Target object mining method and device, equipment and storage medium

Abstract

This application discloses a method, apparatus, device, and storage medium for target object mining, relating to the field of information processing technology. The method includes: acquiring a standard dataset under a target business scenario, wherein the standard dataset is obtained by associating positive and negative sample customer group data and a set of effective features under the target business scenario; training an ensemble tree model based on the standard dataset, extracting model decision paths from the ensemble tree model and performing escape processing, quantitative evaluation, and screening combination to obtain a set of recommendation rules; performing threshold optimization on the set of recommendation rules according to the target business scenario to obtain optimization rules that satisfy the target business scenario; and performing rule matching and list generation on the full customer group data using the optimization rules to obtain a list of target objects corresponding to the target business scenario. This application can reduce the time spent on rule development processes while deeply mining the business correlation patterns hidden in massive amounts of data, effectively expanding the coverage of target object mining.

Description

Technical Field

[0001] This application relates to the field of information processing technology, and in particular to a method, apparatus, device and storage medium for target object mining. Background Technology

[0002] Current rule mining and model application technologies for financial business, such as target customer group mining and list selection based on tag rules, are mainly divided into expert experience-driven rule development models. That is, data analysts manually design rules based on business experience, and complete the rule launch through requirements submission, IT development implementation, and effect feedback. The whole process relies on manual sorting and offline tools.

[0003] Existing technologies have significant shortcomings in practical applications. Rule development requires high levels of professional expertise from personnel, lacks standardized and modular tools to support the process, and involves lengthy development chains, large human resource investments, and long cycles. Rule generation is highly dependent on expert experience, limiting the scope of data mining and covering only a single customer group, making it difficult to discover hidden business models from massive amounts of high-dimensional data. Summary of the Invention

[0004] The main purpose of this application is to provide a method, apparatus, device and storage medium for target object mining, which aims to break free from the constraints of human experience, shorten the development cycle of customer group mining rules, and mine potential customer group association patterns from massive data, thereby broadening the scope of customer group mining.

[0005] To achieve the above objectives, this application proposes a target object mining method, the method comprising: Obtain a standard dataset for the target business scenario, which is obtained by associating positive and negative sample customer group data and a set of effective features for the target business scenario; An ensemble tree model is trained based on the standard dataset. The model decision paths in the trained ensemble tree model are extracted, escaped, quantitatively evaluated, and screened to obtain a set of recommendation rules. The recommendation rule set is threshold-adjusted according to the target business scenario to obtain the adjusted rules that satisfy the target business scenario; The optimization rules are used to match and generate a list of target objects for the target business scenario by matching the rules with the full customer data.

[0006] In one possible implementation, obtaining the standard dataset for the target business scenario includes: Based on the target business scenario, set positive and negative sample differentiation conditions, and extract and construct positive and negative sample customer group data with positive and negative labels from the business data lake based on the positive and negative sample differentiation conditions. Statistical analysis and feature filtering are performed on the feature data related to the target business scenario, and attribution analysis is used to determine the effective feature set related to the business objectives in the target business scenario. The positive and negative sample customer group data are associated and integrated with the effective feature set according to the object identifier to form the standard dataset required for rule mining.

[0007] In one possible implementation, the ensemble tree model is trained based on the standard dataset. The model decision paths in the trained ensemble tree model are extracted and subjected to escaping, quantitative evaluation, and selection / combination to obtain a set of recommendation rules, including: Model training is conducted based on the aforementioned standard dataset, and a fully fitted ensemble tree model is obtained through distributed hyperparameter search training. The integrated tree model is subjected to path traversal processing, the model decision path is converted into a business-understandable rule expression and rule compression and deduplication are performed to obtain an independent decision path set; A multi-dimensional quantitative evaluation is performed on the set of independent decision paths, and rule selection and combination are completed based on preset evaluation indicators to obtain a recommended rule set.

[0008] In one possible implementation, the path traversal processing performed on the ensemble tree model transforms the model decision paths into business-understandable rule expressions and performs rule compression and deduplication to obtain a set of independent decision paths, including: A depth-first traversal method is used to perform a global path scan on the ensemble tree model, extract all decision paths within the ensemble tree model, and form an initial path set. Based on the preset feature attribute information, each initial decision path is converted into a business-understandable rule expression, completing the escaping process from machine logic to business logic; Redundancy merging and duplication removal are performed on each rule expression to compress the initial rule path containing the correlation, resulting in a set of independent decision paths.

[0009] In one possible implementation, the step of threshold tuning of the recommendation rule set based on the target business scenario to obtain tuning rules that satisfy the target business scenario includes: The rule optimization target is determined based on the customer scale requirements of the target business scenario, and a rule search space is constructed based on the business objectives in the target business scenario; The recommendation rule set is iteratively optimized within the search space using a rule tuning algorithm to determine the optimal threshold combination that makes the rule effect match the business objective; The recommendation rule set is adjusted and optimized based on the optimal threshold combination to obtain optimized rules that meet the requirements of the business scenario.

[0010] In one possible implementation, the step of employing a rule optimization algorithm to iteratively optimize the set of recommendation rules within the search space to determine the optimal combination of thresholds that makes the rule effects match the business objectives includes: For any recommendation rule in the set of recommendation rules, the adjustment direction and fluctuation range corresponding to the rule threshold in the recommendation rule are determined by comparing the deviation between the rule filtering effect of the recommendation rule and the business objective. Based on the adjustment trend and the customer group characteristic distribution pattern reflected in the standard dataset, multiple candidate thresholds are extracted within the floating range to form a combined threshold sample for comparison. The effectiveness of each combined threshold sample is verified item by item, and the optimal threshold combination that fits the business objective is selected.

[0011] In one possible implementation, the step of performing rule matching and list generation on the full customer data using the optimization rules to obtain the target object list corresponding to the target business scenario includes: Basic entry criteria are set based on the target business scenario; Based on the optimization rules and the basic inclusion criteria, the full customer data is double-screened to obtain an initial customer list. The initial customer group list is sorted, arranged, and rendered to form the target object list corresponding to the target business scenario.

[0012] Furthermore, to achieve the above objectives, this application also proposes a target object excavation device, the target object excavation device comprising: The acquisition module is used to acquire a standard dataset under the target business scenario. The standard dataset is obtained by associating positive and negative sample customer group data and a set of effective features under the target business scenario. The escaping module is used to train an ensemble tree model based on the standard dataset, extract the model decision paths in the trained ensemble tree model, and perform escaping, quantitative evaluation, and screening and combination to obtain a set of recommendation rules. The optimization module is used to perform threshold optimization on the recommendation rule set according to the target business scenario to obtain optimization rules that meet the target business scenario; The generation module is used to perform rule matching and list generation on the full customer data using the optimization rules to obtain the target object list corresponding to the target business scenario.

[0013] In addition, to achieve the above objectives, this application also proposes a target object mining device, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the target object mining method as described above.

[0014] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and which, when executed by a processor, implements the steps of the target object mining method described above.

[0015] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the target object mining method described above.

[0016] This application provides a method, apparatus, device, and storage medium for target object mining. The target object mining method acquires a standard dataset under a target business scenario. The standard dataset is obtained by associating positive and negative sample customer group data and effective feature sets under the target business scenario. Then, an ensemble tree model is trained based on the standard dataset. The model decision paths in the trained ensemble tree model are extracted and subjected to escaping processing, quantitative evaluation, and screening combination to obtain a set of recommendation rules. The set of recommendation rules is then threshold-adjusted according to the target business scenario to obtain optimized rules that meet the target business scenario. The optimized rules are then used to perform rule matching and list generation on the full customer group data to obtain a list of target objects corresponding to the target business scenario. Thus, the model is trained on the dataset to automatically generate and optimize the screening rules, without relying entirely on manual experience for customization, which can reduce the development process time and deeply mine the business correlation patterns hidden in massive data, effectively expanding the coverage of target object mining. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0018] 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, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 System architecture diagram provided for the target object mining method of this application; Figure 2This is a flowchart illustrating an embodiment of the target object mining method of this application. Figure 3 A scenario implementation diagram of feature index analysis in feature analysis provided for the target object mining method of this application; Figure 4 A scenario implementation diagram showing the ranking of feature importance in feature analysis, provided for the target object mining method of this application; Figure 5 A scenario implementation diagram of correlation analysis in feature analysis provided for the target object mining method of this application; Figure 6 One of the scenario implementation diagrams for attribution analysis in feature analysis provided for the target object mining method of this application; Figure 7 The second scenario implementation diagram for attribution analysis in feature analysis provided for the target object mining method of this application; Figure 8 Example diagrams of rule compression provided for the target object mining method of this application; Figure 9 This is a schematic diagram of the module structure of the target object mining device in an embodiment of this application; Figure 10 This is a schematic diagram of the device structure of the hardware operating environment involved in the target object mining method in the embodiments of this application.

