Commission settlement abnormal data detection method and system based on gradient boosting decision tree

Abstract

The present application relates to the technical field of data processing, and discloses a commission settlement abnormal data detection method and system based on gradient boosting decision tree, comprising: extracting key positioning fields from repayment records to construct scheme matching coding vectors, calculating monthly recovery rates and performing pre-line verification, performing double-account period routing and commission calculation according to the scheme to generate settlement detail data table for the records passing the verification; constructing project monthly recovery rate stepwise line graph and calculating total projection area, stripping contaminated shadow area in the graph to construct an abnormal judgment feature vector, inputting the abnormal judgment feature vector into a pre-constructed gradient boosting decision tree model to output an abnormal score value to determine abnormal correction events, and then retrieving affected records to perform batch re-accounting and account fuse disposal; the present application expands the accounting of single correction data to global rate pollution quantification combined with time dimension, and realizes effective detection of abnormal behavior of short-time account padding global recovery rate level and accurate recovery of affected funds.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and more specifically, to a method and system for detecting abnormal commission settlement data based on gradient boosting decision trees. Background Technology

[0002] In modern business scenarios involving debt collection or agency sales, commission settlement often relies on multi-dimensional dynamic data and complex tiered fee rules. The settlement system needs to be able to accurately identify and handle various abnormal transactions to ensure the accuracy of fund accounting. Existing technologies, such as the Chinese patent application CN120069951A, disclose an automated commission tracing and inspection method, apparatus, and computer-readable storage medium. This technology preprocesses historical commission data obtained from data sources, extracts target feature information with comprehensive relevance ranking using various statistical analysis methods, and inputs it into an initial commission tracing prediction model constructed using a gradient boosting decision tree machine learning algorithm for training and optimization. Finally, data records whose prediction results are outside the preset range are marked as abnormal data records and output. For example, Chinese patent application CN119067698A discloses a method and related device for detecting anomalies in spot electricity market settlement data. This technology collects historical settlement data and calculates feature correlation information, performs preprocessing, filters out settlement feature data with strong correlation, constructs an anomaly detection model based on random forest and extreme gradient boosting XGBoost models, and outputs the classification results of abnormal settlement data with strong correlation as input.

[0003] However, in actual accounting operations and settlement auditing scenarios, settlement specialists often encounter a hidden settlement anomaly that cannot be detected through routine reconciliation when handling commission projects with tiered fee rates: At the final stage of the month-end settlement cycle, a large sum may suddenly appear in the transaction history of certain suppliers, causing the monthly recovery rate of that asset package to instantly jump to a high commission tier. This prompts the system to complete batch commission settlement for a large number of parallel genuine repayment orders within the same batch, based on this high commission rate. However, after this batch of settlements is processed, the large sum is immediately reversed by a refund under the pretext of "misoperation." At this point, the existing daily reconciliation mechanism can only trigger the reversal of a single transaction, deducting the original commission for the revoked amount. But during the short period that this forged fund survives in the system, it inflates the "monthly recovery rate level" of the entire asset package like a rising tide, causing all other genuine orders calculated during this period of artificially high levels to incorrectly receive commissions at the high tier. If the existing point-to-point verification system fails to detect and quantify the cross-order contamination caused by this single data point on the overall fee rate status, the commissions that are mistakenly generated due to the inflated system water level will not be automatically recovered. This will ultimately lead to the settlement platform bearing the hidden loss of funds, and may even evolve into suppliers using this rule loophole for organized systemic arbitrage. Summary of the Invention

[0004] This invention is mainly applied to multi-level commission settlement and reconciliation audit scenarios in the disposal of non-performing financial assets, agency sales collection, or quasi-financial services. It is suitable for large-scale transaction settlement systems that include dynamic performance evaluation indicators (such as monthly recovery rate and monthly collection amount) in business rules and implement a tiered fee rate linkage mechanism.

[0005] In a tiered commission settlement architecture, the monthly recovery rate serves as a core prerequisite for determining the fee tier, and its value exhibits a monotonically non-decreasing dynamic change as daily repayments accumulate. When a large abnormal fund at the month-end threshold is mixed into the incoming data, the system recalculates the global recovery rate based on the actual incoming amount, driving the overall fee rate to a higher tier. Existing reversal and rebalancing mechanisms are limited to a discrete, single-record dimension, with their logic chain only covering "identifying reversal instructions—locking the original incoming order—executing single-transaction amount reconciliation and commission deduction." This static, point-to-point accounting method assumes that the impact of a reversal event is limited to the single transaction itself. The flaw in this logic lies in the fact that a large abnormal fund, within the survival period from its arrival to its reversal, objectively alters the system's global fee rate calculation benchmark at that time. During this high benchmark period, other compliant repayment records processed concurrently by the system all use the fee rate parameters raised by the abnormal fund to complete the settlement. When a single reversal is executed, although the physical trace of the original large-amount record is erased, the resulting rate spike during its lifespan is already solidified in the settlement results of other orders. This type of anomaly exploits the blind spot in the time decoupling between individual fund flows and global state calculations, transforming the anomalous amount of a single data entry into time-dimensional rate pollution across orders. Conventional amount reconciliation mechanisms cannot detect this hidden path of fund embezzlement based on time-series state tampering.

[0006] To overcome the aforementioned deficiencies of existing technologies, this invention provides a method and system for detecting abnormal commission settlement data based on gradient boosting decision trees. The method extracts key location fields to construct a scheme matching encoding vector to determine the target settlement scheme. Based on the settlement details data table, it constructs a step-wise line chart of the project's monthly recovery rate and calculates the total projected area. In the step-wise line chart of the project's monthly recovery rate, the area of ​​contamination caused by short-term inflows inflating the water level is removed to construct an anomaly judgment feature vector. This vector is input into a gradient boosting decision tree model pre-built based on machine learning algorithms to output an anomaly score value, thereby identifying abnormal reversal events. This scheme quantifies the global rate contamination of a single reversal data point combined with the time dimension, enabling batch recalculation of affected repayment records and circuit breaker handling of commission accounts, ensuring the rigor of the settlement and billing logic and the security of business funds.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A method for detecting abnormal commission settlement data based on gradient boosting decision trees includes: Extract key positioning fields from repayment records to construct scheme matching encoding vectors. Match the scheme matching encoding vectors with the settlement scheme library to determine the target settlement scheme. Calculate the monthly recovery rate of the project and perform recovery rate front-line verification. For repayment records that pass the recovery rate front-line verification, perform dual-account period routing and commission calculation based on the target settlement scheme to generate a settlement detail data table. Based on the settlement details data table, a step-line chart of the project's monthly recovery rate is constructed, and the total projected area is calculated. The pollution shadow area is removed from the project's monthly recovery rate step-line chart. An anomaly judgment feature vector is constructed based on the pollution shadow area and the total projected area. The anomaly judgment feature vector is input into a pre-constructed gradient boosting decision tree model to output anomaly score values. Anomaly reversal events are identified and judged based on the anomaly score values. For anomaly reversal events, the affected repayment records are retrieved, and batch recalculation of the affected repayment records and commission account circuit breaker processing are performed.

[0008] The key positioning fields include at least the institution name, asset package name, institution tag type, and settlement month; The method for constructing the matching encoding vector includes: mapping the institution name to the institution code, the settlement month to the month code, the asset package name to the asset package code, and the institution tag type to the tag type code, and concatenating the matching encoding vector in a fixed order of institution code, month code, asset package code, and tag type code.

[0009] The method for determining the target settlement plan includes: The scheme matching code vector is compared one by one with the pre-configured scheme code vectors in the settlement scheme library. When the scheme matching code vector is completely consistent with the scheme code vector of a certain settlement scheme in the settlement scheme library, the settlement scheme with complete consistency is determined as the target settlement scheme. When the scheme matching code vector is inconsistent with all scheme code vectors in the settlement scheme library, the repayment record is routed to the backup settlement scheme and a scheme missing alarm signal is issued.

[0010] The method for performing a recovery rate front-line verification and routing repayment records that pass the recovery rate front-line verification according to the target settlement plan includes: The monthly recovery rate of the project is compared with the preset recovery rate threshold. If the monthly recovery rate of the project is greater than or equal to the recovery rate threshold, it is determined that the recovery rate threshold verification is passed. When the recovery rate threshold verification is passed, the commission settlement scenario identifier is read from the repayment record, and the repayment record is routed to the daily settlement rate matching branch or the monthly settlement rate matching branch according to the commission settlement scenario identifier.

[0011] The method for generating a settlement details data table by performing dual-billing period routing and commission calculation includes: In the daily settlement rate matching branch, the daily settlement rate is extracted from the target settlement plan, the amount credited to the repayment record is obtained, and the daily settlement commission and frozen commission are calculated based on the amount credited and the daily settlement rate. In the monthly settlement rate matching branch, determine the monthly settlement rate and calculate the monthly settlement commission based on the incoming amount, daily settlement commission, and monthly settlement rate; write the daily settlement commission, frozen commission, monthly settlement commission, along with the incoming amount, incoming time, monthly settlement rate, and total monthly settlement amount into the settlement details data table.

[0012] The method further includes: Before commission calculation is executed, the input parameter range is validated on the incoming amount. After commission calculation is executed, the output parameter range is validated on the daily commission, frozen commission, and monthly commission. Repayment records that fail the input or output parameter range validation are marked as abnormal and an alarm signal is issued.

[0013] The method for constructing a step-wise line graph of the monthly recovery rate of the project and calculating the total projected area includes: The horizontal axis is constructed with the first day of the current settlement month as the starting point and the last day as the ending point. The cumulative amount of payment received for each day in the current settlement month is extracted from the settlement details data table. Based on the cumulative amount of payment received, the monthly recovery rate of the project is calculated for each day. A step line chart of the monthly recovery rate of the project is constructed with the day as the horizontal axis scale, the monthly recovery rate of the project as the vertical axis coordinate, and horizontal step line segments connecting the monthly recovery rate values ​​of adjacent days. Multiply the monthly project recovery rate for each calendar day by the time unit length of one calendar day to obtain the daily projected area, and then accumulate the projected area of ​​the monthly project recovery rate step line graph and the horizontal axis to calculate the total projected area of ​​the enclosed area.

[0014] The method for removing the pollution shadow area in the project's monthly recovery rate step line chart includes: Repayment records with associated reversal records are identified from the repayment records and used as original repayment records, marked as reversal events. The posting time and reversal time are read from the original repayment record and the corresponding reversal record, respectively. The survival time interval is determined by the left and right closed intervals from the posting time to the reversal time. The recovery rate increment is calculated. A pollution shadow rectangle is constructed in the project's monthly recovery rate step line chart with the survival time interval as the width and the recovery rate increment as the height. The pollution shadow area is calculated by multiplying the recovery rate increment by the number of natural days in the survival time interval.

