Resource allocation method, apparatus, device, and storage medium

Abstract

This invention belongs to the field of artificial intelligence technology and discloses a resource allocation method, apparatus, device, and storage medium. The allocation method includes: invoking a target business rule set and a target machine learning model that match the description information; generating a first initial allocation scheme based on rules and a second initial allocation scheme based on the model for the items to be allocated based on the target business rule set and the target machine learning model; generating a root cause diagnosis report when the difference between the first initial allocation scheme and the second initial allocation scheme exceeds a preset threshold; constructing a rule parameter optimization space based on the root cause diagnosis report; and searching for target rule parameters within the rule parameter optimization space that give the preset objective function an optimal solution, generating a final allocation scheme for the items to be allocated corresponding to the rule parameter optimization suggestions. This invention can be applied to business management systems in fintech, healthcare, and other fields, solving the technical problem of inaccurate resource reallocation.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology and is applied in the fields of financial technology and healthcare. In particular, it relates to a resource allocation method, device, equipment and storage medium. Background Technology

[0002] In the financial operations and cost management of the property insurance industry, Tongxin's cost allocation is a crucial link in the rational allocation and collection of public resources such as common channel expenses, centralized operating costs, and joint marketing investments across institutions, product lines, and business units. Its rationality directly affects the accuracy of cost accounting, the effectiveness of business analysis, and the fairness of internal performance evaluation. With the continuous expansion of business scale, increasingly complex channel structures, and increasingly refined insurance product types, the traditional Tongxin cost allocation model, which relies on manual configuration and execution of fixed business rules, has gradually revealed significant limitations in actual operation.

[0003] In traditional models, the allocation of public funds and resources typically relies on pre-defined proportions, coefficients, or business definitions, resulting in a single allocation outcome. When faced with diverse data sources, varied allocation dimensions, and changing business scenarios, this approach merely executes allocation actions according to predetermined logic, lacking the ability to quantitatively assess the rationality of the allocation results. When allocation results deviate, become abnormal, or do not match actual business conditions, the system cannot truly identify the root cause of the deviation, thus failing to correct the allocation based on the root cause and achieve optimal allocation of the items to be allocated. Summary of the Invention

[0004] This invention provides a resource allocation method, apparatus, device, and storage medium, which can solve the technical problem in the prior art that resources cannot be accurately redistributed when there is a deviation in resource allocation.

[0005] In a first aspect, the present invention provides a resource allocation method, comprising: Receive description information of the items to be assigned; Invoke the pre-configured target business rule set and target machine learning model that match the description information; Based on the target business rule set and the target machine learning model, parallel computation is performed on the items to be assigned to generate a first initial allocation scheme based on rules and a second initial allocation scheme based on the model for the items to be assigned. A difference quantification analysis is performed on the first initial allocation scheme and the second initial allocation scheme. When the difference between the first initial allocation scheme and the second initial allocation scheme exceeds a preset threshold, a difference root cause diagnosis report is generated. Based on the differential root cause diagnosis report, at least one rule parameter optimization space is constructed; Within the rule parameter optimization space, find the target rule parameters that give the preset objective function the optimal solution, and generate rule parameter optimization suggestions based on the target rule parameters; Based on the optimization suggestions of the rule parameters, the first initial allocation scheme and / or the second initial allocation scheme are optimized to obtain the final allocation scheme of the item to be allocated.

[0006] In a second aspect, the present invention provides a resource allocation device, comprising: The information receiving module is used to receive the description information of the items to be allocated; The calling module is used to call the pre-configured target business rule set and target machine learning model that match the description information; The allocation module is used to perform parallel computation on the items to be allocated based on the target business rule set and the target machine learning model, respectively, to generate a first initial allocation scheme based on rules and a second initial allocation scheme based on the model for the items to be allocated. The difference analysis module is used to perform difference quantification analysis on the first initial allocation scheme and the second initial allocation scheme. When the difference between the first initial allocation scheme and the second initial allocation scheme exceeds a preset threshold, a difference root cause diagnosis report is generated. The module is used to construct at least one rule parameter optimization space based on the differential root cause diagnosis report; The suggestion generation module is used to find target rule parameters within the rule parameter optimization space that enable the preset objective function to have an optimal solution, and to generate rule parameter optimization suggestions based on the target rule parameters. An optimization module is used to optimize the first initial allocation scheme and / or the second initial allocation scheme according to the optimization suggestions of the rule parameters, so as to obtain the final allocation scheme of the item to be allocated.

[0007] Thirdly, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the resource allocation method described above.

[0008] Fourthly, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the steps of the resource allocation method described above.

[0009] It is understood that the present invention provides a basis for comparative analysis of different allocation logic results by generating two allocation schemes in parallel. The present invention generates a first initial allocation scheme (rule logic) based on the target business rule set and a second initial allocation scheme (model logic) based on the target machine learning model. These two schemes are themselves the output results of two different allocation logics. By comparing these two allocation schemes, the preliminary cause of the deviation in the allocation results can be located.

[0010] Furthermore, the present invention achieves a quantitative judgment of the degree of difference through difference quantification analysis. The present invention further performs the following on the first and second initial allocation schemes: performing difference quantification analysis on the first and second initial allocation schemes, and generating a difference root cause diagnosis report when the difference exceeds a preset threshold, thereby achieving quantitative identification and location of the source and influencing factors of the difference, and thus locating the root cause of the deviation.

[0011] Furthermore, the present invention, based on a root cause diagnosis report identifying the fundamental factors causing the deviation, ultimately formulates a final allocation scheme for the items to be allocated. This final allocation scheme corrects for the fundamental factors causing the deviation, thus enabling accurate reallocation of resources. In summary, the present invention solves the technical problem in the prior art where accurate reallocation of resources is impossible when deviations occur in resource allocation. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart illustrating a resource allocation method in one embodiment of the present invention.

[0014] Figure 2 yes Figure 1 A flowchart of step S150.

[0015] Figure 3 This is a schematic diagram of a resource allocation device in one embodiment of the present invention.

[0016] Figure 4 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention.

[0017] Figure 5 This is another structural schematic diagram of a computer device according to one embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Figure 1 A flowchart of the resource allocation method provided in the embodiments of the present invention is shown below. Figure 1 As shown, the resource allocation method provided in this embodiment of the invention includes the following steps.

[0020] Step S110: Receive the description information of the item to be assigned; Specifically, in this step, a standardized data receiving interface can be built. The description information of the items to be allocated can include four core modules: the type of resource to be allocated, the core attributes of the items to be allocated, allocation constraints, and allocation targets. Among them, the core attributes need to be quantified into calculable feature parameters (such as numerical or categorical coding) to avoid vague natural language descriptions. After receiving the information, the legality of the information is verified (verifying parameter completeness, data format correctness, and reasonableness of numerical range). If the verification passes, it is stored in a temporary database; if the verification fails, a specific error message is returned (such as "Missing core attribute of the item to be allocated: XX", "Numerical parameter XX exceeds the reasonable range") to ensure the accuracy of subsequent calculations.

[0021] As a concrete example, in the property insurance field, the item to be allocated can be "property insurance claims resources." The description information of the item to be allocated is received through the claims system interface, and can specifically include: 1. Type of resource to be allocated: claims specialist manpower, claims inspection vehicles, and damage assessment amount; 2. Core attributes of the item to be allocated (quantified): reported amount (numerical, unit: yuan), reported location (e.g., latitude and longitude), type of accident (classification code: 1=vehicle collision, 2=fire, 3=natural disaster, 4=other), reported time (timestamp, e.g., 1711234567), number of accidents of the insured (numerical, e.g., 0 times, 1 time, greater than or equal to 2 times); 3. Allocation constraints: the average number of cases handled per claims specialist per day is less than or equal to 8, the radius coverage of inspection vehicles is less than or equal to 5 kilometers, and the error of damage assessment amount is less than or equal to 10%; 4. Allocation targets: claims processing time is less than or equal to 24 hours, and customer satisfaction is greater than or equal to 90%. Upon receipt, the system verifies that the reported amount is greater than or equal to 0, the latitude and longitude are within a reasonable range, and the type of accident is within the preset code. If the verification is successful, the claim is stored in the claims business database. If the verification fails, the system will prompt "The reported amount is abnormal. Please check and resubmit".

[0022] As another concrete example, in the healthcare field, the item to be allocated can be "medical resources." The description information of the item to be allocated can be received through the hospital's HIS system interface, and can specifically include: 1. Type of resource to be allocated: hospital beds, medical staff (doctors, nurses), medical equipment (ventilators, monitors), and medicines; 2. Core attributes of the item to be allocated (quantified): patient age (numerical, unit: years), severity of illness (classification code: 1=mild, 2=severe, 3=critical), confirmed disease (classification code: 1=pneumonia, 2=hypertensive emergency, 3=stroke, etc.), admission time (timestamp), patient weight (numerical, unit: kg), and whether there is a history of underlying medical conditions (binary code: 0=none, 1=yes); 3. Allocation constraints: each bed for critically ill patients must be staffed with at least one nurse; ventilators must only be used for critically ill patients; and drug dosage must be matched with patient weight (error less than or equal to 5%); 4. Allocation goals: the time from patient admission to bed placement must be less than or equal to 30 minutes; and the response time for critically ill patients must be less than or equal to 5 minutes. Upon receipt, verification is performed: the age must be between 0 and 120 years old, and the severity of the illness must be within the preset code. If the verification is successful, the information will be stored in the hospital resource management database. If the verification fails, the message "Incorrect severity of illness code, please select 1-3" will be displayed.

[0023] Step S120: Invoke the pre-configured target business rule set and target machine learning model that match the description information; Specifically, in this step, a mapping table of "description information - business rule set - machine learning model" (stored in a relational database) can be pre-constructed. The mapping relationship is indexed by "resource type to be allocated + core attribute keywords of the item to be allocated" (e.g., resource type = claims resource, keywords = reported amount, accident type). Based on the description information received in step S110, index keywords can be extracted, and the mapping table can be queried using a query statement to match the unique target business rule set and target machine learning model. The target business rule set consists of pre-configured structured rules, containing at least three executable rules, each with clearly defined conditions and execution results. The target machine learning model is a pre-trained and deployed model. The model input is the core attributes quantified in step S110, and the output is the relevant parameters of the allocation scheme. The model has been validated using a test set (accuracy greater than or equal to 85%) and can be directly called. If no matching rule set or model is found, the system returns "No matching business rule set / model, please supplement the configuration" and terminates the current process.

[0024] As a concrete example, in the property insurance field, based on the "property insurance claims resources" description information received in step S110, index keywords can be extracted: resource type = claims resources, core attribute keywords = reported amount, claim type. The mapping relationship table is then queried to match the target business rule set (claims resource allocation rule set, configured using the Drools rule engine) and the target machine learning model (claims resource allocation model, trained based on the XGBoost algorithm, deployed as an API interface). Among these: The target business rule set (which can be executed directly) can be as follows: Rule 1: IF Reported amount ≤ 10,000 yuan AND Accident type = 1 (vehicle collision), THEN assign 1 junior claims specialist, no need to dispatch a survey vehicle, and the damage assessment amount is initially allocated at 90% of the reported amount; Rule 2: IF 10,000 yuan < Reported amount ≤ 50,000 yuan AND Accident type = 2 (fire), THEN assign 1 intermediate claims specialist, 1 survey vehicle, and the damage assessment amount is initially allocated at 85% of the reported amount; Rule 3: IF Reported amount > 50,000 yuan OR Accident type = 3 (natural disaster), THEN assign 1 senior claims specialist, 2 survey vehicles, and the damage assessment amount is initially allocated at 80% of the reported amount.

[0025] The target machine learning model takes the reported amount, number of accidents, latitude and longitude of the reported location, and time of the report as input, and outputs the claims specialist level (junior / intermediate / senior), number of vehicles inspected, and damage assessment amount as output. The model has a test accuracy rate of 88%. The output results can be obtained by calling the API interface and passing in the quantified core attributes.

[0026] As another concrete example, in the healthcare field, based on the "medical resource" description information received in step S110, index keywords can be extracted: resource type = medical resource, core attribute keywords = severity of illness, diagnosed disease type. The mapping relationship table is then queried to match the target business rule set (medical resource allocation rule set, configured using the Easy Rules rule engine) and the target machine learning model (medical resource allocation model, trained based on a neural network, deployed as an API interface). Among these: The target business rule set (which can be executed directly) can be as follows: Rule 1: IF Severity of illness = 1 (mild) AND Diagnostic disease = 1 (pneumonia), THEN allocate 1 general ward bed, 0.2 nurses (i.e., 5 mild patients share 1 nurse), no ventilator required, medications are allocated according to the standard dosage (500mg / dose for adults); Rule 2: IF Severity of illness = 2 (severe) AND Diagnostic disease = 2 (hypertensive emergency), THEN allocate 1 severe illness ward bed, 1 nurse, 1 ventilator (standby), medication dosage is calculated based on patient weight (10mg / kg); Rule 3: IF Severity of illness = 3 (critical) OR Diagnostic disease = 3 (stroke), THEN allocate 1 ICU bed, 2 nurses, 1 ventilator (continuous use), medication dosage is calculated based on patient weight (15mg / kg). The target machine learning model takes patient age, disease severity, weight, and underlying medical history as input, and outputs bed type (general / critical / ICU), number of medical staff, medical equipment configuration, and drug dosage. The model has a testing accuracy of 90%. The output results can be obtained by calling the API interface and passing in the quantified core attributes.

[0027] Step S130: Based on the target business rule set and the target machine learning model, perform parallel computation on the items to be assigned to generate a first initial allocation scheme based on rules and a second initial allocation scheme based on the model for the items to be assigned. Specifically, in this step, a multi-threaded parallel computing approach can be adopted, starting two independent threads to execute rule calculation and model calculation respectively, ensuring parallel execution without interference and improving computational efficiency. Thread 1: Loads the target business rule set, uses the description information of the items to be allocated quantified in step S110 as rule input, executes the logical judgment through the rule engine, matches the rule conditions one by one, executes the allocation logic of the corresponding rule, and outputs a structured first initial allocation scheme (including quantifiable parameters such as resource allocation type, quantity, specific allocation object, allocation priority, and constraint satisfaction status); Thread 2: Calls the target machine learning model API interface, uses the core attributes quantified in step S110 as model input, receives the model output results, and converts the output results into a structured second initial allocation scheme (the format is consistent with the first initial allocation scheme to ensure comparability in subsequent difference analysis); After the two threads complete their calculations, the two initial allocation schemes are stored in the database, and the calculation completion time is marked to ensure the traceability of the schemes. The timeout for parallel computing is set to 30 seconds. If the timeout occurs, the corresponding thread is terminated, and "Calculation timed out, please try again" is returned.

[0028] As a concrete example, in the property insurance field, two parallel threads can be started: thread 1 performs rule calculations, and thread 2 performs model calculations. Thread 1 (rule calculation): Loads the target business rule set, inputs the description information of the item to be assigned (reported amount = 30,000 yuan, accident type = 2 (fire), number of accidents = 1), and matches it with rule 2 (10,000 yuan < reported amount < 50,000 yuan AND ...) through the rule engine. (Claim Type = 2), Execution Rule Output: Assign 1 intermediate claims specialist (ID: LZ002), 1 survey vehicle (ID: CK015), loss assessment amount = 30000 × 85% = 25500 yuan, allocation priority = medium, constraint satisfaction: number of cases handled by the claims specialist per day = 5 cases (< 8 cases), survey vehicle coverage radius = 3 kilometers (< 5 kilometers), generate the first initial allocation plan (structured format: {"Resource Type":"Claim Resource","Claim Specialist":"LZ002 (Intermediate)","Survey Vehicle":"CK015","Loss Assessment Amount": 25500","Priority":"Medium","Constraint Satisfaction":"Yes"}). Thread 2 (Model Calculation): Calls the target machine learning model API interface, inputting parameters (reported amount = 30,000, number of accidents = 1, latitude and longitude of the reported location = 116.4074, 39.9042, reported time = 1711234567). The model outputs: claims specialist level = intermediate, number of survey vehicles = 1, damage assessment amount = 24,000 yuan, which is then converted into a second initial allocation scheme (structured format: {"resource type":"claims resource","claims specialist":"intermediate (to be allocated)","survey vehicle":"1 (to be dispatched)","damage assessment amount": 24,000","priority":"medium","constraint satisfied":"yes"}). Both threads complete the calculation (taking 12 seconds), and the two schemes are stored in the claims database, with the calculation time marked as 1711234579.

[0029] As another specific example, in the field of healthcare, two parallel threads can be started. Thread 1 performs rule calculation, and thread 2 performs model calculation: Thread 1 (rule calculation): Load the target business rule set, input the description information of the item to be assigned (patient age = 65 years old, severity of illness = 2 (severe illness), diagnosed disease = 2 (hypertensive emergency), weight = 70kg, with underlying medical history (1)), and match rule 2 (severity of illness = 2 AND) through the Easy Rules rule engine. (Confirmed disease type = 2), execute the rule output: allocate 1 intensive care bed (No.: CZ012), 1 nurse (No.: HS035), 1 ventilator (No.: HB008, spare), drug dosage = 70 × 10 = 700 mg / dose, allocation priority = high, constraint satisfaction: 1 nurse, ventilator used for critically ill patients, generate the first initial allocation plan (structured format: {"resource type":"medical resource","bed":"CZ012 (critical care)","nurse":"HS035","ventilator":"HB008 (spare)","drug dosage":700","priority":"high","constraint satisfied":"yes")}). Thread 2 (Model Calculation): Calls the target machine learning model API interface, inputting parameters (age=65, disease severity=2, weight=70, basic medical history=1). The model outputs: bed type=intensive care, number of nurses=1, ventilator=1 (standby), drug dosage=720mg / dose, converting it into a second initial allocation scheme (structured format: {"resource type":"medical resources","bed":"intensive care (to be allocated)","nurse":"1 (to be scheduled)","ventilator":"1 (standby)","drug dosage":720","priority":"high","constraint satisfied":"yes")}). Both threads complete the calculation (taking 15 seconds), storing both schemes in the hospital resource management database, marking the calculation time as 1711234582.

[0030] In some embodiments of the present invention, after step S130, the method further includes: Step S141: Based on the description information of the items to be assigned, classify the items to be assigned to obtain the categories corresponding to the items to be assigned; Specifically, in this step, a feature extraction module can be built. Based on the description information of the items to be assigned received in step S110, the core features that can be used for classification can be automatically extracted. The feature types are divided into numerical types (such as amount, age, weight) and classification coding types (such as accident type, disease type, and disease severity). After extraction, standardization processing is performed (numerical types are mapped to the [0,1] interval using min-max standardization, and classification coding types are converted into calculable values ​​using one-hot coding or label coding) to avoid the impact of feature magnitude differences or inconsistent types on classification accuracy. More specifically, a multi-level classification system can be pre-established based on business scenario requirements, adopting a "first-level major category + second-level minor category" structure, clarifying the feature thresholds and judgment conditions of each category, and storing the classification system in a classification configuration database to support subsequent dynamic updates; among them, the first-level major category is divided according to core resource type or business scenario, and the second-level minor category is divided according to the key attributes of the items to be assigned, ensuring that the classification granularity fits the actual business, without redundancy or omission; Furthermore, a combination of "rule-based coarse classification + K-means clustering fine classification" can be used to ensure classification accuracy and efficiency. Coarse classification can be based on preset classification rules (configured in advance by business personnel, such as "reported damage less than 5000 yuan with no injuries is classified as small-amount pure vehicle damage"), performing rule matching on standardized features to quickly assign items to be assigned to the corresponding first-level category. Fine classification can then be performed on items to be assigned within the same first-level category using the K-means clustering algorithm (K value determined by the elbow rule, ranging from 3 to 5, with a clustering iteration count greater than or equal to 50). Next, the convergence threshold is set to 0.001. Clustering is performed according to the feature vector, and items with high feature similarity are grouped into the same second-level sub-class. Furthermore, a unique category label can be assigned to each item to be assigned (in the format of "first-level major category code + second-level sub-class code", such as "CL-B01", where CL represents the property insurance claim major category and B01 represents the medium-amount vehicle damage sub-class). A mapping relationship table of "item ID-category label-feature vector" is generated and stored in the classification result database for subsequent steps S142 to call. At the same time, the classification time and classification algorithm parameters are recorded to ensure that the classification process is traceable and reproducible.

