Method and device for determining incentive strategy for issuing service, and electronic equipment

Abstract

The invention discloses an incentive strategy determination method and device for issuing business and electronic equipment, and relates to the field of artificial intelligence or other related fields, and the incentive strategy determination method comprises the steps: collecting issuing business data of a target enterprise under the condition of obtaining the authorization of the target enterprise, and determining the issuing business of the target enterprise based on the issuing business data; and determining cycle characteristics corresponding to a plurality of cycles by adopting multi-scale sliding window convolution, determining an enterprise loss score of the target enterprise and a personal loss score of the personal customer based on the dispatched business data and the cycle characteristics corresponding to different cycles, and determining a target incentive strategy by adopting a preset learning model based on the enterprise loss score and the personal loss score. The technical problem that the customer retention rate is low due to the fact that an effective incentive strategy cannot be determined in the prior art is solved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence, and more specifically, to a method, apparatus, and electronic device for determining incentive strategies for payroll services. Background Technology

[0002] In bank payroll processing, retaining corporate clients and accumulating funds with individual clients are crucial. On one hand, retaining corporate clients directly impacts the size of corporate deposits and payroll processing revenue, influencing the long-term partnership between banks and businesses. On the other hand, if individual clients immediately transfer their salaries out of the bank's account after receipt, it hinders the accumulation of retail financial business, reducing the potential for conversion into wealth management, deposits, and other value-added services. Therefore, effectively identifying and intervening in the risks of corporate payroll outflows and individual fund outflows is key to enhancing a bank's competitiveness.

[0003] Currently, the following problems exist in risk monitoring and intervention of payroll services: (1) The models used rely on manually constructed static features (such as average monthly payroll amount and number of payroll recipients), which are difficult to depict the dynamic patterns under multi-cycle models; (2) The original payroll transaction data has high-dimensional sparsity, making direct modeling inefficient and difficult to extract cross-cycle behavioral patterns; (3) Currently, most focus on a single account, ignoring the cross-channel transfer path of funds, making it difficult to locate key transfer nodes and causing risk attribution bias; (4) Intervention strategies generally adopt fixed thresholds, lacking real-time response to market and customer feedback, and are prone to lag or over-marketing.

[0004] In related technologies, LSTM models are used to construct corporate payment behavior, or graph neural networks are used to characterize account relationships. However, LSTM models fail to integrate the characteristics of fund interaction between enterprises, and graph neural networks ignore the periodic spatiotemporal structure of payroll settlement. Furthermore, strategy optimization and early warning are disconnected, failing to form an effective closed loop. This makes it impossible to determine effective incentive strategies to improve fund retention and customer value conversion capabilities.

[0005] There is currently no effective solution to the above problems. Summary of the Invention

[0006] This invention provides a method, apparatus, and electronic device for determining incentive strategies for payroll services, to at least solve the technical problem in related technologies where the inability to determine effective incentive strategies leads to low customer retention rates.

[0007] According to one aspect of the present invention, a method for determining an incentive strategy for payroll processing services is provided, comprising: collecting payroll processing data of the target enterprise after obtaining authorization from the target enterprise, wherein the payroll processing data includes at least: enterprise data, individual customer data, and payroll fund transaction data, wherein the individual customers are customers of the target enterprise who receive payroll through financial institutions; determining periodic features corresponding to multiple periods based on the payroll processing data using multi-scale sliding window convolution, wherein the interval length of different periods is inconsistent; determining the enterprise churn score of the target enterprise and the individual churn score of individual customers based on the payroll processing data and the periodic features corresponding to different periods; and determining a target incentive strategy based on the enterprise churn score and the individual churn score using a preset learning model, wherein the target incentive strategy is used to incentivize the target enterprise and all individual customers.

[0008] Furthermore, based on the agency delivery business data, the step of determining the periodic features corresponding to multiple periods using multi-scale sliding window convolution includes: extracting features from the agency delivery business data to obtain target tensor data, wherein the target tensor data includes: a time series, and each agency delivery time in the time series corresponds to agency delivery business features; based on the target tensor data, determining the periodic features corresponding to each period using multi-scale sliding window convolution, wherein the period includes: a first period, a second period, and a third period, the period length of the first period is less than the period length of the second period, the period length of the second period is less than the period length of the third period, and the periodic features include at least: mean, difference, and ratio.

[0009] Further, the step of extracting features from the payroll data to obtain target tensor data includes: extracting features from the payroll data to obtain three-dimensional tensor data, wherein the three-dimensional tensor data includes: time dimension data, enterprise dimension data, and payroll business feature dimension data. The time dimension data includes a time index generated by discretizing the timestamps in the payroll data. The enterprise dimension data includes a time series associated with the payroll accounts of the target enterprise. The payroll business feature dimension data includes statistical features and structural features. The three-dimensional tensor data is compressed using a preset encoder to obtain target tensor data, wherein the preset encoder is an encoder trained by compressing and decompressing tensor data. The compressed tensor data is data compressed by the preset encoder, and the decompressed tensor data is data decoded by the decoder.

[0010] Furthermore, the step of determining the periodic features corresponding to each period based on the target tensor data using multi-scale sliding window convolution includes: processing the target tensor data to obtain periodic data under different periods; and for each period, processing the periodic data using a preset convolution function to obtain periodic features.

[0011] Furthermore, based on payroll data and the periodic characteristics corresponding to different periods, the steps for determining the corporate churn score of the target enterprise and the individual churn score of individual customers include: determining the first probability distribution of payroll funds based on payroll data, where payroll funds refer to funds paid by the enterprise to individual customers through financial institutions; determining the second probability distribution of account funds based on payroll data, where account funds refer to funds stored in different accounts by individual customers after receiving payroll funds; determining the corporate distribution entropy based on the first probability distribution and the individual distribution entropy based on the second probability distribution; determining the corporate churn score using a preset multi-window feature fusion function based on the corporate distribution entropy and the periodic characteristics corresponding to different periods; and determining the individual churn score using a preset multi-window feature fusion function based on the individual distribution entropy and the periodic characteristics corresponding to different periods.

[0012] Furthermore, after determining the corporate churn score of the target enterprise and the individual churn score of the individual customer, the process also includes: abstracting corporate payroll accounts, individual customer accounts, and external accounts as nodes; when there are transaction behaviors between accounts, abstracting the transaction behaviors as directed edges between the nodes indicated by the accounts; determining the node characteristics of each node, wherein the node characteristics include at least: transaction behavior characteristics and periodic behavior characteristics, the transaction behavior characteristics include at least one of the following: payroll amount, balance change, and the periodic behavior characteristics include at least one of the following: corporate churn score, individual churn score, corporate distribution entropy, and individual distribution entropy; and constructing a dynamic network graph of fund flow based on all nodes, the node characteristics corresponding to each node, and all directed edges.

[0013] Furthermore, after constructing a dynamic network graph of fund flows based on all nodes, the node characteristics corresponding to each node, and all directed edges, the process also includes: calculating the cumulative weight of each fund path based on the dynamic network graph of fund flows, and determining the total cumulative weight based on the cumulative weight of all fund paths; determining all node fund paths for each node, where a node fund path refers to a fund path starting from a node; determining the node score of a node based on the cumulative weight of each node's fund path and the total cumulative weight; if the node score is greater than a preset threshold, the node is determined as a preset transfer-out node, and a fund risk network is constructed based on all node fund paths of the preset transfer-out node.

[0014] Furthermore, the step of determining the target incentive strategy based on the enterprise churn score and individual churn score using a pre-set learning model includes: determining an initial incentive strategy set; for each initial incentive strategy in the initial incentive strategy set, processing the enterprise churn score, individual churn score, financial risk network, and initial incentive strategy using the pre-set learning model to obtain the strategy reward value of the initial incentive strategy, wherein the pre-set learning model is an interactive model of state, action, and reward, the state includes: enterprise churn score, individual churn score, and financial risk network, the action refers to the initial incentive strategy, and the reward refers to the strategy reward value, which is determined based on retention rate, cash outflow rate, and product acceptance rate; and determining the initial incentive strategy indicated by the maximum strategy reward value as the target incentive strategy.

[0015] According to another aspect of the present invention, an apparatus for determining an incentive strategy for payroll processing services is also provided, comprising: a collection unit, configured to collect payroll processing data of a target enterprise upon obtaining authorization from the target enterprise, wherein the payroll processing data includes at least: enterprise data, individual customer data, and payroll fund transaction data, wherein the individual customer is a customer of the target enterprise who receives salary payments through a financial institution; a first determination unit, configured to determine periodic features corresponding to multiple periods based on the payroll processing data using multi-scale sliding window convolution, wherein the interval length of different periods is inconsistent; a second determination unit, configured to determine the enterprise churn score of the target enterprise and the individual churn score of individual customers based on the payroll processing data and the periodic features corresponding to different periods; and a third determination unit, configured to determine a target incentive strategy based on the enterprise churn score and the individual churn score using a preset learning model, wherein the target incentive strategy is used to incentivize the target enterprise and all individual customers.

[0016] Further, the first determining unit includes: a first extraction module, used to extract features from the payroll business data to obtain target tensor data, wherein the target tensor data includes: a time series, and each payroll time in the time series corresponds to payroll business features; the first determining module is used to determine the periodic features corresponding to each period based on the target tensor data using multi-scale sliding window convolution, wherein the period includes: a first period, a second period, and a third period, the period length of the first period is less than the period length of the second period, the period length of the second period is less than the period length of the third period, and the periodic features include at least: mean, difference, and ratio.

[0017] Furthermore, the first extraction module includes: a first extraction submodule, used to extract features from the payroll data to obtain three-dimensional tensor data, wherein the three-dimensional tensor data includes: time dimension data, enterprise dimension data, and payroll business feature dimension data. The time dimension data includes a time index generated by discretizing the timestamps in the payroll data. The enterprise dimension data includes a time series associated with the payroll accounts of the target enterprise. The payroll business feature dimension data includes statistical features and structural features. A first compression submodule is used to compress the three-dimensional tensor data using a preset encoder to obtain target tensor data, wherein the preset encoder is an encoder trained by compressing and decompressing tensor data. The compressed tensor data is data compressed by the preset encoder, and the decompressed tensor data is data decoded by the decoder.

