Credit card overdraft account urging notification management method

Abstract

The invention provides a credit card overdraft account urging notification management method, which belongs to the technical field of financial services, and comprises the following steps: obtaining multi-source data of an overdraft account; performing feature extraction on the multi-source data of the overdraft account to obtain key feature data of the overdraft account; constructing an overdraft portrayal model based on a machine learning algorithm by using the key feature data of the overdraft, and generating an overdraft portrayal corresponding to the dimension; and inputting the overdraft account portraits of different dimensions into a strategy matching library so as to output an overdraft account urging notification strategy. According to the method, a comprehensive and detailed overdraft account portrait model is constructed through multi-source data acquisition and feature extraction, the model comprehensively considers consumption behaviors, credit risks and repayment ability factors, features of each overdraft account are accurately described, and compared with traditional extensive notification of a unified format and a single channel, the method better fits the actual situation of the overdraft account, and the method is more suitable for popularization and application. The pertinence of notification is improved, and repayment of overdraft users is effectively triggered.

Description

Technical Field

[0001] The present invention relates to the technical field of financial services, and in particular to a credit card overdraft reminder notice management method. Background Art

[0002] With the vigorous development of the modern financial market, credit cards, as a convenient payment tool and credit consumption means, have been issued and used in a growing range of ways. Credit cards provide consumers with instant financial support, which has promoted the prosperity of the consumer market and boosted economic development. However, the problem of credit card overdrafts has become increasingly prominent, bringing many challenges to financial institutions.

[0003] Traditional reminder notification management methods are often relatively extensive, and mostly use unified formats, fixed times, and single-channel notification methods, such as simple SMS group or bulk email. This method cannot fully consider the individual differences of overdraft accounts, including consumption habits, repayment ability, credit history and other factors, resulting in the notification being ignored or failing to effectively trigger overdraft accounts to repay. In addition, the lack of in-depth analysis of the notification effect and dynamic adjustment mechanism makes it difficult to adapt to the complex and changing credit card business environment and customer needs. Therefore, we propose a credit card overdraft account reminder notification management method to solve the above problems. Summary of the invention

[0004] In view of the shortcomings of the prior art, the present invention provides a credit card overdraft reminder notification management method to solve the problems raised in the above background technology.

[0005] The purpose of the present invention can be achieved by the following technical solution: comprising the following steps: S1, obtain multi-source data of overdraft accounts; S2, extracting features from multi-source data of overdraft households to obtain key feature data of overdraft households; key feature data include consumption behavior features, credit risk features, and repayment ability features, wherein consumption behavior features include high-frequency consumption category features, consumption time distribution features, and consumption location clustering features; credit risk features include credit score feature values, overdue record quantitative feature values, and credit limit change trend feature values; and repayment ability features include income source stability feature values ​​and debt burden ratio feature values; S3, using the key feature data of overdraft customers based on machine learning algorithms to build an overdraft customer portrait model; S4, deploy the overdraft account portrait model to the production environment, input the key feature data of the overdraft account, and generate the overdraft account portrait corresponding to the dimension; S5, setting a strategy matching library, including a plurality of reminder notification strategies; inputting the overdraft account portraits of different dimensions into the strategy matching library to output the reminder notification strategy of the overdraft account; S6. Collect the historical response data of overdrawn customers on different notification channels, and use the Naive Bayes classification algorithm or the Logistic Regression algorithm. Taking the historical response data as the training set and the overdrawn customer portrait as the input variable, construct a channel preference prediction model to predict the acceptance probability of overdrawn customers for different channels and the ranking of preference weights; where the different notification channels include text messages, emails, voice calls, and push notifications from the mobile banking APP, and the historical response data includes the response rate, open rate, click-through rate, call duration, and subsequent behavior feedback. S7. Determine the notification channel combination and push order for each user according to the reminder notification strategy output by the policy generation module and the prediction results of the channel preference model.