[0020] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0021] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0022] While existing machine learning models can output classification or prediction results based on massive amounts of feature data, the decision paths within these models are typically represented by machine feature identifiers, encoded values, and model thresholds. These are difficult to directly translate into rule expressions understandable to business users, resulting in a lack of interpretability and reusability in the model results. Even when extracting some rules manually from the model, there is a lack of a unified mechanism for evaluating rule support, confidence, lift, recall, and stability, making it difficult to automatically identify rule sets with high accuracy, high coverage, and strong generalization capabilities.

[0023] Furthermore, the rule mining process typically involves two types of computational requirements: interactive analysis and offline batch prediction. The former demands low latency and rapid feedback, while the latter requires high throughput and stable operation. Existing single computing platforms struggle to meet both requirements simultaneously, often resulting in slow response times for interactive rule debugging or insufficient stability in offline list generation tasks.

[0024] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0025] It should be noted that the executing entity in this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device, big data service platform, or target object mining system capable of performing the above functions. The following description uses a target object mining system as an example to illustrate this embodiment and the subsequent embodiments. The target object mining system adopts a layered architecture, divided from bottom to top into an infrastructure layer, a computing engine layer, a service component layer, and an application layer. (See reference...) Figure 1 .

[0026] The infrastructure layer provides the runtime environment for software services. The K8S cloud environment provides various resources such as memory, network, and disk required by the platform services. The Ray computing cluster is also deployed based on the cloud environment, providing a large number of elastic computing resources to support various distributed computing tasks in the distributed rule mining process.

[0027] The computing engine layer completes the standard process of rule mining by encapsulating modules such as model rule recommendation, interactive rule evaluation, and rule prediction list generation, shortening and simplifying the rule mining steps. Furthermore, the computing engine is implemented based on the Ray and PySpark distributed frameworks, employing various analytical methods such as statistical analysis and machine learning to provide highly interpretable rule analysis results and possess big data analysis capabilities. Different computing platforms are used according to the different needs of the business strategy mining stage. In the rule development stage (interactive rule mining stage), KubeRay is used to perform sample analysis, feature analysis, model training, rule evaluation, and rule tuning. This fully leverages the elastic resource management characteristics of Ray's distributed computing and containerized environment, providing highly timely and flexible computing support for rule development and offering a fast-response data analysis interactive experience. In the rule deployment stage (rule deployment / prediction stage), PySpark asynchronous job mode is used to perform full list generation and batch rule prediction. Relying on the high throughput of the Hadoop cluster, it provides stable and reliable rule prediction services; the two types of tasks are isolated in terms of computing resources, scheduling queues, and runtime environment. In addition, this architecture design alleviates the problem of scarce computing resources during the development phase, ensures the high timeliness of interactive rule mining, and avoids interference from high-concurrency development tasks on the stability of production tasks. That is, it reduces resource contention between interactive tasks and production batch tasks, and improves response speed and stability.

[0028] The service component layer empowers the application layer through an "API interface + front-end component" approach. The service layer encapsulates multiple rule-mining related machine learning algorithm services, provides standardized front-end components that can be directly embedded into application layer pages, and offers functions such as interactive computation configuration, visualized query of computation results, and interactive analysis of computation results.

[0029] The application layer is implemented using web pages. The rule analysis component is embedded into the application system through service-oriented APIs and front-end components, enabling the application layer to have interactive data analysis capabilities.

[0030] In one embodiment, during the interactive rule mining phase, the system encapsulates sample analysis, feature analysis, ensemble tree model training, model decision path escaping, rule evaluation, and rule tuning tasks into interactive computing tasks, which are executed by an interactive computing engine based on KubeRay or Ray. The computing results are then returned to the front-end components for display and interactive adjustment via API services.

[0031] During the rule deployment or prediction phase, the system takes the optimized rules, basic entry conditions, and full customer data as input to generate a batch prediction task. This task is executed on a Hadoop cluster by an asynchronous computing engine based on PySpark, and outputs a list of target objects, rule hit details, and list ranking results.

[0032] Interactive computing tasks and asynchronous batch prediction tasks are isolated from each other in terms of computing resources, scheduling queues, runtime environment, and task priority to avoid resource contention between interactive exploration tasks and production batch tasks.

[0033] Based on this, embodiments of this application provide a target object mining method, referring to... Figure 2 , Figure 2 This is a flowchart illustrating an embodiment of the target object mining method of this application.

[0034] In this embodiment, the target object mining method includes steps S11 to S14: Step S11: Obtain a standard dataset for the target business scenario. The standard dataset is obtained by associating positive and negative sample customer group data and effective feature set for the target business scenario. It should be noted that the target business scenario refers to application scenarios within the financial field with clear business objectives, such as corporate marketing, risk control, or customer retention; the standard dataset refers to a standardized data set used for rule mining and model training, after labeling and feature association processing; positive and negative sample customer group data refer to positive sample customer group data that meets the expected results and negative sample customer group data that does not meet the expected results, divided according to business objectives; and the effective feature set refers to a set of feature data that is relevant to the business objectives and can support rule mining and customer group judgment. Thus, by clarifying the boundaries of the business scenario, completing sample labeling and feature selection, and avoiding interference from invalid data, the accuracy and stability of rule mining are improved. In one possible implementation, the system can determine the target business scenario through user interactive configuration or automatically load the corresponding business scenario parameters through scenario templates.

[0035] Specifically, the system first determines the criteria for dividing positive and negative samples based on the business needs of the target business scenario, extracts the corresponding customer groups from the business data lake and completes the labeling, then conducts statistical analysis and attribution analysis on the business-related feature data, and filters out the features that can effectively distinguish the samples to form an effective feature set. Finally, the labeled customer group data and the effective feature set are associated and integrated according to a unified object identifier to form a standard dataset with a standardized structure and reliable quality, thereby reducing the interference caused by invalid data from the source and improving the accuracy and reliability of rule mining.

[0036] For example, in the context of corporate financial marketing, the system uses customers who have completed the designated business within the past year as positive sample customer data and customers who have not completed the business as negative sample customer data. It selects account balance, number of transactions, and years of operation as effective feature sets, and associates customer data with feature data according to the customer's unique code to form a standard dataset suitable for this marketing scenario.

[0037] Step S12: Train an ensemble tree model based on the standard dataset, extract the model decision paths from the trained ensemble tree model, and perform escaping, quantitative evaluation, and screening to obtain a set of recommendation rules. It should be noted that the ensemble tree model refers to a machine learning model constructed by integrating multiple decision trees for classification and rule mining; the model decision path refers to the multi-layered decision-making chain that the ensemble tree model traverses when judging customer data; the escaping process refers to the process of converting the machine-recognizable decision path into rule expressions that business personnel can understand; the quantitative evaluation refers to the process of numerically evaluating the effectiveness, stability, and generalization ability of the rules based on preset indicators; the screening and combination process refers to the process of eliminating inferior rules and logically combining high-quality rules; and the recommended rule set refers to the set of high-quality rules that can be directly used in business after evaluation and screening. Thus, by automatically mining the inherent patterns in the data through model training, the logic of the black-box model is transformed into interpretable rules, which can get rid of the dependence on human expert experience and improve the efficiency and coverage of rule generation.

[0038] In one possible implementation, the ensemble tree model can be trained using the LightGBM model, and the model training process can employ a distributed hyperparameter search approach to improve the fitting effect.

[0039] Specifically, the system trains an ensemble tree model using a standard dataset as input. Through hyperparameter optimization, the model is made to fully fit the data distribution. Then, a global path traversal is performed on the trained model to extract all model decision paths and translate them into rule expressions that business personnel can directly understand. Subsequently, the rule expressions are compressed, deduplicated, and redundantly removed to obtain an independent and standardized set of decision paths. Furthermore, the rules are quantitatively evaluated based on multi-dimensional indicators such as support, confidence, and lift. Based on the evaluation results, inferior rules with weak generalization ability and low accuracy are removed, and high-quality rules are logically combined to finally obtain a set of recommendation rules with balanced accuracy and coverage that can be directly used for business, thereby achieving the goal of automatically mining hidden business patterns from massive high-dimensional data.

