Digital finance and family consumption relation mode mining and simulation system

Abstract

The invention discloses a digital finance and family consumption relation mode mining and simulation system, and relates to the technical field of data mining. The digital finance and household consumption relation mode mining and simulation system comprises a data acquisition and preprocessing module used for acquiring a sample data set and preprocessing the sample data set to generate a standardized data set; the structure equation model construction module is used for carrying out association mining processing on the structure equation model based on a preset convolutional neural network and a graph neural network to generate a structure equation model; the causal association inference module is used for performing causal interference control identification processing on the model to generate a visual causal association map; and the digital finance and household consumption simulation module is used for performing simulation processing in a preset simulation scene based on the model and the map and outputting a simulation result report. Through collaborative analysis among the modules, intelligent deep mining of digital finance and household consumption related data and accurate identification and separation of a causal effect can be realized; and the reliability of relation mining is ensured.

Description

Technical Field

[0001] This invention relates to the field of data mining technology, specifically a system for mining and simulating the relationship between digital finance and household consumption patterns. Background Technology

[0002] The rapid development of digital finance, especially the widespread adoption of services such as internet consumer finance, mobile payment, and digital credit, has profoundly changed how families access financial resources and make consumption decisions. Understanding the impact mechanism of digital finance development on family consumption behavior is of great significance for stimulating consumption potential and promoting consumption upgrading.

[0003] Currently, research and practice in this field mainly rely on the following methods, but significant limitations remain: Analysis methods based on traditional econometric models, while capable of identifying statistical associations, struggle to characterize the complex interaction structures between multiple latent variables, and are even less effective in distinguishing between direct causal effects and indirect effects transmitted through mediating variables. Conclusions are easily influenced by endogeneity issues such as omitted variables and reverse causality. Structural equation modeling methods relying on expert experience heavily depend on researchers' prior theories and subjective experience. From setting and revising the model path diagram to evaluation, it is a process of repeated trial and error with high levels of human intervention, resulting in low efficiency and the model's quality being greatly influenced by personal experience. It is difficult to automatically discover unknown, nonlinear, and complex relationship patterns from massive, high-dimensional data, limiting the model's objectivity and ability to discover new knowledge.

[0004] The limitations of existing technologies include at least the following problems: Existing technologies struggle to automate and intelligently mine the deep relationships between massive, multi-source, and heterogeneous data on digital finance and household consumption. Consequently, when faced with complex variable relationships, such as nonlinear associations and the coexistence of multiple intermediary paths, it is difficult to efficiently and objectively construct interpretable theoretical models, leading to low model building efficiency and poor reproducibility. Furthermore, existing technologies also struggle to rigorously identify and separate causal effects from the constructed models. This makes it difficult to effectively eliminate confounding interferences caused by household self-selection or regional differences when evaluating the effectiveness of digital finance, easily leading to misjudging correlations as causal relationships and resulting in biased evaluation conclusions. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a system for mining and simulating the relationship between digital finance and household consumption patterns. This system solves the problem that existing technologies struggle to intelligently mine relevant data and identify and separate causal effects, which can easily lead to misjudgments.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a digital finance and household consumption relationship pattern mining and simulation system, comprising: a data acquisition and preprocessing module for acquiring sample datasets, including digital finance sample association data and household consumption sample datasets for each region, wherein the household consumption sample datasets include household consumption status data and digital service usage data of several households, and performing preprocessing to generate a standardized dataset; a structural equation model construction module for constructing an initial structural equation model based on the standardized datasets through a preset convolutional neural network and graph neural network, and performing iterative optimization and fitting verification to generate a structural equation model; a causal association inference module for performing causal interference control and identification processing on the structural equation model using a causal effect separation method that integrates propensity score matching and mediation effect analysis, and generating a visualized causal association map; and a digital finance and household consumption simulation module for performing simulation processing based on the structural equation model and the visualized causal association map in a preset simulation scenario, and outputting a simulation result report.

[0007] Furthermore, the digital financial sample associated data specifically comprises several digital financial variables; the household consumption status data includes several household consumption variables; and the digital financial service usage data specifically comprises several digital service usage variables.

[0008] Furthermore, the specific steps for generating the standardized dataset are as follows: data cleaning is performed on the digital finance sample association data and household consumption sample dataset for each region; and the missing data imputation data and household consumption sample dataset for each region are normalized to generate the standardized dataset.

[0009] Furthermore, the specific steps for constructing the initial structural equation model are as follows: Nonlinear feature extraction is performed on the standardized dataset using a pre-defined convolutional neural network to obtain a feature association set; topological relationship mining is performed on the feature association set using a pre-defined graph neural network to generate a variable association matrix; based on the variable association matrix, a variable relationship topology graph is constructed with digital financial variables, household consumption variables, and digital service usage variables as nodes; and the variable relationship topology graph is mapped to the initial structural equation model.

[0010] Further, the specific steps for generating the feature association set are as follows: The standardized dataset is input into a pre-defined convolutional neural network. A sliding convolution process is performed on the standardized dataset to extract local nonlinear features between digital financial variables, household consumption variables, and digital service usage variables, resulting in a local nonlinear feature set, which is then batch normalized. The batch-normalized local nonlinear feature set is downsampled to obtain a dimensionality-reduced feature set. Based on a skip connection mechanism, the local nonlinear feature set and the dimensionality-reduced feature set are fused and flattened to output the feature association set.

[0011] Further, the specific steps for generating the variable association matrix are as follows: First, construct initial graph structure data by combining a standardized dataset and a feature association set; second, input the initial graph structure data into a preset graph neural network, which sequentially includes a graph attention layer and a graph convolutional layer; third, in the graph attention layer, perform association enhancement processing on the initial graph structure data based on a multi-head attention mechanism to obtain an attention-enhanced graph; fourth, in the graph convolutional layer, perform graph convolution operations on the attention-enhanced graph to output a graph node feature embedding representation matrix; and fifth, generate the variable association matrix based on the graph node feature embedding representation matrix.

[0012] Furthermore, the specific steps for generating the structural equation model are as follows: the initial structural equation model is iteratively optimized using the adaptive learning rate gradient descent algorithm to obtain an intermediate optimized structural equation model; the goodness-of-fit test is performed on the intermediate optimized structural equation model; and the initial structural equation model is automatically corrected based on the goodness-of-fit test results to generate the structural equation model.

[0013] Furthermore, the specific steps for generating a visualized causal relationship graph are as follows: Based on the structural equation model, the causal effect value set is extracted using the causal effect separation method; based on the structural equation model and the causal effect value set, a causal network graph is constructed with variables as nodes and causal effects as directed edge weights; the causal network graph is visualized, rendered, and labeled to generate a visualized causal relationship graph.

