A rule and model based delinquent debt negotiation solution generation method

Abstract

This invention relates to the field of financial management technology, and in particular to a method for generating negotiation schemes for overdue debts based on rules and models. The method includes the following steps: S1, data acquisition and standardized preprocessing; S2, multi-level initial screening using a rule engine; S3, prediction using a multi-dimensional machine learning model; S4, intelligent generation of dynamic negotiation schemes; and S5, scheme output and closed-loop feedback optimization. This invention replaces manual experience-based decision-making with automated system execution, completing the entire process from data preprocessing to scheme output without human intervention. It enables rapid handling of massive amounts of overdue cases, solving the technical problems of low efficiency, inconsistent standards, high dependence on personnel qualifications, and inability to handle large-scale processing of massive cases through manual decision-making. This significantly reduces labor costs and processing time, while eliminating decision-making biases caused by subjective human judgment.

Description

Technical Field

[0001] This invention relates to the field of financial management technology, and in particular to a method for generating overdue debt negotiation schemes based on rules and models. Background Technology

[0002] With the upgrading of residents' consumption and the widespread adoption of inclusive finance, the scale of personal credit business continues to expand, and debt delinquency has become a core issue commonly faced by the credit industry. As a crucial step in post-loan asset preservation, delinquency debt negotiation, through flexible methods such as deferred repayment, interest and fee reduction, installment repayment, and lump-sum settlement, can alleviate the debtor's repayment pressure while maximizing the creditor's asset recovery benefits. It is an important means of balancing financial institutions' risk control objectives with customer service experience.

[0003] Currently, the generation of overdue debt negotiation plans in the industry mainly relies on the personal experience of collection specialists and asset preservation personnel, who draft negotiation plans on an ad-hoc basis after communicating with debtors. This model suffers from extremely low decision-making efficiency, inconsistent standards, and significant subjective bias. When faced with tens of thousands of overdue cases daily, the manual processing capacity is severely insufficient. Furthermore, the varying professional qualifications of personnel easily lead to risks such as unreasonable plans, violations of compliance commitments, and customer complaints. Moreover, the decision-making process lacks data support and is untraceable, failing to meet the refined management requirements of financial institutions. Therefore, there is a need to design a rule-based and model-based method for generating overdue debt negotiation plans. Summary of the Invention

[0004] In order to overcome the shortcomings of the existing technology, the purpose of this invention is to provide a method for generating overdue debt negotiation schemes based on rules and models.

[0005] The technical solution adopted in this invention is: a method for generating overdue debt negotiation schemes based on rules and models, comprising the following steps: S1. Data Acquisition and Standardized Preprocessing: Acquire the debtor's full-domain raw data, and after standardized preprocessing including outlier removal, duplicate data deduplication, invalid data filtering, intelligent missing value filling, feature encoding, normalization, and feature selection, generate a standardized input feature vector with fixed dimensions; the full-domain raw data includes basic attribute data, lending behavior data, historical repayment data, contact status data, and external credit data; S2. Multi-level initial screening by the rule engine: The input feature vector is input into a three-level rule engine equipped with compliance red line rules, product attribute rules, and customer status rules. The rule engine performs matching step by step according to the priority of compliance red line rules > product attribute rules > customer status rules, following the principle of veto. If the compliance red line rule is triggered, the result of refusal to negotiate is output. If the customer status rule is triggered, the result of manual handling is output. The product attribute rules are only used to constrain the parameter range of subsequent solutions. Cases that trigger the product attribute rules and cases that do not trigger any hard rules are marked as negotiable cases and proceed to the next step. At the same time, the rule engine outputs the parameter compliance boundary of the dual optimization objectives. S3. Multi-dimensional machine learning model prediction: The input feature vectors corresponding to negotiable cases are synchronously input into the pre-trained repayment ability model, repayment willingness model, loss of contact risk model, and life cycle value model, and the prediction results set including repayment ability range, repayment willingness probability, loss of contact risk probability, and life cycle value are obtained in parallel. S4. Intelligent generation of dynamic negotiation schemes: Construct a linear weighted dual-optimization objective function, guided by maximizing the dual-optimization objective function, and make intelligent decisions on the scheme type based on the prediction result set. Then, dynamically calculate the installment period parameters, down payment ratio parameters, interest and fee reduction ratio parameters, and negotiation validity period parameters. After secondary compliance verification by the rule engine, a structured negotiation scheme is generated. S5. Solution Output and Closed-Loop Feedback Optimization: The structured negotiation solution is output synchronously through multiple channels. The debtor's solution response data and performance behavior data are collected in full. The input feature vector, structured negotiation solution, and response results are used as supervised learning samples. The model in step S3 is incrementally trained and fully retrained on a regular basis to complete the model version management and iterative optimization.

[0006] As a further description of the above technical solution: In step S1, intelligent missing value filling includes filling continuous features with the median and filling discrete features with the mode; the feature encoding is to perform one-hot encoding on discrete features; the normalization is to scale continuous features to the [0,1] interval through min-max normalization; and the feature selection is to retain high-discrimination core features based on Pearson correlation coefficient and feature importance ranking.