[0040] For example, the system trains a LightGBM ensemble tree model based on a standard dataset, extracts all decision paths of the model through depth-first traversal, and converts the paths into rule expressions of "account balance greater than the threshold and number of transactions greater than the threshold". Based on support, confidence and lift, the system completes rule evaluation and screening to obtain a set of recommended rules suitable for customer group mining. In one embodiment, complementary rules can be combined using OR. OR combination refers to associating multiple independent rules in a logical OR manner, so that the object to be screened can be determined as a hit if it meets any one of the independent rules. This improves the coverage of the customer group and expands the number of effective customers to be mined while ensuring the accuracy of the rules.

[0041] Step S13: Optimize the threshold of the recommendation rule set according to the target business scenario to obtain optimization rules that satisfy the target business scenario; It should be noted that threshold tuning refers to adjusting and optimizing the judgment thresholds in rules to ensure that the rule's effect matches business objectives. Optimized rules refer to rules whose customer base size, accuracy, and coverage all meet the requirements of the business scenario after threshold adjustment. Therefore, threshold tuning allows rules to align with the actual customer base size requirements of the business, balancing rule accuracy and coverage, and avoiding poor business performance caused by overly strict or lenient rules. In one possible implementation, the rule tuning algorithm can use the TPE (Tree-structured Parzen estimator) algorithm for iterative optimization. The direction of threshold adjustment can be determined based on the deviation between the target coverage and the current coverage, where target coverage = number of target names / number of eligible customers; current coverage = number of current rule hits / number of eligible customers.

[0042] Specifically, the system first determines the target indicators for rule optimization based on the customer base size requirements of the target business scenario, constructs a search space for rule thresholds based on business evaluation standards, and then uses a rule optimization algorithm to iteratively optimize the set of recommended rules within the search space. The system determines the direction of threshold adjustment based on the deviation between the rule selection effect and the business objective, generates multiple sets of candidate threshold combinations within a reasonable range, verifies the effect of each combination, and selects the optimal threshold combination that matches the business objective in terms of rule accuracy, coverage, and customer base size. Finally, the system adjusts and optimizes the recommended rules based on the optimal threshold combination to obtain optimized rules that perfectly meet the needs of the business scenario.

[0043] For example, in a financial marketing customer segmentation scenario, where the business objective is to select 1,000 target customers, the system uses the thresholds corresponding to account balance and number of transactions in the recommendation rules as optimization targets, constructs a reasonable threshold search range, and uses the TPE algorithm to iteratively find the optimal threshold combination so that the number and accuracy of the customer segments selected by the rules meet the requirements of the marketing scenario, thus obtaining the optimized rules.

[0044] Step S14: The optimization rules are used to perform rule matching and list generation on the full customer data to obtain the target object list corresponding to the target business scenario.

[0045] It should be noted that the full customer base data refers to all customer data within the business scope that has not been filtered; rule matching refers to the process of comparing the optimized rules with the full customer base data one by one to determine whether the customer base meets the rule conditions; list generation refers to the process of organizing the matched customer base into a standardized business list; the target object list refers to the final customer base list that meets the needs of the business scenario and can be directly used for marketing or risk control business; thus, through the two-layer screening and list organization output, the target objects that meet the business requirements can be generated quickly and stably, shortening the cycle from rule to business implementation.

[0046] In one possible implementation, the system can use PySpark asynchronous scheduling to complete the matching calculation of the entire customer group data, ensuring processing stability under high throughput.

[0047] Specifically, the system first sets basic entry conditions based on the business needs of the target business scenario, performs preliminary filtering on the full customer data to remove customers who obviously do not meet the requirements, then superimposes the optimization rules with the basic entry conditions, performs double screening and matching judgment on the initially filtered customer groups, organizes the customer groups that meet all conditions into an initial customer group list, and further sorts, organizes and renders the results of the initial customer group list, finally forming a target object list with a clear structure that can be directly used for business deployment.

[0048] For example, for corporate marketing scenarios, the system first sets the basic entry condition of "operating for more than one year", performs preliminary filtering on all corporate customer data, and then uses optimization rules to accurately match the filtered customer groups. The successfully matched customer groups are sorted by customer level and rendered into a visual list, resulting in a list of target objects that can be directly used for marketing and promotion in this scenario.

[0049] This embodiment first constructs a standardized labeled dataset, then automatically mines and transforms interpretable rules based on an ensemble tree model, and intelligently optimizes rule thresholds in conjunction with business scenarios. Finally, it generates a list of target objects through a two-layer screening process. This effectively reduces the reliance on human experience in rule mining, shortens the rule development cycle, improves rule coverage and mining accuracy, and ensures that the rules have good interpretability. It can stably support the efficient operation of corporate marketing, risk control and other businesses in the financial sector.

[0050] In one feasible implementation, obtaining the standard dataset for the target business scenario includes: Step S21: Set positive and negative sample differentiation conditions according to the target business scenario, and extract and construct positive and negative sample customer group data with positive and negative labels from the business data lake based on the positive and negative sample differentiation conditions. It should be noted that the positive and negative sample differentiation criteria refer to the judgment criteria used to divide positive and negative sample customer groups; the business data lake refers to a unified dataset used to store all customer business data in the financial field; positive and negative sample customer group data with positive and negative labels refer to customer group data labeled with positive or negative sample labels according to the differentiation criteria; thus, by clarifying the business scenario and setting unified sample differentiation criteria, the standardization and consistency of sample selection can be guaranteed. In one possible implementation, the system can configure the positive and negative sample differentiation criteria through an interactive page, or it can directly load the differentiation criteria built into the scenario template.

[0051] Specifically, the system first determines the criteria for judging positive and negative samples based on the business objectives of the target business scenario, forming clear and unified conditions for distinguishing positive and negative samples. Then, based on these conditions, it retrieves and extracts all customer data in the business data lake, labels customers who meet the positive sample criteria as positive samples, and labels customers who meet the negative sample criteria as negative samples. Finally, it completes the construction of labeled positive and negative sample customer group data, enabling subsequent model training to be carried out based on clear sample guidance, thereby improving the targeting and accuracy of rule mining.

[0052] For example, in the context of marketing to corporate clients, the system sets customers who have successfully processed corporate settlement products within the past year as positive samples and customers who have not processed such products as negative samples. Based on this condition, the system extracts the corresponding customers from the business data lake and completes the labeling, forming positive and negative sample customer group data with positive and negative labels.

[0053] Step S22: Perform statistical analysis and feature filtering on the feature data related to the target business scenario, and determine the effective feature set related to the business objectives in the target business scenario through attribution analysis; It should be noted that feature data refers to various quantitative and non-quantitative data used to describe customer attributes, behaviors, accounts, transactions, and business operations; statistical analysis refers to the analytical methods used to calculate indicators such as the null value rate, deviation, information content, and correlation of feature data; feature selection refers to the process of eliminating invalid features and retaining highly discriminative features based on statistical indicators; attribution analysis refers to the analytical methods used to determine the degree of contribution and direction of influence of features to business objectives; and the effective feature set refers to the set of features that are highly relevant to business objectives and can effectively distinguish between positive and negative samples. Therefore, through multi-dimensional feature analysis and selection, redundant and interfering features can be eliminated, improving the efficiency of subsequent model training and the reliability of rules. In one possible implementation, attribution analysis can be implemented using the Shapley method, and feature selection can be completed by combining feature importance ranking.

[0054] Specifically, the system first extracts initial feature data related to the target business scenario. This can be done by selecting a feature package related to the target business scenario from the system's recommended feature packages, selecting features from a multi-level feature classification tree, or using a combination of both to select potentially useful candidate features. Then, feature date settings are applied to the initial feature data to unify the time calculation caliber and ensure that the feature calculation basis for all samples is consistent. An initial dataset is formed based on the feature selection and date setting results. In one possible implementation, feature date setting can support three methods, including automatically calculating the feature date based on the sample date and the observation period window, directly specifying a fixed date, and obtaining the corresponding feature date by time offsetting the specified date feature.

[0055] Furthermore, for each feature in the initial dataset, feature metrics such as null value rate, deviation, information content, and chi-square correlation are calculated. A first feature filtering is performed based on preset screening thresholds. Then, the feature importance value of each feature is obtained by calculating the split gain. Features are sorted from high to low importance for a second feature filtering, eliminating less important features to obtain a streamlined final feature set. Based on this final feature set, the Shapley method is used for attribution analysis to analyze the average contribution, monotonicity, and group distribution of each feature to the business objective within its value range. Finally, an effective feature set that is highly relevant to the business objective and has strong discriminative power is determined. (For reference...) Figure 3 - Figure 7 .