[0006] Preferably, construct an overdrawn customer portrait model based on the key feature data of overdrawn customers using a machine learning algorithm. The specific steps are as follows: S31. Use the extracted key features as the model input to generate an overdrawn customer portrait containing multi-dimensional information: S311. Determine the network structure. Determine the number of neurons in the input layer according to the number of key features, denoted as n; set three hidden layers, and set the number of neurons in the hidden layer according to the layer index of the hidden layer, denoted as i; the number of neurons in the first hidden layer can be set to 2n, the number of neurons in the second hidden layer is n, and the number of neurons in the third hidden layer is 2 / n; the activation function of each hidden layer uses the ReLU function, and its formula is ; where x represents the result after weighted summation of the values passed from the neurons in the previous layer; set the number of output layer neurons to be the same as the number of dimensions of the overdrawn customer portrait, denoted as m; Perform normalization processing on the key feature data to make the value ranges of different features within the set interval; S312. Divide the preprocessed key feature data into a training set and a test set according to a set ratio; S313. Initialize the overdrawn customer portrait model and train and optimize it. Specifically: Use the weighted mean square error as the loss function, and the formula is: , where m represents the number of neurons in the output layer, that is, the number of dimensions of the portrait, yi is the true value of the i-th output feature, is the predicted value of the i-th output feature of the model, and wi is the weight of the i-th output feature; Then use the Adam optimization algorithm to update the parameters; Set up the training process: Set the number of iterations for training the model; in each iteration, batch input the training set data into the overdrawn customer portrait model and calculate the loss function value; according to the loss function value, use the Adam optimization algorithm to calculate the gradient and update the model parameters; set the iteration evaluation frequency period, and whenever the number of iterations reaches the iteration evaluation frequency period, evaluate the performance of the overdrawn customer portrait model on the test set and calculate the evaluation metrics, where the evaluation metrics include mean squared error and accuracy; perform a weighted calculation on all evaluation metrics to obtain a comprehensive evaluation value; if the model evaluation value is greater than its preset threshold, it indicates that the model training is completed; otherwise, if the model evaluation value is less than or equal to its preset threshold, it means that the model evaluation metrics do not meet the expected requirements, and then perform a tuning operation; S32. Set the feature update period. When the update period is reached, re-collect the new key feature data of the overdrawn customer, and use the cosine similarity algorithm to calculate the degree of difference between the new key feature data and the key feature data of the previous period to obtain the feature change amplitude; If the feature change amplitude is less than the preset change amplitude threshold, it is determined that no update is required and the original portrait model is maintained; otherwise, it indicates that the feature change is significant, and then use the new key features to perform incremental learning and update on the overdrawn customer portrait model; S33. Deploy the trained overdrawn customer portrait model to the production environment, input the key feature data of the overdrawn customer, and generate the overdrawn customer portrait corresponding to the dimension.

[0007] Preferably, the method for obtaining the consumption behavior characteristics includes: Obtain the transaction data of the overdrawn customer, classify the transaction data according to the consumption categories, and count the transaction times of each category; apply the frequent camera mining algorithm to calculate the support degree of each consumption category; screen the consumption categories with a support degree greater than its preset threshold and mark them as high-frequency consumption categories; form the high-frequency consumption category characteristics from the high-frequency consumption categories; Obtain the transaction time information of the overdrawn customer's credit card, convert the transaction time information into a time series analysis format, including aggregating and counting the transaction times according to the time granularity of hours, days, weeks, and months; select a set time series analysis module to perform model fitting and parameter estimation on the transaction time information, and extract the transaction frequency characteristics of different time granularities according to the results of the model fitting; form the consumption time distribution characteristics from all the transaction frequency characteristics; Obtain the longitude and latitude information of the transaction location of the credit card, use the longitude and latitude information of any transaction location of the credit card as a data point, construct a data point set from all the data points of the overdrawn customer, perform clustering analysis on the data points according to the set neighborhood radius and minimum number of points to obtain the clustering area; count the proportion of transaction times of the overdrawn customer in different clustering areas; mark the proportion of transaction times in the clustering area as the consumption location clustering characteristic; Mark the high-frequency consumer product category characteristics, consumption time distribution characteristics, and consumption location clustering characteristics as consumption behavior characteristics.

[0008] Preferably, the method for obtaining the credit risk characteristics includes: Obtain the current credit score of the overdraw account through a credit assessment agency; set the original usage score of the user, identify the maximum credit score and the minimum credit score during the use of the credit card by the overdraw account, and normalize the current credit score, the original usage score, the maximum credit score, and the minimum credit score during the use process of the overdraw account to obtain a credit score characteristic value; Obtain the overdue information of the overdraw account, including the number of overdue times, the duration and amount of each overdue; set the reasonable allowable value of any overdue parameter in the overdue information, and subtract the corresponding reasonable allowable value from the value of any overdue parameter in the overdue information to obtain a reasonable difference; perform weighted calculation on the reasonable differences of all overdue parameters in the overdue information to obtain an overdue severity index, which is used as the quantitative characteristic value of the overdue record; Obtain the credit limit adjustment record data of the credit card of the overdraw account, and extract any credit limit and its adjustment time from the credit limit adjustment record data; calculate the difference between adjacent adjustment times to obtain the adjacent adjustment time difference; calculate the difference between adjacent credit limits to obtain the adjacent adjustment amount difference; calculate the variance of all adjacent adjustment time differences and adjacent adjustment amount differences during the use of the credit card to obtain the adjacent adjustment time fluctuation value and the adjacent adjustment amount fluctuation value; perform weighted calculation on the adjacent adjustment time fluctuation value and the adjacent adjustment amount fluctuation value to obtain a credit limit change trend characteristic value; Mark the credit score characteristic value, the quantitative characteristic value of the overdue record, and the credit limit change trend characteristic value as credit risk characteristics.

