Intelligent decision support method and system based on causal inference
By constructing a causal directed weighted acyclic graph and a causal forest algorithm, the robustness and interpretability issues of causal inference intelligent decision support methods are solved, generating intuitive decision solutions and improving the accuracy and operational efficiency of advertising strategies.
Patent Information
- Application Number
- CN202511456155.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-13
- Publication Date
- 2026-02-27
AI Technical Summary
Existing intelligent decision support methods based on causal inference lack robustness assessment, are susceptible to data fluctuations, cannot effectively identify the relationship between intermediate paths and operational actions, have contradictions between the black-box nature and interpretability of machine learning models, and lack the ability to process diverse and heterogeneous data, resulting in coarse strategies and difficulty in scaling up.
By constructing a causal directed weighted acyclic graph and combining resampling and causal forest algorithms, this study systematically reveals stable causal paths and heterogeneous effects among variables, quantifies the individual contribution of key influencing factors, and generates intuitive decision-making schemes through NLG technology.
It significantly improves the accuracy and robustness of advertising strategies, realizes a closed loop from data to automated decision-making, optimizes advertising resource allocation and personalized marketing, and improves return on investment and operational efficiency.
Smart Images
Figure CN121581189A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data analysis technology, specifically to an intelligent decision support method and system based on causal inference. Background Technology
[0002] Intelligent decision-making supported by causal inference lacks robustness assessment in its causal discovery process, and its conclusions are easily affected by data fluctuations; it relies too much on average treatment effect (ATE) while ignoring effect heterogeneity, resulting in coarsened strategies; although it can identify intermediate paths, it cannot effectively correlate them with specific operational actions; the black-box nature of machine learning models contradicts the requirements for interpretability; the output results are mostly statistical charts rather than actionable solutions, heavily relying on human interpretation; and it lacks the ability to process diverse and heterogeneous data and domain knowledge, ultimately leading to a huge gap between causal insights and automated decision-making, making it difficult to scale up and apply. Summary of the Invention
[0003] To address the aforementioned technical problems, this paper provides an intelligent decision support method and system based on causal inference, which solves the problems described above.
[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows: An intelligent decision support method based on causal inference includes: S1. Based on the digital advertising marketing backend database, read the historical multi-dimensional structure data of various digital advertising businesses, discover and analyze the causal structure of the multi-dimensional structure data of the operations according to the mixed causal structure, and establish a causal directed acyclic graph of each digital advertising business. S2. Based on the causal directed acyclic graph of each digital advertising business, identify and estimate causal effects according to the causal structure of the multivariate data in the causal directed acyclic graph, mark the effect state between causal variables, and establish a causal directed weighted acyclic graph of each digital advertising business. S3. Based on the causal directed weighted acyclic graph of each digital advertising business, analyze the stability and heterogeneity of the causal structure of each path in the causal directed weighted acyclic graph, substitute it into the causal decision-making model of digital advertising marketing, and generate causal impact decision support schemes for each digital advertising business.
[0005] Preferably, step S1 specifically includes: By using multiple interpolation, interpolation preprocessing is performed on the multi-dimensional structure data of the operation of each historical digital advertising business to generate several multi-dimensional structure datasets of the operation of each historical digital advertising business. Based on several operational multi-structure datasets from various historical digital advertising businesses, variable types were classified according to the cardinality of each structure data, and standardized according to the corresponding data codes. Based on several operational multivariate structure datasets of various historical digital advertising businesses, the cardinality categorical variable in each operational multivariate structure dataset is used as a node. All nodes are fully connected to construct an initial causal undirected acyclic graph of each digital advertising business. Conditional independence tests were conducted on the causal undirected acyclic graphs of the initial digital advertising businesses. The partial correlation coefficients between nodes and connected nodes in the causal undirected acyclic graphs of the initial digital advertising businesses were calculated and substituted into the test statistics to verify the significant correlation between nodes in the causal undirected acyclic graphs of the initial digital advertising businesses. Based on the significant correlations between nodes in the initial causal undirected acyclic graph of each digital advertising business, remove the unrelated independent edges from the initial causal undirected acyclic graph of each digital advertising business to obtain the causal undirected acyclic graph of each digital advertising business.
