Heterogeneity marketing strategy evaluation method based on payment transaction data

Abstract

The invention discloses a heterogeneous marketing strategy evaluation method based on payment transaction data, and relates to the technical field of causal inference. The method comprises the following steps: constructing a data table between merchant transactions and operation conditions according to payment transaction data of merchants; the data table comprises observation data corresponding to a covariable, a processing variable, a tool variable and a result variable; constructing a first estimator and a second estimator according to the observation data of the covariable, the tool variable and the processing variable; according to all the observation variables, the first estimator and the second estimator, obtaining a third estimator used for representing a real condition average processing effect; inputting the covariable data of the to-be-evaluated merchant into a third estimator to obtain a real condition average processing effect calculation result; the real condition average processing effect calculation result is used for representing the influence of the stored value marketing activity on the operation condition of the merchant belonging to the corresponding operation classification. According to the method, the influence of invalid tool variables in marketing strategy evaluation can be solved.

Description

Technical Field

[0001] The present invention relates to the technical field of causal inference, and in particular to a heterogeneous marketing strategy evaluation method based on payment transaction data. Background Art

[0002] At present, in today's data-driven era, accurately evaluating the effects of strategy implementation is the key for enterprises to stand out in the fierce market competition.

[0003] In the field of causal inference, various methods have been used to assess the effects of marketing strategies. For example, instrumental variables are used to estimate causal effects. However, the classic instrumental variable approach requires that the instrumental variable only affects the treatment variable and not directly affects the outcome variable, i.e., it must satisfy the exclusivity condition. However, in practical applications, it is often difficult to find instrumental variables that fully meet this assumption, resulting in biased estimates from traditional causal inference methods, which in turn affects the scientific and reliable nature of decision-making. For example, a merchant's QR code ordering system, as an instrumental variable, ostensibly influences whether customers choose to use the top-up feature (the treatment variable), thereby affecting the merchant's sales (the outcome variable). However, in actual commercial operations, the QR code ordering system may bypass the impact of the top-up strategy and directly affect the merchant's sales through other channels, such as optimizing the ordering process, improving service efficiency, and increasing table turnover. This phenomenon violates the assumptions of the classic instrumental variable approach, making it difficult for traditional instrumental variable methods to provide accurate causal effect estimates.

[0004] Current research primarily focuses on cases where instrumental variables are effective, but fails to effectively address situations where instrumental variables directly influence outcomes, leading to biased estimates of causal effects. When instrumental variables are ineffective, existing methods lack targeted adjustment mechanisms, resulting in unrobust estimates. Therefore, a method that can address the impact of ineffective instrumental variables in marketing strategy evaluation is urgently needed. Summary of the Invention

[0005] Based on this, it is necessary to provide a heterogeneous marketing strategy evaluation method based on payment transaction data to address the above technical problems. This method can solve the impact of invalid instrumental variables in marketing strategy evaluation.

[0006] The present invention adopts the following technical solutions: The present invention provides a heterogeneous marketing strategy evaluation method based on payment transaction data, comprising: Based on the merchant's payment transaction data, a data table is constructed to link merchant transactions and business performance. The data table includes observational data corresponding to covariates, treatment variables, instrumental variables, and outcome variables. Covariates represent the business classification of the three-level industry; treatment variables indicate whether the merchant has enabled the stored value function; instrumental variables indicate whether the merchant has enabled the scan code ordering function; and outcome variables indicate the merchant's monthly order volume and the number of repeat customers. Based on the observed data of covariates, instrumental variables, and treatment variables, a first estimator and a second estimator are constructed; the first estimator is used to estimate the probability of a merchant enabling the scan code ordering function given the covariates; the second estimator is used to estimate the probability of a merchant enabling the stored value function given the covariates and instrumental variables; Based on all observed variables and the first and second estimators, a third estimator is obtained to characterize the true conditional average treatment effect; the true conditional average treatment effect represents the average difference in the outcome variable between individuals who have the same covariate values and those who do not have the storage function enabled; The covariate data of the merchants to be evaluated are input into the third estimator to obtain the calculation results of the true conditional average treatment effect; the calculation results of the true conditional average treatment effect are used to represent the impact of the stored-value marketing activities on the operating conditions of merchants belonging to the corresponding business classification.

[0007] Optionally, constructing a first estimator and a second estimator based on the observed data of the covariates, the instrumental variables, and the treatment variables includes: The covariate is used as the explanatory variable and the instrumental variable as the response variable. The machine learning algorithm is used to train the covariate and instrumental variable observation data to obtain the first estimator. The covariates and instrumental variables are used as explanatory variables and the treatment variables are used as response variables. The second estimator is constructed by training using a machine learning algorithm based on the observed data of the covariates, instrumental variables and treatment variables.

[0008] Optionally, a third estimator for characterizing the true conditional average treatment effect is obtained based on all observed variables and the first estimator and the second estimator, including: Based on all observed variables and the first and second estimators, a machine learning algorithm is used to construct a heterogeneity estimator of the nuisance parameter that affects the estimate of the true conditional mean treatment effect; According to the first estimator, the second estimator, the heterogeneity estimator and the observed data, the true conditional average treatment effect estimation equation is constructed, and the objective function of the generalized random forest is constructed according to the estimation equation; Based on the observed data, a generalized random forest is trained with the goal of minimizing the objective function to obtain a third estimator for characterizing the true conditional average treatment effect.

