A screening method and system based on global feature-derived cross and combination rules iteration

By using a global feature-derived cross- and combined rule iteration method, the shortcomings of risk identification in logistics and credit transaction scenarios are addressed, achieving efficient and accurate customer screening and risk identification, and improving the stability and business adaptability of the rules.

CN123071071APending Publication Date: 2026-11-10WUHAN ZHIPINTANG TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611106725.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-24
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

In logistics and credit transaction scenarios, existing technologies rely on independent probability models that ignore the joint distribution of events, and traditional rules are insufficient to uncover combined risks and lack stability, resulting in limited risk identification coverage and insufficient rule generalization ability.

Method used

By using feature derivation and cross-validation and rule iteration, a global feature combination space is constructed to generate a candidate rule set. Then, a final rule base is formed through temporal validation, and customer screening is completed by combining parallel matching.

Benefits of technology

It has achieved systematic modeling of high-dimensional joint feature space, improved the accuracy of risk identification and the stability of rules, adapted to actual business needs, and connected the model with business applications.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123071071A_ABST
    Figure CN123071071A_ABST
Patent Text Reader

Abstract

This invention discloses a screening method and system based on iterative cross-combination rules derived from global features, belonging to the field of logistics or credit transaction risk control technology. The method includes: collecting historical sample data, training a prediction model and outputting a score; discretizing and binning the original features and predicted scores to obtain derived features, and combining the original features and derived features in multiple orders to construct a global feature combination space. Using the target event as a negative sample, the feature space is traversed, and a candidate rule set is selected based on the purity and support of the negative samples. The training set is divided into K folds according to time sequence, and after multiple rounds of iterative verification, rules that might misjudge positive samples are eliminated, and the effective rules from each round are integrated to obtain the final rule base. Features and scores of the objects to be screened are extracted and matched in parallel with the rule base; objects that match the rules are eliminated, while those that do not match are retained normally, and finally, the screening results are output. This provides reliable technical support for reducing business risks.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of logistics or credit transaction risk control technology, and more specifically, relates to a screening method and system based on global feature-derived cross and combination rule iteration. Background Technology

[0002] In logistics and credit transaction scenarios, customer behavior is mainly reflected in two core events: receipt confirmation and payment collection. The simultaneous occurrence of these two events is crucial for the smooth operation of the business. Throughout the business process, the core optimization goal is to increase the probability of customers confirming receipts and making timely payments, thereby ensuring business revenue and reducing operational risks.

[0003] Currently, the industry mainly uses two mainstream approaches for risk assessment and customer screening in this scenario: independent probability models and single rules. Both of these technologies have significant shortcomings and are difficult to adapt to complex actual business needs.

[0004] When using independent probability models for analysis, the technical approach involves building predictive models for receipt and payment events separately, and then setting thresholds based on the output of a single model to screen customers. This method assumes that the two types of events are independent, splitting the joint probability into two independent probabilities for calculation. However, in real business environments, customer receipt and payment behavior are strongly correlated, and they do not satisfy the condition of independent distribution. This modeling approach severs the inherent connection between events, fails to characterize the joint distribution characteristics of the two types of events, and struggles to identify potential risk groups with good receipt status but insufficient payment ability, which can easily lead to business losses.

[0005] Traditional rule-based methods rely on single features to formulate judgment rules, while most risks in actual business operations are caused by the combined effects of multiple features, representing combined risks formed by the intersection of multiple features. Existing technologies struggle to comprehensively and systematically traverse and mine the high-dimensional feature combination space, failing to proactively discover nonlinear risk patterns hidden in the coupling relationships of multiple features. Furthermore, traditional rule generation lacks a scientific stability verification process, resulting in insufficient rule generalization ability and a tendency for judgment failures across different time periods and batches of data.

[0006] In summary, existing technologies generally suffer from several fundamental shortcomings. First, they lack the ability to systematically model high-dimensional joint feature spaces, resulting in a relatively singular dimension of feature utilization. Second, they struggle to uncover the nonlinear combination risks formed between features, limiting the scope of risk identification. Third, they lack robust rule stability verification mechanisms, leading to insufficient reliability of the generated judgment rules. Fourth, the analytical results output by machine learning models cannot be effectively transformed into implementable business rules, resulting in a disconnect between the technical model and actual business applications. Faced with increasingly complex customer behaviors and risk patterns, existing technologies can no longer meet the practical needs of enterprises for accurate risk control, efficient screening, and stable operation. The industry urgently requires new technological solutions to address these pain points. Summary of the Invention

[0007] This invention aims to address the problems in logistics and credit transaction scenarios, where independent probability models ignore the joint distribution of events, and traditional rules struggle to uncover combined risks and lack stability. By employing feature derivation and cross-validation with iterative rule verification, reliable screening rules are constructed, bridging the gap between the model and business applications, thereby achieving accurate identification and efficient screening of customer risks.

