A bill payment system based on business configuration

Abstract

The present application provides a bill payment system based on business configuration, including: obtaining the transaction risk level and the user credit score according to the business configuration parameters of the bill payment system; judging whether the transaction risk level and the user credit score trigger the insurance mechanism through dynamic decision rules; training the transaction problem probability prediction model based on historical transaction data using a machine learning algorithm; predicting the probability of occurrence of the transaction problem using the transaction problem probability prediction model for the current batch processing bills; triggering the insurance mechanism if the probability of occurrence of the transaction problem exceeds a preset threshold, otherwise dynamically adjusting the batch processing quantity according to the bill amount and the user credit score; and dynamically routing the bills by comprehensively considering the transaction risk level, the user credit score, and the payment guarantee level through an intelligent routing algorithm.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a bill payment system based on business configuration. Background Art

[0002] There is a technical contradiction between the pursuit of batch processing efficiency and the provision of insurance protection in the bill payment system. Batch processing can significantly improve the transaction processing capacity and efficiency of the system by packaging multiple transactions together for processing. However, batch processing also means higher transaction risks. Once a transaction in the batch has a problem, it may lead to the failure of the entire batch, causing greater losses to the system and even users. The original intention of the design of the insurance mechanism is to reduce the transaction risk of users, but providing insurance will inevitably consume system resources and reduce system performance.

[0003] How the system balances batch processing and insurance mechanisms is a question that requires in-depth thinking based on the system's business scenarios. Bill payment systems are often oriented to many merchants and a large number of C-end users, with transaction models and risk levels that vary widely. For transactions with low risk levels and good user credit, the system can lower the priority of insurance triggering and use batch processing as much as possible to pursue efficiency; while for high-risk transactions and users with poor credit records, the priority of insurance triggering should be increased and insurance mechanisms should be used to control risks. This requires the system to be able to perform fine-grained control in multiple dimensions such as transaction risk level and user credit score for different business scenarios, and dynamically decide whether to use batch processing and whether to trigger the insurance mechanism, so as to achieve a dynamic balance between batch efficiency and risk prevention and control, and meet differentiated business needs. Summary of the invention

[0004] The present invention provides a bill payment system based on business configuration, which mainly includes:

[0005] Obtain transaction risk level and user credit score based on bill payment system business configuration parameters;

[0006] Determine whether the transaction risk level and the user credit score trigger an insurance mechanism through dynamic decision rules;

[0007] Use machine learning algorithms to train transaction problem probability prediction models based on historical transaction data;

[0008] For the current batch processing bills, use the transaction problem probability prediction model to predict the probability of transaction problems occurring;

[0009] If the probability of the transaction problem occurring exceeds a preset threshold, the insurance mechanism is triggered; otherwise, the batch processing quantity is dynamically adjusted according to the bill amount and the user's credit score;

[0010] The intelligent routing algorithm comprehensively considers the transaction risk level, the user credit score, and the payment security level to dynamically route the bill.

[0011] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0012] The present invention discloses an intelligent bill processing method based on risk assessment. The method first obtains the transaction risk level and user credit score, and determines whether to trigger the insurance mechanism through dynamic decision-making rules. At the same time, a machine learning algorithm is used to train a transaction problem probability prediction model to predict the current batch bills. If the predicted probability exceeds the threshold, the insurance mechanism is triggered, otherwise the batch processing quantity is dynamically adjusted according to the bill amount and credit score. Finally, the risk level, credit score and payment guarantee level are comprehensively considered, and the intelligent routing algorithm is used to dynamically route the bills. Through multi-dimensional risk assessment and intelligent decision-making, the present invention realizes the automation, intelligence and risk controllability of bill processing, effectively improves the efficiency and security of bill processing, and provides users with a better payment experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 The present invention is a flow chart of a bill payment system based on business configuration.

[0014] Figure 2 A schematic diagram of a bill payment system based on business configuration according to the present invention.

[0015] Figure 3 This is another schematic diagram of a bill payment system based on business configuration according to the present invention. DETAILED DESCRIPTION

[0016] In order to further understand the content of the present invention, the present invention is described in detail in conjunction with the accompanying drawings and embodiments. The present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It is understood that the specific embodiments described herein are only used to explain the relevant inventions, rather than to limit the invention. It is also necessary to explain that, for ease of description, only the parts related to the invention are shown in the accompanying drawings.

[0017] like Figure 1-3 In this embodiment, a bill payment system based on business configuration may specifically include:

[0018] S101. Obtain transaction risk level and user credit score according to the business configuration parameters of the bill payment system.