[0014] Further, the specific steps for extracting the causal effect value set are as follows: The propensity score matching method is used to control causal interference in the structural equation model to obtain a balanced sample set; based on the balanced sample set and combined with mediation effect analysis, the causal effect decomposition of the structural equation model is performed to obtain an initial causal effect value set; the initial causal effect value set is subjected to effect value significance testing to generate the causal effect value set.

[0015] Furthermore, the specific steps for outputting the simulation result report are as follows: Build a multi-agent simulation infrastructure and define the types and core attributes of the home intelligent agent and the regional digital financial service intelligent agent; construct an intelligent agent interaction rule base based on a visualized causal relationship graph; run the simulation in a preset simulation scenario based on the types and core attributes of the home intelligent agent and the regional digital financial service intelligent agent, and in conjunction with the intelligent agent interaction rule base, to generate a simulation result report.

[0016] Compared with existing technologies, the beneficial effects of this invention are as follows: Through collaborative analysis between modules, intelligent in-depth mining of data related to digital finance and household consumption, as well as accurate identification and separation of causal effects, can be achieved to ensure the reliability of relationship mining. The data acquisition and preprocessing module cleans and normalizes the multi-regional digital finance sample association data and household consumption sample dataset to generate standardized datasets, providing a high-quality data foundation. With the help of the nonlinear feature extraction of convolutional neural networks and the variable topology relationship mining capabilities of graph neural networks in the structural equation model building module, the initial structural equation model is automatically constructed and iterative optimization and fitting verification are completed to achieve intelligent mining of complex relationships between variables. At the same time, the causal relationship inference module adopts a causal effect separation method that integrates propensity score matching and mediation effect analysis to effectively control causal interference, complete causal effect decomposition and significance testing, and generate a visual map containing accurate causal links, providing solid causal relationship support for subsequent simulation analysis and decision-making.

[0017] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0018] Figure 1 This is a block diagram of the digital finance and household consumption relationship pattern mining and simulation system of the present invention.

[0019] Figure 2 This is a flowchart illustrating the specific steps involved in constructing an initial structural equation model within the digital finance and household consumption relationship pattern mining and simulation system of this invention. Detailed Implementation

[0020] Please see Figure 1This invention provides a technical solution: a system for mining and simulating the relationship between digital finance and household consumption, comprising: a data acquisition and preprocessing module for acquiring a sample dataset, which includes digital finance sample association data and household consumption sample dataset for each region. The household consumption sample dataset includes household consumption status data (time dimension can be quarterly) and digital (financial) service usage data (matching the time dimension of the household consumption status data), and performs preprocessing to generate a standardized dataset; a structural equation model construction module for performing association mining processing on the standardized dataset (i.e., the standardized dataset) using a preset convolutional neural network and graph neural network to construct an initial structural equation model, and performing iterative optimization and fitting verification processing to generate a structural equation model; a causal association inference module for performing causal interference control and identification processing on the structural equation model using a causal effect separation method that integrates propensity score matching and mediation effect analysis, and generating a visualized causal association map; and a digital finance and household consumption simulation module for performing simulation processing on a preset simulation scenario based on the structural equation model and the visualized causal association map, and outputting a simulation result report.

[0021] The digital finance sample data specifically comprises several digital finance variables, which may include digital credit subsidy standards, digital credit interest rate adjustment ranges, financial literacy coverage, digital financial service outlet distribution density, and digital payment infrastructure coverage. Household consumption status data includes several household consumption variables, which may include total quarterly household consumption expenditure, expenditure amounts and percentages for categories such as food / clothing / housing / education / culture / healthcare, consumption frequency, online consumption percentage, and precautionary savings amount. Digital financial service usage data specifically comprises several digital service usage variables, which may include digital payment transaction amounts and frequency, digital credit loan amounts / frequency / repayment records, digital wealth management product holdings and transaction records, and digital financial app usage time and login frequency.

[0022] Specifically, the steps for generating standardized datasets are as follows: Data cleaning is performed on the digital finance sample association data and household consumption sample dataset for each region. Specifically, for missing parameters in the digital finance sample association data, spatial imputation is prioritized, i.e., using the mean or median of values ​​from regions with similar economic development levels or geographically adjacent regions during the same period. If no suitable reference region is available, forward or backward imputation of the region's indicators in the time series is used. For randomly missing numerical variables (such as missing consumption expenditure in a certain quarter) in household consumption status data and digital financial service usage data, a multiple imputation method based on chain equations is used. Specifically, other complete household attribute variables (such as household income, assets, and population structure) and consumption behavior variables in the dataset are used as predictor variables to generate multiple reasonable estimates for each missing value, forming multiple complete imputation datasets.

[0023] The data was then normalized for the digital finance sample association data and household consumption sample dataset for each region after missing data imputation, in order to generate a standardized dataset. This involved classifying the fields of the cleaned digital finance sample association data and household consumption sample dataset, clarifying the data type of each field. Continuous numerical data refers to numerical data with continuously changing values, such as digital credit subsidy standards, digital financial service outlet distribution density, total quarterly household consumption expenditure, digital payment transaction amount, digital credit loan amount, and digital wealth management product holding amount. Discrete numerical data refers to numerical data with discrete counts, such as consumption frequency, digital finance APP login frequency, financial literacy coverage rate (percentage discrete value), and online consumption ratio (percentage discrete value).

[0024] For continuous numerical data, the Z-score normalization algorithm is used for normalization. For discrete numerical data, the Min-Max normalization algorithm is used for normalization. The normalized digital financial sample data of each region, as well as the household consumption status data and digital service usage data of each household, are used as a standardized dataset.

[0025] like Figure 2 As shown, the specific steps for constructing the initial structural equation model are as follows: Based on the pre-set convolutional neural network, nonlinear feature extraction processing is performed on the standardized dataset to obtain a feature association set (representing the potential association between digital financial variables, household consumption variables and digital service usage variables);

[0026] Based on a pre-defined graph neural network, the topological relationships between variables in the feature association set are mined to generate a variable association matrix. Based on this matrix, a variable relationship topology graph is constructed with digital finance variables, household consumption variables, and digital service usage variables as nodes. Specifically:

[0027] Set a threshold for association strength (the preset threshold is 0.6, which can be flexibly adjusted according to the accuracy requirements of variable association mining, with an adjustment range of 0.5-0.7). Traverse all elements in the variable association matrix and extract variable pairs whose matrix element values ​​are ≥ the preset association strength threshold: if the matrix element values ​​corresponding to variable A and variable B are ≥ 0.6, it indicates that there is a strong association between the two, and the connection relationship between these two variable nodes is retained; otherwise, the corresponding variable pair connection relationship is removed.

[0028] The selected variable nodes are categorized and labeled: digital finance variable nodes, household consumption variable nodes, and digital service usage variable nodes are labeled with different symbols (such as color and shape, with digital finance variable nodes using blue circles, household consumption variable nodes using green squares, and digital service usage variable nodes using orange triangles) to clarify the variable type attributes of the nodes and make it easier to intuitively distinguish the core variable dimensions.

