A method, apparatus and electronic device for generating business operation strategies

Abstract

This application provides a method, apparatus, and electronic device for generating business operation strategies. The method preprocesses source data, including historical records, business constraints, current business environment data, and user attribute data of the object to be analyzed, to obtain a basic feature vector resistant to distribution shift. Then, it uses high-order causal feature extraction, obfuscation path identification, and structured equation modeling to block the influence of obfuscating factors on causal relationships, obtaining pure causal relationship data. This effectively solves the decision bias caused by obfuscated variables in traditional models. Furthermore, based on interpretable identification features, the generated business operation strategy is logically traceable and conforms to business constraints. Even when using traditional models to generate business operation strategies, because the causal relationships in the input data have already been purified, it can effectively cope with environmental fluctuations, resulting in more stable output results and better robustness.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method, apparatus and electronic device for generating business operation strategies. Background Technology

[0002] With the development of big data technology, enterprises and institutions often use big data to generate corresponding strategies (or decisions) for business operations to assist in production and management. For example, e-commerce platforms generate holiday promotion strategies based on large amounts of e-commerce data, internet platforms generate advertising strategies based on traffic data, medical institutions generate medical resource allocation strategies based on existing medical resource data, and logistics companies generate vehicle scheduling strategies based on current transportation capacity data, and so on.

[0003] Furthermore, with the application of artificial intelligence (AI) technology, enterprises and organizations often use AI models to generate corresponding intelligent business operation strategies based on big data. However, because the data used to generate business operation strategies is not static but dynamically changing in real time, specifically, the data distribution in a dynamic environment changes with time and scenario. This makes the environment corresponding to the training phase of the intelligent model inconsistent with the environment corresponding to the application phase of the intelligent model. Consequently, the trained intelligent model cannot stably output reliable intelligent business operation strategies under environmental fluctuations (such as disturbances and sudden changes), resulting in poor model robustness. Summary of the Invention

[0004] In view of this, embodiments of this application provide a business operation strategy generation method, apparatus, and electronic device to solve the problem of poor robustness of existing intelligent models.

[0005] In a first aspect, embodiments of this application provide a method for generating a business operation strategy, wherein the method includes: The source data of the object to be analyzed are obtained and the source data are preprocessed to generate a basic feature vector that is resistant to distribution shift. The source data includes: the historical records, business constraints, current business environment data and user attribute data of the object to be analyzed. Higher-order causal feature extraction is performed on the basic feature vector resisting distribution shift to obtain the interpretability analysis and identification features of the object to be analyzed and the obfuscated paths between business operation strategies. Then, a pre-set structured equation model is called to extract the pure causal relationship data from the source data of the object to be analyzed based on the obfuscated paths between business operation strategies. Under the constraints of the business conditions, based on the pure causal relationship data and the interpretability analysis identification features, a business operation strategy for the object to be analyzed is generated and output, wherein the business operation strategy includes the target object of the optimal solution.

[0006] Secondly, embodiments of this application provide a business operation strategy generation apparatus, wherein the apparatus includes: The preprocessing module is used to acquire various source data of the object to be analyzed and perform data preprocessing on each source data to generate a basic feature vector that is resistant to distribution shift. The source data includes: the historical records, business constraints, current business environment data, and user attribute data of the object to be analyzed. The causal inference module is used to extract high-order causal features from the basic feature vector resisting distribution shift, obtain the interpretability analysis and identification features of the object to be analyzed and the obfuscated paths between business operation strategies, and call a preset structured equation model to extract pure causal relationship data from the source data of the object to be analyzed based on the obfuscated paths between business operation strategies. The strategy generation module is used to generate and output the business operation strategy of the object to be analyzed based on the pure causal relationship data and the interpretability analysis identification features under the constraints of the business constraints, wherein the business operation strategy includes the target object of the optimal solution.

[0007] Thirdly, embodiments of this application provide an electronic device, wherein the electronic device includes: a processor; and a memory storing a program; wherein the program includes instructions, which, when executed by the processor, cause the processor to perform the business operation strategy generation method described in the first aspect.

[0008] Fourthly, embodiments of this application provide a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause a computer to execute the business operation strategy generation method described in the first aspect.

[0009] The beneficial effects of this application are: This application provides a method, apparatus, and electronic device for generating business operation strategies. The method preprocesses source data, including historical records, business constraints, current business environment data, and user attribute data of the object to be analyzed, to obtain a basic feature vector resistant to distribution shifts. This effectively resists data distribution shifts caused by dynamic changes in the business environment, thus preventing systemic model failure due to environmental changes. Simultaneously, by extracting high-order causal features, identifying obfuscated paths, and using structured equation models to block the influence of obfuscating factors on causal relationships, pure causal relationship data is obtained. This effectively solves the decision bias caused by obfuscated variables in traditional models. Furthermore, based on interpretable identification features, the generated business operation strategy is logically traceable and conforms to business constraints. Using the method provided in this application, even when using traditional models to generate business operation strategies, the causal relationships in the model input data have been purified, effectively responding to environmental fluctuations, resulting in more stable output results and better robustness. Attached Figure Description

[0010] Further details, features, and advantages of this application are disclosed in the following description of exemplary embodiments in conjunction with the accompanying drawings, in which: Figure 1 This application provides a flowchart illustrating a method for generating business operation strategies. Figure 2 This illustration shows an application scenario diagram of the business operation strategy generation method provided in this application; Figure 3 This invention provides a schematic diagram of the structure of a business operation strategy generation apparatus. Figure 4 A structural block diagram of an exemplary electronic device that can be used to implement embodiments of this application is shown. Detailed Implementation

[0011] Embodiments of this application will now be described in more detail with reference to the accompanying drawings. While some embodiments of this application are shown in the drawings, it should be understood that this application can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this application. It should be understood that the drawings and embodiments of this application are for illustrative purposes only and are not intended to limit the scope of protection of this application.

