A decision assistance method and system for personalized scenarios

Abstract

The present application relates to the technical field of intelligent decision-making assistance for financial marketing, and particularly relates to a decision-making assistance method and system for personalized scenarios. The method comprises the following steps: constructing a fuzzy complementary judgment matrix, calculating an initial weight vector based on the fuzzy complementary judgment matrix, obtaining a candidate weight vector through row coupling degree and column coupling degree constraint optimization, generating a personalized weight vector according to the vector similarity between a borrower feature vector and the cluster centers of each customer cluster, synthesizing the borrower feature vector and the personalized weight vector to obtain a decision result, performing credit compliance verification on the decision result, and outputting the decision result if the verification is passed. The present application obtains a comprehensive scenario adjustment vector by using the vector similarity between a borrower feature vector and the cluster centers of each customer cluster, and optimizes the candidate weight vector using the comprehensive scenario adjustment vector to obtain a personalized weight vector, so that personalized credit decision-making can be realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent decision support technology in financial marketing, and in particular to a decision support method and system for personalized scenarios applied to home renovation loan business. Background Technology

[0002] With the development of network communication technology, existing technologies have provided solutions for online consultation and processing of home renovation loans. For example, the home renovation loan processing method disclosed in Chinese patent application CN117273910A uses an online processing device with multiple modules to allow users to directly begin consultation and processing of home renovation loans after finalizing a renovation plan with a renovation company, thereby reducing the time spent on renovation inquiries and loan processing and improving user experience. However, the aforementioned method only achieves online processing and does not provide measures to assist relevant personnel in decision-making. It only saves relevant personnel's travel time and expenses, but does not improve the decision-making efficiency in the actual business processing process. Various information affecting decision-making, such as user credit, income status, job stability, property value, and renovation budget, often involve a large number of vague and uncertain factors. Furthermore, the credit constraints and compliance requirements that financial institutions must implement when issuing loans are also complex and diverse. Therefore, traditional credit decision-making still relies on the experience and ability of relevant personnel, and its commonly used static scoring card mechanism cannot deeply identify the personalized characteristics of different borrowers. Faced with multi-dimensional and interdependent evaluation indicators, existing methods often fail to accurately detect the matching relationship between borrowers' true risk exposure and personalized credit needs, resulting in an inability to make scientific and reasonable loan decisions under complex credit constraints.

[0003] To reduce risk and improve decision-making efficiency and accuracy, existing technologies include solutions for loan decisions based on multi-dimensional information. For example, the loan calculation method disclosed in Chinese patent application CN115564569A calculates renovation factor values ​​and property value factor values ​​based on the customer's purchase orders, sets corresponding weights based on customer information such as age and city of residence, and finally calculates the customer's loanable amount based on the renovation factor value, property value factor value, and various weights, thereby improving the accuracy of the loan amount assessment. While this method introduces evaluation indicators corresponding to multiple types of information into the decision-making process, the scoring of each indicator still relies on subjective experience and lacks consideration of the relative importance of each indicator, resulting in insufficient decision-making precision.

[0004] Furthermore, existing technologies also include solutions that organically combine mathematical models with credit indicators to quickly and accurately assess credit risk. For example, the assessment model for influencing loan decisions disclosed in Chinese patent application CN115936841A constructs a fuzzy complementary judgment matrix based on pairwise comparisons of lending-related business indicators. It then calculates the global weights of various indicators based on this matrix, thereby constructing a credit risk assessment model. While this model pre-establishes a fuzzy complementary judgment matrix to reflect the relative importance of each indicator and can integrate multiple expert scores to reduce individual bias, its global indicator weights are fixed and cannot be customized to suit borrower characteristics and credit granting scenarios. This makes it difficult to adapt to the differentiated credit needs of different customer groups, resulting in rigid credit decision-making processes that cannot be dynamically optimized and fail to meet the demands of refined and personalized credit services. Summary of the Invention

[0005] The purpose of this invention is to provide a decision support method and system for personalized scenarios, in order to solve the problem that existing technologies are unable to provide accurate and personalized credit decisions.