[0009] Preferably, the method for obtaining the repayment ability characteristics includes: Obtain the salary income information of the overdraw account, including the salary payment time, salary income, and investment and financial management income; set the income evaluation time zone, calculate the time interval between any two adjacent salary payment times within the income evaluation time zone and mark it as the salary payment interval; calculate the variance of the salary payment intervals within the income evaluation time zone to obtain the salary payment stability value; calculate the variance of the salary income during the income evaluation period to obtain the income fluctuation value; calculate the standard deviation of the investment and financial management income within the income evaluation time zone to obtain the investment stability value; perform weighted calculation on the salary payment stability value, the income fluctuation value, and the investment stability value to obtain an income source stability characteristic value; Calculate the debt burden ratio characteristic value by dividing the sum of various loans and credit card overdrafts of the overdraw account by the total assets; where various loans include the balance of housing loans, the balance of vehicle loans, and the total credit card overdrafts; total assets include the value of real estate, the value of vehicles, the balance of deposits, salary income, and investment and financial management income; The characteristic values ​​of income source stability and debt burden ratio are marked as repayment ability characteristics.

[0010] As a preferred method, the new key feature data is used to perform incremental learning updates on the overdraft household portrait model, specifically: A gradient-based incremental learning algorithm is used to merge the new data with part of the original training set to form a new training set. The overdraft household portrait model is retrained on the new training set using a set multiple of the original learning rate, and the parameters of all layers of the model are updated.

[0011] Preferably, the channel preference prediction model is constructed using a naive Bayes classification algorithm or a logistic regression algorithm, with historical response data of different notification channels as a training set and overdraft customer portraits as input variables, to predict overdraft customers' acceptance probabilities and preference weight rankings for different channels.

[0012] As a preferred method, the overdraft account portraits of different dimensions are input into the strategy matching library to output the reminder notification strategy of the overdraft account, and the specific steps are as follows: Obtain the consumption behavior characteristics, credit risk characteristics, and repayment ability characteristics of overdraft customers, and match the overdraft customer profile with the predefined strategies in the strategy matching library: The strategies in the strategy matching library are classified and stored according to credit risk level, consumption behavior pattern and repayment ability status; the credit risk level includes low risk, medium risk and high risk, the consumption behavior pattern includes stable consumption type, flexible consumption type and high consumption type, and the repayment ability status includes strong, medium and weak; each category contains specific reminder notification strategies for different feature combinations; When an overdraft account profile is input, the system determines the credit risk level range to which the account belongs based on the credit risk characteristics in the profile, and then further screens and matches the strategies under that level based on the consumption behavior characteristics and repayment ability characteristics; Thus, a reminder notification strategy corresponding to the overdraft account portrait is output, and the reminder notification strategy includes notification tone, reminder frequency, preferential measures and notification content customization.

[0013] Preferably, in the present application, after step S8, step S9 is also included to establish a multi-dimensional reminder effect evaluation index system, use big data analysis technology to count and analyze various indicators within the set notification period, and generate an evaluation report within the set notification period; wherein various indicators include notification delivery rate, reading / listening rate, response rate, and repayment conversion rate; according to the evaluation report, a reinforcement learning algorithm is used to optimize and adjust the intelligent reminder notification strategy.

[0014] Compared with the prior art, the present invention has the following beneficial effects: Through multi-source data acquisition and feature extraction, the present invention constructs a comprehensive and detailed portrait model of overdrawn account holders. This model comprehensively considers consumption behavior, credit risk, and repayment ability factors, and can accurately depict the characteristics of each overdrawn account holder. Based on this, the formulated reminder notice strategy is highly personalized. Compared with the traditional extensive notice in a unified format and single channel, it is more in line with the actual situation of overdrawn account holders, improves the pertinence of the notice, thus effectively prompting overdrawn account holders to repay and enhancing the notice effect.

[0015] By constructing a channel preference prediction model, using historical response data based on the Naive Bayes classification algorithm or Logistic Regression algorithm, and analyzing in combination with the portrait of overdrawn account holders, the present invention can accurately predict the acceptance probability and preference weight ranking of overdrawn account holders for different channels such as text messages, emails, voice calls, and push notifications of mobile banking APPs. This enables financial institutions to select the most suitable notification channel combination and push order for each user according to the prediction results, avoiding resource waste on ineffective channels and significantly improving the notification delivery rate and user response rate. Brief Description of the Drawings

[0016] For the convenience of those skilled in the art to understand, the present invention will be further described below in conjunction with the accompanying drawings.