[0056] For example, in a corporate marketing scenario, the system extracts feature data such as years of operation, account balance, number of transactions, and total tax payment. It uses an observation period window to set the feature dates, calculates the null value rate, deviation, and correlation of each feature, and completes the first screening. It then calculates the feature importance through split gain and removes features with lower importance to complete the second screening. Finally, it uses the Shapley method to analyze the contribution of each feature to customer business transactions, and finally obtains the effective feature set.

[0057] Step S23: The positive and negative sample customer group data and the effective feature set are associated and integrated according to the object identifier to form the standard dataset required for rule mining.

[0058] It should be noted that the object identifier refers to a unique code used to identify each customer; association and integration refers to the process of matching and concatenating customer group tags with corresponding characteristics according to the same object identifier. The system achieves accurate association between customer groups and characteristics through object identifiers, avoiding data misalignment, missing data, or duplication, and ensuring the integrity and usability of the dataset. In one possible implementation, the object identifier can be implemented using a unique customer number, unified social credit code, or other unique identifiers.

[0059] Specifically, the system uses the object identifier corresponding to each customer as the basis for association, and matches and splices the positive and negative sample customer group data with the labeled positive and negative labels with the feature data in the effective feature set one by one. This ensures that each sample data contains complete label information and corresponding feature information. The associated data is then deduplicated, normalized, and formatted to form a standard dataset that is structurally standardized, of reliable quality, and can be directly input into the model for training.

[0060] For example, the system uses the customer's unique customer number as the object identifier, and matches and associates the customer group data with positive and negative labels with effective features such as years of operation, account balance, and number of transactions line by line, removing duplicate samples and abnormal data to form a standard dataset with complete fields and uniform format.

[0061] This embodiment first sets sample differentiation conditions based on business scenarios and constructs labeled customer group data. Then, it performs statistical analysis, feature screening, and attribution analysis on the feature data to extract effective features. Finally, it completes the accurate association and integration of customer group data and feature data according to unique object identifiers. This can build a standardized, high-quality, and highly available rule mining dataset, reduce the impact of data noise from the source, and improve the accuracy, stability, and efficiency of subsequent model training and rule generation. At the same time, it gets rid of the subjective limitations of traditional manual selection of samples and features, making the customer group mining process more objective, standardized, and reusable.

[0062] In one feasible implementation, the step of training an ensemble tree model based on the standard dataset, extracting the model decision paths from the trained ensemble tree model and performing escaping processing, quantitative evaluation, and screening and combination, yields a set of recommendation rules, including: Step S31: Based on the standard dataset, perform model training and obtain a fully fitted ensemble tree model through distributed hyperparameter search training; It should be noted that model training refers to the process of learning and fitting an ensemble tree model using a standard dataset to capture the inherent patterns in the data; distributed hyperparameter search refers to a method of traversing or intelligently optimizing the combination of model hyperparameters in a distributed computing environment to improve the model's fitting effect; an ensemble tree model refers to a classification model constructed by integrating multiple decision trees, possessing high stability and high interpretability potential; a fully fitted ensemble tree model refers to a model that has completed sufficient learning on a standard dataset and can stably output decision paths. The system can significantly improve model training efficiency and generalization ability through distributed hyperparameter search, avoiding the instability caused by manual parameter tuning. In one possible implementation, the ensemble tree model can be implemented using the LightGBM model, and the distributed hyperparameter search can be performed using grid search or random search methods.

[0063] Specifically, the system first sets rule parameters according to business needs, and selects either efficiency-first or accuracy-first rule generation methods to control the model training strategy. The efficiency-first mode uses default hyperparameters to shorten training time, while the accuracy-first mode expands the computational scope and performs hyperparameter grid search to ensure rule accuracy. At the same time, the system completes the configuration of rule filtering parameters, filters the standard dataset in various dimensions to find the best set of rules, and then uses the standard dataset, rule generation parameters, and rule filtering parameters as inputs to perform distributed hyperparameter search under the support of distributed computing. Based on the selected training strategy, the system completes the training of the ensemble tree model, obtains an ensemble tree model that fully fits the data distribution and can stably output decision paths, and simultaneously obtains the model's quantitative performance evaluation indicators.

[0064] For example, for the business scenario of marketing products to corporate clients, the system sets the rule generation method to the precision-first mode, expands the rule calculation scope and performs grid search on hyperparameters, and configures rule filtering parameters related to rule support, confidence and lift. Based on the standard dataset that has been built, the system uses the LightGBM model to carry out distributed hyperparameter search and model training, and finally obtains an ensemble tree model with complete fitting and strong generalization ability, and outputs model quantitative performance evaluation indicators related to model accuracy, recall and precision.

[0065] Step S32: Perform path traversal processing on the integrated tree model, convert the model decision path into a business-understandable rule expression, and complete rule compression and deduplication to obtain an independent decision path set; It should be noted that path traversal processing refers to the process of completely scanning and extracting all decision links from the root node to the leaf node within the ensemble tree model; translating into business-understandable rule expressions refers to converting machine node logic into rule statements composed of features, decision operators, and thresholds that can be directly read by business personnel; rule compression refers to the simplification and merging of paths with inclusion or redundancy relationships; rule deduplication refers to eliminating completely duplicated or highly similar decision paths; and the set of independent decision paths refers to the set of normalized rule paths that are mutually exclusive and non-repeating after compression and deduplication.

[0066] The system transforms black-box models into interpretable rules through path traversal, escaping, compression, and deduplication, improving business understandability and feasibility. In one possible implementation, path traversal can be implemented using a depth-first search approach, and rule escaping is completed based on feature attribute information.

[0067] Specifically, the system employs a depth-first traversal approach to perform a global scan of the fully fitted ensemble tree model, extracting all decision paths from the root node to the leaf node and forming an initial path set. Based on feature attribute information, each machine decision path is converted into a business-readable rule expression. Then, redundancy removal, duplicate path deletion, and inclusion relationship compression are performed on the rule expressions to eliminate redundancy and overlap, ultimately resulting in a set of independent decision paths with a clear structure and logical independence. (See reference...) Figure 8 .

[0068] For example, the system performs a depth-first traversal on the trained LightGBM model, extracts all decision paths, converts path nodes into rule statements such as "account balance is greater than the threshold and operating years are greater than the threshold", compresses paths with inclusion relationships and removes duplicates, and obtains a set of independent decision paths.

[0069] Step S33: Perform multi-dimensional quantitative evaluation on the set of independent decision paths, and complete rule screening and rule combination according to preset evaluation indicators to obtain a recommended rule set.

[0070] It should be noted that multi-dimensional quantitative evaluation refers to the process of numerically evaluating the effectiveness of rules using multiple indicators such as support, confidence, lift, recall, and generalization stability. Preset evaluation indicators refer to pre-defined thresholds and standards used to judge the quality of rules. Rule screening refers to eliminating inferior rules with low accuracy, weak generalization ability, and poor stability based on evaluation indicators. Rule combination refers to combining multiple high-quality independent rules in parallel using a logical OR method to improve coverage while ensuring accuracy. The recommended rule set refers to a high-quality set of rules that can be directly used for business customer acquisition after evaluation, screening, and combination. Through multi-dimensional evaluation and rule combination, the system can automatically filter high-quality rules, eliminating reliance on manual experience and improving rule coverage and effectiveness. In one possible implementation, evaluation indicators may include rule support, confidence, lift, recall, and confidence fluctuation; rule combination can be implemented using a logical OR parallel method.

[0071] Specifically, the system calculates multi-dimensional indicators for each independent decision path, including but not limited to rule antecedent support, rule consequent support, rule confidence, rule lift, rule recall, and rule confidence fluctuation. Based on preset indicator thresholds, rules with weak generalization ability, insufficient accuracy, and poor stability are removed, while a set of single recommendation rules with high accuracy, strong universality, and strong generalization ability is retained. Then, rules are combined according to preset rule combination conditions to expand coverage while maintaining high accuracy, ultimately forming a set of recommendation rules with a balance between accuracy and coverage. The preset rule combination conditions include: For the recommendation rule set, based on the metric screening, all independent rules with a precision greater than the minimum precision and a coverage greater than the minimum coverage are obtained, filtering out noise and low-quality rules. Then, the remaining independent rules are paired up, and based on the structure screening, independent rule pairs containing too many overlapping features are removed, as duplicate features can easily lead to overlapping hit samples. Independent rule pairs are connected by the logical connector OR, and independent rules with high accuracy but low coverage are transformed into combined rules with both high accuracy and high coverage, thus obtaining candidate rule combinations. A recommended set of rule combinations is obtained by filtering based on the combined benefit of the rules and the similarity of the rules.