[0018] Furthermore, the first determining module includes: a first processing submodule, used to process the target tensor data to obtain periodic data under different periods; and a second processing submodule, used to process the periodic data using a preset convolution function for each period to obtain periodic features.

[0019] Furthermore, the second determining unit includes: a second determining module, used to determine a first probability distribution of payout funds based on payout business data, wherein payout funds refer to funds paid by enterprises to individual customers through financial institutions; a third determining module, used to determine a second probability distribution of account funds based on payout business data, wherein account funds refer to funds stored in different accounts after individual customers receive payout funds; a fourth determining module, used to determine the enterprise distribution entropy based on the first probability distribution and the individual distribution entropy based on the second probability distribution; a fifth determining module, used to determine the enterprise churn score using a preset multi-window feature fusion function based on the enterprise distribution entropy and the periodic characteristics corresponding to different periods; and a sixth determining module, used to determine the individual churn score using a preset multi-window feature fusion function based on the individual distribution entropy and the periodic characteristics corresponding to different periods.

[0020] Furthermore, the incentive strategy determination device also includes: a first abstraction module, used to abstract corporate payroll accounts, individual customer accounts, and external accounts as nodes after determining the corporate churn score of the target enterprise and the individual churn score of the individual customer; a second abstraction module, used to abstract the transaction behavior as directed edges between the nodes indicated by the accounts when there is transaction behavior between the accounts; a seventh determination module, used to determine the node characteristics of each node, wherein the node characteristics include at least: transaction behavior characteristics and periodic behavior characteristics, the transaction behavior characteristics include at least one of the following: payroll amount, balance change, and the periodic behavior characteristics include at least one of the following: corporate churn score, individual churn score, corporate distribution entropy, and individual distribution entropy; and a first construction module, used to construct a dynamic network graph of fund flow based on all nodes, the node characteristics corresponding to each node, and all directed edges.

[0021] Furthermore, the incentive strategy determination device also includes: a first statistics module, used to construct a dynamic network graph of fund flow based on all nodes, the node characteristics corresponding to each node, and all directed edges, and then to calculate the cumulative weight of each fund path based on the dynamic network graph of fund flow, and determine the total cumulative weight based on the cumulative weight of all fund paths; an eighth determination module, used to determine all node fund paths for each node, wherein a node fund path refers to a fund path starting from a node; a ninth determination module, used to determine the node score of a node based on the cumulative weight of each node's fund path and the total cumulative weight; and a tenth determination module, used to determine a node as a preset transfer-out node if the node score is greater than a preset threshold, and to construct a fund risk network based on all node fund paths of the preset transfer-out node.

[0022] Furthermore, the third determining unit includes: an eleventh determining module, used to determine the initial incentive strategy set; a first processing module, used to process the enterprise churn score, individual churn score, financial risk network, and initial incentive strategy for each initial incentive strategy in the initial incentive strategy set using a preset learning model to obtain the strategy reward value of the initial incentive strategy, wherein the preset learning model is an interactive model of state, action, and reward, the state includes: enterprise churn score, individual churn score, and financial risk network, the action refers to the initial incentive strategy, and the reward refers to the strategy reward value, which is determined based on retention rate, cash outflow rate, and product acceptance rate; and a twelfth determining module, used to determine the initial incentive strategy indicated by the maximum strategy reward value as the target incentive strategy.

[0023] According to another aspect of the present invention, a computer program product is also provided, including a non-volatile computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements any one of the above-described incentive strategy determination methods for payroll services.

[0024] According to another aspect of the present invention, an electronic device is also provided, including one or more processors and a memory, wherein the memory is used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement any of the above-described methods for determining incentive strategies for payroll services.

[0025] In this invention, with the authorization of the target enterprise, the payroll business data of the target enterprise is collected. Based on the payroll business data, multi-scale sliding window convolution is used to determine the periodic features corresponding to multiple periods. Based on the payroll business data and the periodic features corresponding to different periods, the enterprise churn score and the individual customer churn score of the target enterprise are determined. Based on the enterprise churn score and the individual churn score, a preset learning model is used to determine the target incentive strategy, thereby solving the technical problem in related technologies that the inability to determine an effective incentive strategy leads to a low customer retention rate.

[0026] This invention employs an integrated approach of spatiotemporal feature analysis and intelligent strategy optimization. Through multi-scale sliding window convolution, it accurately identifies the churn risk of corporate and individual customers in payroll processing, thereby achieving the dynamic adjustment of intelligent and personalized incentive strategies. This addresses the technical problems of low customer retention and lagging intervention in bank payroll processing. Specifically, while ensuring data security and compliance, it collects payroll processing data from authorized enterprises, including enterprise information, individual customer information, and payroll transaction details. Multi-scale sliding windows are used to extract cross-cycle behavioral characteristics of enterprises and individuals. Then, combined with risk entropy feedback, enterprise churn scores and individual churn scores are calculated. These risk scores are then input into a pre-trained reinforcement learning model to generate customized incentive strategies for enterprises and individuals, such as fee discounts and financial product recommendations. This allows for real-time responses to customer behavior, improving fund retention and customer value conversion, and avoiding the limitations of traditional fixed threshold strategies. Attached Figure Description

[0027] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:

[0028] Figure 1 A hardware structure block diagram of a computer terminal (or mobile device) for implementing an incentive strategy determination method for payroll services is shown.

[0029] Figure 2 This is a flowchart of an incentive strategy determination method for payroll services according to Embodiment 1 of the present invention;

[0030] Figure 3 This is a schematic diagram of an optional customer churn warning and fund retention improvement process for bank payroll services according to an embodiment of the present invention.

[0031] Figure 4 This is a schematic diagram of an optional incentive strategy determination device for payroll services according to an embodiment of the present invention;

[0032] Figure 5 This is a structural block diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0033] 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 should fall within the scope of protection of the present invention.

[0034] It should be noted that the terms "first," "second," etc., used in this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0035] It should be noted that all relevant information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, and displayed data) collected and involved in this invention are information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of this data comply with the relevant laws, regulations, and standards of the relevant regions, necessary confidentiality measures have been taken, and it does not violate public order and good morals. Corresponding operation entry points are provided for users to choose to authorize or refuse. For example, this system has an interface with relevant users or organizations. Before obtaining relevant information, a request to obtain the information needs to be sent to the aforementioned user or organization through the interface. The relevant information is obtained only after receiving consent from the aforementioned user or organization. If the user chooses to refuse, the process proceeds to an expert decision-making process.

[0036] This invention proposes a method for early warning of customer churn and improvement of fund retention in payroll services based on spatiotemporal feature fusion. It integrates multi-scale temporal feature extraction, dynamic fund flow graph modeling and reinforcement learning strategy optimization, which can comprehensively identify the risks of corporate churn and personal fund loss, and realize differentiated and real-time intervention to improve fund retention and customer value conversion capabilities.

[0037] This invention constructs a method for monitoring and intervening in fund outflows in bank payroll services, enabling accurate identification and dynamic response to both corporate and individual payroll fund outflows. By establishing a unified three-dimensional temporal feature tensor, the payroll data for both enterprises and individuals is transformed into structured input. Cross-cycle trends are extracted using a multi-scale sliding window to obtain core features that simultaneously reflect short-term fluctuations and long-term patterns. Based on this, a dynamic fund flow graph is introduced to aggregate cross-cycle transactions between corporate accounts, individual accounts, and other bank accounts, identifying key path nodes for concentrated transfers and revealing the fund outflow chain. Finally, a reinforcement learning strategy optimization module is introduced to form a closed loop between risk identification and customer feedback, enabling adaptive adjustment of transfer incentive parameters and wealth management recommendation thresholds, thereby avoiding the lag and over-intervention caused by fixed rules. This method not only improves the accuracy and timeliness of fund outflow early warning but also generates differentiated intervention strategies for different corporate and individual customers, effectively enhancing the bank's fund retention capacity and customer value conversion level.

[0038] The present invention will now be described in detail with reference to various embodiments.

[0039] Example 1

[0040] According to an embodiment of this application, an embodiment of a method for determining an incentive strategy for payroll services is also provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0041] The method embodiment provided in Embodiment 1 of this application can be executed on a mobile terminal, computer terminal, or similar computing device. Figure 1 A hardware block diagram of a computer terminal (or mobile device) for implementing an incentive strategy determination method for payroll services is shown. Figure 1 As shown, computer terminal 10 (or mobile device) may include one or more ( Figure 1The processor 102 (which may include, but is not limited to, a microprocessor MCU or a programmable logic device FPGA, etc.), a memory 104 for storing data, and a transmission device 106 for communication functions may also be included. In addition, it may include: a display, a keyboard, a cursor control device, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of a BUS bus), a network interface, a power supply, and / or a camera, wherein the network interface can be connected to wired and / or wireless networks. Those skilled in the art will understand that... Figure 1 The structure shown is for illustrative purposes only and does not limit the structure of the aforementioned electronic device. For example, computer terminal 10 may also include... Figure 1 The more or fewer components shown, or having the same Figure 1 The different configurations shown.

[0042] It should be noted that the aforementioned one or more processors 102 and / or other data processing circuits are generally referred to herein as "data processing circuits". These data processing circuits may be embodied, in whole or in part, in software, hardware, firmware, or any other combination thereof. Furthermore, the data processing circuits may be a single, independent processing module, or may be integrated, in whole or in part, into any other element within the computer terminal 10 (or mobile device). As involved in the embodiments of this application, the data processing circuits serve as a processor control mechanism (e.g., selection of a variable resistor termination path connected to an interface).