[0031] As a specific example in the property insurance field, the description of the item to be assigned is "Reported amount = 22,000 yuan, Accident type = 2 (vehicle collision), Accident location = urban area, Number of accidents = 1, Vehicle usage nature = non-commercial". The core classification features are extracted and processed as follows: Numerical features: Reported amount (22,000 yuan, standardized = (22,000 - 5,000) / (50,000 - 5,000) approximately equal to 0.3778), Number of accidents (1, standardized = 0.25); Classification coding features: Accident type (2 - tag code = 2), vehicle usage nature (non-commercial corresponding tag code = 1), accident location (urban area corresponding tag code = 1); the primary category in the pre-defined classification of property insurance claims can be CL (property insurance claims), which can be divided into 3 secondary categories: CL-A01: minor pure vehicle damage (reported amount less than or equal to 5,000 yuan, accident type = 1 / 2, no personal injury, number of accidents less than or equal to 2); CL-B01: medium vehicle damage + simple property damage (5,000 < reported amount < 3 0000 yuan, incident type = 2 / 3, no casualties, incident count less than or equal to 2 times); CL-C01: Large amount / injury / theft (reported amount > 30000 yuan, or includes casualties, or incident type = 4 (theft)); Further, when classifying, we can first use the coarse classification matching rule "5000 < reported amount < 30000 yuan and incident type = 2 and no casualties" to classify it into the first-level major category CL and the second-level subcategory B01; further, we can use the subcategories to classify the items to be assigned under the CL major category. K-means clustering (K=3, number of iterations=50, convergence threshold=0.001), the similarity between the feature vector of the current item to be assigned and the cluster center of CL-B01 is 0.92 (similarity threshold=0.85), confirming that it belongs to CL-B01; further, since the category label is CL-B01, a mapping relationship can be generated (item ID to be assigned: LC2024001, category label: CL-B01, feature vector: [0.3778,0.25,2,1,1]), and stored in the property insurance classification database.

[0032] As another concrete example, in the field of healthcare, the description information of the items to be assigned is "Patient age = 65 years old, diagnosed disease type = 2 (hypertensive emergency), disease severity = 2, vital signs (blood pressure 160 / 100 mmHg), underlying medical history = 1 (diabetes)". The core classification features are extracted and processed as follows: Numerical features: Age (65 years old, standardized = (65-18) / (80-18) ≈ 0.7581), Blood pressure (160 / 100 mmHg, standardized systolic blood pressure...). =0.625, diastolic blood pressure =0.6667); Classification coding characteristics: confirmed disease type (2 corresponding to label code = 2), severity of illness (2 corresponding to label code = 2), underlying medical history (1 corresponding to label code = 1); Further, the primary category in the medical and health field can be ME (medical resources), which is divided into 3 secondary subcategories: ME-A01: Mild internal medicine (severity of illness = 1, no underlying medical history, age < 60 years, stable vital signs); ME-B01: Severe internal medicine (severity of illness = 2, with underlying medical history, age greater than or equal to 60 years, vital signs basically stable); ME-C01: Acute / trauma (severity of illness = 3, regardless of age or underlying medical history, unstable vital signs); Further, when classifying, we can first use the coarse classification matching rule "severity of illness = 2 and with underlying medical history and age greater than or equal to 60 years" to classify it into the primary category ME and the secondary subcategory B01; Further, through fine classification: for the items to be assigned under the ME category, K- Means clustering (K=3, number of iterations=50, convergence threshold=0.001), the similarity between the feature vector of the current item to be assigned and the cluster center of ME-B01 is 0.91, confirming that it belongs to ME-B01; further, when the category label is ME-B01, a mapping relationship is generated (item ID to be assigned: ME2024001, category label: ME-B01, feature vector: [0.7581,0.625,0.6667,2,2,1]), and stored in the medical classification database.

[0033] Step S142: Monitor the allocation decision pattern of the target machine learning model for each category of the items to be allocated; Specifically, in this step, a model decision monitoring module can be deployed to link with the target machine learning model (such as XGBoost, LightGBM) and the database of items to be assigned. It adopts a real-time monitoring + timed aggregation approach to ensure that no decision data is missed. The monitoring module supports configuring the monitoring frequency (real-time monitoring: data is collected synchronously every time an assignment result of an item to be assigned is received; timed aggregation: decision data of each category is aggregated once per hour). Specifically, for each category of items to be allocated (divided according to the category labels generated in step S141), complete allocation decision data of the target machine learning model can be continuously collected. The collected content may include: (1) Input data: standardized feature vectors and original description information of the items to be allocated; (2) Output data: the second initial allocation scheme generated by the model (including resource type, quantity, specific parameters, priority, etc.); (3) Decision process data: the contribution weight of each feature to the model output (calculated by SHAP value or Gini coefficient), and the output results of decision nodes within the model. More specifically, for multiple items to be allocated under the same category (e.g., sample size greater than or equal to 50, to ensure pattern stability), a stable allocation decision pattern can be extracted using statistical analysis methods. The specific steps may be to first perform feature combination analysis, statistically analyze the high-frequency feature combinations among similar items to be allocated, and screen features with an occurrence frequency greater than or equal to 80%. The combination serves as the input condition for the decision-making model. Further, it can be combined with i to perform output distribution analysis: statistical analysis of the model output results (such as loss assessment amount, drug dosage) for similar items to be allocated, calculating the mean and standard deviation, and determining the reasonable range of output parameters as the output result of the decision-making model. Further, feature-output mapping rules can be extracted; through correlation analysis (such as a Pearson correlation coefficient greater than or equal to 0.7), the mapping relationship between core features and output parameters can be determined, clarifying the corresponding rules of "feature value range and output parameter range". Further, the extracted stable decision-making models for each category can be organized into structured data of "category label - feature combination - output range - feature contribution weight" and stored in the model decision-making model library. A timed update mechanism is set (updated once daily at midnight); after adding new decision data for similar items to be allocated, the statistical analysis is repeated, updating the parameter range and feature combination of the decision-making model to ensure that the decision-making model conforms to the actual decision-making logic of the model.

[0034] As a specific example, in the property insurance field, a property insurance model decision monitoring module can be built, linked with the XGBoost claims resource allocation model and the property insurance classification database, and set up real-time monitoring (data is collected synchronously for each claim allocation result received) and summarize the decision data of CL-B01 category (medium-amount vehicle damage) once per hour; further, 50 decision data can be collected for CL-B01 category items to be allocated. The core collection content is as follows: (1) Input features: reported amount (12,000-28,000 yuan), accident type = 2 (collision), number of accidents = 1-2 times, vehicle usage nature = non-commercial; (2) Output results: claims specialist level = 2 (intermediate), number of vehicles inspected = 1, damage assessment ratio = 82%-88%, priority = medium; (3) Feature contribution weight: reported amount (SHAP value = 0.45), accident type (SHAP value = 0.23), number of accidents (SHAP value = 0.18); further, feature combination analysis can be used to filter out the frequency of occurrence. The characteristic combination with a percentage greater than or equal to 80% is defined as "reported amount within (10000, 30000) and accident type = 2, accident count less than or equal to 2, and vehicle usage nature = non-commercial". Further output distribution analysis reveals a mean loss assessment ratio of 85% and a standard deviation of 2%, determining the output range to be 83%-87%. All claims adjusters are at level 2 (intermediate), and all have only one vehicle inspected. Further analysis reveals a mapping pattern: the reported amount and the loss assessment ratio are positively correlated. The correlation coefficient (Pearson correlation coefficient = 0.78) indicates that the higher the reported amount, the closer the loss assessment ratio is to 87%. Finally, the CL-B01 decision-making model is organized into structured data (category label: CL-B01; feature combination: reported amount (10000, 30000], accident type = 2, etc.; output range: loss assessment ratio 83%-87%, 1 intermediate specialist, 1 survey vehicle; feature contribution weight: reported amount 0.45, etc.), and stored in the property insurance model decision-making model library, which is updated every morning.

[0035] As another specific example, in the field of medical and health care, a medical model decision monitoring module can be built, which is linked with the LightGBM medical resource allocation model and medical classification database. Real-time monitoring is set up (data is collected synchronously for each patient resource allocation result received), and the decision data of ME-B01 category (critical internal medicine) is summarized once per hour; for ME-B01 category items to be allocated, 50 decision data are collected. The core collection content is as follows: (1) Input features: age 60-75 years old, diagnosed disease type = 2 (hypertensive emergency), severity of illness = 2, basic medical history = 1 (diabetes / hypertension), weight 60-80kg; (2) Output results: bed type = critical care, number of nurses = 1, ventilator = 1 (standby), drug dosage calculation coefficient = 10.0-10.5mg / kg, priority = high; (3) Feature contribution weight: age (SHAP value = 0.38), severity of illness (SHAP value = 0.42), weight ... =0.15); Further, when extracting decision patterns, feature combination analysis can be performed to select feature combinations with a frequency greater than or equal to 80% as "age ≥ 60 years old, diagnosed disease = 2, disease severity = 2, with underlying medical history, and weight 60-80 kg"; Further, output distribution analysis can be performed to make the mean of drug dosage calculation coefficient = 10.3 mg / kg, standard deviation = 0.15, and determine the output range as 10.15-10.45 mg / kg; the configuration of beds, nurses, and ventilators is fixed (1 intensive care bed, 1 nurse, 1 spare ventilator); Further, the mapping law is found to be that the severity of the disease is positively correlated with the drug dosage calculation coefficient (Pearson correlation coefficient = 0.76), that is, the more severe the disease, the closer the coefficient is to 10.45 mg / kg; Finally, the ME-B01 type decision patterns are organized into structured data, stored in the medical model decision pattern library, updated every morning, and the output range is recalibrated after adding decision data of the same type of patients.

[0036] Step S143: When the similarity of each allocation decision pattern meets the preset condition and the allocation decision pattern has business interpretability, the allocation decision pattern is transformed into at least one structured and parameterized derivative business rule. Specifically, in this step, when performing decision pattern similarity verification, cosine similarity can be used as the core verification index to calculate the similarity between the decision pattern corresponding to newly added decision data and the stored stable decision pattern under the same category. The cosine similarity value ranges from [0,1], and the closer it is to 1, the higher the similarity. Specifically, when setting preset conditions, the similarity preset threshold can be calibrated based on the business scenario, with a value range of 0.85-0.95, and 0.9 is preferred (to ensure the stability of the decision pattern). At the same time, it is required that the similarity between the decision patterns of 20 consecutive items to be assigned and the stable decision patterns is greater than or equal to the preset threshold, and the coefficient of variation of the output parameters of the same type of decision patterns is less than or equal to 10% (coefficient of variation = standard deviation / mean, to ensure output stability). Further specifically, the similarity verification module can automatically calculate that if both "20 consecutive similarities are greater than or equal to 0.9" and "the coefficient of variation of the output parameters is less than or equal to 10%" are met, then the decision pattern similarity is determined to meet the preset conditions; otherwise, return to step S142 to continue monitoring and collecting data. More specifically, a business interpretability verification module can be deployed, with a built-in business constraint rule library (pre-configured by business personnel, including business legal features, output parameter constraints, and rule logic conflict judgment standards). Specifically, when verifying the legality of features, the input features in the decision-making model can be legal bases recognized by the business scenario (such as the reported amount and type of accident in property insurance, and age and disease level in medical care), with no invalid features (such as random features unrelated to the allocation result). Specifically, when verifying the compliance of the output, the output of the decision-making model should meet business constraints (such as the property insurance loss assessment ratio not exceeding 95% and not lower than 70%; the medical drug dosage coefficient should meet medical standards) and have no illegal outputs. Specifically, when verifying the interpretability of the logic, the mapping relationship between the "feature combination and output result" of the decision-making model should be interpretable through natural language or business logic, without black-box logic (such as "the higher the reported amount, the higher the loss assessment ratio," which conforms to the property insurance claims logic; "the more severe the illness, the higher the drug dosage," which conforms to the medical clinical logic), and should not conflict with existing business rules. Finally, the verification module automatically matches the business constraint rule library. If all three conditions are met, it is determined that the model has business interpretability. If any condition is not met, the process returns to step S142 to re-optimize the decision-making model extraction logic. More specifically, when transforming decision-making models into structured, parameterized business rules, the transformation rules can adopt an "IF-THEN" structured format, clearly defining conditions (input feature combinations) and actions (output allocation parameters). The condition part uses a standardized expression of "feature name + operator + value range," while the action part uses a parameterized expression of "resource type + parameter value + unit," avoiding ambiguous expressions. The transformation steps can be: Condition extraction: Convert frequently occurring feature combinations in the decision-making model into conditional statements after IF, using operators such as "belongs to," "greater than or equal to," "less than or equal to," "equal to," and "and," with value ranges using quantified intervals (e.g., "reported amount"). Belongs to (10000, 30000]”); Action Extraction: Transform the stable output results in the decision-making model into action statements after THEN, clarifying the specific parameters of resource allocation; Rule Standardization: Add rule type (derived rule), corresponding category label, and rule priority (lower than the original manually configured rules) to the derived business rules to ensure that the rule format is consistent with the original rules in the target business rule set and can be directly parsed and executed by the rule engine; Transformation and Verification: Substitute the generated derived business rules into the rule engine for syntax and logic verification to ensure that there are no syntax errors and no logical conflicts. If the verification passes, proceed to the next step; if the verification fails, readjust the expression format of the conditions or actions.

[0037] As a concrete example, in the property insurance field, when verifying the similarity of decision patterns, similarity calculations can be performed to ensure that for class CL-B01, the cosine similarity between the decision patterns of 20 consecutive items to be allocated and the stable decision patterns is greater than or equal to 0.92, satisfying the preset threshold of 0.9; furthermore, the coefficient of variation of the loss assessment amount ratio = 2% / 85% is approximately 2.35%, which is less than 10%, indicating stable output; thus, it can be determined that the similarity meets the preset conditions. Furthermore, the interpretability of the business can be verified. First, the legality of the features is verified, ensuring that the input features (reported amount, type of accident, number of accidents, and vehicle usage nature) are all legal grounds for property insurance claims, with no invalid features. Second, the compliance of the output is verified, ensuring that the loss assessment ratio is 83%-87%, which meets the constraints of property insurance claims (70%-95%). The level of the claims specialist and the number of vehicles inspected both meet the business configuration standards. Third, the interpretability of the logic is verified, as "the higher the reported amount and the type of accident is collision, the higher the loss assessment ratio," which is consistent with the logic of property insurance claims and does not conflict with the original rules. Finally, it is determined that the allocation decision model has business interpretability.

[0038] Further, the allocation decision model can be transformed into derivative business rules: (1) Condition extraction: IF The reported amount belongs to (10000, 30000] and the accident type = 2 (collision) and the number of accidents is less than or equal to 2 and the vehicle usage nature = non-commercial; (2) Action extraction: THEN Allocate 1 intermediate claims specialist, 1 survey vehicle, loss assessment ratio = 85% ± 2%, allocation priority = medium; (3) Rule standardization: Rule type = derivative rule, corresponding category label = CL-B01, rule priority = 3 (the original manual rule priority is 1-2); (4) Conversion verification: Substitute into the property insurance rule engine, the syntax and logic verification pass, and it can be directly parsed and executed. Final derivative business rule: IF If the reported amount is within (10000, 30000) and the type of accident is 2, the number of accidents is less than or equal to 2, and the vehicle usage nature is non-commercial, then the allocation is as follows: 1 intermediate claims specialist, 1 inspection vehicle, damage assessment ratio = 85% ± 2%, priority = medium.

[0039] As another concrete example, in the healthcare field, when verifying the similarity of decision-making patterns, similarity calculations can be performed. For the ME-B01 class, the cosine similarity between the decision-making patterns of 20 consecutive items to be assigned and the stable decision-making patterns is greater than or equal to 0.91, which meets the preset threshold of 0.9. Furthermore, the coefficient of variation of parameters can be output. The coefficient of variation of the drug dosage calculation coefficient is calculated to be 0.15 / 10.3, which is approximately 1.46%, less than 10%, indicating that the output is stable. Finally, it is determined that the similarity meets the preset conditions. When performing business interpretability verification, the following steps can be taken: First, verify the legality of the features, i.e., verify that the input features (age, diagnosed disease, severity of illness, underlying medical history, weight) are all legal bases for medical resource allocation and that there are no invalid features. Second, verify the compliance of the output, i.e., the drug dosage calculation coefficient of 10.15-10.45mg / kg complies with medical standards (8-12mg / kg); the configuration of hospital beds, nurses, and ventilators meets the standards for critical care patients. Third, verify the interpretability of the logic, i.e., "high severity of illness, older age, high drug dosage coefficient", which is consistent with the clinical medication logic for hypertensive emergencies and does not conflict with the original medical rules. Finally, it can be determined that the allocation decision model has business interpretability.

[0040] Further transform the allocation decision model into derivative business rules: (1) Condition extraction: IF age greater than or equal to 60 years old and diagnosed disease type = 2 (hypertensive emergency) and severity of condition = 2 and basic medical history = 1 (yes) and weight is [60,80] kg; (2) Action extraction: THEN allocate 1 intensive care bed, 1 nurse, 1 ventilator (standby), drug dosage calculation coefficient = 10.3mg / kg±0.15mg / kg, allocation priority = high; (3) Rule standardization: rule type = derivative rule, corresponding category label = ME-B01, rule priority = 3 (the original manual rule priority is 1-2); (4) Transformation verification: substitute into the medical rule engine, the syntax and logic verification pass, and it can be directly parsed and executed. Final Derivative Business Rules: IF age greater than or equal to 60 years old and diagnosed disease type = 2 and disease severity = 2 and have underlying medical history and weight within [60, 80] kg, THEN allocation: 1 intensive care bed, 1 nurse, 1 ventilator (standby), drug dosage calculation coefficient = 10.3 mg / kg ± 0.15 mg / kg, priority = high.

[0041] Step S144: Store the derived business rules into the pre-configured target business rule set.

[0042] Specifically, in this step, the derived rules can be preprocessed, including: (1) Rule number: assign a unique number to each derived rule in the format of "rule type + category label + serial number" (e.g., property insurance derived rule number: Y-CL-B01-001, where Y represents the derived rule, CL-B01 is the category label, and 001 is the serial number); (2) Rule source mark: add a source mark (marked as "model derived") to the derived rule to distinguish it from the original manually configured rule (marked as "manual configuration"), which is convenient for subsequent traceability and management; (3) Version management: assign an initial version number (e.g., V1.0) to the derived rule, record the rule generation time and the decision mode parameters at the time of generation, which is convenient for subsequent rule updates and rollback; furthermore, rule conflict verification can be performed, deploy a rule conflict verification module, and perform conflict verification between the preprocessed derived rule and the pre-configured target business rule set. The verification content includes: (1) Condition conflict: check whether the IF condition of the derived rule is consistent with the IF condition of the original rule. Completely consistent but THEN actions inconsistent (e.g., the original rule loss assessment ratio is 80%, and the derivative rule is 85%); (2) Logic conflict: check the logic of the derivative rule to see if it conflicts with the core business logic of the original rule (e.g., the original rule prohibits the assignment of intermediate specialists to cases with more than 3 claims, while the derivative rule allows it); (3) Verification processing: if there is a conflict, the principle of "original rule priority" is adopted to adjust the derivative rule (e.g., adjust the conditions or output parameters of the derivative rule) and re-verify; if there is no conflict, proceed to the next step; further, rule storage and synchronization can be performed, which may include: (1) Storage method: use the append writing method to store the derivative rule into the target business rule set, and update the index of the rule set at the same time to ensure that the rule can be quickly retrieved; (2) System synchronization: synchronize the updated target business rule set to the corresponding business system (e.g., property insurance claims system, hospital HIS system) and rule engine to ensure that the rule engine can call the derivative rule when the subsequent items to be allocated are allocated.

[0043] As a specific example, in the field of property insurance, the following preprocessing of derivative rules can be performed first: (1) Rule number: Y-CL-B01-001 (Y = derivative rule, CL-B01 = medium-amount vehicle damage category, 001 = first rule); (2) Source mark: model derivative; (3) Version management: version V1.0, generation time 2024-05-20, corresponding decision mode parameters: loss assessment ratio 83%-87%, 1 intermediate specialist; Secondly, the following rule conflict verification can be performed: (1) Conflict verification: verify Y-CL-B01-001 with the property insurance target business rule set (including 10 manually configured rules), and find that there is no unconditional conflict. Conflicts and logical conflicts (the original rules did not configure relevant rules for this feature combination); (2) Verification result: no conflict, can be stored; furthermore, rules can be stored and synchronized: (1) Storage: in append mode, write the derived rule into the property insurance target business rule set, update the rule set index, and add the "CL-B01 category" search tag; (2) Synchronization: synchronize the updated rule set to the property insurance claims system and Drools rule engine, configure the rule call priority (lower than manual rules); finally, records and traceability can be performed: in the property insurance rule management ledger, record all information of the rule, store it in the property insurance rule management database, and support subsequent queries, updates and rollbacks. Subsequently, when there is a CL-B01 category (medium-amount vehicle damage) item to be allocated, the rule engine will prioritize calling the manually configured rule. If there is no corresponding manual rule, it will automatically call the derived rule to quickly complete the resource allocation.

[0044] As another specific example, in the field of healthcare, the following preprocessing of derived rules can be performed first: (1) Rule number: Y-ME-B01-001 (Y = derived rule, ME-B01 = critical care medicine category, 001 = first rule); (2) Source mark: model derived; (3) Version management: version V1.0, generation time 2024-05-20, corresponding decision mode parameters: drug dosage coefficient 10.15-10.45mg / kg, 1 intensive care bed; Secondly, the following rule conflict verification can be performed: (1) Conflict verification: compare Y-ME-B01-001 with the medical target business rule set (including 15 manually configured rules). Verification revealed no conflict (the original rules only configured relevant rules for patients with a severity level of 3); (2) Verification result: no conflict, can be stored; furthermore, the following rule storage and synchronization can be performed: (1) Storage: In an append manner, write the derived rule into the medical target business rule set, update the rule set index, and add the "ME-B01 category" search tag; (2) Synchronization: Synchronize the updated rule set to the hospital HIS system and rule engine, configure the rule call priority (lower than manual rules); finally, it can be recorded and traced, record all the information of the rule in the medical rule management ledger, store it in the medical rule management database, and support subsequent clinical audits and rule optimization. Subsequently, when there are ME-B01 category (critical internal medicine) patients, the rule engine will prioritize calling the manually configured rule. If there is no corresponding manual rule, it will automatically call the derived rule to achieve rapid and standardized allocation of medical resources.