[0008] To address the aforementioned deficiencies or improvement needs of existing technologies, as a first aspect of this invention, the present invention provides a screening method based on iterative global feature-derived cross and combination rules, comprising: S1. Obtain historical sample data, train at least one target prediction model, and output the predicted score corresponding to the historical sample data; perform discretization binning on the original features of the historical sample data and the predicted score to generate derived features; based on the original features and the derived features, perform multi-level cross combinations to construct a global feature combination space; S2. Using a preset target event as the negative sample label, traverse the global feature combination space in the training set of the historical sample data; calculate the purity of the negative samples under each feature combination, and select feature combinations whose negative sample purity is greater than or equal to a preset purity threshold and whose sample support is greater than or equal to a minimum support threshold to generate a candidate rule set; the target event includes abnormal logistics customer receipt or overdue credit transaction repayment. S3. Divide the training set into K fold datasets in chronological order; perform iterative verification. In each round, use the current fold dataset as the training subset to re-execute step S2 to generate the verification rule set for the current fold, and use the remaining fold datasets as the verification subset to filter the verification rule set, removing rules that hit positive samples in the verification subset; after completing all rounds of iteration, take the intersection of the verification rule sets retained in all rounds to generate the final rule base; S4. Obtain the target object data to be filtered, extract the features and predicted scores corresponding to the final rule base; perform parallel matching between the target object data and the rules in the final rule base; if the target object data matches at least one rule in the final rule base, mark it as a removed object; if it does not match any rule, mark it as a retained object, and output the filtered target object set; the target objects are customers in logistics and credit transaction business, and the removed objects are customers with risks of abnormal receipt or overdue payment.

[0009] Further, in step S1, training at least one target prediction model and outputting the predicted score corresponding to the historical sample data includes: A signing probability prediction model and a payment probability scoring model are constructed respectively; wherein, the signing probability prediction model is constructed using the LightGBM algorithm, and the payment probability scoring model is constructed using the XGBoost algorithm; Simultaneously, the discretization and binning of the original features of the historical sample data and the predicted score to generate derived features includes: The continuous original features and the continuous prediction scores output by the receipt probability prediction model and the payment probability scoring model are discretized using equal-frequency binning, equal-distance binning, or a custom segmentation method based on prior business knowledge. The continuous values ​​are mapped to discrete interval labels, which serve as the derived features.

[0010] Furthermore, in S1, based on the original features and the derived features, multi-level cross-combinations are performed to construct a global feature combination space, including: Set an upper limit N for the combinatorial order, where N is an integer greater than or equal to 2 and less than or equal to 5; Under the constraint of the upper limit N of the combination order, the original features and the derived features are combined by Cartesian products of order 2 to N to generate a global feature combination space; In the process of generating the global feature combination space, conflicting combinations containing different binning intervals of the same feature are eliminated, as well as invalid combinations that are mutually exclusive in business logic.

[0011] Further, in step S2, traversing the global feature combination space within the training set of the historical sample data includes: Each historical sample in the training set is binned according to its feature interval and transformed into a transaction itemset containing multiple feature-interval key-value pairs to construct a transaction database. The global feature combination space is traversed and searched in the transaction database. During the search, for any feature combination currently traversed, the total number of samples in the transaction database that simultaneously contain all "feature-interval" key-value pairs in that feature combination is counted and used as the sample support. The number of samples with negative sample labels in the total number of samples is counted, and the ratio of the number of negative samples to the total number of samples is calculated and used as the negative sample purity. During the traversal search, if the sample support of the current feature combination is less than the minimum support threshold, support pruning is triggered, and the generation of new feature combinations based on the current feature combination to a higher order is stopped. The preset purity threshold is 100%, and the minimum support threshold is an integer greater than or equal to 20. When the sample support of the current feature combination is greater than or equal to the minimum support threshold, and the purity of the negative sample is equal to 100%, the current feature combination is extracted as a strong risk rule and added to the candidate rule set.

[0012] Furthermore, in step S3, the training set is divided into K fold datasets according to time order, including: Obtain the timestamp information of the historical sample data and sort the training set according to the chronological order; The sorted training set is divided into K consecutive fold datasets, where K is not less than 5.

[0013] Further, in S3, the step of filtering the validation rule set using the remaining dataset as a validation subset to remove rules that hit positive samples in the validation subset includes: In each iteration, only the historical fold dataset before the current fold is used as the training subset to generate rules, and the future fold dataset after the current fold is used as the validation subset for filtering, so as to eliminate rules that hit any positive sample in the validation subset and prevent future data leakage.

[0014] Furthermore, in step S3, the intersection of the validation rule sets retained from all rounds is taken to generate the final rule base, including: For the set of verification rules retained after each round of iteration filtering, extract the feature combination identifiers; Calculate the intersection of feature combination identifiers of the verification rule sets corresponding to all rounds of iteration, and retain only the feature combinations that exist in all verification rule sets simultaneously to generate the final rule base; The rules in the final rule base are sorted in descending order according to sample support or the number of negative samples, and each rule is assigned a unique rule identifier.