[0043] The memory 104 can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the incentive strategy determination method for payroll services in the embodiments of this application. The processor 102 executes various functional applications and data processing by running the software programs and modules stored in the memory 104, thereby implementing the above-mentioned incentive strategy determination method for payroll services. The memory 104 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 104 may further include memory remotely located relative to the processor 102, and these remote memories can be connected to the computer terminal 10 via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0044] The transmission device 106 is used to receive or send data via a network. Specific examples of the network described above may include a wireless network provided by the communication provider of the computer terminal 10. In one example, the transmission device 106 includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device 106 may be a Radio Frequency (RF) module, used for wireless communication with the Internet.

[0045] The display may be, for example, a touchscreen liquid crystal display (LCD) that allows the user to interact with the user interface of the computer terminal 10 (or mobile device).

[0046] Under the aforementioned operating environment, this application provides the following: Figure 2 The method for determining incentive strategies for payroll services is shown. Figure 2 This is a flowchart of an incentive strategy determination method for payroll services according to Embodiment 1 of the present invention, as shown below. Figure 2 As shown, the method includes the following steps:

[0047] Step S201: With the authorization of the target company, collect the payroll business data of the target company. The payroll business data includes at least: company data, individual customer data, and payroll transaction data. Individual customers are the customers of the target company who receive payroll through financial institutions.

[0048] In this embodiment of the invention, under the premise of ensuring that the operation complies with regulations and privacy, and after obtaining formal authorization from the target enterprise, the payroll business data of the target enterprise is obtained. This payroll business data is a type of payroll transaction data, involving information from three dimensions: enterprise side, individual side, and transaction side. It includes: Enterprise data: covering basic enterprise information, such as enterprise customer number, enterprise name, enterprise payroll account, industry category, enterprise size, etc. This data helps to construct a feature tensor for the enterprise dimension. Individual customer data: including the name of the payroll recipient, individual customer number, individual payroll account, etc. This individual-dimensional data is used to analyze the cash flow after payroll. Payroll transaction data: recording detailed information such as the time, amount, and frequency of each payroll payment. This transaction data is sequential data in the time dimension and is used to construct a time-series feature tensor.

[0049] Here, individual customers refer to bank customers whose payroll is disbursed through financial institutions by the target company.

[0050] Step S202: Based on the data of the agency business, multi-scale sliding window convolution is used to determine the periodic features corresponding to multiple periods, wherein the interval length of different periods is inconsistent.

[0051] In this embodiment of the invention, after collecting payroll data, multi-scale sliding window convolution technology is used to extract the changing trends of corporate payroll amounts, frequency, and average single transaction amounts at different time scales (e.g., 7 days, 15 days, 30 days), as well as the withdrawal and transfer behavior of individual customers at different time periods after their salaries are deposited. Sliding windows of varying sizes (7 / 15 / 30 days) can be used to convolve the payroll data to calculate the short-term, medium-term, and long-term window mean, difference, and ratio, which can be used to characterize whether companies face potential churn risks and whether individuals will quickly transfer payroll funds out of the banking system. Extracting these cyclical features helps to capture the cyclical patterns of fund behavior, providing a quantitative basis for subsequent churn warnings and fund retention improvement strategies.

[0052] Step S203: Based on the payroll business data and the cycle characteristics corresponding to different cycles, determine the corporate churn score of the target enterprise and the individual churn score of the individual customer.

[0053] In this embodiment of the invention, based on payroll data and extracted periodic characteristics, corporate churn scores and individual churn scores are further calculated. By introducing risk entropy as a feedback factor, the uncertainty and risk level of behavior are measured according to the distribution entropy of enterprises and the distribution entropy of individual fund destinations. That is, a specific mathematical model is used to assess whether each enterprise's payroll business is stable, and whether each individual customer's funds are likely to be quickly withdrawn from the bank after salary is deposited. The determination of corporate churn scores and individual churn scores can quantitatively identify which enterprises or individuals have a high risk of churn, providing a clear target for implementing intervention measures.

[0054] Step S204: Based on the enterprise churn score and the individual churn score, a preset learning model is used to determine the target incentive strategy, wherein the target incentive strategy is used to incentivize the target enterprise and all individual customers.

[0055] In this embodiment of the invention, based on the calculated corporate churn score and individual churn score, a pre-trained reinforcement learning model is used to determine the most effective target incentive strategy. This strategy will guide the bank to provide differentiated marketing and services to target companies and their individual payroll clients. For example, for high-risk companies, fee reductions or customized financial products may be offered; for individual clients with rapid fund transfers, lower thresholds for recommending wealth management products may be set. The incentive strategy is formulated to maximize the bank's fund retention rate and customer loyalty. The model automatically adjusts the strategy parameters based on customer feedback on various incentive measures, forming a closed-loop optimization mechanism. Through such dynamic adjustments, the bank can more flexibly and accurately respond to the needs and behavioral patterns of different customer groups, improving the overall performance of payroll services.

[0056] In summary, by integrating spatiotemporal feature analysis and intelligent strategy optimization, and employing multi-scale sliding window convolution, this approach accurately identifies the churn risk of corporate and individual customers in payroll processing. This enables the dynamic adjustment of intelligent and personalized incentive strategies, thereby addressing the technical challenges of low customer retention and lagging intervention in bank payroll processing. Specifically, while ensuring data security and compliance, the system collects payroll processing data from authorized enterprises, including enterprise information, individual customer information, and payroll transaction details. Multi-scale sliding windows are used to extract cross-cycle behavioral characteristics of enterprises and individuals. Then, combined with risk entropy feedback, enterprise churn scores and individual churn scores are calculated. These risk scores are then input into a pre-trained reinforcement learning model to generate customized incentive strategies for enterprises and individuals, such as fee discounts and financial product recommendations. This allows for real-time responses to customer behavior, improving fund retention and customer value conversion, and avoiding the limitations of traditional fixed threshold strategies.

[0057] To improve the accuracy of determining periodic features, in the incentive strategy determination method for payroll services provided in Embodiment 1 of this application, feature extraction is performed on payroll service data to obtain target tensor data. The target tensor data includes a time series, where each payroll time in the time series corresponds to a payroll service feature. Based on the target tensor data, a multi-scale sliding window convolution is used to determine the periodic features corresponding to each period. The period includes a first period, a second period, and a third period. The period length of the first period is less than the period length of the second period, and the period length of the second period is less than the period length of the third period. The periodic features include at least the mean, difference, and ratio.

[0058] In this embodiment of the invention, deep feature extraction is performed on the collected payroll data of the target enterprise to construct a three-dimensional target tensor data that comprehensively reflects the payroll behavior of both enterprises and individuals. This target tensor data contains multiple time series, each representing the payroll behavior of an enterprise's payroll account at different points in time, covering business characteristics such as total payroll amount, number of payrolls, average single transaction amount, and number of payroll participants. Each payroll moment, i.e., each point in time, contains rich payroll business characteristics, such as the payroll amount and the number of payroll participants at that moment, enabling the time series to precisely describe the historical evolution of the enterprise's payroll behavior.

[0059] Then, to comprehensively capture the dynamic patterns of fund flows, a multi-scale sliding window convolution technique is used to determine the cyclical characteristics corresponding to the first, second, and third cycles. For example, these three cycles are 7 days, 15 days, and 30 days, representing short-term, medium-term, and long-term time windows, respectively. Within each window, the following three core cyclical characteristics are calculated: Mean: Calculates the average amount and average number of fund transfers within the window to depict the overall level of fund transfer behavior. Difference: Calculates the changes in the amount and number of fund transfers between two adjacent time points, helping to identify the growth or decline trend of fund transfer scale in the short term. Ratio: Calculates the ratio of the current time point to the starting time point of the window, revealing the relative magnitude of change in fund transfer business.

[0060] By applying sliding window convolution across the first, second, and third periods, we can not only capture the short-term volatility of corporate payroll services but also observe the changing trends of capital flows in the medium and long term. This is crucial for identifying potential capital outflow risks. In particular, the extraction of periodic features considers the impact of different periods on risk identification. By dynamically adjusting the window length, we can analyze the behavioral patterns of enterprises and individuals in greater detail, improving the accuracy and timeliness of risk warnings.

[0061] In this embodiment, the periodic feature extraction method based on multi-scale sliding window convolution can effectively extract the patterns of cash flow changes for corporate and individual customers from massive amounts of payroll data. In particular, it can focus on short-term sudden changes and long-term trends, which is of great significance for early warning of customer churn in payroll services. By quantitatively analyzing the performance of each corporate payroll account over different time periods, banks can promptly identify signs of customer churn that may be caused by deteriorating financial conditions, competitor attraction, or other factors. This allows for more precise and effective intervention measures, such as providing customized services or personalized financial product recommendations, to improve fund retention rates and customer satisfaction. In this way, not only is the management efficiency of the bank's payroll business improved, but its market competitiveness and profitability are also enhanced.

[0062] To improve the accuracy of determining the target tensor data, in the incentive strategy determination method for payroll services provided in Embodiment 1 of this application, feature extraction is performed on the payroll service data to obtain three-dimensional tensor data. The three-dimensional tensor data includes: time dimension data, enterprise dimension data, and payroll service feature dimension data. The time dimension data includes a time index generated by discretizing the timestamps in the payroll service data. The enterprise dimension data includes a time series associated with the payroll accounts of the target enterprise. The payroll service feature dimension data includes statistical and structural features. A preset encoder is used to compress the three-dimensional tensor data to obtain the target tensor data. The preset encoder is an encoder trained by compressing and decompressing the tensor data. The compressed tensor data is data compressed by the preset encoder, and the decompressed tensor data is data decoded by the decoder.

[0063] In this embodiment of the invention, a quantitative modeling-based processing method is proposed for the raw transaction data of bank payroll disbursement services. Here, the payroll disbursement data (payroll disbursement business data) includes information from three dimensions: enterprise side, individual side, and transaction side. Enterprise side information mainly includes enterprise customer number, enterprise name, enterprise payroll account, industry category, enterprise size, etc.; individual side information mainly includes the name of the individual receiving the payroll, individual customer number, and individual payroll account; transaction information includes payroll time, payroll amount, payroll frequency, etc.