[0072] For example, the system calculates indicators such as support, confidence, and lift for independent decision paths, filters out low-quality rules, and performs an OR combination of the two rules "high balance, low coverage" and "high number of transactions, low coverage" to obtain a set of recommendation rules that are both accurate and have broader coverage.

[0073] In one embodiment, regarding the basic evaluation indicators of the rule evaluation system, let the antecedent of the rule be X, the consequent of the rule be Y, and the total number of samples be N; To satisfy the sample set of the antecedent X, The set of samples that satisfy the consequent Y (positive sample); For a sample set that simultaneously satisfies X and Y, the calculation methods for rule evaluation indicators include: 1) Rule Support: The support of rule X→Y is defined as the support of itemset X. The support of Y represents the probability that both X and Y are satisfied in the data, denoted as support(X→Y):

[0074] 2) Rule Antecedent Support (Rule Coverage): The support of rule antecedent X represents the proportion of data that satisfies rule antecedent X, denoted as support(X):

[0075] 3) Rule Consequence Support: The support of rule conclusion Y represents the probability that the data satisfies rule conclusion Y, denoted as support(Y). In the intelligent rule module of the strategy analysis platform, the rule consequent is fixed as label = positive, and the rule consequent support is also the proportion of positive samples:

[0076] 4) Rule Confidence (Rule Precision): The confidence of rule X→Y is the conditional probability that Y is satisfied given the basic conditions, denoted as Confidence(X→Y). A higher rule precision indicates a stronger correlation between the rule's antecedents and its conclusion.

[0077] 5) Rule Lift: This represents the ratio of the probability that the antecedent and consequent of a rule occur simultaneously to the probability that they occur independently. `lift=1`: indicates that X and Y are independent; `lift>1`: indicates a valid strong association rule; `lift<1`: indicates an invalid strong association rule.

[0078] 6) Rule Recall: The recall of rule X→Y is the conditional probability that the antecedent X is satisfied given that the consequent Y is satisfied. It is denoted as Recall(X→Y). Higher rule precision indicates a stronger correlation between the antecedent and the conclusion of the rule.

[0079] 7) Rule F1Score:

[0080] 8) Rule confidence fluctuation: The absolute value of the difference between the confidence scores of a rule on the validation set and the training set. The smaller the value, the stronger the rule's generalization ability.

[0081] Furthermore, regarding the rule combination evaluation index of the rule evaluation system, for two independent rules, we define the independent rules as R1 and R2, with coverage rates of Cover(R1) and Cover(R2), and precision rates of Precision(R1) and Precision(R2); and rule support rates of Support(R1) and Support(R2). We define the rule combination as Rcomb, with coverage rate of Cover(comb), precision rate of Precision(comb), and support rate of Support(comb). 1) Rule similarity (Jaccard coefficient):

[0082]

[0083]

[0084] 2) Positive Sample Profit Calculation Formula – △Pos Pos Positive Sample Increment: Represents how many positive samples the rule combination captures compared to independent rules.

[0085] 3) Rule combination payoff (positive-negative sample exchange ratio) – △Neg negative sample increment: indicates how many more negative samples the rule combination captures compared to independent rules.

[0086]

[0087] Sbest is the rule with the highest support among the positive samples.

[0088] Therefore, the combined returns of the rules are: or

[0089] in To prevent the default small quantity where the denominator is zero.

[0090] This embodiment obtains a highly fitted ensemble tree model by conducting distributed model training and hyperparameter optimization based on a standard dataset. Then, it obtains standardized independent decision paths by traversing the model execution path, escaping rules, compressing and deduplicating rules. Furthermore, it forms a high-quality recommendation rule set through multi-dimensional quantitative evaluation, rule screening and logical combination. This enables the automatic mining of interpretable, highly accurate and widely covered business rules from high-dimensional data, significantly reducing the dependence of rule development on human expert experience, shortening the rule generation cycle and improving the depth and breadth of customer group mining.

[0091] In one feasible implementation, the path traversal processing performed on the ensemble tree model converts the model decision paths into business-understandable rule expressions and performs rule compression and deduplication to obtain a set of independent decision paths, including: Step S41: Perform a global path scan on the ensemble tree model using a depth-first traversal method to extract all decision paths within the ensemble tree model and form an initial path set; It should be noted that depth-first traversal refers to a tree structure scanning method that starts from the root node of the ensemble tree model, prioritizes traversing to the leaf nodes, and then backtracks to other branches; global path scanning refers to the process of completely traversing and extracting all decision links from the root node to the leaf node in the ensemble tree model; all decision paths within the ensemble tree model refer to all node decision links traversed by the model when classifying samples; the initial path set refers to the original set of decision paths that have not been compressed, deduplicated, or escaped. The system can extract all decision paths completely and without omission through depth-first traversal, ensuring the comprehensiveness of rule mining. In one possible implementation, global path scanning can be implemented using a depth-first traversal algorithm based on a priority strategy.

[0092] Specifically, the system uses the trained ensemble tree model as the processing object, and adopts a depth-first traversal approach to start a global path scan from the root node. It then traverses all branches and nodes of each decision tree in turn, completely collecting all decision judgment links in the model. All the collected original decision paths are summarized to form an initial path set containing all decision paths, providing a complete data foundation for subsequent rule escaping and path optimization.

[0093] For example, the system uses a depth-first traversal approach to perform a global path scan on the LightGBM ensemble tree model trained for corporate marketing scenarios. It sequentially traverses the root node, intermediate nodes, and leaf nodes of each decision tree to extract all decision links used for customer classification and summarizes all the original paths to obtain the initial path set.

[0094] It should be noted that the optimization of the model decision path in this application is divided into two independent processing logics: the first layer is single-path condition compression, which only simplifies multiple redundant constraints generated by the same feature within a single decision path; the second layer is cross-path rule selection, which performs deduplication, inclusion relationship determination, and merit selection on multiple complete rules after internal compression. Step S42 performs the first layer of single-path condition compression, and step S43 performs the second layer of cross-path rule selection. The two processing logics are independent of each other and are not mixed.

[0095] Step S42: Based on the preset feature attribute information, each initial decision path is converted into a business-understandable rule expression, completing the escaping process from machine logic to business logic; It should be noted that the preset feature attribute information refers to the pre-configured feature name, feature type, feature meaning, and other related attribute information; the initial decision path refers to the original judgment link composed of machine nodes and values ​​in the initial path set; the business-understandable rule expression refers to the rule statement composed of business terminology, logical operators, and thresholds, which can be directly read and understood by business personnel; the machine logic to business logic escaping process refers to the process of converting the node expressions inside the model into business-readable rules. The system can eliminate the black-box nature of the model through rule escaping, making the decision logic interpretable. In one possible implementation, the feature attribute information may include feature ID, Chinese name of the feature, feature data type, etc.

[0096] Specifically, a depth-first traversal approach is used to perform a global path scan on the ensemble tree model, extracting all decision paths within the model and forming an initial path set. Based on preset feature attribute information, each initial decision path is converted into a business-understandable rule expression, completing the escaping process from machine logic to business logic. Intra-path condition compression is performed individually on each rule expression to obtain a preliminary simplified path set. Intra-path condition compression only applies to scenarios where multiple judgment constraints exist for the same feature within a single path, uniformly retaining conditions with stricter constraints. The specific compression rules are as follows: Numerical features: If both f1>10 and f1>100 constraints exist in the same path, retain the one with the smaller interval range, f1>100. Categorical features: If both f2∈[1,2] and f2∈[1] exist in the same path, retain f2∈[1] which has a smaller range of values; Date-based features: If multiple time boundary constraints exist for the same path, the condition with the narrower time constraint interval is retained. For example, the original machine decision path contains the conditions f1>100, f2=[1,2], f1>10, and f2=[1]. After compressing the conditions within the path and deleting the loose and redundant conditions, a simplified single-path rule with f1>100 and f2=[1] is obtained.

[0097] Step S43: Redundancy merging and duplication removal are performed on each rule expression to compress the initial rule path containing the correlation relationship, resulting in a set of independent decision paths.