[0045] It is understandable that the above embodiments automatically mine, extract and generate interpretable and executable derivative business rules from the allocation decisions of the machine learning model, thereby realizing the white-box and rule-based output of the black-box model, which not only preserves the accuracy of the model, but also meets the business's requirements for interpretability, auditability and intervention.

[0046] In some embodiments of the present invention, step S143 includes the following steps.

[0047] Step S1431: Apply the model post-interpretation method to analyze the multiple allocation decision patterns of the target machine learning model for each category of the items to be allocated, quantify the contribution of each feature of the items to be allocated to the allocation decision pattern, and generate a feature contribution ranking list of each feature. Specifically, in this step, the characteristics of the target machine learning model (such as XGBoost, LightGBM) can be combined to select a post-hoc interpretation method that can quantify the contribution of features. Among them, the SHAP (SHapley Additive exPlanations) algorithm (which balances interpretability and quantification accuracy and is compatible with various machine learning models) can be used. The SHAP interpretation module can be deployed and linked with the model decision pattern library and the feature database of the items to be assigned to ensure that the decision pattern data and feature data of each category of items to be assigned can be called. More specifically, multiple allocation decision patterns for each category of items to be allocated can be extracted from the model decision pattern library (e.g., the sample size for each category is greater than or equal to 50 to ensure that the analysis results are statistically significant). The corresponding data of "feature vector - model output result" for each decision pattern can be organized. The feature vector contains all the classification features extracted in step S141, and the model output result is the core parameters of the second initial allocation scheme (e.g., loss assessment ratio, drug dosage coefficient). More specifically, when performing feature contribution quantification calculation, the SHAP interpretation module can be used to input the feature vector of each decision pattern into the model, calculate the SHAP value of each feature for that decision pattern (model output result), and the larger the absolute value of the SHAP value, the higher the contribution of the feature to the decision pattern. A positive SHAP value indicates that the feature promotes the decision result to tilt in a certain direction, while a negative SHAP value indicates inhibition. The feature SHAP values ​​of all decision patterns under the same category are statistically analyzed, and the average absolute value of each feature is calculated (eliminating the offsetting effect of positive and negative values), which is taken as the final contribution of the feature to the decision pattern of that category, and the value range is quantified to [0,1]. The features are sorted from high to low according to their final contribution, generating a structured sorted list of "feature name - average absolute value of SHAP (contribution) - feature type - corresponding category", and the sorted list of each category is marked with a unique identifier. The list is stored in the feature contribution database, and the interpretation method and calculation parameters (such as the kernel function type and number of iterations of the SHAP algorithm) are recorded to ensure that the calculation process is traceable and reproducible.

[0048] As a concrete example, in the property insurance field, the SHAP algorithm can be selected, and a SHAP interpretation module can be deployed to link with the XGBoost claims resource allocation model, the property insurance model decision pattern library, and the property insurance pending allocation item feature database. This supports calling decision pattern data for the CL-B01 class (medium-amount vehicle damage). From the property insurance model decision pattern library, 50 decision pattern data points for the CL-B01 class (medium-amount vehicle damage) are extracted. Each data point includes a feature vector (reported amount, accident type, number of accidents, vehicle usage nature) and the model output result (damage assessment ratio). This is then processed through the SHAP module. Calculate the SHAP value for each feature of each data point. Example: For a reported claim of 22,000 yuan, claim type = 2, claim count = 1, and a non-commercial decision-making mode, the SHAP values ​​are: reported claim amount 0.45, claim type 0.23, claim count 0.18, and vehicle usage nature 0.14. Calculate the average absolute SHAP value of 50 data points to obtain the contribution of each feature: reported claim amount 0.42, claim type 0.25, claim count 0.19, and vehicle usage nature 0.14. Sort the features by contribution from highest to lowest to generate a CL-B01 category feature contribution ranking list: 1. 1. Reported amount (contribution 0.42, numerical type, CL-B01 category); 2. Type of accident (contribution 0.25, classification coding type, CL-B01 category); 3. Number of accidents (contribution 0.19, numerical type, CL-B01 category); 4. Vehicle usage nature (contribution 0.14, classification coding type, CL-B01 category); Store this list in the property insurance feature contribution database, identified as “CL-B01-contribution-20240520”, and record the calculation parameters (SHAP kernel function = linear kernel, number of iterations = 100).

[0049] As another concrete example, in the healthcare field, the SHAP algorithm can be selected, and a SHAP interpretation module can be deployed to link with the LightGBM medical resource allocation model, the medical model decision pattern library, and the medical allocation item feature database. This supports accessing decision pattern data for the ME-B01 class (critical care medicine). From the medical model decision pattern library, 50 decision pattern data points for the ME-B01 class (critical care medicine) are extracted. Each data point contains a feature vector (age, disease severity, underlying medical history, weight) and the model output (drug dosage calculation coefficient). The SHAP module calculates the SHAP value for each feature of each data point. For example, for a decision pattern with an age of 65, disease severity = 2, underlying medical history, and weight of 70kg, the SHAP values ​​are: disease severity 0.42, age 0.38, and weight 0.1. 5. Basic medical history 0.05; Calculate the average absolute value of SHAP for 50 data points to obtain the contribution of each feature: disease severity 0.40, age 0.37, weight 0.16, basic medical history 0.07; Sort by contribution from high to low to generate a ranking list of ME-B01 class features: 1. Disease severity (contribution 0.40, categorical coding type, ME-B01 class); 2. Age (contribution 0.37, numerical type, ME-B01 class); 3. Weight (contribution 0.16, numerical type, ME-B01 class); 4. Basic medical history (contribution 0.07, categorical coding type, ME-B01 class); Store this list in the medical feature contribution database, labeled as "ME-B01-contribution-20240520", and record the calculation parameters (SHAP kernel function = linear kernel, iteration count = 100).

[0050] Step S1432: Based on the feature contribution ranking list, select core decision features with a contribution higher than a preset threshold. Specifically, in this step, when setting the preset contribution threshold, the contribution threshold (quantified and dynamically adjustable) can be preset based on the business scenario and feature contribution distribution. The principle for setting the threshold can be: ensuring that the selected core decision features can explain more than or equal to 80% of the decision pattern variation (i.e., the sum of the contributions of the core features is greater than or equal to 0.8), while avoiding the selection of too many features that would lead to redundancy in subsequent mathematical expressions; generally, the threshold value range is 0.15-0.3, and the specific value should be calibrated according to the contribution distribution of each category of features. More specifically, when screening core decision features, the feature contribution ranking list generated in step S1431 can be called to determine whether the contribution of each feature is higher than the preset threshold. At the same time, the sum of the contributions of the screened features is calculated. When the sum of the contributions is greater than or equal to 0.8, the screening stops (even if the contribution of subsequent features is higher than the threshold, they will not be included to avoid redundancy). More specifically, the screening results can be verified, that is, the core decision features selected can be verified. Correlation analysis (such as correlation analysis using Pearson correlation coefficient) can be used to check for multicollinearity among the core features (collinearity threshold is less than or equal to 0.7; if collinearity exists, features with low contribution are removed) to ensure that the core features are independent and free of redundancy and can effectively support the fitting of subsequent mathematical expressions. More specifically, the core decision features that have passed the screening and verification can be organized into structured data of "feature name-contribution-feature type-value range", labeled with corresponding category tags, and stored in the core feature database for subsequent steps S1433 to call. At the same time, the screening threshold and verification results are recorded to ensure that the screening process is traceable.

[0051] As a specific example, in the property insurance field, for CL-B01 category (medium-amount vehicle damage), based on the feature contribution distribution (highest 0.42, lowest 0.14), the preset contribution threshold is 0.2, and the sum of the core feature contributions is required to be greater than or equal to 0.8; further feature screening is carried out: (1) the reported amount (0.42>0.2) is included in the core features, with a cumulative contribution of 0.42; (2) the type of accident (0.25>0.2) is included in the core features, with a cumulative contribution of 0.42+0.25=0.67; (3) the number of accidents (0.19<0.2) is not included; (4) Vehicle usage characteristics (0.14 < 0.2) are not included; (5) Cumulative contribution 0.67 < 0.8, adjust the threshold to 0.18 and re-screen: number of accidents (0.19 > 0.18) are included in the core features, cumulative contribution = 0.42 + 0.25 + 0.19 = 0.86 > 0.8, stop screening; further, calculate the correlation of the three core features, the Pearson correlation coefficient between the reported amount and the number of accidents = 0.58 ≤ 0.7, the correlation coefficient between the accident type and the other two features is < 0.7, there is no multicollinearity, the verification is passed; finally, the core decision features of CL-B01 are: 1. 1. Reported amount (contribution 0.42, numerical type, value range (10000, 30000] yuan); 2. Type of incident (contribution 0.25, classification coding type, value = 2 (collision)); 3. Number of incidents (contribution 0.19, numerical type, value range [1, 2] times); Stored in the core feature database of property insurance, recording the screening threshold of 0.18 and the verification result (no collinearity).

[0052] As another specific example, in the field of healthcare, when setting preset thresholds, for the ME-B01 class (critical care medicine), based on the feature contribution distribution (highest 0.40, lowest 0.07), the preset contribution threshold is 0.3, while requiring the sum of the core feature contributions to be greater than or equal to 0.8; Further feature screening was conducted: (1) Severity of illness (0.40 > 0.3), included as a core feature, with a cumulative contribution of 0.40; (2) Age (0.37 > 0.3), included as a core feature, with a cumulative contribution of 0.40 + 0.37 = 0.77 < 0.8; (3) Weight (0.16 < 0.3), not included; (4) Underlying medical history (0.07 < 0.3), not included; (5) The cumulative contribution of 0.77 is close to 0.8, so the threshold was adjusted to 0.25 and rescreened: the age still met the criteria, with a cumulative contribution of 0.77, and no additional features were needed (weight 0.16 is still < 0.25), confirming the core feature. The criteria were: severity of illness and age (cumulative contribution 0.77, close to 0.8, and no other matching features, meeting business requirements); further, the correlation between the two core features was calculated, and the Pearson correlation coefficient between age and severity of illness was 0.62 < 0.7, with no multicollinearity, thus passing the verification; finally, the core decision features of ME-B01 class were obtained as follows: 1. Severity of illness (contribution 0.40, categorical coding type, value = 2); 2. Age (contribution 0.37, numerical type, value range [60, 75] years); stored in the medical core feature database, recording the screening threshold of 0.25 and the verification result (no multicollinearity).

[0053] Step S1433: Using the core decision features as input variables for the mathematical expression and the prediction results of the target machine learning model as the approximate fitting target for the mathematical expression, the symbolic regression algorithm is used to iteratively search and optimize the mathematical expression. Specifically, in this step, a symbolic regression algorithm that can automatically search mathematical expressions can be selected (GPTune or Eureqa algorithm can be selected to adapt to multi-feature fitting and support linear and nonlinear expression search), deploy the symbolic regression module, link with the core feature database and model decision mode library, and call the core decision features selected in step S1432 and the "feature vector-model output result" data prepared in step S1431; further, the input, output and algorithm parameters of the fitting can be clarified to ensure that the fitting result is accurate and fits the actual business, specifically: (1) Input variables: core decision features selected in step S1432 (numerical features are directly used as input, and classification coding features are converted into numerical variables, such as risk type = 2 is converted into numerical 2); (2) Fitting target: the prediction result of the target machine learning model (i.e. the core parameters of the second initial allocation scheme, such as loss assessment ratio and drug dosage coefficient), as the output value of the mathematical expression; (3) Algorithm parameters: the number of iterations is greater than or equal to 1000 (to ensure that the optimal expression is found), and the fitting error threshold is less than or equal to 5% (i.e. the expression output value). The deviation from the model prediction result must be less than or equal to 5%. Allowed operators include addition, subtraction, multiplication, division, and exponentiation, while complex operators are prohibited (to avoid expressions that cannot be converted into business rules). Furthermore, the symbolic regression module automatically calls core feature data and model prediction result data, iteratively searches according to configured parameters, generates multiple candidate mathematical expressions, calculates the fitting error (mean squared error, MSE) of each expression, filters out expressions with fitting errors less than or equal to 5%, and then selects the expression with the simplest structure and easiest conversion into business rules (prioritizing linear expressions; if the linear fitting error does not meet the requirement, a simple nonlinear expression is selected). Finally, the optimal mathematical expression is substituted into the test data (accounting for 20% of the total data) that was not used for fitting to verify whether the fitting error is still less than or equal to 5%. If it does, it is determined as the final mathematical expression; if not, the algorithm parameters are adjusted (increasing the number of iterations, expanding the range of operators), and the iterative search is repeated until the error requirement is met. The final mathematical expression is stored in the expression database, recording the algorithm parameters and fitting error to ensure reproducibility.

[0054] As a specific example, in the property insurance field, the Eureqa symbolic regression algorithm can be selected, the symbolic regression module can be deployed, and it can be linked with the property insurance core feature database and the property insurance model decision mode library to call the CL-B01 class core features (reported amount, accident type, accident number) and 50 decision mode data (40 for fitting and 10 for testing); further, the fitting parameters can be configured as follows: (1) Input variables: reported amount (x1, unit: 10,000 yuan), accident type (x2=2, fixed value), accident number (x3, value 1-2); (2) Fitting target: loss assessment ratio (y, unit: %); (3) Algorithm parameters: number of iterations = 1500 times, fitting error threshold = 5%, allowed operators: addition, subtraction, multiplication, division; Further iterative search and optimization can be performed, specifically: (1) Algorithm iterative search to generate multiple candidate expressions, among which the linear expression with the smallest fitting error is: y=75+0.5 x1+2 x3 (x2=2, substituting it does not affect the expression structure); (2) Calculate the fitting error: the average fitting error of 40 fitting data points = 3.2% < 5%, which meets the requirements, and the expression structure is simple and easy to convert into business rules; finally, substitute the expression into 10 test data points, for example: x1=22 (ten thousand yuan), x3=1, the expression output y=75+0.5×22+2×1=75+11+2=88%, the model prediction result is 87%, the deviation = 1% is less than 5%; the average fitting error of 10 test data points = 3.5% is less than 5%, the verification is passed; the final mathematical expression: y=75+0.5 x1+2 x3 (x1 = reported amount / 10000, x3 = number of claims) is stored in the property insurance expression database, recording 1500 iterations and a fitting error of 3.2%.

[0055] As another specific example, in the field of healthcare, the Eureqa symbolic regression algorithm can be selected, the symbolic regression module can be deployed, and it can be linked with the medical core feature database and the medical model decision pattern library to call the ME-B01 class core features (severity of illness, age) and 50 decision pattern data (40 for fitting and 10 for testing); the fitting parameters can be further configured as follows: (1) Input variables: severity of illness (x1=2, fixed value), age (x2, unit: years); (2) Fitting target: drug dosage calculation coefficient (y, unit: mg / kg); (3) Algorithm parameters: number of iterations = 1500 times, fitting error threshold = 5%, allowed operators: addition, subtraction, multiplication, division; further iterative search and optimization can be performed: (1) Algorithm iterative search to generate multiple candidate expressions, among which the linear expression with the smallest fitting error is: y=5+0.08 x2 (x1=2, substituting it does not affect the expression structure); (2) Calculate the fitting error: the average fitting error of 40 fitting data points = 2.8% < 5%, which meets the requirements and the expression structure is simple; finally, substitute this expression into 10 test data points, for example: x2 = 65 (years old), the expression output y = 5 + 0.08 × 65 = 5 + 5.2 = 10.2 mg / kg, the model prediction result is 10.3 mg / kg, the deviation = 0.1 mg / kg ≤ 5%; the average fitting error of 10 test data points = 3.0% less than 5%, which is verified; the final mathematical expression: y = 5 + 0.08 x2 (x2 = age) is stored in the medical expression database, recording 1500 iterations and a fitting error of 2.8%.

[0056] Step S1434: Perform structured parsing and parameter extraction on the generated mathematical expression to obtain preliminary logical rules; Specifically, in this step, an expression parsing module can be deployed to perform structured parsing of the final mathematical expression determined in step S1433, breaking down the expression into its components, including: input variables (core decision features), operators, constant terms, and output variables (model prediction results). The correspondence between each component is clarified, and nonlinear expressions (if any) are transformed into decomposable logical structures (such as piecewise linear structures), ensuring that the parsing results can be converted into logical rules. More specifically, adjustable parameters, including constant terms and variable coefficients, can be extracted from the parsed expression. The value range of each parameter is clarified (based on fitting data calibration, ensuring that parameter values ​​conform to business constraints), and the physical meaning of the parameters is labeled (e.g., constant terms are basic parameter allocations, and variable coefficients are the influence coefficients of features on the output). More specifically, it is possible to... Based on the parsing results and parameter extraction results, the mathematical expression is transformed into a preliminary logical rule in the form of "condition-action", wherein: (1) Condition part: clarify the value range of the core decision feature (i.e., the reasonable range of the input variable, based on the value range of the core feature in step S1432); (2) Action part: clarify the calculation method of the output variable (i.e., the simplified form of the mathematical expression), mark the value of the adjustable parameter, and ensure that the action part is quantifiable and calculable; Finally, the generated preliminary logical rule can be subjected to syntax verification (to ensure that the logical structure is complete and there are no syntax errors) and logic verification (to ensure that the rule output is consistent with the mathematical expression output, and the deviation is less than or equal to 5%). After the verification is passed, it is stored in the preliminary rule database; if the verification fails, return to step S1433, adjust the mathematical expression or parsing logic, and re-parse and extract.

[0057] As a concrete example in the property insurance field, the first step can be expression parsing, specifically the mathematical expression y = 75 + 0.5. x1+2 x3 (x1 = reported amount / 10000, x3 = number of accidents) is analyzed and broken down into: input variables: x1 (reported amount / 10000, numerical type), x3 (number of accidents, numerical type); operators: +, ×; constant term: 75; output variable: y (loss assessment ratio); further, parameters are extracted to obtain adjustable parameters: constant term 75 (basic loss assessment ratio), x1 coefficient 0.5 (reported amount influence coefficient), x3 coefficient 2 (accident number influence coefficient); parameter value range: 75±2, 0.5±0.05, 2±0.2 (based on fitted data calibration, conforming to property insurance loss assessment constraints); further, preliminary logical rules are generated, combined with the core feature value range, and transformed into "condition-action" rules: condition The reported amount is between (10000, 30000) yuan (i.e., x1 belongs to (1, 3]) and the accident type is 2 (collision) and the number of accidents is [1, 2] times (x3 belongs to [1, 2]); Action: Loss assessment ratio = 75 + 0.5 × (reported amount / 10000) + 2 × number of accidents (adjustable parameters: basic ratio 75 ± 2, reported amount coefficient 0.5 ± 0.05, number of accidents coefficient 2 ± 0.2); Finally, the verification was performed, and the syntax check was correct; the test data (reported amount 22000 yuan, number of accidents 1 time) was taken, and the rule output = 75 + 0.5 × 2.2 + 2 × 1 = 88.1%, which is 0.1% less than 5% of the mathematical expression output 88%, so the logic check passed; it was stored in the property insurance preliminary rule database.

[0058] As another concrete example, in the healthcare field, the mathematical expression y = 5 + 0.08 × x² (x² = age) can be analyzed and broken down into: input variable: x² (age, numerical); operators: +, ×; constant term: 5; output variable: y (drug dosage calculation coefficient); further extracting adjustable parameters: constant term 5 (basic dosage coefficient), x² coefficient 0.08 (age influence coefficient); parameter value range: 5 ± 0.2, 0.08 ± 0.005 (based on fitted data calibration, conforming to medical standards); further combining the core feature value range, it is transformed into "conditional" -Action" rule: Conditions: Severity of illness = 2 and age is [60, 75] years old (x2 is [60, 75]) and there is a history of underlying medical conditions; Action: Drug dosage calculation coefficient = 5 + 0.08 × age (adjustable parameters: base coefficient 5 ± 0.2, age coefficient 0.08 ± 0.005); Finally, the verification shows no errors in syntax; Take test data (age 65 years old), rule output = 5 + 0.08 × 65 = 10.2 mg / kg, the deviation from the mathematical expression output 10.2 mg / kg = 0% less than 5%, the logic verification passes; Store in the medical preliminary rule database.

[0059] Step S1435: Map the variables, operators, and constants in the preliminary logical rules to predefined business terms and adjustable business parameters to form the derived business rules that can be directly interpreted and executed by the rule engine.