[0064] To achieve efficient modeling and feature extraction of the above data, it is transformed into a three-dimensional tensor structure. The three dimensions of the tensor are the enterprise dimension (…). ), time dimension ( ) and business characteristic dimensions ( As shown in Table 1:

[0065] Table 1

[0066]

[0067] Using the above modeling method, the constructed tensor is: ,in, Represents a vector space.

[0068] In payroll dropshipping scenarios, due to the limited transaction volume of some small and micro enterprises, tensors exhibit sparsity, leading to high computational costs and feature redundancy when directly modeled. Therefore, an autoencoder is introduced to reduce and compress high-dimensional sparse features. Specifically, the autoencoder uses nonlinear mapping to reduce the dimensionality of high-dimensional sparse features. Compression into low-dimensional dense features, the formula is: ,in, Let D represent the tensor before compression, and D represent the feature dimensions before compression. For encoder functions, The compressed feature dimensions, This represents the compressed tensor. Decoder Used to reconstruct the original features, ensuring the effectiveness of compression, and minimizing reconstruction error. As a training objective, among which... Represents the decoder function. This represents the decompressed tensor. The tensor obtained after compression is... Compared to the original tensor The sparsity is significantly reduced, and the correlation between features is more concentrated, which not only preserves the key features of the financial fluctuations on payday, but also effectively reduces data redundancy.

[0069] Specifically, the acquired payroll processing data is first subjected to feature extraction to transform fragmented payroll information into structured three-dimensional tensor data. This three-dimensional tensor consists of time-dimension data, enterprise-dimension data, and payroll processing feature-dimension data. The time-dimension data involves discretizing the timestamps in the payroll processing data, converting continuous time points into preset time periods, such as weeks, bi-weekly periods, or months. This transformation process generates a time-series index that facilitates analysis, allowing tracking of payroll activity trends over time. The time index, as the first dimension of the tensor, ensures that the temporal characteristics of payroll behavior are preserved. The enterprise-dimension data treats each enterprise's payroll account as an independent time series, constituting the second dimension of the tensor. This means that regardless of enterprise size, as long as payroll processing has been performed, a matching time series record will exist documenting its payroll history and behavioral patterns. This allows for personalized analysis tailored to the specific circumstances of each enterprise, rather than simply relying on averages from large-scale data. Data on characteristics of payroll services includes various statistical and structural features closely related to payroll services, such as statistical features (total payroll amount in the current period, average amount per payroll, total number of payrolls) and structural features (number of individual payroll accounts, concentration of fund allocation, proportion of abnormal transactions). Extracting these features is crucial for understanding and predicting changes in payroll behavior, collectively forming the third dimension of the tensor, providing a multi-faceted perspective on the health of the payroll service.

[0070] To address the inherent high-dimensional sparsity issue in payroll data, a pre-defined encoder is used to compress the 3D tensor data, yielding the target tensor data. Here, the encoder is a meticulously trained tool, obtained through training on compressed and decompressed tensor data. It is used to compress the high-dimensional but sparse original features into low-dimensional and dense target features, facilitating subsequent model processing and analysis. The compressed tensor data is the result of the encoder's compression, while the decompressed tensor data is the version the decoder attempts to reconstruct. The goal of training the encoder and decoder is to minimize the reconstruction error; that is, the decompressed data should be as close as possible to the original data, thus ensuring that no key information is lost during the compression process.

[0071] In this embodiment, dynamic temporal data, enterprise-specific information, and various aspects of payroll behavior are effectively integrated to construct target tensor data. This data structure not only comprehensively reflects the current status and development trend of payroll business, but also overcomes the challenges posed by data redundancy and sparsity through compression optimization by a pre-set encoder, laying a solid foundation for subsequent advanced analysis steps such as multi-scale sliding window convolution.

[0072] To further improve the accuracy of determining periodic characteristics, in the incentive strategy determination method for payroll services provided in Embodiment 1 of this application, the target tensor data is processed to obtain periodic data under different periods; for each period, a preset convolution function is used to process the periodic data to obtain periodic characteristics.

[0073] In this embodiment of the invention, the financial behavior in payroll scenarios often exhibits cross-cycle characteristics. At the enterprise level, the changing trends of payroll amounts and the number of employees in different cycles can reveal whether the enterprise has potential employee attrition risks. At the individual level, employee withdrawals and transfers after their salaries are credited to their accounts follow different cyclical patterns. Within a week of receiving their salaries, some employees will withdraw cash, make purchases, or repay credit card debts. Within half a month, employees may gradually use the funds for household expenses. Throughout the month, employees may tend to retain their salaries for a longer period or transfer them to investment products. To capture these behavioral patterns, a compression tensor is used... A multi-scale sliding window convolution structure is introduced to compress the tensor under 7-day, 15-day, and 30-day windows, respectively. Calculations are performed to obtain the corresponding short-term, medium-term, and long-term features. These features include: window mean (characterizing the overall smooth trend), window difference (identifying growth or decline trends), and window ratio (measuring relative volatility). The multi-scale sliding window convolutional structure can use the following convolution function (i.e., the preset convolution function) to extract periodic features: ,in, This represents the periodic characteristics under different periods k. This represents the weights under different periods k. This represents the bias at different periods k. Represents the compressed tensor over the time interval ti. ti represents the time interval between the current time t and the i-th period.

[0074] Specifically, the target tensor data obtained from the previous stage is further analyzed periodically to extract the dynamic characteristics of capital flows from different time windows. The target tensor data is a three-dimensional data structure composed of time, enterprise, and payroll business characteristics, with each dimension carrying key information. By discretizing the time dimension, payroll records are divided into preset periods, generating first-period data (e.g., 7 days), second-period data (e.g., 15 days), and third-period data (e.g., 30 days). Each period is a subset of the target tensor data, containing complete information about the enterprise's payroll activities within that period. In this way, the patterns of capital flows can be captured at different time scales, whether it is short-term fluctuations or long-term trend changes.

[0075] For each generated periodic data point, a pre-defined convolutional function is applied for deep feature extraction to obtain the periodic features corresponding to that period. A convolutional function is a mathematical operation that captures local patterns and trends in data by sliding across it and applying specific filters. Specifically, the pre-defined convolutional function calculates the periodic mean, periodic difference, and periodic ratio, reflecting the average level, growth or decline trend, and volatility of funds disbursed within a specific time window. For example, the periodic mean reveals whether the amount and frequency of payroll disbursements remain consistent within a given period; the periodic difference identifies sudden decreases or increases; and the periodic ratio provides insights into the relative changes in disbursement behavior, which is crucial for identifying periodic patterns in fund flows.

[0076] In this embodiment, the flow patterns of funds in payroll services can be accurately captured over different time periods, generating a series of cyclical characteristics that reflect the behavioral patterns of corporate and individual customers. These characteristics not only reveal rapid withdrawals and outflows of funds in the short term, but also provide insights into retention intentions and fund allocation strategies in the medium to long term, which is of great significance for early warning of corporate customer churn and improving individual fund retention rates.

[0077] To improve the accuracy of determining corporate churn and individual churn, the incentive strategy determination method for payroll services provided in Embodiment 1 of this application determines the following based on payroll data: a first probability distribution of payroll funds is determined, where payroll funds refer to funds paid by the company to individual customers through financial institutions; a second probability distribution of account funds is determined based on payroll data, where account funds refer to funds stored in different accounts by individual customers after receiving payroll funds; corporate distribution entropy is determined based on the first probability distribution, and individual distribution entropy is determined based on the second probability distribution; corporate churn is determined using a preset multi-window feature fusion function based on corporate distribution entropy and the periodic characteristics corresponding to different periods; and individual churn is determined using a preset multi-window feature fusion function based on individual distribution entropy and the periodic characteristics corresponding to different periods.

[0078] In this embodiment of the invention, each record of payroll payments made by a corporate client to an individual client through a financial institution is extracted from the payroll processing dataset. Here, payroll payments refer to salary payments made by a bank on behalf of a company, serving as a crucial link between the corporate and individual clients. By statistically analyzing the distribution of these funds, a first probability distribution can be obtained. This refers to the distribution pattern of corporate payroll funds across different time periods, amounts, and individual accounts. This distribution reflects the preferences and patterns of companies when conducting payroll services; for example, some companies pay salaries at the beginning of the month, while others may choose the middle of the month. Furthermore, fluctuations in payroll amounts may reveal changes in the company's economic situation. Additionally, it allows for the determination of individual customers' account fund operations after receiving payroll funds. These funds may be stored in different accounts within the bank (such as savings accounts and credit card accounts) or transferred to other financial institutions. By analyzing individual customer account activities, a second probability distribution of account funds can be determined. This refers to the pattern of fund allocation and transfer among individual accounts, including fund retention rate, interbank transfer rate, and investment product selection preferences. The second probability distribution reflects individual customers' preferences for fund usage and liquidity needs, and is an important basis for assessing the risk of individual customer churn.

[0079] After obtaining the first and second probability distributions, the distribution entropy of each distribution is calculated. Distribution entropy is a statistical concept used to measure the uncertainty and information content of a probability distribution. In the context of payroll services, enterprise distribution entropy and individual distribution entropy reflect the randomness and complexity of enterprise payroll behavior and individual account fund operations, respectively. Generally, the more dispersed the distribution, the higher the entropy value, indicating that the behavioral patterns are more difficult to predict and the potential risk of loss is greater. The formula for enterprise distribution entropy is: E represents the enterprise. The probability distribution representing different fund transfer patterns (i.e., the first probability distribution); the formula for individual distribution entropy is: P represents an individual. This represents the probability distribution of different fund transfer patterns (i.e., the second probability distribution).