[0015] The method further includes: The monthly project recovery rate is read from the project monthly recovery rate ladder line chart on the day before the accounting date as the pre-accounting monthly recovery rate, and the monthly project recovery rate on the accounting date is read as the post-accounting monthly recovery rate. The pre-accounting monthly settlement tier is determined based on the pre-accounting monthly recovery rate, and the post-accounting monthly settlement tier is determined based on the post-accounting monthly recovery rate. When the monthly settlement rate corresponding to the post-accounting monthly settlement tier is higher than the monthly settlement rate corresponding to the pre-accounting monthly settlement tier, a reversal event is determined, triggering a cross-tier jump and marked as a cross-tier reversal event.

[0016] The method for constructing the anomaly detection feature vector includes: An anomaly judgment feature vector is constructed for cross-grade reversal events. The anomaly judgment feature vector includes seven features, namely, the proportion of the pollution shadow area, the time distance at the end of the month, the cross-grade difference, the proportion of the reversal amount, the number of affected repayment records, the total amount of commission overflow, and the cumulative number of historical reversal operations. The pollution shadow area ratio is the ratio of the pollution shadow area to the total projected area.

[0017] The method for determining abnormal reversal events is as follows: When the abnormal score value is greater than or equal to the preset abnormal judgment threshold, the cross-grade reversal event will be judged as an abnormal reversal event.

[0018] A commission settlement anomaly detection system based on gradient boosting decision trees is provided to implement the aforementioned commission settlement anomaly detection method based on gradient boosting decision trees. The system includes: Settlement details generation module: used to extract key positioning fields from repayment records to construct scheme matching code vectors, match the scheme matching code vectors with the settlement scheme library to determine the target settlement scheme, calculate the monthly recovery rate of the project and perform recovery rate front-line verification, and perform dual-account period routing and commission calculation on repayment records that pass the recovery rate front-line verification according to the target settlement scheme to generate a settlement details data table; Anomaly feature construction module: Based on the settlement details data table, construct a step line chart of the project's monthly recovery rate and calculate the total projected area. Remove the pollution shadow area from the project's monthly recovery rate step line chart and construct anomaly judgment feature vector based on the pollution shadow area and the total projected area. Anomaly detection and handling module: This module is used to input anomaly detection feature vectors into a pre-built gradient boosting decision tree model to output anomaly scores. Based on the anomaly scores, it identifies and determines anomaly reversal events. For anomaly reversal events, it retrieves affected repayment records and performs batch recalculation of affected repayment records and circuit breaker processing of commission accounts.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: To address the issue of existing single-transaction reconciliation mechanisms neglecting the impact of individual reversal data on the overall fee rate status, thus causing temporal contamination, this invention generates a settlement detail data table from repayment records. This table then constructs a stepped line chart of the project's monthly recovery rate and calculates the total projected area, expressing the dynamic changes in the monthly recovery rate level in a geometric figure in conjunction with its survival time. Furthermore, by removing the contaminated shadow area from the stepped line chart and extracting feature vectors, this invention quantifies the impact of a single reversal data point on the overall recovery rate of the same batch of projects during its survival time within the system as an area feature. This feature is then input into a gradient boosting decision tree model for anomaly detection. This solution overcomes the limitations of traditional static methods that rely on flattening amounts for individual repayment records. It leverages machine learning's strengths in handling high-dimensional data features and recognizing complex patterns to identify abnormal reversal events where large inflows at the end of the month are used to fabricate high-level recovery rate tiers to obtain exorbitant commissions. By locating affected repayment records and performing batch recalculation and account suspension, it blocks and recovers commission overflows from other genuine orders that were settled at incorrectly high rates during periods of inflated rates, thus ensuring the security of funds in the settlement business and the rigor of the billing logic. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart of a commission settlement anomaly data detection method based on gradient boosting decision tree provided in an embodiment of the present invention; Figure 2 A schematic diagram of the matching encoding vector for the key positioning field encoding mapping splicing scheme provided in the embodiments of the present invention; Figure 3 This is a schematic diagram of the daily commission revenue sharing structure provided in an embodiment of the present invention; Figure 4 This is a flowchart of the input and output parameter range verification provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the total projected area of ​​the monthly recovery rate step line chart provided in this embodiment of the invention; Figure 6 This is a schematic diagram of gradient boosting decision tree model inference provided in an embodiment of the present invention; Figure 7 This is a functional block diagram of the commission settlement anomaly data detection system based on gradient boosting decision tree provided in an embodiment of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example

[0023] Please see Figure 1 As shown, this embodiment provides a method for detecting abnormal commission settlement data based on gradient boosting decision trees, including: Step S10: Extract key positioning fields from repayment records to construct a scheme matching encoding vector, match the scheme matching encoding vector with the settlement scheme library to determine the target settlement scheme, calculate the monthly recovery rate of the project and perform recovery rate front-line verification, and perform dual billing cycle routing and commission calculation on repayment records that pass the recovery rate front-line verification according to the target settlement scheme to generate a settlement detail data table. Further, step S10 includes: Step S11: Extract the institution name, project name, asset package name, institution tag type, settlement month, and deposit account from the repayment records as key positioning fields. Map the institution name to the institution code, the settlement month to the month code, the asset package name to the asset package code, and the institution tag type to the tag type code. Concatenate the code vector according to the fixed order of institution code, month code, asset package code, and tag type code to construct the matching code vector. In step S11, the repayment record is a structured data record submitted by the supplier to the settlement system after each customer repayment is completed in the collection business. A repayment record contains multiple fields, including: the institution name is the full registered name of the supplier submitting the repayment record; the project name is the identifier of the debt project to which the repayment record belongs; the asset package name is the identifier of the debt asset package to which the repayment record belongs; the institution tag type is the classification label of the supplier's business type by the settlement system, and the value range is limited to three types: telephone collection, field visits, and legal proceedings; and the settlement month is the identifier of the settlement cycle to which the repayment record belongs, denoted by yyyym. The format is a six-digit fixed-length year and month in m format, such as 202603. The receiving account is the account identifier for receiving repayment funds. The system pre-maintains a mapping table between receiving accounts and invoicing methods, associating each receiving account with the corresponding invoicing method. The invoicing method can be either a special invoice or a general invoice. The system also maintains a mapping table between suppliers and tax points, associating the name of each supplier with the actual tax point value agreed in its business contract. The supplier's tax point and the invoicing method of the receiving account are used together for the tax point adjustment calculation in the monthly settlement rate matching branch of step S14. The repayment record also includes a repayment order number and a commission settlement scenario identifier. The repayment order number is a unique identifier for the repayment record in the settlement system, assigned by the system according to a time sequence rule when the repayment record is generated. Positive value entries corresponding to the same original repayment and subsequent negative value reversals share the same repayment order number, so that in step S22, the original repayment record and its reversal record can be linked by the order number. The entry amount is the amount of funds entered this time; a positive value indicates funds entering the account, and a negative value indicates funds outflow due to a reversal. The entry time is the system registration time of the funds entering the account. The commission settlement scenario identifier is the settlement period type identifier corresponding to the repayment record, filled in by the supplier according to the business contract when submitting the repayment record. The value range is limited to one of two types: daily settlement and monthly settlement. Daily settlement means that the repayment record is settled on the day of entry, while monthly settlement means that the repayment record is settled uniformly at the end of the month according to the tiered fee rate. The process of constructing the scheme matching encoding vector is as follows: The system pre-maintains a mapping table between institution names and institution codes, a mapping table between asset package names and asset package codes, and a mapping table between institution tag types and tag type codes. The institution name in the repayment record is converted to an institution code through the institution name mapping table. The settlement month is directly used as the month code, as it is already in a structured year-month format. The asset package name is converted to an asset package code through the asset package name mapping table. The institution tag type is converted to a tag type code through the tag type mapping table. The four codes are concatenated in a fixed order: institution code, month code, asset package code, and tag type code, forming a fixed-length encoding sequence, i.e., the scheme matching encoding vector. For example, refer to Table 1 for the encoding mapping process of a repayment record: Table 1. Encoding Mapping Process of Repayment Records Organization Name A Digital Technology Co., Ltd. 000000 Settlement month 202512 202512 Asset Package Name (Asset package coded 110) 110 Institutional label types Electrocatalysis 1 See Figure 2 This is a schematic diagram of the matching encoding vector of the key positioning field encoding mapping splicing scheme provided in the embodiments of this application. Figure 2 The diagram illustrates the complete execution process of mapping key positioning fields to their corresponding codes and then concatenating them in a preset fixed order to form a scheme matching code vector. The concatenated scheme matching code vector is [000000, 202512, 110, 1]. The scheme matching code vector does not include project name codes and billing account codes for two reasons: Firstly, the same settlement scheme may cover multiple projects simultaneously. For example, a settlement scheme named "April X External Visit" includes dozens of projects such as Institution D, Institution E, Institution F, and Institution G, along with their corresponding rate configurations. If project names were included in the scheme matching code vector, each project would require a separate settlement scheme configuration, and the total number of schemes would increase linearly with the number of projects. In operation, the number of schemes in a month could expand from dozens to hundreds, increasing the configuration and maintenance burden and easily leading to configuration omissions. The project name matching is performed in step S... Step 14 is executed as a secondary retrieval. After the target settlement plan is determined in step S12, step S14 extracts the corresponding rate from the project plan entry list of the target settlement plan using the project name as the search key. This two-level positioning structure of "first determine the plan, then determine the project" separates the coarse positioning at the plan level from the fine positioning at the project level, maintaining the simplicity of the plan matching coding vector while retaining the flexibility of project-level rate differentiation configuration. On the other hand, the role of the receiving account is limited to determining the tax point adjustment rule based on the invoicing method in step S14, and is unrelated to the positioning logic of the settlement plan. Including it in the plan matching coding vector will introduce redundant dimensions without increasing the distinguishing ability of the plan positioning. By placing the coding mapping operation before the matching operation, each repayment record has obtained a code identifier with a fixed format and uniform length before entering the plan comparison stage. This eliminates text matching ambiguities caused by supplier name changes, names containing special characters, or different suppliers using similar names. This allows the subsequent plan comparison in step S12 to use a bit-by-bit comparison of the coding sequence rather than fuzzy text matching, reducing the probability of plan mismatch. If the encoding mapping and vector construction in step S11 are missing, step S12 will face the dilemma of using unstructured text fields as matching keys. The name of the same supplier in different repayment records may be inconsistent due to differences in entry, resulting in different repayment records of the same supplier being matched to different settlement schemes, causing confusion in scheme attribution.