[0015] Furthermore, in step S4, the target object data is matched in parallel with the rules in the final rule base, including: The final rule base is then converted into a decision tree structure or loaded into an in-memory rule engine; Using a multi-threaded concurrent or distributed computing framework, the target object data to be filtered is sharded and then matched with the rules in the final rule base. During the matching process, a filter is used to perform a preliminary rapid screening of the feature combinations of the target object data. If the filter determines that there is no match, the detailed matching of the memory rule engine is skipped, and the filtered target object set is output.

[0016] As a second aspect of the present invention, a screening system based on global feature derivation, crossover, and combination rule iteration is also provided, comprising: The feature derivation and space construction unit is used to acquire historical sample data, train at least one target prediction model, and output the prediction score corresponding to the historical sample data; the original features of the historical sample data and the prediction score are discretized and binned to generate derived features; based on the original features and the derived features, multi-level cross combinations are performed to construct a global feature combination space. The candidate rule set filtering and generation unit is used to traverse the global feature combination space in the training set of the historical sample data with a preset target event as the negative sample label; to count the purity of the negative samples under each feature combination; to filter out the feature combinations whose negative sample purity is greater than or equal to a preset purity threshold and whose sample support is greater than or equal to a minimum support threshold, and to generate a candidate rule set; the target event includes abnormal logistics customer receipt or overdue credit transaction repayment. The temporal verification and rule filtering unit is used to divide the training set into K fold datasets in chronological order; perform iterative verification, in each round, using the current fold dataset as the training subset, re-execute the method in the candidate rule set filtering and generation unit to generate the verification rule set for the current fold, and use the remaining fold datasets as the verification subset to filter the verification rule set, removing rules that hit positive samples in the verification subset; after completing all rounds of iteration, take the intersection of the verification rule sets retained in all rounds to generate the final rule base; The rule matching and object filtering unit is used to acquire target object data to be filtered, extract features and predicted scores corresponding to the final rule base; perform parallel matching of the target object data with the rules in the final rule base; if the target object data matches at least one rule in the final rule base, it is marked as a removed object; if it does not match any rule, it is marked as a retained object, and the filtered target object set is output; the target objects are customers in logistics and credit transaction business, and the removed objects are customers with risks of abnormal receipt or overdue payment.

[0017] As a third aspect of the invention, a computer-readable storage medium is also provided, on which a computer program is stored, which is executed by a processor to perform a screening method based on global feature derivation, crossover, and combination rule iteration as described in any one of the claims.

[0018] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects: 1. This invention presents a screening method based on global feature derivation, cross-combination, and iterative rule-based selection. It trains two prediction models—receipt and payment collection—to obtain corresponding scores. These scores are then combined with the original features to perform discretization and binning to generate derived features. Furthermore, multi-level cross-combinations of the two types of features are used to construct a global feature combination space. This method transforms continuous data into discrete interval labels, mitigating the interference from extreme values. Simultaneously, it leverages multi-level feature combinations to uncover relationships that cannot be revealed by a single feature. Traditional methods rely solely on original features for analysis, resulting in insufficient utilization of data information and difficulty in capturing implicit connections between features. This method supplements feature dimensions based on model output, and through feature derivation and cross-combination, completes the systematic modeling of a high-dimensional joint feature space, fully preserving the inherent relationships within the data and laying a solid data foundation for subsequent risk identification.

[0019] 2. The screening method of this invention, based on global feature derivation, cross-fertilization, and combination rule iteration, traverses the global feature combination space using two indicators: negative sample purity and sample support. It then filters candidate rule sets according to a set threshold and uses a support pruning strategy to narrow the search range. In logistics and credit transaction scenarios, risks are often formed by the combined effects of multiple features, making effective identification difficult with a single criterion. This method uses actual risk events as negative samples, quantifies the risk orientation and statistical validity of each feature combination, and filters out accidental combinations that lack practical reference value. Simultaneously, it uses pruning techniques to control the size of the combination space, reasonably controlling computational costs while ensuring the integrity of rule screening, and extracting feature combination rules with risk orientation, thus overcoming the shortcomings of traditional rules that cannot systematically traverse the feature combination space.

[0020] 3. The screening method of this invention, based on global feature-derived cross and combination rule iteration, conducts multiple rounds of iterative verification by dividing the dataset into time series. The intersection of the retained rules from multiple rounds forms the final rule base, and then target object screening is completed using parallel matching. This method follows time logic to perform rule verification, distinguishing between training and verification data, and eliminating rules that might incorrectly identify normal samples, thereby verifying the performance of rules in different time series data. Traditional rules lack standardized verification processes, resulting in weak operational stability and generalization ability. This method relies on multiple rounds of time series verification to complete rule purification, ensuring the reliability of rules for long-term use. Simultaneously, it transforms model analysis results into standardized business rules, bridging the gap between the algorithm model and business applications, and relying on parallel matching to complete batch object screening, adapting to the operational needs of actual business. Attached Figure Description

