Customer life cycle-based personalized repayment reminder scheme generation method and system

Abstract

The application discloses a kind of personalized repayment reminding scheme generation method and system based on customer life cycle, which comprises: obtaining customer archives, and modeling the state of customer archives using finite state machine;From state modeling, extract customer behavior trajectory to build state transition sequence, and extract the multi-dimensional features of the customer to build customer portrait;According to customer portrait, define the action space and constraint condition of decision network;According to state modeling and customer portrait, define the current state space of the customer, use the current state space as the input of the decision network, and output the repayment reminding scheme of the optimal action selection;According to the repayment reminding scheme, generate task instruction, uniformly schedule and distribute the task instruction, and update the feedback data collected during task instruction execution to the customer archives.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a method and system for generating personalized repayment reminder schemes based on the customer lifecycle. Background Technology

[0002] In internet finance, when creditors remind debtors of the need to repay non-performing assets, existing repayment reminder systems use a rule engine-based approach. By inputting the customer's overdue days and amount, the system performs rule matching using predefined if-then statements and outputs the corresponding reminder level. Based on the fixed reminder channels and frequencies set for each level, the system completes the repayment reminder for non-performing assets.

[0003] However, existing solutions still have some limitations: (1) The lack of customer profile dimensions and reliance solely on basic customer information and overdue performance make it difficult for the solution to generate accurate and effective evaluation results; (2) A single segmentation strategy cannot reflect the customer’s true capabilities and desires. The generated solution will be difficult to match the actual customer situation, resulting in serious waste of execution resources and low solution efficiency. (3) The scheme generation lacks multi-dimensional collaboration, and there may be strategic conflicts between independent decisions. Especially in the scenario of a single segmented strategy, single-dimensional optimization cannot output the optimal solution of the overall strategy, making it difficult to form a coordinated and unified scheme among the various dimensions, thus reducing the overall energy efficiency of the scheme.

[0004] It is evident that existing technologies have shortcomings that urgently need to be addressed. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a method and system for generating personalized repayment reminder schemes based on the customer lifecycle, which can improve the overall effectiveness and output efficiency of the schemes.

[0006] To address the aforementioned technical problems, the first aspect of this invention discloses a method for generating personalized repayment reminder schemes based on the customer lifecycle, the method comprising: Obtain customer profiles and use a finite state machine to model the state of the customer profiles; The customer behavior trajectory is extracted from state modeling to construct a state transition sequence, and multi-dimensional features of the customer are extracted to construct a customer profile; Define the action space and constraints of the decision-making network based on customer profiles; The current state space of a customer is defined based on state modeling and customer profile. The current state space is used as the input to the decision network to output the repayment reminder scheme with the optimal action selection. Based on the repayment reminder plan, task instructions are generated, and these instructions are uniformly scheduled and distributed. Feedback data from the execution of these task instructions is collected and updated to the customer's profile.

[0007] As an optional implementation, in the first aspect of the present invention, the finite state machine defines the state modeling as a quintuple, the quintuple including a set of states, a set of events, a transition function, an initial state, and a set of final states; The result of the state modeling is a triple, which includes a lifecycle state, the duration of the state stay, and the event type that triggers the state transition.

[0008] As an optional implementation, in the first aspect of the present invention, the feature dimensions of the customer profile include the overdue stage dimension, repayment history dimension, credit status dimension, consumption habit dimension, and social behavior dimension. The overdue stage dimension is used to describe the customer's current overdue status and historical overdue performance; The repayment history dimension is used to depict the customer's historical repayment behavior patterns; The credit status dimension is used to assess a customer's overall credit level; The consumption habit dimension is used to reflect customers' consumption behavior patterns.

[0009] As an optional implementation, in the first aspect of the present invention, the action space includes a timing dimension, a channel dimension, a frequency dimension, and a planning dimension; By defining the overall action space as the Cartesian product of four subspaces, the problem of generating repayment reminder schemes is modeled as a combinatorial optimization problem of four-dimensional action space. The decision network then outputs the optimal combination to obtain the repayment reminder scheme.

[0010] As an optional implementation, in the first aspect of the invention, defining the customer's current state space based on state modeling and customer profile includes: The discrete states modeled by state are encoded into binary vectors to obtain the lifecycle state code; Interaction record features are obtained through the historical interaction record window. The dimensions of the interaction record features include the time of contact, the channel of contact, the result of contact, and the repayment amount. The current state space is defined by vector concatenation using lifecycle state coding, customer profiles, and interaction record features, so that the solution selection is modeled as a Markov decision problem. For the Markov decision problem, a decision network is set up. The current state space is input into the trained decision network to solve for the Q-value estimation of each action. The optimal action selection is determined by maximizing the Q-value. The optimal action selection is decoded to determine the values ​​in the action space, resulting in a repayment reminder scheme.

[0011] As an optional implementation, in the first aspect of the present invention, the current state space is input into a trained decision network to solve for the Q-value estimates of each action, the optimal action selection is determined by maximizing the Q-value, and the optimal action selection is decoded to determine the values ​​of the action space, thereby obtaining a repayment reminder scheme, including: Using the given current state space as input to the decision network, the output yields determined network parameters. In the state of being Next candidate action Q-value estimation; By maximizing the Q-value, we can estimate all actions in the network at the current state and determine the optimal action that maximizes the Q-value. ; For optimal action Decoding yields action values ​​across timing, channel, frequency, and planning dimensions; The repayment reminder schemes under all Q-value estimates are ranked by confidence based on the difference between each Q-value estimate, and the repayment reminder scheme is determined by the action value of the optimal action.

[0012] As an optional implementation, in the first aspect of the present invention, the unified scheduling and distribution of task instructions includes: The repayment reminder program is broken down into execution tasks, each of which includes the target customer, reminder content, execution time, and channel parameters; Each task is sorted by priority and then enters the execution queue to await scheduling and execution; Message templates are adapted and format converted according to the selected channels, and tasks are executed according to the execution queue through the access interfaces of each channel.

[0013] As an optional implementation, in the first aspect of the invention, after the step of collecting feedback data of the task execution instructions and updating the customer profile, the method includes: An evaluation system is set up for the repayment reminder program, which includes indicators such as recovery rate, response rate, cost rate, and complaint rate. The network update is triggered by an event-driven mechanism based on the indicators in the evaluation system. When any event-driven condition is met, the collected feedback data is injected into the experience replay buffer for incremental training to update the decision network.

[0014] A second aspect of this invention discloses a personalized repayment reminder scheme generation system based on the customer lifecycle, the system comprising: The state modeling module is used to obtain customer profiles and uses a finite state machine to model the state of the customer profiles. The profile building module is used to extract customer behavior trajectories from state modeling to construct state transition sequences, and to extract multi-dimensional features of customers to build customer profiles. The action definition module is used to define the action space and constraints of the decision network based on customer profiles. The solution output module is used to define the customer's current state space based on state modeling and customer profile, and use the current state space as the input of the decision network to output the repayment reminder solution with the optimal action selection. The execution feedback module is used to generate task instructions based on the repayment reminder plan, uniformly schedule and distribute the task instructions, and collect feedback data on the execution of task instructions to update the customer profile.