[0080] By leveraging enterprise distribution entropy and previously extracted periodic features (such as mean, difference, and ratio) corresponding to different periods, a pre-defined multi-window feature fusion function is used to comprehensively assess the churn risk of enterprise customers. This multi-window feature fusion function considers changes in short-term and long-term payroll patterns, ensuring that the assessment results accurately reflect the dynamics of enterprise behavior. The enterprise churn score is a quantitative indicator measuring the stability of the enterprise's payroll business and the likelihood of future churn. Furthermore, based on individual distribution entropy and periodic features, an individual churn score is determined using the multi-window feature fusion function. The individual churn score focuses on individual customers' fund handling behavior after receiving payroll funds, such as rapid withdrawal, retention rate, and investment preferences, and is a key indicator for predicting individual customer retention intentions and the risk of fund outflow.

[0081] Here, the formula for the multi-window feature fusion function is: k represents the period. This represents the features (mean, difference, ratio) under the k-window. The weight is defined as: Where j represents the period k, Let REt represent the coefficient corresponding to period k, and REt represent the distribution entropy. A higher entropy value indicates a complex behavioral pattern, requiring an increase in the weight of long-period features; conversely, a lower entropy value indicates a concentrated pattern, necessitating a stronger role for short-period features. Through this dynamic weight adjustment mechanism, interpretable characteristics of fund outflow and leakage risks can be extracted from payroll scenarios for different corporate and individual clients, providing a basis for subsequent risk identification and intervention.

[0082] In this embodiment, a comprehensive analysis of the fund disbursement behavior of corporate and individual clients is achieved through in-depth mining of disbursement business data. The calculation of the first and second probability distributions, along with the quantification of distribution entropy, allows for a statistical understanding of the regularity and uncertainty of fund flows. Furthermore, the corporate churn score and individual churn score determined by the multi-window feature fusion function further transform this statistical analysis into specific early warning signals, helping banks to promptly identify potential fund loss risks and providing data support for implementing differentiated and precise customer retention strategies.

[0083] To improve the accuracy of constructing a dynamic network graph of fund flows, in the incentive strategy determination method for payroll services provided in Embodiment 1 of this application, after determining the corporate churn score of the target enterprise and the personal churn score of the individual customer, the corporate payroll account, the individual customer account, and the external account are abstracted as nodes; when there are transaction behaviors between accounts, the transaction behaviors are abstracted as directed edges between the nodes indicated by the accounts; the node characteristics of each node are determined, wherein the node characteristics include at least: transaction behavior characteristics and periodic behavior characteristics, the transaction behavior characteristics include at least one of the following: payroll amount, balance change, and the periodic behavior characteristics include at least one of the following: corporate churn score, individual churn score, corporate distribution entropy, and individual distribution entropy; based on all nodes, the node characteristics corresponding to each node, and all directed edges, a dynamic network graph of fund flows is constructed.

[0084] In this embodiment of the invention, to further enhance the characterization of capital outflow risk, based on extracting cross-cycle behavioral characteristics of enterprises and individuals and identifying potential outflow risks, the fund transfer relationship between accounts is introduced to establish a dynamic graph structure that evolves with the cycle (i.e., a dynamic network graph of fund flow) to identify key transfer nodes and abnormal paths. Specifically, corporate payroll accounts, personal payroll accounts, and external accounts are abstracted into three types of nodes. , , ,time The transaction behavior is abstracted as a directed edge This forms a dynamic neural network diagram G of fund flows that evolves cyclically. Where G1, G2, ..., GT represent neural network graphs at different times, and the nodes of the neural network graph Gt corresponding to time t are V, and the directed edges are... Furthermore, each node contains two types of features: transaction behavior features such as disbursement amount and balance changes, and cross-cycle behavior features such as corporate churn points, individual outflow points, and risk entropy. Aggregating node features and analyzing cross-cycle paths allows us to characterize changes in corporate disbursement recipients, the destination and channel flow of individual funds. The node update process is as follows: ,in, Represents a node In the cycle The state vector, Let N(v) represent the edge attributes (transfer amount, number of transactions, etc.), N(v) represent the set of edges for node v, and u represent the edge. For aggregate functions, It is a non-linear update function.

[0085] Specifically, corporate payroll accounts, individual customer accounts, and external accounts can be abstracted as nodes. Here, a "node" represents an entity in the network, which can be any identifiable and trackable object, such as an internal bank personal savings account, a corporate payroll account, or an external account outside the bank. In the scenario of bank payroll business, corporate payroll accounts and individual customer accounts are the starting and ending points of fund flows, while external accounts may involve interbank transfers of funds. The transaction behavior between accounts is abstracted as directed edges. A directed edge represents a unidirectional connection from one node to another, used to depict the direction of information or material flow. In this embodiment, the existence of directed edges signifies fund transfers from corporate payroll accounts to individual customer accounts, or from individual customer accounts to external accounts. Each directed edge carries transaction behavior characteristics; the most direct relevant information is the payroll amount and balance changes. These characteristics reflect the specific flow and amount of funds and are the basic data for constructing a dynamic network graph of fund flows. Then, the node characteristics of each node are determined. Node characteristics are a comprehensive indicator that contains key information about a node's behavioral patterns in payroll processing, including transactional and cyclical behavioral characteristics. Transactional behavioral characteristics are directly related to fund transactions; for example, the payroll amount in a corporate account reflects the total amount of each payroll payment, while changes in the balance of an individual customer account reveal the use or retention of funds after receiving payroll. Cyclical behavioral characteristics assess the dynamic changes in behavior from a more macroscopic perspective, including corporate and individual churn scores, quantifying the churn risk of corporate and individual customers in different periods. They also include corporate and individual distribution entropy, which measure the uncertainty of fund usage in corporate payroll and customer accounts, reflecting the complexity and randomness of the behavior. Subsequently, all abstract nodes, their corresponding node characteristics, and directed edges representing transaction behavior are combined to construct a dynamic network graph of fund flows. This network graph is a dynamic, multi-dimensional data structure that not only depicts the flow path of funds between internal and external bank accounts but also presents the patterns and potential risks of fund flows through node characteristics and the risk information carried by directed edges. The structure and weights of a network graph can be updated over time, reflecting the dynamic changes in capital flow patterns.

[0086] In this embodiment, not only can the flow path of funds during corporate payroll, personal receipt, and external transfers be visualized, but key nodes and abnormal paths of fund outflow can also be identified based on node characteristics and directed edge weights in the network graph. This identification capability enables banks to more accurately pinpoint potential fund loss risks, such as personal accounts frequently making large transfers or corporate clients whose payroll amounts suddenly decrease. Furthermore, the inclusion of cyclical behavioral characteristics gives this risk identification a temporal dimension, enabling the capture of risk signals from short-term fluctuations and long-term trends, providing more comprehensive and timely information support for bank decision-making.

[0087] To improve the accuracy of constructing a fund risk network, in the incentive strategy determination method for payroll services provided in Embodiment 1 of this application, after constructing a dynamic fund flow network graph based on all nodes, the node characteristics corresponding to each node, and all directed edges, the cumulative weight of each fund path is calculated based on the dynamic fund flow network graph, and the total cumulative weight is determined based on the cumulative weight of all fund paths; all node fund paths of each node are determined, wherein a node fund path refers to a fund path starting from a node; the node score of a node is determined based on the cumulative weight of each node fund path and the total cumulative weight; if the node score is greater than a preset threshold, the node is determined as a preset transfer-out node, and a fund risk network is constructed based on all node fund paths of the preset transfer-out node.

[0088] In this embodiment of the invention, to identify critical paths, a graph attention mechanism is further employed to calculate the cumulative weight of each funding path. Let P represent a funding path, and (u, v) ∈ P represent the edge u and node v belonging to that funding path. The proportion of nodes in all outgoing funding paths is calculated. Obtain the node criticality index. Defined as: Through the analysis of By sorting the data, key outflow points of funds can be identified, and abnormal cross-channel transfer paths can be located. This process expands risk identification from a single dimension to a chain dimension, improving interpretability and monitoring accuracy.

[0089] Specifically, weight analysis can be performed based on the constructed dynamic network graph of fund flows to quantify the importance of fund transfers along different paths. First, the cumulative weight of each fund path can be calculated. This involves calculating the cumulative weight of each fund path (composed of a series of connected directed edges) from one account to another in the network graph. This is typically based on characteristics such as the transaction amount, frequency, or risk score of each directed edge in the path. The cumulative weight reflects the relative importance of the fund path in the overall fund flow. Then, the cumulative weights of all fund paths are summed to obtain a quantitative indicator reflecting the overall importance of all fund flows in the entire network—the total cumulative weight. For each node in the network graph (corporate payroll account, individual customer account, or external account), the set of fund paths originating from that node is identified. For each node, the sum of the cumulative weights of all its node fund paths is calculated and compared with the total cumulative weight to obtain the node score (or node criticality indicator). The node score measures the centrality and influence of a node in the fund network, revealing which accounts are key transit points for fund outflows. Next, a threshold can be set; when a node's score exceeds this threshold, it is marked as a preset outflow node. The identification of preset outflow nodes is based on their pivotal role in the fund flow network. These nodes may frequently or significantly transfer funds to external accounts, representing a key source of fund retention risk. Following this, focusing on the preset outflow nodes, all fund paths between these nodes and external accounts are analyzed, constructing a fund risk network. This network not only includes the preset outflow nodes but also covers all related fund outflow paths, providing banks with an intuitive and targeted view of outflow risk.

[0090] In this embodiment, key nodes of capital outflow risk, i.e., preset transfer-out nodes, can be accurately located from the dynamic network diagram of capital flows. By statistically analyzing the cumulative weights of paths and calculating node scores, it is possible to identify which corporate payroll accounts or personal accounts play a significant role in capital flows, especially those accounts that frequently or significantly transfer funds to accounts at other banks. The construction of the capital risk network extends risk identification from the single account level to the path link level, enhancing the depth and breadth of risk analysis and providing banks with a more detailed risk management perspective. This not only helps banks to promptly detect and intervene in high-risk capital outflow behaviors but also enables them to optimize capital retention strategies and improve customer capital retention rates by analyzing the path characteristics in the capital risk network. For example, by understanding typical capital outflow paths, banks can offer more attractive wealth management products or services, reduce external transfers of customer funds, and thus effectively improve the bank's capital retention capacity and customer satisfaction.