[0024] Step S12: Compare the scheme matching code vector with the pre-configured scheme code vectors in the settlement scheme library one by one. When the scheme matching code vector is completely consistent with the scheme code vector of a certain settlement scheme in the settlement scheme library, the settlement scheme with complete consistency is determined as the target settlement scheme. When the scheme matching code vector is inconsistent with all scheme code vectors in the settlement scheme library, the repayment record is routed to the backup settlement scheme and a scheme missing alarm signal is issued. In step S12, the settlement scheme library is a set of schemes pre-configured by system administrators before the start of each settlement cycle and stored in the system database. Each settlement scheme includes a scheme name, scheme type (telegraphic collection, external visit, or legal action), whether it is marked as special, applicable month, and one or more project scheme entries. Each project scheme entry includes a list of project names, a monthly settlement tier calculation method, and corresponding rate configuration (including daily settlement rate and monthly settlement tier rate range table). The monthly settlement tier calculation method can be one of three: fixed, based on recovery rate, or based on the amount recovered. This method is used to specify the calculation basis for determining the monthly settlement tier rate in the monthly settlement rate matching branch in step S14. Each settlement scheme is pre-bound with a scheme code vector. The generation rule of the scheme code vector is consistent with the generation rule of the scheme matching code vector in step S11, that is, it is composed of the applicable institution code, applicable month code, applicable asset package code, and applicable tag type code concatenated in the same fixed order. The scheme matching process is as follows: The scheme matching code vector output in step S11 is compared bit by bit with the scheme code vectors of all settlement schemes in the settlement scheme library. This comparison is performed digit by digit: institution code with institution code, month code with month code, asset package code with asset package code, and tag type code with tag type code. A successful match is determined when all four codes match, and the matched settlement scheme is the target settlement scheme. Settlement schemes in the settlement scheme library are divided into two categories based on whether they are marked as special-purpose: regular schemes generate a scheme code vector using a single applicable institution code. Each regular scheme corresponds to a unique quadruple of institution, month, asset package, and tag type. Special-purpose schemes do not participate in the bit-by-bit comparison matching based on the scheme code vector in step S12. Instead, they are stored as an independent set. Each special-purpose scheme includes a list of applicable institution names, applicable months, applicable asset package codes, applicable tag type codes, and a list of project scheme entries. The matching and switching of special-purpose schemes are performed separately in the rate matching branch of step S14, using the institution name as the search key. After S12 confirms the regular scheme matched by the scheme coding vector as the target settlement scheme, step S14 then determines whether there is a special scheme covering the current institution and switches accordingly. When the scheme matching coding vector cannot match any of the scheme coding vectors in the settlement scheme library, the repayment record is routed to the backup settlement scheme. The backup settlement scheme is a conservative rate scheme pre-configured by the system, whose rate is lower than or equal to the lowest rate of all regular schemes in the settlement scheme library. The basis for setting it is: when a scheme configuration is missing, completing the settlement with a lower rate can avoid irreversible financial losses due to overpayment at a high rate. At the same time, the system issues a scheme missing alarm signal to notify the management personnel to supplement the missing scheme configuration before the next settlement batch.Different suppliers may have completely different rate policies in different months and under different asset packages. The bit-by-bit comparison of the scheme matching code vector locks each repayment record to a unique settlement scheme, ensuring that the settlement rules between different suppliers, between different months of the same supplier, and between different asset packages of the same supplier do not affect each other. The rate adjustment of one settlement scheme will not affect the repayment records covered by other schemes. The correspondence between the target settlement scheme and the repayment record used in step S14 when reading the rate is a one-to-one binding that has been verified bit by bit through coding, eliminating the risk of rate misuse caused by cross-reference of schemes. When step S20 detects whether the reversal event triggers a cross-tier jump, the monthly settlement tiered rate range table referenced comes from the isolated and confirmed target settlement scheme, so that anomaly detection and commission calculation are based on the same scheme configuration, and the judgment criteria of the two stages are consistent. If the scheme isolation matching in step S12 is missing, step S14 will not be able to determine which settlement scheme to extract the rate configuration from, the settlement process will be interrupted in the rate reading stage, or in the absence of isolation, the rates of multiple schemes will be confused and referenced, causing the commission calculation results to deviate from expectations.

[0025] Step S13: Calculate the total monthly settlement amount based on the organization name, project name and settlement month summary in the key positioning fields. Calculate the monthly project recovery rate based on the total monthly settlement amount. Compare the monthly project recovery rate with the recovery rate front line to perform recovery rate front line verification. When the recovery rate front line verification is passed, read the commission settlement scenario identifier from the repayment record and route the repayment record to the daily settlement rate matching branch or the monthly settlement rate matching branch based on the commission settlement scenario identifier. The condition for passing the recovery rate front line verification is: the project's monthly recovery rate is greater than or equal to the recovery rate front line.

[0026] In step S13, the pre-verification of the recovery rate is performed earlier than the settlement scenario routing. The purpose is to exclude repayment records whose recovery effect does not meet the basic requirements before the rate calculation. The calculation process of the monthly recovery rate of the project is as follows: From the settlement details data table, using the institution name and project name in the repayment record as the filtering conditions, the total amount of all credited repayment records in the settlement month is summarized, and the total amount of credited amounts is used as the total recovery amount for the settlement month; From the project basic principal data maintained by the system, using the project name in the repayment record as the query condition, the project basic principal is read. The rule for determining the value of the project basic principal is to take the larger value between the total debt-to-principal of the project at the beginning of the settlement month and the current total debt-to-principal of the project. "Total Principal Transferred to Debt Collection" refers to the total principal amount of debt transferred by the client to suppliers for collection. The larger value is used because: if there are withdrawals or partial debt recovery from suppliers during the month, the current total principal transferred to debt collection may be less than the initial total principal transferred to debt collection at the beginning of the month. Using the larger value as the denominator avoids artificially inflating the recovery rate due to a smaller denominator. Similarly, if new debt collections occur during the month, the current total principal transferred to debt collection may be greater than the initial total principal transferred to debt collection at the beginning of the month. Using the larger value as the denominator also avoids artificially low recovery rates due to dilution of old repayments on a larger base. The project's base principal is updated daily within the month to reflect changes in principal caused by new or withdrawn debt collections. It is frozen and no longer updated after the following month. The purpose of freezing is to ensure that the historical monthly recovery rate calculation benchmark remains stable after settlement, ensuring the traceability and auditability of historical settlement results. The project's monthly recovery rate equals the total repayment amount for the settlement month divided by the project's base principal. The recovery rate threshold is a preset minimum recovery rate value set by the system. It is based on the following: when the monthly recovery rate of a project is below a very low level, the supplier's collection activities in that month have not generated any substantial contribution to the collection of payments. The value of the recovery rate threshold should ensure that the recovery rate of the vast majority of normal collection activities is higher than the threshold, filtering only extremely inefficient cases. For example, the recovery rate threshold can be set to 0.05%. When the monthly recovery rate of a project is lower than the recovery rate threshold, both the daily and monthly settlement status of the repayment record remain in the pending settlement status, and it does not proceed to step S14. When the monthly recovery rate of a project is greater than or equal to the recovery rate threshold, and the repayment record successfully completes the daily or monthly commission calculation in step S14 and passes the verification in step S15, the status field of the corresponding payment period is updated to settled. The reason for keeping the project in a pending settlement status instead of marking it as "not settled" is that the monthly recovery rate of the project is a dynamic value that continues to rise as the monthly payments accumulate. The recovery rate, which is lower than the lead rate at a certain point at the beginning of the month, may rise above the lead rate in the middle or end of the month due to the accumulation of subsequent payments. Keeping the project in a pending settlement status allows these records to be reprocessed in subsequent settlement batches after the recovery rate reaches the target without manual intervention.When the monthly recovery rate of a project is greater than or equal to the recovery rate threshold, the system reads the commission settlement scenario identifier from the repayment records. The commission settlement scenario identifier can be either daily or monthly. Daily settlement means that each repayment record recorded on the same day is settled and 70% of the commission is issued. Monthly settlement means that all repayment records for the month are settled at the end of the month according to the tiered fee rate. Based on the value of the commission settlement scenario identifier, the repayment records are routed to the corresponding branch in step S14. Repayment records with the commission settlement scenario identifier of daily settlement are routed to the daily settlement fee rate matching branch in the daily settlement batch and to the monthly settlement fee rate matching branch in the monthly settlement batch to perform monthly settlement netting calculation. Repayment records with the commission settlement scenario identifier of monthly settlement are only routed to the monthly settlement fee rate matching branch in the monthly settlement batch and participate in the monthly settlement commission calculation according to the rule of taking zero value for daily settlement commission. The monthly recovery rate and total monthly payment amount calculated in step S13 are not only directly referenced in the monthly settlement tier matching in step S14, but also used as the source of the vertical axis value for constructing the monthly recovery rate tier line chart in step S20, as well as the denominator for calculating the proportion of reversal amounts. Therefore, step S13 establishes a data connection between the pre-processing process in step S10 and the anomaly detection process in step S20. If the recovery rate pre-line verification in step S13 is missing, repayment records with extremely low monthly recovery rates will directly enter step S14. During the monthly settlement tier matching, they may not be able to match any tier because the recovery rate is lower than the minimum limit of the tier interval table, triggering the boundary anomaly not preset in step S14. If the dual-period routing judgment is missing, daily settlement records may be mistakenly entered into the monthly settlement branch and subjected to tiered fee rate matching and tax point adjustment, while monthly settlement records may be mistakenly entered into the daily settlement branch and directly settled according to the daily settlement fee rate. The calculation logic and profit sharing ratio of the two scenarios are completely different, and scenario confusion will directly produce commission calculation deviations.

