Client classification method, business processing method and system based on causal inference and interpretability

Abstract

The invention discloses a customer classification method, business processing method and system based on causal inference and interpretability, and the method comprises the steps: obtaining a customer classification model through pre-training: collecting historical data to obtain a training data set, the historical data comprising customer features corresponding to customers, business processing variables and result variables; inputting the training data set into a basic model based on a causal forest model and interpretability for training until the model converges to obtain a customer classification model; based on the causal forest model, determining an individual processing effect value of a customer according to customer characteristics, business processing variables and result variables; and calculating contribution values of the customer features to individual processing effects of the customers based on the interpretability, classifying the customers according to the individual processing effect values, the contribution values and the customer features, collecting the customer features to be classified, and inputting the customer features to a customer classification model to output a customer classification result. The accuracy and rationality of the customer classification result can be improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a customer classification method, business processing method and system based on causal inference and interpretability. Background Technology

[0002] Causal modeling is a core method in data science for exploring causal relationships between variables, and its development stems from overcoming the limitations of correlation analysis. Early causal inference relied primarily on basic statistical methods such as linear regression and hypothesis testing. However, with in-depth research into controlling for confounding variables and counterfactual reasoning, modern causal modeling has gradually incorporated theories such as probabilistic graphical models and counterfactual frameworks, enabling more precise separation of causal effects from confounding correlations. This technique has significant value in a wide range of fields, including medicine, social sciences, and business decision-making. For example, in policy effectiveness evaluation, causal modeling can distinguish between the actual impact of interventions and external environmental interferences, providing a scientific basis for strategy formulation.

[0003] The application of causal modeling needs to be combined with specific scenarios and data characteristics. For example, in the field of medical technology, it can be used to assess the causal effect of treatment methods on patient prognosis; in the social sciences, it can be used to analyze the impact of education policies on employment rates. However, practical applications still face challenges such as insufficient data quality and difficulty in observing hidden confounding variables, which limit the accuracy of causal effect estimation. In addition, traditional methods have limited performance in handling high-dimensional data and nonlinear relationships, and it is difficult to capture the dynamic interactions between variables in complex systems.

[0004] Current research is moving towards higher-order causal inference methods. Techniques such as causal forest models and dual robust estimation, based on nonparametric modeling and adaptive algorithms, are gradually overcoming the limitations of traditional methods in estimating effects with heterogeneity. Simultaneously, the integration of causal modeling with machine learning is becoming increasingly significant; for example, deep learning is used to capture nonlinear causal relationships, or reinforcement learning is used to simulate dynamic intervention strategies. These advances are driving the application of causal modeling in complex systems, providing more reliable support for multi-domain decision-making.

[0005] Model interpretability is a core element for the practical application of artificial intelligence, especially in high-risk decision-making scenarios. Its value lies in making decision-making processes transparent, meeting regulatory requirements, and enhancing user trust. With the widespread adoption of complex models such as deep learning, their "black box" nature makes the decision-making logic difficult to trace, leading to ethical controversies and compliance risks. Traditional model interpretation methods often rely on feature importance ranking or visualization techniques, but these methods have limited applicability to high-dimensional data and complex models. In recent years, researchers have proposed tools such as game theory-based feature contribution analysis (e.g., SHAP values) and locally interpretable methods (e.g., LIME) to reveal the internal decision-making logic of the model by quantifying the contribution of input features to the output.

[0006] In the financial sector, model interpretability has become a mandatory requirement for compliance reviews. For example, in scenarios such as loan approval and investment decisions, models must provide clear decision-making justifications to meet regulatory requirements for fairness and transparency. Furthermore, interpretability technologies allow business stakeholders to verify whether model logic aligns with domain common sense, such as detecting implicit bias or data bias. However, existing methods still have limitations when handling dynamic data and complex feature interactions. For instance, they struggle to capture the synergistic effects between nonlinear features or lack a unified interpretive framework in multi-model integration scenarios. This limits the potential of model interpretability in dynamic financial risk assessment and real-time decision-making.

[0007] Currently, research on model interpretability is focusing on improving the granularity and dynamic adaptability of explanations. Techniques such as feature association analysis based on attention mechanisms and incremental explanation frameworks are gradually solving the sensitivity problem of static explanation methods to changes in data distribution. Meanwhile, the combination of interpretability and causal modeling has become a hot topic; for example, revealing the dependencies between features through causal graphs further enhances the logicality of the explanation. However, existing machine learning techniques still lack interpretability in the financial field.

[0008] The above background information is provided only to assist in understanding the inventive concept and technical solution of this invention. It does not necessarily belong to the prior art of this application, nor does it necessarily provide technical teaching. In the absence of clear evidence that the above information was disclosed before the filing date of this application, the above background information should not be used to evaluate the novelty and inventiveness of this application. Summary of the Invention

[0009] The purpose of this invention is to provide a customer classification method, business processing method, and system based on causal inference and interpretability, which can improve the accuracy and rationality of customer classification results.