[0006] To achieve the above objectives, the present invention provides a decision support method for personalized scenarios, comprising: S1. Select multiple evaluation indicators that influence the decision-making process for home renovation loan credit, and construct a fuzzy complementary judgment matrix; S2. Solve the weight vector of the fuzzy complementary judgment matrix to obtain the initial weight vector of each evaluation index; construct an adjustment matrix constrained by inter-row coupling degree and inter-column coupling degree. The inter-row coupling degree is the rank correlation coefficient extracted from the fuzzy complementary judgment matrix, and the inter-column coupling degree is the sample correlation degree obtained by solving and normalizing the feature vector of the evaluation index in the historical sample data through Euclidean distance; multiply the initial weight vector with the adjustment matrix to obtain the candidate weight vector. S3. Obtain the borrower feature vector, perform cluster training on historical customer data to generate multiple customer clusters, each customer cluster is configured with a cluster center and a scenario preference adjustment vector, calculate the vector similarity between the borrower feature vector and the cluster center of each customer cluster, generate an initial similarity vector and normalize it to obtain a weighted coefficient vector, perform a weighted summation of the weighted coefficient vector and the scenario preference adjustment vector of the corresponding customer cluster to obtain a comprehensive scenario adjustment vector, multiply the comprehensive scenario adjustment vector by the corresponding element of the candidate weight vector and normalize it to obtain a personalized weight vector; S4. Perform a weighted synthesis operation on the borrower feature vector and the personalized weight vector to obtain a comprehensive evaluation score. Output the decision result according to the correspondence between the comprehensive evaluation score and the preset mapping benchmark table. Perform a credit compliance check on the decision result. If the check passes, output the decision result.

[0007] This invention introduces sample relevance from historical sample data as the inter-coupling degree to optimize the initial weight vector and obtain candidate weight vectors. Since the candidate weight vectors integrate expert preferences and objective historical sample preferences, they can suppress the over-assignment of subjective expert preferences, making the weight allocation more scientific and reasonable. At the same time, this invention uses a comprehensive scenario adjustment vector to optimize the candidate weight vectors, thereby obtaining personalized weight vectors that fit the actual situation of the borrower, and thus enabling personalized credit decisions.

[0008] Furthermore, when constructing the fuzzy complementary judgment matrix, experts use the Delphi method to compare and score the importance of each evaluation index pairwise, obtaining the fuzzy importance between any two indicators, with a value ranging from 0.1 to 0.9.

[0009] Furthermore, during the construction of the fuzzy complementary judgment matrix, the fuzzy importance is checked and outliers are adjusted. These checks include complementarity checks, consistency checks, and corresponding adjustments. This eliminates logical contradictions in the expert subjective evaluation process, improving the consistency of the judgment matrix and the reliability of the data.

[0010] Furthermore, when constructing the adjustment matrix, the diagonal elements of the adjustment matrix M are adjusted. =1, off-diagonal element , , For adjustment coefficients, For inter-line coupling, For inter-coupling degree, the first adjustment coefficient The second adjustment coefficient is 0.6. It is 0.4.

[0011] Furthermore, the credit compliance verification is performed on the decision-making results based on credit constraint information, which includes the debt-to-income ratio and / or the upper limit of the total loan amount.

[0012] Furthermore, the step of obtaining the sample correlation by solving and normalizing the Euclidean distance of the feature vectors of the evaluation indicators in historical sample data includes: collecting historical sample data from the credit records of financial institutions, extracting the feature vectors corresponding to each evaluation indicator from the historical sample data, and calculating the Euclidean distance between the feature vector of the i-th indicator and the feature vector of the j-th indicator. Then calculate the Euclidean distance. The reciprocal of the sum of preset small positive numbers is used to map the result to the interval [0,1], thus obtaining the sample relevance. This invention obtains sample relevance based on objective historical sample data and uses the sample relevance to optimize the initial weight vector, achieving an organic integration of subjective and objective evaluation and improving the rationality of the initial weight vector.

[0013] Furthermore, the step of calculating the similarity between the borrower's feature vector and the cluster centers of each customer group cluster, generating an initial similarity vector, and obtaining a weighted coefficient vector through normalization includes: calculating the similarity between the borrower's feature vector and each cluster center one by one using a Gaussian kernel function, obtaining an initial similarity vector after traversing the calculation, obtaining a weighted coefficient vector through normalization, and constraining the sum of each coefficient to 1.