[0006] Preferably, step S1 further includes: Traverse each pair of adjacent nodes in the causal undirected acyclic graph of each digital advertising business, mark each pair of adjacent nodes as having legal edge addition operations according to the legal edge addition check rules, and generate a set of candidate edge addition operations for the causal undirected acyclic graph of each digital advertising business. Based on the set of candidate edge-adding operations for each digital advertising business's causal undirected acyclic graph, a temporary edge-adding causal undirected acyclic graph for each digital advertising business is generated. By utilizing the causal undirected acyclic graphs of temporary edge additions for each digital advertising business, candidate edge addition operations under local node connection structure changes in the causal undirected acyclic graphs of each digital advertising business are screened. The maximum score difference of each candidate edge addition operation is verified using the Bayesian information criterion. The node connection edges in the causal undirected acyclic graphs of each digital advertising business are then updated, as follows:
[0007] in, Add edge operations to the candidate nodes and score differences under the changes in the local node connection structure of the causal undirected acyclic graphs of various digital advertising businesses. This is the set of new parent nodes for node Y after adding a new edge. This is the set of parent nodes of the original node Y. For the Bayesian information criterion function, Iterating the rules for checking the legal addition of edges until no legal edge addition operation can improve the score of the causal undirected acyclic graph with temporary edge additions for various digital advertising businesses.
[0008] Preferably, step S1 further includes: Traverse each pair of adjacent nodes in the causal undirected acyclic graph of each digital advertising business, mark each pair of adjacent nodes as having a valid deletion operation according to the valid deletion edge check rules, and generate a set of candidate deletion edge operations for each causal undirected acyclic graph of each digital advertising business. Based on the set of candidate deletion edge operations for the causal undirected acyclic graph of each digital advertising business, a temporary deletion edge causal undirected acyclic graph of each digital advertising business is generated. By utilizing the temporary deletion edge causal undirected acyclic graph of each digital advertising business, candidate deletion edge operations under the local node connection structure changes between the causal undirected acyclic graphs of each digital advertising business are screened. The maximum score difference of each candidate deletion edge operation is verified by Bayesian information criterion, and the node connection edges in the causal undirected acyclic graph of each digital advertising business are updated. Iterate through the rules for checking the legal edge deletion until no legal edge deletion operation can improve the score of the temporary edge deletion causal undirected acyclic graph for each digital advertising business, and obtain the causal directed acyclic graph for each digital advertising business.
[0009] Preferably, step S2 specifically includes: Based on the causal directed acyclic graphs of various digital advertising businesses, and following the backdoor criterion, the weighted average expected value of the choice paths between all operational decision variables A and advertising business outcome variables B is used to select the adjustment set S of the causal directed acyclic graphs of each digital advertising business, as follows:
[0010] in, The weighted average expected value of the choice paths among all operational decision variables A and advertising business outcome variables B. For the value of the advertising business outcome variable, The weighted average expected value of all selected paths in the adjustment set is distributed as a whole average. Based on the causal directed acyclic graphs of various digital advertising businesses, and following the front-door criterion, the weighted average expected value of the mediating variable C along the choice path between all operational decision variables A and advertising business outcome variables B is used to select the variable set U of the causal directed acyclic graphs of each digital advertising business. The method is as follows:
[0011] in, The weighted average expected value of the mediating variable C, representing the choice path between all operational decision variables A and advertising business outcome variables B. Let the set of variables be the weighted average expected value of all chosen paths, distributed as a whole. The value of the mediating variable for advertising business. The conditional probability of the nearest mediator variable C between all operational decision variables A and advertising business outcome variables B.
[0012] Preferably, step S2 further includes: Based on the adjustment set S of the causal directed acyclic graph of each digital advertising business and the variable set U of the causal directed acyclic graph of each digital advertising business, a path decision covariate vector of the causal directed acyclic graph of each digital advertising business is constructed. For each digital advertising business, the path decision covariate vector of the causal directed acyclic graph is divided into a mediated exposure array and a control non-mediated exposure array. A decision tree of expected value of digital advertising business decision path for mediated influencing factors and a decision tree of expected value of digital advertising business decision path for non-mediated influencing factors are established. A random forest of expected factors of decision path for each digital advertising business is constructed, and a set of causal concerns for non-mediated path and mediated path decision path of digital advertising business is generated. Based on Logistic Regression, the non-mediated path and the mediated path of the decision-making path of digital advertising business are taken as input, and the non-mediated path and the mediated path of the causal directed acyclic graph of each digital advertising business are taken as output. By using Blending, the non-mediated path and mediated path decision probabilities of the causal directed acyclic graph of each digital advertising business are linearly weighted and fused to obtain the optimal path of the causal directed acyclic graph of each digital advertising business, and generate the causal directed weighted acyclic graph of each digital advertising business.
