A method and system for building and analyzing a causal decision tree with fusion simulation

Abstract

The application discloses a kind of fusion simulated causality decision tree construction and analysis method and system, belong to simulation experiment data analysis technical field, by individual causality effect distribution as feedback, the importance weight of key characteristic variable is dynamically adjusted, form the self-adapting iteration of reconfiguration causality forest, effectively avoid the generation of bias solidification;Through the iteration of reconfiguration causality forest in multiple rounds, output final causality forest, and generate corresponding causality decision tree, to transform the analysis result of complex causality effect into intuitive grouping rule, provide clear and understandable decision basis for user;In addition, the consistency test of causality effect is carried out to initial data grouping rule set using verification set, so as to complete rule optimization, ensure that each executable decision rule output finally has reproducible simulation empirical support, significantly improve the accuracy and reliability of output executable decision rule.

Description

Technical Field

[0001] This invention belongs to the field of simulation experiment data analysis technology, specifically relating to a method and system for constructing and analyzing causal decision trees that integrates simulation. Background Technology

[0002] In simulation experiments, accurately evaluating the causal effects of experimental variables is a crucial aspect of causal relationship analysis. Traditional methods often use the average causal effect of the entire sample as the inference target, assuming all samples share the same characteristics. However, in practical applications, the causal effects of the same variable in different samples or under different conditions often exhibit systematic differences, i.e., differentiated causal effects. Accurately identifying these differentiated effects and the conditions under which they occur is of great significance for refined decision-making.

[0003] Currently, common methods for differential causal effect analysis mainly employ a combination of causal forests and decision trees. However, existing techniques still largely rely on manual experience or exhaustive methods to select key differentiating variables among high-dimensional feature variables. This approach is highly subjective, computationally complex, and, after each causal decision tree modeling, it cannot automatically correct the output rules based on the estimation results, nor can it definitively assess the accuracy of the rules. Therefore, it is highly susceptible to biases caused by initial parameters and random sampling, making it difficult to guarantee the reliability of the output rules.

[0004] As mentioned above, how to provide a causal decision tree construction and analysis method and system that can automatically select key features and realize rule adaptive correction and accuracy judgment through fusion simulation has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for constructing and analyzing causal decision trees based on integrated simulation, in order to solve the above-mentioned problems existing in the prior art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for constructing and analyzing causal decision trees that integrates simulation, including: Acquire simulation experimental data, define processing variables, response variables and feature variables from the simulation experimental data, sample the simulation experimental data to form training set and test set, and use the honesty method to iteratively generate multiple causal trees based on the training set and test set to construct an initial causal forest. Traverse the initial causal forest and calculate the importance weight of each feature variable in node splitting to select the initial key feature variable as the initial differentiation variable. The initial differentiation variable is used as the reconstruction feature variable to reconstruct the simulation experimental data, resulting in reconstructed simulation experimental data. A reconstructed causal forest is constructed based on the reconstructed simulation experimental data. For each piece of reconstructed simulation experimental data, the corresponding reverse control data is matched in the reconstructed causal forest using a tree structure population matching strategy. Based on the reverse control data, the individual causal effect estimate and the corresponding confidence interval of each piece of reconstructed simulation experimental data are calculated to form the individual causal effect distribution. The individual causal effect distribution is fed back to the simulation experimental data to adjust the importance weight of each feature variable in node splitting. Based on the adjusted importance weight of each feature variable, the differentiation variable is screened, the causal forest is reconstructed, and the individual causal effect distribution is calculated and fed back until the individual causal effect distribution converges, and the final causal forest is output. The feature set is extracted from the final causal forest, and the individual causal effect estimate of each reconstructed simulation experimental data is used as the outcome variable. Based on the feature set and the outcome variable, the regression tree method is used to generate a causal decision tree, and the initial data grouping rule set is extracted from the causal decision tree. A validation set is obtained by sampling simulation experimental data. The validation set is used to verify the consistency of causal effects of the initial data grouping rule set, so as to optimize the rules of the initial data grouping rule set, obtain a data grouping rule set, and transform the data grouping rule set into an executable decision rule set for simulation data analysis.

[0007] In one possible design, simulation experimental data is acquired, and processing variables, response variables, and feature variables are defined from the simulation experimental data. The simulation experimental data is divided into training sets, test sets, and validation sets. Based on the training and test sets, multiple causal trees are iteratively generated using an honesty method to construct an initial causal forest. The initial causal forest is then traversed, and the importance weight of each feature variable in node splitting is calculated to select initial key feature variables as initial differentiation variables, including: Collect various index variables in the simulation experiment, integrate them to obtain the original simulation experiment data, and perform outlier cleaning and standardization on the original simulation experiment data to obtain the simulation experiment data. For each piece of simulation experiment data, define the corresponding processing variable, response variable, and feature variable; Obtain a preset sampling ratio, and perform random sampling without replacement on the simulation experiment data according to the sampling ratio to form a training set and a test set. Based on the training set, the honesty method is used to construct the causal tree. Starting from the root node, the feature variables of each simulation experimental data are traversed for each node, and the point that maximizes the difference in treatment effect is selected as the split point. At the split point, each node is split into paths until the number of each simulation experimental data in the end node of each path is lower than the preset minimum data threshold. The end node of each path is marked as a leaf node to form a causal tree. Based on the test set, the average causal effect value corresponding to each leaf node of the causal tree is calculated, and the average causal effect value corresponding to each leaf node is marked on each leaf node. According to the sampling ratio, the simulation experiment data is resampled multiple times. After each round of resampling, a causal tree is trained using the training set formed by the resampling. The average causal effect value of each leaf node corresponding to each causal tree is calculated using the test set formed by the resampling. All causal trees are numbered and integrated to construct an initial causal forest. From the initial causal forest, each causal tree is traversed to use the number of each causal tree to count the total number of splits and the corresponding split node positions of each feature variable in each causal tree. The hierarchical decay weight allocation method is adopted. Based on the total number of splits and the corresponding split node position of each feature variable in each causal tree, the corresponding variable weight is assigned to each feature variable as the importance weight of each feature variable in the node split. A preset key feature weight threshold is obtained, and the importance weight of each feature variable in node splitting is compared and filtered using the key feature weight threshold, so as to select each feature variable whose importance weight is higher than the key feature weight threshold as the initial differentiation variable.

