Collection effect prediction method based on artificial intelligence

Abstract

The invention discloses a collection effect prediction method based on artificial intelligence, belongs to the technical field of debt collection, and aims to solve the problem of fund loss caused by incapability of accurately predicting the repayment probability of a customer in a current collection method. The method integrates multi-source data, covers customer basic information, financial transaction records, credit reports and past interaction records, lays a foundation for comprehensively describing customer features by carefully constructing multi-dimensional features of debt paying ability, credit portraits and interaction behaviors, realizes customer feature grouping by applying a K-Means + + clustering algorithm, and improves customer feature grouping efficiency. Through loop iteration grouping matching and probability calculation, the relation between individual features of customers and group generality is fully considered, historical data statistics and adjustment factors are combined, and finally the customer repayment probability is accurately calculated, so that a financial institution can accurately identify customers with high repayment possibility and high-risk customers in advance, and the customer repayment efficiency is improved. And meanwhile, targeted measures are taken in time for high-risk customers, so that the collection success rate is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of debt collection, and in particular to a collection effect prediction method based on artificial intelligence. Background Art

[0002] In today's financial field, debt collection business faces many challenges and difficulties. With the continuous expansion of the credit market, loan business has become more frequent and the non-performing loan rate has also increased, which has brought huge economic loss risks to financial institutions. Debt collection work is a key link in controlling risks and ensuring the return of funds. The effectiveness of the work directly affects the financial status and operational stability of financial institutions.

[0003] The traditional collection model mainly relies on manual experience and simple rule judgment to determine the collection strategy. The collection personnel usually work based on limited basic customer information and rough credit records. This method has obvious limitations: on the one hand, manual analysis of massive customer data is inefficient, and it is difficult to quickly identify high-risk or high-repayment-likely customers who really need to be focused on among many customers, resulting in unreasonable allocation of collection resources, often wasting a lot of manpower and time on low-risk customers, while the collection of high-risk customers is not timely and accurate enough; on the other hand, manual judgment is highly subjective, and different collection personnel may have large differences in their interpretation and decision-making of the same customer information. There is a lack of unified and objective standards, and it is impossible to accurately predict the probability of successful repayment of customers, making it difficult to formulate personalized and effective collection plans, which in turn affects the collection effect and overall recovery rate; therefore, it is necessary to accurately predict the probability of successful repayment of customers before collection, so as to carry out collection in different modes. Summary of the invention

[0004] The purpose of the present invention is to provide a collection effect prediction method based on artificial intelligence, which solves the problem that the current existing collection methods in the background technology cannot accurately predict the repayment probability of customers, thereby causing capital loss.

[0005] To achieve the above object, the present invention provides the following technical solution: a collection effect prediction method based on artificial intelligence, comprising the following steps:

[0006] S1. Data collection: Collect customer-related data from multiple data sources, including basic customer information, financial transaction records, credit reports, and past interaction records;

[0007] S2. Data preprocessing: For missing values ​​of numerical type, fill them based on the statistical characteristics of similar customers or supplement them with business logic inference; for outliers, modify or eliminate them according to business rules; construct the debt repayment ability dimension feature D, including calculating the debt-to-income DTI ratio and asset liquidity index ALI; credit profile dimension feature C, through credit score standardization CSstd and OLS calculation of overdue severity; interactive behavior dimension feature B, calculation of SMS response initiative SRI and call cooperation degree TCI;

[0008] S3. Customer feature grouping: K-Means++ clustering algorithm is used to cluster customers in the three-dimensional feature space of debt repayment ability, credit profile and interactive behavior. The initial number of cluster centers is set to K, where K is a preset integer, and G1, G2, …, G K Equal groups are formed, core feature statistics of each group are calculated, and group feature vectors are constructed for subsequent matching;

[0009] S4. Loop iterative group matching and probability calculation, specifically:

[0010] S41. For the customer to be predicted, set the matching threshold θ (0<θ<1) according to its feature vector and assign the initial grouping probability

[0011] S42. Calculate the cosine similarity S (V c ,V i );

[0012] S43. Update the grouping probability according to the comparison result between the similarity and the threshold. If S(V c ,V i )≥0, then P(G i c)=P(G i c)+α(1-P(G i c)), α is the learning rate, 0<α<1, otherwise P(G i c)=P(G i c) (1-β), β is the decay rate, 0<β<1;