[0013] Preferably, step S3 specifically includes: Resample the multi-dimensional structure data of the operation of each digital advertising business in history, execute the entire process from step S1 to step S2, count the frequency of occurrence of causal paths of each digital advertising business in the causal directed weighted acyclic graph, and use the Bootstrap distribution to calculate the confidence interval of the influencing factors of the causal paths of each digital advertising business. Based on the confidence intervals of the influencing factors of the causal paths of various digital advertising businesses, the sequence of influencing factors of digital advertising businesses in the causal directed weighted acyclic graph of each digital advertising business is screened, a CATE causal forest is trained, a causal decision model for digital advertising marketing is established, and the individual contribution values of the influencing factors of the causal paths of each digital advertising business are analyzed.
[0014] Preferably, step S3 further includes: Based on the confidence intervals of the influencing factors in the causal paths of various digital advertising businesses and the individual contribution values of the influencing factors in the causal paths of various digital advertising businesses, the NLG (Natural Language Generating Machine Learning) is used to generate analysis reports for each digital advertising business according to the average increase in the individual contribution values of the influencing factors based on the confidence intervals of the influencing factors in the causal paths of various digital advertising businesses, thus obtaining the causal impact decision support scheme for each digital advertising business.
[0015] Furthermore, an intelligent decision support system based on causal inference, used to implement the intelligent decision support method based on causal inference as described above, includes: The causal relationship analysis module and the directed graph analysis module are used to read the historical multi-dimensional operational data of various digital advertising businesses based on the digital advertising marketing backend database, discover and analyze the causal structure of the multi-dimensional operational data according to the mixed causal structure, and establish a causal directed acyclic graph for each digital advertising business. The causal weighted analysis module is electrically connected to the causal association analysis module. The causal weighted analysis module is used to identify and estimate causal effects based on the causal directed acyclic graph of each digital advertising business, according to the causal structure of the multivariate data in the causal directed acyclic graph, marking the effect state between causal variables, and establishing the causal directed weighted acyclic graph of each digital advertising business. The decision support generation module is electrically connected to the causal weighted analysis module. The decision support generation module is used to analyze the stability and heterogeneity of the causal structure of each path in the causal directed weighted acyclic graph of each digital advertising business, and then input it into the causal decision model of digital advertising marketing to generate causal impact decision support schemes for each digital advertising business.
[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention proposes an intelligent decision support scheme based on causal inference. This method constructs a causal directed weighted acyclic graph (NLG) and combines resampling and causal forest algorithms to systematically reveal stable causal paths and heterogeneous effects among variables in digital advertising, quantifying the individual contributions of key influencing factors. Finally, NLG technology transforms complex statistical results into intuitive decision-making solutions. This effectively overcomes the limitations of traditional correlation analysis, significantly improves the accuracy and robustness of advertising strategies, and achieves a closed loop of automated data-driven decision-making. It provides a reliable basis for the optimal allocation of advertising resources and personalized marketing, greatly improving ROI and operational efficiency. Attached Figure Description
[0017] Figure 1 This is a flowchart of an intelligent decision support method based on causal inference. Figure 2 This is the same framework diagram for intelligent decision support based on causal inference. Detailed Implementation
[0018] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.
[0019] Reference Figure 1 As shown, an intelligent decision support method based on causal inference includes: S1. Based on the digital advertising marketing backend database, read the historical multi-dimensional structure data of various digital advertising businesses, discover and analyze the causal structure of the multi-dimensional structure data of the operations according to the mixed causal structure, and establish a causal directed acyclic graph of each digital advertising business. Step S1 specifically includes: By using multiple interpolation, interpolation preprocessing is performed on the multi-dimensional structure data of the operation of each historical digital advertising business to generate several multi-dimensional structure datasets of the operation of each historical digital advertising business. Based on several operational multi-structure datasets from various historical digital advertising businesses, variable types were classified according to the cardinality of each structure data, and standardized according to the corresponding data codes. Based on several operational multivariate structure datasets of various historical digital advertising businesses, the cardinality categorical variable in each operational multivariate structure dataset is used as a node. All nodes are fully connected to construct an initial causal undirected acyclic graph of each digital advertising business. Conditional independence tests were conducted on the causal undirected acyclic graphs of the initial digital advertising businesses. The partial correlation coefficients between nodes and connected nodes in the causal undirected acyclic graphs of the initial digital advertising businesses were calculated and substituted into the test statistics to verify the significant correlation between nodes in the causal undirected acyclic graphs of the initial digital advertising businesses. As a further development, since the operational multivariate structure dataset contains continuous, categorical, and mixed variables, when conducting conditional independence tests on the causal undirected acyclic graphs of the initial digital advertising businesses, conditional mutual information needs to be used to verify the categorical and mixed variables to ensure the accuracy of the causal relationships between nodes.