[0009] Optionally, the interference parameters that affect the estimation of the true conditional average treatment effect include the conditional average treatment effect, the residual after eliminating the conditional average treatment effect, and the unmeasured confounders in the conditional average treatment effect; the heterogeneity estimator includes a fourth estimator, a fifth estimator, and a sixth estimator; based on all observed variables and the first estimator and the second estimator, a machine learning algorithm is used to construct a heterogeneity estimator for the interference parameters that affect the estimation of the true conditional average treatment effect, including: Construct an explicit expression for the conditional mean treatment effect based on the first estimator, the second estimator, and all variables; The covariates are used as explanatory variables and the explicit expressions are used as response variables. A fourth estimator is constructed using a machine learning algorithm based on the observed data of all variables. The fourth estimator is used to obtain the estimated value of the conditional mean treatment effect. The covariates and instrumental variables are used as explanatory variables, and the estimated value of the residual after eliminating the conditional average treatment effect is used as the response variable. The fifth estimator is constructed using a machine learning algorithm based on the observed data of all variables. The sixth estimator was constructed using a machine learning algorithm based on the observed data of all variables, with the covariates as explanatory variables and the estimated values of the unmeasured confounders in the conditional mean treatment effect as the response variable.

[0010] Optionally, an explicit expression for the conditional mean treatment effect for: ; in, represents the observation data, X represents the observed data of the covariate, A represents the observed data of the treatment variable, Z represents the observation data of the instrumental variable, Y represents the observed data of the outcome variable, represents the first estimator, represents the second estimator, represents the estimate of the treatment regression residual, Indicates that in a given X and Z The output of the second estimator is Indicates that in a given X The output of the first estimator is, Represents an instrumental variable Z =1 when processing variable A The variance estimate of Represents an instrumental variable Z =0 when processing variables A The variance estimate of , , Indicates a givenX 、 Z = 1, the output of the second estimator, Indicates a given X 、 Z =0 when the output of the second estimator.

[0011] Optionally, use as the residual after removing the conditional average treatment effect of which, For a given X The output value of the fourth estimator.

[0012] Optionally, use Acts as an unmeasured confounder in the conditional mean treatment effect Estimates.

[0013] Alternatively, the true conditional mean treatment effect estimation equation is: ; ; in, Indicates that in a given X Down expectations, is the fourth estimator, is the fifth estimator, is the sixth estimator, For a given X and Z The output value of the fifth estimator is, For a given X The output value of the sixth estimator is, is the true conditional average treatment effect to be estimated, E(·) represents the expectation, For a given X The observed data of the outcome variable when the treatment variable is 1, For a given X The observed data of the outcome variable when the treatment variable is 0.

[0014] Optionally, the objective function of the generalized random forest is for: ; in, For the i The weight of each individual in the generalized random forest, n is the total number of individuals.

[0015] The present invention provides a heterogeneous marketing strategy evaluation device based on payment transaction data, comprising: The first construction module is used to construct a data table linking merchant transactions and business performance based on merchant payment transaction data. The data table includes observational data corresponding to covariates, treatment variables, instrumental variables, and outcome variables. Covariates represent the business classification of three-level industries; treatment variables indicate whether merchants have enabled the stored value function; instrumental variables indicate whether merchants have enabled the scan code ordering function; and outcome variables indicate the merchant's monthly order volume and the number of repeat customers. The second construction module is used to construct a first estimator and a second estimator based on the observed data of the covariates, instrumental variables, and treatment variables; the first estimator is used to estimate the probability of the merchant enabling the scan code ordering function given the covariates; the second estimator is used to estimate the probability of the merchant enabling the stored value function given the covariates and instrumental variables; a third building block for obtaining a third estimator for representing the true conditional average treatment effect based on all observed variables, the first estimator, and the second estimator; the true conditional average treatment effect represents the average difference in the outcome variable between individuals with the same covariate values who have the stored value function enabled and those who have not enabled the stored value function; The determination module is used to input the covariate data of the merchant to be evaluated into the third estimator to obtain the calculation result of the true conditional average treatment effect; the calculation result of the true conditional average treatment effect is used to represent the impact of the stored-value marketing activity on the operating conditions of the merchants belonging to the corresponding business classification.

[0016] The present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-mentioned heterogeneous marketing strategy evaluation method based on payment transaction data.

[0017] The present invention provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the above-mentioned heterogeneous marketing strategy evaluation method based on payment transaction data is implemented.

