference module 130 reads the causal network topology and performs a reverse topology traversal along the directed edges of the network topology.

[0091] Step S325: During the reverse traversal calculation, the core tax risk inference module 130 extracts all incoming adjacent nodes of the current node, and the core tax risk inference module 130 reads the attention weight values ​​between the current node and each incoming adjacent node from the attention weight matrix.

[0092] Step S326: The core tax risk inference module 130 compares the attention weight value with the preset connection threshold. The core tax risk inference module 130 retains the incoming adjacent nodes whose attention weight value is greater than the preset connection threshold and retains the corresponding directed edges.

[0093] Step S327: The core tax risk inference module 130 updates the selected incoming adjacent nodes to the current node. The core tax risk inference module 130 performs reverse topology calculation and iteratively backtracks to the initial node in the causal variable set that has no incoming causal edge.

[0094] Step S328: The core tax risk inference module 130 integrates all nodes and directed edges retained during the reverse traversal calculation process, and generates a risk causal path;

[0095] Step S329: The core tax risk inference module 130 sends the baseline tax risk dynamic score and risk causal path to the dynamic interactive training module 150 for display. The core tax risk inference module 130 synchronously transmits the baseline tax risk dynamic score to the game adversarial counterfactual generation engine 140 as the optimization benchmark for the generation process.

[0096] After obtaining the decoupled feature representation and baseline tax risk dynamic score, the game-adversarial counterfactual generation engine 140 enters the game-adversarial counterfactual generation stage. In step S400, the game-adversarial counterfactual generation engine 140 uses the following method to generate counterfactual perturbations:

[0097] Step S411: The game-theoretic counterfactual generation engine 140 receives the training start command sent by the dynamic interactive training module 150. The game-theoretic counterfactual generation engine 140 acquires the causal signal feature matrix output by the neural symbolic causal decoupling module 120 and the baseline tax risk dynamic score output by the core tax risk inference module 130.

[0098] Step S412: The game-theoretic counterfactual generation engine 140 initializes the adversarial optimization agent in the internal computing environment. The adversarial optimization agent is configured with a maximum-minimum game optimization objective function, which is used to calculate the optimized counterfactual perturbation within the preset causal graph dominance boundary.

[0099] Step S413: The mathematical expression of the minimax game optimization objective function against the optimization agent is shown in formula (3):

[0100] (3)

[0101] In formula (3), This represents the counterfactual perturbation vector applied to the feature matrix of the causal signal; This represents the weight parameters of the risk classification prediction network in the core tax risk inference module 130; This represents the original causal signal feature matrix; This represents the updated counterfactual cause feature matrix after applying counterfactual perturbation processing; This represents the predicted value of counterfactual tax risk. The risk classification loss function represents the deviation between the counterfactual tax risk prediction and the true label classification boundary. This represents the maximum mean difference distance between the distribution differences between the original cause signal feature matrix and the updated counterfactual cause feature matrix. This represents the distribution constraint penalty weight coefficient;

[0102] Step S414: The game-theoretic counterfactual generation engine 140 executes an alternating iterative optimization calculation program. In the inner loop, the game-theoretic counterfactual generation engine 140 fixes the weight parameters and performs a single-step update calculation along the positive gradient direction of the counterfactual perturbation to maximize the risk classification loss function.

[0103] Step S415: After completing the single-step update calculation, the game adversarial counterfactual generation engine 140 fixes the counterfactual perturbation in the outer loop and performs the single-step update calculation along the negative gradient direction of the weight parameters. The game adversarial counterfactual generation engine 140 drives the counterfactual perturbation to optimize in the direction of reducing the exposure of tax supervision risks by minimizing the risk classification loss function.

[0104] Step S416: The game-playing counterfactual generation engine 140 simultaneously limits the range of change of the counterfactual perturbation by calculating the maximum mean difference distance. The game-playing counterfactual generation engine 140 uses the projection truncation function to forcibly limit the updated counterfactual perturbation within the initial legal financial value range. The game-playing counterfactual generation engine 140 determines whether the alternating iterative calculation meets the preset convergence condition.