[0012] It should be understood that the steps described in the method embodiments of this application may be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of this application is not limited in this respect.

[0013] The term "comprising" and its variations as used herein are open-ended, meaning "including but not limited to". The term "based on" means "at least partially based on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". Definitions of other terms will be given in the following description. It should be noted that the concepts of "first", "second", etc., mentioned in this application are used only to distinguish different devices, modules, or units, and are not intended to limit the order of functions performed by these devices, modules, or units or their interdependencies.

[0014] It should be noted that the terms "a" and "a plurality of" used in this application are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0015] As described in the background section, the data used to generate business operation strategies is not static but dynamically changing in real time. Specifically, the data distribution in a dynamic environment changes with time and scenario. This causes a discrepancy between the environment during the training phase and the environment during the application phase of the intelligent model. Consequently, the trained intelligent model cannot stably output reliable intelligent business operation strategies under environmental fluctuations (such as disturbances or sudden changes), resulting in poor model robustness. The following example from a credit business scenario illustrates the fundamental reason for the poor robustness of the aforementioned model: Credit institutions often use artificial intelligence models to generate credit strategies based on existing credit-related data, such as determining the price at which credit products are offered. In this case, the credit price, as a core parameter of the entire credit strategy, directly impacts the financial institution's risk and return, as well as its resource allocation. Current mainstream methods rely on statistical models and machine learning algorithms to assess risk levels and set corresponding credit prices by analyzing data such as the historical credit history, asset and liability structure, income, and behavioral characteristics of credit product users. A core premise of these models is that the model training environment must be consistent with the application environment, i.e., the independent and identically distributed (ICD) assumption must be met.

[0016] In practice, common strategies and techniques for determining credit prices (hereinafter referred to as pricing) include: rule-based pricing based on risk scores, model-based pricing that optimizes profit / loss, and pricing that considers market competition. However, these techniques cannot flexibly and accurately generate corresponding prices when dealing with dynamically changing credit environments. Specifically, factors such as macroeconomic fluctuations, policy adjustments, changes in market competition patterns, and evolution of user behavior can easily lead to a significant deviation between the data distribution input to the model after model training and deployment and the data distribution input to the model during the training phase. This data distribution deviation will severely weaken the accuracy and stability of pricing models built on historical correlations. For example, a model trained on boom data during an economic downturn may underestimate risk, resulting in underpricing; new data regulations restricting the use of key features can directly harm model performance; and changes in competitors' strategies may invalidate the original price-response relationship, etc.

[0017] Technical approaches to address this data bias include feature engineering, online learning, and domain adaptation. These techniques often rely on unrealistic assumptions such as known patterns or strong heterogeneity of source domain data, and online learning itself can introduce instability and high computational costs. At a deeper level, the presence of numerous confounding variables in credit scenarios leads models to learn spurious associations. Traditional models struggle to distinguish between causal and correlational relationships between features and credit outcomes (such as default and response rates). For example, the correlation between "frequent use of a specific app" and a low default rate might stem from the fact that high-income, stable professionals (the confounding variable) simultaneously possess both a preference for that app and low-risk characteristics, rather than the app's use itself reducing risk. When environmental changes disrupt these spurious associations, the model's pricing decisions will exhibit significant biases.

[0018] Furthermore, credit price regulation is essentially an intervention, and price changes directly affect customers' willingness to borrow (demand elasticity) and repayment behavior (risk). Most existing technologies fail to effectively model the heterogeneous causal effects (HTE) of price intervention on different customer groups, making it difficult to accurately identify price-sensitive customers, accurately predict the net impact of price adjustments on overall business metrics (scale, risk, profit), and potentially raise questions about fairness. In addition, achieving both strong predictive capabilities and high interpretability to meet the needs of regulatory compliance and the embedding of business rules is also a challenge currently facing existing intelligent models.

[0019] In recent years, causal inference techniques have been introduced into intelligent models. These techniques utilize dual machine learning, propensity score matching estimation, and counterfactual reasoning by constructing causal graphs to identify confounding effects and improve model robustness. However, the application scenarios of these models are often complex business scenarios, and existing causal inference methods still have significant limitations in handling these complex applications. For example, relying on strong assumptions such as "no unmeasured confounding" is difficult to satisfy; accurate estimation of heterogeneous effects requires high-quality, large-scale data, even randomized experimental data, which is costly; model complexity leads to high computational and engineering implementation difficulties and weak scalability; and there is still a lack of mature solutions for seamlessly and reliably integrating causal effect estimation into pricing optimization processes that meet multiple business constraints.

[0020] In summary, current intelligent model generation business operation strategies rely excessively on statistical association modeling, are highly sensitive to spurious associations caused by data distribution shifts and confounding variables, and have difficulty assessing the causal effects on the target prediction object (such as price). Although causal inference provides a solution, its practical application is still limited by the strength of assumptions, data requirements, model complexity, and the difficulty of business integration.

[0021] In view of this, this application provides a business operation strategy generation method, apparatus, and electronic device to improve the accuracy of intelligent models in generating business operation strategies and enhance the robustness of intelligent models. Specifically, in a first aspect, this application provides a business operation strategy generation method, which can be applied to any electronic device with business operation strategy generation capabilities, including but not limited to personal mobile terminals, computers, or servers. Figure 1 As shown, the method includes the following steps: S11. Obtain the source data of the object to be analyzed and perform data preprocessing on each source data to generate a basic feature vector that resists distribution shift. The source data includes: the historical records, business constraints, current business environment data, and user attribute data of the object to be analyzed. S12. Perform high-order causal feature extraction on the basic feature vector resisting distribution shift to obtain the interpretability analysis and identification features of the object to be analyzed and the obfuscated paths between business operation strategies, and call the preset structured equation model to extract the pure causal relationship data in the source data of the object to be analyzed based on the obfuscated paths between business operation strategies. S13. Under the constraints of the business constraints, a pre-set intelligent strategy processing model is invoked to generate and output the business operation strategy of the object to be analyzed based on the pure causal relationship data and the interpretability analysis identification features. The business operation strategy includes the target object of the optimal solution.