[0010] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0011] A customer classification method based on causal inference and interpretability includes the following steps:

[0012] Collect customer information to be classified and input it into a pre-trained customer classification model to output customer classification results. The customer information includes customer features. The customer classification model is trained in the following way:

[0013] Collect historical data and obtain a training dataset based on the historical data. The historical data includes customer characteristics, business processing variables, and outcome variables corresponding to customers. The business processing variables include those that have been recommended and those that have not been recommended. The outcome variables include those that have been successfully converted and those that have failed to convert.

[0014] The training dataset is input into the base model based on the causal forest model and interpretability to train the base model. The training steps include:

[0015] Based on the causal forest model, the individual treatment effect value of the customer is determined according to the customer characteristics, business processing variables, and outcome variables.

[0016] The contribution value of customer characteristics to the individual treatment effect of customers is calculated based on interpretability, and customers are classified according to the individual treatment effect value, the contribution value, and the customer characteristics;

[0017] Repeat the above training steps until the model converges to obtain the customer classification model.

[0018] Furthermore, following any one or a combination of the aforementioned technical solutions, the training dataset is obtained based on historical data, including the following steps:

[0019] The historical dataset is divided into a processing group and a control group. The business processing variables corresponding to the historical data in the processing group have been recommended, while the business processing variables corresponding to the historical data in the control group have not been recommended.

[0020] Predict each customer's preference score for the service based on the customer characteristics and the business processing variables;

[0021] The treatment group and the control group are matched based on the propensity scores to obtain a matched dataset;

[0022] The matched dataset is subjected to a balance test. If the test fails, customer groups with matching scores exceeding the preset value (too large) are removed and the balance test is repeated. If the test passes, the matched dataset is configured as a balanced dataset to be used as the training dataset.

[0023] Furthermore, following any one or a combination of the aforementioned technical solutions, a logistic regression model is used to estimate the business preference score for each customer. The calculation formula is as follows:

[0024]

[0025] Among them, β1, β2...β p X1, X2, ..., Xp are the regression coefficients of logistic regression, X1, X2, ..., Xp are customer feature 1, customer feature 2, ..., customer feature p, and β0 is the intercept term.

[0026] Furthermore, following any one or a combination of the aforementioned technical solutions, the control group and the control group are matched in the following manner:

[0027] For customer i in the treatment group, determine customer j in the control group. Customers i and j satisfy: Distance(i,j)=|P i -P j |Minimum, where P i P represents the propensity score corresponding to customer i in the treatment group. j This represents the propensity score corresponding to customer j in the control group.

[0028] Furthermore, following any one or a combination of the aforementioned technical solutions, the balance of the matched dataset is checked in the following manner:

[0029] Calculate the standardized mean difference of the matched dataset using the following formula:

[0030]

[0031] Among them, SMD k This represents the standardized mean difference. This indicates that the control group is based on customer characteristic X. k The mean of the above, This indicates that the control group is based on customer characteristic X. k The mean of the above, The pooled variance representing customer characteristics;

[0032] If|SMD k |<α, where α is a preset threshold. If 0<α<0.15, the balance test of the matched dataset passes; otherwise, it fails.

[0033] Furthermore, following any one or a combination of the aforementioned technical solutions, based on the causal forest model, the individual treatment effect value ITE for each customer i is estimated using the following formula:

[0034]

[0035] in, This represents the ITE value of customer i, where, This represents the expected value estimate when an individual receives treatment. This represents the expected value estimate when the individual did not receive treatment.

[0036] Furthermore, following any one or a combination of the aforementioned technical solutions, the base model is trained based on a causal forest model with the objective of minimizing the following loss function:

[0037]

[0038] Where θ represents the parameters of the causal forest model, including the number of trees and the number of split nodes. This is the individual treatment effect value for customer i. It is the baseline expected value before treatment, Y i This indicates whether customer i has converted.

[0039] Furthermore, based on any one or a combination of the aforementioned technical solutions, the contribution value of customer characteristics to the individual treatment effect of the customer is calculated based on interpretability. The formula for calculating the contribution value is as follows:

[0040]

[0041] Where, φ i,j Let S be a subset of customer features, |S| represent the number of customer features in subset S, ! denotes factorial operation, and p represent the total number of permutations of all customer features. This means that the individual treatment effect value of customer i is predicted using only a subset of customer features S. Let X represent the individual treatment effect value using the union of a subset S of customer characteristics and customer characteristic j, and let X represent the universal set of all customer characteristics. -j This represents the subset of all customer features excluding customer feature j.

[0042] Furthermore, following any one or a combination of the aforementioned technical solutions, classifying customers based on the individual treatment effect value, the contribution value, and the customer characteristics includes the following steps:

[0043] Constructing the clustering feature matrix Z i , Among them, X i This represents the customer characteristics of customer i. This represents the individual treatment effect value for customer i, and Top SHAP Features represent the contribution value.

[0044] The optimal number of clusters K for the clustering feature matrix is ​​determined by using K-Means or hierarchical clustering algorithms and combining them with the elbow method to divide the training sample set into K clusters.

[0045] Initialize the centroid of each cluster sample and assign the samples in the training sample set to the nearest centroid. After each assignment, iteratively update the centroid of each cluster sample until the centroids of the K cluster samples converge.