[0014] Furthermore, when the decision result fails the credit compliance check, proceed to step S5: The decision result that fails the check is taken as the limit parameter. Combined with the weighted synthesis formula of the comprehensive evaluation score and the weight of each evaluation indicator, the absolute value of the partial derivative of the limit parameter with respect to each evaluation indicator weight is calculated using the chain rule. The calculated constraint sensitivity is sorted in descending order, and the evaluation indicators corresponding to the top-ranked constraint sensitivity are selected as the weight reduction set. The remaining evaluation indicators are assigned to the compensation set. A fixed weight reduction value is deducted from the weight of each indicator in the weight reduction set, and the weight reduction value is added to each indicator in the compensation set to generate a constraint weight vector. Then, the constraint weight vector replaces the personalized weight vector in step S4, and a weighted synthesis operation is performed with the borrower feature vector obtained in step S4 to calculate the updated comprehensive evaluation score of the borrower. The subsequent steps are executed iteratively until the decision result passes the check or the preset maximum number of iterations threshold is reached, at which point the program is terminated and a credit rejection signal is output. After a credit compliance check fails, this invention can recalculate the constraint weight vector based on the specific circumstances of the constraint sensitivity related to the out-of-limit parameters and evaluation indicators, and replace the original personalized weight vector to achieve constraint iterative loop. This enables the decision-making mechanism of this invention to flexibly obtain compliant loan solutions and better meet the borrower's credit service needs. It overcomes the shortcomings of existing fuzzy decision-making processes, which can only output in one direction and lack an iterative adjustment mechanism, resulting in rigid and difficult-to-adjust decision results.

[0015] Furthermore, the step of adding the weight deduction value to each indicator in the compensation set includes: using the sample relevance obtained in step S2, determining the coupling degree between a certain indicator i in the weighted set and each indicator k in the compensation set. Calculate the weight transfer ratio , To prevent tiny positive numbers with a denominator of 0; and based on the calculated transfer ratio The sum of the weighted deduction values ​​will be transferred according to the transfer ratio. It is added to each indicator of the compensation set.

[0016] The present invention provides a decision support system for personalized scenarios, including a processor and a memory. The memory stores a computer program. When the computer program is executed by the processor, it performs the above-mentioned decision support method for personalized scenarios and achieves the same effect as the above method.

[0017] Summary of the beneficial effects of this invention: This invention obtains initial weights based on a fuzzy complementary judgment matrix, extracts inter-row coupling degree and inter-column coupling degree to deeply constrain and adjust the initial weights, and integrates the borrower's characteristics with the scenario similarity weighted conversion to obtain a personalized weight vector, realizing differentiated customer profile assessment. It can be customized in combination with borrower characteristics and specific scenarios, improving the objectivity and scientific nature of approval decision results and enhancing user experience. Attached Figure Description

[0018] Figure 1 A flowchart illustrating an embodiment of a decision support method for personalized scenarios according to the present invention; Figure 2 This is a schematic diagram showing the comparison between an embodiment of the decision support method for personalized scenarios of the present invention and a comparative example after conducting experiments. Detailed Implementation

[0019] An embodiment of the decision support method for personalized scenarios according to the present invention, such as... Figure 1 As shown, it includes: S1. Select multiple evaluation indicators that influence the credit decision of home renovation loans, obtain the fuzzy importance between any two indicators through expert scoring, and construct a fuzzy complementary judgment matrix based on the fuzzy importance.

[0020] In this embodiment, five evaluation indicators are used: number of defaults, longest overdue days, debt-to-income ratio, renovation budget, and estimated construction period. These indicators reflect various factors influencing credit decisions, such as the borrower's creditworthiness, repayment ability, and renovation budget. Experts use the Delphi method to score the pairwise importance of each indicator, obtaining the fuzzy importance between any two indicators, with values ​​ranging from 0.1 to 0.9. A 5×5 fuzzy complementary judgment matrix is ​​constructed based on the fuzzy importance. The Delphi method is a structured expert prediction and opinion solicitation method, mainly used in situations where historical data is lacking or high uncertainty exists. Through multiple rounds of anonymous questionnaires, feedback aggregation, and iterative correction, the opinions of the expert group gradually converge, ultimately reaching a relatively consistent judgment result.

[0021] To ensure the logical consistency of the fuzzy complementary judgment matrix, it is necessary to check the fuzzy importance and adjust outliers. The checks include complementarity checks, consistency checks, and corresponding adjustments, resulting in a valid fuzzy complementary judgment matrix. The following example illustrates the process of complementarity and consistency checks.