[0007] As a further description of the above technical solution: The compliance red line rule in step S2 is used to intercept illegal negotiation behaviors such as reduction of principal of written-off loans, full exemption of interest and fees in the judicial stage, and prohibited by regulations; the product attribute rule is used to limit the number of installments and the upper limit of reduction ratio for different loan products; the customer status rule is used to divert special customer cases such as bankruptcy, litigation, malicious debt evasion, and long-term loss of contact.

[0008] As a further description of the above technical solution: In step S3, the repayment ability model, the loss of contact risk model, and the lifetime value model adopt the LightGBM model, while the repayment willingness model adopts the XGBoost model.

[0009] As a further description of the above technical solution: The formula for the dual-optimization objective function in step S4 is: ,in, , ; In the formula: The sub-target is a 30-day negotiated signing rate. The sub-target is the net present value of debt recovery within 90 days. These are the weighting coefficients. This represents the probability of willingness to repay. This represents the probability of losing contact. For the sake of solution adaptability, This represents the total amount owed. This represents the average of the repayment ability range. For the number of periods in the installment, For the down payment ratio, For the interest and fee reduction ratio, The monthly discount rate is This refers to the number of repayment periods.

[0010] As a further description of the above technical solution: The intelligent decision-making logic for the scheme type in step S4 is as follows: when the probability of loss of contact is greater than 0.7, the one-time reduction and settlement scheme is selected; when the probability of willingness to repay is less than 0.3 and the lower limit of repayment ability is greater than or equal to 10% of the total debt, the scheme with no reduction or exemption of principal installment interest and fees is selected; in other cases, the custom installment scheme is selected.

[0011] As a further description of the above technical solution: In step S4, the number of installments is determined by rounding up the ratio of the total outstanding amount to the average of the repayment ability range, and then matching it with the standard number of installments allowed by the rules; the down payment ratio is inversely proportional to the probability of repayment willingness and directly proportional to the probability of loss of contact; the interest and fee reduction ratio is directly proportional to the lifetime value and does not exceed the upper limit set by the rule engine; the negotiation validity period is directly proportional to the probability of repayment willingness.

[0012] As a further description of the above technical solution: The response data in step S5 includes acceptance, rejection, proposed modification, loss of contact, and failure to confirm within time limit. The performance data includes timely repayment, overdue repayment, early settlement, and default.

[0013] As a further description of the above technical solution: In step S5, incremental training is performed once a week, using only new samples to update model weights; full retraining is performed once a month, using both new and historical samples for joint training.

[0014] The present invention has the following beneficial effects: 1. This invention replaces human experience-based decision-making with automatic system execution, completing the entire chain from data preprocessing to solution output without human intervention. It enables the rapid handling of massive overdue cases, solving the technical problems of low efficiency, inconsistent standards, high dependence on personnel quality, and inability to handle large-scale cases in human decision-making. It significantly reduces labor costs and processing time, while eliminating decision-making bias caused by subjective human judgment.

[0015] 2. This invention employs a three-tiered rule engine. The rule engine utilizes a dual mechanism of pre-screening and post-verification, incorporating three layers of hard rules: compliance red lines, product attributes, and customer status. This pre-screening intercepts cases of non-compliance negotiations, while the post-verification checks the legality of the solution parameters. This ensures that all generated solutions comply with regulatory requirements and institutional business policies, effectively mitigating the risks of compliance violations, customer complaints, and regulatory penalties arising from manual decision-making.

[0016] 3. This invention constructs a comprehensive profile of debtors from four dimensions: repayment ability, repayment willingness, risk of loss of contact, and lifecycle value, enabling the generation of customized solutions tailored to each case and individual. The solution parameters are highly compatible with the debtor's actual repayment amount, strength of repayment willingness, probability of loss of contact, and long-term customer value, fundamentally improving the debtor's acceptance of the solution.

[0017] 4. This invention, based on a dual-optimization objective function and standardized implementation logic, dynamically quantifies and calculates core parameters such as installment periods, down payment ratio, interest and fee reduction ratio, and plan validity period. It prioritizes one-time settlement plans for customers with high risk of losing contact, offers appropriate concessions to retain high-value customers, and increases down payment constraints for customers with low willingness to pay, thereby maximizing both short-term revenue and long-term customer value. Attached Figure Description

[0018] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0019] Reference Figure 1 This invention provides a method for generating overdue debt negotiation schemes based on rules and models, comprising the following steps: S1. Data Acquisition and Standardization Preprocessing: This step is the data foundation of the entire solution generation system. It is responsible for transforming raw business data into standardized data that can be recognized by models and rules. The entire process is automated and requires no human intervention. It consists of five sub-steps: full-domain data collection, multi-level data cleaning, intelligent missing value imputation, feature encoding and normalization, and feature selection and vector construction.