[0027] Step S14: In the daily settlement rate matching branch, extract the daily settlement rate from the target settlement plan, obtain the credited amount of the repayment record, and calculate the daily settlement commission and frozen commission based on the credited amount and the daily settlement rate; In the monthly settlement rate matching branch, determine the monthly settlement rate, and calculate the monthly settlement commission based on the credited amount, daily settlement commission, and monthly settlement rate; Write the daily settlement commission, frozen commission, monthly settlement commission, along with the credited amount, credited time, monthly settlement rate, and total monthly repayment amount, into the settlement details data table; In step S14, the daily settlement rate matching branch and the monthly settlement rate matching branch are two mutually exclusive calculation paths. The same repayment record will only enter one of them in the same settlement batch. In the daily settlement rate matching branch, the institution tag type in the repayment record is read: When the institution tag type is external visit or legal action, suppliers of external visit and legal action types do not set up special plans in their business rules, that is, there is no distinction between regular plans and special plans under the same type. The matching project plan entries are directly searched from the project plan entry list of the target settlement plan determined in step S12 using the project name of the repayment record as the search condition, and the daily settlement rate configured in the project plan entry is extracted. When the institution tag type is telephone collection, suppliers of the telephone collection type may have both regular and special plan rate systems. The target settlement plan determined in step S12 based on the plan encoding vector is the current regular plan. The institution name in the repayment record needs to be compared one by one with the list of institution names in all configured special plans in the system. If a match is found, the daily settlement rate is extracted from the project plan entry list of the matched special plan using the project name of the repayment record as the search key. If no match is found, the target regular plan determined in step S12 is used, and the daily settlement rate is extracted from the project plan entry list of the target regular plan using the project name of the repayment record as the search key. See also Figure 3 This is a schematic diagram of the daily commission revenue sharing structure provided in the embodiments of this application. Figure 3 The diagram illustrates the calculation basis of daily settlement commissions, the profit-sharing ratio, and the complete execution process of the final flow of funds. After obtaining the daily settlement rate, the daily settlement commission is equal to the amount credited to the repayment record multiplied by the daily settlement rate, and then multiplied by a settlement ratio of 70%. The frozen commission is equal to the amount credited to the repayment record multiplied by the daily settlement rate, and then multiplied by a freezing ratio of 30%. Figure 3The above-mentioned revenue sharing calculation process is presented intuitively: After multiplying the received amount by the daily settlement rate to obtain the total commission, it is split into two parts according to a fixed ratio. Seventy percent is paid to the withdrawable balance as daily settlement commission, and 30% is temporarily stored in the frozen amount sub-account as frozen commission. The 70% and 30% allocation constitutes the revenue sharing structure of the commission: During the daily settlement stage, only the majority of the total commission is paid to the supplier's withdrawable balance, and the remaining part is temporarily stored in the frozen amount sub-account of the commission account. During the monthly settlement stage, the daily settlement commission is netted according to the final determined monthly settlement rate (i.e., the total commission payable under the monthly settlement rate is minus the daily settlement commission already paid; if the difference is positive, additional payments are made; if the difference is negative, it is deducted from the frozen amount). This revenue sharing mechanism ensures that the amount paid during the daily settlement stage never exceeds 70% of the final commission payable. Even if the rate is reduced during the monthly settlement stage, it only needs to be deducted from the frozen amount sub-account balance without having to recover the funds already in the withdrawable balance from the supplier, avoiding the accounting complexity and financial dispute risks caused by reverse recovery operations. In daily settlement scenarios, invoice type and tax point calculations are not performed. In the monthly settlement rate matching branch, the monthly settlement tier calculation method is read from the target settlement plan, and the value is one of three: fixed, based on recovery rate, or based on payment amount. In the monthly settlement rate matching branch, the monthly settlement tier rate specifically refers to the rate value obtained from the monthly settlement tier rate range table, and the monthly settlement rate specifically refers to the final rate value after tax point adjustment. When the monthly settlement tier calculation method is fixed, the monthly settlement tier rate is directly taken from the monthly settlement tier rate configured in the target settlement plan, which is suitable for projects where the rate does not fluctuate with performance indicators. When the monthly settlement tier calculation method is based on recovery rate, the project's monthly recovery rate calculated in step S13 is used to find the tier range that the project's monthly recovery rate falls into in the monthly settlement tier rate range table configured in the target settlement plan. The monthly settlement tiered fee rate range table consists of multiple consecutive and non-overlapping recovery rate ranges and the corresponding monthly settlement tiered fee rate for each range. For example, the monthly settlement tiered fee rate range table for a certain project can be shown in Table 2: Table 2. Examples of Monthly Settlement Tiered Rate Ranges 0.05% 10,000 0.35 10,000 0.02% 0.4 0.02% 0.03% 0.45 0.03% No upper limit 0.5 The monthly settlement rate is determined by the range in which the project's monthly recovery rate falls. When the monthly settlement tier is calculated based on the amount received, the project configured in the target settlement plan is used as the query condition. The sum of the inflow amounts of all pending settlement records of the supplier within the settlement month is summarized from the settlement details data table as the project's monthly recovery amount. The same range lookup operation is then performed in the monthly settlement tier rate range table, replacing the project's monthly recovery rate with the project's monthly recovery amount, and the corresponding monthly settlement rate is retrieved. After obtaining the monthly settlement rate tier, the tax point is adjusted based on the invoicing method of the account receiving the payment in the repayment record to determine the monthly settlement rate: When the invoicing method is a special VAT invoice and the supplier's tax point is equal to 6%, the monthly settlement rate is equal to the monthly settlement rate tier and no adjustment is made; when the invoicing method is a special VAT invoice and the supplier's tax point is not equal to 6%, the monthly settlement rate is equal to the monthly settlement rate tier minus an adjustment item. This adjustment item is equal to the difference between the monthly settlement rate tier multiplied by 6% and the supplier's actual tax point (i.e., adjustment item = monthly settlement rate tier × (6% - ...). The supplier's actual tax point is 6%, which is the system's preset standard tax point benchmark. When the supplier's actual tax point is lower than 6%, the difference is positive, and the monthly settlement fee rate is reduced from the monthly settlement tier fee rate. The reduction is proportional to both the monthly settlement tier fee rate and the tax point difference. The logic is that suppliers with lower tax points bear lower tax burdens. By deducting the portion corresponding to the tax point difference from the fee rate, the actual after-tax income of suppliers with different tax points tends to be consistent. When the invoicing method is a general invoice, the monthly settlement fee rate is equal to the monthly settlement tier fee rate. Monthly commission equals the amount credited to the repayment record multiplied by the monthly settlement rate minus the daily settlement commission. The daily settlement commission is determined as follows: when the repayment record's commission settlement scenario is marked as daily, the daily settlement commission is the value calculated and written into the settlement details table according to the daily settlement rate matching branch in step S14 within the daily settlement batch; when the repayment record's commission settlement scenario is marked as monthly, the repayment record has not been processed by the daily settlement rate matching branch, the daily settlement commission is zero, and the monthly settlement commission equals the amount credited to the repayment record multiplied by the monthly settlement rate. The daily settlement commission is deducted because the monthly settlement rate determines the final rate based on the full amount credited to the repayment record. Since 70% of the commission has already been paid according to the daily settlement rate during the daily settlement phase, the monthly settlement commission only takes the difference between the final commission payable and the commission already paid in the daily settlement phase, avoiding double-counting of commission for the same amount credited to the repayment record. Daily commissions, frozen commissions, or monthly commissions, along with the amount credited, crediting time, monthly tiered rate, monthly rate, and total monthly repayment amount, are written into the settlement details data table. The monthly tiered rate range table is the same table used in step S20 to determine whether a reversal event triggers a jump in tiers. The monthly tier into which the repayment record falls in step S14 directly corresponds to the determination criteria for the pre-credit or post-credit monthly tier in step S20. The two steps share the same tier system to ensure consistency between rate calculation and anomaly detection standards.If the multi-branch rate matching step S14 is missing, the differentiated rate configurations for different supplier types, different special statuses, and different performance levels in the target settlement plan will not be able to be distinguished and referenced, the deduction relationship between monthly commission and daily commission will not be established, and the commission splitting structure will be incomplete.

[0028] Step S15: Before commission calculation is executed, the input parameter range is checked on the amount received. After commission calculation is executed, the output parameter range is checked on the daily commission, frozen commission and monthly commission. Repayment records that fail the input parameter range check or the output parameter range check are marked as abnormal and an alarm signal is issued. In step S15, see Figure 4Before the commission calculation in step S14, the parameter range validation checks the validity of the credited amount in the repayment record and each coded value in the scheme matching encoding vector. The lower bound of the valid value range of the credited amount is zero, excluding zero, that is, the credited amount must be strictly greater than zero. The upper bound is the system's preset upper limit value for a single credited amount. The upper limit value for a single credited amount is set based on the statistical distribution of single repayment amounts in the historical operating data of the settlement system. The upper limit value is selected to cover the normal repayment amount range and exclude extremely large outliers. For example, the upper limit value for a single credited amount can be set to 1.5 times the 99th percentile value of a single repayment amount in the historical data. The valid value range for institution codes, asset package codes, and tag type codes is the set of registered code values ​​in the institution name to institution code mapping table, asset package name to asset package code mapping table, and institution tag type to tag type code mapping table in step S11. Code values ​​not registered in the mapping table are judged as illegal values. The valid value range for month codes is the set of settlement months that conform to the six-digit fixed-length year and month format (yyyymm) and belong to the current open settlement month set of the system. Month codes that do not conform to the format or are not in the open settlement month set are judged as illegal values. When the value of the credited amount is not within the valid value range for the credited amount or any code value in the scheme matching code vector is not in the corresponding set of registered code values, the repayment record is marked as an input parameter abnormality state, the system refuses to perform commission calculation in step S14 on the repayment record and issues an input parameter abnormality alarm signal. If the incoming amount is within the legal range, then step S14 is executed to calculate the daily commission, frozen commission, and monthly commission. After the commission calculation in step S14, the parameter range is checked for reasonableness: the lower bound of the legal range for the daily commission is zero (inclusive), and the upper bound is the product of the incoming amount of the repayment record multiplied by the system's preset daily fee rate cap and then multiplied by a 70% settlement ratio. The daily fee rate cap is set based on the maximum daily fee rate configured in all settlement schemes in the settlement scheme library multiplied by a margin coefficient. This margin coefficient ensures that the commission calculation results under normal fee rate configurations pass the check, while only those due to configuration errors are blocked. For example, the margin coefficient can be set to 1.2 to prevent extreme overestimation of the output due to errors. The lower bound of the legal range of frozen commission is zero (inclusive), and the upper bound is the product of the amount credited to the repayment record multiplied by the system's preset daily settlement rate limit and then multiplied by the 30% freezing ratio. The lower bound of the legal range of monthly settlement commission is the opposite of the product of the amount credited to the repayment record multiplied by the system's preset daily settlement rate limit and then multiplied by the 70% settlement ratio (i.e., the monthly settlement commission can be negative in the scenario of rate reduction, and a negative value indicates the amount to be deducted from the frozen amount sub-account), and the upper bound is the product of the amount credited to the repayment record multiplied by the system's preset monthly settlement rate limit.When the daily, frozen, or monthly commission value is outside the corresponding legal range, the repayment record is marked as having an abnormal output parameter status. The commission result calculated in step S14 is withdrawn, and the system issues an abnormal output parameter alarm signal. Input parameter range validation intercepts non-compliant input data, preventing it from participating in commission calculation and avoiding amplified erroneous outputs after the multiplication operation in step S14. Output parameter range validation checks the output result after calculation. Even if some hidden anomalies are not detected by input parameter range validation, the output parameter range validation can still intercept abnormal results if the commission value generated after commission calculation exceeds the legal output parameter range. Repayment records not marked as having an abnormal input or output parameter status will have their daily, frozen, or monthly commissions written to the settlement details data table. Repayment records marked as having abnormal input or output parameters will have their daily and monthly settlement status fields updated to abnormal. These records will not be written into the official settlement records in the settlement details table, nor will they participate in the area integral calculation of the recovery rate step line chart in step S20. This is to prevent abnormal data from interfering with the calculation of the total projected area and the subsequent measurement of the pollution shadow area ratio. If the parameter range verification in step S15 is missing, repayment records with extremely large incoming amounts will directly participate in commission calculation and be written into the settlement details table. This will cause an unrealistic jump in the vertical axis value of the project's monthly recovery rate step line chart in step S20, artificially inflating the total projected area. This will dilute the sensitivity of the pollution shadow area ratio feature in step S20 and weaken the gradient boosting decision tree model's ability to identify abnormal positive events.