[0060] Specifically, in this step, a mapping table of "mathematical expression elements - business terms / business parameters" can be built in advance and configured by business personnel. The input variables, operators, constant terms and output variables in the mathematical expression are clearly defined and correspond to the predefined business terms and adjustable business parameters, respectively, to ensure that the mapping relationship fits the actual business and is unambiguous. Among them, the adjustable business parameters need to be marked with a reasonable range of values ​​and units, consistent with the business constraints. Further, the preliminary logic rules and preset mapping table generated in step S1434 can be called to map the elements in the preliminary rules one by one: (1) Input variables (core decision features): mapped to business terms commonly used in business scenarios (such as x1=reported amount / 10000, mapped to "reported amount (unit: yuan)"); (2) Operators: mapped to expressions that can be understood by the business (such as × mapped to "multiply", + mapped to "add", range symbol (3) Constant terms and variable coefficients: mapped to adjustable business parameters (e.g., constant term 75, mapped to "basic loss assessment ratio", with the range of values ​​marked); (4) Output variables: mapped to resource allocation parameters in the business scenario (e.g., y = loss assessment ratio, mapped to "loss assessment ratio (unit: %)"); Further, the mapped rules can be organized into a standardized format that the rule engine can directly parse and execute (using the "IF-THEN" structure), clarifying the logical relationship (AND / OR) of the condition part, the specific parameters and calculation methods of the action part, and adding rule types. (Derived rules), corresponding category tags, and rule priority (lower than the original manually configured rules) are used to ensure that the rule format is consistent with the original rules in the target business rule set. Finally, the standardized derived business rules can be substituted into the rule engine for syntax parsing and logic execution testing to ensure that the rules can be directly interpreted and executed by the engine, and that the deviation between the execution results and the model prediction results and the mathematical expression output results is less than or equal to 5%. At the same time, the rules are verified to have no conflict with the original business rules. After the verification is passed, they are determined to be the final executable derived business rules and stored in the derived rule database for use in step S144.

[0061] As a concrete example, in the property insurance field, a pre-defined mapping table can be used: Mathematical elements - business terms / business parameters: x1 (reported amount / 10000) - reported amount (unit: yuan); x3 (number of accidents) - number of accidents (unit: times); 75 - basic loss assessment ratio (%); 0.5 - impact coefficient of reported amount; 2 - impact coefficient of number of accidents; y - loss assessment amount ratio (%); + - add; × - multiply by; - Belongs to; Further, the preliminary logical rules can be mapped to business statements: Condition: The reported amount belongs to (10000, 30000) yuan and the accident type is collision (code 2) and the number of accidents belongs to [1, 2] times; Action: Loss assessment ratio = basic loss assessment ratio (75±2%) + reported amount influence coefficient (0.5±0.05) × (reported amount / 10000) + accident number influence coefficient (2±0.2) × accident number; Further, adopt the "IF-THEN" format Add rule information: Rule type: Derived rule; Category tag: CL-B01; Priority: 3 (1-2 levels lower than manual rules); Finally, input the rule into the rule engine, and the syntax parsing passes; Take test data (reported amount of 22,000 yuan, number of claims 1), the rule execution result = 75 + 0.5 × 2.2 + 2 × 1 = 88.1%, the deviation from the model prediction result of 87% = 1.1% < 5%, there is no conflict with the original rule, and the verification passes; Final derived business rule: IF reported amount (10000, 30000) yuan and the type of accident = collision and the number of accidents [1,2] times, THEN loss assessment ratio = basic loss assessment ratio (75±2%) + impact coefficient of reported amount (0.5±0.05) × (reported amount ÷ 10000) + impact coefficient of number of incidents (2±0.2) × number of incidents.

[0062] As another concrete example, in the healthcare field, a pre-defined mapping table could be: Mathematical elements - Business terms / Business parameters: x2 (age) - Patient age (unit: years); 5 - Basic drug dosage coefficient (mg / kg); 0.08 - Age influence coefficient; y - Drug dosage calculation coefficient (mg / kg); + - Add; × - Multiply; - Belongs to; Further, the preliminary logical rules are mapped to business statements: Condition: Severity of illness = 2 and patient age is [60, 75] years old and patient has underlying medical history; Action: Drug dosage calculation coefficient = basic drug dosage coefficient (5 ± 0.2 mg / kg) + age influence coefficient (0.08 ± 0.005) × patient age; Further, using the "IF-THEN" format, rule information is added: Rule type: derived rule; Category label: ME-B01; Priority: 3 (lower than manual rules 1-2 levels); Finally, the generated rule is substituted into the rule engine, and the syntax parsing is passed; Take test data (age 65 years old), the rule execution result = 5 + 0.08 × 65 = 10.2 mg / kg, the deviation from the model prediction result 10.3 mg / kg = 0.1 mg / kg ≤ 5%, there is no conflict with the original medical rules, and the verification is passed; The final derived business rule is: IF Severity of illness = 2 and patient age [60,75] patients with underlying medical history, THEN drug dosage calculation coefficient = basic drug dosage coefficient (5±0.2mg / kg) + age influence coefficient (0.08±0.005) × patient age.

[0063] Understandably, by using post-model interpretation, feature contribution ranking, and symbolic regression fitting, complex model decisions can be automatically transformed into structured mathematical expressions and standard business rules, achieving end-to-end automated transformation from model decisions to executable rules, and significantly reducing the cost of manual rule sorting and logical bias.

[0064] Step S150: Perform a difference quantification analysis on the first initial allocation scheme and the second initial allocation scheme. When the difference between the first initial allocation scheme and the second initial allocation scheme exceeds a preset threshold, generate a difference root cause diagnosis report. Specifically, in this step, firstly, the core indicators for difference quantification analysis are determined (i.e., comparable quantitative parameters in the two initial allocation schemes, such as resource quantity, allocation amount, priority, etc.). Each core indicator is assigned a corresponding weight (the total weight is 1, determined using the analytic hierarchy process, e.g., resource quantity weight = 0.4, allocation amount weight = 0.3, priority weight = 0.3). Secondly, the absolute and relative difference values ​​of each core indicator are calculated: absolute difference value = |first scheme indicator value -second scheme indicator value|, relative difference value = absolute difference value / (second scheme indicator value + 1e-6) (avoiding a denominator of 0). Then, the comprehensive difference degree is calculated as follows: (relative difference value) The sum of the corresponding weights is used to determine the overall difference. A preset threshold is set according to the business scenario (the value range is 0.1-0.3, which can be calibrated through historical data, such as a comprehensive difference threshold of 0.2). Finally, the comprehensive difference is compared with the preset threshold. If the comprehensive difference is less than or equal to the preset threshold, it is determined that there is no significant difference and no report needs to be generated. If the comprehensive difference is greater than the preset threshold, it is determined that there is a significant difference, and a root cause diagnosis report is generated (a structured report, including: core difference indicators, absolute / relative difference values ​​of each indicator, comprehensive difference, preset threshold, and preliminary description of the difference, without involving in-depth analysis of the root cause to avoid mixing in weighted features), and stored in the database for subsequent steps.

[0065] As a specific example, in the property insurance field: 1. Determine the core indicators and weights for the quantitative difference: loss assessment amount (weight = 0.4), claims specialist level (quantification code: primary = 1, intermediate = 2, advanced = 3, weight = 0.3), number of vehicles inspected (weight = 0.3); preset the comprehensive difference threshold = 0.2 (calibrated by claims data from the past 6 months to ensure that the difference of 90% of the reasonable allocation schemes is less than or equal to 0.2). 2. Calculate the difference values ​​of each indicator: (1) loss assessment amount: first scheme = 25,500 yuan, second scheme = 24,000 yuan, absolute difference value = 1,500 yuan, relative difference value = 1,500 / 24,000 = 0.0625; (2) Claims Specialist Level: Option 1 = 2 (Intermediate), Option 2 = 2 (Intermediate), Absolute Difference = 0, Relative Difference = 0; (3) Number of Vehicles Inspected: Option 1 = 1, Option 2 = 1, Absolute Difference = 0, Relative Difference = 0. 3. Calculate the overall difference = 0.0625 0.4+0 0.3+0 0.3 = 0.025, which is less than the preset threshold of 0.2, therefore no significant difference is determined, and no root cause diagnosis report is generated; (Supplement: If we assume the assessed loss amount for Option 1 = 25,500 yuan and Option 2 = 18,000 yuan, the relative difference value = 7500 / 18000 ≈ 0.4167, and the overall difference degree = 0.4167) 0.4 = 0.1668, still less than 0.2; if the second option is 15000 yuan, the relative difference is 10500 / 15000 = 0.7, and the overall difference is 0.7. 0.4 = 0.28 > 0.2, generate a root cause diagnosis report for the difference. The report content is as follows: {“Core Indicator of Difference”: “Loss Assessment Amount”, “Absolute Difference Value”: 10500, “Relative Difference Value”: 0.7, “Comprehensive Difference”: 0.28, “Preset Threshold”: 0.2, “Difference Description”: “The rule-based loss assessment amount (25500 yuan) is significantly higher than the model-based loss assessment amount (15000 yuan). The difference exceeds the preset threshold and further optimization is required”}).

[0066] As another specific example, in the field of healthcare: 1. Determine the core indicators and weights for quantifying differences: drug dosage (weight = 0.4), number of nurses (weight = 0.3), and ventilator usage status (quantification code: not used = 0, standby = 1, continuous use = 2, weight = 0.3); preset the comprehensive difference threshold = 0.15 (calibrated by the hospital's medical resource allocation data over the past 3 months to ensure the rationality of medical resource allocation). 2. Calculate the difference values ​​of each indicator: (1) Drug dosage: first plan = 700mg / time, second plan = 720mg / time, absolute difference value = 20mg / time, relative difference value = 20 / 720≈0.0278; (2) Number of nurses: first plan = 1, second plan = 1, absolute difference value = 0, relative difference value = 0; (3) Ventilator usage status: first plan = 1 (standby), second plan = 1 (standby), absolute difference value = 0, relative difference value = 0. 3. Calculate the comprehensive difference value = 0.0278 0.4+0 0.3+0 0.3 is approximately equal to 0.011, which is less than the preset threshold of 0.15. Therefore, no significant difference is determined, and no root cause diagnosis report is generated.

[0067] In some embodiments of the present invention, such as Figure 2 As shown, step S150 includes the following steps.

[0068] Step S151: Calculate the distribution distance metric between the first initial allocation scheme and the second initial allocation scheme; Specifically, in this step, firstly, the core quantitative indicators in the first and second initial allocation schemes are standardized (e.g., using the min-max standardization method to map all indicator values ​​to the [0,1] interval, the formula being: Standardized value = (Actual indicator value - Minimum indicator value) / (Maximum indicator value - Minimum indicator value), avoiding interference from indicators of different magnitudes on distance calculation); secondly, a commonly used distribution distance metric in this field, Euclidean distance, is selected as the core calculation method (simple to calculate, highly interpretable, and suitable for measuring differences in initial allocation schemes). Cosine similarity can also be used as an auxiliary method (to verify the rationality of the distance metric; the closer the cosine similarity is to 1, the smaller the difference; the closer it is to 0, the larger the difference); finally, the Euclidean distance value and cosine similarity value are calculated and stored in the database as the distribution distance metric results for subsequent contribution analysis, ensuring the reproducibility of the calculation process. The maximum and minimum values ​​of the indicators required for standardization are obtained through statistical analysis of allocation scheme data from the past 3-6 months and are updated regularly (once a month).

[0069] As a specific example, in the property insurance field, based on two initial allocation schemes (the first scheme has a loss assessment amount of RMB 25,500 and the second scheme has RMB 15,000; the claims specialist level is 2 for both; and the number of survey vehicles is 1 for both), the following steps are executed: 1. Determine the core quantitative indicators: loss assessment amount, claims specialist level (quantitative code 1-3), and number of survey vehicles (quantitative code 1-3), a total of 3 core indicators; 2. Indicator standardization: by statistically analyzing the claims data of the past 6 months, the maximum and minimum values ​​of each indicator are obtained: the minimum loss assessment amount = RMB 5,000 and the maximum value = RMB 50,000; the minimum claims specialist level = 1 and the maximum value = 3; the minimum number of survey vehicles = 1 and the maximum value = 3; (1) Standardization of loss assessment amount: the first scheme = (25,500-5,000) / (50,000-5,000) =20500 / 45000≈0.4556; Second scheme =(15000-5000) / 45000=10000 / 45000 is about 0.2222; (2) Standardization of claims specialist level: First scheme =(2-1) / (3-1)=0.5; Second scheme =(2-1) / (3-1)=0.5; (3) Standardization of the number of survey vehicles: First scheme =(1-1) / (3-1)=0; Second scheme =(1-1) / (3-1)=0; 3. Calculate the distribution distance measure: (1) Euclidean distance is about 0.2335; (2) Cosine similarity is about 0.952; 4. Result storage: The Euclidean distance of about 0.2335 and the cosine similarity of about 0.952 are used as the distribution distance measure results and stored in the claims database for subsequent analysis.

[0070] As a specific example, in the field of medical and health care, based on the two initial allocation schemes in the supplementary example of medical and health care in claim 1 (first scheme drug dosage = 700mg / time, second scheme = 1200mg / time; the number of nurses is 1; the ventilator usage status is 1), the following steps S151 can be performed: 1. Determination of core quantitative indicators: drug dosage, number of nurses (quantitative code 1-2), ventilator usage status (quantitative code 0-2), a total of 3 core indicators; 2. Standardization of indicators: by statistically analyzing the hospital drug use and resource allocation data over the past 3 months, the maximum and minimum values ​​of each indicator are obtained: minimum drug dosage = 500mg / time, maximum = 1500mg / time; minimum number of nurses = 1, maximum = 2; minimum ventilator usage status = 0, maximum = 2; (1) Standardization of drug dosage: first Scheme 1 = (700-500) / (1500-500) = 200 / 1000 = 0.2; Scheme 2 = (1200-500) / 1000 = 700 / 1000 = 0.7; (2) Standardization of nurse number: Scheme 1 = (1-1) / (2-1) = 0; Scheme 2 = (1-1) / (2-1) = 0; (3) Standardization of ventilator usage status: Scheme 1 = (1-0) / (2-0) = 0.5; Scheme 2 = (1-0) / (2-0) = 0.5; 3. Calculate the distribution distance measure: (1) Euclidean distance is about 0.5; (2) Cosine similarity is about 0.851; 4. Result storage: The Euclidean distance of about 0.5 and the cosine similarity of about 0.851 are used as the distribution distance measure results and stored in the hospital resource management database for subsequent analysis.

[0071] Step S152: Analyze at least one business rule and core decision feature of the model that contributes the most to the distribution distance metric; Specifically, in this step, firstly, the single-indicator contribution analysis method is used to calculate the contribution of each core quantitative indicator to the distribution distance metric (Euclidean distance). The contribution calculation formula is: Single indicator contribution = (Standardized difference value of the indicator²) / Euclidean distance² 100% (Standardized Difference Value = |Standardized Index Value of First Scheme - Standardized Index Value of Second Scheme|), the higher the contribution, the greater the impact on the distribution distance; secondly, screen out core indicators with a contribution greater than or equal to 50% (determined as the indicators with the greatest contribution to the distribution distance; if multiple indicators have a contribution greater than or equal to 50%, all are screened); then, for the screened core indicators, reverse-associate them with the corresponding business rules in the target business rule set (i.e., which / which business rules directly determine the core indicator), and simultaneously extract the core decision features that have the greatest impact on the core indicator from the target machine learning model (through model feature importance analysis, calculated using Gini coefficient or SHAP value; features with an importance greater than or equal to 30% are determined as core decision features); finally, determine the business rules (1-2) and model core decision features (1-2) that contribute the most to the distribution distance measurement, forming an association list to ensure that each association is traceable and verifiable.

[0072] As a specific example, in the property insurance field, based on the distribution distance measurement result calculated in step S151 (Euclidean distance is approximately 0.2335), step S152 can be executed: 1. Calculate the contribution of each core indicator: (1) Loss assessment amount: the standardized difference value is approximately 0.2334, and the contribution is approximately 99.92%; (2) Claims specialist level: the standardized difference value = 0, and the contribution = 0%; (3) Number of vehicles inspected: the standardized difference value = 0, and the contribution = 0%; 2. Select the core indicator with the largest contribution: the loss assessment amount (contribution 99.92% greater than 50%), which is the only indicator with the largest contribution; 3. Related business rules: the loss assessment amount is directly determined by rule 2 of the target business rule set (rule 2: 10,000 yuan < reported amount less than 50,000 yuan AND 4. Extract core decision features of the model: Perform feature importance analysis on the target machine learning model (XGBoost algorithm), and use Gini coefficient to calculate the importance of each input feature: reported amount (Gini coefficient = 0.42), number of accidents (0.21), latitude and longitude of the reporting location (0.20), and reporting time (0.17); Select the core decision feature with a feature importance greater than or equal to 30% as "reported amount" (Gini coefficient = 0.42); 5. Form an association list: the business rule with the greatest contribution = rule 2 (85% of the loss assessment amount), the core decision feature of the model with the greatest contribution = reported amount, and store it in the database for use in subsequent steps.

[0073] As a specific example, in the field of healthcare, based on the distribution distance metric calculated in step S151 (Euclidean distance = 0.5), step S152 is executed: 1. Calculate the contribution of each core indicator: (1) Drug dosage: Standardized difference value = |0.2-0.7| = 0.5, contribution = (0.5²) / (0.5²) × 100% = 100%; (2) Number of nurses: Standardized difference value = 0, contribution = 0%; (3) Ventilator usage status: Standardized difference value = 0, contribution = 0%; 2. Screen the core indicator with the largest contribution: Drug dosage (contribution = 100% ≥ 50%), which is the only indicator with the largest contribution; 3. Associate business rules: Drug dosage is directly determined by rule 2 of the target business rule set (rule 2: severity of illness = 2 AND 4. Extract core decision features of the model: Perform feature importance analysis on the target machine learning model (neural network), and use SHAP value to calculate the importance of each input feature: patient weight (SHAP value = 0.38), severity of illness (0.29), age (0.21), and basic medical history (0.12); Select the core decision feature with a feature importance greater than or equal to 30% as "patient weight" (SHAP value = 0.38); 5. Form an association list: the business rule with the greatest contribution = rule 2 (drug dosage calculation coefficient 10mg / kg), the core decision feature of the model with the greatest contribution = patient weight, and store it in the database for subsequent steps.

[0074] Step S153: Associate the business rules with the core decision features of the model, and generate the root cause diagnosis report describing the deviation between the business rules and the core decision features of the model.

[0075] Specifically, in this step, firstly, the association list generated in step S152 is retrieved to identify the business rule (including its core adjustable parameters) that contributes the most and the core decision features of the model, establishing a correlation between the two, and focusing on analyzing the deviations between "the parameter settings of the business rules" and "the influence logic of the model's core decision features"; secondly, the specific manifestations of the deviations are clarified: it is clarified that the business rules output core indicator results based on fixed parameters (such as proportions and coefficients), while the model's core decision features are dynamic influence logic formed based on historical data training, and the deviation between the two is mainly reflected in "the fixed parameters of the business rules not being fully adapted to the influence weights of the model's core decision features"; then, combined with the distribution distance measurement results from step S151 and step S15... The contribution analysis in step 2 supplements the quantitative data related to deviations (such as the standardized difference value and contribution of core indicators) to clarify the specific impact of deviations on the distribution distance measurement. Finally, a structured root cause diagnosis report of deviations is generated. The report should include five core modules: basic information (distribution distance measurement results, core indicators with the greatest contribution), a list of related factors (business rules, core decision features of the model), a detailed description of deviations (deviation points and manifestations between business rules and core decision features of the model), quantitative data of deviations, and an explanation of the impact of deviations (clarifying the logic that the deviation leads to differences in core indicators and thus increases the distribution distance measurement). The report format is consistent with step S150 and is stored in the database for subsequent steps to call, ensuring that the report can be directly used for rule parameter optimization.

[0076] As a specific example, in the property insurance field, based on the related list formed in step S152 (business rule = rule 2, core parameter = loss assessment ratio 85%; model core decision feature = reported amount, Gini coefficient = 0.42), combined with the distribution distance measurement result of step S151 (Euclidean distance ≈ 0.2335), step S153 is executed: 1. Establish the relationship: It is clear that the core parameter of business rule 2, "loss assessment ratio 85%", directly determines the loss assessment amount (core indicator), and the model core decision feature "reported amount" is the most critical factor affecting the model's output loss assessment amount (feature importance 42%). The two work together on the same core indicator and are directly related; 2. Analyze the deviation points: Business rule 2 uses a fixed 85% loss assessment ratio, which is only set according to "reported amount range + accident type", without considering the specific numerical differences in the reported amount (i.e., it does not adapt to the dynamic influence weight of "reported amount" in the model); while the model The core decision feature, "reported amount," dynamically adjusts the loss assessment amount based on specific values ​​(e.g., 30,000 yuan, 15,000 yuan), leading to a significant difference between the two output loss assessment amounts, thus increasing the distribution distance metric; 3. Supplementing deviation quantification data: The standardized difference value of the core indicator (loss assessment amount) is approximately 0.2334, contributing approximately 99.92% to the Euclidean distance, with an Euclidean distance of approximately 0.2335, exceeding the preset difference threshold corresponding to the Euclidean distance critical value (combined with the previously preset comprehensive difference threshold of 0.2, converted to an Euclidean distance critical value of approximately 0.18); 4. Generating a root cause diagnosis report: {"Basic Information":{"Distribution Distance Metric":{"Euclidean Distance":0.2335,"Cosine Similarity":0.952},"Core Indicator with the Largest Contribution":"Loss Assessment Amount"},"Related List":{"Business Rule with the Largest Contribution":"Rule 2 (10,000 yuan < reported amount less than 50,000 yuan AND..." "Claim Type = 2, Loss Assessment Ratio 85%", "Core Decision Feature of the Model": "Reported Amount (Gini Coefficient = 0.42)"}, "Specific Description of Deviation": "Business Rule 2 uses a fixed 85% loss assessment ratio, set only based on the reported amount range and claim type, failing to adapt to the dynamic influence weight of the model's core decision feature 'Reported Amount'; the model dynamically adjusts the loss assessment ratio based on the specific value of the reported amount, resulting in a significant difference between the two output loss assessment ratios." Deviation Quantification Data: {"Standardized Difference Value of Loss Assessment Amount": 0.2334, "Contribution of Core Indicators": 99.92%, "Euclidean Distance": 0.2335}, "Explanation of Deviation Impact": "This deviation directly leads to a significant difference in the loss assessment amount, which is the core reason why the distribution distance measurement exceeds the critical value. It is necessary to optimize the loss assessment amount ratio of Rule 2, adapt to the dynamic impact logic of the reported amount in the model, and reduce the difference between the two."} The report is stored in the claims database for use in step S170.

