Machine learning-based cancel-after-verification receipt detection method and device, and medium
Through the machine learning-based verification document detection method, matching the write-off amount prediction model for automated review, the efficiency and accuracy problems of traditional methods in the write-off scenario are solved, and a high accuracy and compliance verification process is achieved.
Patent Information
- Application Number
- CN202411185531.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-27
- Publication Date
- 2024-12-20
AI Technical Summary
The traditional verification and inspection methods that rely on manual rules and experience cannot meet the needs of batch verification and rapid verification in various verification scenarios, and are inefficient, prone to missed inspections and low audit accuracy.
Using machine learning-based verification document detection method, by obtaining the information of documents to be detected, matching the pre-constructed verification amount prediction model, performing verification amount prediction and detection, identifying abnormal documents, and realizing automated processing and intelligent review.
It improves the accuracy and processing efficiency of the write-off amount prediction, reduces manual intervention, ensures the accuracy and compliance of the write-off process, adapts to the needs of different write-off scenarios, and promptly discovers abnormal documents.
Smart Images
Figure CN119168787A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of financial management technology, and in particular to a method, device and medium for detecting written-off documents based on machine learning. Background Art
[0002] Write-offs cover asset depreciation, prepaid accounts, bad debt reserves and other aspects, which directly affect the accuracy and compliance of corporate financial statements. When conducting financial management, companies need to conduct strict audits on the write-off amounts to ensure the rational use of funds. Through the write-off amount audit and anomaly detection methods, abnormal write-off records can be automatically identified to improve the accuracy and efficiency of the audit.
[0003] Traditional write-off amount audits rely on manual rules and experience. When faced with increasingly complex and large amounts of financial data, there are problems such as low efficiency, easy omissions, and low audit accuracy. In addition, when facing a rapidly changing market environment, enterprises have various write-off scenarios. For different write-off types and write-off entities, faster and more accurate write-off audit methods are needed to cope with changing business needs and risk challenges. Therefore, traditional write-off detection methods that rely on manual rules and experience cannot meet the write-off needs of batch write-offs and rapid write-offs in various write-off scenarios. Summary of the invention
[0004] One or more embodiments of this specification provide a method, device and medium for detecting write-off documents based on machine learning, which are used to solve the following technical problems: Traditional write-off detection methods that rely on manual rules and experience cannot meet the write-off requirements of batch write-offs and rapid write-offs in various write-off scenarios.
[0005] One or more embodiments of this specification adopt the following technical solutions:
[0006] One or more embodiments of the present specification provide a method for detecting cancellation documents based on machine learning, the method comprising: obtaining multiple cancellation documents to be detected, and determining the document information of each of the cancellation documents to be detected, wherein the document information comprises the cancellation type, the amount to be cancelled of the document, and the cancellation subject; matching in a pre-built cancellation amount prediction model library according to the cancellation type and the cancellation subject, and determining at least one corresponding cancellation amount prediction model; predicting each of the cancellation documents to be detected by each of the cancellation amount prediction models, and determining a cancellation amount prediction data set corresponding to each of the cancellation documents to be detected, so as to determine cancellation amount reference data based on the cancellation amount prediction data set and pre-acquired business rules; and performing cancellation amount detection on the amount to be cancelled of the cancellation documents to be detected based on the cancellation amount reference data, so as to identify abnormal cancellation documents.
[0007] One or more embodiments of this specification provide a write-off document detection device based on machine learning, including:
[0008] at least one processor; and,
[0009] a memory communicatively connected to the at least one processor; wherein,
[0010] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the above method.
[0011] One or more embodiments of the present specification provide a non-volatile computer storage medium storing computer executable instructions, wherein the computer executable instructions are configured to execute the above method.