[0020] Based on the significant correlation between nodes in the initial causal undirected acyclic graph of each digital advertising business, remove the unrelated independent edges in the initial causal undirected acyclic graph of each digital advertising business to obtain the causal undirected acyclic graph of each digital advertising business. Step S1 also includes: Traverse each pair of adjacent nodes in the causal undirected acyclic graph of each digital advertising business, mark each pair of adjacent nodes as having legal edge addition operations according to the legal edge addition check rules, and generate a set of candidate edge addition operations for the causal undirected acyclic graph of each digital advertising business. As a further development, the rules for legally adding edges are as follows: nodes must be adjacent to each other in the causal undirected acyclic graph of each digital advertising business, and no cycles can be created between nodes after adding edges. Based on the set of candidate edge-adding operations for each digital advertising business's causal undirected acyclic graph, a temporary edge-adding causal undirected acyclic graph for each digital advertising business is generated. By utilizing the causal undirected acyclic graphs of temporary edge additions for each digital advertising business, candidate edge addition operations under local node connection structure changes in the causal undirected acyclic graphs of each digital advertising business are screened. The maximum score difference of each candidate edge addition operation is verified using the Bayesian information criterion. The node connection edges in the causal undirected acyclic graphs of each digital advertising business are then updated, as follows:
[0021] in, Add edge operations to the candidate nodes and score differences under the changes in the local node connection structure of the causal undirected acyclic graphs of various digital advertising businesses. This is the set of new parent nodes for node Y after adding a new edge. This is the set of parent nodes of the original node Y. For the Bayesian information criterion function, As a further development, if the attribute values of a node are continuous variables, the core algorithm for validating the Bayesian information criterion function should use linear regression likelihood estimation to obtain the maximum score difference. Iterating the rules for checking the legal addition of edges until there are no more legal edge addition operations can improve the score of the temporary edge addition causal undirected acyclic graph for various digital advertising businesses; Step S1 also includes: Traverse each pair of adjacent nodes in the causal undirected acyclic graph of each digital advertising business, mark each pair of adjacent nodes as having a valid deletion operation according to the valid deletion edge check rules, and generate a set of candidate deletion edge operations for each causal undirected acyclic graph of each digital advertising business. Based on the set of candidate deletion edge operations for the causal undirected acyclic graph of each digital advertising business, a temporary deletion edge causal undirected acyclic graph of each digital advertising business is generated. By utilizing the temporary deletion edge causal undirected acyclic graph of each digital advertising business, candidate deletion edge operations under the local node connection structure changes between the causal undirected acyclic graphs of each digital advertising business are screened. The maximum score difference of each candidate deletion edge operation is verified by Bayesian information criterion, and the node connection edges in the causal undirected acyclic graph of each digital advertising business are updated. Iterate through the rules for checking the legal edge deletion until no legal edge deletion operation can improve the score of the temporary edge deletion causal undirected acyclic graph of each digital advertising business, and obtain the causal directed acyclic graph of each digital advertising business. When using it, please refer to the steps outlined above: As a further development, this study utilizes conditional independence tests to quickly prune irrelevant edges to determine the causal framework, and then uses greedy search to optimize the Bayesian Information Criterion (BIC) score to determine the causal direction. This allows for the efficient and accurate construction of a causal directed acyclic graph (DAG) from multivariate heterogeneous data. By combining multiple imputation to scientifically address the problem of missing data, this approach not only significantly improves computational efficiency for high-dimensional data but also enhances the robustness and reliability of the findings by integrating results from multiple data versions, providing a highly interpretable structural foundation for subsequent causal effect estimation.