[0019] According to the transaction request, the transaction amount, transaction time and transaction object information are obtained, and the transaction risk is evaluated using a logistic regression model in combination with historical transaction data and user credit records to obtain a transaction risk score. Different measures are taken according to the transaction risk level: if the transaction risk score is greater than or equal to the preset first risk threshold and less than the preset second risk threshold, the user is required to perform identity authentication; if the transaction risk score is greater than or equal to the preset second risk threshold, the upper limit of the transaction amount is limited; the evaluated transaction risk score, together with the transaction amount and transaction object information, is used as the input of the logistic regression model, and the user's credit score is calculated in real time using the random forest algorithm, and the user's credit score is dynamically updated according to the calculation result; the transaction risk score and user credit score are incorporated into the business rule engine of the bill payment, and the system strategy including transaction limit and transaction authority is dynamically adjusted according to the preset rules; after the transaction is completed, the transaction results and user feedback are obtained, and the transaction results and user feedback are used as sample data to regularly retrain the logistic regression model and continuously optimize the model performance; the business rules and threshold settings of the bill payment system are regularly reviewed, the matching degree of risk rules and business requirements is sorted out, and the rules are adjusted when the matching degree of risk rules and business requirements does not meet the requirements, so as to improve the risk prevention and control capabilities of the system.

[0020] Specifically, in the bill payment system, when a shopping transaction request of 500 yuan is received, the system first obtains the amount, time and identity information of the transaction object. Then, the system retrieves the transaction records of the user in the past 6 months from the historical transaction database and finds that the user has made 3 similar shopping transactions, with the amounts of 480 yuan, 550 yuan and 520 yuan respectively, and all payments were completed on time. At the same time, the system also learned from the user's credit file that the user's credit score is 750 points (out of 800 points), which belongs to the user group with good credit. Based on the above information, the system uses a logistic regression model to evaluate the risk of the transaction. After calculation, the risk score of the transaction is 35, which is greater than or equal to the preset first risk threshold of 30, and less than the preset second risk threshold of 40, so the transaction is judged to be low risk. Despite this, the system still requires the user to perform identity authentication of the SMS verification code to confirm the user's identity. If the calculated risk score of the transaction is 45, which is greater than or equal to the preset second risk threshold of 40, the upper limit of the transaction amount will be limited; after identity authentication, the system uses the random forest algorithm to calculate the user's credit score in real time using 10 features such as the user's transaction frequency, transaction amount, and credit score in the past 6 months as input, and obtains the user's latest credit score of 770 points. According to the preset business rules, the system dynamically adjusts the user's single transaction limit to 800 yuan and successfully releases the transaction request of 500 yuan. After the transaction is completed, the system also obtains the user's 5-star praise for this shopping experience and includes it as sample data in the model training set. Through periodic model retraining, the system optimizes the performance of the logistic regression risk assessment model and the random forest credit scoring model, while ensuring business continuity, it also improves the system's risk prevention and control capabilities.

[0021] S102: Determine whether the transaction risk level and the user credit score trigger an insurance mechanism through dynamic decision rules.

[0022] Acquire transaction-related data in the transaction system, wherein the transaction-related data includes transaction amount, transaction time, transaction object, transaction type and transaction location;

[0023] Preprocessing the transaction-related data, wherein the preprocessing includes data cleaning, feature extraction, and normalization of the transaction-related data;

[0024] Inputting the preprocessed transaction-related data into a preset transaction risk assessment model, wherein the transaction risk assessment model adopts a random forest algorithm to output a transaction risk level corresponding to the transaction-related data;

[0025] Obtaining the user's historical transaction data and credit records in the user information system, wherein the user's historical transaction data and credit records include the user's past transaction frequency, transaction amount, transaction type, and credit default status;

[0026] Inputting the user's historical transaction data and credit record into a preset credit scoring model, wherein the credit scoring model uses a logistic regression algorithm to output a credit score corresponding to the user;

[0027] According to the preset risk control rules, the transaction risk level and the credit score are weighted and calculated to obtain a comprehensive risk index;

[0028] If the comprehensive risk index exceeds a preset threshold, a corresponding insurance mechanism is triggered, which includes requiring additional identity verification, limiting the transaction amount, and delaying transaction execution.