[0012] At least one of the above technical solutions adopted in the embodiments of this specification can achieve the following beneficial effects: through the above technical solution, by matching in the write-off amount prediction model library according to the write-off type and the write-off subject, it can be ensured that the most appropriate prediction model is selected for each write-off document to be tested, and the targeted prediction method can more accurately reflect the write-off amount characteristics under different write-off types and write-off subjects, thereby improving the accuracy of the prediction and adapting to the detection requirements of different write-off scenarios; the entire process from the acquisition of document information, matching of prediction models, prediction of write-off amounts to detection of write-off amounts, has realized automated processing, reducing manual intervention and improving processing efficiency It also improves the accuracy and reliability of the write-off process in an intelligent way. Based on the write-off amount forecast data set and the write-off amount reference data determined by business rules, it can detect the write-off amount of the document to be written off, and can promptly discover abnormal write-off documents that do not conform to the forecast value or business rules, and determine the write-off amount reference data in combination with the forecast amount, ensuring that the write-off amount reference data conforms to the write-off characteristic rules. It avoids the workload generated by comparing a single write-off document through automatic anomaly detection, and can meet the detection needs of a large number of write-off documents. Through the detection of the write-off amount and the identification of abnormal write-off documents, enterprises can ensure the compliance of the write-off process. BRIEF DESCRIPTION OF THE DRAWINGS
[0013] In order to more clearly illustrate the technical solutions in the embodiments of this specification or the prior art, the following briefly introduces the drawings required for use in the embodiments or the prior art description. Obviously, the drawings described below are only some embodiments recorded in this specification. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative labor. In the drawings:
[0014] Figure 1A flowchart of a method for detecting cancellation documents based on machine learning provided in an embodiment of this specification;
[0015] Figure 2 A schematic diagram of the structure of a verification document detection device based on machine learning provided in an embodiment of this specification. DETAILED DESCRIPTION
[0016] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the drawings in the embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this specification, not all of the embodiments. Based on the embodiments of this specification, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this specification.
[0017] The embodiment of this specification provides a method for detecting verification documents based on machine learning. It should be noted that the execution subject in the embodiment of this specification can be a server or any device with data processing capabilities. Figure 1 A flowchart of a method for detecting cancellation documents based on machine learning is provided in an embodiment of this specification, such as Figure 1 As shown, it mainly includes the following steps:
[0018] Step S101, obtaining a plurality of documents to be checked and cancelled, and determining the document information of each document to be checked and cancelled.
[0019] The document information includes the write-off type, the amount to be written off and the write-off subject;
[0020] In one embodiment of the present specification, in the write-off document generation node, under the trigger of write-off document generation, multiple write-off documents to be detected are obtained, and the document information of each write-off document to be detected is determined, and the document information includes the write-off type, the amount to be written off of the document, and the write-off subject; the write-off type: usually a descriptive text field, indicating the type of write-off of the document, such as "expense reimbursement", "accounts receivable write-off", etc. The amount to be written off of the document is a numerical field, indicating the amount to be written off on the document; the write-off subject can be an individual, department or organization name, indicating who or which entity will perform the write-off operation, and the write-off subject here includes two types: external transaction subject and internal transaction subject.
[0021] Step S102: Matching is performed in a pre-built write-off amount prediction model library according to the write-off type and the write-off subject, and at least one corresponding write-off amount prediction model is determined.
[0022] According to the write-off type and the write-off subject, matching is performed in a pre-built write-off amount prediction model library respectively, and before determining the corresponding at least one write-off amount prediction model, the method further includes: acquiring multiple historical write-off data, constructing a historical write-off data set, and determining the write-off type and write-off subject information corresponding to each of the historical write-off data, the write-off subject information including any one of an external transaction subject and an internal transaction subject; according to the write-off type corresponding to each of the historical write-off data, classifying the multiple historical write-off data, and determining the same type of historical write-off data set corresponding to each of the write-off types; according to the write-off subject information corresponding to each of the historical write-off data, classifying the multiple historical write-off data Classify and determine the historical write-off data set of the same subject corresponding to each of the write-off subject information; use the historical write-off data set, each of the same type of historical write-off data set and each of the same subject historical write-off data set to perform model training to obtain a general write-off amount prediction model, a first write-off amount prediction model corresponding to each of the write-off types and a second write-off amount prediction model corresponding to each of the write-off subject information; set a general identifier for the general write-off amount prediction model, and set a write-off type identifier for each of the first write-off amount prediction models according to the write-off type, and set a subject identifier for each of the second write-off amount prediction models according to the write-off subject information, so as to store them in the write-off amount prediction model library.