[0015] As an optional implementation, in a second aspect of the invention, the finite state machine defines state modeling as a quintuple, the quintuple including a set of states, a set of events, a transition function, an initial state, and a set of final states; The result of the state modeling is a triple, which includes a lifecycle state, the duration of the state stay, and the event type that triggers the state transition.

[0016] As an optional implementation, in the second aspect of the present invention, the feature dimensions of the customer profile include the overdue stage dimension, repayment history dimension, credit status dimension, consumption habit dimension, and social behavior dimension. The overdue stage dimension is used to describe the customer's current overdue status and historical overdue performance; The repayment history dimension is used to depict the customer's historical repayment behavior patterns; The credit status dimension is used to assess a customer's overall credit level; The consumption habit dimension is used to reflect customers' consumption behavior patterns.

[0017] As an optional implementation, in the second aspect of the present invention, the action space includes a timing dimension, a channel dimension, a frequency dimension, and a planning dimension; By defining the overall action space as the Cartesian product of four subspaces, the problem of generating repayment reminder schemes is modeled as a combinatorial optimization problem of four-dimensional action space. The decision network then outputs the optimal combination to obtain the repayment reminder scheme.

[0018] As an optional implementation, in a second aspect of the invention, defining the customer's current state space based on state modeling and customer profile includes: The discrete states modeled by state are encoded into binary vectors to obtain the lifecycle state code; Interaction record features are obtained through the historical interaction record window. The dimensions of the interaction record features include the time of contact, the channel of contact, the result of contact, and the repayment amount. The current state space is defined by vector concatenation using lifecycle state coding, customer profiles, and interaction record features, so that the solution selection is modeled as a Markov decision problem. For the Markov decision problem, a decision network is set up. The current state space is input into the trained decision network to solve for the Q-value estimation of each action. The optimal action selection is determined by maximizing the Q-value. The optimal action selection is decoded to determine the values ​​in the action space, resulting in a repayment reminder scheme.

[0019] As an optional implementation, in a second aspect of the invention, the current state space is input into a trained decision network to solve for the Q-value estimates of each action. The optimal action selection is determined by maximizing the Q-value. The optimal action selection is then decoded to determine the values ​​in the action space, resulting in a repayment reminder scheme, including: Using the given current state space as input to the decision network, the output yields determined network parameters. In the state of being Next candidate action Q-value estimation; By maximizing the Q-value, we can estimate all actions in the network at the current state and determine the optimal action that maximizes the Q-value. ; For optimal action Decoding yields action values ​​across timing, channel, frequency, and planning dimensions; The repayment reminder schemes under all Q-value estimates are ranked by confidence based on the difference between each Q-value estimate, and the repayment reminder scheme is determined by the action value of the optimal action.

[0020] As an optional implementation, in the second aspect of the invention, the unified scheduling and distribution of task instructions includes: The repayment reminder program is broken down into execution tasks, each of which includes the target customer, reminder content, execution time, and channel parameters; Each task is sorted by priority and then enters the execution queue to await scheduling and execution; Message templates are adapted and format converted according to the selected channels, and tasks are executed according to the execution queue through the access interfaces of each channel.

[0021] As an optional implementation, in a second aspect of the invention, after the step of collecting feedback data of the task execution instructions and updating the customer profile, the method includes: An evaluation system is set up for the repayment reminder program, which includes indicators such as recovery rate, response rate, cost rate, and complaint rate. The network update is triggered by an event-driven mechanism based on the indicators in the evaluation system. When any event-driven condition is met, the collected feedback data is injected into the experience replay buffer for incremental training to update the decision network.

[0022] A third aspect of this invention discloses another personalized repayment reminder scheme generation system based on the customer lifecycle, the system comprising: Memory containing executable program code; A processor coupled to the memory; The processor calls the executable program code stored in the memory to execute some or all of the steps in the personalized repayment reminder scheme generation method based on customer lifecycle disclosed in the first aspect of the present invention.

[0023] The fourth aspect of the present invention discloses a computer storage medium storing computer instructions, which, when invoked, are used to execute some or all of the steps in the method for generating a personalized repayment reminder scheme based on the customer lifecycle disclosed in the first aspect of the present invention.

[0024] Compared with the prior art, the embodiments of the present invention have the following beneficial effects: This invention models customer behavior changes during repayment reminders as a series of ordered state transitions, extracts the trajectory of these state transitions, and integrates multi-dimensional customer feature information to form a dynamic and evolving customer profile. This profile accurately reflects the customer's true state changes and potential intentions within the overdue period, providing rich and accurate personalized data for subsequent intelligent decision-making. A decision network dynamically explores and optimizes within the action space defined by the customer profile, based on the current customer state. Constraints ensure that decisions always operate within compliance and business boundaries. The solution output by the decision network combines action values ​​and network parameters to achieve the optimal result. The optimal solution effectively avoids conflicts and resource consumption caused by independent channel strategies and action strategies, significantly improving the coordination and execution efficiency of the overall solution. Through a closed-loop process from data analysis to strategy execution and effect feedback, the execution layer obtains specific tasks and schedules them uniformly, ensuring orderly and conflict-free multi-channel outreach. The collected real feedback is updated in real time to the customer profile, synchronously correcting the customer profile and status. This closed-loop mechanism enables the system to continuously accumulate effective solution experience and automatically adapt to the latest customer status, thereby achieving dynamic adaptation to the performance replacement of the static system that lags behind the evolution of customer behavior. This not only improves the effectiveness of the solution but also increases the efficiency of solution output. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 This is a flowchart illustrating a method for generating a personalized repayment reminder scheme based on the customer lifecycle, as disclosed in an embodiment of the present invention.

[0027] Figure 2 This is a schematic diagram of a personalized repayment reminder scheme generation system based on the customer lifecycle disclosed in an embodiment of the present invention.

[0028] Figure 3 This is a schematic diagram of another personalized repayment reminder scheme generation system based on the customer lifecycle disclosed in an embodiment of the present invention. Detailed Implementation

[0029] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, apparatus, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0031] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0032] This invention discloses a method and system for generating personalized repayment reminder schemes based on the customer lifecycle. It models the changes in customer behavior during the repayment reminder process as a series of ordered state transitions, extracts the trajectory of these state transitions, and integrates multi-dimensional customer feature information to form a dynamic and evolvable customer profile. This profile accurately reflects the customer's true state change trends and potential intentions within the overdue period, providing rich and accurate personalized data for subsequent intelligent decision-making. A decision network dynamically explores and optimizes within the action space defined by the customer profile matching, based on the current customer state. Constraints ensure that the decision-making always operates within compliance and business boundaries. The output of the decision network... The solution integrates action values ​​and network parameters to obtain the optimal solution, effectively avoiding conflicts and resource consumption caused by independent channel strategies and action strategies. This significantly improves the overall coordination and execution efficiency of the solution. Through a closed-loop process from data analysis to strategy execution and effect feedback, the execution layer obtains specific tasks and schedules them uniformly, ensuring orderly and conflict-free multi-channel outreach. The collected real feedback is instantly updated in the customer profile, synchronously correcting customer profiles and statuses. This closed-loop mechanism allows the system to continuously accumulate effective solution experience and automatically adapt to the latest customer status, thereby achieving dynamic adaptation to the static system's performance update that lags behind customer behavior evolution. This not only improves the effectiveness of the solution but also increases the efficiency of solution output. The following sections provide detailed explanations.