[0091] To improve the accuracy of determining the target incentive strategy, in the incentive strategy determination method for payroll services provided in Embodiment 1 of this application, an initial incentive strategy set is determined. For each initial incentive strategy in the initial incentive strategy set, a preset learning model is used to process the enterprise churn score, individual churn score, financial risk network, and the initial incentive strategy to obtain the strategy reward value of the initial incentive strategy. The preset learning model is an interactive model of state, action, and reward. The state includes: enterprise churn score, individual churn score, and financial risk network. The action refers to the initial incentive strategy, and the reward refers to the strategy reward value, which is determined based on retention rate, cash outflow rate, and product acceptance rate. The initial incentive strategy indicated by the maximum strategy reward value is determined as the target incentive strategy.

[0092] In this embodiment of the invention, a reinforcement learning mechanism can be introduced to achieve a dynamic optimization closed loop for incentive strategies targeting enterprise and individual customers. Currently, fixed thresholds are used for triggering, such as recommending financial products to customers when salary deposits exceed a threshold. However, such static rules lack specificity and are difficult to adapt to the behavioral differences among different customer groups. This embodiment models this problem as a state-action-reward interaction process, where the state... Defined as Input includes enterprise turnover risk score (Company turnover score), individual turnover risk score (Personal attrition points), financial network risks (Funding Risk Network), Incentive Feedback (Initial incentive strategy). Action This indicates an adjustment to the incentive strategy, in which, Incentive parameters for enterprise clients (such as exclusive product discounts, fee reductions, etc.). Incentive parameters for individual customers (such as wealth management product recommendation thresholds). After each incentive program, collect customer feedback, including customer retention rate for direct deposit services, interbank fund outflow rate, and product recommendation acceptance rate, and calculate the reward. Determined by customer fund retention rate and incentive strategy adoption feedback, with the goal of maximizing the reward function: Where λ1, λ2, and λ3 represent weights, and ReasonRate is... t Retention rate is represented by Baseline, which represents the standard value (which can be set according to actual conditions), and OutRate. t AccepRate represents the outflow rate. tThis represents the acceptance rate. Reinforcement learning iterates strategies through a continuous cycle of updating the state-action-reward function. Through this mechanism, immediate intervention can be implemented when the risk of rapid fund loss is detected, and incentive strategies can be continuously optimized based on customer feedback, forming a dynamic "identification-intervention-feedback-optimization" closed loop to improve fund retention and customer loyalty in payroll processing.

[0093] Specifically, the initial set of incentive strategies should be created first. This set includes various incentive strategies that the bank can implement to promote capital retention and customer value conversion. Incentive strategies include, but are not limited to: offering interest rate discounts, awarding points or gifts, launching exclusive wealth management products, and conducting customized marketing campaigns. Each strategy has its specific goals and expected results, but the effectiveness needs to be verified through subsequent data analysis.

[0094] Then, for each initial incentive strategy in the set, a pre-defined learning model is used for analysis. The pre-defined learning model is a decision-making process model that interacts and evaluates based on three key components: state, action, and reward. In this model, the state mainly includes: corporate churn score, individual churn score, and financial risk network. These three elements together constitute the starting point for model evaluation, reflecting the potential risk of current customer churn and the pressure on fund retention. Action refers to the initial incentive strategy in the set, i.e., any specific intervention measure the bank can take. The reward is set as a strategy reward value, a comprehensive indicator used to quantify the positive feedback obtained after adopting a certain strategy, including improved retention rate, reduced fund outflow rate, and increased product acceptance rate.

[0095] Here, the model learns from historical data and, given the current state (corporate churn score, individual churn score, and financial risk network), predicts the possible rewards (strategy reward value) after performing different actions (incentive strategies) in order to find the strategies that can best improve retention rate, reduce cash outflow, and increase product acceptance.

[0096] To obtain the most effective strategy, a corresponding strategy reward value is calculated for each initial incentive strategy. The level of the strategy reward value reflects the expected benefits of implementing the strategy and is directly related to the bank's retention targets and customer relationship management effectiveness. Once the strategy reward values ​​of all initial incentive strategies have been calculated, they can be compared and the strategy with the highest strategy reward value can be selected. This strategy, also known as the target incentive strategy, is the action plan most likely to bring the best retention results, the lowest churn rate, and the highest product acceptance under the current circumstances.

[0097] In this embodiment, by analyzing customer behavior patterns and cash flow trends, appropriate incentive measures can be matched more accurately, avoiding resource waste and over-marketing. Furthermore, the pre-set learning model can quickly evaluate the effectiveness of various strategies, adjust intervention measures in a timely manner, and improve the speed of responding to customer churn and cash outflow events. Moreover, with the accumulation of time and data, the model can continuously learn and improve itself, constantly enhancing the accuracy of strategy reward values, thereby better serving the goals of cash retention and customer value conversion.

[0098] Figure 3 This is a schematic diagram of an optional customer churn warning and fund retention improvement process for bank payroll services according to an embodiment of the present invention, as shown below. Figure 3 As shown, the process includes the following steps: After obtaining the transaction history, feature tensors are extracted and sparse features are compressed. Then, multi-scale sliding window convolution is used to extract cross-cycle trends to obtain cyclical features that simultaneously reflect short-term fluctuations and long-term patterns. Based on these cyclical features, enterprise churn risk scores and individual churn risk scores are calculated. On this basis, a dynamic network graph of fund flows is constructed, aggregating cross-cycle transaction behaviors between enterprise accounts, individual accounts, and other bank accounts to identify key path nodes for concentrated transfers, revealing the outflow of funds and generating fund network risk. Finally, a reinforcement learning strategy optimization module is introduced, taking the enterprise churn risk score, individual churn risk score, and fund network risk as inputs, and outputting an incentive strategy through incentive feedback.

[0099] The incentive strategy determination method for payroll processing provided in this application can effectively solve the problems of untimely identification of fund outflows, limited intervention methods, and low fund retention rates in bank payroll processing. Through three-dimensional tensor modeling, multi-scale feature extraction, dynamic graph neural network analysis, and reinforcement learning closed-loop optimization, comprehensive monitoring and dynamic intervention of fund flows can be achieved in bank payroll processing. In this embodiment, the original payroll transaction data is first converted into a three-dimensional temporal feature tensor (enterprise × time period × business dimension). Sparse features are compressed using an autoencoder, effectively reducing data redundancy and improving modeling efficiency. Furthermore, multi-scale sliding window convolutional layers (e.g., window sizes of 7 / 15 / 30 days) are used to extract cross-period behavioral features, with window weights adjusted by dynamic risk entropy feedback, improving the fund outflow risk identification rate. Then, a dynamic graph network of fund flows is constructed, aggregating cross-channel transaction sequences between payroll accounts and other bank accounts to identify key path nodes for fund transfers, achieving precise location of fund outflow risks. In addition, the strategy optimization module based on reinforcement learning updates the transfer incentive parameters and financial recommendation thresholds in real time according to intervention feedback, generating a dynamic decision-making closed loop.

[0100] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0101] Example 2

[0102] This application also provides an incentive strategy determination device for payroll services. It should be noted that the incentive strategy determination device for payroll services in this application can be used to execute the incentive strategy determination method for payroll services provided in this application. The following describes the incentive strategy determination device for payroll services provided in this application.

[0103] According to an embodiment of this application, an apparatus for implementing the above-described method for determining incentive strategies for payroll services is also provided. Figure 4 This is a schematic diagram of an optional incentive strategy determination device for payroll services according to an embodiment of the present invention, such as... Figure 4 As shown, the excitation strategy determination device may include: a data acquisition unit 40, a first determination unit 41, a second determination unit 42, and a third determination unit 43.

[0104] The collection unit 40 is used to collect the payroll business data of the target enterprise when authorized by the target enterprise. The payroll business data includes at least: enterprise data, individual customer data, and payroll transaction data. Individual customers are customers of the target enterprise who receive their salaries through financial institutions.

[0105] The first determining unit 41 is used to determine the periodic features corresponding to multiple periods based on the agency business data by using multi-scale sliding window convolution, wherein the interval length of different periods is inconsistent.

[0106] The second determining unit 42 is used to determine the corporate churn score of the target enterprise and the individual churn score of the individual customer based on the payroll business data and the cycle characteristics corresponding to different cycles.

[0107] The third determining unit 43 is used to determine the target incentive strategy based on the enterprise churn score and the individual churn score using a preset learning model. The target incentive strategy is used to incentivize the target enterprise and all individual customers.

[0108] The incentive strategy determination device for payroll processing provided in this application adopts an integrated spatiotemporal feature analysis and intelligent strategy optimization approach. Through multi-scale sliding window convolution, it accurately identifies the churn risk of corporate and individual customers in payroll processing, thereby achieving the technical effect of intelligent and personalized dynamic adjustment of incentive strategies. This solves the technical problems of low customer retention and lagging intervention in bank payroll processing. Specifically, under the premise of ensuring data security and compliance, it collects payroll processing data from authorized enterprises, including enterprise information, individual customer information, and payroll transaction details. Multi-scale sliding windows are used to extract cross-cycle behavioral characteristics of enterprises and individuals. Then, combined with risk entropy feedback, enterprise churn scores and individual churn scores are calculated. These risk scores are then input into a pre-trained reinforcement learning model to generate customized incentive strategies for enterprises and individuals, such as fee discounts and wealth management product recommendations, to respond to customer behavior in real time, improve fund retention and customer value conversion, and avoid the limitations of traditional fixed threshold strategies.