[0013] S44. Normalize the updated probability to ensure

[0014] S45. Repeat the above steps until the preset maximum number of iterations T is reached or the probability change of each group after t consecutive iterations is less than the preset threshold ε, 0<ε<1, T and t are both positive integers, and the final grouping probability of the customer is obtained;

[0015] S5. Repayment probability calculation and model evaluation: Based on historical data statistics, the repayment success ratio of each group is used as the basic repayment probability P repay (G i ) ; Introduce adjustment factors, calculate the customer's individual repayment probability P according to the customer's credit score, debt-to-income ratio and the deviation from the group mean, combined with the preset weight coefficient repay(c) Divide the set into training set, validation set, and test set according to the preset ratio, build the model on the training set, tune the parameters on the validation set, and use the test set to calculate the accuracy, recall, and F1 value evaluation indicators.

[0016] Furthermore, in the data collection step, basic customer information includes name, age, occupation and salary level, among which the salary level data comes from the salary report of the human resources department or the tax return record; financial transaction records cover historical loan amounts, terms, repayment times, and overdue situations; credit reports are obtained from authoritative credit reporting agencies, including credit scores, number of credit inquiries, and historical default details; past interaction records are integrated from the SMS platform, customer service system and email server, including SMS sending and receiving time, content, and related information of customer service call records.

[0017] Furthermore, in the cleaning step of data preprocessing, for missing numerical values, if the salary level is missing, the value is filled by finding the mean of the same type by double grouping by occupation and age.

[0018] Furthermore, in data preprocessing, the debt-to-income ratio calculation formula of the debt-paying capacity dimension is: L is the total liabilities, I is the monthly stable income, and the asset liquidity index calculation formula is: A liquid is an asset that can be quickly converted into cash, and E is monthly expenditure. The credit score standardization formula of the credit profile dimension is: CS is the original credit score, CS min The lowest credit score in history, CS max is the highest credit score in history, and the calculation formula for the severity of overdue payment is N late is the number of overdue payments, is the maximum number of overdue payments in history, T late is the overdue time, is the maximum overdue time in history; the SMS response enthusiasm calculation formula in the interactive behavior dimension is SRI=0.7×RR+0.3×SES, RR is the SMS reply rate within 24 hours, SES is the reply sentiment tendency score, and the call cooperation calculation formula is TCI=TL×CPD, TL is the customer service call duration, and CPD is the clarity of the promised repayment.

[0019] Furthermore, in the customer feature grouping step, the input of the clustering algorithm is the customer's feature vector, and the output is the grouping result and the group feature vector. During the clustering process, the cluster center is continuously updated to minimize the intra-group sum of square errors and maximize the inter-group distinction.

[0020] Furthermore, in the repayment probability calculation and model evaluation steps, the adjustment factor calculation formula is P repay (c) = P repay (G i)·(1+λ·ΔCS+δ·ΔDTI), where λ and δ are preset weight coefficients, ΔCS is the difference between the customer's credit score and the mean credit score of the group, and ΔDTI is the difference between the customer's debt-to-income ratio and the mean debt-to-income ratio of the group; the training set, validation set, and test set ratios in the repayment probability calculation and model evaluation are 70%, 15%, and 15%, respectively. The accuracy calculation formula in the evaluation index is: TP is the number of samples with the correct prediction of successful repayment, TN is the number of samples with the correct prediction of failed repayment, N is the total number of samples, and the calculation formula of recall rate is: FN is the number of samples with successful repayment but failed prediction. The calculation formula of F1 value is: in FP is the number of samples where repayment was predicted to be successful but actually failed.

[0021] Furthermore, in the customer feature grouping step, the value range of the initial clustering center number K of the clustering algorithm is 3-10, the value range of the matching threshold θ is 0.7-0.95, the value range of the learning rate α is 0.05-0.2, the value range of the decay rate β is 0.01-0.1, the value range of the number of iterations T is 5-20, the value range of the continuous stable iteration number t is 2-5, the value range of the probability change threshold ε is 0.005-0.02, the value range of the preset weight coefficient γ is 0.1-0.3, the value range of δ is 0.2-0.4, and the parameter values ​​are determined through experimental optimization based on the historical data scale, data quality and collection business scenario characteristics.