[0022] When performing complementarity testing, for any element in the matrix Calculate the complementary deviation ;like If the deviation exceeds the complementary deviation threshold, a correction is performed: Let Otherwise keep Unchanged. In an ideal state. It should approach 0; in this embodiment, the threshold for complementary deviation is set to 0.01.

[0023] Then a consistency check is performed on any element. Calculate the transmission difference ,in, The number of evaluation indicators, This represents the fuzzy importance of the i-th indicator and the k-th indicator. The fuzzy importance of the k-th and j-th indicators is represented by the transitivity; the transitivity difference is used to represent the degree of logical contradiction between the direct comparison result and the indirect derivation result. In this embodiment, the first deviation threshold for the transitivity difference is set to 0.1, and the second deviation threshold is set to 0.3. If the transitivity difference is less than the first deviation threshold, the logical consistency is considered good, and the consistency is maintained. Unchanged. If the transit difference falls between the first and second deviation thresholds, a slight logical conflict is considered to exist. A smoothing adjustment mechanism is then activated, using a linear smoothing function to fuse the original value of the element with the transit difference, and calculating the sum of all indirect comparison paths. Adjusted elements , This is the step size coefficient, which is set in this embodiment. The value is 0.5. If the transmission difference exceeds the second deviation threshold, it is considered an extreme logical conflict and will be... and its symmetric elements Change them all to 0.5 to eliminate the logical conflict.

[0024] Example of the fuzzy complementary judgment matrix in this embodiment ( See Table 1 below. The evaluation indicators are: 1. Number of defaults, 2. Longest overdue days, 3. Debt-to-income ratio, 4. Renovation budget price, and 5. Expected construction period.

[0025] Table 1

[0026] S2. Solve the weight vector of the fuzzy complementary judgment matrix to obtain the initial weight vector of each evaluation index; construct an adjustment matrix constrained by inter-row coupling degree and inter-column coupling degree. The inter-row coupling degree is the rank correlation coefficient extracted from the fuzzy complementary judgment matrix, and the inter-column coupling degree is the sample correlation obtained by calculating the pairwise Euclidean distance based on the feature vectors of the evaluation indexes in the historical sample data and normalizing it; multiply the initial weight vector with the adjustment matrix to obtain the candidate weight vector.

[0027] In this embodiment, the fuzzy complementary judgment matrix is ​​calculated using row sum normalization to obtain the initial weight vector W corresponding to each evaluation index. In other embodiments, other algorithms, such as eigenvector normalization, can also be used to calculate the initial weight vector.

[0028] Construct an adjustment matrix M, where the diagonal elements of M are... =1, off-diagonal element , , These are the first adjustment coefficient and the second adjustment coefficient. For inter-line coupling, To determine the inter-coupling degree, a first adjustment coefficient is set in this embodiment. The second adjustment coefficient is 0.6. The first and second adjustment coefficients are 0.4. They are used to set the weights of expert evaluation and historical data constraints. In this embodiment, expert evaluation is the main factor and historical data constraints are the secondary factor.

[0029] The following is the process for calculating the inter-row and inter-column coupling degree: Calculate the Spearman rank correlation coefficient between each row vector of the fuzzy complementary judgment matrix, and use it as the inter-row coupling degree. , which represents the strength of the relationship between the i-th row and the j-th row; the Spearman rank correlation coefficient is a non-parametric statistic used to measure the strength and direction of the monotonic relationship between two variables, with a value range of [-1,1], representing the degree of consistency of different indicators in the decision-maker's psychological preferences. Its calculation process is existing technology and will not be elaborated here.

[0030] Historical sample data is collected from the credit records of financial institutions. For example, 5,000 home renovation loan credit records from the past year are selected. The feature vectors corresponding to each evaluation indicator are extracted from the historical sample data, and the Euclidean distance between the feature vector of the i-th indicator and the feature vector of the j-th indicator is calculated. Then calculate the Euclidean distance. The reciprocal of the sum of the preset infinitesimal positive numbers is then used to perform max-min normalization on the calculated result to map it to the interval [0,1], yielding a dimensionless correlation. This correlation is used as the inter-column coupling degree for constructing the adjustment matrix. (Preset infinitesimal positive numbers) =0.001, correlation = This correlation is defined as sample correlation because it is derived from historical sample data.

[0031] After obtaining the adjustment matrix, the initial weight vector W is multiplied by the adjustment matrix M to obtain a candidate weight vector that integrates expert preferences and objective historical sample preferences.