[0008] In one possible design, the initial differentiation variable is used as a reconstruction feature variable to reconstruct the simulation experimental data, resulting in reconstructed simulation experimental data. A reconstructed causal forest is then constructed based on this reconstructed simulation experimental data. For each piece of reconstructed simulation experimental data, a tree-structured population matching strategy is used to match the corresponding reverse control data within the reconstructed causal forest. Based on the reverse control data, the estimated individual causal effect and corresponding confidence interval for each piece of reconstructed simulation experimental data are calculated, forming an individual causal effect distribution, including: The initial differentiation variables are used as reconstructed feature variables to replace the feature variables in the simulation experimental data, thereby obtaining the reconstructed simulation experimental data. The reconstruction simulation experiment data is sampled to form a reconstruction training set and a reconstruction test set. Based on the reconstruction training set and the reconstruction test set, multiple reconstruction causal trees are generated iteratively using the honesty method to construct a reconstruction causal forest. Obtain a preset tree structure group matching strategy. Based on the tree structure group matching strategy, take each reconstructed simulation experiment data as the data to be matched. Input the reconstruction feature vector corresponding to the data to be matched into the reconstruction causal forest. Select the leaf node with the reconstruction feature vector in each reconstruction causal tree as the reconstruction feature leaf node. Traverse each existing reconstructed simulation experiment data in each reconstruction feature leaf node and select the reconstructed simulation experiment data with the opposite processing variable to the data to be matched as the reverse control data of the data to be matched. Based on the reconstructed test set, the average causal effect value of each reconstructed feature leaf node is calculated; The average causal effect value of each reconstructed feature leaf node is assigned to each reconstructed simulation experiment data of each reconstructed feature leaf node, as the individual effect estimate of each reconstructed simulation experiment data in each reconstructed causal tree; Based on the reconstructed simulation experimental data, the arithmetic mean of the individual effect estimates of each reconstructed simulation experimental data in each reconstructed causal tree in the reconstructed causal forest is calculated, so that the arithmetic mean corresponding to each reconstructed simulation experimental data is used as the individual causal effect estimate corresponding to each reconstructed simulation experimental data. Based on the reverse control data corresponding to each piece of reconstructed simulation experimental data, the corresponding effect variance is calculated for the individual causal effect estimate of each piece of reconstructed simulation experimental data, and the variance average is calculated for each reconstructed causal tree in the reconstructed causal forest to obtain the confidence interval corresponding to each piece of reconstructed simulation experimental data. Based on the individual causal effect estimate and corresponding confidence interval of each reconstructed simulation experiment data, the individual causal effect distribution of each reconstructed simulation experiment data is formed.

[0009] In one possible design, the individual causal effect distribution is fed back to the simulation experimental data to adjust the importance weights of each feature variable in node splitting. Based on the adjusted importance weights of each feature variable, the process iteratively filters differentiation variables, reconstructs the causal forest, and calculates and feeds back the individual causal effect distribution until the individual causal effect distribution converges, outputting the final causal forest, including: The individual causal effect distribution is fed back to the simulation experimental data to extract the split contribution and residual correlation of each reconstructed feature variable from the individual causal effect distribution. The split contribution and residual correlation of each reconstructed feature variable are weighted and summed according to the preset contribution attention weight to obtain the feature contribution score of each reconstructed feature variable. Obtain a preset learning rate, and according to the learning rate, use the feature contribution score of each reconstructed feature variable as an adjustment term to adjust the importance weight of each feature variable in node splitting, so as to obtain the adjusted importance weight of each feature variable. Based on the adjusted importance weights of each feature variable, the feature variables are further filtered to select the current key feature variable as the current differentiation variable. The current differentiation variable is used as the feature variable for the current round of reconstruction. The simulation experimental data is reconstructed again to obtain the simulation experimental data for the current round of reconstruction. Based on the simulation experimental data for the current round of reconstruction, a causal forest for the current round of reconstruction is constructed. For each piece of simulation experimental data for the current round of reconstruction, the corresponding reverse control data is matched in the causal forest for the current round of reconstruction using a tree structure population matching strategy. Based on the reverse control data, the individual causal effect estimate and the corresponding confidence interval for each piece of simulation experimental data for the current round of reconstruction are calculated to form the individual causal effect distribution for the current round. Obtain a preset convergence rule, and use the convergence rule to determine the convergence of the individual causal effect distribution in the current round, and obtain the convergence determination result; If the current round of individual causal effect distribution does not meet the convergence rule, the convergence determination result is non-convergence. The current round of individual causal effect distribution is fed back to the simulation experimental data to update the adjusted importance weights of each feature variable in node splitting. Based on the updated adjusted importance weights, the differentiation variable screening, reconstruction of the causal forest construction, and calculation and feedback of individual causal effect distribution are iteratively performed until the current round of individual causal effect distribution converges. If the distribution of individual causal effects in the current round satisfies the convergence rule, then the convergence determination result is convergence, and the causal forest reconstructed in the current round is used as the final causal forest output.

[0010] In one possible design, reconstructed feature variables and outcome variables are extracted from the final causal forest. Based on the feature set and outcome variables, a causal decision tree is generated using a regression tree method. An initial data grouping rule set is then extracted from the causal decision tree, including: For the final causal forest, the corresponding reconstruction feature variables of the current round are extracted, and the reconstruction feature variables of each current round are integrated to form a feature set; For the final causal forest, the corresponding individual causal effect distribution of the current round is extracted, and the estimated value of the individual causal effect of each piece of the current round reconstructed simulation experimental data is extracted from the individual causal effect distribution of the current round as the result variable; The feature set and the result variables are integrated into a simulation experiment analysis dataset. Using the regression tree method, with the goal of minimizing the mean square error of the result variables of each simulation experiment analysis data within the child node, a causal decision tree is built for the simulation experiment analysis dataset. Starting from the root node of the causal decision tree, the splitting conditions of each splitting node in each branch path are collected, and the splitting conditions of each splitting node in each branch path are logically ANDed to form the grouping rules corresponding to each branch path. Integrate the grouping rules corresponding to each branch path to form an initial data grouping rule set.

[0011] In one possible design, a validation set is obtained based on sampling of simulation experiment data. This validation set is then used to verify the consistency of causal effects in the initial data grouping rule set, thereby optimizing the initial data grouping rule set to obtain a data grouping rule set, including: The simulation experiment analysis dataset is grouped according to the initial data grouping rule set to obtain multiple simulation experiment analysis data groups. For each simulation experiment analysis data group, the average individual causal effect estimate of each simulation experiment analysis data group is calculated. A preset verification data ratio is obtained, and the simulation experimental data is randomly sampled based on the verification data ratio to obtain a verification set. The simulation experimental data in the verification set is grouped according to the initial data grouping rule set to calculate the average causal verification effect value of each simulation experimental data corresponding to each simulation experimental data group. Based on each leaf node of the causal decision tree, each simulation experiment analysis data group and each simulation experiment data group are matched one-to-one to form multiple pairs of groups with the same node. For each pair of groups with the same node, the absolute value of the difference between the average individual causal effect estimate of the simulation experiment analysis data group and the average causal verification effect value of each simulation experiment data corresponding to the simulation experiment data group is calculated as the absolute deviation of each pair of groups with the same node. Obtain a preset deviation threshold, and perform threshold determination on the absolute deviation of each group of the same node based on the deviation threshold to obtain the threshold determination result; If the threshold determination result is that the absolute deviation of the same node group pair is lower than the deviation threshold, then the grouping rule corresponding to the same node group pair is considered to have passed the consistency verification, and the grouping rule is retained in the initial data grouping rule set; If the threshold determination result is that the absolute deviation of the same node group pair is not lower than the deviation threshold, then the grouping rule corresponding to the same node group pair is considered to have failed the consistency verification, and the grouping rule is removed from the initial data grouping rule set. Based on the grouping rules retained in the initial data grouping rule set, rule optimization is performed to obtain the data grouping rule set.

[0012] In one possible design, the data grouping rule set is transformed into an executable decision rule set for simulation data analysis, including: For each grouping rule in the data grouping rule set, the grouping rule is prioritized and numbered according to the average causal verification effect value of each simulation experimental data in each simulation experimental data group corresponding to each grouping rule, so as to obtain the data grouping priority rule set; Each grouping rule in the data grouping priority rule set is converted into a corresponding conditional judgment statement. Each conditional judgment statement is sorted according to the priority and number of each grouping rule in the data grouping priority rule set to form an executable decision rule set, which is then used for simulation data analysis.