[0021] Figure 1 This is a flowchart of a screening method based on global feature-derived cross and combination rule iteration according to an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the relationship between the predicted probability and the actual acceptance rate of the LightGBM acceptance probability model in an embodiment of the present invention. Figure 3 This is a schematic diagram of feature importance ranking in the XGBoost repayment probability scoring model of this invention. Figure 4 This is a schematic diagram of the system units in an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0023] Example 1 Please refer to Figure 1 This embodiment 1 provides a screening method based on global feature derivation, cross-fertilization, and combination rule iteration, including: S1. Obtain historical sample data, train at least one target prediction model, and output the predicted score corresponding to the historical sample data; perform discretization binning on the original features of the historical sample data and the predicted score to generate derived features; based on the original features and the derived features, perform multi-level cross combinations to construct a global feature combination space; S2. Using a preset target event as the negative sample label, traverse the global feature combination space in the training set of the historical sample data; calculate the purity of the negative samples under each feature combination, and select feature combinations whose negative sample purity is greater than or equal to a preset purity threshold and whose sample support is greater than or equal to a minimum support threshold to generate a candidate rule set; the target event includes abnormal logistics customer receipt or overdue credit transaction repayment. S3. Divide the training set into K fold datasets in chronological order; perform iterative verification. In each round, use the current fold dataset as the training subset to re-execute step S2 to generate the verification rule set for the current fold, and use the remaining fold datasets as the verification subset to filter the verification rule set, removing rules that hit positive samples in the verification subset; after completing all rounds of iteration, take the intersection of the verification rule sets retained in all rounds to generate the final rule base; S4. Obtain the target object data to be filtered, extract the features and predicted scores corresponding to the final rule base; perform parallel matching between the target object data and the rules in the final rule base; if the target object data matches at least one rule in the final rule base, mark it as a removed object; if it does not match any rule, mark it as a retained object, and output the filtered target object set; the target objects are customers in logistics and credit transaction business, and the removed objects are customers with risks of abnormal receipt or overdue payment.

[0024] This embodiment 1 further elaborates on the above steps.

[0025] (1) Feature Derivation and Spatial Construction Obtain historical sample dataset The historical sample dataset Includes The first historical sample, recording the first One historical sample is , where sample index Each historical sample Contains a set of original features The original feature set Includes The first continuous primitive feature, denoted as the... The continuous original features are , where feature index . No. Historical samples In the Continuous original features The specific values ​​taken on the eigenvalue are denoted as the original eigenvalues. Based on the aforementioned historical sample dataset Two target prediction models are trained separately. In this embodiment, the LightGBM algorithm is specifically used to construct the receipt probability prediction model. And a repayment probability scoring model constructed using the XGBoost algorithm. .

[0026] The first Historical samples Input the signed receipt probability prediction model Output the predicted receipt score for this sample. At the same time, the first Historical samples Input the repayment probability scoring model Output the corresponding repayment prediction score for this sample. .like Figure 2 As shown, the predicted acceptance rate segments and the corresponding actual average acceptance rate show a significant positive correlation trend, verifying the effectiveness of the model's predictive scoring. Figure 3 This demonstrates the importance ranking of feature selection for model scoring upgrades, with features such as "overdue amount" and "overdue days" having the highest importance values ​​(2402 and 2227 respectively), indicating that these features have a crucial impact on receipt and payment prediction. Based on this, each historical sample... All received and These are two continuous predictive rating variables. To transform the continuous black-box ratings output by the machine learning model into discrete labels that can be identified by subsequent white-box rules, and to enhance the data's robustness to extreme outliers, it is necessary to process the historical sample dataset... The continuous original features and the predicted scores are discretized and binned. A set of features to be binned is defined. The set of features to be binned Includes all continuous primitive features And two predicted scores and For the set of features to be binned... Any continuous binning feature in (in (For the index of the feature to be binned), obtain the set of values ​​for all historical samples on that feature, and set the target number of bins. According to the continuous type of bins to be separated characteristics Depending on the data distribution characteristics or business requirements, select one of the following discretization methods: equal-frequency binning, equal-distance binning, or a custom segmentation method based on prior business knowledge.

[0027] If the equal-frequency binning method is used to divide the continuous-type binning features... The feature is processed by arranging its values ​​across all historical samples in ascending order, resulting in a sorted feature value sequence. This is then applied based on the target binning quantity. Calculate the ideal number of samples that each bin should contain. Its calculation formula is ,in This represents the floor function. Based on the stated ideal sample size. The sorted feature value sequence is segmented to determine... The boundary value of the equal-frequency bin is denoted as the first... The boundary value of each equal-frequency box is Boundary index Through the aforementioned equal-frequency bin boundary values The continuous type of bin to be divided features The range of values ​​is divided into A discrete equal-frequency interval.