[0018] At least one of the above technical solutions adopted by the present invention can achieve the following beneficial effects: By constructing a first estimator to estimate the probability of a merchant enabling the scan-to-order feature (instrumental variable), and a second estimator to estimate the probability of a merchant enabling the stored-value feature (treatment variable) given the covariates and instrumental variables, this stratified estimation method controls for differences in merchants' underlying characteristics, such as covariates, resulting in a more accurate estimate of the relationship between the instrumental variable (scan-to-order feature) and the treatment variable (stored-value feature). The first and second estimators are then used to estimate the true conditional average treatment effect. This approach, while accounting for the instrumental variables, integrates all observed variables, including covariates and treatment variables, capturing the heterogeneity of confounding factors across merchant groups. This reduces the influence of invalid instrumental variables on the final estimation results, resulting in a more accurate estimate of the true conditional average treatment effect and a precise assessment of the impact of stored-value marketing activities on merchant performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings: Figure 1 A schematic flow chart of a heterogeneous marketing strategy evaluation method based on payment transaction data provided by the present invention; Figure 2 A flow chart of another heterogeneous marketing strategy evaluation method based on payment transaction data provided by the present invention; Figure 3 A schematic diagram of a heterogeneous marketing strategy evaluation device based on payment transaction data provided by the present invention; Figure 4 A schematic diagram of a computer device for implementing a heterogeneous marketing strategy evaluation method based on payment transaction data provided by the present invention. DETAILED DESCRIPTION

[0020] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with specific embodiments of the present invention and corresponding drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0021] Researchers have increasingly focused on more refined causal effect metrics, including the Conditional Average Treatment Effect (CATE) to better reflect the heterogeneity of treatment effects and support personalized strategies. CATE is used to assess the heterogeneous impact of treatments on outcomes under given covariates (or conditions). In heterogeneous marketing strategy evaluations based on large-scale payment transaction data, CATE can reveal treatment effects across different merchant strategies, thus supporting personalized strategies.

[0022] While existing methods such as dual machine learning, causal forests, and generalized random forests perform well in handling heterogeneity and nonlinearity, they have significant limitations in the case of invalid instrumental variables. These methods typically rely on the validity assumption of instrumental variables, which is difficult to meet in practice, and lack mechanisms to adjust the estimated treatment effect of decisions when instrumental variables fail.

[0023] While various methods exist for evaluating the effectiveness of marketing strategies in causal inference, these methods often rely on a series of stringent assumptions that are difficult to meet in real-world scenarios, compromising the accuracy and robustness of the estimates. In a highly competitive business environment, misjudgments can have serious, even fatal, consequences. If a merchant miscalculates the impact of a stored-value feature on sales, they may overinvest or neglect it, leading to irrational allocation of funds and even impacting long-term profitability. Decisions based on inaccurate causal effect estimates can deviate investment from market demand, causing merchants to misdirect their responses at critical marketing moments, miss out on traffic dividends, and weaken their competitiveness. Many businesses rely on data analysis to refine their operations. Inaccurate causal effect estimates can impact pricing, promotional, and user engagement strategies, ultimately making the entire business model unsustainable.

[0024] Current research primarily focuses on cases where instrumental variables are valid, while addressing invalid instrumental variables is limited, making it difficult to adapt to the real-world causal inference requirements. Traditional causal inference methods struggle to address invalid instrumental variables. Existing instrumental variable methods (such as doubly robust estimation and generalized equation of moments estimation) assume that instrumental variables strictly satisfy exclusivity conditions and cannot effectively address cases where instrumental variables directly influence outcomes, leading to biased causal effect estimates. Existing causal forest methods lack adaptability to invalid instrumental variables. Machine learning methods such as causal forests and instrumental causal forests primarily focus on improving the efficiency and flexibility of treatment effect estimation. However, when instrumental variables are invalid, existing methods lack targeted adjustment mechanisms, resulting in unrobust estimates and making them difficult to apply to business decisions. Although some studies have proposed methods using partially valid or modified instrumental variables, these methods often rely on specific data structures or make strong assumptions about the impact pathways of instrumental variables, limiting their applicability in complex business decision-making scenarios and hindering widespread application.

[0025] Therefore, to address the above issues, the present invention proposes a heterogeneous marketing strategy evaluation method based on payment transaction data. This method is a causal inference method with multiple robust estimates for situations where valid instrumental variables are unavailable. Specifically, it uses a generalized random forest (GRF) for estimation to address the impact of invalid instrumental variables in marketing strategy evaluation, and establishes statistical theory to guarantee its effectiveness, thereby accurately calculating the conditional average treatment effect. This method combines the flexibility of machine learning algorithms with robust estimation theory to address the challenges of high-dimensional data, heterogeneous treatment effects, and invalid instrumental variables. For large-scale marketing strategy evaluation using payment transaction data, it ensures the interpretability and applicability of causal inference results, providing more scientific decision support for businesses. This method overcomes the estimation bias caused by invalid instrumental variables, fills the application gap of existing causal inference methods in complex business scenarios, and provides important theoretical and practical support for precision marketing and business optimization.

[0026] The technical solutions provided by various embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0027] Figure 1 The figure is a flow chart of a heterogeneous marketing strategy evaluation method based on payment transaction data in the present invention, which specifically includes the following steps: S101: Based on the merchant's payment transaction data, a data table is constructed to link the merchant's transactions and business conditions. The data table includes observation data corresponding to covariates, treatment variables, instrumental variables, and outcome variables. Covariates represent the business classification of the three-level industry. Treatment variables represent whether the merchant has enabled the stored value function. Instrumental variables represent whether the merchant has enabled the scan code ordering function. The outcome variable represents the merchant's monthly order quantity and the number of repeat purchasers.

[0028] Among them, the merchant's payment transaction data meets the preset conditions, for example, the number of transactions is not less than 30 and the number of transaction days is not less than 3 days.

[0029] Deal with feature imbalance issues, such as downsampling the QR code ordering function variables to ensure estimation effects.