[0020] 1.1 Comprehensive Multi-Dimensional Data Collection: The system retrieves raw data from the entire debtor database in real time through internal business data interfaces, post-loan management system interfaces, and external credit compliance interfaces, covering five core data dimensions, specifically including: Basic attribute data: age, gender, marital status, occupation type, average monthly income, residential stability, and social security payment status; Loan behavior data: loan product type, credit limit, remaining principal, total interest and fees, annualized interest rate, number of overdue days, number of overdue payments; Historical repayment data: historical performance records, historical negotiation records, historical repayment amounts, and repayment frequency; Contact status data: call connection rate, SMS response rate, APP activity frequency, number of effective contact channels; External credit data: central bank credit score, overdue records of other banks, information on dishonest persons subject to enforcement, and third-party big data risk control score.

[0021] 1.2 Multi-level data cleaning: Outlier removal: For continuous numerical features such as monthly income, repayment amount, and overdue days, the 3σ principle is used for anomaly detection, and extreme values ​​that exceed the range of [μ-3σ,μ+3σ] are directly removed. Deduplication of data: Using the debtor's unique identifier (ID number + loan contract number) as the primary key, automatically delete duplicate reports and duplicate data retrievals for the same case; Invalid data filtering: Remove invalid samples with a field missing rate of more than 50%, and retain valid data that meets the data integrity standard and can be used for rule matching and model prediction.

[0022] 1.3 Intelligent missing value imputation: Continuous feature missing: Monthly income, number of overdue days, repayment amount, etc. are filled with the median to avoid the mean being affected by extreme values; Discrete feature missing: Occupation type, product type, contact channel, etc. are filled with the mode to ensure the consistency of feature distribution; Missing core features: If core features such as credit score, valid contact number, and loan contract number are missing, the sample will be directly marked as a high-risk invalid sample, and the subsequent manual intervention rule will be triggered.

[0023] 1.4 Feature Encoding and Normalization: Discrete feature encoding: For non-numerical features such as gender, occupation type, and loan product type, one-hot encoding is used to convert them into binary numerical vectors that the model can process; Continuous feature normalization: Continuous features of different dimensions, such as income, number of overdue days, and repayment amount, are scaled to the [0,1] interval through min-max normalization to eliminate the impact of differences in dimensions on model prediction; Time feature standardization: Convert time features such as length of residence, number of overdue days, and repayment cycle into standardized numerical features with a unified unit.

[0024] 1.5 Feature Selection and Input Vector Construction: By combining Pearson correlation coefficient analysis and model feature importance ranking, redundant features with correlation below 0.1 and no contribution to prediction are removed, and 48 high-discrimination core features are retained. Finally, a 1×48-dimensional standardized input feature vector X is generated as the unified input data for the rule engine and machine learning model.

[0025] Specifically; This step involves fully automated processing of data across the entire domain, including automatic data collection, multi-level noise cleaning, intelligent missing value imputation, standardized feature encoding and normalization, and high-value feature screening. This process completely transforms the raw, messy, inconsistently formatted, unstructured business data containing missing and anomalies from financial institutions' systems into standardized input feature vectors X that are uniformly formatted, compliant in quality, noise-free, dimensionless, and highly discriminative. This step fundamentally solves the technical problem of raw data being unusable for rule matching and model calculation. It provides a unique, stable, high-quality, and reusable data input foundation for subsequent steps S2 (rule engine verification) and S3 (multi-dimensional model prediction), serving as the starting point and data guarantee for the entire negotiation scheme generation process.

[0026] S2, rule engine multi-level initial screening: This step serves as a compliance barrier and case triage link in the entire solution generation system. It leverages the Drools business rule engine to achieve visualized configuration and real-time matching. The rule engine's implementation logic is as follows: it constructs a three-tiered rule base consisting of compliance red line rules, product attribute rules, and customer status rules, arranged irreversibly from highest to lowest priority; it receives standardized input feature vectors and performs precise matching and interval matching, adhering to the veto principle, triggering any high-priority rule to immediately terminate the process and output the corresponding decision result; the rule engine also generates parameter compliance boundaries for subsequent dual optimization objectives.

[0027] 2.1 Visual configuration of the three-tier rule base: The rules engine supports visual configuration in the backend and takes effect in real time. Rules are divided into three levels according to compliance priority from high to low. All rules comply with regulatory requirements and internal business policies of the organization. Compliance red line rules (highest priority): Principal reduction is prohibited for written-off loans; promises of full interest and fee waivers are prohibited during judicial proceedings / enforcement; zero down payment / ultra-long-term installment clauses that are expressly prohibited by regulators will result in immediate refusal of negotiation. Product Attribute Rules (Medium Priority): For short-term overdue cases with no more than 30 days of overdue payment, the installment period is limited to no more than 6 installments; for credit-based consumer loan products, the principal reduction ratio is limited to a maximum of 20%; for mortgage loan products, the collateral disposal process is prioritized, and flexible negotiation is not initiated. This type of rule only restricts the range of parameter values ​​for subsequent solutions; hitting the rule will not terminate the case from entering the negotiation process. Customer status rules (basic priority): Bankruptcy liquidation, declared death, or customers who have been out of contact for 180 days will be automatically transferred to special handling process; customers who are determined by the court to be maliciously evading debts will be directly refused negotiation; cases under litigation will be automatically transferred to manual legal affairs processing, and after being triggered, the case will be transferred to the manual channel.