[0017] Figure 1 It is a flowchart of a method for managing reminder notices for credit card overdrawn account holders of the present invention. Detailed Embodiments

[0018] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0019] Please refer to Figure 1 As shown, a method for managing reminder notices for credit card overdrawn account holders includes the following steps: S1. Obtain multi-source data of overdrawn account holders; S2. Extract features from the multi-source data of overdrawn account holders to obtain key feature data of overdrawn account holders; the key feature data includes consumption behavior features, credit risk features, and repayment ability features. Among them, the consumption behavior features include high-frequency consumption category features, consumption time distribution features, and consumption location clustering features; the credit risk features include credit score feature values, overdue record quantification feature values, and credit limit change trend feature values; the repayment ability features include stable income source feature values and debt burden ratio feature values; S3. Use the key feature data of overdrawn account holders to construct a portrait model of overdrawn account holders based on machine learning algorithms; S4. Deploy the overdrawn account portrait model to the production environment, input the key feature data of the overdrawn account, and generate the overdrawn account portrait corresponding to the dimension; S5. Set up a policy matching library, including several reminder notice policies; input the overdrawn account portraits of different dimensions into the policy matching library to output the reminder notice policy for the overdrawn account; S6. Collect the historical response data of the overdrawn account on different notification channels, and use the Naive Bayes classification algorithm or the Logistic Regression algorithm. Using the historical response data as the training set and the overdrawn account portrait as the input variable, construct a channel preference prediction model to predict the acceptance probability and preference weight ranking of the overdrawn account for different channels; where the different notification channels include text messages, emails, voice calls, and mobile banking APP push notifications, and the historical response data includes the response rate, open rate, click-through rate, call duration, and subsequent behavior feedback; S7. Determine the notification channel combination and push order for each user according to the reminder notice policy output by the policy generation module and the prediction results of the channel preference model.

[0020] In this application, an overdrawn account portrait model is constructed based on machine learning algorithms using the key feature data of the overdrawn account. The specific steps are as follows: S31. Use the extracted key features as the model input to generate an overdrawn account portrait containing multiple dimension information: S311. Determine the network structure. Determine the number of neurons in the input layer according to the number of key features, and set it as n; set three hidden layers, and set the number of neurons in the hidden layer according to the layer index of the hidden layer, with the layer index as i; the number of neurons in the first hidden layer can be set as 2n, the number of neurons in the second hidden layer is n, and the number of neurons in the third hidden layer is 2 / n; the activation function of each hidden layer uses the ReLU function, and its formula is ; where x represents the result after weighted summation of the values passed from the neurons in the previous layer; set the number of output layer neurons to be the same as the number of dimensions of the overdrawn account portrait, denoted as m; Normalize the key feature data so that the value ranges of different features are within the set interval; S312. Divide the preprocessed key feature data into a training set and a test set according to the set ratio; S313. Initialize the overdrawn account portrait model and train and optimize it. Specifically: Use the weighted mean square error as the loss function, and the formula is: , where m represents the number of neurons in the output layer, that is, the number of portrait dimensions, yi is the true value of the i-th output feature, is the predicted value of the i-th output feature of the model, and wi is the weight of the i-th output feature. The weight of the output feature can be preset according to the importance of the feature; Then use the Adam optimization algorithm to update the parameters, and its update formula is as follows: ; ; ; ; ; where and are the first-order moment estimate and the second-order moment estimate respectively, , are the first-order moment estimate and the second-order moment estimate at step t - 1 respectively, and are the decay factors, and usually take the values of 0.9 and 0.999 respectively, , is the t-th power of the decay factor, is the gradient of the loss function with respect to the parameter , is the parameter value at step t + 1, and are the corrected first-order moment estimate and the second-order moment estimate, is the learning rate, usually taking the value of 0.001, is used to prevent the denominator from being zero; Then set the training process: set the number of iterations for training the model; in each iteration, batch the training set data and input it into the overdrawn household portrait model to calculate the loss function value; according to the loss function value, use the Adam optimization algorithm to calculate the gradient and update the model parameters; set the iteration evaluation frequency period, and whenever the number of iterations reaches the iteration evaluation frequency period, evaluate the performance of the overdrawn household portrait model on the test set, calculate the evaluation metrics, and the evaluation metrics include mean squared error and accuracy; perform weighted calculation on all evaluation metrics to obtain the comprehensive evaluation value; if the model evaluation value is greater than its preset threshold, it means the model training is completed; otherwise, if the model evaluation value is less than or equal to its preset threshold, it means the model evaluation metrics do not meet the expected requirements, and then perform the tuning operation; the tuning methods include adjusting the number of hidden layers and the number of neurons in the hidden layers, adjusting the learning rate, decay factor, and batch size; S32, set the feature update period, and when the update period is reached, re-collect the new key feature data of the overdrawn household, and use the cosine similarity algorithm to calculate the degree of difference between the new key feature data and the key feature data in the previous period to obtain the feature change amplitude; If the change amplitude of the feature is less than the preset change amplitude threshold, it is determined that no update is required, and the original portrait model is maintained; otherwise, it indicates that the feature change is significant, and incremental learning update is performed on the overdrawn household portrait model using the new key features; S33. Deploy the trained overdrawn household portrait model to the production environment, input the key feature data of the overdrawn household, and generate the corresponding overdrawn household portrait in dimensions.