[0022] S2. Based on the causal directed acyclic graph of each digital advertising business, identify and estimate causal effects according to the causal structure of the multivariate data in the causal directed acyclic graph, mark the effect state between causal variables, and establish a causal directed weighted acyclic graph of each digital advertising business. Step S2 specifically includes: Based on the causal directed acyclic graphs of various digital advertising businesses, and following the backdoor criterion, the weighted average expected value of the choice paths between all operational decision variables A and advertising business outcome variables B is used to select the adjustment set S of the causal directed acyclic graphs of each digital advertising business, as follows:
[0023] in, The weighted average expected value of the choice paths among all operational decision variables A and advertising business outcome variables B. For the value of the advertising business outcome variable, The weighted average expected value of all selected paths in the adjustment set is distributed as a whole average. Based on the causal directed acyclic graphs of various digital advertising businesses, and following the front-door criterion, the weighted average expected value of the mediating variable C along the choice path between all operational decision variables A and advertising business outcome variables B is used to select the variable set U of the causal directed acyclic graphs of each digital advertising business. The method is as follows:
[0024] in, The weighted average expected value of the mediating variable C, representing the choice path between all operational decision variables A and advertising business outcome variables B. Let the set of variables be the weighted average expected value of all chosen paths, distributed as a whole. The value of the mediating variable for advertising business. The conditional probability of the nearest mediator variable C between all operational decision variables A and advertising business outcome variables B; Step S2 also includes: Based on the adjustment set S of the causal directed acyclic graph of each digital advertising business and the variable set U of the causal directed acyclic graph of each digital advertising business, a path decision covariate vector of the causal directed acyclic graph of each digital advertising business is constructed. For each digital advertising business, the path decision covariate vector of the causal directed acyclic graph is divided into a mediated exposure array and a control non-mediated exposure array. A decision tree of expected value of digital advertising business decision path for mediated influencing factors and a decision tree of expected value of digital advertising business decision path for non-mediated influencing factors are established. A random forest of expected factors of decision path for each digital advertising business is constructed, and a set of causal concerns for non-mediated path and mediated path decision path of digital advertising business is generated. Based on Logistic Regression, the non-mediated path and the mediated path of the decision-making path of digital advertising business are taken as input, and the non-mediated path and the mediated path of the causal directed acyclic graph of each digital advertising business are taken as output. By using Blending, the non-mediated path and mediated path decision probabilities of the causal directed acyclic graph of each digital advertising business are linearly weighted and fused to obtain the optimal path of the causal directed acyclic graph of each digital advertising business, and generate the causal directed weighted acyclic graph of each digital advertising business.
[0025] When using it, please refer to the steps outlined above: As a further development, traditional methods stop at effect estimation, while this solution provides clear, data-driven decision-making recommendations ("which path to take"), greatly enhancing the action guidance value of the analysis results. By distinguishing between mediated and non-mediated paths, operators can implement more refined strategies. For example, for users sensitive to the "brand awareness" mediated path, brand image ads should be placed; for users sensitive to the direct conversion path, promotional ads should be placed, achieving precise operation of different groups. The application of ensemble learning (random forest, blending) reduces the risk of model overfitting, making path discovery and selection strategies more stable and reliable, suitable for large-scale data environments. The final weighted causal graph provides strong business interpretability. At the same time, the entire process is highly automated and can be systematically applied to various business lines of digital advertising (search ads, display ads, feed ads) to form a unified decision analysis framework.
[0026] S3. Based on the causal directed weighted acyclic graph of each digital advertising business, analyze the stability and heterogeneity of the causal structure of each path in the causal directed weighted acyclic graph, substitute it into the causal decision-making model of digital advertising marketing, and generate causal impact decision support schemes for each digital advertising business.
[0027] Step S3 specifically includes: Resample the multi-dimensional structure data of the operation of each digital advertising business in history, execute the entire process from step S1 to step S2, count the frequency of occurrence of causal paths of each digital advertising business in the causal directed weighted acyclic graph, and use the Bootstrap distribution to calculate the confidence interval of the influencing factors of the causal paths of each digital advertising business. Based on the confidence interval of the influencing factors of the causal path of each digital advertising business, the sequence of influencing factors of digital advertising business in the causal directed weighted acyclic graph of each digital advertising business is screened, CATE causal forest is trained, a causal decision model of digital advertising marketing is established, and the individual contribution value of the influencing factors of the causal path of each digital advertising business is analyzed. Step S3 also includes: Based on the confidence intervals of the influencing factors in the causal paths of various digital advertising businesses and the individual contribution values of the influencing factors in the causal paths of various digital advertising businesses, the NLG (Natural Language Generating Machine Learning) is used to generate analysis reports for each digital advertising business according to the average increase in the individual contribution values of the influencing factors based on the confidence intervals of the influencing factors in the causal paths of various digital advertising businesses, thus obtaining the causal impact decision support scheme for each digital advertising business.
[0028] Use this in conjunction with the steps outlined above: As part of that step, a company evaluated the causal effect of "search engine keyword advertising" on "software trial registration." This included historical ad impression data, user behavior data, and registration conversion data.
[0029] S3 workflow execution: Bootstrap resampling and stability assessment: Perform 500 Bootstrap resampling operations, and reconstruct the causal graph and estimate the effect after each resampling.