[0013] Secondly, this invention provides a causal decision tree construction and analysis system that integrates simulation, including: The initial forest building unit is used to acquire simulation experiment data, define processing variables, response variables and feature variables from the simulation experiment data, sample the simulation experiment data to form training set and test set, and use the honest method to iteratively generate multiple causal trees based on the training set and test set to construct the initial causal forest. The initial causal forest is then traversed to calculate the importance weight of each feature variable in node splitting, so as to select the initial key feature variable as the initial differentiation variable. The reconstructed forest building unit is used to reconstruct the simulation experimental data by taking the initial differentiation variable as the reconstruction feature variable, thereby obtaining reconstructed simulation experimental data. Based on the reconstructed simulation experimental data, a reconstructed causal forest is constructed. For each piece of reconstructed simulation experimental data, the corresponding reverse control data is matched in the reconstructed causal forest using a tree structure population matching strategy. Based on the reverse control data, the individual causal effect estimate and the corresponding confidence interval of each piece of reconstructed simulation experimental data are calculated to form the individual causal effect distribution. The reconstructed forest iteration unit is used to feed back the individual causal effect distribution to the simulation experimental data, so as to adjust the importance weight of each feature variable in node splitting. Based on the adjusted importance weight of each feature variable, the unit iteratively performs the screening of differentiation variables, the construction of the reconstructed causal forest, and the calculation and feedback of the individual causal effect distribution until the individual causal effect distribution converges, and outputs the final causal forest. The initial rule extraction unit is used to extract a feature set from the final causal forest, take the individual causal effect estimate of each reconstructed simulation experimental data as the result variable, generate a causal decision tree using the regression tree method based on the feature set and the result variable, and extract an initial data grouping rule set from the causal decision tree; The decision rule generation unit is used to obtain a validation set based on simulation experiment data sampling, use the validation set to verify the consistency of causal effects of the initial data grouping rule set, optimize the rules of the initial data grouping rule set to obtain a data grouping rule set, and transform the data grouping rule set into an executable decision rule set for simulation data analysis.

[0014] Thirdly, the present invention provides an electronic device comprising a memory, a processor, and a transceiver connected in sequence and communication, wherein the memory is used to store a computer program, the transceiver is used to send and receive messages, and the processor is used to read the computer program and execute the causal decision tree construction and analysis method of fusion simulation as described in the first aspect or any possible design of the first aspect.

[0015] Fourthly, the present invention provides a computer-readable storage medium storing instructions that, when executed on a computer, perform the causal decision tree construction and analysis method of the fusion simulation described in the first aspect or any possible design of the first aspect.

[0016] Fifthly, the present invention provides a computer program product containing instructions that, when executed on a computer, cause the computer to perform a causal decision tree construction and analysis method based on fusion simulation as described in the first aspect or any possible design of the first aspect.

[0017] Beneficial Effects: This invention provides a method and system for constructing and analyzing causal decision trees that integrates simulation. By using the distribution of individual causal effects as feedback, it dynamically adjusts the importance weights of key feature variables to form an adaptive iteration of reconstructing the causal forest, effectively avoiding the solidification of biases. Through multiple rounds of causal forest reconstruction iterations, it outputs the final causal forest and generates the corresponding causal decision tree, transforming complex causal effect analysis results into intuitive grouping rules, providing users with clear and understandable decision-making basis. In addition, it uses a validation set to check the consistency of causal effects in the initial data grouping rule set, thereby completing rule optimization and ensuring that each executable decision rule output has reproducible simulation empirical support, significantly improving the accuracy and reliability of the output executable decision rules. Attached Figure Description

[0018] Figure 1A flowchart illustrating the causal decision tree construction and analysis method based on fusion simulation provided in this embodiment of the invention; Figure 2 A schematic diagram of the functional structure of the causal decision tree construction and analysis system with fusion simulation provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the present invention will be briefly introduced below in conjunction with the accompanying drawings and descriptions of the embodiments or the prior art. Obviously, the following description of the structure of the accompanying drawings is only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0020] It should be understood that although the terms first, second, etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are only used to distinguish one unit from another. For example, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit, without departing from the scope of the exemplary embodiments of the invention.

[0021] It should be understood that the term "and / or" that may appear in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" that may appear in this document describes another relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " that may appear in this document generally indicates that the related objects before and after it are in an "or" relationship.

[0022] Example: like Figure 1 As shown, the first aspect of this embodiment provides a method for constructing and analyzing causal decision trees based on simulation, which may include, but is not limited to, the following steps: S1. Obtain simulation experiment data, define processing variables, response variables and feature variables from the simulation experiment data, sample the simulation experiment data to form training set and test set, and use the honesty method to iteratively generate multiple causal trees based on the training set and test set to construct an initial causal forest. Traverse the initial causal forest and calculate the importance weight of each feature variable in node splitting to select the initial key feature variable as the initial differentiation variable. S2. Using the initial differentiation variable as the reconstruction feature variable, the simulation experimental data is reconstructed to obtain reconstructed simulation experimental data. Based on the reconstructed simulation experimental data, a reconstructed causal forest is constructed. For each piece of reconstructed simulation experimental data, the corresponding reverse control data is matched in the reconstructed causal forest using a tree structure population matching strategy. Based on the reverse control data, the individual causal effect estimate and the corresponding confidence interval of each piece of reconstructed simulation experimental data are calculated to form the individual causal effect distribution. S3. Feed the individual causal effect distribution back to the simulation experimental data to adjust the importance weight of each feature variable in node splitting. Based on the adjusted importance weight of each feature variable, iteratively screen the differentiation variables, reconstruct the causal forest, and calculate and feed back the individual causal effect distribution until the individual causal effect distribution converges, and output the final causal forest. S4. Extract the feature set from the final causal forest, use the individual causal effect estimate of each reconstructed simulation experimental data as the outcome variable, generate a causal decision tree using the regression tree method based on the feature set and the outcome variable, and extract the initial data grouping rule set from the causal decision tree; S5. A validation set is obtained based on the sampling of simulation experimental data. The validation set is used to verify the consistency of causal effects of the initial data grouping rule set, so as to optimize the rules of the initial data grouping rule set, obtain a data grouping rule set, and transform the data grouping rule set into an executable decision rule set, so as to use the executable decision rule set for simulation data analysis.

[0023] In one possible implementation, step S1 involves acquiring simulation experiment data, defining processing variables, response variables, and feature variables from the simulation experiment data, dividing the simulation experiment data into training sets, test sets, and validation sets, iteratively generating multiple causal trees based on the training and test sets using an honesty method to construct an initial causal forest, and traversing the initial causal forest to calculate the importance weight of each feature variable in node splitting, in order to select initial key feature variables as initial differentiation variables. This can be decomposed into, but is not limited to, the following steps S11-S19, specifically including: S11. Collect various index variables in the simulation experiment, integrate them to obtain the original simulation experiment data, and perform outlier cleaning and standardization on the original simulation experiment data to obtain the simulation experiment data. S12. For each piece of simulation experimental data in the simulation experimental data, define the corresponding processing variable, response variable, and feature variable; S13. Obtain a preset sampling ratio, and perform random sampling without replacement on the simulation experiment data according to the sampling ratio to form a training set and a test set. S14. Based on the training set, the honesty method is used to construct the causal tree. Starting from the root node, the feature variables of each simulation experimental data are traversed for each node, and the point that maximizes the difference in processing effect is selected as the split point. At the split point, each node is split into paths until the number of each simulation experimental data in the end node of each path is lower than the preset minimum data threshold. The end node of each path is marked as a leaf node to form a causal tree. S15. Based on the test set, calculate the corresponding average causal effect value for each leaf node of the causal tree, and mark the average causal effect value corresponding to each leaf node on the corresponding leaf node. S16. According to the sampling ratio, the simulation experiment data is resampled multiple times, and after each round of resampling, a causal tree is trained using the training set formed by the resampling, and the average causal effect value of each leaf node corresponding to each causal tree is calculated using the test set formed by the resampling. S17. Number and integrate all causal trees to construct an initial causal forest, and traverse each causal tree from the initial causal forest to use the number of each causal tree to count the total number of splits and the corresponding split node positions of each feature variable in each causal tree. S18. Using the hierarchical decay weight allocation method, based on the total number of times each feature variable participates in splits in each causal tree and the corresponding split node position, assign corresponding variable weights to each feature variable as the importance weights of each feature variable in node splits; S19. Obtain a preset key feature weight threshold, and use the key feature weight threshold to compare and filter the importance weights of each feature variable in node splitting, so as to select each feature variable whose importance weight is higher than the key feature weight threshold as the initial differentiation variable.