[0046] When the centroid converges, the customer set corresponding to each cluster sample set is determined as the customer classification result.

[0047] Furthermore, following any one or a combination of the aforementioned technical solutions, classifying customers based on the individual treatment effect value, the contribution value, and the customer characteristics includes the following steps:

[0048] The contribution value of customer characteristics to the individual treatment effect of customers is calculated based on interpretability. The customer characteristics are ranked from largest to smallest according to the absolute value of the contribution value. The customers are then classified according to the customer characteristic ranking to obtain the customer classification result.

[0049] And / or,

[0050] The historical data is updated periodically, and the customer classification model is retrained using the updated historical data. The retrained customer classification model is then used to update the existing customer classification model.

[0051] According to another aspect of the present invention, a business processing method based on causal inference and interpretability is provided, which obtains customer classification results based on the customer classification method based on causal inference and interpretability as described in any one or a combination of the above technical solutions;

[0052] There are multiple services, and for each service, customers can choose to apply it or not. Based on the customer classification results, the corresponding services are assigned to the customers.

[0053] Furthermore, following any one or a combination of the aforementioned technical solutions, the method further includes the following steps:

[0054] Based on the services assigned to customers, calculate the resources required to apply the service processing method to the k-th customer category using the following formula:

[0055]

[0056] Where resource k represents, and K represents the total number of customer categories. denoted as the average individual treatment effect value of the k-th customer group, where customer size k represents the number of customers in the k-th customer group;

[0057] The corresponding business involves redistributing resources to customers with the aim of reducing the aforementioned resources.

[0058] According to another aspect of the present invention, the present invention provides a customer classification system based on causal inference and interpretability, including a processor configured to determine a customer classification result according to the customer classification method based on causal inference and interpretability as described in any one or a combination of the above technical solutions.

[0059] The beneficial effects of the technical solution provided by this invention are as follows:

[0060] a. This invention enables accurate decision-making based on causal drive: Existing solutions typically rely on statistical models or machine learning algorithms to formulate strategies solely based on the correlation between customer characteristics and conversion results. This fails to distinguish between "correlation" and "causation," potentially leading to strategies deviating from true effectiveness, such as spurious associations caused by confounding variables. This invention moves from correlation analysis to causal effect modeling. Through propensity score matching (PSM) and causal forest methods, it constructs a causal model to eliminate covariate distribution differences and accurately estimate individual treatment effects (ITE). This improvement allows strategy design to be based on the true causal impact of marketing activities on customer conversion, rather than superficial correlations. Compared to traditional methods that rely solely on feature correlation, this technical solution can accurately identify true causal effects, avoid resource waste, and significantly improve the scientific rigor and effectiveness of strategy design.

[0061] b. This invention achieves a transition from a black-box model to an interpretable closed loop: While existing models (such as random forests and deep learning) have strong predictive capabilities, they lack a clear explanation of feature contributions, making the strategy logic difficult for business personnel to understand. This solution quantifies the marginal contribution of each feature to ITE through contribution values ​​(SHAP values), revealing key driving factors; at the same time, combined with the traceability of clustering labels, it directly links customer classification results with original customer characteristics. This improvement makes the model output transparent, allowing business personnel to clearly understand the core reasons for customer responses, ensuring that technological achievements are consistent with business needs, and improving the credibility and implementability of the strategy;

[0062] c. This invention realizes the transformation from static stratification to dynamic resource optimization: Existing customer stratification solutions rely on static rules or a single dimension, which is difficult to adapt to market changes. This solution integrates the ITE and SHAP values ​​output by causal forest and the original features to construct a multi-dimensional clustering matrix, uses algorithms such as K-Means to identify differentiated customer groups, and dynamically allocates the budget by combining the average ITE of the cluster and the number of customers to ensure that resources are tilted towards high ROI groups.

[0063] d. This invention forms a continuous iteration mechanism by regularly updating model parameters and verifying the effectiveness of the strategy. This improvement enables customer segmentation and resource allocation to be dynamically adaptable, enhancing the agility and scientific nature of marketing for small and medium-sized banks. Attached Figure Description

[0064] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0065] Figure 1A flowchart of a customer classification method based on causal inference and interpretability provided as an exemplary embodiment of the present invention;

[0066] Figure 2 A block diagram illustrating the module principle of a customer classification model provided as an exemplary embodiment of the present invention. Detailed Implementation

[0067] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0068] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, apparatus, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0069] To address the shortcomings of existing technologies, this invention proposes a customer classification method and a business processing method that combine causal modeling and interpretability tools, aiming to achieve transparency in model decisions for customer classification and business processing and ensure the effectiveness of strategies.

[0070] In one embodiment of the present invention, a customer classification method based on causal inference and interpretability is provided, see [link to relevant documentation]. Figure 1 and Figure 2 The customer classification method includes the following steps:

[0071] Collect customer information to be classified and input it into a pre-trained customer classification model to output customer classification results. The customer information includes customer features. The customer classification model is trained in the following way:

[0072] Collect historical data and obtain a training dataset based on the historical data. The historical data includes customer characteristics, business processing variables, and outcome variables corresponding to customers. The business processing variables include those that have been recommended and those that have not been recommended. The outcome variables include those that have been successfully converted and those that have failed to convert.