[0029] Specifically, in the bill payment system of an e-commerce platform, when a shopping transaction request of 1,000 yuan is received, the system first obtains the transaction amount of 1,000 yuan, the transaction time is 14:30 on May 10, 2023, the transaction object is a well-known brand clothing store, the transaction type is clothing consumption, and the transaction location is Haidian District, Beijing. Then, the system preprocesses these transaction-related data, removes outliers in the transaction amount and transaction time through data cleaning, and extracts five key features such as transaction amount and transaction type, and normalizes the feature data using the maximum and minimum normalization method. Then, the system inputs the preprocessed transaction data into the transaction risk assessment model based on the random forest algorithm. The model conducts a comprehensive analysis of the transaction amount, transaction type and other features, and outputs the risk level of the transaction as "low risk". At the same time, the system also obtains from the user information system that the user's transaction frequency in the past year is an average of 5 times per month, the average monthly transaction amount is 3,000 yuan, the main transaction type is clothing consumption, and there is no credit default record. The system inputs these user historical transaction data and credit records into the credit scoring model using a logistic regression algorithm. After calculation by the model, the user's credit score is 750 points (out of a total of 800 points). Next, the system performs a weighted calculation based on the risk level of the transaction, "low risk", and the user's credit score of 750 points according to the preset risk control rules. The resulting comprehensive risk index is 20, which does not exceed the preset risk threshold of 50. Therefore, the transaction is judged to be a normal transaction, and it is successfully released and the payment process is completed. Through this bill payment system, the e-commerce platform can monitor the risk status of each transaction in real time, and dynamically adjust the risk control strategy according to the user's credit status, which effectively prevents fraud risks while improving payment efficiency.

[0030] S103. Use a machine learning algorithm to train a transaction problem probability prediction model based on historical transaction data.

[0031] Obtain historical transaction data and preprocess the data. Preprocessing includes deleting duplicate records, filling missing values, and standardizing numerical features. Extract key features from the preprocessed data, such as transaction amount, transaction frequency, and transaction time. Combine the extracted features with known problem transaction labels to form a training data set. Based on the characteristics of the training data set, select the random forest algorithm to build a transaction problem probability prediction model. Random forest can handle high-dimensional features, is insensitive to outliers, and has generalization capabilities. Input the training data into the random forest algorithm, and train the initial model by adjusting the number of trees and the depth of the tree. Use the five-fold cross-validation method to evaluate the model performance. Calculate the accuracy, precision, recall, and F1 value on each fold validation set. Based on the evaluation results, use the grid search method to optimize the model parameters, such as the number of trees and the maximum depth, to improve the prediction performance of the model. Apply the same preprocessing and feature extraction methods as the training data to the new transaction data. Input the extracted features into the optimized transaction problem probability prediction model to obtain the probability value of the transaction having problems. Set the risk probability threshold to 0.7. If the predicted probability is greater than the threshold, it is determined to be a high-risk transaction. For high-risk transactions, the system automatically sends risk warning information to relevant operators and suspends the subsequent processing of the transaction. Collect new transaction data and problem feedback every day. When the accumulated new data volume reaches 10% of the original training set, trigger the model update. During the update, merge the new data with the original training set and repeat the model training and optimization process. If the F1 value of the updated model increases by more than 1%, replace the existing model; otherwise, retain the original model. Regularly (such as monthly) manually review the model prediction results. Adjust the feature extraction method or risk probability threshold based on the review feedback. If a new fraud pattern is found, incorporate it into the feature engineering process to improve the model's ability to identify new problematic transactions.

[0032] Specifically, in the transaction risk monitoring system of an e-commerce platform, 5 million historical transaction records in the past year were first obtained from the transaction database. Then, the data were preprocessed, and 2% of duplicate records and 5% of missing value records were removed by data cleaning algorithm, and the maximum and minimum values ​​of numerical features such as transaction amount and transaction time were normalized. Then, 10 key features such as transaction amount, transaction frequency, transaction time, and transaction object were extracted from the preprocessed data, and combined with the corresponding labels of whether there are problems in the transaction to form a training data set containing 4.5 million records. Considering the high-dimensional characteristics of the data set and the possible outliers, the random forest algorithm was selected to construct a transaction problem prediction model. By setting the number of trees to 100 and the maximum depth to 20, the training was performed on the training data set to obtain the initial random forest model. In order to evaluate the performance of the model, the five-fold cross-validation method was used to randomly divide the training set into 5 subsets, and 4 subsets were used to train the model each time, and the remaining 1 subset was used for verification. After 5 cross-validations, the average accuracy of the model on the validation set was 95%, the precision was 91%, the recall was 89%, and the F1 value was 90%. In order to improve the performance of the model, the grid search method was used to optimize the number of trees and the maximum depth, and finally a random forest model with 150 trees and a maximum depth of 25 was obtained. When new transaction data arrives, the system automatically performs the same preprocessing and feature extraction as the training data, and then inputs the extracted features into the optimized random forest model for prediction. If the probability that the transaction is predicted to be a problem transaction is greater than 70%, the system determines it as a high-risk transaction, and automatically sends risk warning information to the relevant personnel, and suspends the payment process of the transaction. In addition, the system will automatically collect new transaction data and problem transaction data reported by users every day. When the accumulated new data reaches 10% of the original training set, that is, 450,000, the system automatically triggers the model update process, merges the new data into the original training set, and retrains and optimizes the model. After a model update, the F1 value of the updated model on the validation set increased from 90% to 92%, an increase of more than 1%, so the system automatically replaced the updated model with a new online prediction model. At the same time, the system will extract 1,000 high-risk transaction data every month for manual review, and dynamically adjust the risk determination threshold and feature extraction method based on the review results to continuously improve the model's risk identification ability.