[0024] Specifically, for each piece of simulation experiment data, the defined processing variable is used to characterize the causal variable for which the causal effect needs to be examined, the response variable is used to characterize the observed result of the simulation experiment, and the feature variables are the remaining indicator variables in each piece of simulation experiment data excluding the processing variable and the response variable (forming a high-dimensional feature variable set); in a possible implementation, if the processing variable is multi-valued, it can be converted into multiple binary sub-tasks, and one level is selected as the processing group each time, and the rest are used as the control group.

[0025] It should be noted that in step S1, in order to eliminate the overfitting of the causal tree to its own training data, each causal tree is constructed using the Honest Approach based on the training set and the test set. The training set is used to determine the split structure of the causal tree, and the test set is used to calculate the causal effect estimate of each leaf node. Using completely different data samples to construct the causal tree and calculate the causal effect estimate can significantly reduce the estimation bias of individual causal effects.

[0026] In specific application scenarios, for the current node, the feature variables of each simulation experimental data and its corresponding candidate split threshold are traversed. The candidate split threshold is derived from the deduplicated set of feature variables in each simulation experimental data contained in the current node. This deduplicated set is sorted, and the midpoint between two adjacent feature variables is taken as the candidate split threshold. The point that maximizes the difference in processing effect between the left and right child nodes after splitting is selected as the split point. The current node is split along the path until the number of simulation experimental data in each end node of each path is lower than the preset minimum data threshold (or the gain brought by splitting is less than the preset tolerance value), at which point the splitting stops.

[0027] Since splits closer to the root node have a greater impact on the overall system, but deeper splits can reflect more subtle differences, this embodiment uses a hierarchical decay weighting method to assign weights to each feature variable. Specifically, the total number of times each feature variable participates in splits in each causal tree is multiplied by the corresponding split node position (the tree depth of the node's location), and the reciprocal of the product is used as the importance weight of each feature variable. The more times a feature variable participates in splits (serves as a split node) in the initial causal forest, and the deeper its corresponding split node position (representing more subtle differentiation), the higher its importance weight.

[0028] In one possible implementation, step S2 involves using the initial differentiation variable as a reconstruction feature variable to reconstruct the simulation experimental data, obtaining reconstructed simulation experimental data, and constructing a reconstructed causal forest based on the reconstructed simulation experimental data. For each piece of reconstructed simulation experimental data, a tree-structured population matching strategy is used to match the corresponding reverse control data in the reconstructed causal forest, and the individual causal effect estimate and corresponding confidence interval for each piece of reconstructed simulation experimental data are calculated based on the reverse control data to form an individual causal effect distribution. This can be, but is not limited to, decomposed into the following steps S21-S28, specifically including: S21. The initial differentiation variable is used as the reconstructed feature variable to replace the feature variable in the simulation experiment data to obtain the reconstructed simulation experiment data; S22. The reconstruction simulation experiment data is sampled to form a reconstruction training set and a reconstruction test set. Based on the reconstruction training set and the reconstruction test set, multiple reconstruction causal trees are generated iteratively using the honesty method to construct a reconstruction causal forest. S23. Obtain a preset tree structure group matching strategy. Based on the tree structure group matching strategy, take each piece of reconstructed simulation experimental data as the data to be matched, and input the reconstruction feature vector corresponding to the data to be matched into the reconstruction causal forest. Select the leaf node with the reconstruction feature vector in each reconstruction causal tree as the reconstruction feature leaf node. Traverse each piece of existing reconstructed simulation experimental data in each reconstruction feature leaf node and select the reconstructed simulation experimental data with the opposite processing variable to the data to be matched as the reverse control data of the data to be matched. S24. Based on the reconstructed test set, calculate the average causal effect value of each reconstructed feature leaf node; S25. Assign the average causal effect value of each reconstructed feature leaf node to each reconstructed simulation experiment data of each reconstructed feature leaf node, as the individual effect estimate of each reconstructed simulation experiment data in each reconstructed causal tree; S26. Based on the reconstructed simulation experimental data, calculate the arithmetic mean of the individual effect estimates of each reconstructed simulation experimental data in each reconstructed causal tree in the reconstructed causal forest, so that the arithmetic mean corresponding to each reconstructed simulation experimental data is used as the individual causal effect estimate corresponding to each reconstructed simulation experimental data. S27. Based on the reverse control data corresponding to each piece of reconstructed simulation experimental data, calculate the corresponding effect variance for the individual causal effect estimate of each piece of reconstructed simulation experimental data, and perform variance averaging on each reconstructed causal tree in the reconstructed causal forest to obtain the confidence interval corresponding to each piece of reconstructed simulation experimental data. S28. Based on the individual causal effect estimate and the corresponding confidence interval of each reconstructed simulation experiment data, form the individual causal effect distribution of each reconstructed simulation experiment data.

[0029] It should be noted that when reconstructing the simulation experimental data, only the feature variables of the simulation experimental data need to be replaced with the initial differentiation variables generated in step S1. The processing variables and response variables of each other simulation experimental data remain unchanged, thereby eliminating noise variables that are not related to the differentiation of causal effects. This allows the subsequent matching of reverse control data and estimation of individual causal effects to focus on the feature space that truly plays a differentiating role, avoiding the interference of too many dimensions on the matching accuracy of reverse control data.

[0030] Specifically, each piece of reconstructed simulation experimental data is used as the data to be matched. The reconstruction feature vector corresponding to the data to be matched is input into the reconstruction causal forest. In each reconstruction causal tree, the data to be matched will fall into a unique leaf node according to the splitting rules, that is, the reconstruction feature leaf node. The reconstruction feature leaf node contains reconstruction simulation experimental data from both the treatment group and the control group. Since these reconstruction simulation experimental data have undergone layers of splitting on the key reconstruction feature variables, they have a high degree of similarity. Therefore, the reconstruction simulation experimental data in the reconstruction feature leaf node that has the opposite processing variables to the data sample to be matched constitutes the reverse control data of the data to be matched. In the reconstruction feature leaf node, for each piece of data to be matched, there may be multiple reverse control data.