[0030] Given a covariate X , instrumental variables Z , processing variables A and the outcome variable Y Based on the true conditional average treatment effect, that is, the effect of a certain behavior, decision or treatment on a specific individual X The outcome variable Y In the context of large-scale payment transaction data, the covariate X is the merchant's characteristics, and the processing variable A For business behavior, decision-making, outcome variables Y For business indicators that merchants care about, instrumental variables Z Should choose to influence A To influence the outcome variable Y As shown in Table 1, Table 1 is a data table.

[0031] Table 1

[0032] When the instrumental variable fails, the conditional average treatment effect of heterogeneity is considered. This method is based on the following assumptions: (1) Unconfoundedness of treatment effects (1); in, Represents an instrumental variable Z The value of Indicates processing variables A The value of Indicates that when the processing variable takes the value , the outcome variable under the condition that the instrumental variable takes the value z Y The value of represents unobserved confounding factors, , , Indicates Z 、 X 、 U Processing variables under value A .

[0033] The above formula (1) indicates that and processing variables A are independent of each other, that is, under control Z 、 and After these variables, process the variables A The value of will not affect the result variable Y distribution.

[0034] (2) Randomness of instrumental variables (2); in, Indicates that when the processing variable takes the value A , the instrumental variable takes the value Z Conditional outcome variable Y The value of X and U Next Z Independent of each other, Indicates that in a given X Next Z and U Independent of each other.

[0035] (3) Heterogeneous Conditional Structural Equation (3); (4); in, and is the observed covariate A measurable and square-integrable function of Unobserved confounders The residual effect.

[0036] (4) Heteroskedasticity and positivity , (5); in, .

[0037] True conditional mean treatment effect ;in, is the heterogeneous treatment effect, that is, given X The true conditional average treatment effect after , E(·) represents the expectation, are the values of the result variable when the processing variable is 1 and the processing variable is 0 respectively.

[0038] S102, constructing a first estimator and a second estimator based on the observed data of covariates, instrumental variables and treatment variables; the first estimator is used to estimate the probability of a merchant enabling the scan code ordering function under given covariates; the second estimator is used to estimate the probability of a merchant enabling the stored value function under given covariates and instrumental variables.

[0039] Optionally, a first estimator and a second estimator are constructed based on the observed data of covariates, instrumental variables and treatment variables, including: taking the covariates as explanatory variables and the instrumental variables as response variables, and training them using a machine learning algorithm based on the observed data of the covariates and instrumental variables to obtain the first estimator; taking the covariates and instrumental variables as explanatory variables and the treatment variables as response variables, and training them using a machine learning algorithm based on the observed data of the covariates, instrumental variables and treatment variables to construct the second estimator.

[0040] Specifically, we use regression forest or other machine learning algorithms to learn interference parameters: the probability of merchants enabling the scan code ordering function and the processing propensity score. X (Third-level industry) as explanatory variables and instrumental variables Z (whether the scan code ordering function is enabled, etc.) as the response variable, using X 、 Z The observed data is trained using machine learning algorithms such as regression forest method to estimate the probability that merchants enable the scan code ordering function. , and get the first estimator ; Use merchant covariates X (Third-level industry), instrumental variables Z As the explanatory variable, the merchant enables the stored value function as the response variable, using X 、 Z 、 A The observation data is trained using machine learning algorithms such as regression forests to estimate the treatment propensity score , and obtain the second estimator , processing propensity score It represents the probability of estimating the merchant enabling the stored value function under given covariates and instrumental variables.

[0041] S103, based on all observed variables and the first estimator and the second estimator, obtain a third estimator for characterizing the true conditional average treatment effect; the true conditional average treatment effect represents the average difference in the outcome variable between individuals with the same covariate values who enable the storage function and those who do not enable the storage function.

[0042] Among them, according to all the observed variables and the first estimator and the second estimator, a third estimator for characterizing the true conditional average treatment effect is obtained, including: according to all the observed variables and the first estimator and the second estimator, a heterogeneity estimator of the interference parameter that affects the estimation of the true conditional average treatment effect is constructed using a machine learning algorithm; according to the first estimator, the second estimator, the heterogeneity estimator and the observed data, a true conditional average treatment effect estimation equation is constructed, and the objective function of the generalized random forest is constructed according to the estimated equation; based on the observed data, the generalized random forest is trained with the goal of minimizing the objective function to obtain the third estimator for characterizing the true conditional average treatment effect.

[0043] Optionally, the interference parameters affecting the estimation of the true conditional average treatment effect include the conditional average treatment effect, the residual after eliminating the conditional average treatment effect, and the unmeasured confounders in the conditional average treatment effect; the heterogeneity estimator includes a fourth estimator, a fifth estimator, and a sixth estimator; based on all observed variables and the first estimator and the second estimator, a machine learning algorithm is used to construct a heterogeneity estimator of the interference parameters affecting the estimation of the true conditional average treatment effect, including: constructing an explicit expression of the conditional average treatment effect based on the first estimator, the second estimator, and all variables; taking the covariate as the explanatory variable and the explicit expression as the response variable, and using the machine learning algorithm to construct the fourth estimator based on the observed data of all variables; the fourth estimator is used to obtain the estimated value of the conditional average treatment effect; taking the covariate and the instrumental variable as the explanatory variables and the estimated value of the residual after eliminating the conditional average treatment effect as the response variable, and using the machine learning algorithm to construct the fifth estimator based on the observed data of all variables; taking the covariate as the explanatory variable and the estimated value of the unmeasured confounder in the conditional average treatment effect as the response variable, and using the machine learning algorithm to construct the sixth estimator based on the observed data of all variables.