[0077] As another concrete example, in the healthcare field, based on the association list formed in step S152 (business rule = rule 2, core parameter = drug dosage calculation coefficient 10mg / kg; model core decision feature = patient weight, SHAP value = 0.38), combined with the distribution distance measurement result (Euclidean distance = 0.5) from step S151, step S153 can be executed: 1. Establish the association: It is clear that the core parameter of business rule 2, "drug dosage calculation coefficient 10mg / kg", directly determines the drug dosage (core indicator), and the model core decision feature "patient weight" is the most critical factor affecting the drug dosage output by the model (feature importance 38%). Both work together on the same core indicator and are directly related; 2. Analyze the deviation points: Business rule 2 uses a fixed calculation coefficient of 10mg / kg, which is only set according to "severity of illness + diagnosed disease", and does not fully adapt to the dynamic influence weight of "patient weight" in the model ( The model adjusts the dosage coefficient based on the patient's specific weight value, rather than using a fixed value; the model's core decision feature, "patient weight," dynamically optimizes the drug dosage based on a specific value (e.g., 70kg), leading to a significant difference in the drug dosage outputs of the two models, thus increasing the distribution distance metric; 3. Supplementing deviation quantification data: The standardized difference value of the core indicator (drug dosage) = 0.5, its contribution to Euclidean distance = 100%, and its Euclidean distance = 0.5, exceeding the preset difference threshold corresponding to the critical Euclidean distance value (combined with the previously preset comprehensive difference threshold of 0.15, the critical Euclidean distance value is approximately 0.13); 4. Generating a root cause diagnosis report: {"Basic Information":{"Distribution Distance Metric":{"Euclidean Distance":0.5,"Cosine Similarity":0.851},"Maximum Contributing Core Indicator":"Drug Dosage"},"Association List":{"Maximum Contributing Business Rule":"Rule 2 (Severity of Illness = 2)} AND Confirmed Disease Type = 2, Drug Dosage Calculation Coefficient 10mg / kg), “Model Core Decision Feature”: “Patient Weight (SHAP value = 0.38)”, “Specific Description of Deviation”: “Business Rule 2 uses a fixed drug dosage calculation coefficient of 10mg / kg, which is set only based on the severity of the illness and the confirmed disease type. It does not adapt to the dynamic influence weight of the model's core decision feature, “Patient Weight”; the model dynamically adjusts the drug dosage according to the specific value of the patient's weight, resulting in a significant difference between the two output drug dosages”, “Quantitative Data of Deviation”: {“Standardized Difference Value of Drug Dosage”: 0.5, “Contribution of Core Indicator”: 100%, Euclidean Distance: 0.5}, “Explanation of the Impact of Deviation”: This deviation directly leads to a significant difference in drug dosage and is the core reason why the distribution distance measurement exceeds the critical value. It is necessary to optimize the drug dosage calculation coefficient of Rule 2 to adapt to the dynamic influence logic of patient weight in the model and reduce the difference between the two}. The report is stored in the hospital resource management database for use in step S170.

[0078] It is understandable that the above embodiments quantify the differences between the two allocation schemes by using the distribution distance metric, and locate the business rules and core decision features of the model that contribute the most to the distribution distance. This upgrades the traditional method of simply comparing the differences in results to root cause localization at the level of rule logic and model features. This achieves a breakthrough from numerical comparison to mechanism-level diagnosis of differences, and can accurately and quantitatively identify the essential reasons for the discrepancies between the model and the rules, rather than just superficial result deviations.

[0079] In some embodiments of the present invention, the following steps are included after step S150.

[0080] Step S161: Based on the differential root cause diagnosis report, when it is found that there is a discrepancy between the first initial allocation scheme and the second initial allocation scheme, and the business rule corresponding to the first initial allocation scheme is logically correct, the first initial allocation scheme, the second initial allocation scheme, the description information of the item to be allocated, and the differential root cause diagnosis report are combined with the logically correct business rule to form a rule verification sample. Specifically, in this step, a sample screening module can be deployed to call the differential root cause diagnosis report generated in step S150. First, it is determined whether there is a discrepancy between the first initial allocation scheme (generated by the rule engine) and the second initial allocation scheme (generated by the target machine learning model). The discrepancy judgment criteria are: the deviation of the core allocation parameters (such as resource quantity, parameter value, priority) of the two schemes is >5%. Second, through the business rule verification module, it is verified whether the logic of the business rules corresponding to the first initial allocation scheme (rules in the pre-configured target business rule set) is correct. The verification content includes: no errors in rule syntax, no logical contradictions, and compliance with business constraints. Business personnel assist in the confirmation (automatic verification + manual review dual mechanism can be configured). More specifically, when both "disagreement exists" and "the first scheme corresponds to the correct rule logic" are met, the sample components are organized to ensure that the components are complete and can be used for model training and verification: (1) Description information of the items to be assigned: complete original descriptions (such as property insurance claim information, medical patient information) and standardized feature vectors (features extracted in step S141); (2) First initial assignment scheme: complete assignment scheme output by the rule engine, with the corresponding business rule number and rule content marked; (3) Second initial assignment scheme: complete assignment scheme output by the target machine learning model, with the parameters and feature contribution of the model prediction marked; (4) Root cause diagnosis report: complete root cause analysis results, with key points of disagreement (such as which assignment parameter has a large deviation) and root cause of disagreement (such as the model not recognizing a certain core business constraint); 5) Logically correct business rules: The business rules corresponding to the first initial allocation scheme are marked with rule priority and core constraints. More specifically, the above 5 types of elements are organized in a structured format of "sample number - information to be allocated - first scheme - second scheme - root cause of divergence - correct business rule", and each sample is assigned a unique number (format: Y-business domain-serial number, such as Y-CL-001), and the sample generation time and divergence type are marked to ensure that the sample format is uniform and can be directly imported into the model training dataset. Finally, the constructed rule verification samples are verified to ensure that there are no missing elements, the divergence judgment is accurate, and the business rule logic is correct. After the verification is passed, they are stored in the rule verification sample library for subsequent steps to call. If the verification fails, the divergence judgment result or the correctness of the business rule is rechecked, the elements are supplemented and improved, and then the verification is performed again.

[0081] As another specific example, in the property insurance field, the first step is to determine the discrepancy and the correctness of the rules: (1) Discrepancy determination: For a certain item to be allocated (reported amount of RMB 25,000, type of accident = 2, number of accidents = 2, non-operation), the loss assessment ratio of the first initial allocation scheme (generated by the rule engine) is 82%, and the loss assessment ratio of the second initial allocation scheme (generated by the XGBoost model) is 75%. The deviation is 7% > 5%, and it is determined that there is a discrepancy; (2) Rule correctness verification: The business rule corresponding to the first scheme (number R-CL-003, content: reported amount) (20000,30000) yuan and the number of accidents = 2 times, the loss assessment ratio is 82%±2%). After automatic verification by the business rule verification module and manual review, there are no grammatical errors and no logical contradictions. It meets the property insurance claim constraints and the judgment logic is correct. Next, the sample elements are sorted out: (1) Description information of the items to be allocated: original information (reported amount 25000 yuan, accident type = 2 (collision), number of accidents = 2 times, vehicle use nature = non-commercial), standardized feature vector [0.4444,2,0.5,1] (reported amount standardization: (25000-5000) / (50000-5000) is about 0 .4444); (2) First initial allocation scheme: loss assessment ratio 82%, 1 intermediate claims specialist, 1 inspection vehicle, corresponding rule number R-CL-003; (3) Second initial allocation scheme: loss assessment ratio 75%, 1 intermediate claims specialist, 1 inspection vehicle, model prediction parameters (iteration number 50, convergence threshold 0.001); (4) Root cause diagnosis report: the point of divergence is the loss assessment ratio (deviation 7%), the root cause is that the model did not fully identify the impact of "number of accidents = 2 times" on the loss assessment ratio, and did not match the corresponding business rule R-CL-003; (5) Correct business rule: R-CL-003 (IF The reported amount is ∈ (20000, 30000) yuan and the number of claims is 2, THEN loss assessment ratio is 82% ± 2%; furthermore, sample standardization can be performed: sample number Y-CL-001, generation time 2024-05-21, the discrepancy type is loss assessment ratio deviation, and it is organized in a structured format; finally, sample verification can be performed: verify that there are no missing elements, the discrepancy judgment is accurate, and the rule logic is correct, and store it in the property insurance rule verification sample library.

[0082] As another specific example, in the field of medical and health care, the first step is to determine the discrepancy and the correctness of the rules: (1) Discrepancy determination: For a certain item to be allocated (patient age 68 years, disease severity = 2, history of diabetes, weight 72kg), the drug dosage calculation coefficient of the first initial allocation scheme (generated by the rule engine) is 10.4mg / kg, and the drug dosage calculation coefficient of the second initial allocation scheme (generated by the LightGBM model) is 9.8mg / kg. The deviation is 0.6mg / kg > 5% (relative deviation is approximately 5.8%), and it is determined that there is a discrepancy; (2) Correctness of the rules: The business rule corresponding to the first scheme (number R-ME-007, content: age greater than or equal to 65 years and disease severity = 2 and history of diabetes, drug dosage calculation coefficient 10.4±0.2mg / kg) has been verified. The syntax is correct and conforms to the medical drug use standard. The judgment logic is correct; Further, the sample elements can be sorted out: (1) Description information of the item to be allocated: Original information (age 68 years, diagnosed disease = 2) (hypertensive emergency), severity of illness = 2, underlying medical history = 1 (diabetes), weight 72kg), standardized feature vector [0.8065, 2, 1, 0.2] (age standardization: (68-18) / (80-18) approximately equals 0.8065); (2) first initial allocation scheme: drug dosage calculation coefficient 10.4mg / kg, 1 intensive care bed, 1 nurse, corresponding rule number R-ME-007; (3) second initial allocation scheme: drug dosage calculation coefficient 9.8 mg / kg, 1 intensive care bed, 1 nurse, model prediction parameters (iteration number 50, convergence threshold 0.001); (4) Root cause diagnosis report: the divergence point is the drug dosage calculation coefficient (deviation 0.6mg / kg), the root cause is that the model calculates the feature weight of "age greater than or equal to 65 years old" too low, and does not match the corresponding business rule R-ME-007; (5) Correct business rule: R-ME-007 (IF age ≥ 65 years old and disease severity = 2 and have underlying medical history, THEN drug dosage calculation coefficient 10.4±0.2mg / kg); further, sample standardization can be performed: sample number Y-ME-001, generation time 2024-05-21, divergence type is drug dosage coefficient deviation, organized in structured format; finally, sample verification can be performed: verify that each element is not missing, the divergence judgment is accurate, the rule logic is correct, and store it in the medical rule verification sample library.

[0083] Step S162: Add the rule verification sample to the training dataset of the target machine learning model; Specifically, in this step, the training dataset can be preprocessed first. The existing training dataset of the target machine learning model (such as XGBoost or LightGBM) can be called, and the dataset can be standardized to ensure that the data format is consistent with the rule validation sample format. This includes: feature vector standardization (consistent with the standardization method in step S141), label unification (using the core parameters of the first initial allocation scheme as the correct labels, since the first scheme corresponds to the correct rule logic), and missing value handling (using mean imputation or interpolation, consistent with the processing method of the original dataset). Further, when adding samples, the "append" method can be used to add the rule validation samples constructed in step S161 to the preprocessed training dataset. When adding, it is necessary to ensure that the feature dimensions and label types of the samples are consistent with the original dataset to avoid data format conflicts. At the same time, the added validation samples are labeled "validation sample" to distinguish them from the original training samples, facilitating targeted optimization during subsequent model training. More specifically, after adding samples, the training, validation, and test sets can be re-divided, maintaining the same proportions as before (e.g., 70% training, 20% validation, and 10% test). This ensures that rule validation samples are evenly distributed across the training and validation sets (validation samples should account for 5%-10% of the total training set samples to avoid overfitting due to an excessively high proportion). Furthermore, the updated training dataset can be validated to check if the samples were added completely, if the feature dimensions are consistent, if the labels are correct, and if there are no duplicate samples. After successful validation, the index of the model training dataset is updated and synchronized to the training module of the target machine learning model, ensuring that the dataset can be used during subsequent training. Simultaneously, the sample addition time, sample quantity, and dataset version are recorded for easy tracking and rollback later.

[0084] As a concrete example in the property insurance field, the existing training dataset of the XGBoost claims resource allocation model (containing 1000 property insurance claims samples) can be used. The feature vector of the rule verification sample (Y-CL-001) is validated using the standardization method consistent with step S141 to ensure consistency with the feature dimensions of the original dataset (4 features: reported amount, accident type, number of accidents, and vehicle usage). The correct label is set to the loss assessment ratio of 82% for the first scheme, with no missing values, completing the preprocessing. Further, the Y-CL-001 sample can be added to the existing training dataset using an append-only method, labeled as a "verification sample," ensuring the data format is consistent with the original sample (feature vector + correct label + auxiliary information). Further, the original dataset can be divided into a training set of 70% (700 samples), a validation set of 20% (200 samples), and a test set of 10% (100 samples). After adding one validation sample, the total number of samples is 1001. After the re-division, the training set has 701 samples (including the validation sample), the validation set has 200 samples, and the test set has 100 samples. The validation sample accounts for approximately 0.14% of the training set, which meets the requirement of 5%-10%. Finally, the validation samples can be added completely, with consistent feature dimensions, correct labels, and no duplicate samples. The training dataset index is updated, synchronized to the XGBoost model training module, and the addition time (2024-05-21), the number of samples added (1), and the dataset version (V2.1) are recorded and stored in the property insurance model training database.

[0085] As another concrete example in the healthcare field, the first step is dataset preprocessing. The existing training dataset of the LightGBM medical resource allocation model (containing 800 patient resource allocation samples) is used. The feature vector of the rule validation sample (Y-ME-001) is verified using the standardization method consistent with step S141 to ensure consistency with the feature dimensions of the original dataset (4 features: age, disease severity, basic medical history, and weight). The correct label is 10.4 mg / kg, calculated from the drug dosage of the first scheme, with no missing values. Preprocessing is then complete. Further, the Y-ME-001 sample can be appended to the existing training dataset, labeled as a "validation sample," to ensure the data... The format is consistent with the original samples. Further, the original dataset can be divided into a training set of 70% (560 samples), a validation set of 20% (160 samples), and a test set of 10% (80 samples). After adding one validation sample, the total number of samples is 801. After re-dividing, the training set has 561 samples (including validation samples), the validation set has 160 samples, and the test set has 80 samples. The validation samples account for approximately 0.18% of the training set, which meets the requirements. Finally, the samples are verified to be complete, with consistent feature dimensions, correct labels, and no duplicate samples. The training dataset index is updated and synchronized to the LightGBM model training module, recording the addition time as 2024-05-21, the number of samples added as 1, and the dataset version as V3.1, and stored in the medical model training database.

[0086] Step S163: In the subsequent training process of the target machine learning model, a rule consistency loss term is introduced, wherein the rule consistency loss term is calculated based on the predicted output of the target machine learning model according to the business rules, and is used to penalize the deviation of the output of the target machine learning model from the logic of the business rules; Specifically, in this step, the rule consistency loss term is defined based on logically correct business rules (the correct rules in step S161). The core function of the rule consistency loss term is to penalize the deviation between the predicted output of the target machine learning model and the logical business rules. The larger the loss value, the more severe the deviation between the model output and the business rules. The specific definition is as follows: Rule Consistency Loss Term ;in: This is the penalty coefficient (which can be dynamically adjusted, ranging from 0.5 to 2.0, and calibrated according to the business scenario, with 1.0 being the preferred value to ensure a moderate penalty). The predicted output of the target machine learning model (the core parameter of the second initial allocation scheme); The output corresponding to the logically correct business rules (the core parameters of the first initial allocation scheme); The absolute deviation between the model output and the rule output; To mitigate relative deviations and avoid inaccurate loss calculations due to differences in parameter magnitudes, a rule consistency loss calculation module is deployed, linked to the target machine learning model training module and the rule validation sample library. This module automatically calls the "correct business rule output y_rule" and "model prediction output y_pred" from the rule validation samples and calculates the rule consistency loss term L_rule in real time according to the formula described above. Furthermore, the rule consistency loss term can be weighted and combined with the original loss function of the target machine learning model (such as mean squared error (MSE) and cross-entropy loss) to form a new total loss function. This ensures that the prediction error and deviation from the business rules are minimized simultaneously during model training. The total loss function formula is as follows: Total Loss ;in: The original loss function for the model (e.g., MSE for regression tasks and cross-entropy loss for classification tasks); This is the rule consistency loss term; further, the loss term parameters can be calibrated, and the penalty coefficient can be adjusted according to the business scenario and model training effect. If the model output deviates too much from the business rules, the value can be increased. (e.g., adjust to 1.5); if the model overfits to business rules, the value can be reduced. (For example, adjust to 0.8) to ensure that the model conforms to business rules and logic while also possessing generalization ability; at the same time, record the loss term parameters ( (Value selection and calculation methods) to ensure the training process is reproducible.

[0087] As a concrete example, in the property insurance field, for property insurance claims scenarios, the target machine learning model (XGBoost) is a regression task (predicting the loss assessment ratio), the original loss function is the mean squared error (MSE), and the rule consistency loss term is defined as follows: ;in: =1.0 (initial penalty coefficient); The loss percentage predicted by the model (e.g., 75%). The loss assessment percentage for correct business rules (e.g., 82%) is then calculated. Further, loss items are calculated, taking rule verification sample Y-CL-001 as an example. =75%, =82%, substitute into the formula to calculate: =1.0 × |75% - 82%| / 82% is approximately equal to 0.085; further, the original loss =MSE( , ),in The actual reasonable loss assessment ratio (and) Consistent (82%), MSE = (75% - 82%)² = 0.0049; Total loss =0.0049 + 0.085 is approximately equal to 0.0899; furthermore, the initial... =1.0. During training, it was found that the model output still deviated somewhat from the business rules. λ was adjusted to 1.2, and L_rule was recalculated as 1.2 × 0.07 / 0.82, approximately equal to 0.102. The total loss was... =0.0049 + 0.102 is approximately equal to 0.1069, which strengthens the penalty and ensures that the model conforms to the business rules; record. The value is 1.2, the calculation method is 1.2, and the result is stored in the model training parameter library.

[0088] As another concrete example, in the healthcare field, specifically for the scenario of medical resource allocation, the target machine learning model (LightGBM) is a regression task (predicting drug dosage calculation coefficients), the original loss function is the mean squared error (MSE), and a rule consistency loss term is defined: in: =1.0 (initial penalty coefficient); Calculate the coefficient for the drug dosage predicted by the model (e.g., 9.8 mg / kg); Calculate the drug dosage coefficient for the correct business rule (e.g., 10.4 mg / kg); further, taking rule validation sample Y-ME-001 as an example, =9.8mg / kg, =10.4mg / kg, substitute into the formula to calculate: =1.0×|9.8-10.4| / 10.4 is approximately equal to 0.058; further, the original loss =MSE(y_pred,y_true), y_true=10.4mg / kg, MSE=(9.8-10.4)²=0.36; Total loss =0.36 + 0.058 is approximately equal to 0.418; furthermore, during parameter calibration, the initial... =1.0, during training, the model's generalization ability was found to be normal, and the deviation from the business rules was small, maintaining... =1.0 remains unchanged; record The value is 1.0, the calculation method is used, and the result is stored in the model training parameter library.

[0089] In some embodiments of the present invention, after step S163 and before step S165, the following steps are further included.

[0090] Step S1641: Extract the target business rule that caused the divergence from the root cause diagnosis report; Specifically, in this step, the root cause diagnosis report can be read to locate the discrepancy between the first and second initial allocation schemes. This triggers the business rule number, rule conditions, and rule output of the first scheme, and marks that rule as the target business rule—the rule whose model prediction is inconsistent with the rule, but whose logic is correct. The extracted content can include rule conditions (feature combinations, thresholds, intervals), rule outputs (allocation results / scores / coefficients), and rule logic (greater than, less than, belongs to, and, or).