[0022] By employing the embodiments of this application, data preprocessing is performed on source data such as historical records, business constraints, current business environment data, and user attribute data of the object to be analyzed to obtain a basic feature vector resistant to distribution shifts. This effectively resists data distribution shifts caused by dynamic changes in the business environment, thereby preventing systemic model failure due to environmental changes. Simultaneously, by extracting high-order causal features, identifying obfuscated paths, and using structured equation models to block the influence of obfuscating factors on causal relationships, pure causal relationship data is obtained. This effectively solves the decision bias caused by obfuscated variables in traditional models. Furthermore, based on interpretability identification features, the logic of the generated business operation strategy is traceable and conforms to business constraints. Using the method provided in this application, even when using traditional models to generate business operation strategies, the causal relationships in the model input data have been purified, effectively responding to environmental fluctuations, resulting in more stable output results and better robustness.

[0023] The following will provide a detailed explanation of steps S11 to S13 with specific examples: In some possible embodiments, the method provided in this application can be applied to a business operation strategy generation system, which can, as follows: Figure 2 As shown, it includes the following parts: data input layer, data preprocessing layer, causal inference network layer, classical prediction network layer, constraint policy optimization layer, and output layer.

[0024] The system comprises the following components: a data input layer for obtaining source data in step S11; a data preprocessing layer for preprocessing the obtained source data; a causal inference network layer for extracting high-order causal features from the anti-distribution-shift basic feature vector output by the data preprocessing layer in step S12; a classical prediction network layer corresponding to the intelligent strategy processing model in step S13, used to determine different objects to be analyzed and their corresponding prediction probability values ​​based on the pure causal relationships output by the causal inference network; a constraint strategy optimization layer for optimizing the prediction probability values ​​output by the classical prediction network layer based on business constraints in step S13, determining the optimal solution for the object to be analyzed, and outputting it to the output layer; and an output layer for generating the corresponding business operation strategy based on the optimal solution calculated by the constraint strategy optimization layer.

[0025] In this application, the object to be analyzed refers to the strategic object of concern in the business operation strategy, which depends on the application scenario. As one implementation method, if the method is applied to a credit business scenario, the object to be analyzed is the credit price; if the method is applied to a medical resource business scenario, the object to be analyzed is medical resources; if the method is applied to a logistics business scenario, the object to be analyzed is transportation cost. The following will provide a detailed explanation of the object to be analyzed being the credit price. For other scenarios, the object to be analyzed can be understood by substituting the credit price. Taking the credit price as an example, the business operation strategy of the object to be analyzed generated and output in step S13 of this application can specifically refer to calculating the optimal credit price through an intelligent model. This credit price can be a credit percentage, such as 3.5%.

[0026] In this application, the source data for the object to be analyzed can refer to various types of data that can influence or affect the fluctuations of the object, specifically including: the object's historical records, business constraints, current business environment data, and user attribute data. For example, taking credit pricing as the object to be analyzed, the historical records can be historical credit prices, specifically historical pricing records, which can be determined by obtaining historical test logs. Business constraints specifically refer to constraints set according to relevant laws, regulations, and industry restrictions. For example, the object to be analyzed, credit pricing, needs to meet the upper and lower limits of interest rates stipulated by the state, as well as the compliance range requirements, risk thresholds, etc., stipulated by the industry. In the scenario where the object to be analyzed is logistics and transportation costs, the business constraints can be the upper and lower limits of transportation costs stipulated by the industry.

[0027] Current business environment data refers to the objective environmental data of the business to which the object of analysis belongs. Taking credit pricing as an example, current business environment data includes macroeconomic indicators such as GDP growth rate, unemployment rate, and industry prosperity index. User attribute data can mainly be obtained through customer characteristics, which include objective personal attributes and behavioral characteristics. In the scenario where the object of analysis is credit pricing, objective personal attributes include credit score, income level, and occupation type. Behavioral characteristics refer to the behavioral data generated when customers use credit products.

[0028] When performing step S11, the corresponding source data can be obtained through various source data interfaces or storage paths, and then the obtained source data can be preprocessed. The types of data preprocessing include: wavelet transform, evidence weight encoding (known in the industry as WOE encoding), causal graph construction, and constraint standardization. In some possible embodiments, step S11 can be implemented through the following steps: Frequency domain transformation and multi-scale analysis are performed on each of the source data to remove periodic disturbances and short-term noise contained in the source data, and to generate the basic feature vector that resists distribution migration.

[0029] The source data includes discrete, continuous, time-series, and static data. Frequency domain transformation and multi-scale analysis are performed on each type of source data to convert them into standardized data with a unified format and sampling frequency, free of outliers. Specifically, time-series data with different sampling frequencies are unified into data with the same time dimension, and missing values ​​are filled using linear interpolation. For outliers, the 3σ principle is used to remove extreme outliers to avoid excessive perturbation. For static data, especially discrete static data, WOE encoding is used to convert it into numerical features.

[0030] Furthermore, time-frequency domain transformation is performed on the periodic source data, converting the periodic source data from the time domain to the frequency domain. Based on the frequency domain distribution, the corresponding essential trend term and periodic disturbance term are determined. The essential trend term refers to the part of the data that inherently has a trend, which depends on the attributes of the data itself. The periodic disturbance term refers to the part of the data that has a periodic cycle caused by short-term fluctuations, which depends on the short-term fluctuations of the data. Then, the periodic disturbance term is deleted, which can achieve the effect of removing the periodic disturbances and short-term noise contained in the source data.