[0023] In one embodiment of the present specification, a large amount of historical write-off data is collected from various data sources (such as financial systems, transaction records, etc.). A large amount of historical write-off data is organized into a unified format to form a historical write-off data set. Write-off data usually contains a large number of transaction records, which may contain duplicate data or inconsistent formats. Data cleaning technology is used to remove duplicate records and inconsistent formats to ensure data consistency and accuracy. The original write-off data set D is input, and after removing duplicate records, format normalization (such as date format, amount format) and processing missing values, the cleaned write-off data set D is output. clean .
[0024] In addition, there may be outliers in the write-off data, which may be caused by data entry errors or business anomalies. Statistical methods and machine learning model prediction methods can be used to identify and process outliers to improve data quality and the accuracy of model prediction. clean , by using statistical methods (such as the 3σ principle) to identify outliers, using machine learning models (such as isolation forests) to detect outliers, and processing outliers (such as replacement, deletion) based on the detection results, outputting a write-off data set D without outliers processedIn order to improve the training effect of the machine learning model, the write-off data is transformed and normalized. For example, the amount data is logarithmically transformed or standardized to eliminate the dimensional effects between different features. Input the write-off data set D without outliers processed , perform logarithmic transformation on the amount data, and use standardization or minimum-maximum scaling normalization to output the transformed write-off data set D transformed .
[0025] For each piece of historical write-off data, determine its write-off type (such as returns, discounts, bad debts, etc.) and write-off subject information (external transaction subjects such as customers, internal transaction subjects such as departments or employees). According to the write-off type, the historical write-off data is divided into different subsets, each subset contains all data of the same write-off type, forming a historical write-off data set of the same type. Further, according to the write-off subject information (external or internal), each historical write-off data set of the same type is further subdivided to form a historical write-off data set of the same subject.
[0026] The entire historical write-off data set is used for training to obtain a general write-off amount prediction model that can handle all types and entities. The general write-off amount prediction model is trained based on the entire historical write-off data set. This data set contains all types of write-off data and records of different write-off entities. By comprehensively analyzing this data, the model can learn the general relationship between the write-off amount and various possible influencing factors. The advantage lies in its wide applicability. Since it is not optimized for a specific type or entity, it can handle any type of write-off request.
[0027] For each write-off type, the corresponding historical write-off data set of the same type is used for training to obtain the respective first write-off amount prediction model, which is used for the prediction of the specific write-off type. The type-specific write-off amount prediction model is trained separately for each write-off type. For each write-off type (such as returns, discounts, bad debts, etc.), there will be a dedicated model to handle the write-off request of that type. These models are trained using the corresponding historical write-off data set of the same type, so they can have a deeper understanding of the characteristics and rules of this type of write-off. Because these models focus on a specific type of write-off, they can provide more accurate prediction results when processing this type of write-off request.
[0028] For each write-off subject information, the corresponding historical write-off data set of the same subject is used for training to obtain the second write-off amount prediction model, so as to provide more accurate predictions for specific write-off subjects. The subject-specific write-off amount prediction model is trained based on the write-off subject information. The write-off subject can be an external trading partner (such as a customer) or an internal department or employee. For each write-off subject, there will be a dedicated model to predict its write-off amount. The model is trained using the corresponding historical write-off data set of the same subject, so it can capture the unique behavior and pattern of the subject in the write-off process. Since these models are optimized for specific subjects, they can provide more personalized prediction results when processing write-off requests related to the subject.
[0029] It should be noted that the above-mentioned model training and optimization phase uses advanced machine learning algorithms such as XGBoost. The XGBoost algorithm is known for its efficient parallel computing capabilities and excellent predictive performance, and can process large-scale data and improve the generalization ability of the model. During the model training process, the present invention further optimizes the performance of the model through cross-validation and parameter tuning to ensure outstanding performance in different write-off scenarios. For example, by optimizing the model's learning rate, tree depth, and number of leaf nodes, the accuracy and stability of the model when processing write-off data are improved.