[0091] As a specific example, in the property insurance field, the root cause diagnosis report pointed out the discrepancy reason: the model did not comply with rule R-CL-003: IF The reported amount belongs to (20000, 30000) and the number of claims = 2, THEN The loss assessment ratio = 82%; Extract the target business rule: R-CL-003: Condition: Reported amount 20000–30000 yuan, number of claims = 2, Output: Loss assessment ratio 82%.

[0092] As another concrete example, in the healthcare field, the root cause diagnosis report indicates the discrepancy is due to the model not adhering to rule R-ME-007: IF age ≥ 65 and disease severity = 2 and underlying medical history, THEN drug dosage coefficient = 10.4 mg / kg. The extracted target business rule is: R-ME-007: Conditions: age greater than or equal to 65, disease severity = 2, underlying medical history, output: drug dosage coefficient 10.4.

[0093] Step S1642: Transform the target business rule into a differentiable constraint. Specifically, in this step, discrete / logic rules can be transformed into a continuous form that is differentiable and can be placed into a loss function. The rule output is y_rule, the model output is y_pred, and the constraint objective is to make y_pred as close as possible to y_rule. A differentiable constraint term is constructed: Lconstraint=|y_pred−y_rule| / y_rule. This expression is differentiable, smooth, and sensitive to relative deviations, making it suitable as a constraint.

[0094] As a concrete example, in the property insurance field, the target rule output y_rule=82%, the model output y_pred, and the differentiability constraint Lconstraint=∣y_pred−0.82∣ / 0.82.

[0095] As another concrete example, in the field of healthcare, the target rule output is y_rule=10.4, and the differentiable constraint is Lconstraint=|y_pred-10.4| / 10.4.

[0096] Step S1643: Take the constraint condition as a regularization term and combine it with the original loss function of the target machine learning model through weighting to form a new composite loss function. Specifically, in this step, the original loss Loriginal = (y_pred - y_true) 2 , introduce a weight coefficient a (0 < a ≤ 2, usually 1.0), and construct the composite loss Lcomposite = Loriginal + a ⋅ Lconstraint, which means reducing the prediction error while punishing the deviation from the business rules.

[0097] As a specific example, in the property insurance field, take a = 1.0, Lcomposite = (y_pred - 0.82) 2 + 1.0 |y_pred - 0.82| / 0.82.

[0098] As another specific example, in the medical and health field, take a = 1.0, Lcomposite = (y_pred - 10.4) 2 + 1.0 |y_pred - 10.4| / 10.4.

[0099] Step S1644: Train the target machine learning model based on the new composite loss function, so that the target machine learning model minimizes the prediction error and minimizes the deviation between the output of the target learning model and the target business rule.

[0100] Specifically, in this step, the Adam / SGD optimizer can be used with the composite loss Lcomposite as the optimization target. The training objective can be to minimize the prediction error while minimizing the deviation between the model output and the target business rule. The convergence condition can be that the composite loss drops by < 0.001 for several consecutive rounds, and the relative deviation between the model output and the rule output is < 5%.

[0101] As a specific example, in the property insurance field, before model training, y_pred = 75%, deviating from the rule by 82% which is relatively large. After training, y_pred converges to 81.5% - 82.5%, meeting the rule constraints while maintaining the overall prediction accuracy.

[0102] As another specific example, in the medical and health field, before model training, y_pred = 9.8, deviating from the rule by 10.4. After training, y_pred stabilizes at 10.3 - 10.5, conforming to the medical rules, and the prediction error decreases synchronously.

[0103] It is understandable that the above embodiments transform the target business rules into differentiable constraints and embed them into a composite loss function, so that the business rules can directly participate in the model gradient optimization in the form of mathematical constraints. This enables the model training to simultaneously meet the requirements of prediction accuracy and rule compliance, thus solving the technical pain point that traditional models have difficulty in forcibly adhering to hard rules.

[0104] Step S165: Based on the total loss function that incorporates the rule consistency loss term, train and optimize the target machine learning model until the output value of the target machine learning model satisfies the error.

[0105] Specifically, in this step, the training parameters of the target machine learning model can be configured based on the updated training dataset (including rule validation samples) and the total loss function to ensure that the training process is stable and convergent. The parameter configuration is as follows: (1) Number of iterations: 50-200 times (adjusted according to the model complexity, preferably 100 times to avoid overfitting or underfitting); (2) Learning rate: 0.01-0.1 (preferably 0.05 to ensure that the training convergence speed is moderate); (3) Convergence threshold: 0.001 (when the change in the total loss function is less than or equal to 0.001, the model is judged to have converged); (4) Regularization parameter: L2 regularization coefficient 0.01-0.1 (preferably 0.05 to prevent the model from overfitting); (5) Optimizer: select gradient descent optimizer (such as SGD, Adam, preferably Adam, which has a faster convergence speed and better stability). Furthermore, during model training, the target machine learning model training module is started, the updated training dataset, total loss function, and configured training parameters are called to begin model training. During the training process, the changing trends of total loss L_total, original loss L_original, and rule consistency loss L_rule are monitored in real time, and the loss value is recorded every 10 iterations to ensure that the training process is traceable. Furthermore, model training is terminated when any of the following conditions are met: (1) the preset number of iterations (e.g., 100) is reached; (2) the change in the total loss function is less than or equal to the convergence threshold (0.001) for 5 consecutive iterations, and the model tends to stabilize; (3) The model's prediction error (e.g., MSE) on the validation set is less than or equal to a preset error threshold (0.005), meeting business requirements. Further, after model training terminates, the optimized target machine learning model is obtained and deployed to the test environment. Test set data (including test data corresponding to rule validation samples) is used for validation. Validation metrics include: the deviation between the model's predicted output and the business rule output is less than or equal to 5%, the model accuracy is greater than or equal to 85%, and the F1 score is greater than or equal to 85%. Once validation is passed, it is determined as the final optimized model. Further, the optimized model replaces the original target machine learning model, the model version is updated (e.g., XGBoost V2.2, LightGBM V3.2), training parameters, total loss function, and validation results are recorded and stored in the model database. Simultaneously, the data is synchronized to the business system to ensure that the optimized model is used during subsequent resource allocation.

[0106] As a concrete example in the property insurance field, the training parameters can be configured as follows: XGBoost model training parameters, number of iterations = 100, learning rate = 0.05, convergence threshold = 0.001, L2 regularization coefficient = 0.05, optimizer = Adam; during model training, XGBoost model training is started, the updated training dataset (including Y-CL-001 samples) and the total loss function L_total = MSE + L_rule are called, and the loss changes are monitored in real time. The loss value is recorded every 10 iterations. For example: the total loss is 0.08 in the 10th iteration, 0.03 in the 50th iteration, 0.008 in the 80th iteration, and so on. The total iterative loss is 0.0009 (less than the convergence threshold of 0.001). Further, if the total loss change is less than or equal to 0.001 for five consecutive iterations during the 85th iteration, the model converges and training terminates. Further, the optimized XGBoost model is deployed to the test environment and validated using 100 data points from the test set. For the rule validation sample Y-CL-001, the model predicts a loss assessment ratio of 81.5%, with a deviation of 0.5% from the business rule output of 82%. The overall accuracy of the test set is 88% > 85%, and the F1 score is 87.5% > 85%, indicating successful validation. Further, the optimized model version is updated to XGBoost V2.2, and the training parameters, total loss function, and validation results are recorded, stored in the property insurance model database, synchronized to the property insurance claims system, and used to replace the original model.

[0107] As another concrete example in the healthcare field, the LightGBM model training parameters are: number of iterations = 100, learning rate = 0.05, convergence threshold = 0.001, L2 regularization coefficient = 0.05, optimizer = Adam. Furthermore, LightGBM model training can be started, using the updated training dataset (containing Y-ME-001 samples) and the total loss function L_total = MSE + L_rule. Loss changes are monitored in real time, with the loss value recorded every 10 iterations. Example: total loss 0.35 in the 10th iteration, 0.12 in the 60th iteration, and 0.0008 in the 90th iteration (less than the convergence threshold). Furthermore, in the 90th iteration, the total loss change was less than or equal to 0.001 for 5 consecutive iterations, the model converged, and training was terminated. Further, the optimized LightGBM model was deployed to the test environment, and 80 data points from the test set were used for validation. The test data corresponding to rule validation sample Y-ME-001 showed a model prediction drug dosage calculation coefficient of 10.3 mg / kg, with a deviation of 0.1 mg / kg ≤ 5% from the business rule output of 10.4 mg / kg. The overall accuracy of the test set was 90% > 85%, and the F1 score was 89.5% > 85%, indicating successful validation. Further, the optimized model version was updated to LightGBM V3.2, and the training parameters, total loss function, and validation results were recorded, stored in the medical model database, synchronized to the hospital's HIS system, and the original model was replaced.

[0108] It is understandable that the above embodiments construct rule verification samples by combining divergent samples with correct business rules and introduce rule consistency loss terms to achieve rule-supervised model training. This enables the model to proactively align with authoritative business logic during iterative optimization, thereby reducing conflicts between model output and compliance rules from the root.

[0109] Step S170: Based on the differential root cause diagnosis report, construct at least one rule parameter optimization space; Specifically, this step is executed only when a root cause diagnosis report is generated in step S150. First, the core difference indicator is extracted from the root cause diagnosis report, and the associated rule parameters (i.e., adjustable parameters in the rules affecting the allocation result of the core indicator) in the target business rule set corresponding to that core indicator are determined. Second, a reasonable range of values ​​for the rule parameter is determined (based on business scenario constraints and historical data statistics; the range must be specific and quantifiable to avoid an unlimited range). This range is the rule parameter optimization space. If multiple core difference indicators exist, a corresponding rule parameter optimization space is constructed for each core indicator. Each optimization space exists independently, specifying the optimization parameter name, value range, and step size (the step size is the smallest unit of parameter adjustment, ensuring that those skilled in the art can test it step by step). The optimization space is stored in a structured format for easy retrieval in subsequent steps.

[0110] As a concrete example in the property insurance field, based on the differential root cause diagnosis report generated in step S150 (the core differential indicator = loss assessment amount), the related rule parameter can be extracted: the "loss assessment amount ratio" of rule 2 in the target business rule set (85% in rule 2). This parameter is adjustable and directly affects the allocation result of the loss assessment amount. The reasonable range of this parameter is determined: based on the constraints of property insurance claims business (the loss assessment ratio cannot be lower than 70%, otherwise it will lead to claims losses; it cannot be higher than 95%, otherwise it will increase claims risks), combined with historical loss assessment data from the past 6 months (the optimal loss assessment ratio range is 75%-90%), the rule parameter optimization space is determined to be "loss assessment amount ratio ∈ [75%, 90%]", with a step size of 1% (i.e., the parameter can start from 75% and be adjusted by 1% each time until 90%). The optimization space is structured and stored as {"optimization parameter":"loss assessment amount ratio","value range":[0.75,0.9],"step size":0.01}. If there are multiple core indicators with differences (such as damage assessment amount + number of survey vehicles), then optimization spaces are constructed separately. For example, if the rule parameter corresponding to the number of survey vehicles is "upper limit of the number of survey vehicles", the optimization space is "upper limit of the number of survey vehicles ∈ [1,3]", and the step size is 1.

[0111] As another specific example, in the healthcare field, based on the differential root cause diagnostic report (differential core indicator = drug dosage) generated in the supplementary example of step S150, the association rule parameter can be extracted: the "drug dosage calculation coefficient" of rule 2 in the target business rule set (10mg / kg in rule 2). This parameter is an adjustable parameter that directly affects the drug dosage allocation result. Determine the reasonable range of values ​​for this parameter: Based on medical guidelines (the drug dosage coefficient for hypertensive emergencies should not be lower than 8 mg / kg, otherwise the efficacy will be insufficient; it should not be higher than 12 mg / kg, otherwise side effects will occur), combined with the hospital's medication data over the past 3 months (the optimal coefficient range is 8.5-11.5 mg / kg), the optimization space for the rule parameter is determined as "the drug dosage calculation coefficient is located in [8.5, 11.5]", with a step size of 0.1 mg / kg (that is, the parameter can start from 8.5 and be adjusted by 0.1 mg / kg each time until it reaches 11.5). The optimization space is structured and stored as {"optimization parameter": "drug dosage calculation coefficient", "value range": [8.5, 11.5], "step size": 0.1}. If there are multiple core indicators of difference (such as drug dosage + number of nurses), then the rule parameter corresponding to the number of nurses is "standard of nurse staffing for critically ill patients", the optimization space is "standard of nurse staffing for critically ill patients is located in [1,2]", and the step size is 0.5 (that is, 1 or 1.5 or 2 nurses, 1.5 means 3 patients share 2 nurses).

[0112] Step S180: Find the target rule parameters in the rule parameter optimization space that make the preset objective function have the optimal solution, and generate rule parameter optimization suggestions based on the target rule parameters; Specifically, in this step, firstly, a predefined objective function is established (constructed based on business objectives, it is a quantifiable mathematical function containing at least one objective indicator; the optimal solution of the objective function is defined as "maximizing the objective indicator" or "minimizing the objective indicator," and the criteria for determining the optimal solution are clearly defined). Secondly, a grid search method (a commonly used parameter optimization method in this field, simple to implement, and suitable for small-scale optimization spaces) is used. Within the rule parameter optimization space constructed in step S170, all candidate rule parameters are traversed at a predefined step size, and each candidate parameter is substituted into the objective function to calculate the objective function value. Then, the objective function values ​​corresponding to all candidate parameters are compared, and candidate parameters that enable the objective function to reach the optimal solution (such as maximum or minimum value) are selected as target rule parameters. Finally, based on the target rule parameters, optimization suggestions for rule parameters are generated (structured suggestions, including: original rule parameter value, target rule parameter value, parameter adjustment range, adjustment reason, and expected effect after adjustment), ensuring that those skilled in the art can directly adjust the rule parameters according to the suggestions. If multiple optimization spaces exist, optimization is performed separately to generate corresponding optimization suggestions.

[0113] As a concrete example in the property insurance field: 1. Preset objective function: Objective function = Claims cost control rate (maximize) - Customer complaint rate (minimize), where claims cost control rate = (Assessed loss amount / Actual loss amount) 100% (target is greater than or equal to 80%), Customer complaint rate = (Number of complaints / Total number of cases) 100% (target less than or equal to 5%), the optimal solution is defined as "claims cost control rate greater than or equal to 80% and customer complaint rate less than or equal to 5%, while maximizing the claims cost control rate". 2. Grid search optimization: The rule parameter optimization space is "the proportion of loss assessment amount is located in [75%, 90%], step size = 1%", traverse all candidate parameters (75%, 76%, ..., 90%), substitute each candidate parameter into the objective function, and calculate the objective function value (based on historical data simulation calculation, such as when candidate parameter = 82%, claims cost control rate = 83%, customer complaint rate = 3%; when candidate parameter = 83%, claims cost control rate = 84%, customer complaint rate = 3.2%; when candidate parameter = 84%, claims cost control rate = 85%, customer complaint rate = 3.5%; when candidate parameter = 85%, claims cost control rate = 86%, customer complaint rate = 4%; when candidate parameter = 86%, claims cost control rate = 87%, customer complaint rate = 5.2% (exceeding the target)). 3. Select target rule parameters: When the candidate parameter = 85%, the claims cost control rate = 86% (greater than 80%), the customer complaint rate = 4% (less than 5%), and the claims cost control rate reaches the current optimal level. Therefore, the target rule parameter is determined to be 85% (consistent with the original rule parameter; if the original parameter is 80%, then the target parameter is 85%). 4. Generate optimization suggestions: {"Original rule parameter": "Loss assessment amount ratio = 80%", "Target rule parameter": "Loss assessment amount ratio = 85%", "Adjustment range": "+5%", "Reason for adjustment": "After adjustment, the claims cost control rate increases from 81% to 86%, the customer complaint rate remains below 4%, meeting business objectives, and simultaneously reducing the difference from the model allocation scheme", "Expected effect after adjustment": "The difference in loss assessment amount decreases from 0.28 to 0.0625, lower than the preset threshold of 0.2, so there is no need to generate a root cause diagnosis report"}.

[0114] As another concrete example, in the field of healthcare, 1. Preset objective function: Objective function = drug efficacy achievement rate (maximize) - side effect incidence rate (minimize), where drug efficacy achievement rate = (number of patients whose condition improved after medication / total number of patients using medication) 100% (target is greater than or equal to 90%), side effect incidence rate = (number of patients experiencing side effects / total number of patients using the medication) 100% (target less than or equal to 3%), the optimal solution is defined as "the efficacy achievement rate is greater than or equal to 90% and the side effect incidence rate is less than or equal to 3%, while maximizing the efficacy achievement rate". 2. Grid Search Optimization: The optimization space for the rule parameters is "the drug dosage calculation coefficient is located in [8.5, 11.5], with a step size of 0.1 mg / kg". Iterate through all candidate parameters (8.5, 8.6, ..., 11.5), substitute each candidate parameter into the objective function, and calculate the objective function value (based on simulation calculation of historical hospital medication data, such as when candidate parameter = 10.2 mg / kg, the efficacy achievement rate = 92%, and the side effect incidence rate = 2.5%; when candidate parameter = 10.3 mg / kg, the efficacy achievement rate = 93%, and the side effect incidence rate = 2.8%; when candidate parameter = 10.4 mg / kg, the efficacy achievement rate = 94%, and the side effect incidence rate = 3.1% (exceeding the target); when candidate parameter = 10.1 mg / kg, the efficacy achievement rate = 91%, and the side effect incidence rate = 2.2%). 3. Screening Target Rule Parameters: When the candidate parameter = 10.3 mg / kg, the efficacy achievement rate = 93% (greater than 90%), the side effect incidence rate = 2.8% (less than 3%), and the efficacy achievement rate reaches the current optimal level. Therefore, the target rule parameter is determined to be 10.3 mg / kg (the original rule parameter is 10 mg / kg). 4. Generating Optimization Suggestions: {"Original Rule Parameter": "Drug Dosage Calculation Coefficient = 10 mg / kg", "Target Rule Parameter": "Drug Dosage Calculation Coefficient = 10.3 mg / kg", "Adjustment Range": "+0.3 mg / kg", "Reason for Adjustment": "After adjustment, the efficacy achievement rate increases from 90% to 93%, and the side effect incidence rate remains below 2.8%, meeting the medical business objectives, while reducing the drug dosage difference from the model allocation scheme", "Expected Effect After Adjustment": "The drug dosage difference decreases from 0.1667 to 0.0083, which is lower than the preset threshold of 0.15, so there is no need to generate a root cause diagnosis report"}.

[0115] In some embodiments of the present invention, step S180 includes the following steps.

[0116] Step S181: Construct a multi-objective evaluation function that includes at least one preset target indicator; Specifically, in this step, firstly, based on the business scenario of the items to be allocated (property insurance claims, medical resource allocation), at least one core target indicator is preset (prioritize indicators that are strongly correlated with the differentiated core indicators and fit the core business needs; 2-3 indicators can be set to avoid function redundancy due to too many indicators); secondly, clarify the quantitative calculation method and reasonable value range of each target indicator to ensure that the indicator can be calculated practically; finally, construct a multi-objective evaluation function, using a linear weighted summation method (the total weight is 1, and the weight of each indicator is determined by the analytic hierarchy process, with the core target indicator having a higher weight than the secondary indicators). The function expression is: Multi-objective evaluation function = (target indicator) / (target value) The sum of the corresponding weights is used to set a preset threshold for the function output value (calibrated based on historical business data and used for subsequent screening of candidate parameter sets), ensuring that the function fits the actual business, is quantifiable, and verifiable.

[0117] As a specific example, in the property insurance sector, we can optimize the rule parameters built in step S170 of the previous section (optimization parameter = loss assessment amount ratio, value range [75%, 90%], step size = 1%), and execute step S181: 1. Preset target indicators (2 core indicators, aligned with claims business needs): Indicator 1: Claims cost control rate (core target), quantitative calculation method = (loss assessment amount / actual loss amount) 100%, with a reasonable range of 70%-95%, the higher the better; Indicator 2: Customer Satisfaction (Secondary Objective), Quantitative Calculation Method = (Number of Satisfactory Cases / Total Number of Cases) 1. **Target Values:** 100%, with a reasonable range of 85%-100%, the higher the better; 2. **Indicator Weights:** Calculated using the Analytic Hierarchy Process (AHP), the weights for claims cost control rate and customer satisfaction are 0.6 and 0.4 respectively, with a total weight of 1; 3. **Multi-Objective Evaluation Function:** The multi-objective evaluation function F(x) = 0.6. Claims cost control rate (x) +0.4 Customer satisfaction (x) (x is a candidate rule parameter: the proportion of loss assessment amount); 4. Set the preset threshold for the output value of the function: based on the historical data of property insurance claims in the past 6 months, the preset threshold = 88 (that is, when F(x) is greater than or equal to 88, the candidate parameter set is considered valid).