[0032] S3. Obtain borrower feature vectors based on borrower information, perform cluster training on historical customer data to generate multiple customer clusters, each customer cluster is configured with a cluster center and a scenario preference adjustment vector, calculate the vector similarity between the borrower feature vector and the cluster center of each customer cluster, generate an initial similarity vector and normalize it to obtain a weighted coefficient vector, perform a weighted summation of the weighted coefficient vector and the scenario preference adjustment vector of the corresponding customer cluster to obtain a comprehensive scenario adjustment vector, multiply the comprehensive scenario adjustment vector by the corresponding element of the candidate weight vector obtained in step S2 above and normalize it to obtain a personalized weight vector.

[0033] By connecting the credit reporting system and business service interface, various evaluation indicators of borrowers are collected: number of defaults, longest overdue days, debt-to-income ratio, renovation budget, and estimated construction period. The maximum-minimum normalization algorithm is used to complete the dimensionless processing of each indicator, eliminating the differences in dimensions and orders of magnitude between different indicators. The five preprocessed indicators are then concatenated horizontally to obtain the borrower's feature vector.

[0034] Clustering training is performed on historical customer data to generate multiple standardized customer clusters, each with a corresponding cluster center and scenario preference adjustment vector. This embodiment employs the KMeans clustering algorithm, pre-training multiple customer clusters corresponding to typical borrowers from historical data. The scenario preference adjustment vector for each cluster is obtained offline: combining the risk control rules of home renovation loans with the risk characteristics of each customer cluster, adjustment coefficients are configured for each of the five evaluation indicators, generating a five-dimensional scenario preference adjustment vector for the corresponding customer cluster, used for subsequent scenario-based correction calculations of personalized weight vectors.

[0035] In one embodiment, combining the risk control rules of home renovation credit business with the risk characteristics of each customer group cluster, adjustment coefficients are configured for each of the five evaluation indicators to generate a five-dimensional scenario preference adjustment vector for the corresponding customer group cluster. Specifically, historical credit samples of a specific customer group cluster (e.g., "high income but occasional overdue payments") are extracted, and the importance of the five evaluation indicators "number of defaults, longest overdue days, debt-to-income ratio, renovation budget price, and expected construction period" to whether or not a default occurs is evaluated using a random forest algorithm or by calculating the information value (IV value). More specifically, the relationship between the above five evaluation indicators and default is established through historical data. For example, if the debt-to-income ratio is more correlated with whether or not a default occurs, a higher importance is set.

[0036] After obtaining the numerical value of the importance of the feature, the penalty mechanism in the risk control rules is further utilized. For example, if the risk control rules are extremely strict on the "number of defaults", then a minimum multiplier weight is assigned to the initial importance of the indicator.

[0037] The importance values ​​of the five features after rule correction are normalized so that their sum equals 1 or a preset constant. The resulting five values ​​are the adjustment coefficients for the corresponding five indicators, forming a five-dimensional scene preference adjustment vector specific to this customer group.

[0038] The similarity between the borrower's feature vector and the cluster centers of each customer group is calculated to generate an initial similarity vector, which is then normalized to obtain a weighted coefficient vector. In this embodiment, a Gaussian kernel function is used to calculate the similarity between the borrower's feature vector and each cluster center one by one, and the initial similarity vector is obtained after traversing the calculation. The weighted coefficient vector is obtained by normalization using the Softmax function, with the sum of all coefficients constrained to 1. This weighted coefficient vector reflects the fit between the borrower and each customer group cluster. The weighted coefficient vector is then summed with the corresponding cluster's scenario preference adjustment vector to obtain the comprehensive scenario adjustment vector. ,in, Weighted coefficient vector The k-th element in the matrix is ​​the normalized similarity value between the borrower's feature vector and the k-th customer cluster center. For the pre-configured scene preference adjustment vector corresponding to the k-th customer group cluster; the comprehensive scene adjustment vector The candidate weight vector calculated in step S2 above is multiplied element-wise, and the multiplication result is normalized to obtain a personalized weight vector.

[0039] This operation retains the structural characteristics of the individual borrower assessment matrix while introducing scenario-specific adjustments for the indicator dimensions, outputting a personalized weight vector that fits the specific risk and qualification scenario of the borrower.