[0031] As one implementation method, the step of performing frequency domain transformation and multi-scale analysis on each of the source data to remove periodic disturbances and short-term noise contained in the source data includes: The current business environment data is windowed using a combination of Fast Fourier Transform and Blackman window function to separate the periodic trends and sudden fluctuations of the business in the source data. Specifically, the windowing process for data refers to inputting the standardized periodic time-series data into a Blackman window function. The Blackman window function then performs calculations on the input periodic time-series data, which can suppress spectral leakage and thus prevent the spread of frequency components of the periodic signal, ensuring accurate identification of periodic disturbances.

[0032] Furthermore, a time-domain to frequency-domain transformation is performed on the windowed data, converting the time-domain signal (time-value) into a frequency-domain signal (frequency-amplitude). A higher amplitude in the frequency-domain signal indicates a stronger frequency disturbance. Then, based on an effective frequency threshold set by the business logic, frequencies in the frequency domain with amplitudes exceeding this threshold are identified. Strong periodic frequency components are filtered and removed, low-frequency components are retained, and high-frequency periodic components are removed, thereby separating the periodic trends and sudden fluctuations of the business in the source data.

[0033] Furthermore, the retained low-frequency components are subjected to inverse Fourier transform to restore them to time-domain signals.

[0034] The user behavior data in the user attribute data is decomposed into multiple scales to remove long-term essential features and short-term noise features from the user behavior data.

[0035] User behavior data within user attribute data is often accompanied by short-term random noise, such as sudden user spending or device malfunctions. Multi-scale decomposition can effectively separate long-term essential features from short-term noise features, thus avoiding the impact of short-term noise on feature stability. Specifically, an appropriate mother wavelet function can be selected based on the type of user behavior data; for example, the db4 wavelet can be chosen for abnormal user behavior, allowing the mother wavelet function to simultaneously capture both the temporal and frequency features of the signal.

[0036] Furthermore, an N-level wavelet decomposition is performed on the standardized dynamic behavior data, where the value of N depends on the data length. Each level is decomposed into low-frequency similarity coefficients and high-frequency detail coefficients. The low-frequency similarity coefficients characterize the overall trend of the signal, such as a user's long-term debt repayment ability in a credit scenario or stable product sales in an e-commerce scenario. The high-frequency detail coefficients characterize noise at different scales; for example, high-frequency detail coefficient D1 reflects high-frequency short-term noise, and D2 reflects sub-high-frequency fluctuations.

[0037] Furthermore, the high-frequency detail coefficients obtained from multi-scale wavelet decomposition are filtered based on a preset high-frequency detail coefficient threshold to obtain filtered high-frequency detail coefficients and low-frequency coefficients. Finally, the filtered high-frequency detail coefficients and low-frequency coefficients are subjected to inverse fast Fourier transform to obtain the corresponding filtered high-frequency detail coefficients and low-frequency coefficients in the time domain.

[0038] Finally, the low-frequency signal in the time domain, the filtered high-frequency detail coefficients, and the low-frequency coefficients are concatenated to obtain a feature vector, which is the basic feature vector for resisting distribution migration.

[0039] In one implementation, the data preprocessing layer is used to perform the following steps S11-1 to S11-4 to preprocess the source data: S11-1. Perform wavelet transform on the current business environment data to obtain the first sub-feature vector.

[0040] Business environment data (such as GDP growth rate in credit scenarios, disease progression rate in medical scenarios, industry demand data in supply chain scenarios, etc.) is inherently time-series data that fluctuates dynamically over time or across different scenarios, accompanied by periodic disturbances and noise. By performing wavelet transform on this business data—specifically, selecting a mother wavelet function suitable for the time-series data, such as db4 wavelet—multi-scale wavelet decomposition is performed on the current business environment data. This decomposes the data into low-frequency trend terms and high-frequency noise terms. The low-frequency trend terms reflect the essential patterns of the business environment, such as long-term economic growth trends, while the high-frequency noise reflects short-term random fluctuations, such as temporary policy disturbances. By retaining the low-frequency trend terms and filtering out invalid high-frequency noise terms, the processed low-frequency trend terms are quantized into a fixed-dimensional vector form as the first sub-feature vector. In this way, the multi-scale analytical capability of wavelet transform removes interfering factors from the business environment data, ensuring the stability of the data and preventing subsequent model distribution shifts due to short-term fluctuations or periodic disturbances. This lays the foundation for providing anti-shift properties for the basic feature vector.

[0041] S11-2. Perform evidence weight encoding on the user attribute data to obtain the second sub-feature vector.

[0042] User attribute data includes data describing static user attributes. The specific nature of this static attribute data depends on the business scenario. For example, in a credit scenario, user attribute data includes: occupation, education level, credit rating, age, etc. In a medical scenario, user attribute data includes: basic medical history, physical type, etc. It is evident that user static attribute data contains a large amount of data of different types and attributes. Different types of data cannot be directly input into the model. The core of evidence weight encoding for user attribute data is to transform multi-dimensional user static attribute data into a numerical vector with a unified dimension. Specifically, different discrete attributes can be grouped, and continuous attributes can be appropriately binned. Then, the evidence weight (WOE) for each group or bin is calculated: WOE = probability of the target event occurring in that group / probability of the target event not occurring in that group. Here, the target event depends on the application scenario. For example, in a credit scenario, the target event refers to credit default, in which case WOE = probability of credit default occurring / probability of credit default not occurring.

[0043] The original attribute values ​​are replaced with the evidence weights (WOE) values ​​of each attribute group, transforming the multi-dimensional attributes into a numerical vector with a unified dimension. This numerical vector is the second sub-feature vector.

[0044] S11-3. Construct a cause-effect graph based on the historical records; perform standardization processing on the business constraints to obtain standardized constraints.

[0045] In this application, historical records include user historical records and business historical records. User historical records include user historical behavior records and historical attribute records, while business historical records include historical business strategy records and corresponding result records. Business constraints include semi-structured and unstructured business constraint rules. Constructing a cause-effect graph based on historical records can be achieved through the following steps: Extract key variables from historical records, such as user behavior, business strategies, and the results of those strategies.