[0029] The pre-processing flow of step S10 forms a five-layer progressive processing link: "encoding standardization—scheme binding—qualification verification and scenario diversion—differentiated rate calculation—parameter bidirectional interception." Step S11's encoding mapping transforms unstructured text fields into structured fixed-length codes, eliminating text matching ambiguities and providing a reliable retrieval key for step S12. Step S12's scheme isolation matching compares the encoded vector bit-by-bit with the scheme library and locks in a unique target settlement scheme, providing a clear scheme source for rate reading in step S14, and a fallback scheme ensures the continuity of the settlement process in case of missing scheme configurations. Step S13's pre-processing line verification of the recovery rate excludes repayment records that do not meet basic recovery requirements in advance, reducing the computational burden of processing invalid records in step S14. Simultaneously, the calculation results of the project's monthly recovery rate and the total monthly settlement amount serve both the monthly settlement tier matching in step S14 and provide basic values ​​for the area integration and feature construction in step S20, connecting the pre-processing flow and the anomaly detection flow at the data level. Step S14's multi-branch rate matching determines the rate reading path based on a combination of supplier type and project status. The use of the monthly settlement tiered rate range table allows commission settlement amounts to dynamically fluctuate with collection performance. The 70% and 30% revenue sharing structure provides a buffer mechanism for rate netting during the monthly settlement phase. The shared use of the monthly settlement tiered rate range table in steps S14 and S20 ensures that rate calculation and anomaly detection operate based on the same set of tier standards. Step S15's input parameter range verification and output parameter range verification set up interception layers at the commission calculation entry and exit points to capture anomalies caused by data transmission errors and system configuration defects before and after calculation. The progressive coordination of the above five layers of processing ensures that each record written into the settlement details data table undergoes dual verification of scheme attribution confirmation, recovery rate qualification verification, scenario routing distribution, rate matching calculation, and commission value rationality. When step S20 performs area integration and pollution shadow area stripping based on the recovery rate ladder line chart, the settlement details dataset on which it relies does not contain scheme mismatch records, recovery rate failure records, or parameter out-of-bounds records. This enables step S20 to accurately calculate the total projected area and pollution shadow area based on data that has been filtered through multiple layers, improves the accuracy of the gradient-enhancing decision tree model in determining the problem of falsely high water levels, and reduces false alarms or missed alarms caused by underlying data quality issues.The 70% and 30% revenue sharing structure in step S10 forms a protective connection with the abnormal reversal event detection in step S20: When step S20 detects an abnormal reversal event and triggers batch recalculation, it is necessary to deduct the total amount of commission recovery from the supplier's commission account. If 100% of the commission has been distributed to the withdrawable balance during the daily settlement stage and the supplier has completed the withdrawal, the deduction operation will not be able to be executed due to insufficient account balance. The existence of the 30% frozen amount ensures that the funds not yet released during the monthly settlement stage are still within the system's control range, providing an executable fund buffer for the recovery deduction operation in step S20.

[0030] Step S20: Construct a step-line chart of the project's monthly recovery rate based on the settlement details data table and calculate the total projected area. Remove the pollution shadow area from the project's monthly recovery rate step-line chart. Construct an anomaly judgment feature vector based on the pollution shadow area and the total projected area. Input the anomaly judgment feature vector into the pre-constructed gradient boosting decision tree model to output anomaly score values. Identify and judge abnormal reversal events based on the anomaly score values. For abnormal reversal events, retrieve the affected repayment records and perform batch recalculation of the affected repayment records and circuit breaker processing of commission accounts.

[0031] Further, step S20 includes: Step S21: Construct a horizontal axis with the first day of the current settlement month as the starting point and the last day as the ending point. Extract the cumulative amount of payment received for each day in the current settlement month up to that day from the settlement details data table. Calculate the monthly project recovery rate for each day based on the cumulative amount of payment received. Construct a project monthly recovery rate stepped line chart with the day as the horizontal axis scale, the project monthly recovery rate as the vertical axis coordinate, and horizontal step lines connecting the project monthly recovery rate values ​​of adjacent days. Multiply the project monthly recovery rate value for each day by the time unit length of one day to obtain the daily projected area. Calculate the total projected area of ​​the closed area enclosed by the project monthly recovery rate stepped line chart and the horizontal axis by accumulating the data day by day. In step S21, the project's monthly recovery rate step line chart is a geometric graph obtained by unfolding the dynamically changing project monthly recovery rate values ​​along the time axis. See also Figure 5 This is a schematic diagram of the total projected area of ​​the project's monthly recovery rate step-line graph provided in the embodiments of this application. Figure 5The diagram illustrates the components, graphic form, and geometric representation of the total projected area of ​​the project's monthly recovery rate tiered line chart. The construction process uses the first day of the current settlement month as the starting point of the horizontal axis and the last day of the current settlement month as the ending point. The horizontal axis scale interval is set to one day, and the vertical axis scale unit is set to 0.01%, consistent with the recovery rate scale used in the monthly settlement tiered fee rate interval table in step S14. For each day on the horizontal axis, the settlement details data table is used to filter by the institution name and project name in the repayment records. The sum of the amounts received for all repayment records with an invoice date no later than the current day and that have passed the parameter interval verification in step S15 is summarized to obtain the cumulative repayment amount up to the current day. The cumulative repayment amount is divided by the project's basic principal determined in step S13 to obtain the project's monthly recovery rate value up to the current day, and this value is marked as the corresponding coordinate point on the vertical axis for the current day. The cumulative repayment amount is calculated using a cumulative summary up to the current day rather than a daily increment method. This is because the project's monthly recovery rate is essentially the ratio of the cumulative repayments within the month to the project's initial principal. This value monotonically increases with the accumulation of repayments within the month, only briefly declining when there is a positive reversal. Therefore, the stepped line chart typically shows a horizontal or upward trend on most days. Connecting the project's monthly recovery rate coordinates of two adjacent days with a horizontal stepped line segment is defined as follows: extending a horizontal line segment to the right from the coordinate point of the previous day to the horizontal axis position of the next day, and then jumping to the project's monthly recovery rate coordinate point of the next day with a vertical line segment. The two line segments connect end-to-end to form a stepped shape. Figure 5 The two core line segments that make up the stepped line chart are clearly marked: horizontal stepped segments and vertical jump segments, which intuitively present the step-like growth pattern of the project's monthly recovery rate over time. Horizontal stepped segments indicate that the project's monthly recovery rate remains unchanged between two adjacent calendar days, while vertical jump segments indicate that the recovery rate changes abruptly due to the net inflow of funds into the system on the next calendar day.

[0032] The closed area enclosed between the stepped line graph and the horizontal axis forms a polygon composed of several single-sided rectangles. The total area of ​​the polygon is the total projected area. Figure 5The shaded area visually identifies the total projected area enclosed by the stepped line chart and the horizontal axis, clearly demonstrating the geometric feature that the total projected area is composed of multiple daily rectangles. The calculation process for the total projected area is as follows: for each calendar day on the horizontal axis, the monthly recovery rate corresponding to the current calendar day is multiplied by the time unit length of one calendar day to obtain the daily projected area corresponding to the current calendar day; the daily projected areas of all calendar days in the current settlement month are accumulated day by day to obtain the total projected area. The geometric meaning of the area below the stepped line chart differs from the traditional recovery rate measurement method that takes the final value at the end of the month: the final value at the end of the month only reflects the proportion of the cumulative recovery to the project's basic principal, ignoring the time difference of the recovery amount entering the system at different times within the month; while the total projected area below the stepped line chart simultaneously carries information about the scale of the recovery amount and the length of time the recovery survives in the system. The same final value at the end of the month can correspond to different total projected areas. The total projected area corresponding to a high recovery rate achieved at the beginning of the month is much larger than the total projected area corresponding to the same high recovery rate achieved at the end of the month. This geometric representation, which incorporates the time dimension into the recovery rate metric, provides a calculable graphical basis for step S22, which uses rectangular areas to peel away the contamination impact of the reversal amount within its survival time interval. This extends anomaly detection beyond the static point-to-point model of individual transaction reconciliation to a global state change tracing across time periods. The rule in step S13, which uses the larger value of the project's base principal, keeps the denominator relatively stable within the month, avoiding spurious fluctuations in the total projected area caused by denominator jumps. Without the construction of the stepped line chart and the calculation of the total projected area in step S21, the contamination shadow area in step S22 would lack a comparable benchmark, and the proportion of the contamination shadow area in step S24 would fail to form an effective normalized dimension, resulting in a lack of comparability between different sized asset packages for anomalies in the gradient boosting decision tree model.

[0033] Step S22: Based on the settlement details data table, detect repayment records with associated reversal records from the repayment records as original repayment records and mark them as reversal events. Read the entry time point and reversal time point from the original repayment record and the corresponding reversal record respectively. Determine the survival time interval by the left and right closed intervals from the entry time point to the reversal time point. Calculate the recovery rate increment. Construct a pollution shadow rectangle in the project's monthly recovery rate step line chart with the survival time interval as the width and the recovery rate increment as the height. Multiply the recovery rate increment by the number of natural days in the survival time interval to calculate the pollution shadow area. In step S22, the detection basis for the reversal event is the original repayment record with a related reversal record in the settlement details data table. The identification rule for related reversal records is as follows: for each repayment record belonging to the current settlement month in the settlement details data table, check whether there is another record with the same repayment order number and a negative amount. If so, the negative record is regarded as the reversal record, and the corresponding positive amount record is regarded as the original repayment record. A negative amount indicates that the corresponding record has performed a fund outflow operation at the account level. The association with the repayment order number indicates that the outflow operation is aimed at the original repayment record's corresponding fund inflow. The fund inflow time is read from the original repayment record as the fund inflow time point, and the fund inflow time is read from the reversal record as the reversal time point. The left-closed and right-closed time interval between the two time points is defined as the survival time interval of the original repayment record's fund inflow amount in the system. The length of the survival time interval is measured in calendar days. The calculation process for the recovery rate increment is as follows: Divide the original repayment record's credited amount by the project's base principal determined in step S13 to obtain the recovery rate increment corresponding to the original repayment record's credited amount. The recovery rate increment reflects the immediate increase in the project's monthly recovery rate due to the original repayment record's credited amount. In the stepped line graph constructed in step S21, locate the horizontal axis range corresponding to the survival time interval. Construct a shading rectangle with the starting natural day of the survival time interval as the left boundary, the ending natural day of the survival time interval as the right boundary, and the recovery rate increment as the rectangle's height. The area of ​​the shading rectangle is equal to the recovery rate increment multiplied by the number of natural days in the survival time interval. The geometric meaning of the shading rectangle is: the original repayment record's credited amount contributes a constant incremental height to the project's monthly recovery rate within its survival time interval, and the rectangle's area quantifies the contribution of the incremental height to the total projected area of ​​the stepped line graph within the survival time interval. The reason for using a geometric rectangle to isolate the contamination shadow area is that traditional transaction-by-transaction reconciliation mechanisms only balance the amounts between the original repayment record and the reversal record when a reversal occurs. After the reversal record is recorded, the commission of the original repayment record is deducted, and the balance is restored at the book level. However, the commission overflow generated by other repayment records settled at the higher tier due to the jump in tiered rates during the survival period of the original repayment record has flowed into the supplier's account and cannot be captured by regular reconciliation. After quantifying the contamination impact as a rectangular area, the height dimension of the recovery rate increment characterizes the strength of the increase, and the width dimension of the survival time interval characterizes the duration of the increase. The product of the two reflects the intensity of contamination on the overall recovery rate status during the existence of the original repayment record. The contamination shadow area and the total projected area in step S21 are dimensionally unified, both using the recovery rate scale multiplied by the length of the natural day as the basic unit, so that the numerator and denominator can be directly divided when forming the contamination shadow area proportion feature in step S24, avoiding the problem of feature scale imbalance.Steps S22 and S21 form a graphical overlay: the stepped line chart provides the area basis for the global recovery rate, and the contaminated shadow rectangles embedded inside the stepped line chart represent local contaminated areas. The division of the areas of the two is the numerical source of the first feature in step S24. Without the contaminated shadow area stripping in step S22, the global impact of the reversal event on other orders in the same batch cannot be measured, and the input feature set of the gradient boosting decision tree model will lack the core discriminative dimension used to distinguish between the artificially high manufacturing levels at the end of the month and normal reversals within the month.