[0021] In this application, the method for obtaining consumption behavior characteristics includes: Obtain the transaction data of the overdrawn household, classify the transaction data by consumption categories, and count the transaction times of each category; apply the frequent itemset mining algorithm to calculate the support degree of each consumption category; screen the consumption categories with a support degree greater than the preset threshold, and mark them as high-frequency consumption categories; form the high-frequency consumption category characteristics from the high-frequency consumption categories; Obtain the transaction time information of the overdrawn household's credit card, convert the transaction time information into a time series analysis format, including aggregating and counting the transaction times according to the time granularity of hours, days, weeks, and months; select a set time series analysis module to perform model fitting and parameter estimation on the transaction time information, and extract the transaction frequency characteristics of different time granularities according to the results of the model fitting; form the consumption time distribution characteristics from all the transaction frequency characteristics; Obtain the longitude and latitude information of the transaction location of the credit card, use the longitude and latitude information of any transaction location of the credit card as a data point, construct a data point set from all the data points of the overdrawn household, perform clustering analysis on the data points according to the set neighborhood radius and minimum number of points to obtain a clustering area; count the proportion of transaction times of the overdrawn household in different clustering areas; mark the proportion of transaction times in the clustering area as the consumption location clustering characteristic; Mark the high-frequency consumption category characteristics, consumption time distribution characteristics, and consumption location clustering characteristics as consumption behavior characteristics.

[0022] In this application, the method for obtaining credit risk characteristics includes: Obtain the current credit score of the overdrawn household through a credit assessment agency; set the original usage score of the user, identify the maximum credit score and the minimum credit score during the usage process of the overdrawn household for this credit card, and perform normalization processing on the current credit score, original usage score, maximum credit score, and minimum credit score during the usage process of the overdrawn household to obtain the credit score characteristic value; Obtain the overdue information of the overdrawn household, including the number of overdue times, the duration and amount of each overdue; set the reasonable allowable value of any overdue parameter in the overdue information, subtract the numerical value of any overdue parameter in the overdue information from its corresponding reasonable allowable value to obtain a reasonable difference; perform weighted calculation on the reasonable differences of all overdue parameters in the overdue information to obtain the overdue severity index, which is used as the overdue record quantification characteristic value; Obtain the credit limit adjustment record data of the credit card of the overdrawn household, extract any credit limit and its adjustment time from the credit limit adjustment record data; calculate the difference between adjacent adjustment times to obtain the adjacent adjustment time difference; calculate the difference between adjacent credit limits to obtain the adjacent adjustment amount difference; calculate the variances of all the adjacent adjustment time differences and adjacent adjustment amount differences during the use of the credit card to obtain the adjacent adjustment time fluctuation value and adjacent adjustment amount fluctuation value; perform weighted calculation on the adjacent adjustment time fluctuation value and adjacent adjustment amount fluctuation value to obtain the credit limit change trend eigenvalue; Mark the credit score eigenvalue, overdue record quantization eigenvalue, and credit limit change trend eigenvalue as credit risk characteristics.

[0023] In this application, the method for obtaining the repayment ability characteristics includes: Obtain the salary income information of the overdrawn household, including salary payment time, salary income, and investment and financial management income; set the income evaluation time zone, calculate the time interval between any two adjacent salary payment times within the income evaluation time zone and mark it as the salary payment interval; calculate the variance of the salary payment intervals within the income evaluation time zone to obtain the salary payment stability value; calculate the variance of the salary income during the income evaluation period to obtain the income fluctuation value; calculate the standard deviation of the investment and financial management income within the income evaluation time zone to obtain the investment stability value; perform weighted calculation on the salary payment stability value, income fluctuation value, and investment stability value to obtain the income source stability eigenvalue; Calculate the debt burden ratio eigenvalue by dividing the sum of various loans of the overdrawn household and the total credit card overdraft by the total assets; where various loans include the balance of housing loans, the balance of car loans, and the total credit card overdraft; total assets include the value of real estate, the value of vehicles, the balance of deposits, salary income, and investment and financial management income; Mark the income source stability eigenvalue and the debt burden ratio eigenvalue as repayment ability characteristics.