[0098] It should be noted that redundancy merging refers to merging and simplifying rule paths with inclusion or subordinate relationships; duplication removal refers to deleting identical or highly consistent duplicate rule expressions; initial rule paths with inclusion relationships refer to rule paths where multiple paths have relationships such as conditional inclusion or logical coverage; and the set of independent decision paths refers to a set of standardized rule paths that, after merging, removal, and compression, are mutually exclusive, non-duplicative, and structurally simple. The system can reduce the number of rules and improve rule usage efficiency and readability through path compression and deduplication. In one possible implementation, the relationship can be determined by the rule condition coverage range.

[0099] Specifically, a cross-path rule-based optimization selection process is performed on the initial simplified path set after all intra-path compression has been completed, resulting in an independent decision path set. The cross-path rule-based optimization selection process is performed in two sub-steps: The first step is to standardize and regularize all the simplified rule expressions, directly eliminating rule statements that are completely duplicated. The second step is to identify the inclusion and coverage relationships between multiple independent and complete rules. For example, there are two independent rules: rule one is that the average daily balance of the main account is greater than 100,000 yuan, and rule two is that the average daily balance of the main account is greater than 100,000 yuan and the operating period is greater than 3 years. This application does not rely on simple inclusion relationships to directly delete any rule, but instead calculates the confidence, coverage, lift, and recall of the two rules separately, and retains the rule with the better comprehensive evaluation index based on the preset evaluation index threshold.

[0100] After the above deduplication and inclusion relationship selection, a set of independent decision paths with no duplication, no invalid redundancy, and logical independence is finally formed.

[0101] This embodiment extracts all decision paths of the ensemble tree model through depth-first traversal, then transforms the machine paths into business-understandable rules based on feature attribute information, and finally performs redundancy merging, duplicate removal, and association compression on the rules. This can transform the black-box logic of the model into clear, standardized, and non-redundant interpretable rules, significantly improving the business readability and usage efficiency of the rules.

[0102] In one feasible implementation, the step of threshold tuning of the recommendation rule set according to the target business scenario to obtain tuning rules that satisfy the target business scenario includes: Step S51: Determine the rule optimization target based on the customer scale requirements of the target business scenario, and construct a rule search space based on the business target in the target business scenario; It should be noted that the customer base size requirement refers to the range of target customers expected to be selected in the business scenario; the rule optimization goal refers to adjusting the rule coverage or accuracy to an optimization direction that matches business needs; and the rule search space refers to the adjustable numerical range or interval defined for the rule threshold. By clearly defining the optimization goal and constructing a reasonable search space, the system can ensure that the threshold optimization process is efficient and closely aligned with actual business needs. In one possible implementation, the rule optimization goal can be determined by converting the expected customer base size into target coverage.

[0103] Specifically, the system first converts the expected customer base size given in the target business scenario into corresponding rule coverage or accuracy requirements, thus forming a clear rule optimization goal. Then, combining the business goal and feature distribution, it defines reasonable upper and lower limits for the thresholds in each recommendation rule, constructing a complete and searchable rule threshold search space. This provides a legitimate and valid range of values ​​for subsequent iterative optimization. For example, in a corporate product marketing business scenario, the business expects to select approximately 1000 target customers. Based on this, the system sets the rule optimization goal to achieve a coverage rate within the corresponding range, and constructs a reasonable numerical search space for each rule threshold based on the distribution of features such as account balance and number of transactions.

[0104] Step S52: Use a rule optimization algorithm to iteratively optimize the set of recommendation rules in the search space to determine the optimal combination of thresholds that makes the rule effect match the business objective; It should be noted that the rule optimization algorithm refers to an iterative optimization algorithm used to quickly locate the optimal combination within the threshold search space; iterative optimization refers to the process of continuously approaching the optimal solution through multiple rounds of sampling, verification, and comparison; rule effect refers to the quantitative performance of the rule, such as coverage, precision, and confidence; business objective refers to the expected indicators such as customer base size and accuracy required by the business scenario; and the optimal threshold combination refers to a set of threshold values ​​that best matches the rule effect with the business objective. The system can automatically find the optimal threshold through iterative optimization, avoiding the inefficiency and bias caused by manual trial and error. In one possible implementation, the rule optimization algorithm can be implemented using the TPE algorithm.

[0105] Specifically, the system uses the constructed rule search space as the value range, and performs multiple rounds of iterative sampling using a rule optimization algorithm. In each round, it generates corresponding threshold combinations and verifies the rule effects. It compares the rule coverage and accuracy with the business objectives, continuously adjusts the search direction based on the deviation, and finally selects the optimal threshold combination that makes the rule effects perfectly match the business objectives.

[0106] For example, the system uses the TPE rule tuning algorithm to perform multiple rounds of iterative optimization within the constructed threshold search space, continuously adjusting the thresholds of features such as account balance, number of transactions, and years of operation, and finally obtaining the optimal combination of thresholds that ensures both the number of rules selected and the accuracy meet the business objectives.

[0107] In one embodiment, based on the number of customers in the specified list (1000), an optimization target is generated to automatically define the search space. Then, based on the TPE (Tree-structured Parzen estimator) rule optimization algorithm, the optimal combination of threshold parameters is found to obtain the optimization rules.

[0108] The first step is to convert the number of optimized lists into target coverage. The second step is to compare the target coverage with the rule coverage to determine the search direction, for example: Target coverage > rule coverage: Improve coverage; Target coverage < rule coverage: Improves accuracy; The third step is to define the search space based on the feature type, for example: Numerical type: Within the search direction, samples are taken from a semi-normal distribution with mean as the threshold and standard deviation as abs(extreme value - threshold) / 1.96. 95% of the values ​​will fall between [extreme value, threshold]. Date type: Within the search direction, samples are taken from a semi-normal distribution with mean as the threshold and standard deviation as abs(extreme value - threshold) / 1.96. 95% of the values ​​will fall between [extreme value, threshold]. Discrete type: within the search direction, permutations and combinations with values ​​fluctuating up and down by 5 based on a threshold. Fourth, search for the best combination and filter according to the following rules: Improve coverage: Coverage ≥ current rule coverage, and coverage ≥ target coverage; Increase confidence: Confidence ≥ current rule confidence, coverage ≤ current rule coverage, and the deviation between coverage and target coverage is less than a preset threshold.

[0109] The fifth step is to obtain the optimal combination of thresholds that ensures both the number of rules selected and their accuracy meet the business objectives.

[0110] As an optional embodiment, this application can configure a quantized optimization objective function when using the TPE algorithm for iterative optimization. This objective function is used to uniformly quantify and evaluate the comprehensive business adaptability of each threshold combination. The objective function expression is as follows:

[0111] in: The rule confidence level corresponding to the current threshold combination represents the accuracy of customer group screening. This represents the current rule coverage. The target coverage rate is calculated based on the expected customer base size. , , Customizable non-negative weighting coefficients for business operations: To prioritize the confidence level of the adjustment rules during the optimization process, the weighting is adjusted based on accuracy. To determine the weight of rewards for achieving coverage targets, the incentive intensity is adjusted to reach the target target size; The penalty weight for scale deviation is used to adjust the severity of the penalty for customer group size deviating from business objectives. This ensures that the TPE algorithm, when iteratively sampling and validating each candidate threshold combination, optimizes by maximizing the value of the objective function, automatically selecting the optimal threshold combination that balances screening accuracy and target object size constraints.

[0112] Step S53: Adjust and optimize the recommendation rule set according to the optimal threshold combination to obtain optimized rules that meet the requirements of the business scenario.

[0113] It should be noted that the optimal threshold combination refers to a set of threshold values ​​obtained through iterative optimization that maximizes the rule's effectiveness; adjustment and optimization refers to the process of replacing the original thresholds in the recommendation rules with the optimal thresholds; and rule tuning refers to the final usable rule whose coverage, accuracy, and customer base size all meet the requirements of the business scenario after threshold optimization. The system can quickly complete rule optimization through threshold replacement, enabling the rules to directly meet the requirements for business deployment.

[0114] Specifically, the system iteratively optimizes the optimal threshold combination and replaces the corresponding feature thresholds in the recommendation rules one by one, completing the update and optimization of all rules. This ensures that the judgment conditions of the rules fully meet the needs of the business scenario, ultimately forming optimized rules that can be directly used for customer group selection and list generation.

[0115] For example, the system replaces the corresponding thresholds in the original recommendation rules with the account balance threshold, operating years threshold, and number of transactions threshold in the optimal threshold combination, thus completing the rule adjustment and optimization and obtaining optimized rules that can be directly used in corporate marketing scenarios.