[0030] Findings: The path "keyword advertising -> visit landing page -> register" appeared 450 times (frequency 90%), with a 95% confidence interval of [8%, 12%] (i.e., an increase in registration rate of 8 to 12 percentage points). The path "keyword advertising -> direct registration" appeared 50 times (frequency 10%), with a CI of [-1%, 2%).
[0031] Conclusion: Path 1 (via landing page) is stable and effective; Path 2 (direct effect) is unstable and its effect is not significant and can be ignored.
[0032] Training the causal forest and contribution analysis: Based on path 1, select the sequence of influencing factors with high confidence: [user's industry, search keyword matching type, ad creative version, user's company size].
[0033] Using these factors as characteristics, a causal forest model is trained to predict the CATE (Certification Optimization) for each potential user.
[0034] Calculate feature contribution: Search keyword matching type: Contribution 0.40 User's industry: Contribution 0.35 Ad creative version: Contribution 0.20 User company size: Contribution 0.05 Analysis: Keyword matching degree is the most important driver of heterogeneity in results.
[0035] NLG generates decision-making schemes: Input: All of the above results (path stability, effect CI, contribution ranking).
[0036] Excerpt from NLG output report: Search Advertising Causal Impact Decision Support Solution Key findings: Search ads primarily drive registrations by attracting users to landing pages, a robust and reliable path (90% occurrence rate) that delivers an average 10% increase in registration rate (95% CI: 8%-12%).
[0037] Performance-driven analysis: Keyword match type is the primary driver of performance (contributing 40%). Ads using exact match keywords resulted in up to a 15% increase in registration rates, while broad match only achieved a 5% increase.
[0038] User industry is the second driving factor (contributing 35%). User response rates from the finance and technology sectors are significantly higher than those from other industries.
[0039] In advertising creatives (contributing 20%), versions emphasizing "free trial" performed better than versions emphasizing "features".
[0040] Decision Recommendation: Budget Allocation: Allocate more than 80% of your search advertising budget to "exact match" keywords.
[0041] Targeting strategy: Prioritize targeting individuals in the "finance" and "technology" industries and increase bids.
[0042] Creative optimization: Advertise entirely with "free trial" oriented creatives and suspend ineffective creatives.
[0043] Landing page optimization: Since registration conversion is highly dependent on the landing page, it is recommended to start A / B testing to further optimize the landing page design.
[0044] By evaluating the stability (frequency of occurrence) and effect confidence of each path in the causal graph of digital advertising through Bootstrap resampling, highly reliable paths are selected. Then, a causal forest model is used to quantify the individual contribution of each influencing factor to heterogeneous effects, identifying key driving factors. Finally, NLG technology is used to automatically generate a decision support solution containing key findings and actionable suggestions from the above quantitative insights. Its core value lies in transforming the statistical conclusions of causal inference into robust, interpretable, and directly actionable marketing action guidelines, achieving end-to-end automation from data validation to strategy generation, greatly improving the accuracy and efficiency of advertising decisions.
[0045] Reference Figure 2 As shown, an intelligent decision support system based on causal inference includes: The causal relationship analysis module and the directed graph analysis module are used to read the historical multi-dimensional operational data of various digital advertising businesses based on the digital advertising marketing backend database, discover and analyze the causal structure of the multi-dimensional operational data according to the mixed causal structure, and establish a causal directed acyclic graph for each digital advertising business. The causal weighted analysis module is electrically connected to the causal association analysis module. The causal weighted analysis module is used to identify and estimate causal effects based on the causal directed acyclic graph of each digital advertising business, according to the causal structure of the multivariate data in the causal directed acyclic graph, marking the effect state between causal variables, and establishing the causal directed weighted acyclic graph of each digital advertising business. The decision support generation module is electrically connected to the causal weighted analysis module. The decision support generation module is used to analyze the stability and heterogeneity of the causal structure of each path in the causal directed weighted acyclic graph of each digital advertising business, and then input it into the causal decision model of digital advertising marketing to generate causal impact decision support schemes for each digital advertising business.
[0046] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.
Claims
1. An intelligent decision support method based on causal inference, characterized in that, include: S1. Based on the digital advertising marketing backend database, read the historical multi-dimensional structure data of various digital advertising businesses, discover and analyze the causal structure of the multi-dimensional structure data of the operations according to the mixed causal structure, and establish a causal directed acyclic graph of each digital advertising business. S2. Based on the causal directed acyclic graph of each digital advertising business, identify and estimate causal effects according to the causal structure of the multivariate data in the causal directed acyclic graph, mark the effect state between causal variables, and establish a causal directed weighted acyclic graph of each digital advertising business. S3. Based on the causal directed weighted acyclic graph of each digital advertising business, analyze the stability and heterogeneity of the causal structure of each path in the causal directed weighted acyclic graph, substitute it into the causal decision-making model of digital advertising marketing, and generate causal impact decision support schemes for each digital advertising business.