[0024] In this application, use the new key feature data to perform incremental learning and update on the overdrawn household portrait model, specifically: Adopt the gradient-based incremental learning algorithm, combine the new data with a part of the original training set to form a new training set; use a set multiple of the original learning rate to retrain the overdrawn household portrait model on the new training set and update the parameters of all layers of the model.

[0025] In this application, the channel preference prediction model is constructed using the Naive Bayes classification algorithm or the Logistic Regression algorithm. Using the historical response data of different notification channels as the training set and the overdrawn household portrait as the input variable, predict the acceptance probability and preference weight ranking of the overdrawn household for different channels.

[0026] In this application, input the overdrawn household portraits of different dimensions into the policy matching library to output the reminder notice policy for this overdrawn household, and its specific steps are: Obtain the consumption behavior characteristics, credit risk characteristics, and repayment ability characteristics of overdraft customers, and match the overdraft customer profile with the predefined strategies in the strategy matching library: The strategies in the strategy matching library are classified and stored according to credit risk level, consumption behavior pattern and repayment ability status; the credit risk level includes low risk, medium risk and high risk, the consumption behavior pattern includes stable consumption type, flexible consumption type and high consumption type, and the repayment ability status includes strong, medium and weak; each category contains specific reminder notification strategies for different feature combinations; When an overdraft account profile is input, the system determines the credit risk level range to which the account belongs based on the credit risk characteristics in the profile, and then further screens and matches the strategies under that level based on the consumption behavior characteristics and repayment ability characteristics; Thus, a reminder notification strategy corresponding to the overdraft account portrait is output, and the reminder notification strategy includes notification tone, reminder frequency, preferential measures and notification content customization.

[0027] In the present application, based on step S8, step S9 is also included to establish a multi-dimensional reminder effect evaluation index system, use big data analysis technology to count and analyze various indicators within the set notification period, and generate an evaluation report within the set notification period; wherein various indicators include notification delivery rate, reading / listening rate, response rate, and repayment conversion rate; according to the evaluation report, a reinforcement learning algorithm is used to optimize and adjust the intelligent reminder notification strategy.

[0028] It should be noted that by establishing a multi-dimensional reminder effect evaluation index system, covering key indicators such as notification delivery rate, reading / listening rate, response rate, repayment conversion rate, etc., and using big data analysis technology to conduct statistics and analysis on various indicators within the set notification cycle, an evaluation report is generated, and then based on the reinforcement learning algorithm, the shortcomings of the current reminder notification strategy can be discovered in a timely manner based on the evaluation results, and intelligent optimization and adjustment can be made to ensure that the strategy always adapts to the ever-changing business environment and customer needs, continuously improves the efficiency and quality of credit card overdraft management, and reduces overdue risks.

[0029] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. The above embodiments and descriptions are only preferred examples of the present invention and are not intended to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A credit card overdraft reminder notification management method, characterized in that: The following steps are involved: S1, obtain multi-source data of overdraft households; S2, extracting features from multi-source data of overdraft households to obtain key feature data of overdraft households; Key feature data include consumption behavior characteristics, credit risk characteristics, and repayment ability characteristics. Among them, consumption behavior characteristics include high-frequency consumption category characteristics, consumption time distribution characteristics, and consumption location clustering characteristics. Credit risk characteristics include credit score characteristic values, overdue record quantitative characteristic values, and credit limit change trend characteristic values. Repayment ability characteristics include income source stability characteristic values ​​and debt burden ratio characteristic values. S3, using the key feature data of overdraft customers based on machine learning algorithms to build an overdraft customer portrait model; S4, deploy the overdraft account portrait model to the production environment, input the key feature data of the overdraft account, and generate the overdraft account portrait corresponding to the dimension; S5, setting a strategy matching library, including a plurality of reminder notification strategies; inputting the overdraft account portraits of different dimensions into the strategy matching library to output the reminder notification strategy of the overdraft account; S6, collect historical response data of overdraft users in different notification channels, apply naive Bayes classification algorithm or logistic regression algorithm, use historical response data as training set and overdraft user portrait as input variable, build channel preference prediction model, predict overdraft users' acceptance probability and preference weight ranking for different channels; different notification channels include SMS, email, voice call, mobile banking APP push, historical response data include reply rate, open rate, click-to-read rate, answering time and subsequent behavior feedback; S7, determining a notification channel combination and a push order for each user according to the reminder notification strategy output by the strategy generation module and the prediction result of the channel preference model.