[0040] S4. Perform a weighted synthesis operation on the borrower's feature vector and the personalized weight vector obtained in step S3 to obtain the borrower's comprehensive evaluation score; output the decision result according to the correspondence between the comprehensive evaluation score and the preset mapping benchmark table; obtain the borrower's credit constraint information, and perform a credit compliance check on the decision result according to the credit constraint information. If the check passes, output the decision result as the final decision result; otherwise, proceed to step S5.

[0041] For the comprehensive evaluation score, multiple mapping benchmark tables are configured. In one embodiment, the method for determining the segmentation points of the mapping benchmark table is as follows: the comprehensive evaluation scores of historical non-default samples are normally fitted and key statistical quantiles are extracted, such as the scores corresponding to 25%, 50%, and 75%. Combined with the existing tiered levels of credit business, the overall score interval [60, 100] is divided into multiple corresponding discrete sub-intervals.

[0042] By using the retained historical sample set for backtesting, the actual scores of historical customers are substituted into the proposed benchmark table to simulate credit granting, and the simulated expected return in each segment interval is calculated. The expected return is the interest spread income minus the expected loss calculated based on the actual bad debt rate of that interval. If the expected bad debt loss of a certain interval exceeds the risk provision limit, the benchmark limit corresponding to the endpoint of that interval is slightly adjusted downward or the benchmark interest rate is increased until the return and risk of each segment interval reach the optimal balance.

[0043] When applying the calibrated mapping benchmark table, the comprehensive assessment score is mapped to the loan amount and loan term through forward piecewise linear interpolation. The specific form of the interpolation formula is: Let the score at the endpoint of the current sub-interval be... and The corresponding benchmark amount or term is and For the borrower's actual score X, the mapping result is... For example, the score range [60, 100] is mapped to a loan amount of [100,000, 500,000] and a loan term of [12 installments, 60 installments]. A score of 82.5, calculated using the above interpolation, is mapped to a loan amount of 325,000 and a term of 39 months. Similarly, the comprehensive assessment score is mapped to the loan interest rate through reverse piecewise linear interpolation. For example, a score of [60, 100] is mapped to an annualized loan interest rate of [6.5%, 4.2%]. The higher the score, the lower the interest rate; for example, a score of 82.5 corresponds to an annualized loan interest rate of 5.21%. After obtaining the initial loan amount, loan term, and loan interest rate, the expected monthly payment is calculated using the equal principal and interest repayment formula or equal principal repayment formula. For example, based on 325,000, 39 installments, and 5.21%, the expected monthly payment is approximately 9,086 yuan. The decision result, including the expected monthly payment, is then obtained.

[0044] The process involves obtaining the borrower's credit constraint information. In this embodiment, the credit constraint information is the debt-to-income ratio set by the financial institution; in other embodiments, it could be the upper limit of the total loan amount. After obtaining the decision result, the debt-to-income ratio is calculated based on the borrower's average monthly after-tax income and expected monthly payment. Specifically, the ratio of expected monthly payment to average monthly income is calculated. If the debt-to-income ratio does not exceed the limit, the credit compliance check is deemed passed, and the decision result is output. If the check fails, a failure flag and corresponding over-limit parameters are generated, and the process proceeds to step S5. For example, if the calculated debt-to-income ratio is 45.43%, and the upper limit for monthly payment affordability set by the financial institution is 45%, then the credit compliance check fails, a failure flag is generated, and the over-limit parameters are recorded. In this embodiment, the over-limit parameters are the decision result, corresponding to the comprehensive evaluation score.

[0045] S5. Calculate the constraint sensitivity based on the limit parameter, and obtain the constraint weight vector by combining the sample correlation in step S2. Replace the personalized weight vector in step S4 with the constraint weight vector, and perform a weighted synthesis operation with the borrower feature vector obtained in step S4 to calculate the updated comprehensive evaluation score of the borrower. Then, perform subsequent steps to iterate until the decision result passes the test or reaches the preset maximum number of iterations threshold, and then terminate the program and output a credit rejection signal.

[0046] Decisions that fail the test are considered as limits-crossing parameters. Based on these limits-crossing parameters, the comprehensive evaluation score, and the weights of each evaluation indicator, the final decision is made. The weighted composite formula is used to calculate the weight of the limit parameter relative to each evaluation index using the chain rule. The absolute value of the partial derivative. For example, if the expected monthly payment exceeds the limit, the excess parameter is the expected monthly payment y, then calculate... , where variables This represents the borrower's overall assessment score. The partial derivative value is expressed as the sensitivity of each evaluation indicator to the constraint parameters exceeding the limit.