[0028] 2.2 Hierarchical rule matching and execution: The standardized input feature vector X generated by S1 is input into the rule engine for layer-by-layer verification and veto. If any high-level rule is triggered, subsequent rule matching is immediately terminated. Prioritize verification of compliance red line rules: cases have not been written off, have not entered judicial litigation process, and have not triggered any red line rules; Secondly, the product attribute rules were verified: overdue for 92 days > 30 days, with no short-term overdue installment restrictions, and the product rules were not triggered. Finally, the customer status rules were verified: the debtor can be contacted normally, there are no special abnormalities, and the customer status rules have not been triggered.

[0029] 2.3 Case Result Distribution and Output: Based on the matching results, the rules engine categorizes cases into three types: Triggering the non-negotiable compliance red line rule: Output the result {"result":"reject","reason":"compliance_violation"}, and the process will terminate immediately; Triggering special customer status rules: Output the result {"result":"transfer","reason":"special_customer_status"}, and the process will terminate immediately; Only if the product attribute rule is matched: the case is not terminated, only the parameter constraints are retained, and the case continues to flow to S3; No hard rules were triggered: the case status was marked as "negotiable", allowing it to enter the subsequent model prediction stage. At the same time, the compliance boundaries of installment period, interest and fee reduction, and down payment ratio were output, providing constraints for the optimization of the dual optimization target parameters in step S4.

[0030] Specifically: This step utilizes a rule engine equipped with three levels of compliance and business rules as its core, strictly adhering to a three-tiered priority, step-by-step matching, and veto logic. It rigorously receives the standardized input feature vector X from step S1, prioritizing compliance red lines, followed by product rules, and then customer status as a fallback, executing a hard-line verification logic of step-by-step matching and veto. Its core function is to quickly complete compliance screening, business verification, and special status triage of cases before entering model calculation. It intercepts cases that violate regulatory requirements and institutional policies at the source, diverts customers in special states such as bankruptcy, litigation, and malicious debt evasion, and only marks cases that trigger product attribute rules and those that do not trigger any hard rules as negotiable cases, allowing them to proceed to the subsequent model prediction stage. Simultaneously, the rule engine generates compliance boundaries for the dual optimization objectives of step S4, limiting the legal range of installment periods, reduction / exemption ratios, and down payment ratios, preventing illegal parameters from entering the optimization stage. This step not only strengthens the compliance defenses of the entire process, effectively avoiding compliance risks and regulatory penalties, but also significantly reduces the unnecessary computational resource consumption of the S3 model by diverting invalid cases in advance, thereby improving the overall process efficiency. As a key screening node connecting data preparation and model prediction, S2 directly receives the data output from S1, while providing clean, valid, and compliant case inputs for S3 and parameter compliance constraints for S4. It is the core hub that ensures the compliance and efficiency of the process.

[0031] S3, Multi-dimensional Machine Learning Model Prediction: This step is the user profiling stage of the entire solution generation system. The effective input feature vector X selected in step S2 is input into four pre-trained machine learning models. Through parallel prediction, the quantitative results of the debtor's core capabilities and risks are output. The model adopts a hybrid model architecture of LightGBM, XGBoost and deep neural network (DNN), and is trained based on the TensorFlow framework, supporting millisecond-level batch prediction.

[0032] 3.1 Model Selection and Industrial-Grade Pre-training: Model selection: LightGBM (adapts to tabular data and has high prediction efficiency) is used for the repayment ability model, loss of contact risk model, and lifetime value model; XGBoost (high prediction accuracy and strong stability) is used for the repayment willingness model; Deep neural networks (DNN) can be used to replace complex nonlinear feature scenarios. Dataset partitioning: Historical overdue case samples were divided into training set, validation set, and test set in a 7:2:1 ratio, and 5-fold cross-validation was used to optimize the model hyperparameters; Model evaluation: Accuracy, AUC, and mean squared error (MSE) are the core indicators to ensure that the model AUC is ≥ 0.85 and meets the business forecast accuracy requirements. Model Deployment: Models are deployed in a standardized ONNX format, supporting high-concurrency, low-latency online prediction services.

[0033] 3.2 Independent implementation logic of the four existing models: All existing models share the 1×48-dimensional standardized input feature vector X output by S1. The features are divided into 5 categories, providing a unified input for the models: Basic attributes and characteristics: age, marital status, occupation, monthly income, debt expenditure, and residential stability; Characteristics of lending behavior: product type, amount owed, number of overdue days, number of historical overdue payments, annualized interest rate; Historical repayment characteristics: normal repayment rate, proportion of voluntary repayments, historical average monthly repayments, and negotiation records; Contact status characteristics: call connection rate, SMS response rate, APP activity, effective contact channels; External credit characteristics: credit score, risk control score, record of dishonesty, and overdue payment status at other banks.