[0022] Furthermore, in the data preprocessing stage, the weight ratio of each dimensional feature vector constructed by feature engineering is dynamically updated on a quarterly or annual basis according to business scenario requirements. The weight of the debt-paying ability dimension and the credit portrait dimension can be appropriately increased when the market is volatile and the credit environment is tightening. The weight of the interactive behavior dimension can be increased during the period of business expansion when customer communication and feedback are important.

[0023] Furthermore, in the customer feature grouping stage, after clustering is completed, the newly accumulated data is regularly re-clustered and grouped, and the group feature vector is updated to make the grouping structure fit the latest feature distribution of the customer group, ensuring that subsequent iterative matching and prediction are in line with the actual business evolution trend.

[0024] Compared with the prior art, the present invention has the following beneficial effects:

[0025] The present invention provides an artificial intelligence-based collection effect prediction method, which integrates multi-source data, covering basic customer information, financial transaction records, credit reports and past interaction records, and carefully constructs multi-dimensional features of debt repayment ability, credit portraits and interactive behaviors to lay the foundation for comprehensively characterizing customer characteristics. The K-Means++ clustering algorithm is used to realize customer feature grouping, and through cyclic iterative group matching and probability calculation, the relationship between individual customer characteristics and group commonalities is fully considered, and historical data statistics and adjustment factors are combined to finally accurately calculate the customer repayment probability, so that financial institutions can accurately identify customers with high repayment possibilities and high-risk customers in advance, avoid wasting resources on low-risk customers, and take targeted measures for high-risk customers in a timely manner to improve the collection success rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a module diagram of the present invention. DETAILED DESCRIPTION

[0027] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0028] In order to solve the technical problem that the existing collection methods cannot accurately predict the repayment probability of customers, thus leading to capital loss, such as Figure 1 As shown, the following preferred technical solutions are provided:

[0029] A collection effect prediction method based on artificial intelligence, characterized by comprising the following steps:

[0030] S1. Data collection: Collect customer-related data from multiple data sources, including basic customer information, financial transaction records, credit reports, and past interaction records;

[0031] S2. Data preprocessing: For missing values ​​of numerical type, fill them based on the statistical characteristics of similar customers or supplement them with inferences based on business logic; for outliers, modify or eliminate them according to business rules; construct debt repayment capacity dimension features, including calculating debt-to-income ratio and asset liquidity indicators; credit profile dimension features, through credit score standardization and overdue severity calculation; interactive behavior dimension features, calculate SMS response enthusiasm and call cooperation;

[0032] S3. Customer feature grouping: K-Means++ clustering algorithm is used to cluster customers in the three-dimensional feature space of debt repayment ability, credit profile and interactive behavior. The initial number of cluster centers is set to K, where K is a preset integer, and G1, G2, …, G K Equal groups are formed, core feature statistics of each group are calculated, and group feature vectors are constructed for subsequent matching;

[0033] S4. Loop iterative group matching and probability calculation, specifically:

[0034] S41. For the customer to be predicted, set the matching threshold θ (0<θ<1) according to its feature vector and assign the initial grouping probability

[0035] S42. Calculate the cosine similarity S (V c ,V i );

[0036] S43. Update the grouping probability according to the comparison result between the similarity and the threshold. If S(V c ,V i )≥0, then P(G i c)=P(G i c)+α(1-P(G i c)), α is the learning rate, 0<α<1, otherwise P(G i c)=P(G i c) (1-β), β is the decay rate, 0<β<1;

[0037] S44. Normalize the updated probability to ensure

[0038] S45. Repeat the above steps until the preset maximum number of iterations T is reached or the probability change of each group after t consecutive iterations is less than the preset threshold ε, 0<ε<1, T and t are both positive integers, and the final grouping probability of the customer is obtained;

[0039] S5. Repayment probability calculation and model evaluation: Based on historical data statistics, the repayment success ratio of each group is used as the basic repayment probability P repay (G i ) ; Introduce adjustment factors, calculate the customer's individual repayment probability P according to the customer's credit score, debt-to-income ratio and the deviation from the group mean, combined with the preset weight coefficient repay (c) Divide the set into training set, validation set, and test set according to the preset ratio, build the model on the training set, tune the parameters on the validation set, and use the test set to calculate the accuracy, recall, and F1 value evaluation indicators.