[0031] In one possible implementation, step S3 involves feeding back the individual causal effect distribution to the simulation experimental data to adjust the importance weights of each feature variable in node splitting. Based on the adjusted importance weights of each feature variable, the process iteratively filters differentiation variables, reconstructs the causal forest, and calculates and feeds back the individual causal effect distribution until the individual causal effect distribution converges, outputting the final causal forest. This can be, but is not limited to, decomposed into the following steps S31-S37, specifically including: S31. Feed the individual causal effect distribution back to the simulation experimental data to extract the split contribution and residual correlation of each reconstructed feature variable from the individual causal effect distribution, and perform a weighted summation of the split contribution and residual correlation of each reconstructed feature variable according to the preset contribution attention weight to obtain the feature contribution score of each reconstructed feature variable. S32. Obtain a preset learning rate, and according to the learning rate, use the feature contribution score of each reconstructed feature variable as an adjustment term to adjust the importance weight of each feature variable in node splitting, so as to obtain the adjusted importance weight of each feature variable. S33. Based on the adjusted importance weights of each feature variable, the feature variables are screened again to select the current key feature variable as the current differentiation variable; S34. Using the current differentiation variable as the current round reconstruction feature variable, the simulation experiment data is reconstructed again to obtain the current round reconstruction simulation experiment data. Based on the current round reconstruction simulation experiment data, a current round reconstruction causal forest is constructed. For each current round reconstruction simulation experiment data, the corresponding reverse control data is matched in the current round reconstruction causal forest using a tree structure population matching strategy. Based on the reverse control data, the individual causal effect estimate and corresponding confidence interval of each current round reconstruction simulation experiment data are calculated to form the current round individual causal effect distribution. S35. Obtain a preset convergence rule, and use the convergence rule to determine the convergence of the individual causal effect distribution in the current round, and obtain the convergence determination result; S36. If the current round individual causal effect distribution does not satisfy the convergence rule, the convergence determination result is non-convergence. The current round individual causal effect distribution is fed back to the simulation experimental data to update the adjusted importance weights of each feature variable in node splitting. Based on the updated adjusted importance weights, the differentiation variable screening, reconstruction of the causal forest construction, and calculation and feedback of the individual causal effect distribution are iteratively performed until the current round individual causal effect distribution converges. S37. If the distribution of individual causal effects in the current round satisfies the convergence rule, then the convergence determination result is convergence, and the causal forest reconstructed in the current round is used as the final causal forest output.

[0032] Wherein, the split contribution is the difference in the average causal effect values ​​of the left and right child nodes when the reconstructed feature variable is a split node of a causal tree in the current reconstructed causal forest, which is used to reflect its actual differentiation ability; the residual correlation is the correlation coefficient between the reconstructed feature variable and the individual causal effect distribution, which is used to reflect the feature's ability to explain differential effects.

[0033] The individual causal effect distribution output in step S2 is used as feedback input to step S1, thus forming a closed-loop iterative chain of screening, estimation, feedback, and re-screening. This allows the screening criteria for the differentiation variables to dynamically evolve with data characteristics, eliminating the need for repeated manual adjustments. It also avoids the solidification of biases from single estimations and prevents misjudgments of feature importance weights due to random sampling or noise factors in a single round of screening (such as misjudging noise variables as key feature variables). Through multiple iterations, weight smoothing and threshold adaptation gradually approximate the true causal structure, reducing the risk of misjudgment.

[0034] In one possible implementation, step S4 involves extracting reconstructed feature variables and outcome variables from the final causal forest, generating a causal decision tree using a regression tree method based on the feature set and outcome variables, and extracting an initial data grouping rule set from the causal decision tree. This can be decomposed, but is not limited to, the following steps S41-S45, specifically including: S41. Extract the corresponding current-round reconstruction feature variables from the final causal forest, and integrate the current-round reconstruction feature variables to form a feature set; S42. For the final causal forest, extract the corresponding individual causal effect distribution of the current round, and extract the estimated value of the individual causal effect of each piece of the current round reconstructed simulation experimental data from the individual causal effect distribution of the current round as the result variable; S43. Integrate the feature set and the result variables into a simulation experiment analysis dataset, and use the regression tree method to establish a causal decision tree for the simulation experiment analysis dataset with the goal of minimizing the mean square error of the result variables of each simulation experiment analysis data within the child node; S44. Starting from the root node of the causal decision tree, collect the splitting conditions of each splitting node in each branch path, and perform a logical AND connection on the splitting conditions of each splitting node in each branch path to form the grouping rules corresponding to each branch path. S45. Integrate the grouping rules corresponding to each branch path to form an initial data grouping rule set.

[0035] It should be noted that in the causal decision tree construction and analysis method provided in this embodiment, when using the regression tree method to construct the causal decision tree, it is first necessary to traverse the feature set and all its possible values ​​(candidate split thresholds), divide each simulation experiment analysis data in the simulation experiment analysis dataset into a left child node and a right child node, and then select the point (current round reconstructed feature variable) that minimizes the mean square error of the result variable (individual causal effect estimate) of the simulation experiment analysis data inside the left and right child nodes after splitting as the optimal split point, perform path splitting on each node, and finally, repeat the above feature set traversal and path splitting process for each split child node until the preset stopping condition is met, and generate a causal decision tree; In specific application scenarios, in order to prevent the regression tree from overfitting (creating too many fragmented groups), various stopping conditions (pre-pruning) can be preset during the growth process. Specifically: when the number of simulation experimental analysis data in a node is less than the preset simulation experimental analysis data threshold (such as 10 data), the splitting stops; when the mean square error of a split decreases less than the preset iteration benefit threshold, the splitting stops.

[0036] Specifically, starting from the root node of the causal decision tree, the splitting conditions of each splitting node in each branch path are collected. Each branch path is a complete decision tree branch path that starts from the root node and ends at a leaf node. Therefore, the number of branch paths corresponds to the number of leaf nodes in the causal decision tree, and thus the number of grouping rules that can be extracted. The splitting conditions of each splitting node in each branch path are logically ANDed. This "logical AND" connection is actually an "AND" connection, indicating that only when a simulation experimental analysis data satisfies all the splitting conditions will it enter the corresponding group. Therefore, all simulation experimental analysis data will eventually be divided into multiple non-overlapping subsets (groups) with the same number of leaf nodes. The differential causal effect values ​​of the simulation experimental analysis data in each group are very close, while the differential causal effect values ​​between each group are very significant.

[0037] In one possible implementation, step S5 involves obtaining a validation set based on simulation experiment data sampling. This validation set is then used to verify the consistency of causal effects in the initial data grouping rule set, thereby optimizing the initial data grouping rule set to obtain a data grouping rule set. This can be, but is not limited to, decomposed into the following steps S51-S57, specifically including: S51. The simulation experiment analysis dataset is grouped according to the initial data grouping rule set to obtain multiple simulation experiment analysis data groups. For each simulation experiment analysis data group, the average individual causal effect estimate of each simulation experiment analysis data group is calculated. S52. Obtain a preset verification data ratio, randomly sample the simulation experiment data based on the verification data ratio to obtain a verification set, and group each simulation experiment data in the verification set according to the initial data grouping rule set to calculate the average causal verification effect value of each simulation experiment data corresponding to each simulation experiment data group. S53. Based on each leaf node of the causal decision tree, each simulation experiment analysis data group and each simulation experiment data group are matched one-to-one to form multiple pairs of groups with the same node. For each pair of groups with the same node, the absolute value of the difference between the average individual causal effect estimate of the simulation experiment analysis data group and the average causal verification effect value of each simulation experiment data corresponding to the simulation experiment data group is calculated as the absolute deviation of each pair of groups with the same node. S54. Obtain a preset deviation threshold, and perform threshold determination on the absolute deviation of each group of the same node based on the deviation threshold to obtain the threshold determination result; S55. If the threshold determination result is that the absolute deviation of the same node group pair is lower than the deviation threshold, then the grouping rule corresponding to the same node group pair is considered to have passed the consistency verification, and the grouping rule is retained in the initial data grouping rule set; S56. If the threshold determination result is that the absolute deviation of the same node group pair is not lower than the deviation threshold, then it is considered that the grouping rule corresponding to the same node group pair has not passed the consistency verification, and the grouping rule is removed from the initial data grouping rule set. S57. Based on the grouping rules retained in the initial data grouping rule set, complete the rule optimization to obtain the data grouping rule set.