[0047] Based on the calculated constraint sensitivities, the indicators are sorted in descending order. The top two constraint sensitivities are selected as the weighted set, and the remaining indicators are assigned to the compensation set. Using the sample correlation results obtained in step S2, the coupling degree between an indicator i in the weighted set and each indicator k in the compensation set is determined. Calculate the weight transfer ratio , To prevent tiny positive numbers with a denominator of 0, a value of 0.001 is preferred. A predetermined weight deduction value is fixedly subtracted from the weights of each indicator in the weighted set. In this embodiment, the weight deduction value is set to 0.03, and the calculated transfer ratio is used as the basis for the reduction. The total weighted share deducted will be transferred according to the proportion. The weights are added to each indicator in the compensation set. Through the above transfer strategy, a constraint weight vector is generated while strictly ensuring that the total weight of all evaluation indicators remains constant at 1.

[0048] Then, the constraint weight vector is used to replace the personalized weight vector in step S4, and a weighted synthesis operation is performed with the borrower feature vector obtained in step S4 to calculate the updated comprehensive assessment score of the borrower. The subsequent steps are executed to iterate until the decision result passes the test or reaches the preset maximum number of iterations threshold, at which point the program is terminated and a credit rejection signal is output.

[0049] The above embodiment conducts an ablation experiment based on 1000 historical real-world renovation credit sample data, dividing the sample data into training and test sets in a 7:3 ratio. The experiment uses three groups for comparison: the basic model group, the non-feedback iterative group, and the present invention (this embodiment). The basic model group uses traditional fuzzy hierarchical analysis for weight calculation without any consistency adjustment or over-limit constraint feedback mechanism. The non-feedback iterative group adds fuzzy complementary judgment matrix segmented mapping adjustment and constraint weight allocation to the basic model but removes the iterative loop mechanism. Evaluation metrics include model logic consistency achievement rate, credit over-limit occurrence rate, and overall decision accuracy. Five rounds of repeated cross-validation tests were conducted to obtain specific experimental data and results, as follows: Figure 2 As shown, the above embodiments have advantages in logical consistency, credit limit exceeding rate, and decision correctness. Specifically, the basic model group achieved a logical consistency rate of 65.2%, a credit limit exceeding rate as high as 23.4%, and an overall decision accuracy of only 71.5%. The feedback-free iteration group achieved a logical consistency rate of 94.1%, a credit limit exceeding rate of 8.7%, and an overall decision accuracy of 85.3%. The experimental group achieved a logical consistency rate of 98.6%, a credit limit exceeding rate further reduced to 1.2%, and an overall decision accuracy that improved to a maximum of 93.8%.

[0050] In other embodiments of the present invention, step S5 can be omitted, i.e., feedback-free iteration, which can also improve decision accuracy compared to the prior art.

[0051] An embodiment of the decision support system for personalized scenarios according to the present invention includes a processor and a memory. The memory stores a computer program, and when the computer program is executed by the processor, it performs a decision support method for personalized scenarios as described above.

Claims

1. A decision support method for personalized scenarios, characterized in that, include: S1. Select multiple evaluation indicators that influence the decision-making process for home renovation loan credit, and construct a fuzzy complementary judgment matrix; S2. Solve the weight vector of the fuzzy complementary judgment matrix to obtain the initial weight vector of each evaluation index; construct an adjustment matrix constrained by inter-row coupling degree and inter-column coupling degree. The inter-row coupling degree is the rank correlation coefficient extracted from the fuzzy complementary judgment matrix, and the inter-column coupling degree is the sample correlation degree obtained by solving and normalizing the feature vector of the evaluation index in the historical sample data through Euclidean distance; multiply the initial weight vector with the adjustment matrix to obtain the candidate weight vector. S3. Obtain the borrower feature vector, perform cluster training on historical customer data to generate multiple customer clusters, each customer cluster is configured with a cluster center and a scenario preference adjustment vector, calculate the vector similarity between the borrower feature vector and the cluster center of each customer cluster, generate an initial similarity vector and normalize it to obtain a weighted coefficient vector, perform a weighted summation of the weighted coefficient vector and the scenario preference adjustment vector of the corresponding customer cluster to obtain a comprehensive scenario adjustment vector, multiply the comprehensive scenario adjustment vector by the corresponding element of the candidate weight vector and normalize it to obtain a personalized weight vector; S4. Perform a weighted synthesis operation on the borrower feature vector and the personalized weight vector to obtain a comprehensive evaluation score. Output the decision result according to the correspondence between the comprehensive evaluation score and the preset mapping benchmark table. Perform a credit compliance check on the decision result. If the check passes, output the decision result.