[0118] As another concrete example, in the healthcare field, based on the rule parameter optimization space constructed in step S170 of the previous section (optimization parameter = drug dosage calculation coefficient, value range [8.5, 11.5] mg / kg, step size = 0.1 mg / kg), step S181 can be executed: 1. Preset target indicators (2 core indicators, tailored to medical business needs): Indicator 1: Drug efficacy achievement rate (core target), quantitative calculation method = (number of patients whose condition improved after medication / total number of patients using medication) × 100%, reasonable value range 8 1. 5%-98%, the higher the better; Indicator 2: Medication safety (secondary objective), quantitative calculation method = 100% - side effect incidence rate, side effect incidence rate = (number of patients with side effects / total number of patients using medication) × 100%, reasonable value range 97%-100%, the higher the better; 2. Determine indicator weights: calculated by AHP analytic hierarchy process, drug efficacy achievement rate weight = 0.7, medication safety weight = 0.3, total weights = 1; 3. Construct multi-objective evaluation function: multi-objective evaluation function F(y) = 0.7 × drug efficacy achievement rate (y) + 0.3 × medication safety (y) (y is candidate rule parameter: drug dosage calculation coefficient); 4. Set preset threshold for function output value: calibrated based on the hospital's medication history data for the past 3 months, preset threshold = 92 (that is, when F(y) is greater than or equal to 92, the candidate parameter set is considered valid).

[0119] Step S182: Sampling of rule parameters within each rule parameter optimization space to form at least one set of candidate rule parameters; Specifically, in this step, firstly, the rule parameter optimization space constructed in step S170 is invoked to clarify the value range and step size of the adjustable rule parameters, and to determine the sampling range (within the optimization space); secondly, a random sampling method is adopted (a commonly used sampling method in this field, simple to implement, and sampling uniformity can be guaranteed by setting a sampling seed). The sampling quantity is determined according to the size of the optimization space (when the optimization space is small, the sampling quantity can cover more than 50% of the candidate parameters; when the optimization space is large, the sampling quantity is greater than or equal to 20 groups to ensure that the sample is representative); then, the validity of each rule parameter obtained by sampling is verified (parameters that exceed the optimization space or do not meet business constraints are removed); finally, the verified rule parameters are organized into a candidate rule parameter set. Each parameter set contains one adjustable rule parameter (single parameter optimization scenario). If it is a multi-parameter optimization, each parameter set contains all adjustable parameters. Each parameter set is labeled with a unique number and stored in the database for subsequent simulation evaluation.

[0120] As a concrete example, in the property insurance field, the optimization space based on the rule parameters in step S170 (the proportion of loss assessment amount is between [75%, 90%], step size = 1%, a total of 16 candidate parameters) can be executed in step S182: 1. Define the sampling range: 75%-90%, step size 1%, and set the sampling seed to 123 (to ensure sample reproducibility); 2. Random sampling: Select 8 sets of candidate parameters (covering more than 50% of the candidate parameters to ensure representativeness), and the sampling results are as follows: 76%, 79%, 82%, 83%, 85%, 87%, 88%, 89%; 3. Validity verification: All sampling parameters are within the range of [75%, 90%], which is in line with the property insurance principle. Claims business constraints (loss assessment ratio 70%-95%), all verified and passed; 4. Form candidate rule parameter sets: a total of 8 groups, numbered 1-8, as follows: Group 1: {loss assessment ratio = 76%}; Group 2: {loss assessment ratio = 79%}; Group 3: {loss assessment ratio = 82%}; Group 4: {loss assessment ratio = 83%}; Group 5: {loss assessment ratio = 85%}; Group 6: {loss assessment ratio = 87%}; Group 7: {loss assessment ratio = 88%}; Group 8: {loss assessment ratio = 89%}; Store the 8 parameter sets in the claims database, noting the sampling time and sampling method for use in subsequent steps.

[0121] As another concrete example, in the healthcare field, the optimization space can be based on the rule parameters established in step S170 (the drug dosage calculation coefficient is located in [8.5, 11.5] mg / kg, step size = 0.1 mg / kg, with a total of 31 candidate parameters). Step S182 is then executed: 1. Define the sampling range: 8.5-11.5 mg / kg, step size 0.1 mg / kg, and set the sampling seed to 456 (to ensure reproducibility of the sampling); 2. Random sampling: select 20 sets of candidate parameters (covering more than 60% of the candidate parameters to ensure representativeness). The sampling results are as follows: 8.7, 9.2, 9.5, 9.8, 10.0, 10.1, 10.3, 10.5, 10.7, 10.8, 10.9, 11.0, 11.1, 11.2, 11.3, 8.9, 9.1, 9.6, 10.2, 10.4; 3. Validity verification: All sampling parameters are within the range of [8.5, 11.5] mg / kg, which conforms to medical standards (8-12 mg / kg), and all verifications are passed; 4. Formation of candidate rule parameter sets: A total of 20 sets, numbered 1-20, each set of parameters is {drug dosage calculation coefficient = XX mg / kg}, stored in the hospital resource management database, and labeled with sampling-related information.

[0122] Step S183: Simulate and evaluate the function output value of each of the candidate rule parameter sets in the multi-objective evaluation function; Specifically, in this step, firstly, the candidate rule parameter set formed in step S182 and the multi-objective evaluation function constructed in step S181 are invoked; secondly, for each set of candidate rule parameter sets, the parameters are substituted into the target business rule set, the first initial allocation scheme is recalculated, and combined with historical business data (relevant data from the past 3-6 months), the target indicator value corresponding to the parameter set is simulated (ensuring that the simulated data closely matches the actual business and is traceable); then, each target indicator value is substituted into the multi-objective evaluation function to calculate the function output value corresponding to the candidate parameter set; finally, all candidate parameter sets, corresponding target indicator values, and function output values ​​are organized to form a candidate parameter set-function output value comparison table, the validity of each set of parameter sets is marked (not judged at this time, to be screened in subsequent steps), and stored in the database to ensure that the calculation process is reproducible and the results are verifiable.

[0123] As a concrete example, in the property insurance field, the evaluation can be based on the 8 sets of candidate rule parameters from step S182 and the multi-objective evaluation function F(x) = 0.6 from step S181. Claims cost control rate +0.4 Customer satisfaction, execution step S183: 1. Simulate and calculate target indicator values: Based on the historical property insurance claims data of the past 6 months, simulate the indicator values ​​corresponding to each group of parameters one by one (examples are as follows): Group 1 (76%): Claims cost control rate = 79%, Customer satisfaction = 88%; Group 3 (82%): Claims cost control rate = 83%, Customer satisfaction = 91%; Group 5 (85%): Claims cost control rate = 86%, Customer satisfaction = 93%; Group 6 (87%): Claims cost control rate = 88%, Customer satisfaction = 92%; Group 8 (89%): Claims cost control rate = 90%, Customer satisfaction = 89%; 2. Calculate the function output value: Substitute each group of indicator values ​​into the function, The calculation results are as follows: Group 1: F(x) = 0.6 × 79 + 0.4 × 88 = 47.4 + 35.2 = 82.6; Group 3: F(x) = 0.6 × 83 + 0.4 × 91 = 49.8 + 36.4 = 86.2; Group 5: F(x) = 0.6 × 86 + 0.4 × 93 = 51.6 + 37.2 = 88.8; Group 6: F(x) = 0.6 × 88 + 0.4 × 92 = 52.8 + 36.8 = 89.6; Group 8: F(x) = 0.6 × 90 + 0.4 × 89 = 54 + 35.6 = 89.6; 3. Compile a comparison table: Compile the 8 sets of parameters, index values, and function output values ​​into a table and store it in the claims database for subsequent screening.

[0124] As another concrete example, in the healthcare field, this can be based on the 20 sets of candidate rule parameters from step S182 and the multi-objective evaluation function F(y) = 0.7 from step S181. The efficacy target achievement rate increased by 0.3%. Medication safety, execution step S183: 1. Simulate and calculate target indicator values: Based on the hospital's historical medication data over the past 3 months, simulate the indicator values ​​corresponding to each group of parameters one by one (examples are as follows): Group 5 (10.0 mg / kg): efficacy achievement rate = 90%, medication safety = 97.8% (side effect incidence rate 2.2%); Group 7 (10.3 mg / kg): efficacy achievement rate = 93%, medication safety = 97.2% (side effect incidence rate 2.8%); Group 10 (10.8 mg / kg): efficacy achievement rate = 95%, medication safety = 96.7% (side effect incidence rate 3.3%); Group 15 (11.3 mg / kg) Group 10 (10.2 mg / kg): Efficacy rate = 96%, drug safety = 96.1% (side effect incidence rate 3.9%); Group 11 (10.2 mg / kg): Efficacy rate = 92%, drug safety = 97.5% (side effect incidence rate 2.5%); 2. Calculate the function output value: Substitute the index values ​​of each group into the function, and the calculation results are as follows: Group 5: F(y) = 0.7 × 90 + 0.3 × 97.8 = 63 + 29.34 = 92.34; Group 7: F(y) = 0.7 × 93 + 0.3 × 97.2 = 65.1 + 29.16 = 94.26; Group 10: F(y) = 0.7 × 95 + 0.3 × 96.7 = 66.5 + 29.01=95.51; Group 15: F(y)=0.7×96 + 0.3×96.1=67.2 + 28.83=96.03; Group 19: F(y)=0.7×92 + 0.3×97.5=64.4 + 29.25=93.65; 3. Compile a comparison table: Compile the 20 sets of parameters, index values, and function output values ​​into a table and store it in the hospital resource management database for subsequent screening.

[0125] Step S184: The set of candidate rule parameters corresponding to function output values ​​higher than preset values ​​is taken as the target rule parameters; Specifically, in this step, firstly, the candidate parameter set-function output value comparison table compiled in step S183 is called to clarify the preset threshold for function output values ​​set in step S181; secondly, candidate rule parameter sets with function output values ​​> preset thresholds are selected (if the function output value = preset threshold, it is considered to meet the requirements and is selected together); then, if multiple sets of candidate parameter sets are selected, the function output values ​​of each set are compared, and the parameter set with the highest output value is selected as the target rule parameter; if multiple sets of parameter sets have the same output value (tied for the highest), the parameter set closest to the original rule parameter is selected (reducing the parameter adjustment range and lowering the business adaptation cost); finally, the selected candidate rule parameter set is determined as the target rule parameter, the name and value of the target rule parameter are clarified, it is compared with the original rule parameter, the differences are recorded, and stored in the database to ensure that the target rule parameter meets the business requirements and is executable.

[0126] As a specific example, in the property insurance field, based on the comparison table in step S183 and the preset threshold set in step S181 = 88, step S184 can be executed as follows: 1. Filter the candidate parameter sets that meet the conditions: the parameter sets with function output values ​​> 88 are group 5 (88.8), group 6 (89.6), and group 8 (89.6), a total of 3 groups; 2. Compare the function output values: the output values ​​of group 6 and group 8 are the highest (89.6), and are tied for best; 3. Select the target rule parameter: the original rule parameter is 80%, the difference between group 6 (87%) and the original parameter is +7%, and the difference between group 8 (89%) and the original parameter is +9%. The group 6 (87%) with the smaller difference is selected as the target rule parameter; 4. Record the target parameter information: target rule parameter = loss assessment amount ratio 87%, original parameter = 80%, difference = +7%, and store it in the claims database for subsequent use in generating optimization suggestions.

[0127] As another specific example, in the field of healthcare, based on the comparison table in step S183 and the preset threshold of 92 set in step S181, step S184 is executed: 1. Screening candidate parameter sets that meet the conditions: There are 15 groups of parameter sets with function output values ​​> 92, including group 5 (92.34), group 7 (94.26), group 10 (95.51), group 15 (96.03), group 19 (93.65), etc.; 2. Comparing function output values: the output value of group 15 (11.3mg / kg) is the highest (96.03); 3. Select target rule parameters: The original rule parameter is 10.0 mg / kg. Group 15 (11.3 mg / kg) differs from the original parameter by +1.3 mg / kg, which complies with medical standards (8-12 mg / kg) and does not exceed the constraints. Therefore, the target rule parameter is determined to be the drug dosage calculation coefficient of 11.3 mg / kg. 4. Record the target parameter information: Target rule parameter = drug dosage calculation coefficient 11.3 mg / kg, original parameter = 10.0 mg / kg, difference = +1.3 mg / kg, and store it in the hospital resource management database.

[0128] Step S185: Generate rule parameter optimization suggestions based on the target rule parameters.

[0129] Specifically, in this step, firstly, the target rule parameters and original rule parameters determined in step S184 are called, and the parameter adjustment range is calculated (percentage-type parameters: adjustment range = (target parameter - original parameter) / original parameter × 100%; numerical parameters: adjustment range = target parameter - original parameter); secondly, combined with the function output value and target indicator changes in step S183, the adjustment reasons are written (clearly stating how the adjustment will improve the multi-objective evaluation function output value, how to meet business objectives, and how to narrow the difference between the two initial allocation schemes); then, the expected effect after the adjustment is predicted (quantifying the changes in function output value, target indicator changes, and the difference between the two initial allocation schemes, based on historical data simulation to ensure verifiability); finally, a structured rule parameter optimization suggestion is generated, containing five core modules: "original rule parameters, target rule parameters, adjustment range, adjustment reasons, and expected effect," with concise language and clear logic, for those skilled in the art to directly use for rule parameter adjustment, and stored in the database, linked to the root cause diagnosis report and optimization space to ensure traceability.

[0130] As a specific example, in the property insurance field, based on the target rule parameter (loss assessment amount ratio of 87%) determined in step S184, step S185 can be executed to generate optimization suggestions: {"Original rule parameter": "loss assessment amount ratio = 80%", "Target rule parameter": "loss assessment amount ratio = 87%", "Adjustment range": "+8.75% (absolute adjustment +7 percentage points)", "Reason for adjustment": "After adjustment, the output value of the multi-objective evaluation function increases from 82.0 corresponding to the original parameter to 89.6, exceeding the preset threshold of 88; the claims cost control rate increases from 81% to 88%, and customer satisfaction increases from 87% to 92%..." The difference between rule-based and model-based loss assessments was reduced, decreasing from 0.28 to 0.15, below the preset threshold of 0.2, thus meeting business objectives. Expected results: 1. The multi-objective evaluation function output value stabilizes at around 89.6; 2. Claims cost control rate remains at 88%-89%, and customer satisfaction remains at 92%-93%; 3. The overall difference between the two initial allocation schemes is less than or equal to 0.15, eliminating the need to generate a root cause diagnosis report; 4. Claims business adaptation costs are low, requiring no adjustment of other rule parameters. The optimization suggestions are stored in the claims database and synchronized to the claims rule management system, allowing staff to directly adjust rule parameters.

[0131] As another concrete example, in the healthcare field, based on the target rule parameter (drug dosage calculation coefficient 11.3 mg / kg) determined in step S184, step S185 can be executed to generate optimization suggestions: {"Original rule parameter": "Drug dosage calculation coefficient = 10.0 mg / kg", "Target rule parameter": "Drug dosage calculation coefficient = 11.3 mg / kg", "Adjustment range": +13% (absolute adjustment +1.3 mg / kg), "Reason for adjustment": "After adjustment, the output value of the multi-objective evaluation function increased from 90.5 to 96.03, exceeding the preset threshold of 92; the efficacy achievement rate increased from 90% to 96%, and the medication safety remained at 96.1% (side effect incidence rate of 3.9%, not exceeding the medical standard threshold of 5%). Simultaneously, the difference between rule-based and model-based drug dosages was reduced, with the difference decreasing from 0.1667 to 0.08, below the preset threshold of 0.15, meeting the medical business objectives." "Expected Results": "1. The output value of the multi-objective evaluation function stabilizes at around 96.0; 2. The efficacy achievement rate remains at 95%-96%, and the side effect incidence rate is controlled at 3.9%-4.2%, complying with medical standards; 3. The overall difference between the two initial allocation schemes is ≤0.08, eliminating the need to generate a root cause diagnosis report; 4. Medical staff do not need to adjust medication procedures, resulting in low adaptation costs." The optimization suggestions will be stored in the hospital resource management database and synchronized to the hospital HIS system for medical staff to refer to when adjusting rule parameters.

[0132] It is understandable that the above embodiments, by generating rule-based schemes and model-based schemes in parallel and making independent decisions, avoid the strong dependence and error propagation of traditional rule-first or model-first schemes, achieve decoupling of dual-path decision-making, and significantly improve the system's fault tolerance and the reliability of allocation results.

[0133] Step S190: Optimize the first initial allocation scheme and / or the second initial allocation scheme according to the rule parameter optimization suggestions to obtain the final allocation scheme for the item to be allocated.

[0134] Specifically, in this step, firstly, the rule parameter optimization suggestions generated in step S180 are analyzed to determine the parameter adjustment objects (the corresponding rule parameters in the target business rule set). The rule parameters are adjusted according to the suggestions, and the rule calculation in step S130 is re-executed to generate the optimized first initial allocation scheme. Secondly, it is determined whether the second initial allocation scheme needs to be optimized: if the difference between the optimized first initial allocation scheme and the original second initial allocation scheme is less than or equal to a preset threshold, then the second scheme does not need to be optimized. If the difference is still greater than the preset threshold, the second initial allocation scheme can be fine-tuned based on the optimized rule parameters (the fine-tuning range does not exceed 10% to ensure that it does not deviate from the model prediction logic). Finally, the optimized first and / or second initial allocation schemes are integrated to generate the final allocation scheme. The final scheme must include all resource allocation details (resource type, quantity, specific object, allocation time, constraint satisfaction, optimization description), be structured and stored, and synchronized to the corresponding business system (such as resource management system, business execution system) for execution by those skilled in the art. At the same time, the optimization process is recorded to ensure traceability.

[0135] As a specific example in the property insurance field: 1. Adjust rule parameters: Based on optimization suggestions, adjust the "loss assessment amount ratio" of rule 2 in the target business rule set from 80% to 85%; 2. Optimize the first initial allocation scheme: Re-execute the rule calculation, input the description information of the item to be allocated (reported amount = 30,000 yuan, accident type = 2), and the rule outputs the optimized first scheme: loss assessment amount = 30,000 yuan. 85% = 25,500 yuan, with other parameters (1 intermediate claims specialist, 1 inspection vehicle) remaining unchanged; 3. Comparison of differences: The relative difference between the optimized first plan (damage assessment amount of 25,500 yuan) and the original second plan (damage assessment amount of 15,000 yuan) = 10,500 / 15,000 = 0.7, and the overall difference = 0.7. 0.4 = 0.28 > 0.2, therefore, fine-tune the second initial allocation scheme: adjust the loss assessment amount output by the model from 15,000 yuan to 24,000 yuan (fine-tuning range = 60%, this is an example, the actual fine-tuning range can be adjusted according to business needs, if the difference after fine-tuning is less than or equal to 0.2); 4. Generate the final allocation scheme: integrate the optimized first and second schemes, the final scheme is {"Resource Type": "Claims Resource", "Claims Specialist": "LZ002 (Intermediate)"} The following data is generated: "Inspection Vehicle": "CK015", "Damage Assessment Amount": 25,000 yuan (average of both, taking into account both rules and model), "Priority": "Medium", "Constraint Satisfaction": "Yes", "Optimization Explanation": "Adjust the rule-based damage assessment ratio to 85%, fine-tune the model-based damage assessment amount to 24,000 yuan, and finally take the average, reducing the difference to 0.02, which meets the preset threshold". This data is synchronized to the claims system for claims personnel to execute, and the optimization process (original parameters, adjusted parameters, and difference changes) is recorded.

[0136] As a specific example in the healthcare field: 1. Adjust rule parameters: Based on optimization suggestions, the "drug dosage calculation coefficient" of rule 2 in the target business rule set is adjusted from 10 mg / kg to 10.3 mg / kg; 2. Optimize the first initial allocation scheme: Re-execute the rule calculation, input the description information of the item to be allocated (patient weight 70 kg), and the rule outputs the optimized first scheme: drug dosage = 70 × 10.3 = 721 mg / dose, with other parameters (1 intensive care bed, 1 nurse, 1 ventilator on standby) remaining unchanged; 3. Compare the degree of difference: The relative difference between the optimized first scheme (drug dosage 721 mg / dose) and the original second scheme (drug dosage 1200 mg / dose) is approximately 0.399 (479 / 1200), and the overall degree of difference is 0.399. Since 0.4 is approximately 0.1596 > 0.15, the second initial allocation scheme is fine-tuned: the drug dosage output by the model is adjusted from 1200mg / dose to 750mg / dose (fine-tuning range = 37.5%). 4. Generate the final allocation scheme: integrate the optimized first and second schemes, and the final scheme is {"Resource Type": "Medical Resources", "Bed": "CZ012 (Intensive Care)", "Nurse": "HS035", "Ventilator": "HB008 (Spare)", "Drug Dosage": 735mg / dose (take the average of the two), "Priority": "High", "Constraint Satisfaction": "Yes", "Optimization Description": "Adjust the rule drug dosage coefficient to 10.3mg / kg, fine-tune the model drug dosage to 750mg / dose, and finally take the average value, the difference is reduced to 0.019, which meets the preset threshold"}, and synchronized to the hospital HIS system for medical staff to execute, and record the optimization process (original parameters, adjusted parameters, difference changes).

[0137] Understandably, the first step of this invention is to acquire at least one customer segment, dividing customers into multiple segments based on their core characteristics. This upgrades the fragmented mapping of "single feature - single product" to a unified management of a group-level strategy pool. In this approach, when the mapping relationship changes, there's no need for manual adjustments to the matching relationship for each individual feature value; only the relevant rules need to be adjusted for each customer segment. This significantly reduces the workload of basic maintenance and solves the problem of high workload at the management granularity level.