[0044] Among them, the first estimator is used and the second estimator As input, and the remaining variables X 、 Y 、 A 、 Z The observed variables are used to estimate the interference parameters using a sequence of machine learning algorithms such as regression forests: the conditional mean treatment effect, the residual after eliminating the conditional mean treatment effect, and the unmeasured confounders in the conditional mean treatment effect.

[0045] Alternatively, the conditional mean treatment effect Explicit expression of for: (6); It can be proved that in If the estimate is correct The expectation of the conditional mean treatment effect .in, represents the observation data, X represents the observed data of the covariate, A represents the observed data of the treatment variable, Z represents the observation data of the instrumental variable, Y represents the observed data of the outcome variable, represents the first estimator, represents the second estimator, represents the estimate of the treatment regression residual, Indicates that in a given X and Z The output of the second estimator is Indicates that in a given X The output of the first estimator is, Represents an instrumental variable Z =1 when processing variable A The variance estimate of Represents an instrumental variable Z =0 when processing variables A The variance estimate of , , Indicates a given X 、 Z = 1, the output of the second estimator, Indicates a given X 、 Z =0 when the output of the second estimator.

[0046] The covariates are used as explanatory variables and the explicit expressions are used as response variables. The fourth estimator is constructed using a machine learning algorithm based on the observed data of all variables, including: using covariates X As explanatory variables, the function As the response variable, using X , Z , A The observed data is trained using machine learning algorithms such as regression forests to estimate , and obtain the fourth estimator , the fourth estimator is used to obtain the conditional mean treatment effect estimate .

[0047] Optionally, given The residual after removing the conditional mean treatment effect is defined as: ;in, is the conditional average treatment effect under a given X. The residual after removing the conditional average treatment effect is It reflects the result of the combined effect of other factors that are not explained by the treatment effect.

[0048] In order to estimate the given The residual after removing the conditional mean treatment effect As an estimate of the residual after removing the conditional mean treatment effect For a given X The output of the fourth estimator, specifically, using the covariate X and instrumental variables Z As an explanatory variable, As The estimated value of the response variable is trained using a machine learning algorithm (such as regression forest) to estimate , and get the fifth estimator ,in, For the general Substitution Define the resulting estimate, For a given X The output value of the fourth estimator.

[0049] Optionally, use Acts as an unmeasured confounder in the estimation of the conditional mean treatment effect The estimate of represents the estimate of the treatment regression residual, Indicates a given X 、 Z =1 when the output of the second estimator.

[0050] To estimate , using covariates X As an explanatory variable, As the response variable, a machine learning method (such as regression forest) is used for training to estimate , and obtain the sixth estimator ;in, reflects unmeasured confounding factors in a particular condition, For use , Substitute into Define the resulting estimate, , 、 are all estimated values obtained in the previous steps.

[0051] It should be noted that, given assumptions (2) and (3), it can be proved that: , after appropriate mathematical transformation, we can get , 、 、 represents the true parameter to be estimated, 、 Respectively indicate that in the given X Next and .

[0052] This paper uses multiple robust estimation methods to estimate the true conditional average treatment effect of business decisions, combining multiple estimation functions to ensure that even if there are deviations in the estimation of some model parameters, the final results remain consistent. The key estimation function of this method is the efficiency impact function, which incorporates the propensity score , processing regression residuals , thus providing a robust estimate of the true conditional mean treatment effect, Indicates that in a given Z, X under conditions A = 1. The multiple robust influence functions are defined as follows: (7); in, Indicates a given Down The probability of , Indicates that in a given X and Z =1 when processing variable A The variance of Indicates that in a given X and Z =0 when processing variables A The variance of , , 、 、 、 and represents the real interference parameter to be estimated, Indicates that in a given X and Z Next .

[0053] Optionally, based on multiple robust influence functions, a true conditional average treatment effect estimation equation is constructed, and the true conditional average treatment effect estimation equation is: (8); (9); in, Indicates that in a given X Down expectations, is the fourth estimator, is the fifth estimator, is the sixth estimator, For a given X and Z The output value of the fifth estimator is, For a given X The output value of the sixth estimator is, For a given X The output value of the fourth estimator is, represents the final estimated true conditional average treatment effect, E(·) represents the expectation, For a given X The observed data of the outcome variable when the treatment variable is 1, For a given X The observed data of the outcome variable when the treatment variable is 0.

[0054] The true conditional average treatment effect under the multiple robust effective influence functions proposed in this method The implicit expression of , is the true conditional average treatment effect to be estimated.

[0055] According to the estimation equation, define the objective function of the generalized random forest. Optionally, the objective function of the generalized random forest is for: (10); in, For the i The weight of each individual in the generalized random forest, n is the total number of individuals.

[0056] Based on the observed data, the generalized random forest is trained with the objective function minimized to obtain the third estimator for characterizing the true conditional average treatment effect. Specifically, the generalized random forest is constructed with the objective function minimized to obtain the third estimator for characterizing the true conditional average treatment effect. , the observation data Input to the generalized random forest, train the generalized random forest, and obtain the third estimator used to characterize the true conditional average treatment effect , the output of the third estimator is the multi-robust heterogeneity estimate.