[0033] S104: For the current batch processing bills, use the transaction problem probability prediction model to predict the probability of transaction problems occurring.

[0034] For batch-processed bill transactions, first obtain the attributes of each transaction from the transaction database, including the transaction object, transaction time, transaction amount, and recent transaction frequency. The acquired transaction attribute data is cleaned and standardized, including removing outliers, filling missing values, and unifying data formats. The processed transaction attribute data is input into a pre-trained transaction problem probability prediction model, which is trained based on historical transaction data and is used to predict the probability of problem occurrence for each transaction. The problem occurrence probability threshold is set to 0.7, and it is determined whether the predicted probability of each transaction exceeds the threshold. For transactions exceeding the threshold, they are marked as high-risk transactions, and additional features of the transaction are obtained, including the credit score of the transaction object, the difference between the transaction amount and the average transaction amount of the account in the past three months, and whether the transaction time is in an unconventional period. All features of high-risk transactions (including initial attributes and additional features) are input into a pre-trained gradient boosting decision tree model, which is used to calculate the suspiciousness score of high-risk transactions. High-risk transactions are sorted in descending order according to the suspiciousness score, and the top 10% transactions with the highest scores are selected as suspicious transactions that need to be focused on and verified. For the suspicious transactions screened out, an early warning report containing transaction details and suspiciousness scores is generated. The early warning report is pushed to the person in charge of the relevant business department in the form of a message through the application program interface of the enterprise's internal instant messaging system. At the same time, a verification task is created in the pending task list of the transaction monitoring system and assigned to a specific business person. After receiving the early warning information, the business person logs into the transaction monitoring system to view the detailed information of the suspicious transaction. According to the preset verification process, the business person needs to contact the transaction parties, verify the transaction background, and review the relevant documents to verify the authenticity and compliance of the transaction. The verification results will be recorded in the transaction monitoring system, including the transaction status (normal, abnormal, fraud, etc.), the verification process description and processing suggestions. For transactions confirmed to be abnormal or fraudulent, the system automatically generates a processing work order and transfers it to the risk control department for investigation and processing. At the same time, the relevant transaction information is fed back to the model training module for regular updating and optimization of the transaction problem probability prediction model to improve the accuracy and adaptability of the model.

[0035] Specifically, in the transaction risk monitoring system of a cross-border e-commerce platform, the attribute data of each cross-border transaction is first obtained in real time through the data interface with the transaction system, including the registered country of the transaction object, the transaction time, the transaction amount (in US dollars) and the number of transactions in the account in the past week. Then, the natural language processing technology is used to standardize the registered country name of the transaction object, unify different languages, synonyms, and abbreviations into standardized English country names; the transaction time is converted to Greenwich Mean Time; the exchange rate is converted to US dollars; the transaction frequency is detected for outliers, and extreme values ​​exceeding 3 times the standard deviation of the mean are eliminated. Then, the processed transaction attribute data is input into a random forest model with 100 trees trained based on 10 million historical transaction data. The model has an AUC (Area Under Curve, i.e., area under the ROC curve) of 0.95 on the test set, which is used to predict the probability of problems in each transaction. If the probability of a transaction problem exceeds the threshold of 0.7, it will be marked as a high-risk transaction, and the transaction object's credit score (provided by a third-party credit reporting agency, ranging from 0-100), the difference between the transaction amount and the average daily transaction amount of the account in the past three months, whether the transaction time is in the high-risk period from 23:00 to 5:00 the next day local time, and other additional features are obtained through the query interface of the transaction system. Then, the 7 features of high-risk transactions (4 initial attributes + 3 additional features) are input into a gradient boosting decision tree model consisting of 500 trees and a maximum tree depth of 8. The accuracy of the model on the validation set is 92%, which is used to calculate the suspiciousness score of high-risk transactions (ranging from 0-100). All high-risk transactions of the day are sorted in descending order according to the suspiciousness score, and the top 10% (if less than 100 transactions, all are selected, and no more than 500 transactions) are taken as suspicious transactions, and an early warning report containing information such as transaction serial number, transaction object account number, transaction amount, transaction time, and suspiciousness score is generated. Through the RestFul API interface of the internal IM system, the early warning report is pushed to the person in charge of the cross-border business department in the form of a rich text message in Json format, and sent with the message as an attachment in CSV format. At the same time, the early warning report information is automatically written into the pending task table in the Mysql database of the transaction monitoring system, and the trigger is automatically assigned to a specific business person, and the status is set to "pending verification". After the business person receives the early warning message through the IM system, he can click the link in the message to log in to the Web page of the transaction monitoring system, find the corresponding early warning transaction in the pending task list, and view the transaction details. The page is embedded with data interfaces of the company's CRM system, ERP system, logistics system, etc. Business personnel can retrieve the historical transaction data, commodity inventory status, logistics order number and other internal information of the transaction object with one click, and can also call the transaction object or send an email with one click to complete the verification efficiently.After the verification is completed, the business personnel select the transaction status (normal / abnormal / fraud) on the page, enter the description of the verification results and processing suggestions, and click Submit. If abnormal or fraud is selected, the system automatically creates a risk event record and pushes it to the risk control department for processing. At the same time, the actual status and related information of the transaction are transmitted to the model training module in Json format through the Kafka real-time data stream. After receiving the feedback data, the model training module starts the feature engineering process, redesigns the features and merges them into the model training set. When the accumulated feedback data reaches 100,000, the model retraining is automatically started, the gradient boosting decision tree model is updated, and the updated model parameters are written to the configuration file to take effect in real time.