2. The intelligent decision support method based on causal inference according to claim 1, characterized in that, Step S1 specifically includes: By using multiple interpolation, interpolation preprocessing is performed on the multi-dimensional structure data of the operation of each historical digital advertising business to generate several multi-dimensional structure datasets of the operation of each historical digital advertising business. Based on several operational multi-structure datasets from various historical digital advertising businesses, variable types were classified according to the cardinality of each structure data, and standardized according to the corresponding data codes. Based on several operational multivariate structure datasets of various historical digital advertising businesses, the cardinality categorical variable in each operational multivariate structure dataset is used as a node. All nodes are fully connected to construct an initial causal undirected acyclic graph of each digital advertising business. Conditional independence tests were conducted on the causal undirected acyclic graphs of the initial digital advertising businesses. The partial correlation coefficients between nodes and connected nodes in the causal undirected acyclic graphs of the initial digital advertising businesses were calculated and substituted into the test statistics to verify the significant correlation between nodes in the causal undirected acyclic graphs of the initial digital advertising businesses. Based on the significant correlations between nodes in the initial causal undirected acyclic graph of each digital advertising business, remove the unrelated independent edges from the initial causal undirected acyclic graph of each digital advertising business to obtain the causal undirected acyclic graph of each digital advertising business.
3. The intelligent decision support method based on causal inference according to claim 2, characterized in that, Step S1 also includes: Traverse each pair of adjacent nodes in the causal undirected acyclic graph of each digital advertising business, mark each pair of adjacent nodes as having legal edge addition operations according to the legal edge addition check rules, and generate a set of candidate edge addition operations for the causal undirected acyclic graph of each digital advertising business. Based on the set of candidate edge-adding operations for each digital advertising business's causal undirected acyclic graph, a temporary edge-adding causal undirected acyclic graph for each digital advertising business is generated. By utilizing the causal undirected acyclic graphs of temporary edge additions for each digital advertising business, candidate edge addition operations under local node connection structure changes in the causal undirected acyclic graphs of each digital advertising business are screened. The maximum score difference of each candidate edge addition operation is verified using the Bayesian information criterion. The node connection edges in the causal undirected acyclic graphs of each digital advertising business are then updated, as follows: ； in, Add edge operations to the candidate nodes and score differences under the changes in the local node connection structure of the causal undirected acyclic graphs of various digital advertising businesses. This is the set of new parent nodes for node Y after adding a new edge. This is the set of parent nodes of the original node Y. For the Bayesian information criterion function, Iterating the rules for checking the legal addition of edges until no legal edge addition operation can improve the score of the causal undirected acyclic graph with temporary edge additions for various digital advertising businesses.
4. The intelligent decision support method based on causal inference according to claim 3, characterized in that, Step S1 also includes: Traverse each pair of adjacent nodes in the causal undirected acyclic graph of each digital advertising business, mark each pair of adjacent nodes as having a valid deletion operation according to the valid deletion edge check rules, and generate a set of candidate deletion edge operations for each causal undirected acyclic graph of each digital advertising business. Based on the set of candidate deletion edge operations for the causal undirected acyclic graph of each digital advertising business, a temporary deletion edge causal undirected acyclic graph of each digital advertising business is generated. By utilizing the temporary deletion edge causal undirected acyclic graph of each digital advertising business, candidate deletion edge operations under the local node connection structure changes between the causal undirected acyclic graphs of each digital advertising business are screened. The maximum score difference of each candidate deletion edge operation is verified by Bayesian information criterion, and the node connection edges in the causal undirected acyclic graph of each digital advertising business are updated. Iterate through the rules for checking the legal edge deletion until no legal edge deletion operation can improve the score of the temporary edge deletion causal undirected acyclic graph for each digital advertising business, and obtain the causal directed acyclic graph for each digital advertising business.