[0038] It should be noted that the validation set described in this embodiment should be completely independent of the training set and test set in steps S1 and S2 in actual applications to ensure the independence of rule optimization. Therefore, when sampling the simulation experimental data to obtain the validation set, experimental data that has not been used in the training set and test set should be selected from the simulation experimental data, and this part of the experimental data should be randomly sampled to obtain the validation set.

[0039] In one possible implementation, when optimizing the initial data grouping rule set, not only the elimination method in steps S53-S56 can be used, but also the node truncation method. Specifically: for the grouping rule corresponding to the same node grouping pair, if it fails the consistency verification, the leaf node corresponding to this grouping rule is pruned, and the end node of this pruned branch path is taken as the new leaf node to re-form the corresponding grouping rule. The causal effect consistency verification is performed again using the re-formed grouping rule. If it passes, the re-formed grouping rule is used to cover the corresponding grouping rule in the initial data grouping rule set, and the rule optimization is completed; if it fails, pruning continues until the consistency verification passes.

[0040] In one possible implementation, step S5, converting the data grouping rule set into an executable decision rule set for simulation data analysis, can be broken down into steps S58-S59, specifically including: S58. For each grouping rule in the data grouping rule set, prioritize and number each grouping rule according to the average causal verification effect value of each simulation experimental data in each simulation experimental data group corresponding to each grouping rule, to obtain a data grouping priority rule set; S59. Convert each grouping rule in the data grouping priority rule set into a corresponding conditional judgment statement, and sort each conditional judgment statement according to the priority order and number of each grouping rule in the data grouping priority rule set to form an executable decision rule set, so as to use the executable decision rule set for simulation data analysis.

[0041] In one feasible implementation, the causal decision tree can be updated by updating the branch paths according to each grouping rule in the data grouping rule set to obtain the updated causal decision tree. When outputting the executable decision rule set, the updated causal decision tree can be visualized in a graphical form to help users gain an intuitive and comprehensive understanding of the causal decision relationships in the simulation experiment data.

[0042] In practical applications, each conditional statement in the executable decision rule set can be converted into a corresponding natural language description. This facilitates data analysis and grouping of simulation data by the user.

[0043] like Figure 2 As shown, the second aspect of this embodiment provides a hardware system for implementing the causal decision tree construction and analysis method based on fusion simulation described in the first aspect of the embodiment, including: The initial forest building unit is used to acquire simulation experiment data, define processing variables, response variables and feature variables from the simulation experiment data, sample the simulation experiment data to form training set and test set, and use the honest method to iteratively generate multiple causal trees based on the training set and test set to construct the initial causal forest. The initial causal forest is then traversed to calculate the importance weight of each feature variable in node splitting, so as to select the initial key feature variable as the initial differentiation variable. The reconstructed forest building unit is used to reconstruct the simulation experimental data by taking the initial differentiation variable as the reconstruction feature variable, thereby obtaining reconstructed simulation experimental data. Based on the reconstructed simulation experimental data, a reconstructed causal forest is constructed. For each piece of reconstructed simulation experimental data, the corresponding reverse control data is matched in the reconstructed causal forest using a tree structure population matching strategy. Based on the reverse control data, the individual causal effect estimate and the corresponding confidence interval of each piece of reconstructed simulation experimental data are calculated to form the individual causal effect distribution. The reconstructed forest iteration unit is used to feed back the individual causal effect distribution to the simulation experimental data, so as to adjust the importance weight of each feature variable in node splitting. Based on the adjusted importance weight of each feature variable, the unit iteratively performs the screening of differentiation variables, the construction of the reconstructed causal forest, and the calculation and feedback of the individual causal effect distribution until the individual causal effect distribution converges, and outputs the final causal forest. The initial rule extraction unit is used to extract a feature set from the final causal forest, take the individual causal effect estimate of each reconstructed simulation experimental data as the result variable, generate a causal decision tree using the regression tree method based on the feature set and the result variable, and extract an initial data grouping rule set from the causal decision tree; The decision rule generation unit is used to obtain a validation set based on simulation experiment data sampling, use the validation set to verify the consistency of causal effects of the initial data grouping rule set, optimize the rules of the initial data grouping rule set to obtain a data grouping rule set, and transform the data grouping rule set into an executable decision rule set for simulation data analysis.

[0044] The working process, working details and technical effects of the system provided in this embodiment can be found in the first aspect of the embodiment, and will not be repeated here.

[0045] like Figure 3 As shown, the third aspect of this embodiment provides an electronic device, including: a memory, a processor, and a transceiver that are sequentially and communicatively connected, wherein the memory is used to store a computer program, the transceiver is used to send and receive messages, and the processor is used to read the computer program and execute the causal decision tree construction and analysis method of fusion simulation as described in the first aspect of the embodiment.

[0046] For specific examples, the memory may include, but is not limited to, random access memory (RAM), read-only memory (ROM), flash memory, first-in-first-out (FIFO) memory, and / or first-in-last-out (FILO) memory, etc.; specifically, the processor may include one or more processing cores, such as a 4-core processor, an 8-core processor, etc. The processor may be implemented using at least one hardware form of DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), PLA (Programmable Logic Array). The processor may also include a main processor and a coprocessor. The main processor, also known as the CPU (Central Processing Unit), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state.

[0047] In some embodiments, the processor may integrate a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the screen. For example, the processor may not be limited to microprocessors of the STM32F105 series, reduced instruction set computer (RISC) microprocessors, x86 architecture processors, or processors with integrated neural network processing units (NPUs). The transceiver may be, but is not limited to, a Wi-Fi transceiver, a Bluetooth transceiver, a General Packet Radio Service (GPRS) transceiver, a ZigBee (a low-power LAN protocol based on the IEEE 802.15.4 standard) transceiver, a 3G transceiver, a 4G transceiver, and / or a 5G transceiver. Furthermore, the electronic device may also include, but is not limited to, a power module, a display screen, and other necessary components.

[0048] The working process, working details and technical effects of the electronic device provided in this embodiment can be found in the first aspect of the embodiment, and will not be repeated here.

[0049] The fourth aspect of this embodiment provides a storage medium that stores instructions containing the causal decision tree construction and analysis method of fusion simulation as described in the first aspect of the embodiment. That is, the storage medium stores instructions that, when executed on a computer, perform the causal decision tree construction and analysis method of fusion simulation as described in the first aspect of the embodiment.

[0050] The storage medium refers to a carrier for storing data, which may include, but is not limited to, floppy disks, optical disks, hard disks, flash memory, USB flash drives, and / or memory sticks. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices.

[0051] The working process, working details and technical effects of the storage medium provided in this embodiment can be found in the first aspect of the embodiment, and will not be repeated here.

[0052] The fifth aspect of this embodiment provides a computer program product containing instructions that, when executed on a computer, cause the computer to perform the causal decision tree construction and analysis method of fusion simulation as described in the first aspect of the embodiment, wherein the computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device.