[0034] 3.2.1 Repayment Ability Model M_afford: Model type: Existing LightGBM regression model (supervised learning, outputting a continuous numerical range); Core task: Based on existing models, predict the range of average monthly repayment amounts that debtors can afford over the next three months; Input characteristics: monthly income, household fixed debt, historical average monthly repayment, credit score, and number of overdue days; Training implementation logic: Dataset: Historical overdue case samples, labeled with the debtor's actual monthly repayment amount (real business label); Preprocessing: median imputation of missing values, feature normalization, outlier removal; Training: The existing LightGBM framework is used, with 5-fold cross-validation, and the optimization objective is to minimize the mean squared error (MSE). Hyperparameters: learning rate 0.05, tree depth 6, leaf nodes 31, number of iterations 100 (existing standard parameters); Evaluation: The mean absolute error (MAE) is used as the indicator, and an error of ≤10% is considered acceptable; Output definition: Repayment capacity range A_cap=[A_min,A_max], unit: yuan.

[0035] 3.2.2 Repayment Willingness Model M_willing: Model type: Existing XGBoost binary classification model; Core task: To predict the probability that a debtor will accept the negotiated solution and fulfill its obligations based on existing models; Input characteristics: proactive repayment ratio, number of historical overdue payments, channel response rate, credit score, and plan acceptance history; Training implementation logic: Dataset: Historical overdue case samples, labeled 1 = acceptance of performance, 0 = rejection / default; Preprocessing: one-hot encoding, sample balancing (SMOTE oversampling, existing technology), feature selection; Training: Using the existing XGBoost framework, with 5-fold cross-validation, the optimization objective is binary classification cross-entropy; Hyperparameters: learning rate 0.03, tree depth 5, L2 regularization term 1 (existing standard parameters); Assessment: AUC ≥ 0.85, accuracy ≥ 0.8 (standard threshold for post-loan scenarios); Output definition: Probability of repayment willingness P_willing∈[0,1].

[0036] 3.2.3 Loss of Contact Risk Model M_contact: Model type: Existing LightGBM binary classification model; Core task: To predict the probability of debtors becoming unreachable based on existing models; Input characteristics: connection rate in the past 7 days, APP activity in the past 30 days, length of residence, number of contact channels; Training implementation logic: Dataset: Historical overdue case samples, labeled 1 = out of contact, 0 = contactable; Training: Using the existing LightGBM framework, the optimization objective is binary classification cross-entropy; Assessment: AUC ≥ 0.82, meeting the accuracy requirements for loss of contact prediction; Output definition: the probability of losing contact P_lost∈[0,1].

[0037] 3.2.4 Life Cycle Value Model M_ltv: Model type: Existing LightGBM regression model; Core task: To predict the comprehensive long-term customer value score of debtors based on existing models; Input characteristics: product usage duration, normal repayment period, credit limit, and historical contribution earnings; Training implementation logic: Dataset: Historical customer lifecycle data, labeled with the customer's cumulative net revenue; Training: Using the existing LightGBM framework, the optimization objective is to minimize MSE; The output is normalized to the [0,1] interval; Output definition: Lifetime value score V_ltv∈[0,1].

[0038] 3.3 Parallel prediction execution of four major models: The feature vector X filtered by S2 is simultaneously input into the four models, with no sequential dependency and parallel computation. The output results are retained to 4 decimal places. The repayment capacity model M_afford: Input monthly income, household debt, historical repayments, credit score and other features to predict the range of average monthly repayment amount A_cap that can be afforded in the next 3 months; The repayment willingness model M_willing: Input features such as the proportion of proactive repayments, the number of historical overdue payments, the channel response rate, and credit score, and predict the probability P_willing of accepting and fulfilling the agreement; The model for the risk of being unable to contact someone, M_contact, takes into account features such as call connection rate, APP activity frequency, and residential stability, and predicts the probability P_lost that the person cannot be contacted through the current channel. Lifetime Value Model M_ltv: Input features such as product usage duration, normal repayment period, and credit limit to predict long-term comprehensive value score V_ltv.

[0039] 3.4 Aggregation of Prediction Results Set: The quantization results output by the four models are integrated into a unified prediction result set Y_pred, with the format: Y_pred={A_cap,P_willing,P_lost,V_ltv}, which serves as the core input data for generating the S4 dynamic scheme.

[0040] Specifically: This step strictly follows the valid standardized input feature vector X selected in step S2. It employs a multi-model parallel prediction architecture, utilizing four independent pre-trained models—repayment ability, repayment willingness, risk of default, and lifetime value—to create a comprehensive and quantitative profile of the debtor. This step transforms the abstract user features collected in S1 and the compliant case information verified in S2 into calculable, comparable, and parameter-generating numerical results. This overcomes the technical limitations of traditional techniques that rely solely on a single repayment probability model and cannot comprehensively assess the debtor's true situation. The output prediction result set Y_pred serves as the sole quantitative calculation basis for generating the dynamic negotiation scheme in step S4, directly determining the accuracy of scheme type selection and core parameter calculations.