[0033] Example 1 Please see Figure 1 , Figure 1 This is a flowchart illustrating a method for generating personalized repayment reminder schemes based on the customer lifecycle, as disclosed in an embodiment of the present invention. Figure 1 The described method for generating personalized repayment reminder schemes based on the customer lifecycle can be applied to data processing systems / data processing devices / data processing servers (wherein, the server includes local processing servers or cloud processing servers). For example... Figure 1 As shown, the method for generating a personalized repayment reminder scheme based on the customer lifecycle may include the following operations: 101. Obtain customer files and use a finite state machine to model the state of the customer files.

[0034] Optionally, the state model of a finite state machine is defined as a quintuple, represented as: The quintuple includes a set of states, a set of events, a transition function, an initial state, and a final state.

[0035] Optionally, the result of state modeling is a triple, represented as: ,in Indicates the customer at any time The current stage of its life cycle; This indicates the duration the customer remains in this state. This indicates the event type that triggered the most recent state transition. The triples represent the modeling results of the customer lifecycle states, which will serve as inputs for subsequent customer feature extraction steps and lifecycle state encoding in the decision network.

[0036] Optionally, the set of states can be represented as The set of states includes at least the entry stage, growth stage, maturity stage, decline stage, overdue stage, deep overdue stage, and final state; The access period refers to the period when a customer has just passed the review but has not yet engaged in any business activities or is in the initial stage of business. The growth stage indicates a period in which a customer's business volume continues to grow and their credit performance remains stable. The mature stage refers to the phase in which a customer's business volume reaches its peak and remains stable. The recession phase refers to a stage in which customer business volume continues to decline or early risk signals appear. Among them, the overdue period refers to the stage where a customer has first defaulted on a payment but has not yet reached the standard for deep default. Among them, the "deep overdue period" indicates that the customer's overdue days have exceeded the set threshold and have entered a high-risk state. Additionally, the final state indicates the completion of the customer's repayment or write-off of the transaction.

[0037] Optionally, the event set can be represented as Events may include approval. Stable business Risk signals Repayment resumption First overdue Deeply overdue Partial repayment and write-off .

[0038] Optionally, the transformation function can be expressed as ,in This is the set of actions triggered during state transitions.

[0039] Optional, Indicates the initial state. Indicates the final state.

[0040] As an example, a finite state machine defines the transition rules between different lifecycle stages of a client, where the transition paths are defined in detail: When the current status is the admission period, the input event is "approval", and the conversion condition is "successful first credit granting". Then the target status at this time is the growth period, and the corresponding execution action is to establish customer files and initialize profiles. When the current state is the growth stage, the input event is business stability, and the transition condition is continuous transaction volume growth (M months). Then the target state is the maturity stage, and the corresponding actions are to update customer level and adjust credit limit. Alternatively, if the current state is the growth stage and the input event is a risk signal, where the transition condition is that the risk score is greater than the risk threshold, then the target state at this time is the decline stage, and the corresponding actions are to mark risky customers and start monitoring. When the current state is the mature stage, the input event is a risk signal. The transition condition is that the risk score is greater than the risk threshold or overdue behavior occurs. Then the target state is the decline stage, and the corresponding actions are to reduce the amount and strengthen monitoring. When the current state is in the decline phase, the input event is repayment recovery, where the transition condition is multiple consecutive normal repayments (default is 3 periods). Then the target state at this time is the maturity phase, and the corresponding actions are to restore credit and remove monitoring. Alternatively, if the current state is the decline phase and the input event is the first overdue payment, where the transition condition is the occurrence of any overdue behavior, then the target state at this time is the overdue period, and the corresponding action is to start the repayment reminder process and record the overdue payment. When the current status is overdue, the input event is multiple overdue days. The conversion condition is that the number of overdue days is greater than the preset deep overdue days. Then the target status is deep overdue, and the corresponding actions are to upgrade the reminder level and prepare for litigation. Alternatively, if the current status is overdue, and the input event is partial repayment, where the conversion condition is that the repayment amount is greater than the minimum repayment amount, then the target status is still overdue, and the corresponding action is to update the repayment plan. Alternatively, if the current status is overdue, and the input event is repayment resumption, where the transition condition is that the repayment amount equals the total overdue amount, then the target status at this time is the mature stage, and the corresponding execution actions are ending the reminder and restoring normal status; Alternatively, if the current status is overdue / deeply overdue, the input event is write-off, and the conversion condition is manual write-off / automatic write-off, then the target status is the final state, and the corresponding execution action is to settle accounts and archive.

[0041] Optionally, the increment parameter for overdue days is set to 1 day to indicate that the customer's overdue status is updated daily. The deep overdue threshold is set to 90 days. When the customer's overdue days exceed this threshold, they enter the deep overdue period. The business stability judgment parameter is set to 3 months to assess whether the customer has moved from the growth period to the maturity period. The repayment recovery judgment parameter is set to 3 periods, indicating that customers in the decline period need to make 3 consecutive repayments to recover to the maturity period.

[0042] Furthermore, in addition to the exemplary conversion paths mentioned above, the conversion paths not shown still follow the following default rules: if a customer triggers a risk escalation event (such as first-time delinquency or deep delinquency) in the current state, the customer's state will be upgraded to the corresponding risk state; if a state recovery event is triggered but the condition threshold specified by the conversion path is not met, the customer's state will remain unchanged, and an abnormal event log will be recorded for subsequent data analysis.

[0043] In some embodiments, concurrent events exist in the conversion path, so a priority order for concurrent events is set. When a customer may trigger multiple conversion events within the same time window, they are processed according to a predefined priority order, with the priority from high to low as follows: deep overdue events, first overdue events, risk signal events, and regular business events. This priority design ensures that high-risk events are processed first.

[0044] 102. Extract customer behavior trajectories from state modeling to construct state transition sequences, and extract multi-dimensional features of customers to build customer profiles.

[0045] It should be noted that the triples corresponding to the state modeling obtained in step 101 can be used to obtain a state transition sequence representing the customer's state transition process by arranging the temporal behavior trajectory. This includes overdue status and historical overdue performance, which can serve as the data basis for multi-dimensional feature extraction, which is part of the subsequent construction of customer profiles.

[0046] Optionally, the feature dimensions of the customer profile include the overdue stage dimension, repayment history dimension, credit status dimension, consumption habit dimension, and social behavior dimension. Among them, the customer profile features of the overdue stage dimension can be extracted through the state transition sequence constructed by state modeling, and the features of other dimensions are extracted through customer profiles containing multi-source customer data.