[0053] Finally, it should be noted that the above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing and analyzing causal decision trees that integrates simulation, characterized in that, include: Acquire simulation experimental data, define processing variables, response variables and feature variables from the simulation experimental data, sample the simulation experimental data to form training set and test set, and use the honesty method to iteratively generate multiple causal trees based on the training set and test set to construct an initial causal forest. Traverse the initial causal forest and calculate the importance weight of each feature variable in node splitting to select the initial key feature variable as the initial differentiation variable. The initial differentiation variable is used as the reconstruction feature variable to reconstruct the simulation experimental data, resulting in reconstructed simulation experimental data. A reconstructed causal forest is constructed based on the reconstructed simulation experimental data. For each piece of reconstructed simulation experimental data, the corresponding reverse control data is matched in the reconstructed causal forest using a tree structure population matching strategy. Based on the reverse control data, the individual causal effect estimate and the corresponding confidence interval of each piece of reconstructed simulation experimental data are calculated to form the individual causal effect distribution. The individual causal effect distribution is fed back to the simulation experimental data to adjust the importance weight of each feature variable in node splitting. Based on the adjusted importance weight of each feature variable, the differentiation variable is screened, the causal forest is reconstructed, and the individual causal effect distribution is calculated and fed back until the individual causal effect distribution converges, and the final causal forest is output. The feature set is extracted from the final causal forest, and the individual causal effect estimate of each reconstructed simulation experimental data is used as the outcome variable. Based on the feature set and the outcome variable, the regression tree method is used to generate a causal decision tree, and the initial data grouping rule set is extracted from the causal decision tree. A validation set is obtained by sampling simulation experimental data. The validation set is used to verify the consistency of causal effects of the initial data grouping rule set, so as to optimize the rules of the initial data grouping rule set, obtain a data grouping rule set, and transform the data grouping rule set into an executable decision rule set for simulation data analysis.

2. The causal decision tree construction and analysis method based on fusion simulation according to claim 1, characterized in that, Simulation experiment data is acquired, and processing variables, response variables, and feature variables are defined from the simulation experiment data. The simulation experiment data is divided into training set, test set, and validation set. Based on the training set and test set, multiple causal trees are iteratively generated using the honesty method to construct an initial causal forest. The initial causal forest is traversed, and the importance weight of each feature variable in node splitting is calculated to select initial key feature variables as initial differentiation variables, including: Collect various index variables in the simulation experiment, integrate them to obtain the original simulation experiment data, and perform outlier cleaning and standardization on the original simulation experiment data to obtain the simulation experiment data. For each piece of simulation experiment data, define the corresponding processing variable, response variable, and feature variable; Obtain a preset sampling ratio, and perform random sampling without replacement on the simulation experiment data according to the sampling ratio to form a training set and a test set. Based on the training set, the honesty method is used to construct the causal tree. Starting from the root node, the feature variables of each simulation experimental data are traversed for each node, and the point that maximizes the difference in treatment effect is selected as the split point. At the split point, each node is split into paths until the number of each simulation experimental data in the end node of each path is lower than the preset minimum data threshold. The end node of each path is marked as a leaf node to form a causal tree. Based on the test set, the average causal effect value corresponding to each leaf node of the causal tree is calculated, and the average causal effect value corresponding to each leaf node is marked on each leaf node. According to the sampling ratio, the simulation experiment data is resampled multiple times. After each round of resampling, a causal tree is trained using the training set formed by the resampling. The average causal effect value of each leaf node corresponding to each causal tree is calculated using the test set formed by the resampling. All causal trees are numbered and integrated to construct an initial causal forest. From the initial causal forest, each causal tree is traversed to use the number of each causal tree to count the total number of splits and the corresponding split node positions of each feature variable in each causal tree. The hierarchical decay weight allocation method is adopted. Based on the total number of splits and the corresponding split node position of each feature variable in each causal tree, the corresponding variable weight is assigned to each feature variable as the importance weight of each feature variable in the node split. A preset key feature weight threshold is obtained, and the importance weight of each feature variable in node splitting is compared and filtered using the key feature weight threshold, so as to select each feature variable whose importance weight is higher than the key feature weight threshold as the initial differentiation variable.

3. The causal decision tree construction and analysis method based on fusion simulation according to claim 1, characterized in that, The initial differentiation variables are used as reconstruction feature variables to reconstruct the simulation experimental data, resulting in reconstructed simulation experimental data. A reconstructed causal forest is constructed based on this data. For each piece of reconstructed simulation experimental data, a tree-structured population matching strategy is used to match the corresponding reverse control data within the reconstructed causal forest. Based on the reverse control data, the estimated individual causal effect and corresponding confidence interval for each piece of reconstructed simulation experimental data are calculated, forming the individual causal effect distribution, including: The initial differentiation variables are used as reconstructed feature variables to replace the feature variables in the simulation experimental data, thereby obtaining the reconstructed simulation experimental data. The reconstruction simulation experiment data is sampled to form a reconstruction training set and a reconstruction test set. Based on the reconstruction training set and the reconstruction test set, multiple reconstruction causal trees are generated iteratively using the honesty method to construct a reconstruction causal forest. Obtain a preset tree structure group matching strategy. Based on the tree structure group matching strategy, take each reconstructed simulation experiment data as the data to be matched. Input the reconstruction feature vector corresponding to the data to be matched into the reconstruction causal forest. Select the leaf node with the reconstruction feature vector in each reconstruction causal tree as the reconstruction feature leaf node. Traverse each existing reconstructed simulation experiment data in each reconstruction feature leaf node and select the reconstructed simulation experiment data with the opposite processing variable to the data to be matched as the reverse control data of the data to be matched. Based on the reconstructed test set, the average causal effect value of each reconstructed feature leaf node is calculated; The average causal effect value of each reconstructed feature leaf node is assigned to each reconstructed simulation experiment data of each reconstructed feature leaf node, as the individual effect estimate of each reconstructed simulation experiment data in each reconstructed causal tree; Based on the reconstructed simulation experimental data, the arithmetic mean of the individual effect estimates of each reconstructed simulation experimental data in each reconstructed causal tree in the reconstructed causal forest is calculated, so that the arithmetic mean corresponding to each reconstructed simulation experimental data is used as the individual causal effect estimate corresponding to each reconstructed simulation experimental data. Based on the reverse control data corresponding to each piece of reconstructed simulation experimental data, the corresponding effect variance is calculated for the individual causal effect estimate of each piece of reconstructed simulation experimental data, and the variance average is calculated for each reconstructed causal tree in the reconstructed causal forest to obtain the confidence interval corresponding to each piece of reconstructed simulation experimental data. Based on the individual causal effect estimate and corresponding confidence interval of each reconstructed simulation experiment data, the individual causal effect distribution of each reconstructed simulation experiment data is formed.