[0029] Based on the filtered variable pair connections and node classification labels, the basic structure of the topology graph is constructed: each variable is an independent node, and the retained strong associations are the connecting edges between nodes. The corresponding association strength values ​​in the variable association matrix are labeled on one side of each edge, forming the initial topology structure of nodes and edges with strength labels.

[0030] The nodes are arranged according to the logical partitioning of "digital financial variable cluster - digital service usage variable cluster - household consumption variable cluster", so that the correlation edges between digital financial variables and digital service usage variables, and between digital service usage variables and household consumption variables are presented in a concentrated manner. The topology clearly shows the distribution of the three types of variable nodes, the strong correlation between nodes, and the correlation strength. Among them, the core correlation links (such as the digital credit interest rate adjustment range and digital credit loan amount, digital wealth management product holding amount and education and culture consumption ratio) are highlighted with high brightness edges.

[0031] The variable relationship topology is mapped to the initial structural equation model. Specifically, the digital finance variables, household consumption variables, and digital service usage variables in the variable relationship topology are mapped to the observed variables in the initial structural equation model. At the same time, latent variables are extracted based on the common dimensions of the variables. Specifically, the latent variable of "digital finance development level" (covering observed variables such as digital credit subsidy standards and the distribution density of digital financial service outlets) is extracted from the digital finance variables; the latent variable of "digital finance usage depth" (covering observed variables such as the frequency of digital payment transactions and the amount of digital wealth management product holdings) is extracted from the digital service usage variables; and the latent variable of "upgrading of household consumption structure" (covering observed variables such as the proportion of education and culture consumption expenditure and the proportion of online consumption) is extracted from the household consumption variables.

[0032] The connection edges between nodes in the variable relationship topology graph are mapped to the path relationships between variables in the initial structural equation model: the correlation edge between the digital finance variable and the digital service usage variable is mapped to the driving path of the potential variable "digital finance development level" on the potential variable "digital finance usage depth"; the correlation edge between the digital service usage variable and the household consumption variable is mapped to the transmission path of the potential variable "digital finance usage depth" on the potential variable "household consumption structure upgrading"; and the direct correlation edge between the digital finance variable and the household consumption variable (if it exists) is mapped to the direct influence path of the potential variable "digital finance development level" on the potential variable "household consumption structure upgrading".

[0033] The correlation strength values ​​of the edges in the topological graph of variable relationships are directly used as the initial path coefficients of the corresponding path relationships and assigned to the initial structural equation model. For example, if the correlation strength between "digital credit interest rate adjustment range" and "digital credit loan amount" in the topological graph is 0.85, then this value is used as the initial coefficient of the path corresponding to the potential variable "digital finance development level" pointing to the potential variable "digital finance usage depth".

[0034] Based on the above mapping results, a complete framework for the initial structural equation model is constructed, including: measurement equations, defining the attribution relationship between latent variables and corresponding observed variables (e.g., the latent variable "level of digital finance development" is measured by observed variables such as digital credit subsidy standards and financial literacy coverage); structural equations, defining the path relationships and initial coefficients between latent variables (e.g., the chain structural equation from "level of digital finance development to depth of digital finance use to upgrading of household consumption structure"); the model includes observed variables, latent variables, path relationships between variables, and initial path coefficients.

[0035] The specific steps for generating a structural equation model are as follows: The initial structural equation model is iteratively optimized using an adaptive learning rate gradient descent algorithm to obtain an intermediate optimized structural equation model. Specifically, the core parameters for iterative optimization are configured as follows: The initial value of the adaptive learning rate can be set to 0.01 (which can be dynamically adjusted according to the model's convergence speed, with an adaptation range of 0.001-0.05), the upper limit of the number of iterations is set to 1000 (to avoid overfitting due to excessive iteration), and the loss function is the mean squared error (MSE) function, which quantifies the deviation between the predicted value output by the model and the actual observed value of the standardized dataset.

[0036] Based on the adaptive learning rate gradient descent algorithm, the partial derivative of the loss function with respect to each path coefficient in the model (i.e., the gradient) is first calculated. The direction of this gradient reflects the direction of adjustment of the path coefficients (adjusting along the negative gradient direction can reduce the loss value). The path coefficients are then initially corrected according to the gradient direction, and the correction magnitude is determined by the current learning rate.

[0037] Monitor the trend of the loss function in real time. If the decrease in the loss value is ≥10% in three consecutive iterations, it indicates that the current learning rate is well-suited. Increase the learning rate to 1.2 times the current value to accelerate the optimization convergence speed. If the loss value increases in a certain iteration, or the decrease in the loss value is <0.5% in two consecutive iterations, it indicates that the learning rate is too high and the convergence is unstable. Reduce the learning rate to 0.8 times the current value to ensure that the model converges stably towards the minimum value of the loss function.

[0038] After each iteration, record the current loss value and path coefficient. When the number of iterations reaches 50, if the loss value is stable below 0.05 (i.e. the model prediction bias is small), the iteration can be terminated early, and the current optimized model can be output as an intermediate optimized structural equation model. If the condition is not met, the iteration continues until the upper limit of 1000 iterations is reached, at which point the iteration is forcibly terminated and the intermediate optimized structural equation model is output.

[0039] The intermediate optimized structural equation model was subjected to a goodness-of-fit test, which involved selecting key indicators for the structural equation model test, including root mean square approximation error (RMSEA), comparison fit index (CFI), and root mean square standardized residual (SRMR). RMSEA is used to measure the approximate fit of the model, CFI is used to compare the fit advantage of the model with the baseline ineffective model, and SRMR is used to measure the magnitude of the residuals between the observed data and the model's predicted data.

[0040] Set the goodness-of-fit criteria: Referring to industry standards and the application scenarios of this system, the acceptable thresholds can be set as follows: RMSEA≤0.08 (the smaller the value, the better the model fit), CFI≥0.9 (the closer the value is to 1, the better the model fit), and SRMR≤0.06 (the smaller the residual, the stronger the model fit with the data). All three indicators must meet the threshold requirements simultaneously to determine that the model fit is acceptable.

[0041] Based on standardized datasets, the specific values ​​of RMSEA, CFI, and SRMR of the intermediate optimized structural equation model are calculated using specialized structural equation modeling tools (such as LISREL and AMOS); at the same time, the residual matrix of the model is output to help determine the source of the model-data fit deviation.

[0042] The calculated index values ​​are compared with the preset pass standards. If all three indexes meet the threshold requirements, the model fit is deemed to be qualified; if any index does not meet the requirements, the model fit is deemed to be unqualified.