[0073] The training dataset is input into the base model based on the causal forest model and interpretability to train the base model. The training steps include:

[0074] Based on the causal forest model, the individual treatment effect value of the customer is determined according to the customer characteristics, business processing variables, and outcome variables.

[0075] The contribution value of customer characteristics to the individual treatment effect of customers is calculated based on interpretability, and customers are classified according to the individual treatment effect value, the contribution value, and the customer characteristics;

[0076] Repeat the above training steps until the model converges to obtain the customer classification model.

[0077] In one embodiment of the present invention, taking the implementation of customer segmentation and an explainable marketing method as an example, customer data is first collected and processed to obtain the historical data. It should be noted that this embodiment describes whether or not a marketing method is recommended to customers. In other embodiments, customers can be users or other objects, and the business may not be a marketing method, but rather whether or not a corresponding technical solution is applied to users.

[0078] The historical data consists of multiple records, each including customer characteristics, business processing variables, and outcome variables corresponding to the customer. The customer characteristics include, but are not limited to, basic information such as the customer's age, deposit balance, and monthly transaction frequency. Accordingly, in this application, the customer information includes, but is not limited to, the customer characteristics. The business processing variables include whether the customer has been recommended or not, and the outcome variables include successful conversion and failed conversion. The key to this step is ensuring the accuracy and completeness of the historical data, as all subsequent analyses will be based on it. Furthermore, preprocessing of this historical data is required, such as missing value imputation and outlier handling, to ensure the effectiveness of the analysis. For each historical data record, it is necessary to clearly distinguish whether it belongs to the treatment group or the control group, which is crucial for subsequent propensity score matching. In this application, a training dataset is obtained based on the historical data; therefore, the training dataset contains multiple samples, each including customer characteristics, processing variables, and outcome variables, used for training model parameters. For new predictions, i.e., when using a customer classification model to determine the customer classification result, the trained model parameters are used to obtain the customer classification result based on customer information. The customer information includes, but is not limited to, customer features and processing variables to obtain the result variable, or the customer information includes, but is not limited to, customer features to obtain the processing variable and the result variable.

[0079] In this embodiment, each piece of historical data is represented as follows:

[0080] Customer characteristics (covariates): X = {X1, X2, ..., X} p},X1,X2,…,X p These represent parameters such as age, bank balance, and monthly transaction frequency.

[0081] Processing variables: T∈{0,1}, T=1 indicates that a marketing campaign has been promoted (business processing methods have been applied / assigned to customers), and T=0 indicates that a marketing campaign has not been promoted (business processing methods have not been applied / assigned to customers);

[0082] Outcome variable: Y∈{0,1}, where Y=1 indicates successful conversion and Y=0 indicates failure.

[0083] In one embodiment of the present invention, historical data is processed by propensity score matching to obtain the training dataset. This step mainly consists of three sub-steps: calculating propensity scores, matching the processing group and the control group, and performing a balance test.

[0084] First, the historical dataset is divided into a processing group and a control group. In the processing group, the business processing variables corresponding to the historical data are those that have been recommended, while in the control group, the business processing variables are those that have not been recommended. In one specific embodiment, the historical dataset is divided into processing and control groups as shown in Table 1. In the processing group, all business processing variables are T=1, and in the control group, all business processing variables are T=0. The variables related to customer characteristics and conversion results are the corresponding collected actual data.

[0085] Table 1. Comparison table of historical datasets divided into treatment and control groups.

[0086]

[0087] While causal inference can capture causal effects, its accuracy depends on the balance of covariates in the training data. Historically, significant differences between customers who participated in marketing campaigns and those who did not lead to imbalanced training data. For example, typically, customers in the treatment group (who participated in marketing campaigns) have an average asset of 100,000 and are 30 years old, while customers in the control group (who did not participate in marketing campaigns) have an average asset of 10,000 and are 40 years old. Directly comparing the conversion results of the two groups might lead to the erroneous conclusion that the campaign was ineffective, because the treatment group is inherently younger and has higher assets (higher customer loyalty). This difference is selection bias, which can cause bias in model training. To reduce causal inference errors caused by customer self-selection bias, propensity score matching (PSM) is used to construct balanced samples to eliminate selection bias.

[0088] Propensity score matching (PSM) eliminates selection bias caused by differences in covariates in observational data by balancing the distribution of covariates between the treatment and control groups, providing more reliable input data for causal inference. PSM enforces similarity in covariates between the two groups by matching propensity scores (the probability of an individual receiving a treatment), thereby reducing spurious associations caused by systematic differences in covariates. This allows causal inference to focus more accurately on the impact of the treatment (whether or not to generalize) itself, rather than being interfered with by the imbalance of covariates, ultimately improving the reliability of causal effect estimation.