[0109] Optionally, the first determining unit includes: a first extraction module, used to extract features from the payroll data to obtain target tensor data, wherein the target tensor data includes: a time series, and each payroll time in the time series corresponds to payroll business features; the first determining module is used to determine the periodic features corresponding to each period based on the target tensor data using multi-scale sliding window convolution, wherein the period includes: a first period, a second period, and a third period, the period length of the first period is less than the period length of the second period, the period length of the second period is less than the period length of the third period, and the periodic features include at least: mean, difference, and ratio.

[0110] Optionally, the first extraction module includes: a first extraction submodule, used to extract features from the payroll data to obtain three-dimensional tensor data, wherein the three-dimensional tensor data includes: time dimension data, enterprise dimension data, and payroll business feature dimension data. The time dimension data includes a time index generated by discretizing the timestamps in the payroll data. The enterprise dimension data includes a time series associated with the payroll accounts of the target enterprise. The payroll business feature dimension data includes statistical features and structural features. A first compression submodule is used to compress the three-dimensional tensor data using a preset encoder to obtain target tensor data, wherein the preset encoder is an encoder trained by compressing and decompressing tensor data. The compressed tensor data is data compressed by the preset encoder, and the decompressed tensor data is data decoded by the decoder.

[0111] Optionally, the first determining module includes: a first processing submodule, used to process the target tensor data to obtain periodic data under different periods; and a second processing submodule, used to process the periodic data using a preset convolution function for each period to obtain periodic features.

[0112] Optionally, the second determining unit includes: a second determining module, used to determine a first probability distribution of payout funds based on payout business data, wherein payout funds refer to funds paid by an enterprise to individual customers through a financial institution; a third determining module, used to determine a second probability distribution of account funds based on payout business data, wherein account funds refer to funds stored in different accounts after an individual customer receives the payout funds; a fourth determining module, used to determine the enterprise distribution entropy based on the first probability distribution and the individual distribution entropy based on the second probability distribution; a fifth determining module, used to determine the enterprise churn score based on the enterprise distribution entropy and the periodic characteristics corresponding to different periods using a preset multi-window feature fusion function; and a sixth determining module, used to determine the individual churn score based on the individual distribution entropy and the periodic characteristics corresponding to different periods using a preset multi-window feature fusion function.

[0113] Optionally, the incentive strategy determination device further includes: a first abstraction module, used to abstract corporate payroll accounts, individual customer accounts, and external accounts as nodes after determining the corporate churn score of the target enterprise and the individual churn score of the individual customer; a second abstraction module, used to abstract the transaction behavior into directed edges between the nodes indicated by the accounts when there is transaction behavior between the accounts; a seventh determination module, used to determine the node characteristics of each node, wherein the node characteristics include at least: transaction behavior characteristics and periodic behavior characteristics, the transaction behavior characteristics include at least one of the following: payroll amount, balance change, and the periodic behavior characteristics include at least one of the following: corporate churn score, individual churn score, corporate distribution entropy, and individual distribution entropy; and a first construction module, used to construct a dynamic network graph of fund flow based on all nodes, the node characteristics corresponding to each node, and all directed edges.

[0114] Optionally, the incentive strategy determination device further includes: a first statistics module, used to construct a dynamic network graph of fund flow based on all nodes, the node characteristics corresponding to each node, and all directed edges, and then to calculate the cumulative weight of each fund path based on the dynamic network graph of fund flow, and determine the total cumulative weight based on the cumulative weight of all fund paths; an eighth determination module, used to determine all node fund paths of each node, wherein a node fund path refers to a fund path starting from a node; a ninth determination module, used to determine the node score of a node based on the cumulative weight of each node's fund path and the total cumulative weight; and a tenth determination module, used to determine a node as a preset transfer-out node if the node score is greater than a preset threshold, and to construct a fund risk network based on all node fund paths of the preset transfer-out node.

[0115] Optionally, the third determining unit includes: an eleventh determining module, used to determine the initial incentive strategy set; a first processing module, used to process the enterprise churn score, individual churn score, financial risk network, and initial incentive strategy for each initial incentive strategy in the initial incentive strategy set using a preset learning model to obtain the strategy reward value of the initial incentive strategy, wherein the preset learning model is an interactive model of state, action, and reward, the state includes: enterprise churn score, individual churn score, and financial risk network, the action refers to the initial incentive strategy, and the reward refers to the strategy reward value, which is determined based on retention rate, cash outflow rate, and product acceptance rate; and a twelfth determining module, used to determine the initial incentive strategy indicated by the maximum strategy reward value as the target incentive strategy.

[0116] The above-mentioned excitation strategy determination device may also include a processor and a memory. The acquisition unit 40, the first determination unit 41, the second determination unit 42, the third determination unit 43, etc. are all stored in the memory as program units, and the processor executes the above-mentioned program units stored in the memory to realize the corresponding functions.

[0117] The aforementioned processor contains a kernel, which retrieves the corresponding program units from memory. One or more kernels can be configured, and by adjusting kernel parameters, a target incentive strategy is determined using a pre-defined learning model based on enterprise churn scores and individual churn scores.

[0118] The aforementioned memory may include non-permanent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.

[0119] It should be noted that the aforementioned acquisition unit 40, first determining unit 41, second determining unit 42, and third determining unit 43 correspond to steps S201 to S204 in Embodiment 1. The instances and application scenarios implemented by these units and their corresponding steps are the same, but they are not limited to the content disclosed in Embodiment 1. It should be noted that the aforementioned units can be hardware or software components stored in a memory (e.g., memory 104) and processed by one or more processors (e.g., processors 102a, 102b, ..., 102n). These units can also be part of a device and run in the computer terminal 10 provided in Embodiment 1.

[0120] Example 3

[0121] Embodiments of this application may provide an electronic device. Figure 5 This is a structural block diagram of an electronic device according to an embodiment of the present invention. Figure 5 As shown, the electronic device may include: one or more ( Figure 5 (Only one is shown) processor 502, memory 504, memory controller, and peripheral interface, wherein the peripheral interface is connected to the radio frequency module, audio module and display.

[0122] The memory can be used to store software programs and modules, such as the program instructions / modules corresponding to the incentive strategy determination method and apparatus for payroll services in this embodiment. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory, thereby implementing the aforementioned incentive strategy determination method for payroll services. The memory may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory may further include memory remotely located relative to the processor, and these remote memories can be connected to the terminal via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0123] The processor can access information and applications stored in the memory via a transmission device to execute the following steps: With authorization from the target company, collect the target company's payroll business data, which includes at least: company data, individual customer data, and payroll transaction data; individual customers are those whose salaries are paid through financial institutions. Based on the payroll business data, use multi-scale sliding window convolution to determine periodic features corresponding to multiple periods, where the intervals between different periods are inconsistent. Based on the payroll business data and the periodic features corresponding to different periods, determine the target company's churn score and the individual customer's churn score. Based on the company churn score and the individual churn score, use a pre-set learning model to determine a target incentive strategy, which is used to incentivize the target company and all individual customers.

[0124] The processor can access the information and application programs stored in the memory via the transmission device to perform the following steps: extracting features from the agency business data to obtain target tensor data, wherein the target tensor data includes: a time series, and each agency delivery time in the time series corresponds to agency business features; based on the target tensor data, using multi-scale sliding window convolution to determine the periodic features corresponding to each period, wherein the period includes: a first period, a second period, and a third period, the period length of the first period is less than the period length of the second period, the period length of the second period is less than the period length of the third period, and the periodic features include at least: mean, difference, and ratio.

[0125] The processor can access information and applications stored in the memory via a transmission device to perform the following steps: Extracting features from the payroll data to obtain three-dimensional tensor data. This three-dimensional tensor data includes: time dimension data, enterprise dimension data, and payroll business feature dimension data. The time dimension data contains a time index generated by discretizing the timestamps in the payroll data. The enterprise dimension data contains a time series associated with the payroll accounts of the target enterprise. The payroll business feature dimension data includes statistical and structural features. The three-dimensional tensor data is then compressed using a preset encoder to obtain target tensor data. The preset encoder is trained by compressing and decompressing the tensor data. The compressed tensor data is data compressed by the preset encoder, and the decompressed tensor data is data decoded by the decoder.

[0126] The processor can access the information and application programs stored in the memory via the transmission device to perform the following steps: process the target tensor data to obtain periodic data under different periods; for each period, process the periodic data using a preset convolution function to obtain periodic features.

[0127] The processor can access information and applications stored in the memory via a transmission device to execute the following steps: Based on payroll data, determine a first probability distribution of payroll funds, where payroll funds refer to funds paid by the enterprise to individual customers through financial institutions; based on payroll data, determine a second probability distribution of account funds, where account funds refer to funds stored in different accounts by individual customers after receiving payroll funds; based on the first probability distribution, determine the enterprise distribution entropy, and based on the second probability distribution, determine the individual distribution entropy; based on the enterprise distribution entropy and the cycle characteristics corresponding to different cycles, determine the enterprise churn score using a preset multi-window feature fusion function; based on the individual distribution entropy and the cycle characteristics corresponding to different cycles, determine the individual churn score using a preset multi-window feature fusion function.

[0128] The processor can invoke information and applications stored in memory via a transmission device to perform the following steps: abstracting corporate payroll accounts, individual customer accounts, and external accounts as nodes; when there are transaction behaviors between accounts, abstracting the transaction behaviors as directed edges between nodes indicated by the accounts; determining the node characteristics of each node, wherein the node characteristics include at least: transaction behavior characteristics and periodic behavior characteristics, the transaction behavior characteristics including at least one of the following: payroll amount, balance change, and the periodic behavior characteristics including at least one of the following: corporate churn points, individual churn points, corporate distribution entropy, and individual distribution entropy; and constructing a dynamic network graph of fund flow based on all nodes, the node characteristics corresponding to each node, and all directed edges.