[0043] Based on the goodness-of-fit test results, the initial structural equation model is automatically corrected to generate a new structural equation model, specifically as follows:

[0044] If the goodness of fit is satisfactory: Perform a significance test on the path coefficients of the intermediate optimized structural equation model using the t-test method, with a significance level of 0.05 (i.e., 95% confidence level). Calculate the t-value for each path coefficient. If the t-value is ≥1.96, it indicates that the path coefficient is statistically significant (i.e., the corresponding path relationship between variables truly exists), and the path is retained. If the t-value is <1.96, it indicates that the path coefficient is not statistically significant, and the corresponding path relationship is automatically removed. After removal, recalculate the model's goodness of fit. If it still meets the acceptable criteria, the model can be output as the final structural equation model.

[0045] If the goodness of fit is unsatisfactory: Activate the automated correction mechanism. The specific correction logic is as follows: Refer to the residual matrix to locate the variable pairs with large residuals (i.e., the model has a large prediction bias for this variable pair). Prioritize checking the path relationship between the corresponding variables. If there are redundant paths (such as bidirectional paths between two variables), automatically delete one of the paths. If there are no redundant paths, adjust the attribution relationship between the latent variables and the observed variables (e.g., change "digital payment infrastructure coverage" from an observed variable of the latent variable "digital financial development level" to an auxiliary observed variable). Reinitialize the path coefficients that failed the significance test, setting their initial value to 0.05 (1 / 2 of the original initial value). Return to the "iterative optimization step" and re-execute iterative optimization based on the corrected initial structural equation model to obtain a new intermediate optimized structural equation model. Perform a goodness of fit test on the new intermediate optimized structural equation model again, and repeat the above correction-optimization-test process until the model's goodness of fit meets the qualified standard.

[0046] The corrected, optimized, and verified structural equation model must include all core elements: verified relationships between observed and latent variables, significant path relationships between variables (such as "digital finance development level to digital finance usage depth" and "digital finance usage depth to household consumption structure upgrade"), and optimized optimal path coefficients (quantifying the influence strength of each path). At the same time, it outputs the model fit index value and the significance test results of the path coefficients to generate the structural equation model.

[0047] In this implementation plan, through refined data processing and intelligent model construction optimization, high-quality and reliable core support is provided for the mining of the relationship between digital finance and household consumption. In the data preprocessing stage, differentiated imputation methods are adopted for different types of missing data characteristics, combined with data type accurate matching and normalization algorithms, to effectively improve data integrity and consistency and lay a high-quality data foundation. In the model construction stage, convolutional and graph neural networks are used to automatically mine the nonlinear features and topological relationships of variables. With the help of threshold screening and logical partitioning, core correlation links are accurately captured. Furthermore, the initial model is systematically constructed through variable mapping to reduce subjective bias. In the model optimization and verification process, the adaptive learning rate gradient descent algorithm ensures convergence efficiency and stability. Multi-index goodness-of-fit testing combined with an automated correction mechanism accurately eliminates invalid paths and optimizes the model structure. The final output model contains complete verification elements, ensuring model reliability and providing accurate and solid model support for subsequent causal inference and simulation analysis.

[0048] Specifically, a convolutional neural network includes an input layer that receives a standardized dataset, several convolutional layers, a normalization layer, a max-pooling layer, a fusion layer, and an output layer. The specific steps for generating a feature association set are as follows:

[0049] The standardized dataset is input into a pre-defined convolutional neural network. (Within the convolutional layers) sliding convolution is performed on the standardized dataset to extract local nonlinear features between digital financial variables, household consumption variables, and digital service usage variables. This yields a local nonlinear feature set (i.e., by sliding the 3×3 convolutional kernels of each convolutional layer along the width and height of the feature matrix with a stride of 1, performing convolution operations on the variable features within each sliding window. This process extracts features layer by layer through the stacking of multiple convolutional kernels: shallow convolutional layers capture the basic nonlinear features of digital financial variables, household consumption variables, and digital service usage variables (such as the numerical trend of a single variable); deep convolutional layers capture the complex correlation features between the three types of variables (such as the relationship between the frequency of digital credit use and household education consumption). The algorithm identifies the positive correlation characteristics of expenditures and the coupling characteristics between digital payment amounts and the proportion of online consumption. The final output is a set of local nonlinear features containing multi-dimensional local nonlinear characteristics, including but not limited to the regional clustering characteristics of the distribution density of digital financial service outlets, the marginal growth characteristics of total quarterly household consumption expenditures, the rigid fluctuation characteristics of food / health care consumption expenditures, the intraday peak distribution characteristics of digital payment transaction amounts, the linkage characteristics between the frequency of digital credit borrowing and the timeliness of repayment, the positive correlation characteristics between the proportion of online consumption and the usage time of digital financial APPs, and the preliminary negative correlation characteristics between the amount of precautionary savings and the amount of digital wealth management product holdings. The algorithm then performs batch normalization processing (i.e., eliminates the dimensional differences between different features through batch normalization algorithms) in the normalization layer.

[0050] (In the max pooling layer) the batch-normalized local nonlinear feature set is downsampled to obtain a dimensionality-reduced feature set (i.e., sliding through a 2×2 pooling window with a stride of 2, taking the maximum value of the feature value within each window); based on the skip connection mechanism, the local nonlinear feature set and the dimensionality-reduced feature set are fused and flattened (i.e., fusion is performed in the fusion layer, and flattening is performed in the output layer), outputting a feature association set, specifically:

[0051] Based on the skip connection mechanism, the fusion layer fuses the original local nonlinear feature set output by the convolutional layer with the dimensionality-reduced feature set output by the max pooling layer: first, the dimension of the local nonlinear feature set is adjusted to match that of the dimensionality-reduced feature set using a 1×1 convolutional kernel, and then residual addition is used to achieve feature fusion. This fusion process avoids the vanishing gradient of features in deep networks, while preserving the detailed information of the original local features and the core correlation information of the dimensionality-reduced features, strengthening the representation ability of potential correlations between the three types of variables. The fused features are then transmitted to the output layer.

[0052] By flattening the multidimensional feature tensors into one-dimensional feature vectors, the final output is a set of feature associations representing the potential relationships between digital finance variables, household consumption variables, and digital service usage variables. These include, but are not limited to: positive driving association features between digital finance policy variables (digital credit interest rate adjustment range, financial literacy coverage rate) and digital service usage variables (digital credit loan amount, digital payment transaction frequency); mediating transmission association features between digital service usage variables (digital wealth management product holding amount, digital finance APP usage time) and household consumption variables (proportion of education and culture consumption expenditure, proportion of online consumption); regional matching association features between the distribution density of digital financial service outlets and household medical and health consumption expenditure; and non-linear association features between digital credit usage and housing consumption expenditure moderated by household structure (implied in the proportion of categorized expenditures of household consumption variables).