[0036] S105. If the probability of the transaction problem occurring exceeds a preset threshold, the insurance mechanism is triggered; otherwise, the batch processing quantity is dynamically adjusted according to the bill amount and the user credit score.

[0037] Obtain the probability of transaction problems and compare it with the preset threshold. If the probability of transaction problems exceeds the preset threshold, the insurance mechanism is triggered and the corresponding risk prevention measures are implemented; if the probability of transaction problems does not exceed the preset threshold, the subsequent transaction processing process continues. Obtain the bill amount and user credit score data of the current transaction, use them as input features, use the pre-trained CART decision tree model to predict, and obtain the adjustment range of the batch processing quantity for the current transaction. The batch processing quantity represents the number of transaction records contained in each batch when the transaction data is batch processed. The adjustment range of the batch processing quantity predicted by the decision tree model is weighted averaged with the currently set batch processing quantity, and the weight is set according to historical data to obtain the new batch processing quantity. Apply the adjusted batch processing quantity to the transaction processing process, perform a corresponding number of batch processing operations on the transaction data, and improve the transaction processing efficiency. Continuously monitor the changes in the probability of transaction problems. When the probability increases, use the grid search algorithm to tune the preset threshold and the hyperparameters of the CART decision tree model, and dynamically update the threshold and batch processing quantity adjustment strategy to adapt to changes in transaction patterns and ensure a dynamic balance between risk prevention and transaction efficiency.

[0038] Specifically, in the transaction risk monitoring system of a cross-border e-commerce platform, the attribute data of each cross-border transaction is first obtained in real time through the data interface with the transaction system, including the registered country of the transaction object, the transaction time, the transaction amount (in US dollars) and the number of transactions in the account in the past week. Then, the natural language processing technology is used to standardize the registered country name of the transaction object, unify different languages, synonyms, and abbreviations into standardized English country names; the transaction time is converted to Greenwich Mean Time; the exchange rate is converted to US dollars; the transaction frequency is detected for outliers, and extreme values ​​exceeding 3 times the standard deviation of the mean are eliminated. Then, the processed transaction attribute data is input into a random forest model of 100 trees trained based on 10 million historical transaction data. The AUC of this model on the test set reaches 0.95, which is used to predict the probability of problems in each transaction. If the probability of problems in a transaction exceeds the preset threshold of 0.8, the insurance mechanism is triggered, and the relevant information of the transaction is automatically sent to the insurance company to apply for the start of the corresponding risk compensation process; if it does not exceed the threshold, the subsequent transaction processing continues. At the same time, the system obtains the bill amount and user credit score of the current transaction as input features, and uses a CART decision tree model trained based on 50,000 historical data to predict the adjustment range of the batch processing quantity for the current transaction. For example, if the current batch processing quantity is set to 50 transactions / batch, and the adjustment range predicted by the decision tree model is +10%, it will be weighted averaged with the original batch quantity at a weight of 7:3 to obtain a new batch processing quantity of 52 transactions / batch. The system continuously monitors the average probability of transaction problems occurring every hour. When the probability rises by more than 15% compared with the previous hour, a grid search algorithm is automatically started, and a 10-fold cross-validation is performed within the parameter range of threshold 0.7-0.9, maximum depth of decision tree 3-8, and minimum number of leaf node samples 10-100 to obtain the optimal model hyperparameter combination, and the preset threshold and batch quantity adjustment strategy are updated in real time to dynamically adapt to changes in transaction patterns.

[0039] S106. Dynamically route the bills by comprehensively considering the transaction risk level, the user credit score, and the payment security level through an intelligent routing algorithm.