[0041] S4. Intelligent generation of dynamic negotiation solutions: This step is the core generation stage of the entire solution generation system. With the dual optimization objectives of maximizing the 30-day negotiation and signing rate and maximizing the 90-day debt recovery net present value, a linear weighted objective function is constructed. Following the standardized implementation logic of linear weighted quantization conversion, rule-constrained parameter optimization, secondary compliance verification, and closed-loop data iterative correction, based on the Y_pred quantization result output in step S3, the solution type is first determined, then the core parameters are calculated in detail, and after secondary compliance verification by the rule engine, a feasible structured solution is generated.

[0042] 4.1 Construction of the dual optimization objective function: This step uses a linearly weighted multi-objective function as the core optimization principle, as shown in the following formula:

[0043] The sub-target for the 30-day negotiated signing rate: ,

[0044] The sub-objective for 90-day debt recovery net present value:

[0045] In the above formula The down payment is collected in the period of signing the contract, without the need for discounting, which meets the financial cash flow accounting rules; the constraints are: W≤Wmax, N∈3,6,9,12,18,24, Downmin≤Down≤1, T∈[24,72] hours, which simultaneously meets the compliance red line and product attribute rules.

[0046] In the formula: These are the weighting coefficients. This represents the probability of willingness to repay. This represents the probability of losing contact. For the sake of solution adaptability, This represents the total amount owed. This represents the average of the repayment ability range. For the number of periods in the installment, For the down payment ratio, For the interest and fee reduction ratio, This is the maximum reduction / exemption. The monthly discount rate is This refers to the number of repayment periods.

[0047] 4.2 Execution of the logic for achieving the dual optimization objectives: Target scalarization transformation: The dual objectives of 30-day contract signing rate and 90-day net present value are integrated into a single comprehensive objective function through linear weighting, solving the problem of parallel optimization of multiple objectives; Compliance parameter space generation: Using the rule engine in step S2 as the boundary, the legal range of installment period, reduction ratio, and down payment ratio is limited, and illegal parameter combinations are eliminated; Traversal optimization calculation: Enumerate all compliant parameter combinations, substitute them into the objective function to calculate the score, and select the optimal parameter combination with the highest score; Secondary compliance verification: The optimal parameters are sent back to the rule engine for re-verification to ensure that the solution is 100% compliant; Iterative correction preparation: The solution data is retained for subsequent model iterative optimization in step S5.

[0048] 4.3 Intelligent Decision-Making Based on Solution Type: Execute unique matching decision logic based on preset thresholds to ensure a high degree of compatibility between the solution type and the user's state: If the probability of loss of contact P_lost > 0.7, choose a one-time reduction and settlement plan to quickly lock in and recover funds; If the probability of repayment willingness P_willing < 0.3 and the lower limit of the repayment ability range is ≥ 10% of the total debt: choose the principal installment plan with no reduction or exemption of interest and fees to guarantee the amount recovered; Other options: Choose a custom installment plan to suit the repayment needs of most debtors.

[0049] 4.4 Quantitative Calculation of Core Parameters: All parameters are calculated using mathematical formulas and are strictly constrained by the rules engine's limits to ensure compliance and relevance to actual user needs. Installment period N: Formula N=ceil(D / A_cap_mean), where A_cap_mean is the average of the repayment ability range; for example: if D is 10000 and A_cap_mean is 1000, the calculated ceil(10000 / 1000)=10. The upward matching rule allows the standard number of installments, so the final N=12 installments; Down payment ratio: Formula Down = α × (1 - P_willing) + β × P_lost (α = 0.4, β = 0.3 are preset business weights); For example: P_willing is 0.65, P_lost is 0.25, calculated as 0.4 × 0.35 + 0.3 × 0.25 = 0.215, rounded down = 22%; Interest and fee reduction ratio W: Formula W=γ×V_ltv (γ=0.2), and W≤W_max (rule upper limit W_max=30%); For example: V_ltv is 0.7, calculated as 0.2×0.70=0.14, that is, W=14%, which meets the upper limit constraint; Negotiated validity period T: Formula T = δ × P_willing (δ = 72 hours); For example: P_willing is 0.65, calculated as 72 × 0.65 = 46.8, rounded up to the standard duration, the final T = 48 hours.

[0050] 4.5 Secondary compliance verification of the solution: The calculated optimal parameters are then sent back to the rule engine in step S2 for secondary target validation. The validation includes: 12 installments: This meets the installment rules for cases overdue for 90 days. Interest and fee reduction of 14%: lower than the 30% cap, compliant; Down payment ratio of 22%: No violations or restrictions, compliant; If all checks pass, the final solution can be generated; if the checks fail, the parameter combination is discarded and the optimization process begins again.