[0116] This embodiment determines the optimization target and constructs a threshold search space based on the customer base size requirements of the business scenario. It uses a rule optimization algorithm to perform iterative optimization to obtain the optimal threshold combination, and then completes rule adjustment and optimization based on the optimal threshold. This can automatically and accurately complete the rule threshold optimization, making the rule effect highly match the business goal, significantly shortening the rule optimization cycle, reducing the cost of manual parameter tuning, and improving the practicality and production efficiency of the rules.

[0117] In one feasible implementation, the step of employing a rule optimization algorithm to iteratively optimize the set of recommendation rules within the search space to determine the optimal combination of thresholds that makes the rule effects match the business objectives includes: Step S61: For any recommendation rule in the set of recommendation rules, compare the rule filtering effect of the recommendation rule with the deviation of the business objective, and determine the adjustment trend and floating range corresponding to the rule threshold in the recommendation rule; It should be noted that the recommendation rule refers to the rule to be optimized after model mining, evaluation, and screening; the rule screening effect refers to the coverage, precision, confidence, and other performance of the recommendation rule at the current threshold; the business objective refers to the expected indicators such as customer base size and accuracy required by the target business scenario; the deviation refers to the gap between the current screening effect of the rule and the business objective; the rule threshold refers to the critical value used for judgment in the rule; the adjustment direction refers to the direction in which the threshold is raised or lowered to reduce the deviation; and the floating range refers to the upper and lower limits of the threshold that can be legally adjusted. The system determines the adjustment direction and range by comparing the effect deviation, which can ensure that the optimization process does not deviate from the business objective and improve optimization efficiency. In one possible implementation, the adjustment direction can be determined as either increasing coverage or increasing precision based on the relationship between the target coverage and the current coverage.

[0118] Specifically, the system reads each recommendation rule in the recommendation rule set one by one, calculates the actual filtering effect of the rule at the current threshold, compares it with the expected effect set by the business objective, determines the magnitude and type of the deviation between the two, determines the adjustment direction of the threshold to be adjusted upward or downward based on the type of deviation, and defines a reasonable and searchable floating range for the threshold in combination with the feature value range to ensure that subsequent threshold adjustments are carried out within the effective range.

[0119] For example, when the system detects that the current rule coverage rate is lower than the coverage rate required by the business target in the recommendation rules for corporate marketing scenarios, it determines that the adjustment direction is to expand the coverage range, and defines the threshold fluctuation range as a reasonable value range based on the value distribution of the main account's average daily balance characteristics.

[0120] Step S62: Combining the adjustment direction and the customer group characteristic distribution pattern reflected in the standard dataset, extract multiple sets of candidate thresholds within the floating range to form a combined threshold sample that can be compared. It should be noted that the customer group characteristic distribution pattern refers to the statistical regularities such as the mean, standard deviation, extreme values, and distribution pattern of each feature value in the standard dataset; the candidate threshold refers to the threshold to be verified extracted within a floating range according to a certain strategy; and the combined threshold sample refers to a combination of thresholds composed of candidate thresholds from multiple rules, which can be used for effect comparison. By combining feature distribution with sample extraction, the system can ensure the rationality and representativeness of the candidate thresholds and avoid invalid sampling. In one possible implementation, the candidate threshold can be extracted based on a semi-normal distribution, discrete value floating, or other methods.

[0121] Specifically, based on the determined adjustment trend and combined with the distribution patterns of customer characteristics such as mean, extreme values, and dispersion reflected in the standard dataset, the system generates multiple sets of different candidate thresholds within the defined floating range according to the preset sampling strategy. The candidate thresholds corresponding to each rule are combined to form multiple sets of combined threshold samples that can be verified one by one and compared with each other, providing sufficient test objects for subsequent effect verification.

[0122] For example, based on the distribution patterns of features such as the average daily balance of the main account and the number of transactions in the past 12 months, the system extracts multiple sets of candidate thresholds in a semi-normal distribution within the corresponding floating range. The candidate thresholds with different rules are combined to form multiple sets of combined threshold samples that can be used for comparison and verification.

[0123] Step S63: Verify the judgment effectiveness of each of the combined threshold samples one by one, and select the optimal threshold combination that fits the business objective.

[0124] It should be noted that "judgment effectiveness" refers to the actual results obtained after applying each set of combined threshold samples to the recommendation rules, such as coverage, precision, confidence, and positive sample capture; "alignment with the business objectives" means that the judgment effectiveness meets the pre-set requirements for customer base size, accuracy, and stability; and "optimal threshold combination" refers to the set of threshold samples that best aligns with the business objectives and exhibits the best overall performance. The system can automatically determine the optimal solution through item-by-item verification and selection, improving the accuracy and objectivity of rule tuning. In one possible implementation, the verification process can be completed based on a standard dataset or a verification dataset.

[0125] Specifically, the system substitutes each set of combined threshold samples into the recommendation rules, runs and calculates the corresponding coverage, precision, confidence fluctuation and other performance indicators, compares each indicator with the business objectives, retains the combinations that meet the requirements according to the preset screening conditions, and then selects the set with the best overall effect and the smallest deviation as the optimal threshold combination that finally adapts to the business objectives.

[0126] For example, the system verifies the effectiveness of each set of threshold samples, calculates indicators such as coverage, precision, and positive sample returns, and selects the set of thresholds that meets the coverage target and has the best precision, which is then determined as the optimal threshold set that fits the corporate marketing business objectives.

[0127] This embodiment determines the threshold adjustment trend and fluctuation range by comparing the deviation between the rule effect and the business objective. It generates candidate threshold combinations by sampling based on the feature distribution law, and then verifies and judges the effectiveness of each one and selects the best. This can realize the automation, accuracy and high efficiency of rule threshold optimization, significantly reduce the cost of manual parameter adjustment, make the rule effect stable and fit the business needs, and improve the accuracy and practicality of customer group mining.

[0128] In one feasible implementation, the step of performing rule matching and list generation on the full customer data using the optimization rules to obtain the target object list corresponding to the target business scenario includes: Step S71: Set basic entry conditions based on the target business scenario; It should be noted that the basic entry criteria refer to pre-set screening conditions based on business compliance requirements, business access standards, or operational restrictions, used for initial filtering of the entire customer base. By setting basic entry criteria, the system can preemptively eliminate customer groups that clearly do not meet business requirements, reducing the computational workload of subsequent rule matching and improving overall screening efficiency and business compliance. In one possible implementation, the basic entry criteria may include operating years, account status, customer type, compliance whitelist / blacklist restrictions, etc.

[0129] Specifically, the system configures one or more basic screening conditions based on the business attributes, access requirements, and operational strategies of the target business scenario, forming a unified pre-access standard. This standard is used to quickly and coarsely filter the entire customer base before formal rule matching, ensuring that all customers entering the precise matching stage meet the minimum access requirements. For example, for the business scenario of marketing products to corporate clients, the system sets the basic access conditions to corporate clients with more than one year of operation, normal account status, not on the business blacklist, and with normal transaction behavior, thereby completing the initial access filtering of the entire customer base.

[0130] Step S72: Based on the optimization rules and the basic entry conditions, perform double screening on the full customer data to obtain an initial customer list; It should be noted that the dual screening refers to a two-layer screening mechanism: first, preliminary filtering based on basic entry criteria, and then precise matching through optimized rules. The initial customer list refers to the raw, unsorted list of matched customers obtained after the two layers of screening. The system uses dual screening to balance screening efficiency and accuracy, ensuring that the final customer group meets both business access requirements and the precise requirements of the rules. In one possible implementation, dual screening can be implemented using PySpark distributed asynchronous scheduling to support high-throughput processing of large datasets.

[0131] Specifically, the system first filters all customer data through basic entry criteria, eliminating those that do not meet the requirements. Then, it matches the remaining customer groups against the optimization rules one by one, and summarizes the customer groups that meet both the basic entry criteria and the optimization rules to form an initial customer list.

[0132] For example, the system first filters all corporate clients based on basic criteria such as years of operation and account status, and then uses optimized rules to accurately match the remaining clients. Clients that pass both screenings are then aggregated to obtain the initial client list.

[0133] Step S73: Organize, sort, and render the initial customer group list to form the target object list corresponding to the target business scenario.