[0105] Step S417: When the alternating iterative calculation reaches the preset convergence condition, the game adversarial counterfactual generation engine 140 stops the optimization calculation and outputs the converged optimal counterfactual perturbation. The game adversarial counterfactual generation engine 140 inputs the optimal counterfactual perturbation into the symbolic logic engine to perform forward propagation of counterfactual data reconstruction.

[0106] Furthermore, after obtaining the optimal counterfactual perturbation, to ensure that the counterfactual data meets financial logic constraints, in step S400, the game-theoretic counterfactual generation engine 140 employs the following method when performing counterfactual reconstruction:

[0107] Step S421: The game-playing counterfactual generation engine 140 obtains the optimal counterfactual perturbation output by the optimization calculation. The game-playing counterfactual generation engine 140 performs matrix addition operation with the original cause signal feature matrix to generate the counterfactual cause feature matrix.

[0108] Step S422: The game-playing counterfactual generation engine 140 sends the counterfactual cause feature matrix to the neural symbolic causal decoupling module 120, and the neural symbolic causal decoupling module 120 receives the counterfactual cause feature matrix sent by the game-playing counterfactual generation engine 140.

[0109] Step S423: The neural symbolic causal decoupling module 120 starts the forward propagation algorithm according to the directed edge connection order of the causal network topology. During the execution of the forward propagation algorithm, the neural symbolic causal decoupling module 120 traverses each node variable in the causal network topology in the topological order.

[0110] Step S424: The neural symbolic causal decoupling module 120 extracts the node type indicator variable corresponding to the node variable currently performing topology calculation. When the value of the node type indicator variable corresponding to the node variable currently performing topology calculation is 1, the neural symbolic causal decoupling module 120 schedules the multilayer perceptron neural network model to perform node feature mapping.

[0111] Step S425: The multilayer perceptron neural network model calculates and outputs the counterfactual network feature vector corresponding to the node variable currently performing topology calculation. The neural symbolic causal decoupling module 120 uses the counterfactual network feature vector as the input data benchmark for subsequent downstream connected nodes.

[0112] Step S426: When the value of the node type indicator variable corresponding to the node variable currently performing topology calculation is 0, the neural symbolic causal decoupling module 120 stops calling the multilayer perceptron neural network model, and the neural symbolic causal decoupling module 120 calls the internally configured non-parametric symbolic logic operator.

[0113] Step S427: The symbolic logic operator reads the latest counterfactual feature value of the set of parent nodes corresponding to the node variable currently performing topology calculation at the current time step. The symbolic logic operator uses the latest counterfactual feature value as an input parameter and substitutes it into the pre-configured financial logic rule base. The financial logic rule base contains at least the balance sheet balancing identity and the cascading calculation formula for VAT payable.

[0114] Step S428: The symbolic logic operator performs algebraic rule calculations and outputs a counterfactual logic feature vector that satisfies the financial conservation constraint. The neural symbolic causal decoupling module 120 uses the counterfactual logic feature vector to replace the floating-point estimated output value of the multilayer perceptron neural network model, thereby eliminating the data logic inconsistency problem caused by the floating-point estimated output value.

[0115] Step S429: After completing the traversal calculation of all nodes in the causal network topology, the neural symbolic causal decoupling module 120 combines the output representations of the variables of each node to generate overall logically consistent counterfactual tax-related feature data. The neural symbolic causal decoupling module 120 transmits the counterfactual tax-related feature data to the game adversarial counterfactual generation engine 140.

[0116] Step S430: The game-theoretic counterfactual generation engine 140 receives counterfactual tax-related feature data and transmits the counterfactual tax-related feature data to the dynamic interactive training module 150 for data update and rendering display.

[0117] After obtaining risk scores, causal paths, and counterfactual tax-related characteristic data, the system enters the dynamic interactive training phase. In step S500, the dynamic interactive training module 150 can use the following methods during rendering and interaction:

[0118] Step S511: The dynamic interactive training module 150 receives the baseline tax risk dynamic score and risk causal path sent by the core tax risk inference module 130, and at the same time receives the counterfactual tax-related feature data sent by the game adversarial counterfactual generation engine 140.