[0040] Obtain the information of the bill to be processed, including the transaction amount, transaction type, user history, etc. Extract the transaction records of the user in the past six months from the database, and calculate the transaction frequency, average transaction amount and other features. Based on the transaction frequency and average transaction amount features, use the logistic regression model to calculate the transaction risk score. Compare the risk score with the preset threshold to determine the transaction risk level. Obtain the credit score of the user from the credit scoring system. The credit scoring system uses the weighted average method to calculate a score of 0-100 based on the user's repayment history and credit card usage. Determine the payment security level of the transaction. Check whether the transaction supports refunds and whether it is guaranteed. For transactions that support refunds, assign a value of 1, and for transactions that do not support refunds, assign a value of 0. For transactions that are guaranteed, assign a value of 1, and for transactions that are not guaranteed, assign a value of 0. Add the values ​​of whether refunds are supported and whether there are guarantees to obtain the payment security, ranging from 0 to 2. Use the transaction risk level, user credit score, and payment security as input variables. Use the random forest algorithm to build a decision model. The output of the decision model is the probability distribution of different processing channels. Select the channel with the highest probability as the optimal routing solution. According to the routing solution, assign the bill to the corresponding processing channel. The specific implementation is as follows: establish a processing channel mapping table, and correspond the channel ID output by the decision model to the API interface of the actual processing system. Call the corresponding API to transmit the bill information to the target processing system. After the processing system receives the bill, it returns the processing result. The processing result is recorded in the database for subsequent model optimization and system evaluation. Regularly analyze the processing results, evaluate the accuracy and efficiency of the routing solution, and provide a basis for adjusting the model parameters.

[0041] Specifically, in the intelligent bill routing system of a cross-border e-commerce platform, the attribute information of each bill to be processed is first obtained in real time through the data interface with the transaction system, including the transaction amount, transaction type (such as purchasing goods, recharging, etc.) and user ID. Then, the system extracts the transaction records of the user in the past six months from the database, and uses SQL statements to aggregate and count by month to obtain characteristic variables such as the user's average monthly transaction frequency and average monthly transaction amount. Then, these characteristic variables are input into a logistic regression model trained on 100,000 historical data. The model has an accuracy of 85% on the test set and is used to predict the risk score of each transaction, ranging from 0 to 100. The system sets three risk levels: low risk (0-30), medium risk (30-70), and high risk (70-100). At the same time, the system obtains the credit score of the user from the credit scoring system through the API interface, ranging from 350 to 950 points. In addition, the system will also determine the payment security level of the transaction. For transactions that support refunds, the value is 1, and for transactions that do not support refunds, the value is 0; for transactions with third-party guarantees, the value is 1, and for transactions without guarantees, the value is 0. The two values ​​are added together to get the payment security level. Finally, the system uses a random forest model with 100 trees trained based on 50,000 historical data, taking the transaction risk level, user credit score, and payment security as input variables, predicts the probability distribution of different processing channels, and selects the channel with the highest probability as the optimal routing solution. For example, for a low-risk transaction with a transaction amount of US$500, a user credit score of 650 points, and a payment security level of 1, the model calculates that the probability of being assigned to the manual review channel is 2%, the probability of automatic approval is 95%, and the probability of rejection is 3%. In this case, the transaction will be automatically approved. Based on the routing results, the system calls the corresponding API interface, transmits the billing information to the target processing system in JSON format, and waits for the processing results to be returned. The system will automatically analyze the bill processing results of the previous hour every hour and calculate indicators such as the actual pass rate and rejection rate of each channel. If the deviation from the expected rate exceeds 5%, the model parameters will be automatically adjusted to adapt to changes in bill distribution.

[0042] Obviously, those skilled in the art can make various changes and modifications to the embodiments of the present application without departing from the spirit and scope of the embodiments of the present application. Thus, if these modifications and variations of the embodiments of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A bill payment system based on business configuration, characterized in that: The bill payment system comprises: Obtain transaction risk level and user credit score based on bill payment system business configuration parameters; Determine whether the transaction risk level and the user credit score trigger an insurance mechanism through dynamic decision rules; Use machine learning algorithms to train transaction problem probability prediction models based on historical transaction data; For the current batch processing bills, use the transaction problem probability prediction model to predict the probability of transaction problems occurring; If the probability of the transaction problem occurring exceeds a preset threshold, the insurance mechanism is triggered; otherwise, the batch processing quantity is dynamically adjusted according to the bill amount and the user's credit score; Dynamically route bills by taking into account the transaction risk level, the user credit score, and the payment security level through an intelligent routing algorithm; The method also includes: obtaining the probability of transaction problems occurring, comparing it with a preset threshold, and if the probability of transaction problems occurring exceeds the preset threshold, triggering the insurance mechanism and executing corresponding risk prevention measures; If the probability of transaction problems occurring does not exceed the preset threshold, the subsequent transaction processing process will continue; Obtain the bill amount and user credit score data of the current transaction, use them as input features, use the pre-trained CART decision tree model to make predictions, and obtain the adjustment range of the batch processing quantity for the current transaction; The batch processing quantity indicates the number of transaction records contained in each batch when the transaction data is processed in batches; The adjustment range of the batch processing quantity predicted by the decision tree model is weighted averaged with the currently set batch processing quantity, and the weight is set according to historical data to obtain the new batch processing quantity; Apply the adjusted batch processing quantity to the transaction processing flow, perform a corresponding number of batch processing operations on the transaction data, and improve transaction processing efficiency; Continuously monitor changes in the probability of transaction problems occurring. When the probability of occurrence increases, use the grid search algorithm to tune the preset thresholds and hyperparameters of the CART decision tree model, and dynamically update the threshold and batch processing quantity adjustment strategy to adapt to changes in transaction patterns and ensure a dynamic balance between risk prevention and transaction efficiency.