2. A credit card overdraft reminder notification management method according to claim 1, characterized in that: Using the key feature data of overdraft accounts, we build an overdraft account portrait model based on a machine learning algorithm. The specific steps are as follows: S31, using the extracted key features as model input to generate an overdraft account portrait containing multiple dimensions of information: S311, determine the network structure, determine the number of input layer neurons according to the number of key features, set it as n; set three hidden layers, set the number of hidden layer neurons according to the number of hidden layers, and the layer index is i; the number of neurons in the first hidden layer can be set to 2n, the number of neurons in the second hidden layer is n, and the number of neurons in the third hidden layer is 2 / n; the activation function of each hidden layer adopts the ReLU function, and its formula is ; x represents the result of weighted summation of the values ​​transmitted from the previous layer of neurons; the number of neural units in the output layer is set to be consistent with the number of dimensions of the overdraft household portrait, denoted as m; Normalize key feature data so that the value range of different features is within the set range; S312, dividing the preprocessed key feature data into a training set and a test set according to a set ratio; S313, initialize the overdraft account portrait model and train and optimize it, specifically: The weighted mean square error is used as the loss function, and the formula is: , where m represents the number of neurons in the output layer, that is, the number of image dimensions, and yi is the true value of the i-th output feature. is the i-th output feature value predicted by the model, and wi is the weight of the i-th output feature; Then use the Adam optimization algorithm to update the parameters; Then set the training process: set the number of iterations for training the model; In each iteration, the training set data is batch-inputted into the overdraft household portrait model to calculate the loss function value; based on the loss function value, the Adam optimization algorithm is used to calculate the gradient and update the model parameters; Set an iterative evaluation cycle. Whenever the number of iterations reaches the iterative evaluation cycle, evaluate the performance of the overdraft household portrait model on the test set and calculate the evaluation indicators, including mean square error and accuracy. Perform weighted calculation on all evaluation indicators to obtain a comprehensive evaluation value. If the model evaluation value is greater than its preset threshold, it means that the model training is completed. On the contrary, if the model evaluation value is less than or equal to its preset threshold, it means that the model evaluation index does not meet the expected requirements, and then the tuning operation is performed; S32, setting a feature update cycle, and when the update cycle is reached, re-collecting new key feature data of the overdraft account, and using the cosine similarity algorithm to calculate the difference between the new key feature data and the key feature data of the previous cycle to obtain the feature change amplitude; If the feature change amplitude is less than the preset change amplitude threshold, it is determined that no update is required and the original portrait model is maintained; otherwise, it indicates that the feature change is significant, and the new key features are used to perform incremental learning updates on the overdraft household portrait model; S33, deploy the trained overdraft account portrait model to the production environment, input the key feature data of the overdraft account, and generate the overdraft account portrait corresponding to the dimension.

3. A credit card overdraft reminder notification management method according to claim 1, characterized in that: The method for acquiring the consumption behavior characteristics includes: Obtain the transaction data of overdraft customers, classify the transaction data according to consumer categories, and count the number of transactions for each category; apply the frequent camera mining algorithm to calculate the support of each consumer category; filter out consumer categories with support greater than the preset threshold and mark them as high-frequency consumer categories; and form high-frequency consumer category features from high-frequency consumer categories; Obtain the transaction time information of the overdraft credit card, and convert the transaction time information into a time series analysis format, including aggregating and counting the number of transactions by time granularity of hours, days, weeks, and months; select a set time series analysis module to perform model fitting and parameter estimation on the transaction time information, and extract transaction frequency features of different time granularities based on the results of model fitting; all transaction frequency features constitute consumption time distribution features; The latitude and longitude information of the credit card transaction location is obtained, and the latitude and longitude information of any credit card transaction location is used as a data point. A data point set is constructed from all the data points of the overdraft account, and the data points are clustered according to the set neighborhood radius and minimum number of points to obtain the clustering area; the proportion of the number of transactions of the overdraft account in different clustering areas is counted; and the proportion of the number of transactions in the clustering area is marked as the consumption location clustering feature; High-frequency consumption category characteristics, consumption time distribution characteristics, and consumption location clustering characteristics are marked as consumption behavior characteristics.