[0119] Step S512: The dynamic interactive training module 150 converts the baseline tax risk dynamic score and risk causal path into core case view data, and converts the counterfactual tax-related feature data into counter-evolution case view data;

[0120] Step S513: The dynamic interactive training module 150 starts the front-end split-window rendering program. The dynamic interactive training module 150 renders the core case view data in the left area of ​​the training window and renders the adversarial evolution case view data in the right area of ​​the training window, completing the synchronous rendering of the dual-track parallel rendering.

[0121] Step S514: The dynamic interactive training module 150 activates the parameter intervention interaction interface configured in the adversarial evolution case view data, and establishes a binding mapping between the parameter intervention interaction interface and the cause variable nodes in the causal network topology.

[0122] Step S515: The dynamic interactive training module 150 captures the numerical perturbation or state change instructions input by the operator through the parameter interaction interface, and the dynamic interactive training module 150 encapsulates the numerical perturbation or state change instructions into an intervention instruction package containing node identifiers and target values.

[0123] Step S516: The dynamic interactive training module 150 transmits the intervention instruction packet to the game adversarial counterfactual generation engine 140 through the network communication protocol. The game adversarial counterfactual generation engine 140 receives the intervention instruction packet sent by the dynamic interactive training module 150.

[0124] Step S517: The game-theoretic counterfactual generation engine 140 parses the node identifiers and target values ​​in the intervention instruction package. The game-theoretic counterfactual generation engine 140 cuts off the incoming causal edges of the variable nodes corresponding to the target values, writes the target values ​​as constant terms into the corresponding symbolic logic operators, and re-executes causal inference and forward propagation.

[0125] Step S518: The game-theoretic counterfactual generation engine 140 returns the calculated and generated updated counterfactual tax-related feature data and updated risk causal path to the dynamic interactive training module 150;

[0126] Step S519: The dynamic interactive training module 150 receives the updated counterfactual tax-related feature data and updates the risk causal path, partially refreshes the adversarial evolution case view data on the right side of the training window area, and presents the risk cascading change process after the second deduction.

[0127] Step S520: The dynamic interactive training module 150 activates the reset interface, clears the current hypothetical intervention instruction package, and restores the adversarial evolution case view data to the initial state in step S512.

[0128] Furthermore, after the practical training is completed, in order to review the operators' audit decisions and feed back into the model training, in step S500, the dynamic interactive training module 150 adopts the following methods during decision review and sample feedback:

[0129] Step S521: The dynamic interactive training module 150 establishes an operation log. The dynamic interactive training module 150 captures and records the numerical perturbation commands, status change commands and final tax audit decision conclusions executed by the operator in real time according to the time sequence, and generates an audit decision time sequence.

[0130] Step S522: Upon receiving the training end instruction, the dynamic interactive training module 150 retrieves the risk causal path generated by the core tax risk inference module 130, and the dynamic interactive training module 150 compares and calculates the overlap between it and the audit decision time sequence.

[0131] Step S523: The dynamic interactive training module 150 extracts the correct decision nodes that match the risk causal path in the audit decision time sequence, and at the same time extracts the misjudged deviation nodes that do not match, and generates a decision review analysis report. The decision review analysis report includes a comprehensive evaluation score generated by the overlap comparison calculation results. The dynamic interactive training module 150 renders and displays the decision review analysis report in the training window.

[0132] Step S524: The dynamic interactive training module 150 retrieves the preset evaluation rule line, which includes the minimum score boundary for the operator to judge correctly.

[0133] Step S525: The dynamic interactive training module 150 compares the final tax audit judgment with the hidden real violation status inside the game adversarial counterfactual generation engine 140. If the final tax audit judgment fails to identify the real violation status and the operator's comprehensive evaluation score is lower than the minimum score boundary, the dynamic interactive training module 150 activates the hard sample labeling program.