[0046] Based on the statistical relationships between variables and business logic, such as the relationship between overdue records and default risk in a credit scenario, a preliminary causal graph is constructed. This causal graph includes nodes, edges, and directions. Variables are filled into nodes, statistical relationships between variables are filled into corresponding edges, and business logic is used as the direction of the edges.

[0047] Standardization of business constraints specifically includes: Classify the constraint types and determine the boundary constraints, such as the interest rate range of 0.3% to 15%.

[0048] Standardize the units of measurement to convert different business constraints into the same numerical order of magnitude.

[0049] Convert it into a computable mathematical form, such as converting an interest rate of no more than 15% into a constraint of ≤0.15%.

[0050] S11-4. Generate the basic feature vector that resists distribution shift based on the first sub-feature vector, the second sub-feature vector, the causal graph, and the standardized constraint conditions.

[0051] By aligning and concatenating the first and second sub-feature vectors according to their dimensions, a two-dimensional basic feature matrix of environment and user is obtained. The variable correlation information in the initial causal graph is converted into weight factors of the feature matrix. These weight factors are positively correlated with the variable correlation information, specifically, strongly correlated variables have higher weight coefficients, ensuring the feature matrix carries causal correlation information. Furthermore, the edge transformation constraints are converted into boundary masks of the feature matrix, ensuring the basic feature vector does not deviate from business constraints in subsequent processing. Finally, the feature matrix is ​​normalized to eliminate differences in the dimensions of different dimensions, generating a fixed-length vector, which is the basic feature vector resistant to distribution shift. Thus, this basic feature vector resistant to distribution shift contains information from four dimensions: environment, user, causality, and constraints, and is an integrated feature vector that is resistant to shift, computable, and does not deviate from business rules.

[0052] By employing the embodiments of this application, the fragmented state of multi-source preprocessing results can be broken down. Through feature concatenation, causal embedding, and constraint fusion, a basic feature vector is generated that possesses resistance to distribution shift (derived from wavelet transform), quantization computability (derived from WOE encoding), a foundation for causal analysis (derived from causal graphs), and adaptability to business constraints (derived from constraint standardization). This provides high-quality, highly adaptable input for subsequent high-order causal feature extraction and pure causal relationship mining, ensuring the coherence and effectiveness of the entire process.

[0053] Further, step S12 is executed, which can be broken down into two steps: Step 1: Extract higher-order causal features from the basic feature vector resisting distribution shift. These higher-order causal features are characterized by nonlinearity and multi-dimensional interaction.

[0054] Step 2: Accurately identify false and confusing paths between "features - business operation strategies - strategy results".

[0055] Throughout the entire process of steps 1 and 2, it is necessary to ensure the interpretability of causal analysis and to ensure that business compliance and policy traceability requirements are met.

[0056] As one implementation method, the above two steps can be achieved through the following steps: The extraction of higher-order causal features from the underlying feature vector resisting distribution shift includes: A pre-set wavelet domain encoder is used to scale and translate the basic feature vector resisting distribution migration, extracting coarse-grained long-term features and fine-grained short-term features from the basic feature vector resisting distribution migration, and generating a high-order causal feature tensor based on the coarse-grained long-term features and fine-grained short-term features. A causal graph is constructed based on SHAP (SHapley Additive exPlanations) values, and spurious association paths between features and business operation strategies are determined based on the causal graph.

[0057] The fundamental feature vector resisting distribution migration in this application has been stripped of periodic perturbations and short-term noise, and includes environmental trends, behavioral essence, and static attributes. Through this fundamental feature vector resisting distribution migration, In step 1, high-order causal feature extraction is performed on the basic feature vectors that resist distribution shift. The core objective is to capture the nonlinear interactions, spatiotemporal coupling, and hierarchical causal relationships between features. The specific operations are as follows: Multinomial interaction and attention weighting are performed on the underlying feature vector that is resistant to distribution shift. Specifically: 1. Generate second- or third-order interaction features (such as GDP trend × user repayment stability).

[0058] 2. Employ a causal attention mechanism to assign weights to interactive features (emphasizing only interactions strongly correlated with business objectives, such as giving a much higher weight to repayment stability × debt ratio than to region × occupation). This allows for the capture of complex causal relationships among second-order interactive features.

[0059] Furthermore, multi-scale analysis is performed using a wavelet domain encoder (WDE). The wavelet domain encoder is invoked to decompose the increased-dimensional interaction features at both coarse-grained and fine-grained levels. For coarse-grained decomposition, long-term causal features can be extracted, such as the long-term correlation between a user's repayment behavior over the past year and delinquency risk. For fine-grained decomposition, short-term causal features can be extracted, such as the short-term correlation between a user's monthly income fluctuations and delinquency risk. Then, the long-term and short-term causal features are fused to generate a higher-order causal feature tensor. This higher-order causal feature tensor contains information in three dimensions: feature, time, and causal hierarchy. Specifically, this higher-order causal feature tensor can be a three-dimensional tensor, with each dimension corresponding to the basic feature, interaction hierarchy, and time, respectively.

[0060] Furthermore, causal representation denoising is performed on this high-order causal feature tensor. Invalid interaction features within this tensor can be filtered out using a locality-sensitive hash weighted memory. This weighted memory stores the distribution of causal features from similar historical scenarios. Based on this recorded distribution, invalid interaction features belonging to this weighted memory can be identified and removed. Invalid interaction features refer to interactions without causal significance; for example, a user's height and the risk of overdue payments are invalid interaction features.

[0061] In this application, interpretability analysis identifies core features that can directly explain how a feature influences a business operation strategy or the outcome of that strategy. Specifically, SHAP (SHapley Additive Explanations) is used to calculate the SHAP value of each higher-order causal feature, which characterizes the contribution of that higher-order causal feature to the effectiveness of the business operation strategy. Then, for a single object to be analyzed, a locally linear explanation of the feature-business operation strategy effect is generated. Features with SHAP values ​​exceeding a preset SHAP threshold (e.g., the top 20%) are further selected as interpretability analysis features.