[0089] Based on the customer characteristics and the business processing variables, predict each customer's preference score for the business. In a specific application example, the probability of a customer becoming part of the processing group (having participated in a marketing campaign) can be estimated using a logistic regression model. The formula for calculating each customer's preference score for the business using a logistic regression model is as follows:

[0090]

[0091] Among them, β1, β2...β pX1, X2, ..., Xp are the regression coefficients of logistic regression, X1, X2, ..., Xp are customer feature 1, customer feature 2, ..., customer feature p, and β0 is the intercept term.

[0092] The treatment group and the control group are matched based on the propensity score to obtain a matched dataset. The control group and the control group are matched as follows: for customer i in the treatment group, customer j in the control group is determined, where customer i and customer j satisfy: Distance(i,j)=|P i -P j |Minimum, where P i P represents the propensity score corresponding to customer i in the treatment group. j This represents the propensity score corresponding to customer j in the control group. This step aims to find customers in the control group with similar propensity scores for each customer in the treatment group; after matching, the customer characteristics of the two groups will become similar.

[0093] The matched dataset undergoes a balance test. If the test fails, customer groups with matching scores exceeding a preset value (too high) are removed, and the balance test is repeated. If the test passes, the matched dataset is configured as a balanced dataset and used as the training dataset. Specifically, the balance test of the matched dataset can be performed as follows: First, the standardized mean difference of the matched dataset is calculated using the following formula:

[0094]

[0095] Among them, SMD k This represents the standardized mean difference. This indicates that the control group is based on customer characteristic X. k The mean of the above, This indicates that the control group is based on customer characteristic X. k The mean of the above, This represents the pooled variance of customer characteristics.

[0096] If|SMD k |<α, where α is a preset threshold, 0<α<0.15, then the balance test of the matched dataset passes; otherwise, it fails. For example, |SMD k If | < 0.1, then the match is successful.

[0097] Traditional methods typically focus only on basic customer information, such as age and bank balance, during the data collection phase, neglecting the specific impact of marketing activities and customer responses. This leads to insufficient accuracy and completeness of the data, limiting the understanding of customers' true needs and behavioral patterns. This application, based on the aforementioned method, utilizes propensity score matching to first obtain balanced data from historical data as the training sample set. This eliminates covariate selection bias and improves the accuracy of subsequent model training.

[0098] After obtaining a matched and validated balanced dataset, the process proceeds to the construction and training phase of the causal forest model to estimate the Individual Treatment Effect (ITE). During training, the objective of the model is to minimize the loss function. The causal forest model on the training dataset is used to estimate the potential difference in effect between promoting a marketing campaign to each customer and not promoting it. This difference in effect is the ITE, reflecting the specific impact of the marketing campaign on an individual customer. Existing techniques typically rely on statistical methods, correlation indicators (such as spending amount), or machine learning algorithms (such as random forests) to select features, but they do not consider causal relationships, potentially leading to feature selection biased towards correlation rather than true driving factors, and a lack of identification of causal effects. This application, through accurate estimation of ITE, can better understand the different response patterns of different customer groups to marketing campaigns, thus providing a basis for personalized marketing strategy development. Simultaneously, this also provides fundamental data support for further customer segmentation.

[0099] First, based on the causal forest model, the individual treatment effect value (ITE) for each customer i is estimated using the following formula:

[0100]

[0101] in, This represents the ITE value of customer i, where, This represents the expected value estimate when an individual receives treatment. This represents the expected value estimate when the individual did not receive treatment.

[0102] Construct a causal forest model on the matched training dataset The goal of training the causal forest model is to minimize the following loss function:

[0103]

[0104] Where θ represents the parameters of the causal forest model, including the number of trees and the number of split nodes. This is the individual treatment effect value for customer i. It is the baseline expected value before treatment, Y i This indicates whether customer i has converted.

[0105] To enhance model interpretability, the contribution of each customer feature to ITE (Information Tolerance) needs to be quantified using SHAP (Shapley Additive exPlanations) values. Based on cooperative game theory, SHAP values ​​reveal key driving factors by calculating the marginal contribution of each feature to the prediction results. When a feature has a significantly positive SHAP value for customers with high ITE, it indicates that this feature is a core factor in customer responsive marketing. The distribution of SHAP values ​​can identify the non-linear relationship between features and ITE. This step not only verifies the model logic but also provides an interpretable basis for feature selection for the hierarchical strategy, ensuring consistency between marketing strategies and business needs.

[0106] First, the contribution value, or SHAP value, is calculated. Based on interpretability, the contribution value of customer characteristics to the individual treatment effect of the customer is calculated using the following formula:

[0107]

[0108] Where, φ i,j Let S be a subset of customer features, |S| represent the number of customer features in subset S, ! denotes factorial operation, and p represent the total number of permutations of all customer features. This means that the individual treatment effect value of customer i is predicted using only a subset of customer features S. Let X represent the individual treatment effect value using the union of a subset S of customer characteristics and customer characteristic j, and let X represent the universal set of all customer characteristics. -j This represents the subset of all customer features excluding customer feature j.

[0109] Then, key customer characteristics are identified as key features, and the customer characteristics that have the greatest impact on ITE are filtered by the absolute value of the SHAP value.