[0053] The specific steps for generating the variable association matrix are as follows: First, combine the standardized dataset and the feature association set to construct an initial graph structure data. This involves defining each variable in the standardized dataset as a graph node. This includes all digital financial variables (such as the digital credit interest rate adjustment range and financial literacy coverage), all household consumption variables (such as total quarterly household consumption expenditure and the proportion of education and cultural consumption), and all digital service usage variables (such as the frequency of digital payment transactions and the amount of digital credit loans). Each variable becomes an independent node in the graph. Second, based on the potential associations between variables revealed in the feature association set, establish a connection edge between any two variable nodes. Use the specific numerical values ​​representing the strength of the association between each pair of variables in the feature association set to initialize the weights of the corresponding edges.

[0054] The initial graph structure data is input into a pre-defined graph neural network, which includes a graph attention layer and a graph convolutional layer. The network is configured to: receive the initial graph structure data as input; dynamically calculate and assign weights to the edges in the graph based on a multi-head attention mechanism through the graph attention layer to generate an attention-enhanced graph; perform at least one graph convolution operation on the attention-enhanced graph through the graph convolutional layer to update the embedding representation of each node; and output the updated embedding representation of each node.

[0055] In the graph attention layer, the initial graph structure data is enhanced by a multi-head attention mechanism to obtain an attention-enhanced graph. Specifically, the number of attention heads in the multi-head attention mechanism is configurable to 8 (which can be flexibly adjusted according to the total number of variables, with an adaptation range of 4-16 heads). Through linear transformation, the feature information of each node is split into 8 groups of feature descriptions in different dimensions to ensure that the relationship details between variables in different dimensions can be fully captured.

[0056] For each set of split feature descriptions, the attention score between any two nodes with an initial connection edge is calculated (i.e., the contribution of the feature information of one node to the associated feature expression of another node). During the calculation, the feature descriptions of the two nodes are first linearly transformed, and then the transformation results are combined and processed by an activation function to finally obtain the attention score representing the association contribution.

[0057] The attention scores are filtered and restricted by combining the connection edge information in the initial graph structure: only the attention scores of node pairs with initial connection edges are retained, and the attention scores of node pairs without initial edges (i.e. variable pairs with no potential association in the feature association set) can be set to a minimum value to avoid variables with no actual association interfering with the mining of core associations.

[0058] The filtered attention scores are normalized, mapping all scores to the range of 0-1 to obtain normalized attention weights. Based on the normalized attention weights, the feature information of each node is updated: each attention head collects and integrates the feature information of all related nodes in the neighborhood of the corresponding node according to its own calculated weight, generating a set of updated node features. Then, the node features output by the 8 attention heads are integrated and spliced ​​to obtain the final node features that fuse multi-view association information.

[0059] Synchronously update the weights of the connecting edges between nodes: take the average of the normalized weights between the same pair of nodes calculated by the 8 attention heads, and use it as the dynamic weight of the edge between the two nodes to replace the initial edge weight. The final attention-enhanced graph contains the updated node features and dynamic edge weights, which can significantly strengthen the weights of core associations such as digital financial variables and digital service usage variables, and digital service usage variables and household consumption variables.

[0060] In the graph convolutional layer, at least one graph convolution operation is performed on the attention enhancement graph to output the graph node feature embedding representation matrix. Specifically, the graph convolutional layer can be configured to have 2 layers. Graph convolution operation is used to realize the deep aggregation and propagation of node features, so that the features of each node fully integrate the information of all its related nodes and accurately represent the complex relationship between variables.

[0061] Perform the first layer of graph convolution operation: normalize the dynamic edge weights of the attention-enhanced graph, preserving the feature information of each node during the normalization process (to avoid dilution of its own features); then fuse the normalized edge weights with the input node features, adjust the feature dimensions by combining preset weight parameters and bias parameters, and finally process them through the activation function to obtain the node features after the first layer of convolution (at this point, the feature information of directly related nodes has been fused).

[0062] Perform the second layer graph convolution operation: Based on the output features of the first layer, the dynamic edge weights are normalized again. The normalized edge weights are then fused with the node features output from the first layer. The feature dimensions are adjusted by combining new weight parameters and bias parameters. After processing with an activation function, the node features after the second layer convolution are obtained. This achieves the aggregation of indirectly related node features (i.e., the fusion of full-link features of nodes, directly related nodes, and indirectly related nodes), allowing node features to more comprehensively reflect the deep relationships between variables.

[0063] The feature set output by the second convolution operation is the graph node feature embedding representation matrix. In this matrix, each entry corresponds to the fusion association feature of a variable node, which can accurately represent the comprehensive association information between the variable and all other variables (including directly associated and indirectly associated variables).

[0064] Based on the graph node feature embedding representation matrix, a variable association matrix is ​​generated. Specifically, the fusion association feature information corresponding to each variable node is extracted one by one from the graph node feature embedding representation matrix. Each feature information entry uniquely corresponds to a digital financial variable, a household consumption variable, or a digital service usage variable.

[0065] Calculate the similarity between the fused associated feature information corresponding to any two variable nodes, and use this similarity as the association strength between the two variables; specifically, the vector similarity judgment method is adopted: calculate the sum of the corresponding numerical products of the two feature information entries, and then divide it by the product of the numerical moduli of the two feature information entries to obtain the similarity score;

[0066] According to the preset sequence number of the variable nodes (such as the order from digital finance variable to digital service usage variable to household consumption variable), the association strength scores of all variable pairs are organized into a matrix. The number of rows and columns of the matrix are consistent with the total number of variables. Each value in the matrix is ​​the association strength score of the corresponding two variables.

[0067] Normalize all correlation strength scores in the matrix, mapping all scores to the range of 0-1. In the final variable correlation matrix, the closer the value is to 1, the stronger the correlation between the two corresponding variables (such as the digital credit interest rate adjustment range and the amount of digital credit loans, and the amount of digital wealth management product holdings and the proportion of education and cultural consumption).

[0068] In this implementation scheme, the synergistic interaction of convolutional neural networks (CNNs) and graph neural networks (GNNs) enables the accurate extraction of features related to digital finance and household consumption, and the in-depth mining of their correlations. This provides high-quality core support for subsequent model construction. The CNNs extract features layer by layer using multi-layer 3×3 convolutional kernels, accurately capturing basic trends and complex correlation features. Batch normalization eliminates dimensional interference, and the skip connection mechanism effectively avoids gradient vanishing while preserving details and core information, thus improving the comprehensiveness and reliability of feature representation. The GNNs mine correlation details from multiple perspectives through a multi-head attention mechanism, dynamically updating edge weights to strengthen core correlations and eliminate irrelevant interference. Two layers of graph convolution operations achieve full-link fusion of direct and indirect correlation features, allowing node features to fully reflect deep correlations. The resulting variable correlation matrix is ​​standardized to quantify correlation strength, arranged in an orderly manner according to variable type, and accurately locates core correlation links, laying a solid foundation for subsequent variable relationship topology graph construction and model mapping.