[0134] Step S526: The dynamic interactive training module 150 extracts the corresponding counterfactual tax-related feature data in the current training window, and the dynamic interactive training module 150 uses a data labeling tool to add out-of-distribution difficult sample labels to the counterfactual tax-related feature data;

[0135] Step S527: The dynamic interactive training module 150 writes the counterfactual tax-related feature data carrying outlier sample labels into the incremental training set partition of the underlying database through the network transmission protocol, thus completing the data feedback.

[0136] Step S528: The neural symbolic causal decoupling module 120 and the core tax risk inference module 130 periodically read the data in the incremental training set partition. The neural symbolic causal decoupling module 120 and the core tax risk inference module 130 mix the data in the incremental training set partition with the historical training data in the basic database, and use the gradient descent algorithm to perform incremental training on the deep neural network layer in the neural symbolic causal decoupling module 120 and the risk classification prediction network in the core tax risk inference module 130, update the corresponding model weights, and improve the robustness of the system in identifying unseen tax-related data.

Claims

1. A tax risk intelligent identification and training system based on multi-source data fusion, characterized in that, include: The multi-source data fusion and causal graph construction module (110) is used to acquire multimodal tax-related raw data, perform time dimension alignment and standardization processing on the multimodal tax-related raw data, output a standardized underlying feature matrix, and establish a directed acyclic graph between the variables of the standardized underlying feature matrix according to preset tax rules and causal discovery algorithms; divide the nodes of the directed acyclic graph into a set of causal variables, a set of internal variables, a set of interference variables and a set of result variables to generate a causal network topology. The neural symbolic causal decoupling module (120) is connected to the multi-source data fusion and causal graph construction module (110) and is used to receive the standardized low-level feature matrix and the causal network topology. It uses the internally configured deep neural network layer and symbolic logic engine to project the standardized low-level feature matrix to the latent space and outputs a decoupling feature representation that is isomorphic to the causal network topology. The core tax risk inference module (130) is connected to the neural symbolic causal decoupling module (120) and is used to receive the decoupling feature representation, use the backdoor adjustment operator to perform blocking calculation on the set of interference variables in the decoupling feature representation, and calculate the causal effect distribution of the set of cause variables on the set of result variables to generate a baseline tax risk dynamic score and risk causal path. The game-theoretic counterfactual generation engine (140) is connected to the neural symbolic causal decoupling module (120) and the core tax risk inference module (130) respectively. It is used to receive the decoupling feature representation and the baseline tax risk dynamic score. When the training start instruction is received, it iteratively calculates the counterfactual perturbation in the feature space corresponding to the causal variable set according to the preset game optimization objective function. The counterfactual perturbation is input into the symbolic logic engine and forward propagation calculation is performed in combination with the preset financial verification rules to generate counterfactual tax-related feature data. The dynamic interactive training module (150) is connected to the core tax risk inference module (130) and the game-theoretic counterfactual generation engine (140) respectively. It is used to receive and render the baseline tax risk dynamic score, the risk causal path and the counterfactual tax-related feature data in the interface, and receive the simulation parameters input by the operator through the interactive interface to pass through to the game-theoretic counterfactual generation engine (140) to perform secondary inference. When the preset conditions are met, it captures the counterfactual tax-related feature data and attaches the distribution of difficult sample labels for feedback.

2. The tax risk intelligent identification and training system based on multi-source data fusion according to claim 1, characterized in that, In the multi-source data fusion and causal graph construction module (110), the specific method for processing the multimodal tax-related raw data in terms of time dimension and standardization is as follows: the business scope text is converted into a continuous numerical vector using a word embedding algorithm, and the enterprise association graph records are aggregated into a topological node vector using a graph neural network algorithm. The numerical vector, the topological node vector, and the numerical matrix obtained by the one-hot encoding algorithm are spliced ​​together to form an original feature set. A discrete-time sliding window is constructed with natural months as the time step to resample the original feature set. The resampled multidimensional intermediate feature tensor is linearly mapped to a dense floating-point matrix of 0 to 1 using the max-min scaling algorithm to generate the standardized bottom-level feature matrix.