[0034] Step S23: Read the monthly project recovery rate value of the day before the accounting time from the project monthly recovery rate ladder line chart as the pre-accounting monthly recovery rate, and read the monthly project recovery rate value of the day of accounting as the post-accounting monthly recovery rate. Determine the pre-accounting monthly settlement tier based on the pre-accounting monthly recovery rate, and determine the post-accounting monthly settlement tier based on the post-accounting monthly recovery rate. When the monthly settlement tier rate corresponding to the post-accounting monthly settlement tier is higher than the monthly settlement tier rate corresponding to the pre-accounting monthly settlement tier, it is determined that the reversal event triggers a cross-tier jump and is marked as a cross-tier reversal event. In step S23, the jump-ahead determination is a secondary screening step after calculating the pollution shadow area, further filtering the positive events. The determination process relies on the monthly settlement tiered rate range table used in step S14. Specifically, the process is as follows: the monthly recovery rate value of the project corresponding to the day before the accounting entry time point is read from the tiered line chart constructed in step S21 as the pre-accounting monthly recovery rate, and the monthly recovery rate value of the project corresponding to the accounting entry time point is read as the post-accounting monthly recovery rate; the pre-accounting monthly recovery rate is compared with the monthly settlement tiered rate range table to determine the tier range into which the pre-accounting monthly recovery rate falls, and the corresponding monthly settlement tier rate is defined as the pre-accounting monthly settlement tier rate, and the tier range is defined as the pre-accounting monthly settlement tier; the post-accounting monthly recovery rate is compared in the same way to obtain the post-accounting monthly settlement tier. When the monthly settlement rate corresponding to the post-post payment tier is higher than the monthly settlement rate corresponding to the pre-post payment tier, the reversal event is determined to have triggered a cross-tier jump, and the original reversal event is marked as a cross-tier reversal event and proceeds to feature construction in step S24. When the pre-post payment tier and the post-post payment tier are the same, or the monthly settlement rate corresponding to the post-post payment tier is lower than the pre-post payment tier, the reversal event is determined not to have triggered a cross-tier jump, and only a single commission deduction operation is performed on the original repayment record; it does not proceed to the subsequent anomaly determination process. The filtering function of the cross-tier jump determination is to identify reversal events with the potential for substantial rate pollution. In real-world business scenarios, reasonable reversal situations exist. For example, a customer's mistake leading to a duplicate payment and subsequent automatic refund from the bank system, or a supplier's error resulting in data entry and subsequent cancellation, may involve monetary outflows. However, since the monthly recovery rate of the project before and after the original repayment record was entered into the account, it does not trigger a jump in the monthly settlement rate, thus preventing other orders during the current period from being settled at the incorrectly higher rate. Without secondary filtering for cross-rate jumps, all reversal events enter the gradient boosting decision tree model for judgment. Normal reversal samples will dilute the abnormal signal density of the training samples, causing the model to misjudge events with small contamination shadow areas that do not trigger cross-rate jumps. Steps S23 and S14 form a shared relationship in the rate system: Step S14 determines the monthly settlement rate based on the project's monthly recovery rate during the commission calculation stage, while Step S23 determines cross-rate jumps based on the same rate interval table during the anomaly detection stage. The two steps use the same rate standard to ensure logical consistency between the rate calculation rate allocation and the anomaly detection rate change judgment.

[0035] Step S24: Construct an anomaly judgment feature vector for cross-tier reversal events. The anomaly judgment feature vector includes seven features: the proportion of the contaminated shadow area, the time distance at the end of the month, the cross-tier difference, the proportion of the reversal amount, the number of affected repayment records, the total amount of commission overflow, and the cumulative number of historical reversal operations. In step S24, for each event identified as a cross-rate reversal event in step S23, an anomaly judgment feature vector consisting of seven features is constructed as input to the gradient boosting decision tree model. The first feature is the proportion of the contaminated shadow area, calculated by dividing the contaminated shadow area obtained in step S22 by the total projected area obtained in step S21, with a value between 0 and 1. It reflects the proportion of the contamination intensity of the cross-rate reversal event on the overall recovery rate status in the total recovery rate area. The larger the value, the more severe the contamination. The second feature is the time distance to the end of the month, calculated as the number of calendar days between the accounting entry time and the last calendar day of the current settlement month. The smaller the proximity value, the closer the accounting entry time is to the end of the month. Anomalies that create artificially high water levels at the end of the month tend to be concentrated in the last few days of the settlement month so that the artificially high state just covers the month-end settlement time. The third characteristic is the cross-tier difference, calculated as the difference between the monthly settlement tier rate corresponding to the monthly settlement tier after the payment is received and the monthly settlement tier rate corresponding to the monthly settlement tier before the payment is received. The larger the difference, the greater the rate jump, and the greater the single commission overflow generated by other orders settled at the higher tier during the survival period. The fourth characteristic is the proportion of the reversal amount, calculated as the payment amount of the original repayment record divided by the total monthly settlement amount used in step S13 to calculate the project's monthly recovery rate. The higher the proportion, the greater the weight of the original repayment record in the total monthly settlement amount, and the more direct the impact on the final value of the project's monthly recovery rate.

[0036] The fifth characteristic is the number of affected repayment records. The search scope for affected repayment records is limited to: retrieving repayment records from the settlement details table whose posting time falls within the survival period of the current cross-tier reversal event, belong to the same institution and project as the original repayment record of the current cross-tier reversal event, and whose monthly settlement rate is equal to the monthly settlement rate corresponding to the posting monthly settlement rate. The original repayment record of the current cross-tier reversal event itself is excluded from the search results, and the total number of remaining records is counted. The count represents the number of other genuine repayment records that, during the survival period of the original repayment record of the cross-tier reversal event, were settled at the monthly settlement rate of the posting monthly settlement rate because the project's monthly recovery rate was inflated to the posting monthly settlement rate. The higher the number of records, the wider the scope of orders affected by the contamination. The sixth feature is the total commission overflow, calculated as follows: For each affected repayment record retrieved by the fifth feature, the difference between the monthly settlement rate corresponding to the monthly settlement tier after the post-posting and the monthly settlement rate corresponding to the pre-posting monthly settlement tier is the jump rate difference. The jump rate difference is multiplied by the post-posting monthly settlement rate difference by the post-posting monthly settlement rate difference to obtain the commission overflow amount of the single affected repayment record (its value is equivalent to the difference between the two monthly settlement commissions calculated based on the monthly settlement rate after the post-posting and the pre-posting monthly settlement rate, respectively, and determined by the same tax point adjustment rules). The commission overflow amounts of all affected repayment records within the survival period are summed up one by one to obtain the total commission overflow amount, which quantifies the actual financial loss risk exposure caused by the cross-tier reversal event. The seventh feature is the cumulative number of historical reversal operations. This is calculated by counting the total number of reversal records for the same supplier within the past twelve consecutive settlement months from the system's historical settlement data. The frequency of historical reversals reflects the supplier's historical behavioral patterns; a high cumulative number indicates abnormal frequency of reversal operations, increasing the suspicion of anomalous activity. These seven features characterize cross-tier reversal events from seven dimensions: pollution intensity, time concentration, rate jump magnitude, amount weight, scope of collateral impact, financial loss exposure, and historical behavioral patterns. Together, these seven features constitute the anomaly detection feature vector. A single feature is insufficient to determine an anomaly. For example, an event with a high proportion of pollution shadow area may simply have a longer survival period due to the payment occurring at the beginning of the month, and does not necessarily indicate subjective anomaly. Similarly, an event with a large cross-tier difference, if occurring at the beginning of the month and accompanied by a large amount of genuine payments, may also be a normal business payment concentration situation. The parallel arrangement of feature dimensions is not a mechanical listing, but rather a multifaceted portrait of the same event. Events with both high time proximity and high monetary weights are more likely to be manipulated. Events with both large contaminated area percentages and large affected ranges will inevitably have substantial financial exposure. Historical reversal frequency adds longitudinal evidence of supplier behavior patterns on top of the features of a single event. The gradient-boosting decision tree model, through nonlinear combination and split learning of seven features, can capture the combination patterns of concentrated month-end payments, large cross-tier differences, wide affected range, and frequent historical reversals, achieving a discriminative capability that single feature dimensions cannot match.