[0057] In the present invention, based on the first estimator and the second estimator, a machine learning algorithm is used to construct a heterogeneity estimator of the interference parameters that affect the estimation of the true conditional average treatment effect, and finally the true conditional average treatment effect estimation equation is used to estimate the true conditional average treatment effect. In this way, while considering the instrumental variables, all observed variables such as covariates and treatment variables can be integrated to capture the heterogeneity of interference factors in different merchant groups. When the instrumental variables are invalid, they may be mixed with other interference factors to affect the results. The heterogeneity estimator can more comprehensively characterize and separate these interference factors, avoiding the interference caused by invalid instrumental variables from being mistakenly attributed to the effect of the treatment variable (stored value function), reducing the interference of invalid instrumental variables on the final estimation results, making the true conditional average treatment effect estimate more accurate, and accurately evaluating the impact of stored value marketing activities on the business conditions of merchants.

[0058] S104, inputting the covariate data of the merchant to be evaluated into a third estimator to obtain a true conditional average treatment effect calculation result; the true conditional average treatment effect calculation result is used to represent the impact of the stored-value marketing activity on the operating conditions of the merchants belonging to the corresponding business classification.

[0059] For each merchant to be evaluated , and its covariate Substitute into the third estimator Different Next, we get the processing variables A The true conditional mean treatment effect calculation results The true conditional average treatment effect of stored-value marketing activities is the impact on the operating conditions of merchants with different attributes.

[0060] The robustness and consistency of the multiple robust influence functions are guaranteed by the following theorem: Theorem: Given assumptions (1)–(4), the conditional mean treatment effect of a stored-value marketing campaign can be identified by the following three different expressions, each involving a different set of nuisance parameters.

[0061] (1) Explicit expression: , is established almost everywhere. Among them, , Represents the processed regression residuals.

[0062] (2) Implicit expression: , which is established almost everywhere. .

[0063] (3) Implicit expression: , is established almost everywhere. Among them, .

[0064] Multiple robust influence functions , based on the representations under the three different interference parameter sets mentioned above, has multiple robustness, that is, if at least one of the following conditions holds: (i) (ii) (iii) ,but is a consistent asymptotically normal estimator of the conditional mean treatment effect of the stored-value marketing campaign.

[0065] like Figure 2 As shown, Figure 2 Another flow chart of the heterogeneous marketing strategy evaluation method based on payment transaction data provided by the present invention, specifically comprising: first obtaining payment transaction data, preprocessing the data to obtain a data table, and then constructing a first estimator through a machine learning algorithm and the second estimator , the first estimator By covariate X and instrumental variables Z Constructed through X Estimating the probability of merchants enabling the scan-to-order function , the second estimator By covariate X , instrumental variables Z and process variables A Constructed by covariates X and instrumental variables Z Estimating treatment propensity scores .

[0066] Then based on the first estimator , the second estimator and the number of observations using a machine learning algorithm to construct a fourth estimator , the fifth estimator and the sixth estimator ; The fourth estimator By covariate X and function Constructed, mainly through covariates X Estimated conditional mean treatment effect estimate ; The fifth estimator By covariate X , instrumental variables Z and function Constructed, mainly through covariates X and instrumental variables Z Estimate the residual after removing the conditional mean treatment effect .

[0067] Sixth estimator By covariate X and function Constructed, mainly through covariates X Unmeasured confounders in estimating conditional mean treatment effects .

[0068] Then based on the first estimator , the second estimator , the fourth estimator , the fifth estimator and the sixth estimator , construct multiple robust influence functions, and train the third estimator based on the multiple robust influence functions through the generalized random forest algorithm Finally, the third estimator outputs each merchant to be evaluated The true conditional mean treatment effect calculation results .

[0069] This paper extends generalized random forests to the case of invalid instrumental variables, adapting to high-dimensional data and complex interaction effects, making it suitable for causal inference. In the case of invalid instruments, after estimating the nuisance parameters, the GRF framework is introduced to estimate heterogeneous treatment effects. The final CATE estimate is derived using multiple robust influence functions. This ensures that even if some of the model's instrumental variable assumptions fail, the estimated results remain consistent. This improves the robustness of the estimate in complex data scenarios, reduces reliance on instrumental variable assumptions, and provides reliable estimates for causal inference. This approach is suitable for the complex data environment of heterogeneous marketing strategy evaluation based on large-scale payment data.

[0070] This paper introduces a multi-robust estimation framework, enabling accurate estimation of the conditional mean treatment effect even when instrumental variables fail. In the context of large-scale payment transaction data, this method accurately assesses the causal effects of heterogeneous marketing strategies, ensuring robustness and reliability of the estimation results even in the presence of instrumental variable failure, thus providing scientific decision support for businesses.