[0028] For data with a relatively uniform distribution, an equidistant binning method can be used to divide the continuous data into bins. Process the data. Obtain the maximum value of this feature across all historical samples. and minimum value Based on the target number of boxes. Calculate the interval step size of the equidistant bins. Its calculation formula is Based on the minimum value and the interval step size ,Sure The boundary value of the equidistant bins is denoted as the nth... The boundary value of each equidistant compartment is Its calculation formula is Through the equidistant bin boundary values The continuous bin-to-be-separated features The range of values ​​is divided into The discretization process uses discrete, equidistant intervals to preserve the relative proportions of the original values ​​in terms of physical distance. When a feature has a clear business interpretation, a custom segmentation method based on prior business knowledge is used for discretization. This directly obtains the externally input business rule configuration, which is pre-defined. The custom bin boundary value is denoted as the first... Each custom bin boundary value is Through the custom bin boundary value Divide A discrete, custom interval.

[0029] After completing any of the above-mentioned box splitting operations, for the continuous box splitting characteristics... Generate corresponding derived features Traverse the historical sample dataset. Each historical sample Determine its characteristics in the continuous type of box to be divided. The original value on (i.e.) Or predicted score , The discrete interval into which the given information falls is identified, and the corresponding interval label is assigned. Assigned to the historical sample Derivative features Repeat the above mapping process for all continuous features to be binned, generating a derived feature set containing all interval labels. The derived feature set With the original feature set Together, they form the global feature basis for subsequent multi-level cross combinations.

[0030] Based on the constructed global feature set, multi-level cross-combinations are performed to build a global feature combination space. An upper limit N for the combination order is set, where N is an integer greater than or equal to 2 and less than or equal to 5. Under the constraint of the upper limit N, the original feature set is... With derived feature set The features are merged into a global feature pool, and features are extracted from the global feature pool without replacement. The features are then combined using Cartesian products of order 2 to N to generate an initial global feature combination space. Each feature combination consists of features of the corresponding order and their associated value conditions or interval labels.

[0031] During the generation of the initial global feature combination space, a conflict combination removal operation is performed. Each feature combination in the initial global feature combination space is traversed. If a feature combination is found to contain different binning interval labels belonging to the same original feature, the feature combination is determined to be a conflict combination and removed, ensuring that the same object does not simultaneously satisfy two mutually exclusive discrete interval conditions for the same feature.

[0032] After conflict removal is completed, a business logic mutual exclusion removal operation is performed. A preset business mutual exclusion rule dictionary is obtained, which records feature condition pairs that cannot be simultaneously true in business logic. The remaining feature combinations are compared with the business mutual exclusion rule dictionary. If a feature combination contains any pair of mutually exclusive feature conditions defined in the business mutual exclusion rule dictionary, it is determined to be an invalid combination and removed, ultimately resulting in an optimized global feature combination space.

[0033] (2) Generation of candidate rule sets After completing global feature derivation and constructing the derived feature set, the candidate rule set generation stage begins. A preset target event is defined as a specific risky behavior that needs to be identified or intercepted in the business scenario. Historical samples that have occurred under the preset target event are assigned negative sample labels, while historical samples that have not occurred under the preset target event are assigned positive sample labels. First, obtain the following... Training set of historical samples , record the first One historical sample is Sample Index Each historical sample Corresponding to a containing The set of derived features of the i-th derived feature, denoted as the set of derived features of the th derived feature. The derived features are Feature Index .sample In derived features The upper has a unique interval label To transform structured tabular data into a format suitable for association rule mining, each historical sample... Mapped to a transaction item set The transaction item set Depend on The feature-interval key-value pairs constitute a set, denoted as the i-th feature-interval key-value pair. Each feature-interval key pair is And satisfy A transaction database is constructed by aggregating the transaction itemsets of all historical samples. ,satisfy .

[0034] In the transaction database When traversing the global feature combination space, any feature combination to be evaluated is defined as... The feature combination Include A set of distinct feature-interval key-value pairs, denoted as... The index of the combined inner element , Representing feature combinations The first in Each feature-range key-value pair. To measure the prevalence of this feature combination in historical data, statistics are collected from the transaction database. The combination of features is fully contained in The number of historical samples for all key-value pairs is used as the sample support. Its calculation formula is ,in For indicator functions, when yes The value is 1 when it is a subset of the subset, and 0 otherwise. In the sample support... In the corresponding historical samples, the number of samples with negative sample labels is counted and recorded as the negative sample hit count. Based on this, the feature combination is calculated. Negative sample purity Its calculation formula is .

[0035] To overcome the combinatorial explosion problem caused by the intersection of higher-order features during the traversal search process, a downward closure property is introduced for support pruning. A minimum support threshold is set. The minimum support threshold The integer must be greater than or equal to a preset positive integer lower limit to ensure that the rules have a minimum statistical sample size and avoid random pseudo-rules due to insufficient samples. During the expansion of the search tree, if the current feature combination... Sample support Less than the minimum support threshold If this occurs, the support pruning mechanism will be triggered, and the pruning process based on the current feature combination will be stopped immediately. Extend to higher orders to generate new feature combinations.