3. The tax risk intelligent identification and training system based on multi-source data fusion according to claim 2, characterized in that, In the multi-source data fusion and causal graph construction module (110), the specific way to generate the causal network topology structure according to the preset tax rules and causal discovery algorithm is as follows: the features in the standardized bottom feature matrix are analyzed by using the conditional independence test algorithm, and the connection edges that satisfy the conditional independence are removed according to the partial correlation coefficient and the preset significance threshold to generate the initial undirected graph. The conflict node structure is searched in the initial undirected graph to determine the unidirectional connecting edges. The conflict node structure is a V-shaped structure. A preset tax prior rule base is read, and the remaining undirected connecting edges are forced to be directed using the time order constraints, business inclusion constraints, and node attribute labels defined in the tax prior rule base to generate a global directed acyclic graph. The nodes of the global directed acyclic graph are divided into the causal variable set, the internal variable set, the interference variable set, and the result variable set according to the node attribute labels.

4. The tax risk intelligent identification and training system based on multi-source data fusion according to claim 3, characterized in that, In the neural symbolic causal decoupling module (120), the internally configured deep neural network layer and the symbolic logic engine establish a hybrid-driven neural symbolic structure configuration model based on the set of parent nodes of each node variable in the causal network topology, and its mathematical expression is shown in formula (1): （1） In formula (1), Represents the first in the causal network topology. Individual node metrics; Indicates the first The node type indicator variable corresponding to each node indicator, when representing numerical cognitive attributes... When representing symbolic constraint attributes ; Indicates the first The set of parent nodes for each node variable; Represents exogenous noise variables; Indicated by A multilayer perceptron neural network model with weight parameters; Represents non-parametric symbolic logic operators.

5. The tax risk intelligent identification and training system based on multi-source data fusion according to claim 4, characterized in that, When the neural symbolic causal decoupling module (120) performs feature space decoupling, during the model training phase, it reorganizes the node features output by the hybrid-driven neural symbolic structure configuration model based on the batch output of training samples into a causal signal feature matrix, an internal signal feature matrix, and an interference signal feature matrix, and constructs the total loss function of the decoupling representation, the mathematical equation of which is shown in formula (2): （2） This represents the baseline risk prediction cross-entropy loss term; This represents the predicted output value of the network forward propagation; This represents the true label distribution; This represents the hyperparameter of the independence constraint weights; and These represent distinct feature matrix pairs taken from the causal signal feature matrix, the internal signal feature matrix, and the interference signal feature matrix, respectively. This represents the Hilbert-Schmidt independence criterion function based on radial basis kernel function mapping; The neural symbolic causal decoupling module (120) initiates a backpropagation algorithm to iteratively update the weight parameters based on the total loss gradient generated by the total loss function. The decoupled feature representation that satisfies statistical independence is output.

6. The tax risk intelligent identification and training system based on multi-source data fusion according to claim 5, characterized in that, In the core tax risk inference module (130), the specific method of using the backdoor adjustment operator to perform blocking calculation is as follows: The core tax risk inference module (130) is internally configured with a risk classification prediction network containing a graph attention layer. The core tax risk inference module (130) receives the decoupled feature representation and extracts the causal signal feature matrix and interference signal feature matrix from the decoupled feature representation. The causal signal feature matrix and the interference signal feature matrix are input into the risk classification prediction network for forward propagation to output the conditional probability distribution. The empirical marginal probability distribution of the conditional probability distribution is calculated. The conditional probability distribution is weighted and summed using the empirical marginal probability distribution to remove the statistical correlation bias introduced by the interference signal feature matrix. Pure causal effect distribution data is calculated and pure causal effect distribution data is retained.