[0047] Optionally, the overdue stage dimension describes the customer's current overdue status and historical overdue performance. The feature data source for the overdue stage dimension can be a status transition sequence or extracted from the customer profile, specifically including the overdue record table and the account status table. The overdue amount field is extracted from the current overdue balance field of the account status table. Specific features within the overdue stage dimension may include: Current overdue days: This indicates the actual number of days since the customer first defaulted. Maximum number of overdue days in history: This records the longest number of days a customer has been overdue during the observation period. Number of overdue payments in the past 6 months: This counts the number of times overdue payments have occurred within the last 6 months. : Total number of overdue payments, statistics on all historical overdue payments for the customer; The current overdue amount represents the sum of the overdue principal and interest that the customer has not yet repaid. Overdue credit ratio: The ratio of current overdue amount to credit limit; Maximum overdue period: Records the longest consecutive period of overdue payments that a customer has ever made; Standard deviation of overdue days reflects the degree of fluctuation in the number of overdue days for customers; : Number of days since the most recent overdue payment, and number of days since the last overdue payment ended at the current time; : Deep delinquency flag, a binary variable indicating whether the product is in a deep delinquency state.

[0048] Optionally, the repayment history dimension is used to characterize a customer's historical repayment behavior patterns. The data source is a transaction repayment log, and various indicators are obtained by time-series aggregation calculation of repayment records. Specific features within the repayment history dimension may include: On-time payment rate in the past 3 months: the percentage of on-time payments made in the past 3 months out of the total number of payments made. On-time payment rate in the past 6 months: the percentage of on-time payments made in the past 6 months out of the total number of payments made. On-time payment rate in the past 12 months: the percentage of on-time payments made in the past 12 months out of the total number of payments made. : Number of early repayments, total number of early repayments in history; : Prepayment rate, the proportion of prepayments to total repayments; Average number of days before due date: The average number of days before the due date when making an early repayment. : Repayment period after delinquency: The time from when the customer defaults until the actual repayment; : Repayment amount fluctuation coefficient, the ratio of the standard deviation of the repayment amount to the mean; Number of times the minimum payment has been used; the number of times the minimum payment has been used in history. Minimum repayment ratio: The percentage of minimum repayments made out of the total number of repayments. Partial repayment count: The number of times partial repayments have occurred in the past; Partial repayment ratio: the percentage of partial repayments made out of the total number of repayments. Average repayment period: the average number of days between two consecutive repayments; The coefficient of variation of repayment amount reflects the degree of dispersion of repayment amount; Number of loan extensions: The number of times a loan has been extended in the past.

[0049] Optionally, the credit status dimension is used to assess a customer's overall creditworthiness. Data sources include internal scores (de-identified and encrypted) obtained from the credit scoring interface of the credit management system within the customer's profile and external credit data queried. Features within the credit status dimension may include: Internal credit score, derived from the scoring results of the organization's credit assessment model; External credit scoring, derived from the credit score in the People's Bank of China's credit report; Credit history length: the length of time a customer first establishes a credit relationship; Debt ratio: The ratio of a customer's total liabilities to their total assets; Credit utilization rate: the ratio of currently used credit limit to total credit limit; Loan balance to income ratio: the ratio of average monthly loan balance to average monthly income; Number of outstanding loan accounts: The number of accounts currently holding outstanding loans; Number of inquiries in the past 3 months; Number of credit inquiries in the past 3 months. Outstanding balance: Total number of outstanding loans and credit card transactions. : Guarantee indicator, a binary variable representing whether a loan is secured; Overdue account percentage: The proportion of overdue accounts to total accounts in the credit report; : Follow flag, a binary variable indicating whether there are any accounts in the follow status.

[0050] Optionally, the consumption habit dimension reflects customers' consumption behavior patterns. The data source is the core transaction consumption record. Various indicators are obtained by classifying and aggregating consumption details. Features within the consumption habit dimension may include: Average monthly spending: The average total monthly spending over the past 6 months; Average monthly purchase frequency, average monthly number of purchase transactions; Food and beverage consumption percentage: The proportion of food and beverage consumption in total consumption; Shopping-related spending percentage: The proportion of shopping-related spending in total spending; Entertainment spending percentage: The proportion of entertainment spending in total spending; Online consumption share: The proportion of total consumption spent through online channels; Offline consumption share: The proportion of total consumption through offline channels; Weekday spending percentage: The proportion of weekday spending to total spending; Weekend spending percentage: The proportion of total spending to total spending; The proportion of consumption during the morning period, specifically the percentage of total consumption from 6:00 AM to 12:00 PM; Afternoon period percentage: The proportion of total consumption from 12:00 to 18:00. The percentage of spending during the evening hours, specifically the proportion of total spending from 6 PM to 10 PM; The percentage of spending during late-night hours, specifically the proportion of spending from 10 PM to 6 AM the following day; Merchant diversity index: the number of merchant categories involved in consumption; Consumer trend indicator: the trend of changes in consumption amount in the past 3 months compared to the previous 3 months; Consumer stability index: the standard deviation of monthly fluctuations in consumer spending. Large transaction count: The number of times a single transaction exceeds a set threshold; Cross-border consumption indicator, a binary variable representing whether cross-border consumption has occurred; : Percentage of recharge-related consumption; the proportion of recharge-related consumption in total consumption; Merchant concentration: Percentage of merchants with the highest spending.

[0051] Optionally, the social behavior dimension is used to assess customer contact response characteristics. Data sources include contact center verification records and push notification logs in the customer profile. Contact channel interaction data comes from the contact center system, and consumption behavior data comes from push notification platform logs. Features within the social behavior dimension may include: : Mobile number validity, a binary variable representing whether the registered mobile number is valid; Email validity: A binary variable indicating whether the registered email address is valid. Historical response rate: the percentage of customers who responded to historical repayment reminder tasks; : Call answer rate, the percentage of customers who answer historical payment reminder calls; SMS open rate: The percentage of customers who open and read historical SMS repayment reminders; : App push notification open rate, the percentage of customers who open historical app push notifications; Social media activity: a metric indicating the frequency of customer activity on social media platforms; Completeness of emergency contact information: The degree to which emergency contact information is filled out completely.

[0052] In some embodiments, to eliminate discrepancies between different features in a customer's multidimensional features, a Z-score normalization method is used to convert the original values ​​into a standard normal distribution with a mean of 0 and a standard deviation of 1. This normalization process can be expressed as: in, This represents the mean of the corresponding feature across all samples. The standard deviation is the observation window length used to calculate the mean and standard deviation of each feature. The statistical window for multidimensional features is usually set to 90 days.

[0053] In some embodiments, for categorical feature encoding, a One-Hot encoding method is used to convert the categorical features into numerical vectors, where the target number of categorical feature encoding categories is set to be . If the actual number of categories exceeds the threshold, the target encoding method is adopted, such as the category number threshold. Set as When the actual number of categories for a categorical feature exceeds When that happens, it is automatically converted into target encoding based on the target variable.

[0054] In some embodiments, for time-series features, a sliding window statistical method is used to extract features. Within the sliding window, statistical measures such as mean, maximum, minimum, count, and standard deviation are calculated, such as the time-series feature window. Set as The time frame is used to calculate the sliding window length of time series characteristic statistics. In some special embodiments, if the time series data of a new customer is less than one statistical window, the statistics of the same customer group are used to replace it.