[0069] Specifically, the steps for generating a visual causal relationship map are as follows: Based on the structural equation model, the causal effect separation method is used to extract the causal effect value set (the total effect of digital financial variables on household consumption variables, the direct effect, and the indirect effect transmitted through the variable of digital service use).

[0070] Based on structural equation modeling and causal effect value sets, a causal network graph is constructed with variables as nodes and causal effects as directed edge weights. Specifically, the nodes of the causal network graph are the validated observed and latent variables from the structural equation model. The observed variables include all digital finance variables (such as the adjustment range of digital credit interest rates and the coverage rate of financial literacy), household consumption variables (such as total quarterly household consumption expenditure and the proportion of education and cultural consumption), and digital service usage variables (such as the frequency of digital payment transactions and the amount of digital wealth management product holdings). The latent variables include "the level of digital finance development", "the depth of digital finance use", and "the upgrading of household consumption structure". Each node uniquely corresponds to one type of variable, with no duplicate or redundant nodes.

[0071] Digital finance variable nodes are marked with blue circles, household consumption variable nodes with green squares, and digital service usage variable nodes with orange triangles. At the same time, the word "potential" is marked on the right side of potential variable nodes to clearly distinguish between observed variables and potential variables and avoid confusion.

[0072] The path relationships corresponding to the directed edges in the causal effect value set are constructed as follows: the path corresponding to the direct effect value, "from the level of digital finance development to the upgrading of household consumption structure", is constructed as a direct causal directed edge; the path corresponding to the indirect effect value, "from the level of digital finance development to the depth of digital finance use to the upgrading of household consumption structure", is constructed as an indirect causal directed edge; only the significant paths that have passed the significance test (t-value ≥ 1.96) are retained, and the paths corresponding to non-significant effects are removed.

[0073] The direct and indirect effect values ​​in the causal effect value set are used as the weights of the corresponding directed edges, and the weight values ​​are completely consistent with the effect values. At the same time, the line type is distinguished according to the path type—direct causal directed edges are represented by solid lines, and indirect causal directed edges are represented by dashed lines. The thickness of the edge is positively correlated with the absolute value of the optimal path coefficient output by the structural equation model (the larger the absolute value of the optimal path coefficient, the wider the edge).

[0074] The causal network diagram is visualized, rendered, and labeled to generate a visualized causal relationship map, specifically as follows:

[0075] The nodes are arranged hierarchically according to the logical link from "digital financial variable cluster to digital service usage variable cluster to household consumption variable cluster", so that the direct causal path and the indirect causal path are presented together along this link, avoiding edge crossing and node overlap; at the same time, the latent variable node is placed in the center of the corresponding observed variable cluster, highlighting the dominant relationship between the latent variable and the observed variable.

[0076] Supplementing core information annotation: Two sets of key information are annotated on each directed edge—the optimal path coefficient output by the structural equation model (rounded to two decimal places) and the corresponding causal effect value (direct effect value or indirect effect value, rounded to two decimal places); a description of causal interference control is annotated below the graph, clearly specifying the core confounding variables (regional economic development level, household disposable income, population size) and the test criteria for the balanced sample set (t-value < 1.96); the total effect value summary formula (total effect value = direct effect value + indirect effect value) is also annotated to ensure the completeness of the graph information;

[0077] Add standardized legends: The legends include three core elements—node labeling: clearly defining the meaning of blue circles (digital financial variables), green squares (household consumption variables), orange triangles (digital service usage variables), and the "potential" label; line type: solid lines correspond to direct causal effects, and dashed lines correspond to indirect causal effects; edge thickness: divided into three levels based on the absolute value of the optimal path coefficient (below 0.3 is thin, 0.3-0.7 is medium, and above 0.7 is thick).

[0078] The rendered and labeled causal relationship map can intuitively present the significant causal paths between digital finance variables, digital service usage variables, household consumption variables and potential variables, the optimal path coefficient and causal effect value of each path, and the differences in the intensity of path influence. It can clearly show the core transmission link from "the level of digital finance development to the depth of digital finance use to the upgrading of household consumption structure".

[0079] The specific steps for extracting the causal effect value set are as follows: Propensity score matching is used to control for causal interference in the structural equation model, resulting in a balanced sample set, specifically:

[0080] Based on the analysis scenario of the relationship between digital finance and household consumption, regional economic development level, household disposable income, and population size are selected as core confounding variables for causal interference control. Relevant data for these variables can be obtained by supplementing them with publicly available statistical data (such as regional GDP and resident income survey data) or related fields of household consumption sample datasets. After supplementation, they are integrated into the standardized dataset so that each sample (i.e., each household record) in the standardized dataset contains the values ​​of these variables, ensuring data integrity. The core purpose of selecting these three types of variables is to eliminate the interference of non-digital finance factors. These three types of variables simultaneously affect digital finance variables (such as the distribution density of digital financial service outlets and the amount of digital credit loans) and household consumption variables (such as total quarterly household consumption expenditure and the proportion of education and cultural consumption). If their interference is not controlled, it will be difficult to accurately distinguish between "the real impact of digital finance on household consumption" and "the impact of external factors on household consumption".

[0081] Using "whether the sample is affected by digital financial variables" as the dependent variable (if the household corresponding to the sample uses digital financial services such as digital credit and digital wealth management, the dependent variable takes the value of 1, otherwise it is 0), and the core confounding variable as the independent variable, a logistic regression model is constructed; the core confounding variable values ​​of each sample in the standardized dataset are substituted into the model to calculate the propensity score of each sample (representing the probability that the sample is affected by digital financial variables).

[0082] Using a 1:1 nearest neighbor matching rule, for each "sample affected by digital financial variables" in the standardized dataset, one "sample not affected by digital financial variables" with the closest propensity score is matched to form an initial matching sample set;

[0083] The initial matched sample set is subjected to a balance test using a t-test (with a significance level set at 0.05, consistent with the path coefficient significance test standard mentioned earlier) to verify whether there is a significant difference in the distribution of the core confounding variables in the two groups of samples after matching. If the t-values ​​for all confounding variables are <1.96, it indicates that there is no significant difference in the confounding variables between the two groups of samples, and the confounding factors have been effectively controlled. This sample set is the balanced sample set. If there are confounding variables with t-values ​​≥1.96, the matching rules are readjusted (e.g., the matching radius is expanded) and the matching process is repeated until the balance test is passed.

[0084] Based on the balanced sample set and combined with mediation effect analysis, the structural equation model is decomposed into causal effects to obtain the initial causal effect value set, which is as follows:

[0085] From the structural equation model, the optimal path coefficients corresponding to significant causal paths are extracted, including the direct path coefficients from "the level of digital finance development to the upgrading of household consumption structure", the pre-mediation path coefficients from "the level of digital finance development to the depth of digital finance use", and the post-mediation path coefficients from "the depth of digital finance use to the upgrading of household consumption structure". At the same time, the relationship between observed variables and latent variables in the model is clarified to ensure that the dimensions of the corresponding variables in the effect decomposition are consistent.