[0129] The processor can access information and applications stored in the memory via a transmission device to execute the following steps: Based on the dynamic network diagram of fund flows, calculate the cumulative weight of each fund path and determine the total cumulative weight based on the cumulative weight of all fund paths; determine all node fund paths for each node, where a node fund path refers to a fund path starting from a node; determine the node score of a node based on the cumulative weight of each node fund path and the total cumulative weight; if the node score is greater than a preset threshold, designate the node as a preset transfer-out node, and construct a fund risk network based on all node fund paths of the preset transfer-out node.

[0130] The processor can access information and applications stored in memory via a transmission device to execute the following steps: determining an initial incentive strategy set; for each initial incentive strategy in the initial incentive strategy set, using a preset learning model to process the enterprise churn score, individual churn score, financial risk network, and the initial incentive strategy to obtain the strategy reward value of the initial incentive strategy, wherein the preset learning model is an interactive model of state, action, and reward, the state includes: enterprise churn score, individual churn score, and financial risk network, the action refers to the initial incentive strategy, and the reward refers to the strategy reward value, which is determined based on retention rate, cash outflow rate, and product acceptance rate; and determining the initial incentive strategy indicated by the maximum strategy reward value as the target incentive strategy.

[0131] This application provides a scheme for determining incentive strategies for payroll processing services. By integrating spatiotemporal feature analysis and intelligent strategy optimization, and employing multi-scale sliding window convolution, it accurately identifies the churn risk of corporate and individual customers in payroll processing services. This achieves the technical effect of dynamically adjusting intelligent and personalized incentive strategies, thereby solving the technical problems of low customer retention and lagging intervention in bank payroll processing services. Specifically, under the premise of ensuring data security and compliance, payroll processing data from authorized enterprises is collected, including enterprise information, individual customer information, and payroll transaction details. Multi-scale sliding windows are used to extract cross-cycle behavioral characteristics of enterprises and individuals. Then, combined with risk entropy feedback, enterprise churn scores and individual churn scores are calculated. These risk scores are then input into a pre-trained reinforcement learning model to generate customized incentive strategies for enterprises and individuals, such as fee discounts and wealth management product recommendations, to respond to customer behavior in real time, improve fund retention and customer value conversion, and avoid the limitations of traditional fixed threshold strategies.

[0132] Those skilled in the art will understand that Figure 5 The structure shown is for illustrative purposes only. Electronic devices can also be terminal devices such as smartphones, tablets, PDAs, and mobile internet devices (MIDs). Figure 5 This does not limit the structure of the aforementioned electronic device. For example, electronic devices may also include components that are more... Figure 5 The more or fewer components shown (such as network interfaces, display devices, etc.), or having the same Figure 5 The different configurations shown.

[0133] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing the hardware related to the terminal device. The program can be stored in a computer-readable storage medium, which may include: flash drive, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0134] Example 4

[0135] Embodiments of this application also provide a storage medium. Optionally, in this embodiment, the storage medium can be used to store the program code executed by the incentive strategy determination method for payroll services provided in Embodiment 1.

[0136] Optionally, in this embodiment, the storage medium may be located in any computer terminal in a group of computer terminals in a computer network, or in any mobile terminal in a group of mobile terminals.

[0137] This application also provides a computer program product that, when executed on a data processing device, is adapted to perform the steps of a method for determining an incentive strategy for payroll services.

[0138] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0139] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0140] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0141] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0142] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0143] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard drive, magnetic disk, or optical disk.

[0144] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for determining incentive strategies for payroll services, characterized in that, include: With the authorization of the target company, the payroll business data of the target company is collected, wherein the payroll business data includes at least: company data, individual customer data, and payroll transaction data, and the individual customers are the customers of the target company who receive their salaries through financial institutions. Based on the aforementioned payroll service data, a multi-scale sliding window convolution is used to determine the periodic features corresponding to multiple periods, wherein the interval duration of different periods is inconsistent. Based on the payroll business data and the cycle characteristics corresponding to different cycles, the corporate churn score of the target enterprise and the individual churn score of the individual customer are determined. Based on the enterprise churn score and the individual churn score, a preset learning model is used to determine a target incentive strategy, wherein the target incentive strategy is used to incentivize the target enterprise and all the individual customers.

2. The method for determining the incentive strategy according to claim 1, characterized in that, Based on the aforementioned payroll service data, the step of determining the periodic features corresponding to multiple periods using multi-scale sliding window convolution includes: Feature extraction is performed on the delivery service data to obtain target tensor data, wherein the target tensor data includes: a time series, and each delivery time in the time series corresponds to a delivery service feature; Based on the target tensor data, the periodic features corresponding to each period are determined by the multi-scale sliding window convolution. The period includes a first period, a second period, and a third period. The period length of the first period is less than the period length of the second period, and the period length of the second period is less than the period length of the third period. The periodic features include at least the mean, difference, and ratio.

3. The method for determining the incentive strategy according to claim 2, characterized in that, The steps of extracting features from the payroll service data to obtain target tensor data include: Feature extraction is performed on the payroll data to obtain three-dimensional tensor data, wherein the three-dimensional tensor data includes: time dimension data, enterprise dimension data, and payroll business feature dimension data. The time dimension data includes a time index generated by discretizing the timestamps in the payroll data. The enterprise dimension data includes a time series associated with the payroll account of the target enterprise. The payroll business feature dimension data includes statistical features and structural features. The target tensor data is obtained by compressing the three-dimensional tensor data using a preset encoder. The preset encoder is an encoder trained by compressing and decompressing tensor data. The compressed tensor data is data compressed by the preset encoder, and the decompressed tensor data is data decoded by the decoder.

4. The method for determining the incentive strategy according to claim 2, characterized in that, The step of determining the periodic features corresponding to each period based on the target tensor data using the multi-scale sliding window convolution includes: The target tensor data is processed to obtain periodic data under different periods; For each period, the periodic data is processed using a preset convolution function to obtain the periodic features.

5. The method for determining the incentive strategy according to claim 1, characterized in that, The steps for determining the corporate churn score of the target enterprise and the individual churn score of the individual customer based on the payroll service data and the cycle characteristics corresponding to different cycles include: Based on the aforementioned payroll business data, a first probability distribution of payroll funds is determined, wherein the payroll funds refer to funds that enterprises pay to individual customers through financial institutions. Based on the payroll data, a second probability distribution of account funds is determined, wherein the account funds refer to the funds stored in different accounts after an individual customer receives the payroll funds. Based on the first probability distribution, determine the enterprise distribution entropy, and based on the second probability distribution, determine the individual distribution entropy; Based on the enterprise distribution entropy and the cycle characteristics corresponding to different cycles, the enterprise churn score is determined by a preset multi-window feature fusion function. Based on the individual distribution entropy and the periodic features corresponding to different periods, the individual loss score is determined using the preset multi-window feature fusion function.

6. The method for determining the incentive strategy according to claim 1, characterized in that, After determining the corporate churn score of the target enterprise and the individual churn score of the individual customer, the process also includes: Abstract corporate payroll accounts, personal customer accounts, and external accounts into nodes; In the case of transactions between accounts, the transactions are abstracted into directed edges between the nodes indicated by the accounts; Determine the node characteristics of each node, wherein the node characteristics include at least: transaction behavior characteristics and periodic behavior characteristics, wherein the transaction behavior characteristics include at least one of the following: disbursement amount and balance change, and the periodic behavior characteristics include at least one of the following: enterprise churn points, individual churn points, enterprise distribution entropy, and individual distribution entropy; Based on all the nodes, the node features corresponding to each node, and all the directed edges, a dynamic network graph of fund flow is constructed.

7. The method for determining incentive strategies according to claim 6, characterized in that, After constructing a dynamic network graph of fund flow based on all the nodes, the node features corresponding to each node, and all the directed edges, the method further includes: Based on the aforementioned dynamic network diagram of fund flows, the cumulative weight of each fund path is calculated, and the total cumulative weight is determined based on the cumulative weight of all the aforementioned fund paths. Determine all node funding paths for each node, wherein the node funding path refers to the funding path starting from the node mentioned above; The node score is determined based on the cumulative weight of each node's funding path and the total cumulative weight. If the node score is greater than a preset threshold, the node is identified as a preset transfer-out node, and a fund risk network is constructed based on all the fund paths of the preset transfer-out node.

8. The method for determining incentive strategies according to claim 7, characterized in that, The steps for determining the target incentive strategy based on the enterprise churn score and the individual churn score using a preset learning model include: Determine the initial set of incentive strategies; For each initial incentive strategy in the initial incentive strategy set, the preset learning model is used to process the enterprise churn score, the individual churn score, the financial risk network, and the initial incentive strategy to obtain the strategy reward value of the initial incentive strategy. The preset learning model is an interactive model of state, action, and reward. The state includes: the enterprise churn score, the individual churn score, and the financial risk network. The action refers to the initial incentive strategy, and the reward refers to the strategy reward value. The strategy reward value is determined based on retention rate, cash outflow rate, and product acceptance rate. The initial incentive strategy indicated by the maximum strategy reward value is determined as the target incentive strategy.

9. A device for determining incentive strategies for payroll services, characterized in that, include: The data collection unit is used to collect the payroll business data of the target enterprise when authorized by the target enterprise. The payroll business data includes at least: enterprise data, individual customer data, and payroll transaction data. Individual customers are customers of the target enterprise who receive their salaries through financial institutions. The first determining unit is used to determine the periodic features corresponding to multiple periods based on the agency business data using multi-scale sliding window convolution, wherein the interval duration of different periods is inconsistent. The second determining unit is used to determine the corporate churn score of the target enterprise and the individual churn score of the individual customer based on the payroll business data and the period characteristics corresponding to different periods. The third determining unit is used to determine a target incentive strategy based on the enterprise churn score and the individual churn score using a preset learning model, wherein the target incentive strategy is used to incentivize the target enterprise and all the individual customers.

10. An electronic device, characterized in that, It includes one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the incentive strategy determination method for payroll services as described in any one of claims 1 to 8.

Images