2. The bill payment system according to claim 1, characterized in that: The obtaining of the transaction risk level and user credit score according to the business configuration parameters of the bill payment system includes: Obtain transaction amount, transaction time, and transaction object information based on transaction requests, and use a logistic regression model to assess transaction risk based on historical transaction data and user credit records to obtain a transaction risk score; Different measures are taken according to the transaction risk level: if the transaction risk score is greater than or equal to the preset first risk threshold and less than the preset second risk threshold, the user is required to perform identity verification; If the transaction risk score is greater than or equal to the preset second risk threshold, the upper limit of the transaction amount will be limited; The assessed transaction risk score, together with the transaction amount and transaction object information, is used as the input of the logistic regression model to calculate the user's credit score in real time using the random forest algorithm, and the user's credit score is dynamically updated based on the calculation results; Incorporate transaction risk scores and user credit scores into the business rule engine for bill payment, and dynamically adjust system policies including transaction limits and transaction permissions based on preset rules; After the transaction is completed, the transaction results and user feedback are obtained, and the transaction results and user feedback are used as sample data to regularly retrain the logistic regression model and continuously optimize the model performance; Regularly review the business rules and threshold settings of the bill payment system, sort out the degree of match between risk rules and business needs, and adjust the rules when the degree of match between risk rules and business needs does not meet the requirements to improve the system's risk prevention and control capabilities.

3. The bill payment system according to claim 1, characterized in that: The determining, by means of dynamic decision rules, whether the transaction risk level and the user credit score trigger an insurance mechanism includes: Acquire transaction-related data in the transaction system, wherein the transaction-related data includes transaction amount, transaction time, transaction object, transaction type and transaction location; Preprocessing the transaction-related data, wherein the preprocessing includes data cleaning, feature extraction, and normalization of the transaction-related data; Inputting the preprocessed transaction-related data into a preset transaction risk assessment model, wherein the transaction risk assessment model adopts a random forest algorithm to output a transaction risk level corresponding to the transaction-related data; Obtaining the user's historical transaction data and credit records in the user information system, wherein the user's historical transaction data and credit records include the user's past transaction frequency, transaction amount and transaction type, and credit default status; Inputting the user's historical transaction data and credit record into a preset credit scoring model, wherein the credit scoring model uses a logistic regression algorithm to output a credit score corresponding to the user; According to the preset risk control rules, the transaction risk level and the credit score are weighted and calculated to obtain a comprehensive risk index; If the comprehensive risk index exceeds a preset threshold, a corresponding insurance mechanism is triggered, which includes requiring additional identity verification, limiting the transaction amount, and delaying transaction execution.

4. The bill payment system according to claim 1, characterized in that: The method of using a machine learning algorithm to train a transaction problem probability prediction model based on historical transaction data includes: Obtain historical transaction data and pre-process the data; Preprocessing includes removing duplicate records, filling missing values, and standardizing numerical features; Extract key features from the preprocessed data, wherein the key features include transaction amount, transaction frequency, and transaction time; Combine the extracted key features with known problematic transaction labels to form a training dataset; Based on the characteristics of the training data set, the random forest algorithm is selected to build a transaction problem probability prediction model; The random forest algorithm can handle high-dimensional features, is insensitive to outliers, and has generalization capabilities; Input the training data into the random forest algorithm, and train the initial model by adjusting the number of trees and the depth parameters of the trees; The five-fold cross validation method was used to evaluate the model performance; Calculate the accuracy, precision, recall and F1 value on each fold of the validation set; Based on the evaluation results, a grid search method is used to optimize model parameters, including the number and maximum depth of trees, to improve the prediction performance of the model; Apply the same preprocessing and feature extraction methods to the new transaction data as the training data; The extracted features are input into the optimized transaction problem probability prediction model to obtain the probability value of the transaction having a problem; The risk probability threshold is set to 0.7; If the predicted probability is greater than the threshold, it is considered a high-risk transaction; For high-risk transactions, the system automatically sends risk warning information to relevant operators and suspends the subsequent processing of the transaction; Collect new transaction data and problem feedback every day; When the accumulated new data reaches 10% of the original training set, the model update is triggered; When updating, the new data is merged with the original training set and the model training and optimization process is repeated; If the F1 value of the updated model increases by more than 1%, the existing model is replaced; Otherwise, keep the original model; Regularly conduct manual review of model prediction results; Adjust feature extraction methods or risk probability thresholds based on review feedback; If new fraud patterns are discovered, they will be incorporated into the feature engineering process to improve the model's ability to identify new problematic transactions.