[0071] This paper overcomes the limitations of traditional causal inference in the context of invalid instrumental variables by creatively combining multiple robust estimates with generalized random forests, providing a new causal inference method that combines theoretical rigor, computational efficiency, and adaptability to business decisions. Specifically, the innovations of this paper are reflected in the following aspects: The introduction of multiple robust estimates ensures the robustness of causal inference. This method constructs a multiple robust influence function that maintains consistent estimation even when some model assumptions fail. Traditional causal inference methods typically rely on strict assumptions when dealing with invalid instrumental variables. However, this method solves multiple conditional moment estimation equations to ensure that causal effect estimates remain robust even when instrumental variables are invalid. This theoretical innovation fills a gap in the research on robust causal inference under invalid instrumental variables and provides a new paradigm for modeling high-dimensional complex causal relationships.

[0072] This method expands the GRF framework to enhance the ability to model complex causal relationships. This method introduces the GRF framework within the interference parameter estimation mechanism to estimate heterogeneous treatment effects. It also constructs an automatic adjustment mechanism to address the bias that may be introduced by invalid instrumental variables. This allows the GRF to not only adaptively learn heterogeneous treatment effects but also adjust for the impact of invalid instrumental variables. Compared to existing causal forest methods, this method is more robust and adaptable in high-dimensional data and complex causal scenarios, avoiding the structural bias that may occur with single causal inference methods.

[0073] Integrating multiple estimation methods improves model adaptability. By integrating multiple estimation methods, the method of the present invention significantly improves the reliability and accuracy of causal effect estimation, making it particularly suitable for analyzing large-scale payment transaction data in commercial marketing decision-making.

[0074] This method addresses the issue of invalid instrumental variables in commercial marketing decisions and provides support for precise decision-making. The application of traditional causal inference methods in commercial scenarios is often limited by strict assumptions, making it difficult to apply the causal effect estimates. This method specifically addresses the core issue of invalid instrumental variables impacting causal inference, providing a robust solution that enables businesses to optimize marketing strategies based on reliable causal inference results. This method can be applied to a variety of commercial scenarios, including personalized pricing, precision marketing, and user profile optimization, ensuring that businesses can make scientific, accurate, and efficient business decisions in a data-driven environment.

[0075] When applying the heterogeneous marketing strategy evaluation method based on payment transaction data provided by the present invention, it is not necessary to Figure 1 The steps are executed in the order shown. The specific execution order of the steps can be determined according to needs, and the present invention does not limit this.

[0076] The above is a heterogeneous marketing strategy evaluation method based on payment transaction data provided by one or more embodiments of the present invention. Based on the same idea, the present invention also provides a corresponding heterogeneous marketing strategy evaluation device based on payment transaction data, such as Figure 3 shown.

[0077] Figure 3This is a schematic diagram of a device for evaluating heterogeneous marketing strategies based on payment transaction data provided by the present invention. The device 300 includes: The first construction module 301 is used to construct a data table between merchant transactions and business conditions based on the merchant's payment transaction data; the data table includes observation data corresponding to covariates, treatment variables, instrumental variables and outcome variables; covariates represent the business classification of the three-level industry; treatment variables represent whether the merchant has enabled the stored value function; instrumental variables represent whether the merchant has enabled the scan code ordering function; and outcome variables represent the merchant's monthly order quantity and the number of repeat purchasing users.

[0078] The second construction module 302 is used to construct a first estimator and a second estimator based on the observed data of covariates, instrumental variables and treatment variables; the first estimator is used to estimate the probability that the merchant enables the scan code ordering function under given covariates; the second estimator is used to estimate the probability that the merchant enables the stored value function under given covariates and instrumental variables.

[0079] The third construction module 303 is used to obtain a third estimator for characterizing the true conditional average treatment effect based on all observed variables and the first estimator and the second estimator; the true conditional average treatment effect represents the average difference in the outcome variable between individuals with the same covariate value and those without the stored value function enabled.

[0080] The determination module 304 is used to input the covariate data of the merchant to be evaluated into the third estimator to obtain the true conditional average treatment effect calculation result; the true conditional average treatment effect calculation result is used to represent the impact of the stored-value marketing activity on the operating conditions of the merchants belonging to the corresponding business classification.

[0081] The specific limitations of the heterogeneous marketing strategy evaluation device based on payment transaction data can be found in the limitations of the heterogeneous marketing strategy evaluation method based on payment transaction data above and will not be repeated here. The various modules in the aforementioned heterogeneous marketing strategy evaluation device based on payment transaction data can be implemented in whole or in part through software, hardware, or a combination thereof. The aforementioned modules can be embedded in or independent of a processor in a computer device in hardware form, or can be stored in a memory in a computer device in software form, so that the processor can call and execute the corresponding operations of the aforementioned modules.

[0082] The present invention also provides a computer-readable storage medium, which stores a computer program, which can be used to execute the above Figure 1 A heterogeneous marketing strategy evaluation method based on payment transaction data is provided.

[0083] The present invention also provides Figure 4 The structural diagram of the computer equipment shown in FIG. Figure 4As shown in the figure, at the hardware level, the computer device includes a processor, an internal bus, a network interface, a memory, and a non-volatile memory. Of course, it may also include other hardware required for the business. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to achieve the above Figure 1 A heterogeneous marketing strategy evaluation method based on payment transaction data is provided.

[0084] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing the relevant hardware using a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes in the above-described method embodiments. Any reference to memory, storage, database, or other media used in the various embodiments provided herein may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can take various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0085] The technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of the present invention.