[0110] After identifying the customer characteristics that have the greatest impact on ITE, one way to determine the customer classification result is as follows: Based on interpretability, calculate the contribution value of each customer characteristic to the individual treatment effect of the customer. Rank the customer characteristics according to their absolute contribution values ​​from largest to smallest, and classify the customers based on this ranking. Specifically, for a business activity such as marketing campaign M, calculate the contribution value of each customer characteristic to the individual treatment effect of that business. Rank the contribution values ​​according to their absolute values ​​from largest to smallest to obtain the contribution value ranking result. Based on the correspondence between contribution values ​​and customer characteristics, map the contribution value ranking result to the customer characteristic ranking. Classify the customers based on this customer characteristic ranking to obtain the customer classification result. Specifically, customers who meet the first ranked customer characteristic can be identified as customers of the same category, and marketing campaign M can be promoted to all customers of this category. Alternatively, several customer characteristics that meet the first ranked characteristics can be used as a key feature set. Customers who meet all or some of the customer characteristics in this key feature set can be identified as customers of the same category, and marketing campaign M can be promoted to all customers of this category.

[0111] After identifying the customer features that have the greatest impact on ITE, another way to determine the customer classification results is to integrate the ITE and SHAP values ​​output by the causal forest with the original customer features to construct a clustering feature matrix. K-Means or hierarchical clustering algorithms are then used, combined with the Elbow Method, to determine the optimal number of clusters. Each customer category is labeled according to its customer characteristics and response patterns, forming hierarchical customers, thus obtaining the customer classification results and providing a basis for differentiation strategies.

[0112] First, the causal results are combined with the original features to construct a clustering feature matrix.

[0113] Constructing the clustering feature matrix Z i , Among them, X i This represents the customer characteristics of customer i. This represents the individual treatment effect value for customer i, and Top SHAP Features represent the contribution value.

[0114] Then, determine the optimal number of clusters K. Specifically, K-Means or hierarchical clustering algorithms can be used, combined with the elbow method, to determine the optimal number of clusters K of the clustering feature matrix to obtain a K-cluster sample set, that is, to divide the sample training set into K sample sets.

[0115]

[0116] Where WCSS() represents the sum of squares within the cluster, μ k Let C be the centroid of the k-th cluster. kLet represent the set of customers in the k-th cluster.

[0117] Next, perform clustering operations as follows to divide the current customers into K different customer groups, which facilitates differentiated marketing to different customers.

[0118] Initialize the centroids of K clusters. The centroids of the K clusters are the centers of the K sample classes. The K centroids refer to the centers of the K sample classes during the calculation.

[0119] The samples in the training sample set are assigned to the nearest centroid using the following formula:

[0120]

[0121] Where, μ j Let be the centroid of the j-th cluster sample.

[0122] After each assignment, the centroid is iteratively updated until it converges. When the centroid converges, the customer set corresponding to each cluster sample set is the customer classification result.

[0123] Through the above steps, this technical solution achieves end-to-end interpretability support: combining the traceability of SHAP values ​​and clustering labels, this invention achieves full-process transparency from feature contribution to customer segmentation. SHAP values ​​reveal key driving factors, and clustering labels associate with original features, ensuring that the strategy logic aligns with business intuition. This interpretability loop enhances model credibility, facilitating rapid understanding and implementation of strategies by business personnel.

[0124] In one embodiment of the invention, historical data is periodically updated, and the customer classification model is retrained using the updated historical data. The retrained customer classification model is then used to update the previously established customer classification model. This iterative updating of the customer classification model continuously improves its predictive accuracy.

[0125] In one embodiment of the present invention, a business processing method based on causal inference and interpretability is provided, comprising the following steps:

[0126] The customer classification result is obtained based on the causal inference and interpretability-based customer classification method described in any one or a combination of the above embodiments;

[0127] The services are diverse, and for each service, the service is assigned to the customer based on the customer classification results.

[0128] In this embodiment, the following step is also included: calculating the resources required to apply the service to the k-th type of customer based on the service assigned to the customer, using the following formula:

[0129]

[0130] Where resource k represents, and K represents the total number of customer categories. denoted as the average individual treatment effect value of the k-th customer group, where customer size k represents the number of customers in the k-th customer group;

[0131] The corresponding business is redistributed to customers with the aim of reducing the resources mentioned above, so as to allocate the corresponding business to customers more rationally.

[0132] For marketing campaigns, the resources mentioned can be represented as marketing cost / marketing budget, calculated using the following formula:

[0133]

[0134] Where budget k represents the number of clusters, and K represents the number of clusters. denoted as the average individual treatment effect value of the k-th cluster, and cluster size k represents the number of customers in the k-th cluster.

[0135] This step involves designing specific marketing strategies based on the previously obtained customer segmentation / classification results. Existing techniques typically develop uniform or roughly differentiated strategies based on segmentation results, but these strategies lack interpretable analysis of key characteristics (such as age and deposit habits). This application, however, develops differentiated marketing plans for different types of customer groups. For example, it provides customized services for high-value, high-response customers, recommends value-added products to potential customers, and adopts simplified services or reactivation strategies for low-value, low-response customers.