[0036] Set a preset purity threshold This means that no positive samples are allowed in the samples covered by the rule. When the current feature combination... Sample support Integers greater than or equal to 20, and whose negative sample purity Equal to the preset purity threshold At this point, the feature combination is determined to meet the criteria for a high-risk assessment. The current feature combination is then... The rules are extracted as high-risk rules and added to the candidate rule set. After the global feature combination space is traversed and searched, all collected high-risk rule objects are summarized, output, and the final candidate rule set is generated.

[0037] (3) Time-series verification and rule filtering After generating the initial candidate rule set, obtain the set containing... Training set of historical samples The training set is sorted in ascending order based on the timestamp information of the historical samples. Divided into Let the nth continuous fold dataset be denoted as . The dataset is folded index Where K is not less than 5. During the partitioning process, if there are boundary samples with the same timestamp, the boundary samples are forcibly partitioned into the same fold dataset to ensure that the first... The maximum timestamp of the folded dataset is strictly less than the first one. The minimum timestamp of the dataset is used to construct a strict temporal causal isolation boundary, preventing data from crossing the training and validation phases at the same time and causing hidden data traversal. Based on the above... Expand the dataset Let the current iteration round be indexed as the number of iterations. ,in In each iteration, the first... Folded dataset Defined as the current fold dataset, the first fold dataset... Turn to the first The union of the folded datasets is used as the historical folded dataset. and the first Turn to the first The union of folded datasets as future folded datasets .

[0038] In the In each iteration, only the historical fold dataset is used. This subset is used to perform feature derivation and rule generation operations, outputting the corresponding set of validation rules. The verification rule set Include Candidate rule, denoted as the first. The candidate rules are Rule Index Each candidate rule At the underlying data structure level, it is represented as a set containing multiple feature-range key-value pairs.

[0039] To test the generalization ability of the rules in unknown future environments and to remove rules that pose a risk of false positives, the future fold dataset was used. As a subset of the verification rule set Perform filtering. In the future fold dataset... Select all samples with positive sample labels from the data and construct a positive sample validation set. The positive sample validation set Include Let the nth positive sample be denoted as the nth positive sample. The number of positive samples is Positive sample index Each positive sample It was also transformed into a transaction itemset containing multiple feature-range key-value pairs. .

[0040] For the aforementioned verification rule set Each candidate rule in traverse the positive sample validation set Each positive sample Determine the transaction item set Does it fully include candidate rules? All feature-range key-value pairs. If there exists any positive sample transaction itemset If the above inclusion conditions are met, then the candidate rule is determined. If a positive sample is found in the validation subset, the candidate rule is immediately added. From the verification rule set Remove from the positive sample validation set. After that, candidate rules If no positive sample is found, it is retained and added to the next step. The target rule set output by round iteration .

[0041] After completing all rounds of iterative filtering, the target rule set retained after each round of iterative filtering is... The feature combination identifiers are extracted. The intersection of the feature combination identifiers for all iterations of the validation rule sets is calculated, and only feature combinations that exist in all validation rule sets are retained to generate the final rule base. Finally, the rules in the final rule base are sorted in descending order according to sample support or the number of negative samples, and each rule is assigned a unique rule identifier.

[0042] (4) Rule matching and object filtering After generating the final rule base, the next step is to evaluate the objects to be screened in the actual business scenario. First, obtain the rules containing... A set of target objects to be filtered , record the first The target objects are Object index For each target object Extract the rules corresponding to the final rule base. One derived feature, and obtain the target object. In the Derivative features The interval labels on Map it to contain Transaction itemsets with features-range key-value pairs Simultaneously extract the target object Predicted score .

[0043] To improve matching efficiency with massive amounts of data, the final rule base is transformed into a decision tree structure. Or load directly into the memory rule engine In the middle, and build a pre-filter. The pre-filter is used before the formal matching. For the target object transaction itemset Perform a rapid pre-screening; if the pre-filter... Determine the current target object If the feature combination does not match any rule's latent space, then the memory rule engine is skipped directly. The detailed matching logic reduces unnecessary computational overhead.

[0044] Based on multi-threaded concurrency mechanisms or distributed computing frameworks, the target object set... Divided into The data shard, denoted as the _th_ data shard. Each data shard is sharded index Divide the data into shards. The objects are distributed to different computing nodes or execution threads, so that the target objects in each slice are simultaneously connected to the memory rule engine. The final rule base loaded in the middle is used for parallel matching calculation.

[0045] During parallel matching computation, for the target object within the current slice Compare each of its transaction itemsets Does it completely contain all feature-range key-value pairs of any rule in the final rule base? If the target object transaction itemset If at least one rule in the final rule base is matched, the target object will be... Mark as a culling object; if the target object transaction itemset The final rule base was not hit. Any rule within it will target the object. Mark as reserved objects. Summarize all target objects marked as reserved objects and their corresponding predicted scores. It outputs the filtered target object set, completing the full-process filtering based on global feature-derived cross and combination rules iteration.