[0062] Furthermore, based on this interpretability identification feature, a feature-policy-outcome causal graph is constructed. In this graph, nodes are filled with interpretability features, business policies, policy outcomes, and potential confounding variables. Edges are filled with SHAP values, and the direction of the edges is determined based on business logic. Furthermore, confounding paths are identified. In this application, a confounding path refers to a seemingly reasonable but actually false association path formed by a confounding variable (a third-party factor that simultaneously affects both the source and outcome variables) acting as an influencing medium, causing two variables with no direct causal relationship to exhibit a strong statistical correlation.

[0063] Specifically, the d-separation criterion can be used to locate confounding paths. For example, if variable X points to both feature A and result B, then X->A->B can be considered a confounding path. This allows for the identification of confounding paths. For instance, in a credit scenario, identifying a user's education level, income, repayment ability, and default risk, education level is a confounding variable. Since education level does not directly affect default risk, this path constitutes a false path.

[0064] Furthermore, a pre-built structured equation model is invoked to extract pure causal relationship data from the source data of the object to be analyzed based on the obfuscation paths between the business operation strategies. This pre-built structured equation model (SEM) is a model for causal inference based on causal graphs, linear or nonlinear equations. It can quantify the true causal effect of feature-strategy-strategy outcome by controlling obfuscating variables and blocking obfuscation paths. Thus, by controlling obfuscating variables and blocking obfuscation paths through this structured equation model, the high-order causal feature tensor can be purified, removing obfuscation paths and obtaining pure causal relationship data.

[0065] Specifically, confounding paths can be removed (or blocked) through covariate adjustment and full-space integration: Covariate adjustment involves inputting confounding variables as control variables into the SEM (Structural Equation Modeling) equation, fixing the values ​​of the confounding variables (e.g., fixing the credit product lifecycle = maturity period). Full-space integration involves integrating all possible values ​​of the confounding variables to eliminate their interference with the feature-policy-outcome (e.g., integrating all values ​​in the credit product lifecycle during the growth, maturity, and decline periods to calculate the average causal effect between the "credit product lifecycle and credit price").

[0066] Furthermore, the SEM equation is solved to obtain the causal effect matrix, potential business operation strategies, and individual-level causal effect values ​​corresponding to each potential business operation strategy. The obtained information is then organized into a dataset, which corresponds to the pure causal relationship data. In the causal effect matrix, each element represents the causal effect of each interpretable feature on different potential business operation strategies; for example, the potential sales volume of credit products when credit prices increase by 0.3%.

[0067] Choosing this embodiment allows for the capture of the true causal relationship between features and business operation strategies, avoiding directional errors in strategy formulation. Furthermore, by extracting interpretable features through SHAP and LIME, the causal path can be traced and verified, meeting compliance requirements in highly supervised scenarios such as finance and healthcare. Moreover, the output clean causal relationship data supports differentiated and customized business operation strategies, rather than a one-size-fits-all approach, adapting to individual-level operational strategy needs.

[0068] In some possible embodiments, after the step of invoking a pre-set structured equation model to extract clean causal relationship data from the source data of the object to be analyzed based on the obfuscated paths between the business operation strategies, the method further includes: Based on the pure causal relationship data and the historical intervention behavior records in the historical records, the heterogeneous treatment effect of different objects to be analyzed on the intervention behavior is quantified by a dual machine learning algorithm, and a heterogeneous treatment effect vector of each object to be analyzed is generated. Under the constraints of the business conditions, based on the pure causal relationship data and the interpretability analysis identification features, the process of generating and outputting the business operation strategy of the object to be analyzed includes: Based on the heterogeneity processing effect vector and the interpretability analysis identification features, an improved optimization algorithm is used to determine the object to be analyzed that satisfies the optimal solution corresponding to the business constraints as the target object.

[0069] After extracting the pure causal relationship data and identifying features through interpretability analysis, the two need to be deeply integrated. Specifically, the Heterogeneous Treatment Effect (HTE) algorithm is used to quantify the differentiated responses of different subjects to intervention behaviors, i.e., the individual causal effects of different subjects. For example, a 1% interest rate reduction might increase user A's willingness to place an order by 20%, while a 1% interest rate reduction might only increase user B's willingness to place an order by 5%. This avoids the resource waste or poor business operation strategy effects caused by uniform interventions in traditional classic prediction models.

[0070] Specifically, a dual machine learning algorithm is used to split the prediction task and the effect estimation task to reduce model bias and adapt to high-dimensional features and residual unobserved confounding variables. Specifically, historical data can be split in a 7:3 ratio into a training set (for effect estimation) and a validation set (for accuracy verification). Further, two base models (such as gradient boosting trees and neural networks) are initialized. Then, model M1 is trained to predict the propensity score (i.e., the probability of an individual accepting intervention) for intervention behavior (such as whether to lower interest rates), and model M2 is trained to predict the potential outcome without intervention based on pure causal features, such as the user's willingness to borrow when interest rates are not lowered. The intervention propensity residual and outcome prediction residual are calculated for models M1 and M2 respectively, where the intervention propensity residual = actual intervention status - propensity score predicted by M1. The outcome prediction residual = actual outcome (such as loan response rate) - outcome without intervention in M2.

[0071] Furthermore, a linear regression model was constructed using the outcome prediction residuals as the dependent variable and the intervention propensity residuals as the independent variable, and the regression coefficients were solved. During this process, the interpretability analysis was introduced to identify the interaction term between the characteristics and the intervention, extending the regression model. Based on the extended regression model, the individual treatment effect value (denoted as the ITE value) for each subject under analysis was calculated.