4. A credit card overdraft reminder notification management method according to claim 1, characterized in that: The method for obtaining the credit risk characteristics includes: Obtain the current credit score of the overdraft user through a credit rating agency; set the user's original usage score, identify the maximum and minimum credit scores of the overdraft user during the use of the credit card, and normalize the current credit score, original usage score, and the maximum and minimum credit scores of the overdraft user during the use process to obtain a credit score feature value; Obtain overdue information of overdraft accounts, including the number of overdue times, the duration of each overdue time, and the overdue amount; set a reasonable allowable value for any overdue parameter in the overdue information, and subtract the corresponding reasonable allowable value from the value of any overdue parameter in the overdue information to obtain a reasonable difference; perform weighted calculation on the reasonable differences of all overdue parameters in the overdue information to obtain an overdue severity index as a quantitative feature value of the overdue record; Obtain the credit limit adjustment record data of the overdraft account's credit card, extract any credit limit and its adjustment time from the credit limit adjustment record data; calculate the difference between adjacent adjustment times to obtain the adjacent adjustment time difference; calculate the difference between adjacent credit limits to obtain the adjacent adjustment amount difference; calculate the variance of all adjacent adjustment time differences and adjacent adjustment amount differences during the use of the credit card to obtain the adjacent adjustment time fluctuation value and adjacent adjustment amount fluctuation value; perform weighted calculation on the adjacent adjustment time fluctuation value and adjacent adjustment amount fluctuation value to obtain the credit limit change trend characteristic value; The credit score characteristic value, overdue record quantitative characteristic value, and credit limit change trend characteristic value are marked as credit risk characteristics.

5. A credit card overdraft reminder notification management method according to claim 1, characterized in that: The method for obtaining the repayment ability characteristics includes: Obtain the wage income information of the overdraft account, including wage payment time, wage income, and investment and financial management income; set the income assessment time zone, calculate the time interval between any adjacent wage payment times in the income assessment time zone and mark it as the wage payment interval; calculate the variance of the wage payment interval in the income assessment time zone to obtain the wage payment stability value; calculate the variance of the wage income in the income assessment period to obtain the income fluctuation value; calculate the standard deviation of the investment and financial management income in the income assessment time zone to obtain the investment stability value; perform weighted calculation on the wage payment stability value, income fluctuation value, and investment stability value to obtain the stable characteristic value of the income source; The debt burden ratio characteristic value is calculated by dividing the sum of various loans and credit card overdrafts of overdraft households by total assets; various loans include mortgage balances, car loan balances, and total credit card overdrafts; total assets include property value, vehicle value, deposit balances, salary income, and investment and financial management income; The characteristic values ​​of income source stability and debt burden ratio are marked as repayment ability characteristics.

6. A credit card overdraft reminder notification management method according to claim 1, characterized in that: Use the new key feature data to perform incremental learning updates on the overdraft account portrait model, specifically: A gradient-based incremental learning algorithm is used to merge the new data with part of the original training set to form a new training set. The overdraft household portrait model is retrained on the new training set using a set multiple of the original learning rate, and the parameters of all layers of the model are updated.

7. A credit card overdraft reminder notification management method according to claim 1, characterized in that: The channel preference prediction model is constructed using a naive Bayes classification algorithm or a logistic regression algorithm, with historical response data of different notification channels as a training set and overdraft customer portraits as input variables, to predict the acceptance probability and preference weight ranking of overdraft customers for different channels.

8. A credit card overdraft reminder notification management method according to claim 1, characterized in that: Input the overdraft account portraits of different dimensions into the strategy matching library to output the reminder notification strategy for the overdraft account. The specific steps are as follows: Obtain the consumption behavior characteristics, credit risk characteristics, and repayment ability characteristics of overdraft customers, and match the overdraft customer profile with the predefined strategies in the strategy matching library: The strategies in the strategy matching library are classified and stored according to credit risk level, consumption behavior pattern and repayment ability; The credit risk levels include low risk, medium risk, and high risk; the consumption behavior patterns include stable consumption, flexible consumption, and high consumption; and the repayment capacity includes strong, medium, and weak; each category includes specific reminder notification strategies for different feature combinations; When an overdraft account profile is input, the system determines the credit risk level range to which the account belongs based on the credit risk characteristics in the profile, and then further screens and matches the strategies under that level based on the consumption behavior characteristics and repayment ability characteristics; Thus, a reminder notification strategy corresponding to the overdraft account portrait is output, and the reminder notification strategy includes notification tone, reminder frequency, preferential measures and notification content customization.

9. A credit card overdraft reminder notification management method according to claim 1, characterized in that: In the present application, after step S8, step S9 is also included to establish a multi-dimensional reminder effect evaluation index system, use big data analysis technology to count and analyze various indicators within the set notification period, and generate an evaluation report within the set notification period; The indicators include notification delivery rate, reading / listening rate, response rate, and repayment conversion rate. Based on the evaluation report, a reinforcement learning algorithm is used to optimize and adjust the intelligent reminder notification strategy.

Images