7. The tax risk intelligent identification and training system based on multi-source data fusion according to claim 6, characterized in that, The core tax risk inference module (130) outputs the baseline tax risk dynamic score and risk causal path in the following way: it uses a preset Sigmoid numerical mapping function to perform scaling transformation calculation on the pure causal effect distribution data, and nonlinearly maps it to the interval of 0 to 100 to generate the baseline tax risk dynamic score; at the same time, it takes the target node representing the tax risk status as the starting point of the path backtracking, and performs reverse traversal along the directed edges of the causal network topology from the starting point of the path backtracking, iteratively selects the incoming nodes in the attention weight matrix of the graph attention layer whose attention weight values ​​are greater than the preset connection threshold and retains the corresponding directed edges, iteratively backtracks to the initial node in the causal variable set that has no incoming causal edges, thereby integrating and generating the risk causal path.

8. The tax risk intelligent identification and training system based on multi-source data fusion according to claim 7, characterized in that, The game-theoretic counterfactual generation engine (140) is configured with an adversarial optimization agent, and the mathematical expression of the maximum-minimum game optimization objective function of the adversarial optimization agent is shown in formula (3): （3） In the formula, This represents the counterfactual perturbation vector applied to the feature matrix of the causal signal; These represent the weight parameters of the risk classification prediction network; This represents the updated counterfactual cause feature matrix; Represents relative entropy (KL divergence); The distribution constraint penalty weight coefficient is represented; the game-theoretic counterfactual generation engine (140) fixes the weight parameters in the inner loop, performs single-step update calculation along the positive gradient direction of the counterfactual perturbation, and maximizes the risk classification loss function; fixes the counterfactual perturbation in the outer loop, performs single-step update calculation along the negative gradient direction of the weight parameters, and minimizes the risk classification loss function; and uses the projection truncation function to force the updated counterfactual perturbation to be restricted within the preset legal financial value range, and iteratively calculates alternately until the preset convergence condition is met to output the converged counterfactual perturbation.

9. The tax risk intelligent identification and training system based on multi-source data fusion according to claim 8, characterized in that, The specific method by which the game-theoretic counterfactual generation engine (140) generates counterfactual tax-related feature data in combination with preset financial verification rules is as follows: the optimal counterfactual perturbation output by the optimization calculation is added to the original cause signal feature matrix to generate a counterfactual cause feature matrix, and the variables of each node in the causal network topology are solved according to the topological sorting relationship; when the node type indicator variable takes the value of 1, the multilayer perceptron neural network model is called to perform node feature mapping. When the node type indicator variable is 0, the multilayer perceptron neural network model is stopped and the symbolic logic operator is activated. The latest counterfactual feature value is used as the input parameter and substituted into the financial logic rule base containing the balance sheet balancing identity and the cascaded calculation formula of the value-added tax payable for algebraic rule calculation. The counterfactual logic feature vector that satisfies the financial conservation constraint is output. The floating-point estimated output value of the multilayer perceptron neural network model is replaced by the counterfactual logic feature to generate overall logically consistent counterfactual tax-related feature data.

10. The intelligent tax risk identification and training system based on multi-source data fusion according to claim 1, characterized in that, The specific method by which the dynamic interactive training module (150) realizes dual-track parallel rendering and hard sample feedback is as follows: the baseline tax risk dynamic score and the risk causal path are converted into core case view data, the counterfactual tax-related feature data are converted into adversarial evolution case view data, and tiled rendering is performed in the left and right areas of the training window respectively; the numerical perturbation or state change instructions input by the operator are captured through the parameter interaction interface of the bound cause variable node and encapsulated into an intervention instruction package containing node identifier and target value, and transmitted to the game adversarial counterfactual generation engine (140) to cut off the input cause of the variable node corresponding to the target value. The result edge is written as a constant term into the corresponding symbolic logic operator to re-execute the deduction, and a reset interface is configured to restore the initial state; at the end of the training, the system records the audit decision sequence and compares it with the risk causal path to generate a decision review analysis report and a comprehensive evaluation score. When the operator fails to identify the real violation status and the comprehensive evaluation score is lower than the minimum score boundary, the data labeling tool is used to add out-of-distribution difficult sample labels to the current counterfactual tax-related feature data, and writes it into the incremental training set partition of the underlying database through the network transmission protocol, so that it can be mixed with historical training data and incremental training is performed using the gradient descent algorithm during system retraining.

Images