[0055] In some embodiments, missing values ​​are handled as follows: For numerical features, if missing values ​​exist, they are filled using the median of customers in the same customer group; for categorical features, if missing values ​​exist, they are filled with the "unknown" category; for key business fields, if the number of missing fields exceeds a set threshold, the customer's profile update process is suspended. For example, the median for missing value filling uses 90-day window data from the same customer group. Key field missing value threshold. Set as When the number of missing key fields in a single record exceeds The process will be suspended at any time.

[0056] In some embodiments, outlier handling is performed in the following manner: Truncating rules handle outliers and will exceed the limit. The range of values ​​is truncated to boundary values, where outlier determination parameters... Set as , that is to Double the standard deviation is used as the boundary for outlier detection.

[0057] As an optional embodiment, the step of building a customer profile in the above steps includes: By vectorizing the multi-dimensional features corresponding to the overdue stage, repayment history, credit status, consumption habits, and social behavior, a customer profile is obtained. Furthermore, a weighted calculation is performed on all multidimensional features to obtain a comprehensive score for the customer profile; The vector concatenation is represented as follows: in, This represents the feature vector corresponding to the overdue stage dimension. This represents the feature vector corresponding to the repayment history dimension. This represents the feature vector corresponding to the credit status dimension. This represents the feature vector corresponding to the dimension of consumption habits. This represents the feature vector corresponding to the social behavior dimension. This represents a vector concatenation operation. This indicates that the customer profile comprises a total of 65 dimensions.

[0058] The overall score is expressed as follows: It is used to reflect the overall credit level and repayment ability of a customer.

[0059] It should be noted that the customer profile and comprehensive score obtained in step 102 can be used as the input basis for determining the action space constraints in the subsequent step 103. At the same time, together with the triples corresponding to the state model output in step 101, they serve as components of the decision network state vector in step 104.

[0060] 103. Define the action space and constraints of the decision network based on customer profiles.

[0061] Optionally, the action space includes timing, channel, frequency, and planning dimensions. By defining the overall action space as the Cartesian product of the four subspaces, the problem of generating a repayment reminder scheme is modeled as a combinatorial optimization problem of a four-dimensional action space. The decision network then outputs the optimal combination to obtain the repayment reminder scheme.

[0062] Optional, timing dimension definition Several optional time windows, each representing a specific time period within a day (the day can be divided into multiple 3-hour periods, providing execution options for different time periods), for example, the number of timing candidates. Set as This corresponds to eight time windows throughout the day: 00:00-03:00, 03:00-06:00, 06:00-09:00, 09:00-12:00, 12:00-15:00, 15:00-18:00, 18:00-21:00, and 21:00-24:00.

[0063] Optional, channel dimension Define the set of optional contact channels, which may include automated voice calls, human calls, SMS, app push notifications, emails, and WeChat, for example, the number of candidate channels for repayment reminders. Set as These are phone calls, text messages, app push notifications, emails, and WeChat.

[0064] Optional, frequency dimension The option to define the frequency of repayment reminder actions can include frequencies such as daily, every 2 days, weekly, and every 2 weeks, for example, by listing the number of frequency candidates. Set as The corresponding times are daily, every 2 days, weekly, and every 2 weeks.

[0065] Optional, planning dimension Define a set of optional installment repayment plans, which can include installment periods of 1, 2, 3, 6, 9, 12, etc., such as the maximum number of installments. Set as There are 6 installment options: 1, 2, 3, 6, 9, and 12 installments.

[0066] Optionally, the total action space across timing, channel, frequency, and planning dimensions can be represented as the total number of candidate solutions generated by a four-dimensional Cartesian product, such as... Of these, approximately 768 effective methods can be retained after filtering by constraints.

[0067] Optionally, the constraints in the action space include business constraints and compliance constraints, which are used to filter action scheme combinations that do not meet the conditions.

[0068] Optionally, business constraints may include a maximum number of installments that deeply delinquent customers can choose. Set the number of periods, such as a maximum of 6 periods; scoring. Higher than For customers with a credit history, set an upper limit for the frequency of repayment reminders, such as every 2 weeks; for customers with a history of default, set a lower limit for the frequency of repayment reminders, such as daily.

[0069] Optionally, compliance constraints may include setting a daily limit on the number of reminders sent to a single customer, such as 3 times per reminder; setting a minimum time interval between two contacts with the same customer, such as a 2-hour interval; and setting a silent period for channel reminders, during which actions are prohibited, such as setting... By the next day This is a quiet period.

[0070] Understandably, the values ​​of the parameters in the constraints can be adjusted using the corresponding feature values ​​of each dimension in the customer profile to align with the subsequent solution generation for the current customer.

[0071] 104. Define the customer's current state space based on state modeling and customer profile, use the current state space as input to the decision network, and output the repayment reminder scheme with the optimal action selection.

[0072] As an optional embodiment, the step above, defining the customer's current state space based on state modeling and customer profile, includes: The discrete states modeled by state are encoded into binary vectors to obtain the lifecycle state code; Interaction record features are obtained through the historical interaction record window. The dimensions of the interaction record features include the time of contact, the channel of contact, the result of contact, and the repayment amount. The current state space is defined by vector concatenation using lifecycle state coding, customer profiles, and interaction record features, so that the solution selection is modeled as a Markov decision problem. For the Markov decision problem, a decision network is set up. The current state space is input into the trained decision network to solve for the Q-value estimation of each action. The optimal action selection is determined by maximizing the Q-value. The optimal action selection is decoded to determine the values ​​in the action space, resulting in a repayment reminder scheme.

[0073] Optionally, the customer profile includes 65 features across five dimensions: delinquency stage, repayment history, credit status, consumption habits, and social behavior.

[0074] Optionally, the lifecycle state encoding can adopt a One-Hot encoding method, where the 7 discrete states correspond to a 7-dimensional binary vector, and the dimension of the lifecycle state encoding is set to 7. Optionally, the interaction record features are obtained from the customer's most recent H interactions. Each interaction includes information in four dimensions: contact time, contact channel, contact result, and repayment amount. Typically, the historical interaction record window H is set to 5 times, and the dimension of each interaction feature is set to 4 dimensions. Therefore, the total vector dimension of the interaction record features is 20 dimensions.

[0075] As an example, the current state space can be represented as: in, This represents the current state space of the concatenated vectors. A vector representing a customer profile. This represents the lifecycle state code corresponding to state modeling. Indicates interactive record characteristics, This represents the total dimension of the current state space.

[0076] Optionally, the decision network can be a DQN (Deep-Q Network), including an evaluation network. and target network The architecture employs a dual-network structure, where the evaluation network learns the output decision scheme in the current state space, and the target network provides the temporal difference target value and periodically synchronizes parameters from the evaluation network. The weights in the network parameters are initialized using the Xavier method, and the learning rate is... Discount factor .

[0077] Optionally, the decision network may include an input layer, hidden layers, and an output layer. The input layer receives a 92-dimensional current state space as input. The hidden layers are three fully connected layers with 256, 128, and 64 neurons in each layer, respectively, and the activation function for each layer is ReLU. The output layer outputs... A dimensional Q-value vector, where each dimension corresponds to a Q-value estimate for an action plan.