[0040] In the data collection step, basic customer information includes name, age, occupation and salary level, among which the salary level data comes from the human resources department's salary report or tax return records; financial transaction records cover historical loan amounts, terms, repayment times, and overdue situations; credit reports are obtained from authoritative credit reporting agencies, including credit scores, number of credit inquiries, and historical default details; past interaction records are integrated from the SMS platform, customer service system, and email server, including SMS sending and receiving time, content, and related information on customer service call records.

[0041] In the cleaning step of data preprocessing, for missing numerical values, if the salary level is missing, fill it with the same mean by double grouping by occupation and age; when handling outliers, taking the loan amount as an example, if it exceeds the historical mean μ plus or minus 3 times the standard deviation σ, it will be marked for review and correction.

[0042] In data preprocessing, the debt-to-income ratio calculation formula for the debt-paying capacity dimension is: L is the total liabilities, I is the monthly stable income, and the asset liquidity index calculation formula is: A liquid is an asset that can be quickly converted into cash, and E is monthly expenditure. The credit score standardization formula of the credit profile dimension is: CS is the original credit score, CS min The lowest credit score in history, CS max is the highest credit score in history, and the calculation formula for the severity of overdue payment is N late is the number of overdue payments, is the maximum number of overdue payments in history, T late is the overdue time, is the maximum overdue time in history; the SMS response enthusiasm calculation formula in the interactive behavior dimension is SRI=0.7×RR+0.3×SES, RR is the SMS reply rate within 24 hours, SES is the reply sentiment tendency score, and the call cooperation calculation formula is TCI=TL×CPD, TL is the customer service call duration, and CPD is the clarity of the promised repayment.

[0043] In the customer feature grouping step, the input of the clustering algorithm is the customer's feature vector, and the output is the grouping result and the group feature vector. During the clustering process, the cluster center is continuously updated to minimize the intra-group sum of square errors and maximize the inter-group distinction.

[0044] In the repayment probability calculation and model evaluation steps, the adjustment factor calculation formula is P repay (c) = P repay (G i)·(1+λ·ΔCS+δ·ΔDTI), where λ and δ are preset weight coefficients, ΔCS is the difference between the customer’s credit score and the mean credit score of its group, and ΔDTI is the difference between the customer’s debt-to-income ratio and the mean debt-to-income ratio of its group.

[0045] In the repayment probability calculation and model evaluation, the training set, validation set, and test set are divided into 70%, 15%, and 15% respectively. The accuracy calculation formula in the evaluation index is: TP is the number of samples with the correct prediction of successful repayment, TN is the number of samples with the correct prediction of failed repayment, N is the total number of samples, and the calculation formula of recall rate is: FN is the number of samples with successful repayment but failed prediction. The calculation formula of F1 value is: in FP is the number of samples where repayment was predicted to be successful but actually failed.

[0046] In the customer feature grouping step, the initial cluster center number K of the clustering algorithm ranges from 3 to 10, the matching threshold θ ranges from 0.7 to 0.95, the learning rate α ranges from 0.05 to 0.2, the decay rate β ranges from 0.01 to 0.1, the number of iterations T ranges from 5 to 20, the number of continuous stable iterations t ranges from 2 to 5, the probability change threshold ε ranges from 0.005 to 0.02, the preset weight coefficient γ ranges from 0.1 to 0.3, and the value range of δ ranges from 0.2 to 0.4. The parameter values ​​are determined through experimental optimization based on the historical data scale, data quality and collection business scenario characteristics.

[0047] During the data preprocessing stage, the weight ratio of each dimensional feature vector constructed by feature engineering is dynamically updated on a quarterly or annual basis according to business scenario requirements. The weight of the debt-paying ability dimension and the credit portrait dimension can be appropriately increased when the market is volatile and the credit environment is tightening. The weight of the interactive behavior dimension is increased during the period of business expansion when customer communication and feedback are important.