[0086] Substituting the numerical values ​​of variables in the balanced sample set into the structural equation model, and combining them with the extracted optimal path coefficients, the causal effect is decomposed through mediation effect analysis: Direct effect value: The optimal path coefficient of "digital finance development level to household consumption structure upgrade" is used as the direct effect value of the digital finance variable on the household consumption variable, representing the strength of the direct influence between the two without transmission through the digital service usage variable; Indirect effect value: The optimal path coefficients of "digital finance development level to digital finance usage depth" and "digital finance usage depth to household consumption structure upgrade" are multiplied, and the result is used as the indirect effect value of the digital finance variable transmitted to the household consumption variable through the digital service usage variable; Total effect value: The direct effect value and the indirect effect value are added together to obtain the total effect value of the digital finance variable on the household consumption variable, representing the overall causal influence strength of digital finance on household consumption;

[0087] The initial causal effect value set was processed to perform effect value significance testing to generate a new causal effect value set. Specifically, the t-test method (with a significance level set to 0.05, consistent with the path coefficient significance test standard mentioned above) was used to test the significance of the above direct effect value, indirect effect value, and total effect value. If the t-value corresponding to the effect value is ≥1.96, it indicates that the effect is statistically significant and the effect value is retained; otherwise, it is removed.

[0088] To form a causal effect value set: The direct effect value, indirect effect value, and total effect value that have been tested to be significant are organized and summarized in the format of "Effect type (direct / indirect / total effect) - corresponding variable path - effect value - significance test result" to form a complete causal effect value set.

[0089] This implementation plan achieves a clear and precise representation of the causal relationship between digital finance and household consumption through accurate causal effect extraction and visualization. This provides an intuitive and reliable core basis for decision analysis. When extracting the causal effect value set, bias score matching effectively controls the interference of confounding variables such as regional economy and household income, and generates a balanced sample set to ensure the authenticity of the effect. Mediation effect decomposition accurately distinguishes direct, indirect, and total effects, and combines t-tests to screen significant effects, ensuring the scientific nature of the effect values. In the graph generation process, nodes are differentiated according to variable type and arranged hierarchically along the core logical link to avoid confusion. By distinguishing path types and influence intensity through line type and edge thickness, and supplementing complete information such as path coefficients and test criteria, as well as standardized legends, the core transmission link and effect differences are presented intuitively, making complex causal relationships clear at a glance, facilitating quick grasp of the core logic, and providing accurate causal support for subsequent simulations.

[0090] Specifically, the steps for outputting the simulation results report are as follows: Establish a multi-agent simulation infrastructure, define the types and core attributes of the household agent and the regional digital financial service agent. The core attributes correspond to the digital financial variables, household consumption variables, and digital financial service usage variables in the standardized dataset, specifically:

[0091] The system is constructed by mapping each household in the standardized dataset to a one-to-one relationship. The core attributes include household consumption variables (total quarterly household consumption expenditure, amount and percentage of consumption expenditure in categories such as food / clothing / education / culture / health care, consumption frequency, proportion of online consumption, and amount of precautionary savings) and digital financial service usage variables (amount and frequency of digital payment transactions, amount / frequency / repayment records of digital credit loans, amount and transaction records of digital wealth management products, and duration and frequency of use of digital financial apps). The initial values ​​of the attributes are directly taken from the normalized values ​​of the corresponding households in the standardized dataset.

[0092] Define the regional digital financial service agent: It is constructed according to the regional dimension of the digital financial sample associated data. The core attributes are digital financial variables (digital credit subsidy standard, digital credit interest rate adjustment range, financial literacy coverage rate, digital financial service outlet distribution density, digital payment infrastructure coverage rate). The initial values ​​of the attributes are taken from the normalized values ​​of the digital financial variables in the corresponding region in the standardized dataset.

[0093] Establish a basic framework for multi-agent simulation: Configure agent communication interfaces and data interaction channels to ensure that home agents and regional digital financial service agents can achieve real-time transmission and response of attribute data;

[0094] A rule base for agent interaction is constructed based on a visualized causal relationship graph (with significant causal links in it). The causal effect value obtained by causal effect separation is incorporated to quantify the intensity of the interaction influence between agents. Specifically, core causal links that have passed the significance test (t-value ≥ 1.96) are extracted from the visualized causal relationship graph, including "from the level of digital finance development to the depth of digital finance use", "from the depth of digital finance use to the upgrading of household consumption structure", and "from the level of digital finance development to the upgrading of household consumption structure".

[0095] The core logic of building the interaction rule base is to transform the extracted significant causal links into actionable rules for agent behavior—changes in the digital financial variable attributes of the regional digital financial service agent trigger adjustments in the digital financial service usage variable attributes of the household agent; adjustments in the digital financial service usage variable attributes of the household agent further trigger changes in its household consumption variable attributes.

[0096] Quantifying the intensity of interaction impact: The direct effect value, indirect effect value, and total effect value obtained by separating causal effects are integrated into the rule base to clarify the magnitude of behavioral changes corresponding to different effect values ​​(e.g., the larger the indirect effect value, the stronger the transmission impact of changes in the attributes of regional digital financial service agents on household agent consumption variables). At the same time, the optimal path coefficients of the structural equation model are associated to calibrate the quantitative benchmark of interaction impact.

[0097] Based on the types and core attributes of home intelligent agents and regional digital financial service intelligent agents, and combined with the intelligent agent interaction rule base, simulations are run in preset simulation scenarios to generate simulation result reports. Specifically, the preset gradient digital financial policy simulation scenarios include digital credit interest rate adjustment scenarios (increase / decrease by 0.5%, 1%, 1.5%), digital credit subsidy standard adjustment scenarios (increase / decrease by 2%, 5%, 8%), digital financial service outlet distribution density optimization scenarios (increase / decrease by 10%, 20%, 30%), and financial literacy coverage improvement scenarios (increase by 5%, 10%, 15%). At the same time, a control group scenario without policy adjustments is set up.

[0098] Simulation operation: Drive the regional digital financial service agent to adjust core attribute parameters according to preset scenarios, and trigger the dynamic changes of digital financial service usage behavior and household consumption behavior of the household agent through the interaction rule base; record simulation data on a quarterly basis (consistent with the time dimension of household consumption status data), including the consumption / use behavior trajectory of each household agent at the micro level, and the total household consumption expenditure, consumption structure upgrading range, and policy transmission efficiency data of each region at the macro level;

[0099] Simulation data verification: The mean squared error (MSE) is used to compare the simulation data with the predicted values ​​of the structural equation model. If MSE > 0.05 (consistent with the threshold of the loss function for model optimization mentioned above), the effect strength parameters of the agent interaction rule base are recalibrated and the simulation is repeated; if MSE ≤ 0.05, the simulation data is deemed valid.