4. The causal decision tree construction and analysis method based on fusion simulation according to claim 1, characterized in that, The individual causal effect distribution is fed back to the simulation experimental data to adjust the importance weights of each feature variable in node splitting. Based on the adjusted importance weights of each feature variable, the process iteratively filters differentiation variables, reconstructs the causal forest, and calculates and feeds back the individual causal effect distribution until the individual causal effect distribution converges, outputting the final causal forest, including: The individual causal effect distribution is fed back to the simulation experimental data to extract the split contribution and residual correlation of each reconstructed feature variable from the individual causal effect distribution. The split contribution and residual correlation of each reconstructed feature variable are weighted and summed according to the preset contribution attention weight to obtain the feature contribution score of each reconstructed feature variable. Obtain a preset learning rate, and according to the learning rate, use the feature contribution score of each reconstructed feature variable as an adjustment term to adjust the importance weight of each feature variable in node splitting, so as to obtain the adjusted importance weight of each feature variable. Based on the adjusted importance weights of each feature variable, the feature variables are further filtered to select the current key feature variable as the current differentiation variable. The current differentiation variable is used as the feature variable for the current round of reconstruction. The simulation experimental data is reconstructed again to obtain the simulation experimental data for the current round of reconstruction. Based on the simulation experimental data for the current round of reconstruction, a causal forest for the current round of reconstruction is constructed. For each piece of simulation experimental data for the current round of reconstruction, the corresponding reverse control data is matched in the causal forest for the current round of reconstruction using a tree structure population matching strategy. Based on the reverse control data, the individual causal effect estimate and the corresponding confidence interval for each piece of simulation experimental data for the current round of reconstruction are calculated to form the individual causal effect distribution for the current round. Obtain a preset convergence rule, and use the convergence rule to determine the convergence of the individual causal effect distribution in the current round, and obtain the convergence determination result; If the current round of individual causal effect distribution does not meet the convergence rule, the convergence determination result is non-convergence. The current round of individual causal effect distribution is fed back to the simulation experimental data to update the adjusted importance weights of each feature variable in node splitting. Based on the updated adjusted importance weights, the differentiation variable screening, reconstruction of the causal forest construction, and calculation and feedback of individual causal effect distribution are iteratively performed until the current round of individual causal effect distribution converges. If the distribution of individual causal effects in the current round satisfies the convergence rule, then the convergence determination result is convergence, and the causal forest reconstructed in the current round is used as the final causal forest output.

5. The causal decision tree construction and analysis method based on fusion simulation according to claim 4, characterized in that, Reconstructed feature variables and outcome variables are extracted from the final causal forest. Based on the feature set and outcome variables, a causal decision tree is generated using the regression tree method. An initial data grouping rule set is then extracted from the causal decision tree, including: For the final causal forest, the corresponding reconstruction feature variables of the current round are extracted, and the reconstruction feature variables of each current round are integrated to form a feature set; For the final causal forest, the corresponding individual causal effect distribution of the current round is extracted, and the estimated value of the individual causal effect of each piece of the current round reconstructed simulation experimental data is extracted from the individual causal effect distribution of the current round as the result variable; The feature set and the result variables are integrated into a simulation experiment analysis dataset. Using the regression tree method, with the goal of minimizing the mean square error of the result variables of each simulation experiment analysis data within the child node, a causal decision tree is built for the simulation experiment analysis dataset. Starting from the root node of the causal decision tree, the splitting conditions of each splitting node in each branch path are collected, and the splitting conditions of each splitting node in each branch path are logically ANDed to form the grouping rules corresponding to each branch path. Integrate the grouping rules corresponding to each branch path to form an initial data grouping rule set.

6. The causal decision tree construction and analysis method based on fusion simulation according to claim 5, characterized in that, A validation set is obtained based on sampling of simulation experiment data. This validation set is then used to verify the consistency of causal effects in the initial data grouping rule set. This process optimizes the initial data grouping rule set, resulting in a new data grouping rule set, including: The simulation experiment analysis dataset is grouped according to the initial data grouping rule set to obtain multiple simulation experiment analysis data groups. For each simulation experiment analysis data group, the average individual causal effect estimate of each simulation experiment analysis data group is calculated. A preset verification data ratio is obtained, and the simulation experimental data is randomly sampled based on the verification data ratio to obtain a verification set. The simulation experimental data in the verification set is grouped according to the initial data grouping rule set to calculate the average causal verification effect value of each simulation experimental data corresponding to each simulation experimental data group. Based on each leaf node of the causal decision tree, each simulation experiment analysis data group and each simulation experiment data group are matched one-to-one to form multiple pairs of groups with the same node. For each pair of groups with the same node, the absolute value of the difference between the average individual causal effect estimate of the simulation experiment analysis data group and the average causal verification effect value of each simulation experiment data corresponding to the simulation experiment data group is calculated as the absolute deviation of each pair of groups with the same node. Obtain a preset deviation threshold, and perform threshold determination on the absolute deviation of each group of the same node based on the deviation threshold to obtain the threshold determination result; If the threshold determination result is that the absolute deviation of the same node group pair is lower than the deviation threshold, then the grouping rule corresponding to the same node group pair is considered to have passed the consistency verification, and the grouping rule is retained in the initial data grouping rule set; If the threshold determination result is that the absolute deviation of the same node group pair is not lower than the deviation threshold, then the grouping rule corresponding to the same node group pair is considered to have failed the consistency verification, and the grouping rule is removed from the initial data grouping rule set. Based on the grouping rules retained in the initial data grouping rule set, rule optimization is performed to obtain the data grouping rule set.

7. The causal decision tree construction and analysis method based on fusion simulation according to claim 6, characterized in that, The data grouping rule set is transformed into an executable decision rule set for simulation data analysis, including: For each grouping rule in the data grouping rule set, the grouping rule is prioritized and numbered according to the average causal verification effect value of each simulation experimental data in each simulation experimental data group corresponding to each grouping rule, so as to obtain the data grouping priority rule set; Each grouping rule in the data grouping priority rule set is converted into a corresponding conditional judgment statement. Each conditional judgment statement is sorted according to the priority and number of each grouping rule in the data grouping priority rule set to form an executable decision rule set, which is then used for simulation data analysis.

8. A causal decision tree construction and analysis system integrating simulation, characterized in that, The causal decision tree construction and analysis method applied to the fusion simulation as described in any one of claims 1 to 7 includes: The initial forest building unit is used to acquire simulation experiment data, define processing variables, response variables and feature variables from the simulation experiment data, sample the simulation experiment data to form training set and test set, and use the honest method to iteratively generate multiple causal trees based on the training set and test set to construct the initial causal forest. The initial causal forest is then traversed to calculate the importance weight of each feature variable in node splitting, so as to select the initial key feature variable as the initial differentiation variable. The reconstructed forest building unit is used to reconstruct the simulation experimental data by taking the initial differentiation variable as the reconstruction feature variable, thereby obtaining reconstructed simulation experimental data. Based on the reconstructed simulation experimental data, a reconstructed causal forest is constructed. For each piece of reconstructed simulation experimental data, the corresponding reverse control data is matched in the reconstructed causal forest using a tree structure population matching strategy. Based on the reverse control data, the individual causal effect estimate and the corresponding confidence interval of each piece of reconstructed simulation experimental data are calculated to form the individual causal effect distribution. The reconstructed forest iteration unit is used to feed back the individual causal effect distribution to the simulation experimental data, so as to adjust the importance weight of each feature variable in node splitting. Based on the adjusted importance weight of each feature variable, the unit iteratively performs the screening of differentiation variables, the construction of the reconstructed causal forest, and the calculation and feedback of the individual causal effect distribution until the individual causal effect distribution converges, and outputs the final causal forest. The initial rule extraction unit is used to extract a feature set from the final causal forest, take the individual causal effect estimate of each reconstructed simulation experimental data as the result variable, generate a causal decision tree using the regression tree method based on the feature set and the result variable, and extract an initial data grouping rule set from the causal decision tree; The decision rule generation unit is used to obtain a validation set based on simulation experiment data sampling, use the validation set to verify the consistency of causal effects of the initial data grouping rule set, optimize the rules of the initial data grouping rule set to obtain a data grouping rule set, and transform the data grouping rule set into an executable decision rule set for simulation data analysis.

9. An electronic device, characterized in that, The device includes a memory, a processor, and a transceiver that are sequentially and communicatively connected. The memory is used to store a computer program, the transceiver is used to send and receive messages, and the processor is used to read the computer program and execute the causal decision tree construction and analysis method of fusion simulation as described in any one of claims 1 to 7.

10. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or the instructions are executed by the computer, they implement the causal decision tree construction and analysis method of fusion simulation as described in any one of claims 1 to 7.

Images