Claims

1. A heterogeneous marketing strategy evaluation method based on payment transaction data, characterized by: include: Based on the merchant's payment transaction data, a data table is constructed between the merchant's transaction and operating conditions; The data table includes the observation data corresponding to the covariates, treatment variables, instrumental variables, and outcome variables; The covariates represent the three-level industry classification of operations; The treatment variable indicates whether the merchant has enabled the stored value function; the instrumental variable indicates whether the merchant has enabled the scan code ordering function; the result variable indicates the merchant's monthly order quantity and the number of repeat purchasers; Based on the observed data of covariates, instrumental variables, and treatment variables, a first estimator and a second estimator are constructed. The first estimator is used to estimate the probability of a merchant enabling the scan code ordering function given the covariates. The second estimator is used to estimate the probability of a merchant enabling the stored value function given the covariates and instrumental variables; According to all observed variables and the first and second estimators, a third estimator for characterizing the true conditional average treatment effect is obtained; The true conditional mean treatment effect represents the average difference in the outcome variable between individuals who have the same value of the covariate and those who do not have the stored value function enabled; The covariate data of the merchant to be evaluated is input into the third estimator to obtain the calculation results of the true conditional average treatment effect; The calculated results of the true conditional average treatment effect are used to represent the impact of stored-value marketing activities on the operating conditions of merchants belonging to the corresponding business classification.

2. The method according to claim 1, characterized in that According to the observed data of covariates, instrumental variables and treatment variables, the first estimator and the second estimator are constructed, including: The covariate is used as the explanatory variable and the instrumental variable as the response variable. The machine learning algorithm is used to train the covariate and instrumental variable observation data to obtain the first estimator. The covariates and instrumental variables are used as explanatory variables and the treatment variables are used as response variables. The second estimator is constructed by training using a machine learning algorithm based on the observed data of the covariates, instrumental variables and treatment variables.

3. The method according to claim 1, characterized in that According to all observed variables and the first and second estimators, a third estimator is obtained to characterize the true conditional average treatment effect, including: Based on all observed variables and the first and second estimators, a machine learning algorithm is used to construct a heterogeneity estimator of the nuisance parameter that affects the estimate of the true conditional mean treatment effect; According to the first estimator, the second estimator, the heterogeneity estimator and the observed data, the true conditional average treatment effect estimation equation is constructed, and the objective function of the generalized random forest is constructed according to the estimation equation; Based on the observed data, a generalized random forest is trained with the goal of minimizing the objective function to obtain a third estimator for characterizing the true conditional average treatment effect.

4. The method according to claim 3, characterized in that The interference parameters that affect the estimation of the true conditional average treatment effect include the conditional average treatment effect, the residual after eliminating the conditional average treatment effect, and the unmeasured confounders in the conditional average treatment effect; the heterogeneity estimators include the fourth estimator, the fifth estimator, and the sixth estimator; Based on all observed variables and the first and second estimators, a machine learning algorithm is used to construct a heterogeneity estimator of the nuisance parameter that affects the estimate of the true conditional mean treatment effect, including: Construct an explicit expression for the conditional mean treatment effect based on the first estimator, the second estimator, and all variables; The covariates are used as explanatory variables and the explicit expressions are used as response variables. A fourth estimator is constructed using a machine learning algorithm based on the observed data of all variables. The fourth estimator is used to obtain the estimated value of the conditional mean treatment effect. The covariates and instrumental variables are used as explanatory variables, and the estimated value of the residual after eliminating the conditional average treatment effect is used as the response variable. The fifth estimator is constructed using a machine learning algorithm based on the observed data of all variables. The sixth estimator was constructed using a machine learning algorithm based on the observed data of all variables, with the covariates as explanatory variables and the estimated values of the unmeasured confounders in the conditional mean treatment effect as the response variable.

5. The method according to claim 4, characterized in that Explicit expression for conditional mean treatment effect for: ； in, represents the observation data, X represents the observed data of the covariate, A represents the observed data of the treatment variable, Z represents the observation data of the instrumental variable, Y represents the observed data of the outcome variable, represents the first estimator, represents the second estimator, represents the estimate of the treatment regression residual, Indicates that in a given X and Z The output of the second estimator is Indicates that in a given X The output of the first estimator is, Represents an instrumental variable Z =1 when processing variable A The variance estimate of Represents an instrumental variable Z =0 when processing variables A The variance estimate of , , Indicates a given X 、 Z = 1, the output of the second estimator, Indicates a given X 、 Z =0 when the output of the second estimator.

6. The method according to claim 5, characterized in that use as the residual after removing the conditional average treatment effect of which, For a given X The output value of the fourth estimator.

7. The method according to claim 6, characterized in that use Acts as an unmeasured confounder in the conditional mean treatment effect Estimates.

8. The method according to claim 6, characterized in that The true conditional mean treatment effect estimation equation is: ； ； in, Indicates that in a given X Down expectations, is the fourth estimator, is the fifth estimator, is the sixth estimator, For a given X and Z The output value of the fifth estimator is, For a given X The output value of the sixth estimator is, is the true conditional average treatment effect to be estimated, E(·) represents the expectation, For a given X The observed data of the outcome variable when the treatment variable is 1, For a given X The observed data of the outcome variable when the treatment variable is 0.

9. The method according to claim 8, characterized in that Objective function of generalized random forest for: ； in, For the i The weight of each individual in the generalized random forest, n is the total number of individuals.

Images