[0048] In the customer feature grouping stage, after clustering is completed, the newly accumulated data is regularly re-clustered and grouped, and the group feature vector is updated to make the grouping structure fit the latest feature distribution of the customer group, ensuring that subsequent iterative matching and prediction are in line with the actual business evolution trend.

[0049] Specifically, for the customer c to be predicted, according to its preprocessed feature vector DTI c =0.6, CS std,c =0.65, SRI c =0.3 to start the process, set the matching threshold θ = 0.8, and assign the customer an initial grouping probability The cosine similarity formula is used to accurately measure the matching degree between the customer and the group feature vector. Taking customer c and group G1 as an example, the mean debt repayment ability within group G1 is Credit Profile Mean Interaction behavior mean Molecular part Denominator According to this formula Calculate the similarity value; if after the iteration, the customer still belongs to group G1, taking group G1 as an example, if the historical sample size reaches 200 cases, of which 150 cases are successfully repaid, then the basic repayment probability Considering the subtle impact of the difference in individual characteristics of customers deviating from the group mean on the repayment probability, an adjustment factor mechanism is introduced, and its credit score deviates from the mean Debt-to-income ratio deviation Combining the weight coefficients γ=0.2 and δ=0.3 determined through data analysis and field experience, the formula P repay (c)=0.75×(1+0.2×(-0.05)+0.3×0.05)=0.7875.

[0050] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0051] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A collection effect prediction method based on artificial intelligence, characterized in that: The following steps are involved: S1. Data collection: Collect customer-related data from multiple data sources, including basic customer information, financial transaction records, credit reports, and past interaction records; S2. Data preprocessing: For missing values ​​of numerical type, fill them based on the statistical characteristics of similar customers or supplement them with business logic inference; for outliers, modify or eliminate them according to business rules; construct the debt repayment ability dimension feature D, including calculating the debt-to-income DTI ratio and asset liquidity index ALI; credit profile dimension feature C, through credit score standardization CS std and OLS calculation of overdue severity; interactive behavior dimension feature B, calculation of SMS response initiative SRI and call cooperation degree TCI; S3. Customer feature grouping: K-Means++ clustering algorithm is used to cluster customers in the three-dimensional feature space of debt repayment ability, credit profile and interactive behavior. The initial number of cluster centers is set to K, where K is a preset integer, and G1, G2, …, G K Grouping, calculating the core feature statistics of each group, and constructing the group feature vector for subsequent matching; S4. Loop iterative group matching and probability calculation, specifically: S41. For the customer to be predicted, set the matching threshold θ (0<θ<1) according to its feature vector and assign the initial grouping probability S42. Calculate the cosine similarity S (V c ,V i ); S43. Update the grouping probability according to the comparison result between the similarity and the threshold. If S(V c ,V i )≥0, then P(G i |c)=P(G i |c)+α(1-P(G i |c)), α is the learning rate, 0<α<1, otherwise P(G i |c)=P(G i |c)·(1-β), β is the decay rate, 0<β<1; S44. Normalize the updated probability to ensure S45. Repeat the above steps until the preset maximum number of iterations T is reached or the probability change of each group after t consecutive iterations is less than the preset threshold ε, 0<ε<1, T and t are both positive integers, and the final grouping probability of the customer is obtained; S5. Repayment probability calculation and model evaluation: Based on historical data statistics, the repayment success ratio of each group is used as the basic repayment probability P repay (G i ) ; Introduce adjustment factors, calculate the customer's individual repayment probability P according to the customer's credit score, debt-to-income ratio and the deviation from the group mean, combined with the preset weight coefficient repay (c); Divide the dataset into training set, validation set, and test set according to preset proportions, build a model on the training set, tune parameters on the validation set, and use the test set to calculate accuracy, recall, and F1 value evaluation indicators.

2. The method for predicting collection effects based on artificial intelligence according to claim 1, characterized in that: In the data collection step, basic customer information includes name, age, occupation and salary level, among which the salary level data comes from the human resources department's salary report or tax return records; financial transaction records cover historical loan amounts, terms, repayment times, and overdue situations; credit reports are obtained from authoritative credit reporting agencies, including credit scores, number of credit inquiries, and historical default details; past interaction records are integrated from the SMS platform, customer service system, and email server, including SMS sending and receiving time, content, and related information on customer service call records.