[0030] Setting a general identifier for the general write-off amount prediction model, setting a write-off type identifier for each first write-off amount prediction model, and setting a subject identifier for each second write-off amount prediction model will help to quickly identify and use the correct model later. Store all trained models in the write-off amount prediction model library so that they can be matched and called later based on the write-off type and write-off subject information.
[0031] According to the write-off type and the write-off subject, matching is performed in a pre-built write-off amount prediction model library respectively to determine at least one corresponding write-off amount prediction model, specifically including: by using the write-off type and the write-off subject, model identification is performed in the write-off amount prediction model library respectively to determine whether there is a first write-off amount prediction model corresponding to the write-off type, and to determine whether there is a second write-off amount prediction model corresponding to the write-off subject; when the first write-off amount prediction model exists and the second write-off amount prediction model exists, based on the first write-off amount prediction model and the second write-off amount prediction model, at least one corresponding write-off amount prediction model is determined; when the first write-off amount prediction model does not exist and the second write-off amount prediction model exists, a general write-off amount prediction model is obtained, and based on the general write-off amount prediction model and the second write-off amount prediction model, at least one corresponding write-off amount prediction model is determined; when the second write-off amount prediction model does not exist and the first write-off amount prediction model exists, a general write-off amount prediction model is obtained, and based on the general write-off amount prediction model and the first write-off amount prediction model, at least one corresponding write-off amount prediction model is determined.
[0032] In one embodiment of the present specification, according to the provided write-off type, the model library is searched to see whether there is a corresponding first write-off amount prediction model (i.e., type-specific model). If found, the model is recorded. Next, according to the provided write-off subject information, the model library is searched to see whether there is a corresponding second write-off amount prediction model (i.e., subject-specific model). If found, the model is recorded. If both the first write-off amount prediction model and the second write-off amount prediction model exist, the final prediction model combination can be determined based on the two models. If the type-specific first write-off amount prediction model does not exist, but the subject-specific second write-off amount prediction model exists, then a general write-off amount prediction model can be obtained and used in combination with the second write-off amount prediction model. Similarly, if the subject-specific second write-off amount prediction model does not exist, but the type-specific first write-off amount prediction model exists, then the general write-off amount prediction model can be used in combination with the first write-off amount prediction model. After the final prediction model combination is determined, these models can be used to predict new write-off requests, the characteristics of the write-off requests are respectively input into the models, and the predicted amount output by each prediction model is obtained.
[0033] Through the above technical solution, by distinguishing the write-off type and the write-off subject, and adopting type-specific and subject-specific forecasting models, the amount forecasting needs under different write-off scenarios can be more accurately reflected. This fine-grained model design can capture the unique write-off characteristics under a specific type or subject, thereby improving the accuracy and reliability of the forecast. The strategy not only considers type-specific and subject-specific models, but also designs a general write-off amount forecasting model as an alternative. When a specific model does not exist, it can be replaced by a general model to ensure the flexibility of the system and its adaptability to different write-off scenarios. This design reduces the forecast interruption caused by the lack of a model and improves the robustness of the system; the entire forecasting process is highly automated, and the output of the model selection and forecast amount are all completed automatically by the system, reducing the need for manual intervention, improving work efficiency, and reducing the risk of human error, making the forecast of the write-off amount faster and more accurate; by accurately forecasting the write-off amount, enterprises can better plan capital flow and cost control and optimize resource allocation.
[0034] Step S103, predicting each to-be-detected write-off document through each write-off amount prediction model, determining a write-off amount prediction data set corresponding to each to-be-detected write-off document, and determining write-off amount reference data based on the write-off amount prediction data set and pre-acquired business rules.