[0134] It should be noted that sorting and organizing refers to ranking and standardizing the initial customer group list according to dimensions such as customer level, contribution, hit priority, and business value; rendering refers to converting the customer group data into a business-readable, displayable, and exportable visual format or standardized report; and the target list refers to the standardized customer group list that can be directly used for business deployment, operational execution, and list export. Through sorting, organizing, and rendering, the system can improve the readability, usability, and business implementation efficiency of the list. In one possible implementation, the sorting dimensions can be customized according to customer contribution, rule hit priority, customer risk level, etc.

[0135] Specifically, the system deduplicates and normalizes the initial customer list, sorts the customer groups according to preset business dimensions, and then performs visualization rendering and format standardization on the sorted customer data to generate a target object list that is clearly structured, complete in fields, and ready to use.

[0136] For example, the system deduplicates the initial customer list, sorts customers by their overall business contribution from high to low, and visualizes information such as customer name, customer number, years of operation, account balance, and hit rules to ultimately generate a list of corporate business target objects that can be directly used for marketing and promotion.

[0137] This embodiment sets basic entry conditions based on business scenarios, performs double screening on the entire customer base by combining optimization rules, and then organizes, sorts, and renders the screening results. This enables the rapid, stable, and compliant generation of a high-quality target object list, shortens the cycle from rules to business implementation, improves the efficiency and accuracy of customer selection, and ensures that the list has good readability and executability.

[0138] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0139] This application also provides a target object excavation device, please refer to... Figure 9 The target object excavation device includes: The acquisition module 91 is used to acquire a standard dataset under the target business scenario. The standard dataset is obtained by associating positive and negative sample customer group data and a set of effective features under the target business scenario. The escaping module 92 is used to train an ensemble tree model based on the standard dataset, extract the model decision paths in the trained ensemble tree model and perform escaping processing, quantitative evaluation and screening combination to obtain a set of recommendation rules; The optimization module 93 is used to perform threshold optimization on the recommendation rule set according to the target business scenario to obtain optimization rules that meet the target business scenario; The generation module 94 is used to perform rule matching and list generation on the full customer group data using the optimization rules to obtain the target object list corresponding to the target business scenario.

[0140] The target object mining apparatus provided in this application, employing the target object mining method described in the above embodiments, can solve the technical problems in the background art. Compared with the prior art, the beneficial effects of the target object mining apparatus provided in this application are the same as those of the target object mining method described in the above embodiments, and other technical features in the target object mining apparatus are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0141] This application provides a target object mining device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the target object mining method in Embodiment 1 above.

[0142] The following is for reference. Figure 10 The diagram illustrates a structural schematic suitable for implementing the target object mining device in the embodiments of this application. The target object mining device in the embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 10 The target object excavation device shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0143] like Figure 10 As shown, the target object excavation device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory 1002 or a program loaded from a storage device 1003 into a random access memory 1004. The random access memory 1004 also stores various programs and data required for the operation of the target object excavation device. The processing unit 1001, the read-only memory 1002, and the random access memory 1004 are interconnected via a bus 1005. An input / output interface 1006 is also connected to the bus. Typically, the following systems can be connected to the input / output interface 1006: input devices 1007 including, for example, a touch screen, touchpad, keyboard, mouse, image sensor, microphone, accelerometer, gyroscope, etc.; output devices 1008 including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; storage devices 1003 including, for example, magnetic tape, hard disk, etc.; and communication devices 1009. The communication device 1009 allows the target excavation equipment to communicate wirelessly or wiredly with other devices to exchange data. Although the figure shows target excavation equipment with various systems, it should be understood that implementing or having all the systems shown is not required. More or fewer systems may be implemented alternatively.

[0144] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from read-only memory 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0145] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0146] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the target object mining methods provided by the methods described above.

[0147] The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0148] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0149] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A target object mining method characterized by comprising: include: Obtain a standard dataset for the target business scenario, which is obtained by associating positive and negative sample customer group data and a set of effective features for the target business scenario; An ensemble tree model is trained based on the standard dataset. The model decision paths in the trained ensemble tree model are extracted, escaped, quantitatively evaluated, and screened to obtain a set of recommendation rules. The recommendation rule set is threshold-adjusted according to the target business scenario to obtain the adjusted rules that satisfy the target business scenario; The optimization rules are used to match and generate a list of target objects for the target business scenario by matching the rules with the full customer data.

2. The target object mining method of claim 1, wherein, The acquisition of the standard dataset for the target business scenario includes: Based on the target business scenario, set positive and negative sample differentiation conditions, and extract and construct positive and negative sample customer group data with positive and negative labels from the business data lake based on the positive and negative sample differentiation conditions. Statistical analysis and feature filtering are performed on the feature data related to the target business scenario, and attribution analysis is used to determine the effective feature set related to the business objectives in the target business scenario. The positive and negative sample customer group data are associated and integrated with the effective feature set according to the object identifier to form the standard dataset required for rule mining.

3. The target object mining method of claim 1, wherein, The process involves training an ensemble tree model based on the standard dataset, extracting the model decision paths from the trained ensemble tree model, performing escaping, quantitative evaluation, and selection / combination, to obtain a set of recommendation rules, including: Model training is conducted based on the aforementioned standard dataset, and a fully fitted ensemble tree model is obtained through distributed hyperparameter search training. The integrated tree model is subjected to path traversal processing, the model decision path is converted into a business-understandable rule expression and rule compression and deduplication are performed to obtain an independent decision path set; A multi-dimensional quantitative evaluation is performed on the set of independent decision paths, and rule selection and combination are completed based on preset evaluation indicators to obtain a recommended rule set.

4. The target object mining method as described in claim 3, characterized in that, The path traversal processing performed on the ensemble tree model transforms the model decision paths into business-understandable rule expressions and performs rule compression and deduplication to obtain a set of independent decision paths, including: A depth-first traversal method is used to perform a global path scan on the ensemble tree model, extract all decision paths within the ensemble tree model, and form an initial path set. Based on the preset feature attribute information, each initial decision path is converted into a business-understandable rule expression, completing the escaping process from machine logic to business logic; Redundancy merging and duplication removal are performed on each rule expression to compress the initial rule path containing the correlation, resulting in a set of independent decision paths.

5. The target object mining method as described in claim 1, characterized in that, The step of performing threshold tuning on the recommendation rule set according to the target business scenario to obtain tuning rules that satisfy the target business scenario includes: The rule optimization target is determined based on the customer scale requirements of the target business scenario, and a rule search space is constructed based on the business objectives in the target business scenario; The recommendation rule set is iteratively optimized within the search space using a rule tuning algorithm to determine the optimal threshold combination that makes the rule effect match the business objective; The recommendation rule set is adjusted and optimized based on the optimal threshold combination to obtain optimized rules that meet the requirements of the business scenario.

6. The target object mining method as described in claim 5, characterized in that, The step of employing a rule optimization algorithm to iteratively optimize the set of recommendation rules within the search space to determine the optimal combination of thresholds that makes the rule effects match the business objectives includes: For any recommendation rule in the set of recommendation rules, the adjustment direction and fluctuation range corresponding to the rule threshold in the recommendation rule are determined by comparing the deviation between the rule filtering effect of the recommendation rule and the business objective. Based on the adjustment trend and the customer group characteristic distribution pattern reflected in the standard dataset, multiple candidate thresholds are extracted within the floating range to form a combined threshold sample for comparison. The effectiveness of each combined threshold sample is verified item by item, and the optimal threshold combination that fits the business objective is selected.

7. The target object mining method as described in claim 1, characterized in that, The step of performing rule matching and list generation on the full customer data using the optimization rules to obtain the target object list corresponding to the target business scenario includes: Basic entry criteria are set based on the target business scenario; Based on the optimization rules and the basic inclusion criteria, the full customer data is double-screened to obtain an initial customer list. The initial customer group list is sorted, arranged, and rendered to form the target object list corresponding to the target business scenario.

8. A target object excavation device, characterized in that, include: The acquisition module is used to acquire a standard dataset under the target business scenario. The standard dataset is obtained by associating positive and negative sample customer group data and a set of effective features under the target business scenario. The escaping module is used to train an ensemble tree model based on the standard dataset, extract the model decision paths in the trained ensemble tree model, and perform escaping, quantitative evaluation, and screening and combination to obtain a set of recommendation rules. The optimization module is used to perform threshold optimization on the recommendation rule set according to the target business scenario to obtain optimization rules that meet the target business scenario; The generation module is used to perform rule matching and list generation on the full customer data using the optimization rules to obtain the target object list corresponding to the target business scenario.

9. A target object excavation device, characterized in that, The target object mining device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the target object mining method as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the target object mining method as described in any one of claims 1 to 7.

Images