[0136] In addition, it is necessary to consider the rational allocation of resources to ensure that budget and channel selection maximize return on investment (ROI). The entire process should be a dynamic adjustment process, as shown in Table 2, regularly reassessing customer segmentation and continuously optimizing strategies based on market feedback to adapt to rapidly changing market demands.

[0137] Table 2. Differentiated Marketing Strategies Based on Cluster Characteristics

[0138] Cluster type Strategy High-value, high-response clusters Provide high-value services (such as customized financial planning), prioritize access, and maximize resource investment. Potential customer cluster Recommend value-added products (such as installment payment options) and increase user activity through promotions. Low-value, low-response clusters Streamline services (such as basic deposits), or reactivate them through promotional activities.

[0139] Dynamic monitoring and iteration are implemented, with re-clustering and model parameter updates every quarter. A / B testing is used to validate the strategy's effectiveness, ensuring that actual conversion rates match predicted values. Consistent.

[0140] Compared to existing technologies that typically evaluate effectiveness solely based on superficial metrics like conversion rates and click-through rates, failing to differentiate the true impact of marketing activities from confounding factors through causal inference, this embodiment provides a business processing method capable of dynamic stratification and resource optimization: based on multidimensional clustering stratification using ITE, SHAP values, and original features, this invention accurately segments high-value, high-potential, and low-response customer groups. Through dynamic budget allocation rules, resources are tilted towards high ROI groups, and the introduction of A / B testing and iteration mechanisms ensures that the strategy continuously adapts to market changes, maximizing marketing efficiency.

[0141] The intelligent customer operation methods and systems based on causal modeling and machine learning interpretability technologies mainly belong to the technical fields of causal inference and model interpretability. Causal modeling, by quantifying the causal relationships between variables, can overcome the limitations of traditional correlation analysis, extract decision-making basis with business logic from mixed data, and provide banks with more accurate customer behavior prediction and strategy optimization capabilities.

[0142] This technology is particularly applicable to the banking industry: on the one hand, causal modeling can reveal the deep connection between marketing interventions and customer conversion, helping banks identify high-value customer groups and evaluate the actual effectiveness of strategies; on the other hand, machine learning interpretability technologies (such as feature importance analysis and model transparency tools) provide an intuitive explanatory framework for the decision-making process of complex models, enabling banks to trace the key drivers of customer segmentation, risk assessment, or product recommendation, thereby enhancing model credibility and regulatory compliance. This technology is especially important in the context of small and medium-sized banks—which often face challenges such as inconsistent data quality and limited resources. Causal modeling and interpretability technologies can improve operational efficiency and reduce trial-and-error costs through dynamic optimization strategies (such as clustering and hierarchical management, and differentiated resource allocation).

[0143] Furthermore, this technology portfolio is not only applicable to areas such as customer lifecycle management and precision marketing, but can also be extended to risk management, product design, and other aspects, driving the banking industry to transform from experience-driven to data-driven, and ultimately achieving intelligent, compliant, and sustainable development.

[0144] In one embodiment of the present invention, a customer classification system based on causal inference and interpretability is provided, including a processor configured to determine a customer classification result according to the customer classification method based on causal inference and interpretability as described in any one or a combination of the embodiments above.

[0145] It should be noted that the business processing method and customer classification system based on causal inference and interpretability provided by the present invention have the same inventive concept as the above-mentioned customer classification method based on causal inference and interpretability embodiments. The entire contents of the customer classification method based on causal inference and interpretability embodiments are incorporated into the business processing method and customer classification system based on causal inference and interpretability embodiments by means of introduction.

[0146] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0147] The above description is only a specific embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A customer classification method based on causal inference and interpretability, characterized in that, Includes the following steps: Collect customer information to be classified and input it into a pre-trained customer classification model to output customer classification results. The customer information includes customer features. The customer classification model is trained in the following way: Collect historical data and obtain a training dataset based on the historical data. The historical data includes customer characteristics, business processing variables, and outcome variables corresponding to customers. The business processing variables include those that have been recommended and those that have not been recommended. The outcome variables include those that have been successfully converted and those that have failed to convert. The training dataset is input into the base model based on the causal forest model and interpretability to train the base model. The training steps include: Based on the causal forest model, the individual treatment effect value of the customer is determined according to the customer characteristics, business processing variables, and outcome variables. The contribution value of customer characteristics to the individual treatment effect of customers is calculated based on interpretability, and customers are classified according to the individual treatment effect value, the contribution value, and the customer characteristics; Repeat the above training steps until the model converges to obtain the customer classification model.

2. The customer classification method based on causal inference and interpretability according to claim 1, characterized in that, The training dataset is obtained based on historical data, including the following steps: The historical dataset is divided into a processing group and a control group. The business processing variables corresponding to the historical data in the processing group have been recommended, while the business processing variables corresponding to the historical data in the control group have not been recommended. Predict each customer's preference score for the service based on the customer characteristics and the business processing variables; The treatment group and the control group are matched based on the propensity scores to obtain a matched dataset; Perform a balance check on the matched dataset. If the check passes, configure the matched dataset as a balanced dataset to be used as the training dataset.