[0046] Example 2 Please refer to Figure 4 This embodiment 2 provides a screening system based on global feature derivation, cross-fertilization, and combination rule iteration, including: The feature derivation and space construction unit is used to acquire historical sample data, train at least one target prediction model, and output the prediction score corresponding to the historical sample data; the original features of the historical sample data and the prediction score are discretized and binned to generate derived features; based on the original features and the derived features, multi-level cross combinations are performed to construct a global feature combination space. The candidate rule set filtering and generation unit is used to traverse the global feature combination space in the training set of the historical sample data with a preset target event as the negative sample label; to count the purity of the negative samples under each feature combination; to filter out the feature combinations whose negative sample purity is greater than or equal to a preset purity threshold and whose sample support is greater than or equal to a minimum support threshold, and to generate a candidate rule set; the target event includes abnormal logistics customer receipt or overdue credit transaction repayment. The temporal verification and rule filtering unit is used to divide the training set into K fold datasets in chronological order; perform iterative verification, in each round, using the current fold dataset as the training subset, re-execute the method in the candidate rule set filtering and generation unit to generate the verification rule set for the current fold, and use the remaining fold datasets as the verification subset to filter the verification rule set, removing rules that hit positive samples in the verification subset; after completing all rounds of iteration, take the intersection of the verification rule sets retained in all rounds to generate the final rule base; The rule matching and object filtering unit is used to acquire target object data to be filtered, extract features and predicted scores corresponding to the final rule base; perform parallel matching of the target object data with the rules in the final rule base; if the target object data matches at least one rule in the final rule base, it is marked as a removed object; if it does not match any rule, it is marked as a retained object, and the filtered target object set is output; the target objects are customers in logistics and credit transaction business, and the removed objects are customers with risks of abnormal receipt or overdue payment.

[0047] Example 3 This embodiment 3 also provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it can implement any step of a screening method based on global feature derivation, crossover, and combination rule iteration.

[0048] The computer-readable storage medium may include various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0049] For a description of the computer-readable storage medium provided in this application, please refer to the above method embodiments; further details will not be repeated here.

[0050] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A screening method based on iterative rules of global feature derivation, crossover, and combination, characterized in that, include: S1. Obtain historical sample data, train at least one target prediction model, and output the predicted score corresponding to the historical sample data; The original features of the historical sample data and the predicted score are discretized and binned to generate derived features; based on the original features and the derived features, multi-level cross combinations are performed to construct a global feature combination space; S2. Using the preset target event as the negative sample label, traverse the global feature combination space in the training set of the historical sample data; The purity of negative samples under each feature combination is statistically analyzed, and feature combinations with negative sample purity greater than or equal to a preset purity threshold and sample support greater than or equal to a minimum support threshold are selected to generate a candidate rule set; the target events include abnormal logistics customer receipt or overdue credit transaction repayment. S3. Divide the training set into K fold datasets in chronological order; perform iterative verification. In each round, use the current fold dataset as the training subset to re-execute step S2 to generate the verification rule set for the current fold, and use the remaining fold datasets as the verification subset to filter the verification rule set, removing rules that hit positive samples in the verification subset; after completing all rounds of iteration, take the intersection of the verification rule sets retained in all rounds to generate the final rule base; S4. Obtain the target object data to be filtered, and extract the features and predicted scores corresponding to the final rule base; The target object data is matched in parallel with the rules in the final rule base; if the target object data matches at least one rule in the final rule base, it is marked as a removed object; if no rule is matched, it is marked as a retained object, and the filtered target object set is output; the target object is a customer in the logistics and credit transaction business, and the removed object is a customer with the risk of abnormal receipt or overdue payment.

2. The screening method based on global feature derivation, crossover, and combination rule iteration as described in claim 1, characterized in that, In step S1, at least one target prediction model is trained, and the predicted score corresponding to the historical sample data is output, including: A signing probability prediction model and a payment probability scoring model are constructed respectively; wherein, the signing probability prediction model is constructed using the LightGBM algorithm, and the payment probability scoring model is constructed using the XGBoost algorithm; Simultaneously, the discretization and binning of the original features of the historical sample data and the predicted score to generate derived features includes: The continuous original features and the continuous prediction scores output by the receipt probability prediction model and the payment probability scoring model are discretized using equal-frequency binning, equal-distance binning, or a custom segmentation method based on prior business knowledge. The continuous values ​​are mapped to discrete interval labels, which serve as the derived features.

3. The screening method based on global feature derivation, crossover, and combination rule iteration as described in claim 1, characterized in that, In step S1, based on the original features and the derived features, multi-level cross-combinations are performed to construct a global feature combination space, including: Set an upper limit N for the combinatorial order, where N is an integer greater than or equal to 2 and less than or equal to 5; Under the constraint of the upper limit N of the combination order, the original features and the derived features are combined by Cartesian products of order 2 to N to generate a global feature combination space; In the process of generating the global feature combination space, conflicting combinations containing different binning intervals of the same feature are eliminated, as well as invalid combinations that are mutually exclusive in business logic.