[0051] 4.6 Structured Scheme Output: Generate a standardized, ready-to-use negotiation plan, containing all core information: {Plan Type, Total Outstanding Amount, Installment Period N, Down Payment Ratio Down, Down Payment Amount, Monthly Payment, Interest and Fee Reduction Ratio W, Reduction Amount, Negotiation Validity Period T, Repayment Deadline} Specifically: This step focuses on maximizing the dual-optimization objective function as its core business objective. It strictly adheres to the implementation logic of linear weighted scalarization transformation, rule-constrained parameter optimization, secondary compliance verification, and closed-loop data iterative correction. It relies entirely on the quantitative prediction result set Y_pred output from step S3 as its computational foundation. First, a linear weighted multi-objective optimization function is constructed, clearly defining the dual optimization orientation of the 30-day signing rate and the 90-day net present value. Then, intelligent decision-making on the solution type is completed through threshold matching logic. Pre-set mathematical formulas are used to perform refined dynamic calculations of four core parameters: installment period, down payment ratio, interest and fee reduction ratio, and negotiation validity period. The entire calculation process is constrained by the upper limit of the S2 rule engine. After calculation, a secondary compliance verification by the rule engine is performed to ensure that the final solution not only reflects the debtor's actual situation but also fully complies with regulatory and business rules. This step transforms the user quantitative profile generated in S3 into a structured negotiation solution Plan that can be directly pushed, executed, and signed. This is the core output of the invention's technical solution, solving the problems of rigidity, fixed parameters, and lack of personalization in traditional negotiation solutions.

[0052] S5. Solution Output and Closed-Loop Feedback Optimization: This step is the implementation and self-iteration phase of the entire solution generation system. The solution generated in step S4 is pushed to multiple channels, all debtor response and performance data are collected, new samples are formed, and the prediction model in S3 is optimized regularly to build a complete closed loop of data, rules, models, solutions, feedback, and iteration.

[0053] 5.1 Simultaneous output of multi-channel solutions: The system uses a multi-port synchronous push mechanism to ensure that debtors, collection personnel, and customer service agents can all view the information in real time. On the business system side: push to the collection system, intelligent outbound call system, and customer service agent system via RESTful API, and return a standard JSON format solution; User outreach: Push solution details via SMS, APP messages, and mini-program pop-ups, along with one-click confirmation and consultation links; Agreement document: Automatically generates a standardized negotiated agreement in PDF format, supporting download, printing, and signing.

[0054] 5.2 Full collection of response and performance data: We collect debtor response data in its entirety through three methods: front-end tracking, system integration, and manual data entry. Basic response data: Accept, Reject, Proposed Modification, Lost Contact, Timeout and Unconfirmed; Performance data: On-time repayment, overdue repayment, early settlement, default; Interaction behavior data: duration of viewing the solution, number of clicks, number of inquiries.

[0055] 5.3 Valid Sample Entry and Labeling: The input feature vector X, the generated plan, and the response result Result are stored as a complete sample in the sample database, and the following steps are completed simultaneously: Valid sample screening: Remove invalid samples with missing data or timeouts, and retain valid samples that can be used for model training; Sample labeling: Accept = 1, Reject = 0, Disconnection = -1, forming supervised learning labels.

[0056] 5.4 Model Incremental and Full Training: Training cycle: Incremental training is performed once a week (updating model weights with only new samples), and full retraining is performed once a month (joint training with new samples and historical samples). Training logic: Without changing the model structure, only optimize the model parameters to quickly improve prediction accuracy; Performance validation: After training, compare AUC and accuracy. If the performance index improves by ≥1%, replace the old model.

[0057] 5.5 Model Version Management and Retrospection: The system manages model version numbers (V1.0, V1.1, V2.0) and supports one-click rollback to historical versions to ensure business stability. It also records the sample size, metric changes, and parameter adjustments for each training session, making the entire process traceable and auditable.

[0058] Specifically: This step strictly follows the structured negotiation plan generated in step S4, synchronously pushing and disseminating the plan through multiple channels and ports to realize its business implementation. Simultaneously, it collects all debtor responses, interaction behaviors, and repayment data, integrating the input feature vector X, the generated plan, and the actual response results into a complete supervised learning sample. After labeling, this sample is stored in a sample library. Incremental training and full retraining are then performed on the multi-dimensional prediction model from step S3 at fixed intervals to optimize and update model parameters. Version management ensures the stability of model iteration. This step serves as both the final execution point of the entire process, achieving a closed-loop business process from plan generation to dissemination, and the system's self-evolution engine. It feeds back feedback data from real business scenarios to the model stage in step S3, continuously correcting model prediction biases, improving the accuracy of plan generation, and ensuring that the calculation results of the dual-optimization objective function better align with actual business scenarios. It also provides data support for rule engine optimization, addressing the technical shortcomings of traditional static models that cannot adapt to changes in business and customer characteristics.