3. The method for predicting collection effects based on artificial intelligence according to claim 1, characterized in that: In the cleaning step of data preprocessing, for missing numerical values, if the salary level is missing, the value is filled by finding the mean of the same type by double grouping by occupation and age.

4. The method for predicting collection effects based on artificial intelligence according to claim 1, characterized in that: In data preprocessing, the debt-to-income ratio calculation formula for the debt-paying capacity dimension is: L is the total liabilities, I is the monthly stable income, and the asset liquidity index calculation formula is: A liquid is an asset that can be quickly converted into cash, and E is monthly expenditure. The credit score standardization formula of the credit profile dimension is: CS is the original credit score, CS min The lowest credit score in history, CS max is the highest credit score in history, and the calculation formula for the severity of overdue payment is N late is the number of overdue payments, is the maximum number of overdue payments in history, T late is the overdue time, is the maximum overdue time in history; the SMS response enthusiasm calculation formula in the interactive behavior dimension is SRI=0.7×RR+0.3×SES, RR is the SMS reply rate within 24 hours, SES is the reply sentiment tendency score, and the call cooperation calculation formula is TCI=TL×CPD, TL is the customer service call duration, and CPD is the clarity of the promised repayment.

5. The method for predicting collection effects based on artificial intelligence according to claim 1, characterized in that: In the customer feature grouping step, the input of the clustering algorithm is the customer's feature vector, and the output is the grouping result and the group feature vector. During the clustering process, the cluster center is continuously updated to minimize the intra-group sum of square errors and maximize the inter-group distinction.

6. The method for predicting collection effects based on artificial intelligence according to claim 1, characterized in that: In the repayment probability calculation and model evaluation steps, the adjustment factor calculation formula is P repay (c) = P repay (G i )·(1+λ·ΔCS+δ·ΔDTI), where λ and δ are preset weight coefficients, ΔCS is the difference between the customer’s credit score and the mean credit score of its group, and ΔDTI is the difference between the customer’s debt-to-income ratio and the mean debt-to-income ratio of its group.

7. The method for predicting collection effects based on artificial intelligence according to claim 6, characterized in that: In the repayment probability calculation and model evaluation, the training set, validation set, and test set are divided into 70%, 15%, and 15% respectively. The accuracy calculation formula in the evaluation index is: TP is the number of samples with successful repayments predicted correctly, TN is the number of samples with failed repayments predicted correctly, N is the total number of samples, and the calculation formula for recall is: FN is the number of samples with successful repayment but failed prediction. The calculation formula of F1 value is: in FP is the number of samples where repayment was predicted to be successful but actually failed.

8. The method for predicting collection effects based on artificial intelligence according to claim 1, characterized in that: In the customer feature grouping step, the initial cluster center number K of the clustering algorithm ranges from 3 to 10, the matching threshold θ ranges from 0.7 to 0.95, the learning rate α ranges from 0.05 to 0.2, the decay rate β ranges from 0.01 to 0.1, the number of iterations T ranges from 5 to 20, the number of continuous stable iterations t ranges from 2 to 5, the probability change threshold ε ranges from 0.005 to 0.02, the preset weight coefficient γ ranges from 0.1 to 0.3, and the value range of δ ranges from 0.2 to 0.

4. The parameter values ​​are determined through experimental optimization based on the historical data scale, data quality and collection business scenario characteristics.

9. The method for predicting collection effects based on artificial intelligence according to claim 1, characterized in that: During the data preprocessing stage, the weight ratio of each dimensional feature vector constructed by feature engineering is dynamically updated on a quarterly or annual basis according to business scenario requirements. The weight of the debt-paying ability dimension and the credit portrait dimension can be appropriately increased when the market is volatile and the credit environment is tightening. The weight of the interactive behavior dimension is increased during the period of business expansion when customer communication and feedback are important.

10. The method for predicting debt collection effects based on artificial intelligence according to claim 1, characterized in that: In the customer feature grouping stage, after clustering is completed, the newly accumulated data is regularly re-clustered and grouped, and the group feature vector is updated to make the grouping structure fit the latest feature distribution of the customer group, ensuring that subsequent iterative matching and prediction are in line with the actual business evolution trend.

Images