5. The bill payment system according to claim 1, characterized in that: The method of predicting the probability of transaction problems using the transaction problem probability prediction model for the current batch processing bills includes: For batch-processed bill transactions, first obtain the attributes of each transaction from the transaction database, including the transaction object, transaction time, transaction amount, and recent transaction frequency; Clean and standardize the acquired transaction attribute data, including removing outliers, filling missing values ​​and unifying data formats; Input the processed transaction attribute data into a pre-trained transaction problem probability prediction model, which is trained based on historical transaction data and is used to predict the probability of problems occurring in each transaction; Set the problem occurrence probability threshold to 0.7 and determine whether the predicted probability of each transaction exceeds the threshold; For transactions exceeding the threshold, they are marked as high-risk transactions, and additional features of the transaction are obtained, including the credit score of the transaction object, the difference between the transaction amount and the average transaction amount of the account in the past three months, and whether the transaction time is in an unconventional period; All features of high-risk transactions are input into a pre-trained gradient boosting decision tree model, which is used to calculate the suspicion score of high-risk transactions; Sort high-risk transactions in descending order according to the suspiciousness scores, and select the top 10% of transactions with the highest scores as suspicious transactions that require special attention and verification; For the suspicious transactions screened out, an early warning report containing transaction details and suspicion level scores is generated; Through the application program interface of the enterprise's internal instant messaging system, the early warning report is pushed to the person in charge of the relevant business department in the form of a message; At the same time, create a verification task in the pending task list of the transaction monitoring system and assign it to a specific business person; After receiving the warning information, the business personnel log in to the transaction monitoring system to view the detailed information of the suspicious transaction; According to the preset verification process, business personnel need to verify the authenticity and compliance of the transaction by contacting the transaction parties, verifying the transaction background and reviewing relevant documents; The verification results will be recorded in the transaction monitoring system, including transaction status, description of the verification process and handling suggestions; For transactions confirmed to be abnormal or fraudulent, the system automatically generates a processing ticket and transfers it to the risk control department for investigation and processing; At the same time, relevant transaction information is fed back to the model training module to regularly update and optimize the transaction problem probability prediction model to improve the accuracy and adaptability of the model.

6. The bill payment system according to claim 1, characterized in that: The intelligent routing algorithm comprehensively considers the transaction risk level, the user credit score, and the payment guarantee level to dynamically route the bill, including: Obtain pending bill information, including transaction amount, transaction type, and user history; Extract the user's transaction records for the past six months from the database and calculate the transaction frequency and average transaction amount characteristics; Calculate the transaction risk score using a logistic regression model based on the transaction frequency and average transaction amount characteristics; Compare the risk score with the preset threshold to determine the transaction risk level; Obtain the user's credit score from the credit scoring system; The credit scoring system uses a weighted average method to calculate a score of 0-100 based on the user's repayment history and credit card usage; Determine the payment security level of the transaction; Check whether the transaction supports refunds and whether it is guaranteed; For transactions that support refunds, the value is 1; for transactions that do not support refunds, the value is 0. Secured transactions are assigned a value of 1, and unsecured transactions are assigned a value of 0; The payment security degree is obtained by adding the value of whether a refund is supported and the value of whether there is a guarantee, ranging from 0 to 2; Transaction risk level, user credit score and payment security are used as input variables; Use the random forest algorithm to build a decision model; The output of the decision model is the probability distribution of different processing channels; Select the channel with the highest probability as the optimal routing solution; According to the routing plan, the bills are assigned to the corresponding processing channels; The specific implementation is: establish a processing channel mapping table to correspond the channel ID output by the decision model with the API interface of the actual processing system; Call the corresponding API to transfer the billing information to the target processing system; After the processing system receives the bill, it returns the processing result; Record the processing results into the database for subsequent model optimization and system evaluation; Regularly analyze and process results to evaluate the accuracy and efficiency of routing solutions and provide a basis for adjusting model parameters.

Images