[0059] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A rule and model based delinquent debt negotiation scenario generation method, characterized by, Includes the following steps: S1. Data Acquisition and Standardized Preprocessing: Acquire the debtor's full-domain raw data, and after standardized preprocessing including outlier removal, duplicate data deduplication, invalid data filtering, intelligent missing value filling, feature encoding, normalization, and feature selection, generate a standardized input feature vector with fixed dimensions; the full-domain raw data includes basic attribute data, lending behavior data, historical repayment data, contact status data, and external credit data; S2. Multi-level initial screening by the rule engine: The input feature vector is input into a three-level rule engine equipped with compliance red line rules, product attribute rules, and customer status rules. The rule engine performs matching step by step according to the priority of compliance red line rules > product attribute rules > customer status rules, following the principle of veto. If the compliance red line rule is triggered, the result of refusal to negotiate is output. If the customer status rule is triggered, the result of manual handling is output. The product attribute rules are only used to constrain the parameter range of subsequent solutions. Cases that trigger the product attribute rules and cases that do not trigger any hard rules are marked as negotiable cases and proceed to the next step. At the same time, the rule engine outputs the parameter compliance boundary of the dual optimization objectives. S3. Multi-dimensional machine learning model prediction: The input feature vectors corresponding to negotiable cases are synchronously input into the pre-trained repayment ability model, repayment willingness model, loss of contact risk model, and life cycle value model, and the prediction results set including repayment ability range, repayment willingness probability, loss of contact risk probability, and life cycle value are obtained in parallel. S4. Intelligent generation of dynamic negotiation schemes: Construct a linear weighted dual-optimization objective function, guided by maximizing the dual-optimization objective function, and make intelligent decisions on the scheme type based on the prediction result set. Then, dynamically calculate the installment period parameters, down payment ratio parameters, interest and fee reduction ratio parameters, and negotiation validity period parameters. After secondary compliance verification by the rule engine, a structured negotiation scheme is generated. S5. Solution Output and Closed-Loop Feedback Optimization: The structured negotiation solution is output synchronously through multiple channels. The debtor's solution response data and performance behavior data are collected in full. The input feature vector, structured negotiation solution, and response results are used as supervised learning samples. The model in step S3 is incrementally trained and fully retrained on a regular basis to complete the model version management and iterative optimization.

2. The method of claim 1, wherein, In step S1, intelligent missing value filling includes filling continuous features with the median and filling discrete features with the mode; the feature encoding is to perform one-hot encoding on discrete features; the normalization is to scale continuous features to the [0,1] interval through min-max normalization; and the feature selection is to retain high-discrimination core features based on Pearson correlation coefficient and feature importance ranking.

3. The method of claim 1, wherein, The compliance red line rule in step S2 is used to intercept illegal negotiation behaviors such as principal reduction of written-off loans, full exemption of interest and fees during the judicial stage, and prohibited by regulations; the product attribute rule is used to limit the number of installments and the upper limit of reduction ratio for different loan products. The customer status rules are used to divert special customer cases such as bankruptcy, litigation, malicious debt evasion, and long-term loss of contact.

4. The method of claim 1, wherein, In step S3, the repayment ability model, the loss of contact risk model, and the lifetime value model adopt the LightGBM model, while the repayment willingness model adopts the XGBoost model.

5. The method for generating overdue debt negotiation schemes based on rules and models according to claim 1, characterized in that, The formula for the dual-optimization objective function in step S4 is: ,in, , ; In the formula: The sub-target is a 30-day negotiated signing rate. The sub-target is the net present value of debt recovery within 90 days. These are the weighting coefficients. This represents the probability of willingness to repay. This represents the probability of losing contact. For the sake of solution adaptability, This represents the total amount owed. This represents the average of the repayment ability range. For the number of periods in the installment, For the down payment ratio, For the interest and fee reduction ratio, The monthly discount rate is This refers to the number of repayment periods.

6. The method for generating overdue debt negotiation schemes based on rules and models according to claim 1, characterized in that, The intelligent decision-making logic for the scheme type in step S4 is as follows: when the probability of loss of contact is greater than 0.7, the one-time reduction and settlement scheme is selected; when the probability of willingness to repay is less than 0.3 and the lower limit of repayment ability is greater than or equal to 10% of the total debt, the scheme with no reduction or exemption of principal installment interest and fees is selected; in other cases, the custom installment scheme is selected.

7. The method for generating overdue debt negotiation schemes based on rules and models according to claim 1, characterized in that, In step S4, the number of installments is determined by rounding up the ratio of the total outstanding amount to the average of the repayment ability range, and then matching it with the standard number of installments allowed by the rules; the down payment ratio is inversely proportional to the probability of repayment willingness and directly proportional to the probability of loss of contact; the interest and fee reduction ratio is directly proportional to the lifetime value and does not exceed the upper limit set by the rule engine. The validity period of the negotiation is directly proportional to the probability of repayment willingness.

8. The method for generating overdue debt negotiation schemes based on rules and models according to claim 1, characterized in that, The response data in step S5 includes acceptance, rejection, proposed modification, loss of contact, and failure to confirm within time limit. The performance data includes timely repayment, overdue repayment, early settlement, and default.

9. The method for generating overdue debt negotiation schemes based on rules and models according to claim 1, characterized in that, In step S5, incremental training is performed once a week, using only new samples to update model weights; full retraining is performed once a month, using both new and historical samples for joint training.

Images