[0100] Generate simulation results report: Integrate effective simulation data and organize the content according to the framework of "scenario description - micro data - macro data - effect analysis - decision-making suggestions". It includes policy scenario parameters, trends in household consumption indicators, efficiency of digital finance-consumption transmission, comparison of effects in different scenarios, model building efficiency and prediction accuracy data, and outputs the final simulation results report after visualization processing.

[0101] This implementation plan achieves dynamic simulation and accurate assessment of the impact of digital financial policies on household consumption through the construction of a multi-agent simulation architecture and full-process simulation analysis. It provides quantitative and actionable decision support. During the architecture construction phase, the core attributes of two types of agents are defined by a one-to-one mapping with real data to ensure the realism of the simulation scenario. An interaction channel is configured to ensure real-time data transmission, laying the foundation for efficient simulation. The interaction rule base is built based on significant causal links, incorporating causal effect values ​​and optimal path coefficients to quantify the impact intensity, making the agent's behavioral logic conform to the real correlation mechanism. By pre-setting gradient policy scenarios and control groups, micro-level household trajectories and macro-level regional data are recorded from multiple dimensions. The validity of the data is ensured by combining mean square error verification. Finally, the data is integrated to generate a structured report, which intuitively presents the differences in policy effects and transmission efficiency, quantifying the advantages of the model and providing clear and reliable quantitative basis for policy optimization.

[0102] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0103] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A system for mining and simulating the relationship between digital finance and household consumption patterns, characterized in that: include: The data acquisition and preprocessing module is used to acquire sample datasets, which include digital financial sample correlation data for each region and household consumption sample datasets. The household consumption sample datasets include household consumption status data and digital service usage data of several households. The module performs preprocessing to generate standardized datasets. The structural equation model building module is used to construct an initial structural equation model based on a standardized dataset by performing correlation mining on the dataset using a pre-defined convolutional neural network and graph neural network, and then perform iterative optimization and fitting verification to generate a structural equation model. The causal association inference module is used to perform causal interference control and identification processing on structural equation models by adopting the causal effect separation method that integrates propensity score matching and mediation effect analysis, and to generate a visual causal association map. The Digital Finance and Household Consumption Simulation Module is used to perform simulation processing in a preset simulation scenario based on structural equation modeling and visualized causal relationship graphs, and output simulation result reports.

2. The system for mining and simulating the relationship between digital finance and household consumption patterns according to claim 1, characterized in that, The digital financial sample associated data specifically comprises several digital financial variables; the household consumption status data includes several household consumption variables; and the digital financial service usage data specifically comprises several digital service usage variables.

3. The system for mining and simulating the relationship between digital finance and household consumption patterns according to claim 2, characterized in that, The specific steps for generating a standardized dataset are as follows: Data cleaning and processing were performed on the digital finance sample data and household consumption sample dataset for each region. The data associated with digital finance samples and household consumption samples in each region after missing data filling were normalized to generate a standardized dataset.

4. The system for mining and simulating the relationship between digital finance and household consumption patterns according to claim 3, characterized in that, The specific steps for constructing the initial structural equation model are as follows: Based on a pre-defined convolutional neural network, a non-linear feature extraction process is performed on a standardized dataset to obtain a feature association set; Based on a pre-defined graph neural network, the topological relationships between variables in the feature association set are mined to generate a variable association matrix. Based on the variable association matrix, a variable relationship topology graph is constructed with digital finance variables, household consumption variables, and digital service usage variables as nodes; Map the topological graph of variable relationships to an initial structural equation model.

5. The system for mining and simulating the relationship between digital finance and household consumption patterns according to claim 4, characterized in that, The specific steps for generating feature association sets are as follows: The standardized dataset is input into a pre-defined convolutional neural network. The standardized dataset is subjected to sliding convolution processing to extract local nonlinear features between digital financial variables, household consumption variables and digital service usage variables, resulting in a local nonlinear feature set, which is then batch normalized. The local nonlinear feature set after batch normalization is downsampled to obtain a dimension-reduced feature set. Based on the skip connection mechanism, the local nonlinear feature set and the dimensionality-reduced feature set are fused and flattened to output the feature association set.

6. The system for mining and simulating the relationship between digital finance and household consumption patterns according to claim 4, characterized in that, The specific steps for generating the variable correlation matrix are as follows: By combining standardized datasets and feature association sets, an initial graph structure data is constructed; The initial graph structure data is input into a pre-defined graph neural network, which consists of a graph attention layer and a graph convolutional layer. In the graph attention layer, the initial graph structure data is subjected to association enhancement processing based on the multi-head attention mechanism to obtain an attention-enhanced graph; In the graph convolutional layer, graph convolution operation is performed on the attention enhancement graph to output the graph node feature embedding representation matrix; Based on the graph node feature embedding representation matrix, a variable correlation matrix is ​​generated.

7. The system for mining and simulating the relationship between digital finance and household consumption patterns according to claim 4, characterized in that, The specific steps for generating a structural equation model are as follows: The initial structural equation model is iteratively optimized using an adaptive learning rate gradient descent algorithm to obtain an intermediate optimized structural equation model. And perform a goodness-of-fit test on the intermediate optimized structural equation model; The initial structural equation model is automatically corrected based on the goodness-of-fit test results, and a new structural equation model is generated.

8. The system for mining and simulating the relationship between digital finance and household consumption patterns according to claim 1, characterized in that, The specific steps for generating a visual causal relationship graph are as follows: Based on structural equation modeling, a causal effect value set is extracted using a causal effect separation method. Based on structural equation modeling and causal effect value set, a causal network graph is constructed with variables as nodes and causal effects as directed edge weights. Visualize and annotate the causal network diagram to generate a visual causal relationship map.

9. The system for mining and simulating the relationship between digital finance and household consumption patterns according to claim 8, characterized in that, The specific steps for extracting the causal effect value set are as follows: Propensity score matching was used to control causal interference in the structural equation model, resulting in a balanced sample set. Based on the balanced sample set and combined with mediation effect analysis, the structural equation model is decomposed into causal effect values ​​to obtain the initial causal effect value set. The initial causal effect value set is subjected to effect value significance testing to generate a new causal effect value set.

10. The system for mining and simulating the relationship between digital finance and household consumption patterns according to claim 1, characterized in that, The specific steps for generating the simulation results report are as follows: Establish a multi-agent simulation infrastructure and define the types and core attributes of home intelligent agents and regional digital financial service intelligent agents; Construct an intelligent agent interaction rule base based on a visualized causal relationship graph; Based on the types and core attributes of home intelligent agents and regional digital financial service intelligent agents, and combined with the intelligent agent interaction rule base, simulation is run in a preset simulation scenario to generate a simulation result report.

Images