[0035] Through each of the write-off amount prediction models, predictions are made for each of the to-be-tested write-off documents respectively, and a write-off amount prediction data set corresponding to each of the to-be-tested write-off documents is determined, specifically including: constructing write-off document characteristic data corresponding to each of the to-be-tested write-off documents based on the document characteristic information of each of the to-be-tested write-off documents, wherein the write-off document characteristic data includes time characteristic data, write-off type characteristic data, write-off subject behavior characteristic data and transaction amount data; inputting the write-off document characteristic data into each of the write-off amount prediction models respectively, determining multiple write-off amount prediction data corresponding to each of the to-be-tested write-off documents, so as to construct the write-off amount prediction data set.
[0036] Based on the write-off amount prediction data set and pre-acquired business rules, write-off amount reference data is determined, specifically including: determining the first write-off amount prediction data, the second write-off amount prediction data and the model identifier corresponding to each write-off amount prediction data in the write-off amount prediction data set; correcting the first write-off amount prediction data and the second write-off amount prediction data through the model identifier corresponding to each of the write-off amount prediction data to determine the corrected write-off amount prediction data; obtaining the business rules corresponding to each of the write-off documents to be tested, wherein the business rules include any one or more of the upper and lower limits of the write-off amount and the write-off ratio; determining the write-off amount reference data based on the corrected write-off amount prediction data and the pre-acquired business rules.
[0037] In one embodiment of the present specification, two or more (in this case, the first write-off amount prediction data and the second write-off amount prediction data) prediction data generated by different models (trained based on different data sets) are identified from a write-off amount prediction data set. At the same time, it is also necessary to determine the model identifier corresponding to each prediction data for subsequent differentiation and correction of the prediction data. Since different models may be trained based on different assumptions and data sets, their prediction results may differ. In order to obtain more accurate prediction results, these prediction data can be corrected by applying certain correction rules or algorithms through the model identifier corresponding to each prediction data. The correction methods include but are not limited to: weighted average, weight adjustment based on model performance, compensation for model deviations, etc., to obtain corrected write-off amount prediction data.
[0038] Obtain the business rules corresponding to each document to be written off, including the upper and lower limits of the write-off amount, the write-off ratio under specific conditions, the write-off policy for specific customers, etc. The acquisition of business rules may come from system configuration, database query or manual input. Finally, based on the revised write-off amount forecast data and the pre-acquired business rules, determine the reference data for the write-off amount. Match and calculate the forecast data with the business rules, and match the forecast data with the business rules to determine the rule conditions that the forecast data needs to meet. According to the matched rules, calculate the forecast data to obtain a reasonable write-off amount, including simple mathematical operations (such as multiplication and division), and may also include conditional judgment, loop and other logical processing. For example, if the forecast amount exceeds the upper limit set by the business rules, it may need to be adjusted to the upper limit value.
[0039] The first write-off amount forecast data and the second write-off amount forecast data are corrected by using the model identifier corresponding to each of the write-off amount forecast data to determine the corrected write-off amount forecast data, specifically including: performing model deviation evaluation on each of the write-off amount forecast models to determine the deviation adjustment factor corresponding to each of the write-off amount forecast models, and determining the corresponding designated deviation adjustment factor by using the model identifier corresponding to each of the write-off amount forecast data; respectively correcting the first write-off amount forecast data and the second write-off amount forecast data by using each of the designated deviation adjustment factors to determine the first corrected write-off amount forecast data and the second corrected write-off amount forecast data; and integrating the first corrected write-off amount forecast data and the second corrected write-off amount forecast data to determine the corrected write-off amount forecast data.
[0040] In one embodiment of the present specification, the compensation of model deviation is taken as an example for explanation. First, the deviation of each prediction model needs to be evaluated. In the model training stage, the prediction results of the model are compared with the actual write-off amount data to calculate the prediction error. Based on the evaluation results of the model deviation, a corresponding compensation strategy can be formulated to reduce or eliminate the impact of the model deviation on the prediction results. For example, a deviation adjustment factor can be used, which is adjusted according to the degree of deviation of the model and applied to the prediction results. For example, if a model tends to underestimate the write-off amount, a deviation adjustment factor greater than 1 can be set for it so as to add a certain amount to the prediction result to compensate for its underestimated deviation. On the contrary, if the model tends to overestimate the write-off amount, a deviation adjustment factor less than 1 can be used to reduce the predicted amount.