[0072] Furthermore, for each subject to be analyzed, its ITE values ​​under different intervention types and / or intervention intensities are output, forming a heterogeneous treatment effect vector. For example, the vector for user A is: [interest rate cut 0.5% → 10%, interest rate cut 1% → 20%, interest rate cut 1.5% → 22%]). Then, the accuracy of the HTE estimation is verified using a validation set: the model's predicted ITE is compared with historical actual intervention results, requiring an accuracy of ≥85% (e.g., the model predicts a 20% increase in user A's ITE with a 1% interest rate cut, while the actual increase in historical data is 19.5%, meeting the requirement).

[0073] Finally, the heterogeneity treatment effect vector for each subject under analysis is output. This heterogeneity treatment effect vector is a multi-dimensional vector, with the dimensions corresponding to the intervention type and the intervention intensity, respectively. This heterogeneity treatment effect vector quantifies the sensitivity or benefit of each subject to the intervention, providing a quantitative basis for subsequent selection of the optimal subjects to be analyzed.

[0074] Further, step S13 is executed to determine the object to be analyzed with the optimal solution, and to generate the business operation strategy for that object. Specifically, under the condition of meeting the standardized constraints, the object to be analyzed that benefits the most after intervention can be selected as the target object. For example, the credit price that benefits the most after intervention can be selected as the target object.

[0075] In this application, the business operation strategy includes: a list of objects to be analyzed for the optimal solution, corresponding intervention plans, and proof that the constraints are satisfied. As one implementation method, generating and outputting the business operation strategy for the objects to be analyzed can be achieved through the following steps: Based on the target object of the optimal solution, a business revenue analysis report is generated. The business revenue analysis report includes: the degree to which the target object satisfies the business constraints and the sensitivity analysis results of the target object.

[0076] For example, a corresponding business revenue analysis report is generated for the optimal credit price. This report calculates the degree to which the business constraints corresponding to the optimal credit price are met, as well as the user's sensitivity to the target. For instance, if the optimal credit price is 3.68%, a 0.8% interest rate reduction satisfies the interest rate constraint of less than 15%, and the corresponding risk exposure sensitivity is 50%.

[0077] In summary, this application combines a multi-scale causal network architecture with constraint optimization to improve the model's robustness in dealing with distributed offset environments, the accuracy of causal effect estimation, and the efficiency of satisfying business constraints when generating business operation strategies.

[0078] Based on the method provided in the first aspect, in the second aspect, this application provides a business operation strategy generation apparatus, wherein, as... Figure 3 As shown, the device 30 includes: Preprocessing module 301 is used to acquire various source data of the object to be analyzed and perform data preprocessing on each source data to generate a basic feature vector that resists distribution shift. The source data includes: the historical records, business constraints, current business environment data and user attribute data of the object to be analyzed. The causal inference module 302 is used to extract high-order causal features from the basic feature vector resisting distribution shift, obtain the interpretability analysis and identification features of the object to be analyzed and the obfuscated paths between business operation strategies, and call a preset structured equation model to extract pure causal relationship data from the source data of the object to be analyzed based on the obfuscated paths between business operation strategies. The strategy generation module 303 is used to, under the constraints of the business constraints, call a preset intelligent strategy processing model to generate and output the business operation strategy of the object to be analyzed based on the pure causal relationship data and the interpretability analysis identification features, wherein the business operation strategy includes the target object of the optimal solution.

[0079] The collection, storage, use, processing, transmission, provision, and disclosure of user personal information involved in this application comply with relevant laws and regulations and do not violate public order and good morals.

[0080] The names of the messages or information exchanged between multiple devices in the embodiments of this application are for illustrative purposes only and are not intended to limit the scope of these messages or information.

[0081] Thirdly, exemplary embodiments of this application also provide an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor. The memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to cause the electronic device to perform a method according to an embodiment of this application.

[0082] An exemplary embodiment of this application also provides a non-transitory computer-readable storage medium storing a computer program, wherein the computer program, when executed by a computer's processor, is used to cause the computer to perform a method according to an embodiment of this application.

[0083] An exemplary embodiment of this application also provides a computer program product, including a computer program, wherein, when executed by a computer's processor, the computer program is used to cause the computer to perform a method according to an embodiment of this application.

[0084] refer to Figure 4 The present invention describes a structural block diagram of an electronic device 400 that can serve as a server or client of this application, which is an example of a hardware device that can be applied to various aspects of this application. The electronic device is intended to represent various forms of digital electronic computer devices, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the application described and / or claimed herein.

[0085] like Figure 4 As shown, the electronic device 400 includes a computing unit 401, which can perform various appropriate actions and processes based on a computer program stored in a read-only memory (ROM 402) or a computer program loaded from a storage unit 408 into a random access memory (RAM 403). The RAM 403 may also store various programs and data required for the operation of the electronic device 400. The computing unit 401, ROM 402, and RAM 403 are interconnected via a bus 404. An input / output interface (I / O interface 405) is also connected to the bus 404.

[0086] Multiple components in electronic device 400 are connected to I / O interface 405, including: input unit 406, output unit 407, storage unit 408, and communication unit 409. Input unit 406 can be any type of device capable of inputting information to electronic device 400. Input unit 406 can receive input digital or character information and generate key signal inputs related to user settings and / or function control of electronic device. Output unit 407 can be any type of device capable of presenting information and may include, but is not limited to, a display, speaker, video / audio output terminal, vibrator, and / or printer. Storage unit 408 may include, but is not limited to, disks and optical discs. Communication unit 409 allows electronic device 400 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks, and may include, but is not limited to, modems, network cards, infrared communication devices, wireless communication transceivers, and / or chipsets, such as Bluetooth™ devices, WiFi devices, WiMax devices, cellular communication devices, and / or the like.

[0087] The computing unit 401 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 401 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 401 performs the various methods and processes described above. For example, in some embodiments, the aforementioned business operation strategy generation method can be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as storage unit 408. In some embodiments, part or all of the computer program can be loaded and / or installed on the electronic device 400 via ROM 402 and / or communication unit 409. In some embodiments, the computing unit 401 can be configured to perform the aforementioned business operation strategy generation method by any other suitable means (e.g., by means of firmware).