2. The decision support method for personalized scenarios according to claim 1, characterized in that, When constructing the fuzzy complementary judgment matrix, experts use the Delphi method to compare and score the importance of each evaluation index pairwise, and obtain the fuzzy importance between any two indicators, with a value range of 0.1 to 0.

9.

3. The decision support method for personalized scenarios according to claim 2, characterized in that, When constructing the fuzzy complementarity judgment matrix, the fuzzy importance must be checked and abnormal data adjusted. The checks include complementarity checks, consistency checks, and corresponding adjustments.

4. The decision support method for personalized scenarios according to claim 1, characterized in that, When constructing the adjustment matrix, the diagonal elements of the adjustment matrix M are adjusted. =1, off-diagonal element , The first adjustment coefficient, This is the second adjustment coefficient. For inter-line coupling, For inter-coupling degree, the first adjustment coefficient The second adjustment coefficient is 0.

6. It is 0.

4.

5. The decision support method for personalized scenarios according to claim 1, characterized in that, The credit compliance verification is a credit compliance verification of the decision-making results based on credit constraint information, which includes the debt-to-income ratio and / or the upper limit of the total loan amount.

6. The decision support method for personalized scenarios according to claim 1, characterized in that, Methods for obtaining sample relevance include: collecting historical sample data from the credit records of financial institutions, extracting the feature vectors corresponding to each evaluation indicator from the historical sample data, and calculating the Euclidean distance between the feature vector of the i-th indicator and the feature vector of the j-th indicator. Then calculate the Euclidean distance. The reciprocal of the sum of the preset small positive numbers is used to perform maximum and minimum standardization on the calculated result to map it to the interval [0,1], thus obtaining the sample correlation.

7. The decision support method for personalized scenarios according to claim 1, characterized in that, The calculation of the similarity between the borrower's feature vector and the cluster centers of each customer group, generating an initial similarity vector and obtaining a weighted coefficient vector through normalization, includes: calculating the similarity between the borrower's feature vector and each cluster center one by one using a Gaussian kernel function, obtaining an initial similarity vector after traversing the calculation, obtaining a weighted coefficient vector through normalization, and constraining the sum of each coefficient to 1.

8. The decision support method for personalized scenarios according to any one of claims 1-7, characterized in that, If the decision fails the credit compliance check, proceed to step S5: The decision results that fail the test are used as the limit-breaking parameter. A weighted composite formula combining the comprehensive evaluation score and the weights of each evaluation indicator is used. The absolute value of the partial derivative of this limit-breaking parameter with respect to each evaluation indicator weight is calculated using the chain rule. The calculated constraint sensitivity is sorted in descending order, and the evaluation indicators corresponding to the top-ranked constraint sensitivity are selected as the weighted set. The remaining evaluation indicators are assigned to the compensation set. A fixed weight deduction value is deducted from the weights of each indicator in the weighted set, and this deduction value is added to each indicator in the compensation set to generate a constraint weight vector. Then, the constraint weight vector replaces the personalized weight vector in step S4, and a weighted composite operation is performed with the borrower feature vector obtained in step S4 to calculate the updated comprehensive evaluation score of the borrower. Subsequent steps are executed iteratively until the decision result passes the test or reaches the preset maximum iteration threshold, at which point the program terminates and outputs a credit rejection signal.

9. The decision support method for personalized scenarios according to claim 8, characterized in that, The step of adding the weight deduction value to each indicator of the compensation set includes: using the sample relevance obtained in step S2, determining the coupling degree between a certain indicator i in the weighted set and each indicator k in the compensation set. Calculate the weight transfer ratio , To prevent tiny positive numbers with a denominator of 0; and based on the calculated transfer ratio The sum of the weighted deduction values ​​will be transferred according to the transfer ratio. It is added to each indicator of the compensation set.

10. A decision support system for personalized scenarios, characterized in that, It includes a processor and a memory, the memory storing a computer program, which, when executed by the processor, performs the steps in the decision support method for personalized scenarios according to any one of claims 1-9.

Images