[0078] Optionally, the decision network can be trained using an ε-greedy exploration strategy to balance exploration and exploitation, where the ε value decays linearly from the initial value to the final value, which can be expressed as: In the formula, This represents the value of ε as a function of the exploration time step t. Indicates the final value exploration rate. The initial exploration rate is represented by t, and the exploration time step is represented by t. Indicates the number of decay steps; Subsequently, the agent's interaction experience is stored in the revisit buffer through the experience replay mechanism, and gradient updates are calculated by sampling batch_size samples during training. Finally, the target network parameters in the decision network are periodically soft-updated from the evaluation network, which can be represented as: In the formula, Indicates the target network. Indicates the evaluation network, This indicates the target network soft update parameters.

[0079] As an example, the exploration strategy parameter is set to: initial exploration rate. Final exploration rate decay steps .

[0080] As an example, the experience playback parameter is set to: buffer capacity (100,000 records), batch_size Preheating threshold (1000 entries) Target network soft update parameters .

[0081] As an optional embodiment, in the above steps, the current state space is input into the trained decision network to solve for the Q-value estimate of each action, the optimal action selection is determined by maximizing the Q-value, and the optimal action selection is decoded to determine the value of the action space, thus obtaining the repayment reminder scheme, including: Using the given current state space as input to the decision network, the output yields determined network parameters. In the state of being Next candidate action Q-value estimation; By maximizing the Q-value, we can estimate all actions in the network at the current state and determine the optimal action that maximizes the Q-value. ; For optimal action Decoding yields action values ​​across timing, channel, frequency, and planning dimensions; The repayment reminder schemes under all Q-value estimates are ranked by confidence based on the difference between each Q-value estimate, and the repayment reminder scheme is determined by the action value of the optimal action.

[0082] As an example, the Q-value estimate can be expressed as: Where Q represents the Q-value estimate, Indicates the state One of the candidate actions that can be executed. This represents all the weights and bias parameters in the decision network, which determine how the network maps the current state space of the input to a Q-value estimate.

[0083] As an example, the solution for the optimal action can be expressed as: in, Indicates candidate actions The optimal action in the process, This represents the operator that maximizes the Q-value corresponding to the candidate action.

[0084] In some embodiments, diversity processing is employed when Q values ​​are similar, such as when the difference in Q values ​​between the optimal action and the second-best action is less than a threshold. At this time, the system employs a top-k random perturbation mechanism to increase policy diversity, randomly selecting from the top-k candidate actions with a certain probability. The Q-value difference threshold is used as an example. Set as Probability of diversity disturbances Number of candidates .

[0085] In some embodiments, Mahalanobis distance is used for out-of-distribution detection to detect abnormal states. When the Mahalanobis distance between the current state and the training set states exceeds a set threshold, it is determined to be an abnormal state. For the handling of abnormal states, a predefined baseline repayment reminder strategy is invoked to ensure that there is still a strategy to be executed in abnormal situations. The baseline strategy is a simple rule matching based on the number of overdue days and the amount.

[0086] 105. Generate task instructions based on the repayment reminder plan, uniformly schedule and distribute the task instructions, and collect feedback data on the execution of task instructions to update the customer file.

[0087] Optionally, task instructions include task decomposition, queue management, channel adaptation, and execution.

[0088] In an optional embodiment, the step of uniformly scheduling and distributing task instructions in the above steps includes: The repayment reminder program is broken down into execution tasks, each of which includes the target customer, reminder content, execution time, and channel parameters; Each task is sorted by priority and then enters the execution queue to await scheduling and execution; Message templates are adapted and format converted according to the selected channels, and tasks are executed according to the execution queue through the access interfaces of each channel.

[0089] In some embodiments, task scheduling period parameters Set as Hours; Maximum capacity of the execution queue Set as Item; Task execution timeout window Set as Hours, feedback collection cycle Set as Hours, task execution timeout determination time Set as Hour.

[0090] Optionally, feedback data may include outreach feedback, intent feedback, and outcome feedback. Outreach feedback records whether the action successfully reached the customer, including fields such as outreach status code, outreach time, and connection duration. Intent feedback records the customer's willingness to respond to the action, including intent tags such as promising repayment, explicitly refusing, and negotiating an extension. Outcome feedback records the actual result of the action, including fields such as actual repayment amount, partial repayment amount, and repayment time.

[0091] Optionally, the feedback data adopts a structured record format, including the following key fields: task_id, result_code, intent_label, repayment_amount, and timestamp.

[0092] In an optional embodiment, after updating the customer profile with feedback data from the collected task execution instructions in the above steps, the process includes: An evaluation system is set up for the repayment reminder program, which includes indicators such as recovery rate, response rate, cost rate, and complaint rate. The network update is triggered by an event-driven mechanism based on the indicators in the evaluation system. When any event-driven condition is met, the collected feedback data is injected into the experience replay buffer for incremental training to update the decision network.

[0093] Optional, recovery rate metrics The effectiveness of a repayment reminder program is measured by calculating the ratio of successfully recovered funds to the target total overdue amount, which can be expressed as: Optional, response rate metric The effectiveness of the repayment reminder program is measured by the ratio of effective interaction tasks to the total number of tasks, which can be expressed as: Optional, cost ratio metric The operational efficiency of a repayment reminder program is measured by calculating the ratio of total cost to total amount recovered, which can be expressed as: Optional, complaint rate metric The ratio of complaint assignments to total assignments used to measure the compliance and customer satisfaction of a repayment reminder program is determined by calculating the ratio of such assignments to total assignments, and can be expressed as: In some embodiments, the evaluation period of the evaluation system is set to 7 days to calculate the statistical window length of each indicator, and the rolling window is set to 30 days to calculate the rolling average indicator.

[0094] Optionally, the triggering conditions for event-driven events can be: The first trigger condition is when the recovery rate metric is continuously... The number of days below the update threshold When this happens, the model is updated. The second trigger condition is when the number of newly added feedback samples reaches [a certain threshold]. When a new sample triggers a model update, a new sample triggers a threshold. Set as strip.

[0095] The third trigger condition is that the administrator can trigger the model update process through the management interface according to business needs.

[0096] In one alternative embodiment, the network model update mechanism includes new sample injection, incremental training, version management, and smooth transition, wherein: The new sample injection step injects newly collected feedback samples into the experience replay buffer. When the buffer capacity exceeds the limit, the oldest sample is eliminated using a FIFO strategy. The incremental training step employs a warm-restart strategy, starting from the current model parameters and updating the gradients several times using new samples. A small learning rate is used during training to avoid excessive disruption of existing knowledge. The version management process generates a new model version number with each update, retains snapshots of parameters from historical versions, and supports rolling back to historical versions when the new version is not performing well. The smooth transition process employs a strategy smoothing mechanism to achieve a gradual switch between new and old strategies, avoiding the impact of sudden strategy changes on business operations.

[0097] In some embodiments, to ensure training stability, the trend of loss function changes is monitored in real time during incremental training. If the loss divergence exceeds a set threshold, the update process is terminated and the system reverts to the state before the update. The loss divergence threshold is set to... The threshold for unstable epoch count is set to... The learning rate degradation factor is set to ; .

[0098] Furthermore, a rollback mechanism is also set up. When various business indicators deteriorate significantly after the new strategy is launched, the system will automatically or manually trigger a rollback to restore the effective model parameters of the previous version.