5. The intelligent decision support method based on causal inference according to claim 4, characterized in that, Step S2 specifically includes: Based on the causal directed acyclic graphs of various digital advertising businesses, and following the backdoor criterion, the weighted average expected value of the choice paths between all operational decision variables A and advertising business outcome variables B is used to select the adjustment set S of the causal directed acyclic graphs of each digital advertising business, as follows: ； in, The weighted average expected value of the choice paths among all operational decision variables A and advertising business outcome variables B. For the value of the advertising business outcome variable, The weighted average expected value of all selected paths in the adjustment set is distributed as a whole average. Based on the causal directed acyclic graphs of various digital advertising businesses, and following the front-door criterion, the weighted average expected value of the mediating variable C along the choice path between all operational decision variables A and advertising business outcome variables B is used to select the variable set U of the causal directed acyclic graphs of each digital advertising business. The method is as follows: ； in, The weighted average expected value of the mediating variable C, representing the choice path between all operational decision variables A and advertising business outcome variables B. Let the set of variables be the weighted average expected value of all chosen paths, distributed as a whole. The value of the mediating variable for advertising business. The conditional probability of the nearest mediator variable C between all operational decision variables A and advertising business outcome variables B.
6. The intelligent decision support method based on causal inference according to claim 5, characterized in that, Step S2 also includes: Based on the adjustment set S of the causal directed acyclic graph of each digital advertising business and the variable set U of the causal directed acyclic graph of each digital advertising business, a path decision covariate vector of the causal directed acyclic graph of each digital advertising business is constructed. For each digital advertising business, the path decision covariate vector of the causal directed acyclic graph is divided into a mediated exposure array and a control non-mediated exposure array. A decision tree of expected value of digital advertising business decision path for mediated influencing factors and a decision tree of expected value of digital advertising business decision path for non-mediated influencing factors are established. A random forest of expected factors of decision path for each digital advertising business is constructed, and a set of causal concerns for non-mediated path and mediated path decision path of digital advertising business is generated. Based on Logistic Regression, the non-mediated path and the mediated path of the decision-making path of digital advertising business are taken as input, and the non-mediated path and the mediated path of the causal directed acyclic graph of each digital advertising business are taken as output. By using Blending, the non-mediated path and mediated path decision probabilities of the causal directed acyclic graph of each digital advertising business are linearly weighted and fused to obtain the optimal path of the causal directed acyclic graph of each digital advertising business, and generate the causal directed weighted acyclic graph of each digital advertising business.
7. The intelligent decision support method based on causal inference according to claim 6, characterized in that, Step S3 specifically includes: Resample the multi-dimensional structure data of the operation of each digital advertising business in history, execute the entire process from step S1 to step S2, count the frequency of occurrence of causal paths of each digital advertising business in the causal directed weighted acyclic graph, and use the Bootstrap distribution to calculate the confidence interval of the influencing factors of the causal paths of each digital advertising business. Based on the confidence intervals of the influencing factors of the causal paths of various digital advertising businesses, the sequence of influencing factors of digital advertising businesses in the causal directed weighted acyclic graph of each digital advertising business is screened, a CATE causal forest is trained, a causal decision model for digital advertising marketing is established, and the individual contribution values of the influencing factors of the causal paths of each digital advertising business are analyzed.
8. The intelligent decision support method based on causal inference according to claim 7, characterized in that, Step S3 also includes: Based on the confidence intervals of the influencing factors in the causal paths of various digital advertising businesses and the individual contribution values of the influencing factors in the causal paths of various digital advertising businesses, the NLG (Natural Language Generating Machine Learning) is used to generate analysis reports for each digital advertising business according to the average increase in the individual contribution values of the influencing factors based on the confidence intervals of the influencing factors in the causal paths of various digital advertising businesses, thus obtaining the causal impact decision support scheme for each digital advertising business.
9. An intelligent decision support system based on causal inference, characterized in that, An intelligent decision support method based on causal inference as described in any one of claims 1-8 includes: The causal relationship analysis module and the directed graph analysis module are used to read the historical multi-dimensional operational data of various digital advertising businesses based on the digital advertising marketing backend database, discover and analyze the causal structure of the multi-dimensional operational data according to the mixed causal structure, and establish a causal directed acyclic graph for each digital advertising business. The causal weighted analysis module is electrically connected to the causal association analysis module. The causal weighted analysis module is used to identify and estimate causal effects based on the causal directed acyclic graph of each digital advertising business, according to the causal structure of the multivariate data in the causal directed acyclic graph, marking the effect state between causal variables, and establishing the causal directed weighted acyclic graph of each digital advertising business. The decision support generation module is electrically connected to the causal weighted analysis module. The decision support generation module is used to analyze the stability and heterogeneity of the causal structure of each path in the causal directed weighted acyclic graph of each digital advertising business, and then input it into the causal decision model of digital advertising marketing to generate causal impact decision support schemes for each digital advertising business.