[0088] The program code used to implement the methods of this application may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the functions / operations specified in the flowcharts and / or block diagrams are implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0089] In the context of this application, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0090] As used in this application, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, device, and / or apparatus (e.g., disk, optical disk, memory, programmable logic device (PLD)) for providing machine instructions and / or data to a programmable processor, including machine-readable media that receive machine instructions as machine-readable signals. The term "machine-readable signal" refers to any signal for providing machine instructions and / or data to a programmable processor.

[0091] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0092] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0093] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other.

Claims

1. A method for generating business operation strategies, characterized in that, The method includes: The source data of the object to be analyzed are obtained and the source data are preprocessed to generate a basic feature vector that is resistant to distribution shift. The source data includes: the historical records, business constraints, current business environment data and user attribute data of the object to be analyzed. Higher-order causal feature extraction is performed on the basic feature vector resisting distribution shift to obtain the interpretability analysis and identification features of the object to be analyzed and the obfuscated paths between business operation strategies. Then, a pre-set structured equation model is called to extract the pure causal relationship data from the source data of the object to be analyzed based on the obfuscated paths between business operation strategies. Under the constraints of the business conditions, a pre-set intelligent strategy processing model is invoked to generate and output the business operation strategy of the object to be analyzed based on the pure causal relationship data and the interpretability analysis identification features. The business operation strategy includes the target object of the optimal solution.

2. The method according to claim 1, characterized in that, After the step of extracting pure causal relationship data from the source data of the object to be analyzed based on the obfuscated paths between the business operation strategies by calling a pre-set structured equation model, the method further includes: Based on the pure causal relationship data and the historical intervention behavior records in the historical records, the heterogeneous treatment effect of different objects to be analyzed on the intervention behavior is quantified by a dual machine learning algorithm, and a heterogeneous treatment effect vector of each object to be analyzed is generated. Under the constraints of the business conditions, based on the pure causal relationship data and the interpretability analysis identification features, the process of generating and outputting the business operation strategy of the object to be analyzed includes: Based on the heterogeneity processing effect vector and the interpretability analysis identification features, an improved optimization algorithm is used to determine the object to be analyzed that satisfies the optimal solution corresponding to the business constraints as the target object.

3. The method according to claim 1, characterized in that, The process of acquiring various source data of the object to be analyzed and performing data preprocessing on each source data to generate a basic feature vector resistant to distribution shift includes: Frequency domain transformation and multi-scale analysis are performed on each of the source data to remove periodic disturbances and short-term noise contained in the source data, and to generate the basic feature vector that resists distribution migration.

4. The method according to claim 3, characterized in that, The step of performing frequency domain transformation and multi-scale analysis on each of the source data to remove periodic disturbances and short-term noise contained in the source data includes: The current business environment data is windowed using a combination of Fast Fourier Transform and Blackman window function to separate the periodic trends and sudden fluctuations of the business in the source data. The user behavior data in the user attribute data is decomposed into multiple scales to remove long-term essential features and short-term noise features from the user behavior data.

5. The method according to claim 1, characterized in that, The process of acquiring various source data of the object to be analyzed and performing data preprocessing on each source data to generate a basic feature vector resistant to distribution shift includes: Perform wavelet transform on the current business environment data to obtain the first sub-feature vector; Evidence weight encoding is performed on the user attribute data to obtain the second sub-feature vector; A cause-effect graph is constructed based on the historical records; the business constraints are then standardized to obtain standardized constraints. The basic feature vector resisting distribution shift is generated based on the first sub-feature vector, the second sub-feature vector, the causal graph, and the standardized constraints.

6. The method according to claim 1, characterized in that, The extraction of higher-order causal features from the underlying feature vector resisting distribution shift includes: A pre-set wavelet domain encoder is used to scale and translate the basic feature vector resisting distribution migration, extracting coarse-grained long-term features and fine-grained short-term features from the basic feature vector resisting distribution migration, and generating a high-order causal feature tensor based on the coarse-grained long-term features and fine-grained short-term features. A causal graph is constructed based on SHAP value analysis, and false association paths between features and business operation strategies are determined based on the causal graph.

7. The method according to claim 1, characterized in that, The process of generating and outputting the business operation strategy for the object to be analyzed includes: Based on the target object of the optimal solution, a business revenue analysis report is generated. The business revenue analysis report includes: the degree to which the target object satisfies the business constraints and the sensitivity analysis results of the target object.

8. The method according to claim 1, characterized in that, If the method is applied to a credit business scenario, the object to be analyzed is the credit price; if the method is applied to a medical resource business scenario, the object to be analyzed is the medical resource; if the method is applied to a logistics business scenario, the object to be analyzed is the transportation cost.

9. A business operation strategy generation device, characterized in that, The device includes: The preprocessing module is used to acquire various source data of the object to be analyzed and perform data preprocessing on each source data to generate a basic feature vector that is resistant to distribution shift. The source data includes: the historical records, business constraints, current business environment data, and user attribute data of the object to be analyzed. The causal inference module is used to extract high-order causal features from the basic feature vector resisting distribution shift, obtain the interpretability analysis and identification features of the object to be analyzed and the obfuscated paths between business operation strategies, and call a preset structured equation model to extract pure causal relationship data from the source data of the object to be analyzed based on the obfuscated paths between business operation strategies. The strategy generation module is used to, under the constraints of the business constraints, call a pre-set intelligent strategy processing model to generate and output the business operation strategy of the object to be analyzed based on the pure causal relationship data and the interpretability analysis identification features, wherein the business operation strategy includes the target object of the optimal solution.

10. An electronic device, characterized in that, The electronic device includes: a processor and a memory storing a program; wherein the program includes instructions that, when executed by the processor, cause the processor to perform the method according to any one of claims 1-8.

Images