[0138] Secondly, the second step of the solution, "constructing strategy evolution sub-populations corresponding to different customers within each customer segment, with different strategy evolution sub-populations corresponding to differentiated strategy search spaces," builds a dedicated recommendation strategy pool for each customer segment and sets a differentiated strategy search space for each sub-population. When the mapping relationship changes, there is no need to manually modify the product corresponding to each feature value one by one. It is only necessary to simply adjust the boundary of the strategy search space of the corresponding sub-population. All recommendation strategies within the sub-population will autonomously explore and evolve within the new search space boundary, automatically adapting to the new mapping relationship. There is no need for manual intervention to match them one by one, thus eliminating a large amount of manual workload in terms of adaptation method.

[0139] Furthermore, the third step of the solution constructs a strategy fitness function that integrates real-time external risk signals, providing an automatically updatable evaluation criterion for the autonomous evolution of the strategy. This fitness function is not fixed but can integrate external change signals (external risk signals) in real time. These external changes are essentially the core reasons that cause changes in the "user feature value-product" mapping relationship. The function automatically transforms these external changes into evaluation criteria for strategy evolution, eliminating the need for manually redefining the "feature-product" matching rules. This allows the strategy to automatically evolve in the direction of adapting to the new mapping relationship, avoiding the enormous workload of manually reconstructing the rules in terms of evaluation criteria.

[0140] Then, the fourth step of the solution, "based on the policy fitness function, drives multiple policy evolution subpopulations to perform iterative optimization in parallel, and performs policy knowledge transfer operations across policy evolution subpopulations during the iteration process," achieves fully automated optimization of recommended policies and reuse of high-quality policies. Driven by the fitness function, each policy evolution subpopulation autonomously performs iterative optimization in parallel: automatically calculating the fitness of each policy, selecting high-quality policies that fit the new mapping relationship, and eliminating invalid policies that do not conform to the new mapping relationship; at the same time, the cross-population policy knowledge transfer operation can reuse the high-quality policies of one group to other groups with similar needs, without the need for manual establishment of similar mapping relationships for different groups, further reducing the workload of duplication. The entire iterative optimization and knowledge transfer process is fully automated, without the need for manual intervention in screening, matching, and reuse, completely eliminating the large amount of workload of manually re-screening and matching policies when the mapping relationship changes, which is the core link in solving this technical problem.

[0141] Finally, the fifth step of the solution, "outputting a recommended strategy set from the optimized strategy evolution subpopulation and generating explanatory information about the strategy evolution trajectory association of the recommended strategy set," completes the automatic output and explanation of high-quality strategies under the new mapping relationship. The final output of the solution is a high-quality recommended strategy set that has undergone automated iterative optimization, rather than a fixed mapping table. This set can be directly used to recommend products to customers without the need for manual screening and organization of strategies. Simultaneously, the solution automatically generates explanatory information about the strategy evolution trajectory, further reducing subsequent manual workload.

[0142] In summary, when the recommendation mapping relationship between "user feature value - product" changes, existing technologies require manual intervention throughout the entire process, resulting in a huge workload. However, this solution, through the synergistic effect of the above five steps, forms a complete closed loop of "group overall management - search space adaptation - automatic update of evaluation criteria - automatic iterative optimization of strategies - automatic output of high-quality strategies." Only a few core parameters (such as search space boundaries and fitness function weights) need to be adjusted manually. All other processes for adapting to the new mapping relationship are completed automatically by the solution. There is no need to manually re-establish the recommendation mapping relationship between feature values ​​and products one by one, thus fundamentally solving the technical problem of the large workload of recommending products when the mapping relationship changes in existing technologies.

[0143] 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. Software tools or components not belonging to this company that appear in the embodiments of this application are merely illustrative examples and do not represent actual use.

[0144] In one embodiment, a resource allocation device is provided, which corresponds one-to-one with the resource allocation method in the above embodiments. For example... Figure 3As shown, the resource allocation device includes a data acquisition module 510, a calling module 320, an allocation module 330, a difference analysis module 340, a construction module 350, a suggestion generation module 360, and an information generation module 550. Detailed descriptions of each functional module are as follows: The information receiving module 310 is used to receive the description information of the item to be allocated; The module 320 is used to invoke the pre-configured target business rule set and target machine learning model that match the description information; The allocation module 330 is used to perform parallel computation on the items to be allocated based on the target business rule set and the target machine learning model, respectively, to generate a first initial allocation scheme based on rules and a second initial allocation scheme based on the model for the items to be allocated. The difference analysis module 340 is used to perform difference quantification analysis on the first initial allocation scheme and the second initial allocation scheme. When the difference between the first initial allocation scheme and the second initial allocation scheme exceeds a preset threshold, a difference root cause diagnosis report is generated. Module 350 is used to construct at least one rule parameter optimization space based on the differential root cause diagnosis report; It is suggested that the generation module 360 ​​is used to find the target rule parameters that make the preset objective function have the optimal solution within the rule parameter optimization space, and generate rule parameter optimization suggestions based on the target rule parameters; The optimization module 370 is used to optimize the first initial allocation scheme and / or the second initial allocation scheme according to the optimization suggestions of the rule parameters, so as to obtain the final allocation scheme of the item to be allocated.

[0145] In one embodiment, the difference analysis module 340 is specifically used for: Calculate the distribution distance metric between the first initial allocation scheme and the second initial allocation scheme; Analyze at least one business rule and the core decision features of the model that contribute the most to the distribution distance metric; By associating the business rules with the core decision features of the model, a root cause diagnosis report describing the deviation between the business rules and the core decision features of the model is generated.

[0146] In one embodiment, it is suggested that a generation module 360 ​​be used specifically for: Construct a multi-objective evaluation function that includes at least one pre-defined target indicator; Within each of the rule parameter optimization spaces, rule parameters are sampled to form at least one set of candidate rule parameters; The simulation evaluates the output value of each of the candidate rule parameter sets in the multi-objective evaluation function; The set of candidate rule parameters corresponding to function output values ​​higher than preset values ​​is used as the target rule parameters; Based on the target rule parameters, generate rule parameter optimization suggestions.

[0147] In one embodiment, the allocation module 330 is further configured to: Based on the description information of the items to be assigned, the items to be assigned are classified to obtain the categories corresponding to the items to be assigned; Monitor the target machine learning model's allocation decision pattern for each category of the items to be allocated; When the similarity of each allocation decision pattern meets the preset condition and the allocation decision pattern has business interpretability, the allocation decision pattern is transformed into at least one structured and parameterized derivative business rule. The derived business rules are stored in the pre-configured target business rule set.

[0148] In one embodiment, the allocation module 330 is further configured to: The model post-hoc interpretation method is applied to analyze the multiple allocation decision patterns of the target machine learning model for each category of the items to be allocated, quantify the contribution of each feature of the items to be allocated to the allocation decision pattern, and generate a feature contribution ranking list for each feature. Based on the feature contribution ranking list, core decision features with a contribution higher than a preset threshold are selected. Using the core decision features as input variables for the mathematical expression, and the prediction results of the target machine learning model as the approximate fitting target for the mathematical expression, the symbolic regression algorithm is used to iteratively search and optimize the mathematical expression. The generated mathematical expressions are subjected to structured parsing and parameter extraction to obtain preliminary logical rules; The variables, operators, and constants in the initial logical rules are mapped to predefined business terms and adjustable business parameters to form the derived business rules that can be directly interpreted and executed by the rule engine.

[0149] In one embodiment, the difference analysis module 340 is further configured to: Based on the root cause diagnosis report, when the analysis reveals a discrepancy between the first initial allocation scheme and the second initial allocation scheme, and the business rule corresponding to the first initial allocation scheme is logically correct, the first initial allocation scheme, the second initial allocation scheme, the description information of the item to be allocated, and the root cause diagnosis report are combined with the logically correct business rule to form a rule verification sample. Add the rule validation samples to the training dataset of the target machine learning model; In the subsequent training process of the target machine learning model, a rule consistency loss term is introduced, wherein the rule consistency loss term is calculated based on the predicted output of the target machine learning model according to the business rules, and is used to penalize the deviation of the output of the target machine learning model from the logic of the business rules; The target machine learning model is trained and optimized based on the total loss function that incorporates the rule consistency loss term until the output value of the target machine learning model meets the error.

[0150] In one embodiment, the difference analysis module 340 is further configured to: Extract the target business rules that led to the discrepancy from the root cause diagnosis report; The target business rule is transformed into a differentiable constraint. The constraint condition is used as a regularization term and weighted together with the original loss function of the target machine learning model to form a new composite loss function. The target machine learning model is trained based on a new composite loss function, so that the target machine learning model minimizes the prediction error and minimizes the deviation between the output of the target learning model and the target business rule.

[0151] Understandably, the first step of this invention is to acquire at least one customer segment, dividing customers into multiple segments based on their core characteristics. This upgrades the fragmented mapping of "single feature - single product" to a unified management of a group-level strategy pool. In this approach, when the mapping relationship changes, there's no need for manual adjustments to the matching relationship for each individual feature value; only the relevant rules need to be adjusted for each customer segment. This significantly reduces the workload of basic maintenance and solves the problem of high workload at the management granularity level.

[0152] Secondly, the second step of the solution, "constructing strategy evolution sub-populations corresponding to different customers within each customer segment, with different strategy evolution sub-populations corresponding to differentiated strategy search spaces," builds a dedicated recommendation strategy pool for each customer segment and sets a differentiated strategy search space for each sub-population. When the mapping relationship changes, there is no need to manually modify the product corresponding to each feature value one by one. It is only necessary to simply adjust the boundary of the strategy search space of the corresponding sub-population. All recommendation strategies within the sub-population will autonomously explore and evolve within the new search space boundary, automatically adapting to the new mapping relationship. There is no need for manual intervention to match them one by one, thus eliminating a large amount of manual workload in terms of adaptation method.

[0153] Furthermore, the third step of the solution constructs a strategy fitness function that integrates real-time external risk signals, providing an automatically updatable evaluation criterion for the autonomous evolution of the strategy. This fitness function is not fixed but can integrate external change signals (external risk signals) in real time. These external changes are essentially the core reasons that cause changes in the "user feature value-product" mapping relationship. The function automatically transforms these external changes into evaluation criteria for strategy evolution, eliminating the need for manually redefining the "feature-product" matching rules. This allows the strategy to automatically evolve in the direction of adapting to the new mapping relationship, avoiding the enormous workload of manually reconstructing the rules in terms of evaluation criteria.

[0154] Then, the fourth step of the solution, "based on the policy fitness function, drives multiple policy evolution subpopulations to perform iterative optimization in parallel, and performs policy knowledge transfer operations across policy evolution subpopulations during the iteration process," achieves fully automated optimization of recommended policies and reuse of high-quality policies. Driven by the fitness function, each policy evolution subpopulation autonomously performs iterative optimization in parallel: automatically calculating the fitness of each policy, selecting high-quality policies that fit the new mapping relationship, and eliminating invalid policies that do not conform to the new mapping relationship; at the same time, the cross-population policy knowledge transfer operation can reuse the high-quality policies of one group to other groups with similar needs, without the need for manual establishment of similar mapping relationships for different groups, further reducing the workload of duplication. The entire iterative optimization and knowledge transfer process is fully automated, without the need for manual intervention in screening, matching, and reuse, completely eliminating the large amount of workload of manually re-screening and matching policies when the mapping relationship changes, which is the core link in solving this technical problem.

[0155] Finally, the fifth step of the solution, "outputting a recommended strategy set from the optimized strategy evolution subpopulation and generating explanatory information about the strategy evolution trajectory association of the recommended strategy set," completes the automatic output and explanation of high-quality strategies under the new mapping relationship. The final output of the solution is a high-quality recommended strategy set that has undergone automated iterative optimization, rather than a fixed mapping table. This set can be directly used to recommend products to customers without the need for manual screening and organization of strategies. Simultaneously, the solution automatically generates explanatory information about the strategy evolution trajectory, further reducing subsequent manual workload.

[0156] In summary, when the recommendation mapping relationship between "user feature value - product" changes, existing technologies require manual intervention throughout the entire process, resulting in a huge workload. However, this solution, through the synergistic effect of the above five steps, forms a complete closed loop of "group overall management - search space adaptation - automatic update of evaluation criteria - automatic iterative optimization of strategies - automatic output of high-quality strategies." Only a few core parameters (such as search space boundaries and fitness function weights) need to be adjusted manually. All other processes for adapting to the new mapping relationship are completed automatically by the solution. There is no need to manually re-establish the recommendation mapping relationship between feature values ​​and products one by one, thus fundamentally solving the technical problem of the large workload of recommending products when the mapping relationship changes in existing technologies.

[0157] Based on the above resource allocation methods, such as Figure 4 As shown in the diagram, this embodiment of the invention also provides a structural schematic of an apparatus for a resource allocation method, the apparatus including a processor 41 and a memory 42 coupled to the processor 41. The memory 42 stores a computer program, which, when executed by the processor 41, causes the processor 41 to perform the steps of the resource allocation method in the above embodiment.

[0158] For further details regarding the implementation of the above technical solution by the processor 41 in the device for the above resource allocation method steps, please refer to the description of the resource allocation method provided in the above embodiments of the invention, which will not be repeated here.

[0159] The processor 41 can also be called a CPU (Central Processing Unit). The processor 41 may be an integrated circuit chip with signal processing capabilities. The processor 41 may also be a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor may be a microprocessor, or the processor 41 may be any conventional processor.

[0160] like Figure 5As shown in the diagram, this embodiment of the invention also provides a schematic diagram of a computer-readable storage medium, on which a readable computer program 51 is stored. The computer program 51 can be stored in the storage medium in the form of a software product, including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods described in various embodiments of the invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, magnetic disks or optical disks, ROM (Read-Only Memory), RAM (Random Access Memory), or terminal devices such as computers, servers, mobile phones, and tablets.

[0161] In the embodiments provided by this invention, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, apparatuses, or modules, and may be electrical, mechanical, or other forms.

[0162] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0163] Furthermore, the functional modules in the various embodiments of this invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium.

[0164] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product.

[0165] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a server or data center that integrates one or more available media. The available medium may be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium, or a semiconductor medium (e.g., SSD (solid state disk)).

[0166] The technical solution provided by the present invention has been described in detail above. Specific examples have been used to illustrate the principle and implementation of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation and application scope based on the idea of ​​the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

[0167] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, optical storage, etc.) containing computer-usable program code.

[0168] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0169] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0170] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0171] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A resource allocation method, characterized by, The method comprises the following steps: receiving description information of an item to be allocated; calling a pre-configured target business rule set and a target machine learning model matched with the description information; performing parallel calculation on the item to be allocated based on the target business rule set and the target machine learning model, respectively, to generate a first initial allocation scheme of the item to be allocated based on a rule and a second initial allocation scheme of the item to be allocated based on a model; performing difference quantification analysis on the first initial allocation scheme and the second initial allocation scheme, and generating a difference root cause diagnosis report when the difference between the first initial allocation scheme and the second initial allocation scheme exceeds a preset threshold; constructing at least one rule parameter optimization space based on the difference root cause diagnosis report; finding a target rule parameter in the rule parameter optimization space, which makes a preset target function have an optimal solution, and generating a rule parameter optimization suggestion according to the target rule parameter; optimizing the first initial allocation scheme and / or the second initial allocation scheme according to the rule parameter optimization suggestion to obtain a final allocation scheme of the item to be allocated.

2. The resource allocation method of claim 1, wherein, The difference quantification analysis on the first initial allocation scheme and the second initial allocation scheme, and the generation of the difference root cause diagnosis report when the difference between the first initial allocation scheme and the second initial allocation scheme exceeds a preset threshold, comprise the following steps: calculating a distribution distance measure between the first initial allocation scheme and the second initial allocation scheme; analyzing at least one business rule and a model core decision feature that contribute most to the distribution distance measure; associating the business rule with the model core decision feature to generate the difference root cause diagnosis report describing the deviation of the business rule and the model core decision feature.

3. The resource allocation method of claim 1, wherein, The finding of a target rule parameter in the rule parameter optimization space, which makes a preset target function have an optimal solution, and the generation of a rule parameter optimization suggestion according to the target rule parameter, comprise the following steps: constructing a multi-objective evaluation function containing at least one preset target index; sampling rule parameters in each rule parameter optimization space to form at least one set of candidate rule parameters; simulating and evaluating the function output value of each candidate rule parameter set in the multi-objective evaluation function; taking a candidate rule parameter set with a function output value higher than a preset value as the target rule parameter; generating a rule parameter optimization suggestion according to the target rule parameter.

4. The resource allocation method of claim 1, wherein, After the parallel calculation on the item to be allocated based on the target business rule set and the target machine learning model, respectively, to generate a first initial allocation scheme of the item to be allocated based on a rule and a second initial allocation scheme of the item to be allocated based on a model, the method further comprises the following steps: classifying the item to be allocated according to the description information of the item to be allocated to obtain a category corresponding to the item to be allocated; monitoring allocation decision patterns of the item to be allocated of each category by the target machine learning model; when the similarity of each allocation decision pattern meets a preset condition and the allocation decision pattern has business explainability, converting the allocation decision pattern into at least one structured and parameterized derived business rule; store the derived business rule into the target business rule set.

5. The resource allocation method of claim 4, wherein, when the similarity of each of the allocation decision patterns meets a preset condition and the allocation decision pattern has business interpretability, converting the allocation decision pattern into at least one structured and parameterized derived business rule, including: applying a model post-explanation method to analyze the plurality of allocation decision patterns of the target machine learning model for each category of the to-be-allocated items, quantifying the contribution degree of each feature of the to-be-allocated items to the allocation decision pattern, and generating a feature contribution degree ranking list of each feature; based on the feature contribution degree ranking list, filtering out core decision features with a contribution degree higher than a preset threshold; taking the core decision features as input variables of a mathematical expression and taking the prediction result of the target machine learning model as an approximate fitting target of the mathematical expression, and iteratively searching and optimizing the mathematical expression by using a symbolic regression algorithm; structurally analyzing and parameterizing the generated mathematical expression to obtain a preliminary logical rule; mapping variables, operators and constants in the preliminary logical rule into predefined business terms and adjustable business parameters to form the derived business rule that can be directly interpreted and executed by a rule engine.

6. The resource allocation method of claim 1, wherein, after the difference quantification analysis on the first initial allocation scheme and the second initial allocation scheme, when the difference degree of the first initial allocation scheme and the second initial allocation scheme exceeds a preset threshold, generating a difference root cause diagnosis report, the method further includes: based on the difference root cause diagnosis report, when it is analyzed that the first initial allocation scheme and the second initial allocation scheme are different and the logical correctness of the business rule corresponding to the first initial allocation scheme is correct, combining the first initial allocation scheme, the second initial allocation scheme, the description information of the to-be-allocated item, the difference root cause diagnosis report and the business rule that is logically correct as a rule verification sample; adding the rule verification sample to a training data set of the target machine learning model; in a subsequent training process of the target machine learning model, introducing a rule consistency loss term, wherein the rule consistency loss term is calculated according to a prediction output of the target machine learning model based on the business rule, and is used to punish the deviation of the output of the target machine learning model from the logic of the business rule; based on a total loss function combined with the rule consistency loss term, training and optimizing the target machine learning model until the output value of the target machine learning model meets the error.

7. The resource allocation method of claim 6, wherein, after introducing the rule consistency loss term in the subsequent training process of the target machine learning model, before the training and optimization of the target machine learning model based on the total loss function combined with the rule consistency loss term until the output value of the target machine learning model meets the error, the method further includes: extracting the target business rule causing the difference from the difference root cause diagnosis report; converting the target business rule into a differentiable constraint condition; Combine the constraint condition as a regularization term with an original loss function of the target machine learning model to form a new composite loss function; Train the target machine learning model based on the new composite loss function, so that the target machine learning model minimizes the prediction error and minimizes the deviation between the output of the target learning model and the target business rule.

8. A resource allocation apparatus characterized by comprising: Comprise: An information receiving module configured to receive description information of an item to be allocated; A calling module configured to call a pre-configured target business rule set and a target machine learning model matched with the description information; An allocation module configured to perform parallel calculation on the item to be allocated based on the target business rule set and the target machine learning model, respectively, to generate a first initial allocation scheme based on rules and a second initial allocation scheme based on models of the item to be allocated; A difference analysis module configured to perform difference quantification analysis on the first initial allocation scheme and the second initial allocation scheme, and generate a difference root cause diagnosis report when the difference between the first initial allocation scheme and the second initial allocation scheme exceeds a preset threshold; A construction module configured to construct at least one rule parameter optimization space based on the difference root cause diagnosis report; A suggestion generation module configured to find a target rule parameter in the rule parameter optimization space that makes a preset target function have an optimal solution, and generate a rule parameter optimization suggestion according to the target rule parameter; An optimization module configured to optimize the first initial allocation scheme and / or the second initial allocation scheme according to the rule parameter optimization suggestion to obtain a final allocation scheme of the item to be allocated.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The processor executes the computer program to realize the steps of the resource allocation method of any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 9. The computer program is executed by the processor to realize the steps of the resource allocation method of any one of claims 1 to 7.

Images