[0099] It should be noted that after updating the network parameters in step 105, the new network parameters will be... The new parameters are written into the decision network to replace the original network parameters. In the next round of repayment reminder scheme decision-making, the network with updated parameters will be used for inference. For example, the new parameters are written back into the decision network (DQN network) → the new parameters are used in the next round of inference → the new scheme is executed → feedback data is collected → the effect is evaluated → the update is triggered → the new parameters are written back into the decision network (DQN network) again, thus forming a complete closed-loop optimization process.

[0100] As can be seen, the above-described embodiments of the invention model customer behavior changes during repayment reminders as a series of ordered state transitions, extract the trajectory of these state transitions, and integrate multi-dimensional customer feature information to form a dynamic and evolvable customer profile. This accurately reflects the customer's true state change trends and potential intentions within the overdue period, providing rich and accurate personalized data for subsequent intelligent decision-making. The decision network dynamically explores and optimizes within the action space defined by customer profile matching, based on the current customer state. Constraints ensure that decisions always operate within compliance and business boundaries. The solution output by the decision network integrates action values ​​and network parameters. By obtaining the optimal solution, the system effectively avoids conflicts and resource consumption caused by the independence of channel strategies and action strategies, significantly improving the coordination and execution efficiency of the overall solution. Through a closed-loop process from data analysis to strategy execution and effect feedback, the execution layer obtains specific tasks and schedules them in a unified manner to ensure orderly and conflict-free multi-channel outreach. The collected real feedback is updated in real time to the customer profile and synchronize the correction of customer profile and status. This closed-loop mechanism enables the system to continuously accumulate effective solution experience and automatically adapt to the latest customer status, thereby achieving dynamic adaptation to the performance replacement of the static system that lags behind the evolution of customer behavior. This not only improves the effectiveness of the solution but also increases the efficiency of solution output.

[0101] Example 2 Please see Figure 2 , Figure 2 This is a schematic diagram of a personalized repayment reminder scheme generation system based on the customer lifecycle disclosed in an embodiment of the present invention. Figure 2 The described personalized repayment reminder system based on the customer lifecycle can be applied to data processing systems / data processing devices / data processing servers (including local processing servers or cloud processing servers). For example... Figure 2 As shown, the personalized repayment reminder system based on the customer lifecycle can include: State modeling module 201 is used to obtain customer files and use a finite state machine to model the state of the customer files. The profile building module 202 is used to extract customer behavior trajectories from state modeling to construct state transition sequences, and to extract multi-dimensional features of customers to build customer profiles. Action definition module 203 is used to define the action space and constraints of the decision network based on the customer profile; The solution output module 204 is used to define the current state space of the customer based on state modeling and customer profile, and to use the current state space as the input of the decision network to output the repayment reminder solution with the optimal action selection. The execution feedback module 205 is used to generate task instructions based on the repayment reminder plan, uniformly schedule and distribute the task instructions, and collect feedback data on the execution of task instructions to update the customer profile.

[0102] As an optional embodiment, the finite state machine models the state as a quintuple, which includes a set of states, a set of events, a transition function, an initial state, and a set of final states. The result of the state modeling is a triple, which includes a lifecycle state, the duration of the state stay, and the event type that triggers the state transition.

[0103] As an optional embodiment, the feature dimensions of the customer profile include the overdue stage dimension, repayment history dimension, credit status dimension, consumption habits dimension, and social behavior dimension; The overdue stage dimension is used to describe the customer's current overdue status and historical overdue performance; The repayment history dimension is used to depict the customer's historical repayment behavior patterns; The credit status dimension is used to assess a customer's overall credit level; The consumption habit dimension is used to reflect customers' consumption behavior patterns.

[0104] As an optional embodiment, the action space includes timing dimension, channel dimension, frequency dimension, and planning dimension; By defining the overall action space as the Cartesian product of four subspaces, the problem of generating repayment reminder schemes is modeled as a combinatorial optimization problem of four-dimensional action space. The decision network then outputs the optimal combination to obtain the repayment reminder scheme.

[0105] As an optional implementation, the customer's current state space is defined based on state modeling and customer profiles, including: The discrete states modeled by state are encoded into binary vectors to obtain the lifecycle state code; Interaction record features are obtained through the historical interaction record window. The dimensions of the interaction record features include the time of contact, the channel of contact, the result of contact, and the repayment amount. The current state space is defined by vector concatenation using lifecycle state coding, customer profiles, and interaction record features, so that the solution selection is modeled as a Markov decision problem. For the Markov decision problem, a decision network is set up. The current state space is input into the trained decision network to solve for the Q-value estimation of each action. The optimal action selection is determined by maximizing the Q-value. The optimal action selection is decoded to determine the values ​​in the action space, resulting in a repayment reminder scheme.

[0106] As an optional embodiment, the current state space is input into a trained decision network to solve for the Q-value estimate of each action. The optimal action selection is determined by maximizing the Q-value. The optimal action selection is then decoded to determine the values ​​in the action space, resulting in a repayment reminder scheme, including: Using the given current state space as input to the decision network, the output yields determined network parameters. In the state of being Next candidate action Q-value estimation; By maximizing the Q-value, we can estimate all actions in the network at the current state and determine the optimal action that maximizes the Q-value. ; For optimal action Decoding yields action values ​​across timing, channel, frequency, and planning dimensions; The repayment reminder schemes under all Q-value estimates are ranked by confidence based on the difference between each Q-value estimate, and the repayment reminder scheme is determined by the action value of the optimal action.

[0107] As an optional embodiment, the task instructions are uniformly scheduled and distributed, including: The repayment reminder program is broken down into execution tasks, each of which includes the target customer, reminder content, execution time, and channel parameters; Each task is sorted by priority and then enters the execution queue to await scheduling and execution; Message templates are adapted and format converted according to the selected channels, and tasks are executed according to the execution queue through the access interfaces of each channel.

[0108] As an optional embodiment, the system is also used to perform the following steps: After the step of collecting feedback data from task execution instructions and updating the customer profile, the following steps are included: An evaluation system is set up for the repayment reminder program, which includes indicators such as recovery rate, response rate, cost rate, and complaint rate. The network update is triggered by an event-driven mechanism based on the indicators in the evaluation system. When any event-driven condition is met, the collected feedback data is injected into the experience replay buffer for incremental training to update the decision network.

[0109] Example 3 Please see Figure 3 , Figure 3 This is another personalized repayment reminder scheme generation system based on the customer lifecycle disclosed in the embodiments of the present invention. Figure 3 The described personalized repayment reminder system based on the customer lifecycle is applied in a data processing system / data processing equipment / data processing server (wherein, the server includes a local processing server or a cloud processing server). For example... Figure 3 As shown, the personalized repayment reminder system based on the customer lifecycle can include: Memory 301 storing executable program code; Processor 302 coupled to memory 301; The processor 302 calls the executable program code stored in the memory 301 to execute the steps of the personalized repayment reminder scheme generation method based on customer lifecycle described in Embodiment 1.