[0037] Step S25: Input the anomaly judgment feature vector into the pre-trained gradient boosting decision tree model and output the anomaly score value. When the anomaly score value is greater than or equal to the preset anomaly judgment threshold, the cross-gear reversal event is judged as an anomaly reversal event. In step S25, see Figure 6 This is a schematic diagram of the gradient boosting decision tree model inference provided in the embodiments of this application. Figure 6The diagram illustrates the complete reasoning process of the gradient boosting decision tree model, from inputting anomaly detection feature vectors to outputting anomaly scores and completing anomaly detection. The gradient boosting decision tree model is trained using a round-by-round residual fitting method. The construction process of the training sample set is as follows: from the system's historical settlement data, manually audited and confirmed anomaly correction events are labeled as positive samples with a target variable value of 1, and confirmed normal correction events are labeled as negative samples with a target variable value of 0. Following the same process from steps S21 to S24, anomaly detection feature vectors are calculated for each training sample, forming a training sample set with the anomaly detection feature vectors as input and binary labels as the target variable. The loss function is a binary logarithmic loss function, and the corresponding negative gradient is the fitting target for each iteration. The difference between the predicted probability obtained by logical transformation of the current predicted value and the true value of the target variable is used as the pseudo-residual for the current round. The training process is as follows: The initial predicted value is the mean of the target variable from all training samples after logarithmic probability transformation. For each training sample, the pseudo-residual for the current round is calculated. A CART regression tree is trained using the pseudo-residual as the fitting target. Each internal node of the CART regression tree traverses the seven features in the anomaly detection feature vector. For each feature, candidate split thresholds are traversed. The feature and threshold combination that minimizes the sum of the squares of the pseudo-residuals in the two subsets after splitting is selected as the node splitting condition. Node samples are assigned to the left or right child node according to the splitting condition. Recursive splitting continues until the preset maximum tree depth is reached or the number of samples in a node is lower than the minimum number of split samples. The output value of a leaf node is the mean of the pseudo-residuals of all samples within that leaf node. The output value of the leaf node of the CART regression tree is multiplied by the learning step size and accumulated to the current predicted value to obtain the updated predicted value. The difference between the predicted probability after logical transformation of the updated predicted value and the true value of the target variable is used as the new round of pseudo-residual. The next CART regression tree is trained, iterating round by round until the preset maximum number of iterations is reached or the validation set loss stops decreasing for several consecutive rounds, triggering early stopping. The hyperparameters, such as learning step size, maximum tree depth, maximum number of iterations, minimum number of split samples, and L2 leaf node weight regularization coefficient, are set based on the following: a grid search is performed on the validation set to select the hyperparameter combination with the lowest log loss on the validation set. For example, the learning step size can be between 0.05 and 0.1, the maximum tree depth can be between 3 and 6 layers, the maximum number of iterations can be between 100 and 500, the minimum number of split samples can be between 20 and 50, and the L2 leaf node weight regularization coefficient can be between 0.1 and 1. The inference process is as follows: the abnormal judgment feature vector of the cross-gear reversal event to be judged, constructed in step S24, is sequentially passed through each trained CART regression tree. Figure 6The ensemble inference structure of multiple CART regression trees (such as CART regression tree 1, CART regression tree 2, and CART regression tree 3) is clearly demonstrated. Each CART regression tree performs a node splitting operation on the input anomaly detection feature vector and outputs the prediction increment of the corresponding leaf node (such as prediction increment 1, prediction increment 2, and prediction increment 3). Each CART regression tree routes the feature vector to a unique leaf node according to the splitting condition and outputs the prediction increment of the leaf node. The initial prediction value is accumulated with the prediction increments output by all CART regression trees to obtain the logarithmic probability value. After logical transformation, the anomaly score value of the cross-gear reversal event is obtained. Figure 6 The entire computational chain of this cumulative transformation is presented. The anomaly score is compared with the preset anomaly detection threshold. Figure 6 The comparison rules between the anomaly score and the anomaly determination threshold are clearly marked. When the anomaly score is greater than or equal to the anomaly determination threshold, the cross-grade reversal event is judged as an anomaly reversal event, and the batch recalculation process in step S26 is initiated. When the anomaly score is less than the anomaly determination threshold, only a single commission deduction operation is performed on the original repayment record. The anomaly determination threshold is set based on the following: the working characteristic curve of the subjects is plotted on the validation set, and the score corresponding to the working point that best balances the recall and precision is selected as the anomaly determination threshold. As a classic machine learning algorithm, the gradient boosting decision tree model is chosen for the current anomaly detection task because: the number of positive samples in anomaly reversal events is much smaller than the number of negative samples in normal reversal events, and the sample distribution exhibits an imbalanced characteristic. Gradient boosting decision trees achieve a biased fit to the minority class by focusing on the difficult samples with large pseudo residuals in the previous round, which maintains better classification performance in imbalanced scenarios compared to single decision trees or logistic regression models. At the same time, gradient boosting decision trees automatically complete the feature importance ranking and feature combination learning when splitting nodes, without the need for manually pre-setting the cross rules between the seven features.

[0038] Step S26: For cross-tier reversal events determined to be abnormal reversal events, retrieve the set of affected repayment records, replace the monthly settlement tier rate applicable to each repayment record in the set of affected repayment records with the monthly settlement tier rate before the posting of the payment, re-execute the tax point adjustment to determine the monthly settlement rate and recalculate the monthly settlement commission, obtain the total commission recovery amount based on the recalculated monthly settlement commission, deduct the total commission recovery amount from the commission account and trigger the circuit breaker mechanism.

[0039] In step S26, for each cross-tier reversal event determined to be an abnormal reversal event, batch recalculation of affected repayment records and circuit breaker processing of commission accounts are performed. The retrieval criteria for the set of affected repayment records are consistent with the retrieval criteria for the fifth feature in step S24, namely, the posting time falls within the survival time range, they belong to the same institution and the same project, the monthly settlement tier rate is equal to the monthly settlement tier rate corresponding to the posting tier, and the original repayment record itself is excluded from the current abnormal reversal event. For each record in the affected repayment record set, the monthly settlement rate is replaced from the rate corresponding to the post-post monthly settlement rate to the rate corresponding to the pre-post monthly settlement rate. Based on the replaced monthly settlement rate, the monthly settlement rate is re-determined according to the tax adjustment rules of the monthly settlement rate matching branch in step S14, and the monthly settlement commission is recalculated. The calculation rules include invoicing method judgment and tax adjustment. The difference between the recalculated monthly settlement commission and the original monthly settlement commission value stored in the settlement details data table is the single commission recovery amount. The total commission recovery amount is obtained by summing up all single commission recovery amounts in the affected repayment record set. The total commission recovery amount is deducted from the withdrawable balance of the supplier commission account. When the withdrawable balance is greater than or equal to the total commission recovery amount, the withdrawable balance is reduced by the corresponding amount after deduction. When the withdrawable balance is insufficient to cover the total commission recovery amount, the withdrawable balance is reduced to zero, and the remaining unrecovered portion is recorded as the supplier's outstanding payment amount, which will then be processed through manual collection or offset against the next commission settlement. The deduction operation simultaneously triggers the commission account circuit breaker mechanism: freezing all remaining withdrawable balances in the supplier's commission account and the balances of the frozen sub-accounts, rejecting all pending withdrawal requests and subsequent withdrawal requests from the supplier during the circuit breaker period, and issuing an abnormal reversal circuit breaker alarm signal to the system administrator. The alarm signal includes the supplier's organization name, the repayment order number that triggered the abnormal cross-tier reversal event, the reversal amount, the survival time interval, the percentage of the contamination shadow area, the abnormal score value output by the gradient boosting decision tree model, the total number of affected repayment records, and the total commission recovery amount. The circuit breaker mechanism is lifted manually by the system administrator after a manual audit is completed. In the settlement details data table, the monthly settlement rate field of the affected repayment records is updated to the monthly settlement rate corresponding to the monthly settlement rate before the payment, the monthly settlement rate field is updated to the monthly settlement rate re-determined after tax point adjustment, and the monthly settlement commission field is updated to the recalculated value. The event traceability mark of batch recalculation is marked in the remarks field. In the stepped line chart constructed in step S21, the payment amount corresponding to the abnormal reversal event is removed from the cumulative repayment amount of each natural day within the survival time interval. The monthly recovery rate value of the project for each natural day within the survival time interval is recalculated based on the updated cumulative repayment amount, and a corrected stepped line chart is generated for reference in the anomaly detection of subsequent settlement cycles.A recovery guarantee relationship is formed between step S26 and the 70% and 30% commission sharing structure described in step S14 of step S10: the 30% freezing ratio sharing structure in step S14 ensures that the frozen amount sub-account balance before the monthly settlement stage is still within the system's control range. When step S26 triggers commission recovery deduction, even if the supplier has already withdrawn the 70% commission paid in the daily settlement stage from the withdrawable balance, the frozen amount sub-account balance can still serve as a buffer source for recovery funds. This cross-step protection connection avoids the scenario where anomaly detection is successful but no funds can be recovered. The commission account circuit breaker mechanism in step S26 and the input and output parameter range verification in step S15 are complementary in risk interception logic. The parameter range verification in step S15 performs pre-interception for numerical anomalies at the single data level, while the commission account circuit breaker mechanism in step S26 performs post-event blocking for global rate pollution triggered by cross-tier reversal events. The two mechanisms respectively block different risk entry points in the fund outflow process.

[0040] Step S20 abstracts the dynamically changing monthly recovery rate of the project into a geometric figure with time width and recovery rate height using a stepped line graph. It quantifies the impact of reversal amounts on the global recovery rate within the survival time interval into a measurable area value by peeling away the contaminated shaded rectangles. It filters out normal reversal events that do not possess rate contamination capabilities through cross-tier jump judgment. It comprehensively characterizes cross-tier reversal events with rate contamination capabilities through seven multi-dimensional features. It performs non-linear learning and discrimination of multi-dimensional features using a gradient boosting decision tree model. Finally, it completes fund recovery and risk mitigation after anomaly detection through batch recalculation and commission account circuit breaker. Anomaly detection has expanded from a static model of reconciling amounts one by one to a global state tracing model based on the time dimension, and from judging the legality of a single event to judging the cascading impact of the event on other orders in the same batch. The practice of injecting large sums of money at the end of the month to inflate the monthly recovery rate of a project and cover the high-level monthly settlement rate before withdrawing them is difficult to avoid due to the combination of the characteristics of the proportion of the contaminated shadow area and the characteristics of the time distance at the end of the month. The commission overflow generated by other real orders settled at the erroneous high level during the survival period of the reversal event can be accurately recovered after the event is identified because of the combination of the characteristics of the number of affected repayment records and the total amount of commission overflow. Normal reversal events will not be falsely reported as abnormal due to the cross-level leap filtering mechanism and the comprehensive judgment of seven characteristics. Abnormal events between asset packages of different sizes can be compared and evaluated in the same model due to the normalized dimension design of the proportion of the contaminated shadow area. The abnormal judgment result has an executable fund recovery path due to the cross-step cooperation with the commission account splitting mechanism and the withdrawal circuit breaker mechanism in step S10, avoiding the execution break of valid judgment but failed recovery. After the time dimension of the recovery rate is expressed through geometric area, the time information originally hidden behind the end-of-month final value is transformed into graphical features. This allows the gradient boosting decision tree model to capture the differences that the same end-of-month recovery rate value may correspond to different recovery quality within the month. The area ratio relationship between the stepped line graph and the contaminated shaded rectangle allows the contamination intensity of cross-tier reversal events to be quantitatively expressed, rather than just qualitatively judging whether reversal behavior exists. The introduction of the cumulative number of historical reversal operations expands the single-event discrimination to discrimination based on the supplier's vertical behavior pattern, so that the probability of identifying suppliers who repeatedly perform abnormal operations increases with the accumulation of operation counts. The batch recalculation mechanism uses the pre-billing monthly settlement rate to retrospectively correct the monthly settlement commission of affected orders, eliminating the rate contamination traces caused by cross-tier reversal events from the settlement result level. This allows the system state after the anomaly to revert to an uncontaminated state, ensuring the consistency and auditability of historical settlement data in the long term. Example

[0041] This embodiment, based on Embodiment 1, provides a commission settlement anomaly data detection system based on gradient boosting decision trees, such as... Figure 7 As shown, it includes: Settlement details generation module: used to extract key positioning fields from repayment records to construct scheme matching code vectors, match the scheme matching code vectors with the settlement scheme library to determine the target settlement scheme, calculate the monthly recovery rate of the project and perform recovery rate front-line verification, and perform dual-account period routing and commission calculation on repayment records that pass the recovery rate front-line verification according to the target settlement scheme to generate a settlement details data table; Anomaly feature construction module: Based on the settlement details data table, construct a step line chart of the project's monthly recovery rate and calculate the total projected area. Remove the pollution shadow area from the project's monthly recovery rate step line chart and construct anomaly judgment feature vector based on the pollution shadow area and the total projected area. Anomaly detection and handling module: This module is used to input anomaly detection feature vectors into a pre-built gradient boosting decision tree model to output anomaly scores. Based on the anomaly scores, it identifies and determines anomaly reversal events. For anomaly reversal events, it retrieves affected repayment records and performs batch recalculation of affected repayment records and circuit breaker processing of commission accounts.