4. The screening method based on global feature derivation, crossover, and combination rule iteration as described in claim 1, characterized in that, In step S2, traversing the global feature combination space within the training set of the historical sample data includes: Each historical sample in the training set is binned according to its feature interval and transformed into a transaction itemset containing multiple feature-interval key-value pairs to construct a transaction database. The global feature combination space is traversed and searched in the transaction database. During the search, for any feature combination currently traversed, the total number of samples in the transaction database that simultaneously contain all "feature-interval" key-value pairs in that feature combination is counted and used as the sample support. The number of samples with negative sample labels in the total number of samples is counted, and the ratio of the number of negative samples to the total number of samples is calculated and used as the negative sample purity. During the traversal search, if the sample support of the current feature combination is less than the minimum support threshold, support pruning is triggered, and the generation of new feature combinations based on the current feature combination to a higher order is stopped. The preset purity threshold is 100%, and the minimum support threshold is an integer greater than or equal to 20. When the sample support of the current feature combination is greater than or equal to the minimum support threshold, and the purity of the negative sample is equal to 100%, the current feature combination is extracted as a strong risk rule and added to the candidate rule set.

5. The screening method based on global feature derivation, crossover, and combination rule iteration as described in claim 1, characterized in that, In step S3, the training set is divided into K fold datasets according to time order, including: Obtain the timestamp information of the historical sample data and sort the training set according to the chronological order; The sorted training set is divided into K consecutive fold datasets, where K is not less than 5.

6. The screening method based on global feature derivation, crossover, and combination rule iteration as described in claim 1, characterized in that, In step S3, filtering the validation rule set using the remaining dataset as a validation subset to remove rules that hit positive samples in the validation subset includes: In each iteration, only the historical fold dataset before the current fold is used as the training subset to generate rules, and the future fold dataset after the current fold is used as the validation subset for filtering, so as to eliminate rules that hit any positive sample in the validation subset and prevent future data leakage.

7. The screening method based on global feature derivation, crossover, and combination rule iteration as described in claim 1, characterized in that, In step S3, the intersection of the validation rule sets retained from all rounds is taken to generate the final rule base, including: For the set of verification rules retained after each round of iteration filtering, extract the feature combination identifiers; Calculate the intersection of feature combination identifiers of the verification rule sets corresponding to all rounds of iteration, and retain only the feature combinations that exist in all verification rule sets simultaneously to generate the final rule base; The rules in the final rule base are sorted in descending order according to sample support or the number of negative samples, and each rule is assigned a unique rule identifier.

8. The screening method based on global feature derivation, crossover, and combination rule iteration as described in claim 1, characterized in that, In step S4, the target object data is matched in parallel with the rules in the final rule base, including: The final rule base is then converted into a decision tree structure or loaded into an in-memory rule engine; Using a multi-threaded concurrent or distributed computing framework, the target object data to be filtered is sharded and then matched with the rules in the final rule base. During the matching process, a filter is used to perform a preliminary rapid screening of the feature combinations of the target object data. If the filter determines that there is no match, the detailed matching of the memory rule engine is skipped, and the filtered target object set is output.

9. A screening system based on iterative rules of global feature derivation, crossover, and combination, characterized in that, include: Feature derivation and spatial construction unit, used to acquire historical sample data, train at least one target prediction model, and output the prediction score corresponding to the historical sample data; The original features of the historical sample data and the predicted score are discretized and binned to generate derived features; based on the original features and the derived features, multi-level cross combinations are performed to construct a global feature combination space; The candidate rule set filtering and generation unit is used to traverse the global feature combination space in the training set of the historical sample data with a preset target event as the negative sample label. The purity of negative samples under each feature combination is statistically analyzed, and feature combinations with negative sample purity greater than or equal to a preset purity threshold and sample support greater than or equal to a minimum support threshold are selected to generate a candidate rule set; the target events include abnormal logistics customer receipt or overdue credit transaction repayment. The temporal verification and rule filtering unit is used to divide the training set into K fold datasets in chronological order; perform iterative verification, in each round, using the current fold dataset as the training subset, re-execute the method in the candidate rule set filtering and generation unit to generate the verification rule set for the current fold, and use the remaining fold datasets as the verification subset to filter the verification rule set, removing rules that hit positive samples in the verification subset; after completing all rounds of iteration, take the intersection of the verification rule sets retained in all rounds to generate the final rule base; The rule matching and object filtering unit is used to acquire target object data to be filtered, extract features and predicted scores corresponding to the final rule base; The target object data is matched in parallel with the rules in the final rule base; if the target object data matches at least one rule in the final rule base, it is marked as a removed object; if no rule is matched, it is marked as a retained object, and the filtered target object set is output; the target object is a customer in the logistics and credit transaction business, and the removed object is a customer with the risk of abnormal receipt or overdue payment.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program is executed by a processor as described in any one of claims 1-8, a screening method based on iterative global feature derivation, crossover, and combination rules.