3. The customer classification method based on causal inference and interpretability according to claim 2, characterized in that, The logistic regression model is used to estimate the customer's preference score for the service. The calculation formula is as follows: Among them, β1, β2...β p X1, X2, ..., Xp are the regression coefficients of logistic regression, X1, X2, ..., Xp are customer feature 1, customer feature 2, ..., customer feature p, and β0 is the intercept term.

4. The customer classification method based on causal inference and interpretability according to claim 3, characterized in that, The control group and the control group were matched in the following way: For customer i in the treatment group, determine customer j in the control group. Customers i and j satisfy: Distance(i,j)=|P i -P j |Minimum, where P i P represents the propensity score corresponding to customer i in the treatment group. j This represents the propensity score corresponding to customer j in the control group.

5. The customer classification method based on causal inference and interpretability according to claim 2, characterized in that, The balance of the matched dataset is checked using the following method: Calculate the standardized mean difference of the matched dataset using the following formula: Among them, SMD k This represents the standardized mean difference. This indicates that the control group is based on customer characteristic X. k The mean of the above, This indicates that the control group is based on customer characteristic X. k The mean of the above, The pooled variance representing customer characteristics; If|SMD k |<α, where α is a preset threshold. If 0<α<0.15, the balance test of the matched dataset passes; otherwise, it fails.

6. The customer classification method based on causal inference and interpretability according to claim 1, characterized in that, Based on the causal forest model, the individual treatment effect value (ITE) for each customer i is estimated using the following formula: in, This represents the ITE value of customer i, where, This represents the expected value estimate when an individual receives treatment. This represents the expected value estimate when the individual did not receive treatment.

7. The customer classification method based on causal inference and interpretability according to claim 1, characterized in that, Based on the causal forest model, the base model is trained with the objective of minimizing the following loss function: Where θ represents the parameters of the causal forest model, including the number of trees and the number of split nodes. This is the individual treatment effect value for customer i. It is the baseline expected value before treatment, Y i This indicates whether customer i has converted.

8. The customer classification method based on causal inference and interpretability according to claim 1, characterized in that, The contribution value of customer characteristics to the individual treatment effect of customers is calculated based on interpretability. The formula for calculating the contribution value is as follows: Where, φ i,j Let S represent the contribution value, where S is a subset of customer features, |S| represents the number of features in the subset S, ! represents the factorial operation, and p represents the total number of permutations of all customer features. This means that the individual treatment effect value of customer i is predicted using only a subset of features S. Let X represent the individual treatment effect value using the union of a subset S of customer characteristics and customer characteristic j, and let X represent the universal set of all customer characteristics. -j This represents the subset of all customer features excluding customer feature j.

9. The customer classification method based on causal inference and interpretability according to claim 1, characterized in that, Classifying customers based on the individual treatment effect value, the contribution value, and the customer characteristics includes the following steps: Constructing the clustering feature matrix Z i , Among them, X i This represents the customer characteristics of customer i. This represents the individual treatment effect value for customer i, and Top SHAP Features represent the contribution value. The optimal number of clusters K for the clustering feature matrix is ​​determined by using K-Means or hierarchical clustering algorithms and combining them with the elbow method to divide the training sample set into K clusters. Initialize the centroid of each cluster sample and assign the samples in the training sample set to the nearest centroid. After each assignment, iteratively update the centroid of each cluster sample until the centroids of the K cluster samples converge. When the centroid converges, the customer set corresponding to each cluster sample set is determined as the customer classification result.

10. The customer classification method based on causal inference and interpretability according to claim 1, characterized in that, Classifying customers based on the individual treatment effect value, the contribution value, and the customer characteristics includes the following steps: The contribution value of customer characteristics to the individual treatment effect of customers is calculated based on interpretability. The customer characteristics are ranked from largest to smallest according to the absolute value of the contribution value. The customers are then classified according to the customer characteristic ranking to obtain the customer classification result. And / or, The historical data is updated periodically, and the customer classification model is retrained using the updated historical data. The retrained customer classification model is then used to update the existing customer classification model.

11. A business processing method based on causal inference and interpretability, characterized in that, Includes the following steps: The customer classification results are obtained based on the customer classification method based on causal inference and interpretability as described in claim 1; There are multiple services, and for each service, customers can choose to apply it or not. Based on the customer classification results, the corresponding services are assigned to the customers.

12. The business processing method based on causal inference and interpretability according to claim 11, characterized in that, It also includes the following steps: Based on the services assigned to customers, calculate the resources required to apply the service processing method to the k-th customer category using the following formula: Where resource k represents, and K represents the total number of customer categories. denoted as the average individual treatment effect value of the k-th customer group, where customer size k represents the number of customers in the k-th customer group; The corresponding business involves redistributing resources to customers with the aim of reducing the aforementioned resources.

13. A customer classification system based on causal inference and interpretability, characterized in that, The system includes a processor configured to determine a customer classification result according to the customer classification method based on causal inference and interpretability as described in claim 1.

Images