[0042] Furthermore, in the settlement details generation module, the key positioning fields include at least the institution name, asset package name, institution tag type, and settlement month; The method for constructing the matching encoding vector includes: mapping the institution name to the institution code, the settlement month to the month code, the asset package name to the asset package code, and the institution tag type to the tag type code, and concatenating the matching encoding vector in a fixed order of institution code, month code, asset package code, and tag type code.

[0043] Furthermore, in the anomaly feature construction module, the method for constructing a step-wise line graph of the monthly recovery rate of the project and calculating the total projected area includes: The horizontal axis is constructed with the first day of the current settlement month as the starting point and the last day as the ending point. The cumulative amount of payment received for each day in the current settlement month is extracted from the settlement details data table. Based on the cumulative amount of payment received, the monthly recovery rate of the project is calculated for each day. A step line chart of the monthly recovery rate of the project is constructed with the day as the horizontal axis scale, the monthly recovery rate of the project as the vertical axis coordinate, and horizontal step line segments connecting the monthly recovery rate values ​​of adjacent days. Multiply the monthly project recovery rate for each calendar day by the time unit length of one calendar day to obtain the daily projected area, and then accumulate the projected area of ​​the monthly project recovery rate step line graph and the horizontal axis to calculate the total projected area of ​​the enclosed area.

[0044] The method for removing the pollution shadow area in the project's monthly recovery rate step line chart includes: Repayment records with associated reversal records are identified from the repayment records and used as original repayment records, marked as reversal events. The posting time and reversal time are read from the original repayment record and the corresponding reversal record, respectively. The survival time interval is determined by the left and right closed intervals from the posting time to the reversal time. The recovery rate increment is calculated. A pollution shadow rectangle is constructed in the project's monthly recovery rate step line chart with the survival time interval as the width and the recovery rate increment as the height. The pollution shadow area is calculated by multiplying the recovery rate increment by the number of natural days in the survival time interval.

[0045] The method for constructing the anomaly detection feature vector includes: An anomaly judgment feature vector is constructed for cross-grade reversal events. The anomaly judgment feature vector includes seven features, namely, the proportion of the pollution shadow area, the time distance at the end of the month, the cross-grade difference, the proportion of the reversal amount, the number of affected repayment records, the total amount of commission overflow, and the cumulative number of historical reversal operations. The pollution shadow area ratio is the ratio of the pollution shadow area to the total projected area.

[0046] The methods and systems of this application may be implemented in many ways. For example, they may be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware. The above-described order of steps for the method is for illustrative purposes only, and the steps of the method of this application are not limited to the order specifically described above, unless otherwise specifically stated.

[0047] In addition, the parts of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of the corresponding technical solutions in the prior art have not been described in detail, so as to avoid excessive elaboration.

[0048] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting abnormal commission settlement data based on gradient boosting decision trees, characterized in that, The method includes: Extract key positioning fields from repayment records to construct scheme matching encoding vectors. Match the scheme matching encoding vectors with the settlement scheme library to determine the target settlement scheme. Calculate the monthly recovery rate of the project and perform recovery rate front-line verification. For repayment records that pass the recovery rate front-line verification, perform dual-account period routing and commission calculation based on the target settlement scheme to generate a settlement detail data table. Based on the settlement details data table, a step-line chart of the project's monthly recovery rate is constructed, and the total projected area is calculated. The pollution shadow area is removed from the project's monthly recovery rate step-line chart. An anomaly judgment feature vector is constructed based on the pollution shadow area and the total projected area. The anomaly judgment feature vector is input into a pre-constructed gradient boosting decision tree model to output anomaly score values. Anomaly reversal events are identified and judged based on the anomaly score values. For anomaly reversal events, the affected repayment records are retrieved, and batch recalculation of the affected repayment records and commission account circuit breaker processing are performed.

2. The commission settlement anomaly data detection method based on gradient boosting decision tree according to claim 1, characterized in that, The key positioning fields include at least the institution name, asset package name, institution tag type, and settlement month; The method for constructing the matching encoding vector includes: mapping the institution name to the institution code, the settlement month to the month code, the asset package name to the asset package code, and the institution tag type to the tag type code, and concatenating the matching encoding vector in a fixed order of institution code, month code, asset package code, and tag type code.

3. The commission settlement anomaly data detection method based on gradient boosting decision tree according to claim 1, characterized in that, The method for determining the target settlement plan includes: The scheme matching code vector is compared one by one with the pre-configured scheme code vectors in the settlement scheme library. When the scheme matching code vector is completely consistent with the scheme code vector of a certain settlement scheme in the settlement scheme library, the settlement scheme with complete consistency is determined as the target settlement scheme. When the scheme matching code vector is inconsistent with all scheme code vectors in the settlement scheme library, the repayment record is routed to the backup settlement scheme and a scheme missing alarm signal is issued.

4. The commission settlement anomaly data detection method based on gradient boosting decision tree according to claim 1, characterized in that, The method for performing a recovery rate front-line verification and routing repayment records that pass the recovery rate front-line verification according to the target settlement plan includes: The monthly recovery rate of the project is compared with the preset recovery rate threshold. If the monthly recovery rate of the project is greater than or equal to the recovery rate threshold, it is determined that the recovery rate threshold verification is passed. When the recovery rate threshold verification is passed, the commission settlement scenario identifier is read from the repayment record, and the repayment record is routed to the daily settlement rate matching branch or the monthly settlement rate matching branch according to the commission settlement scenario identifier.

5. The commission settlement anomaly data detection method based on gradient boosting decision tree according to claim 4, characterized in that, The method for generating a settlement details data table by performing dual-billing period routing and commission calculation includes: In the daily settlement rate matching branch, the daily settlement rate is extracted from the target settlement plan, the amount credited to the repayment record is obtained, and the daily settlement commission and frozen commission are calculated based on the amount credited and the daily settlement rate. In the monthly settlement rate matching branch, determine the monthly settlement rate and calculate the monthly settlement commission based on the incoming amount, daily settlement commission, and monthly settlement rate; write the daily settlement commission, frozen commission, monthly settlement commission, along with the incoming amount, incoming time, monthly settlement rate, and total monthly settlement amount into the settlement details data table.

6. The commission settlement anomaly data detection method based on gradient boosting decision tree according to claim 5, characterized in that, The method further includes: Before commission calculation is executed, the input parameter range is validated on the incoming amount. After commission calculation is executed, the output parameter range is validated on the daily commission, frozen commission, and monthly commission. Repayment records that fail the input or output parameter range validation are marked as abnormal and an alarm signal is issued.

7. The commission settlement anomaly data detection method based on gradient boosting decision tree according to claim 5, characterized in that, The method for constructing a step-wise line graph of the monthly recovery rate of the project and calculating the total projected area includes: The horizontal axis is constructed with the first day of the current settlement month as the starting point and the last day as the ending point. The cumulative amount of payment received for each day in the current settlement month is extracted from the settlement details data table. Based on the cumulative amount of payment received, the monthly recovery rate of the project is calculated for each day. A step line chart of the monthly recovery rate of the project is constructed with the day as the horizontal axis scale, the monthly recovery rate of the project as the vertical axis coordinate, and horizontal step line segments connecting the monthly recovery rate values ​​of adjacent days. Multiply the monthly project recovery rate for each calendar day by the time unit length of one calendar day to obtain the daily projected area, and then accumulate the projected area of ​​the monthly project recovery rate step line graph and the horizontal axis to calculate the total projected area of ​​the enclosed area.

8. The commission settlement anomaly data detection method based on gradient boosting decision tree according to claim 7, characterized in that, The method for removing the pollution shadow area in the project's monthly recovery rate step line chart includes: Repayment records with associated reversal records are identified from the repayment records and used as original repayment records, marked as reversal events. The posting time and reversal time are read from the original repayment record and the corresponding reversal record, respectively. The survival time interval is determined by the left and right closed intervals from the posting time to the reversal time. The recovery rate increment is calculated. A pollution shadow rectangle is constructed in the project's monthly recovery rate step line chart with the survival time interval as the width and the recovery rate increment as the height. The pollution shadow area is calculated by multiplying the recovery rate increment by the number of natural days in the survival time interval.

9. The commission settlement anomaly data detection method based on gradient boosting decision tree according to claim 8, characterized in that, The method further includes: The monthly project recovery rate is read from the project monthly recovery rate ladder line chart on the day before the accounting date as the pre-accounting monthly recovery rate, and the monthly project recovery rate on the accounting date is read as the post-accounting monthly recovery rate. The pre-accounting monthly settlement tier is determined based on the pre-accounting monthly recovery rate, and the post-accounting monthly settlement tier is determined based on the post-accounting monthly recovery rate. When the monthly settlement rate corresponding to the post-accounting monthly settlement tier is higher than the monthly settlement rate corresponding to the pre-accounting monthly settlement tier, a reversal event is determined, triggering a cross-tier jump and marked as a cross-tier reversal event.

10. The commission settlement anomaly data detection method based on gradient boosting decision tree according to claim 9, characterized in that, The method for constructing the anomaly detection feature vector includes: An anomaly judgment feature vector is constructed for cross-grade reversal events. The anomaly judgment feature vector includes seven features, namely, the proportion of the pollution shadow area, the time distance at the end of the month, the cross-grade difference, the proportion of the reversal amount, the number of affected repayment records, the total amount of commission overflow, and the cumulative number of historical reversal operations. The pollution shadow area ratio is the ratio of the pollution shadow area to the total projected area.

11. The commission settlement anomaly data detection method based on gradient boosting decision tree according to claim 10, characterized in that, The method for determining abnormal reversal events is as follows: When the abnormal score value is greater than or equal to the preset abnormal judgment threshold, the cross-grade reversal event will be judged as an abnormal reversal event.

12. A commission settlement anomaly data detection system based on gradient boosting decision tree, used to implement the commission settlement anomaly data detection method based on gradient boosting decision tree as described in any one of claims 1-11, characterized in that, The system includes: Settlement details generation module: used to extract key positioning fields from repayment records to construct scheme matching code vectors, match the scheme matching code vectors with the settlement scheme library to determine the target settlement scheme, calculate the monthly recovery rate of the project and perform recovery rate front-line verification, and perform dual-account period routing and commission calculation on repayment records that pass the recovery rate front-line verification according to the target settlement scheme to generate a settlement details data table; Anomaly feature construction module: Based on the settlement details data table, construct a step line chart of the project's monthly recovery rate and calculate the total projected area. Remove the pollution shadow area from the project's monthly recovery rate step line chart and construct anomaly judgment feature vector based on the pollution shadow area and the total projected area. Anomaly detection and handling module: This module is used to input anomaly detection feature vectors into a pre-built gradient boosting decision tree model to output anomaly scores. Based on the anomaly scores, it identifies and determines anomaly reversal events. For anomaly reversal events, it retrieves affected repayment records and performs batch recalculation of affected repayment records and circuit breaker processing of commission accounts.

Images