[0110] Example 4 This invention discloses a computer read storage medium that stores a computer program for electronic data interchange, wherein the computer program causes a computer to execute the steps of the method for generating a personalized repayment reminder scheme based on the customer lifecycle described in Embodiment 1.

[0111] Example 5 This invention discloses a computer program product, which includes a non-transitory computer-readable storage medium storing a computer program, and the computer program is operable to cause a computer to perform the steps of the method for generating a personalized repayment reminder scheme based on the customer lifecycle described in Embodiment 1.

[0112] The foregoing has described specific embodiments of this specification; other embodiments are within the scope of the appended claims. In some cases, the actions or steps described in the claims may be performed in a different order than those shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily have to follow the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0113] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, a computer can be, for example, a personal computer, laptop computer, cellular phone, camera phone, smartphone, personal digital assistant, media player, navigation device, email device, game console, tablet computer, wearable device, or any combination of these devices.

[0114] For ease of description, the above devices are described in terms of function, divided into various units. Of course, in implementing this specification, the functions of each unit can be implemented in one or more software and / or hardware components.

[0115] Those skilled in the art will understand that the embodiments of this specification can be provided as methods, systems, or computer program products. Therefore, the embodiments of this specification can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the embodiments of this specification can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0116] This specification is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this specification. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0117] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0118] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0119] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0120] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0121] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0122] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0123] This specification can be described in the general context of computer-executable instructions that are executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. This specification can also be practiced in distributed computing environments, where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0124] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0125] Finally, it should be noted that the personalized repayment reminder scheme generation method and system based on customer lifecycle disclosed in the embodiments of the present invention are merely preferred embodiments of the present invention, and are only used to illustrate the technical solutions of the present invention, not to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for generating a personalized repayment reminder scheme based on the customer lifecycle, characterized in that, The method includes: Obtain customer profiles and use a finite state machine to model the state of the customer profiles; The customer behavior trajectory is extracted from state modeling to construct a state transition sequence, and multi-dimensional features of the customer are extracted to construct a customer profile; Define the action space and constraints of the decision-making network based on customer profiles; The current state space of a customer is defined based on state modeling and customer profile. The current state space is used as the input to the decision network to output the repayment reminder scheme with the optimal action selection. Based on the repayment reminder plan, task instructions are generated, and these instructions are uniformly scheduled and distributed. Feedback data from the execution of these task instructions is collected and updated to the customer's profile.

2. The method for generating a personalized repayment reminder scheme based on the customer lifecycle as described in claim 1, characterized in that, The finite state machine models the state as a quintuple, which includes a set of states, a set of events, a transition function, an initial state, and a set of final states. The result of the state modeling is a triple, which includes a lifecycle state, the duration of the state stay, and the event type that triggers the state transition.

3. The method for generating a personalized repayment reminder scheme based on the customer lifecycle as described in claim 1, characterized in that, The customer profile features include overdue period, repayment history, credit status, consumption habits, and social behavior. The overdue stage dimension is used to describe the customer's current overdue status and historical overdue performance; The repayment history dimension is used to depict the customer's historical repayment behavior patterns; The credit status dimension is used to assess a customer's overall credit level; The consumption habit dimension is used to reflect customers' consumption behavior patterns.

4. The method for generating a personalized repayment reminder scheme based on the customer lifecycle as described in claim 1, characterized in that, The action space includes the timing dimension, channel dimension, frequency dimension, and planning dimension; By defining the overall action space as the Cartesian product of four subspaces, the problem of generating repayment reminder schemes is modeled as a combinatorial optimization problem of four-dimensional action space. The decision network then outputs the optimal combination to obtain the repayment reminder scheme.

5. The method for generating a personalized repayment reminder scheme based on the customer lifecycle as described in claim 1, characterized in that, Define the customer's current state space based on state modeling and customer profiles, including: The discrete states modeled by state are encoded into binary vectors to obtain the lifecycle state code; Interaction record features are obtained through the historical interaction record window. The dimensions of the interaction record features include the time of contact, the channel of contact, the result of contact, and the repayment amount. The current state space is defined by vector concatenation using lifecycle state coding, customer profiles, and interaction record features, so that the solution selection is modeled as a Markov decision problem. For the Markov decision problem, a decision network is set up. The current state space is input into the trained decision network to solve for the Q-value estimation of each action. The optimal action selection is determined by maximizing the Q-value. The optimal action selection is decoded to determine the values ​​in the action space, resulting in a repayment reminder scheme.

6. The method for generating a personalized repayment reminder scheme based on the customer lifecycle as described in claim 5, characterized in that, The current state space is input into the trained decision network to estimate the Q-value of each action. The optimal action selection is determined by maximizing the Q-value. The optimal action selection is then decoded to determine the values ​​in the action space, resulting in a repayment reminder scheme, including: Using the given current state space as input to the decision network, the output yields determined network parameters. In the state of being Next candidate action Q-value estimation; By maximizing the Q-value, we can estimate all actions in the network at the current state and determine the optimal action that maximizes the Q-value. ; For optimal action Decoding yields action values ​​across timing, channel, frequency, and planning dimensions; The repayment reminder schemes under all Q-value estimates are ranked by confidence based on the difference between each Q-value estimate, and the repayment reminder scheme is determined by the action value of the optimal action.

7. The method for generating a personalized repayment reminder scheme based on the customer lifecycle as described in claim 1, characterized in that, Unified scheduling and distribution of task instructions, including: The repayment reminder program is broken down into execution tasks, each of which includes the target customer, reminder content, execution time, and channel parameters; Each task is sorted by priority and then enters the execution queue to await scheduling and execution; Message templates are adapted and format converted according to the selected channels, and tasks are executed according to the execution queue through the access interfaces of each channel.

8. The method for generating a personalized repayment reminder scheme based on the customer lifecycle as described in claim 1, characterized in that, After the step of collecting feedback data from task execution instructions and updating the customer profile, the following steps are included: An evaluation system is set up for the repayment reminder program, which includes indicators such as recovery rate, response rate, cost rate, and complaint rate. The network update is triggered by an event-driven mechanism based on the indicators in the evaluation system. When any event-driven condition is met, the collected feedback data is injected into the experience replay buffer for incremental training to update the decision network.

9. A personalized repayment reminder scheme generation system based on the customer lifecycle, characterized in that, The system includes: The state modeling module is used to obtain customer profiles and uses a finite state machine to model the state of the customer profiles. The profile building module is used to extract customer behavior trajectories from state modeling to construct state transition sequences, and to extract multi-dimensional features of customers to build customer profiles. The action definition module is used to define the action space and constraints of the decision network based on customer profiles. The solution output module is used to define the customer's current state space based on state modeling and customer profile, and use the current state space as the input of the decision network to output the repayment reminder solution with the optimal action selection. The execution feedback module is used to generate task instructions based on the repayment reminder plan, uniformly schedule and distribute the task instructions, and collect feedback data on the execution of task instructions to update the customer profile.

10. A personalized repayment reminder scheme generation system based on the customer lifecycle, characterized in that, The system includes: Memory containing executable program code; A processor coupled to the memory; The processor calls the executable program code stored in the memory to execute the personalized repayment reminder scheme generation method based on the customer lifecycle as described in any one of claims 1-8.

Images