[0041] After the deviation adjustment factor is determined, it can be applied to the original forecast amount to calculate the revised forecast amount. The specific calculation formula may be as follows: first revised forecast amount = first write-off amount forecast data × first deviation adjustment factor, second revised forecast amount = second write-off amount forecast data × second deviation adjustment factor. The first revised write-off amount forecast data and the second revised write-off amount forecast data are integrated to determine the revised write-off amount forecast data, for example, by averaging.
[0042] Through the above technical solution, by constructing write-off document feature data containing multi-dimensional features (time, type, subject behavior and transaction amount) and inputting it into multiple write-off amount prediction models, it is possible to more comprehensively consider various factors affecting the write-off amount, thereby improving the accuracy and reliability of the prediction; multiple prediction models are used for prediction, and the prediction results are corrected according to the model identifier, so that the system can flexibly respond to different types of write-off documents and different business scenarios; on the basis of the prediction results, the write-off amount reference data is determined in combination with pre-acquired business rules (such as the upper and lower limits of the write-off amount, the write-off ratio, etc.), which can ensure that the prediction results meet the actual business needs and financial management specifications of the enterprise; through the prediction and correction process, and the final determination in combination with business rules, the risks caused by inaccurate predictions can be reduced to a certain extent.
[0043] Step S104: Based on the write-off amount reference data, the write-off amount of the document to be written off is checked to identify abnormal write-off documents.
[0044] Based on the write-off amount reference data, a write-off amount detection is performed on the document to be written off of the document to be written off to identify abnormal write-off documents, specifically including: training a pre-built isolation forest model through the write-off amount reference data to determine a write-off amount anomaly detection model; inputting the document to be written off of each document to be written off into the write-off amount anomaly detection model to determine the write-off document status corresponding to each document to be written off, wherein the write-off document status includes any one of an abnormal write-off status and a normal write-off status.
[0045] The amount to be written off of each document to be detected is input into the write-off amount anomaly detection model to determine the document status corresponding to each document to be detected, specifically including: inputting the amount to be written off of the document into the write-off amount anomaly detection model, calculating the number of edges from the root node to the leaf node to determine the path length corresponding to the amount to be written off of the document; determining the anomaly score of each data point corresponding to the amount to be written off of the document according to the path length corresponding to each amount to be written off of the document; and determining the document status corresponding to each amount to be written off of the document according to the anomaly score of the data point.
[0046] In one embodiment of the present specification, the write-off amount data of historical write-off documents are collected and sorted, and a training data set is constructed by using the write-off amount data of historical write-off documents and the write-off amount reference data corresponding to each write-off document to be detected to ensure the accuracy of the data set. Using the Isolation Forest (iForest) algorithm to detect anomalies in the write-off amount is an effective method, especially suitable for processing high-dimensional data and large-scale data sets. The Isolation Forest algorithm isolates abnormal points by constructing multiple isolation trees (iTrees). The basic idea is that abnormal points are usually easier to be isolated (i.e., reach leaf nodes faster). The isolation forest model is trained using the training set data, and the model will learn the characteristic distribution of normal write-off amounts. The amount to be written off and other relevant features of each write-off document to be detected are input into the trained isolation forest model. According to the model output, the write-off status of each write-off document to be detected is determined. The data of the write-off document to be detected is input into the trained isolation forest model. For each document to be checked for write-off, the model calculates its path length (i.e., the number of edges from the root node to the leaf node). The path length reflects the ease with which the data point is isolated. Based on the path length, the model gives an anomaly score for each data point (usually the normalized value of the path length). The higher the anomaly score, the more likely the data point is to be an anomaly. Set an anomaly threshold and determine which documents are abnormal write-off documents based on the anomaly score and threshold. Usually, the model outputs an anomaly score and determines whether the document is a normal write-off or an abnormal write-off based on the preset threshold.
[0047] By training with reference data of write-off amounts in mind, the model can learn the characteristic distribution of normal write-off amounts, thereby more effectively identifying abnormal write-off amounts that deviate significantly from normal patterns. Through the write-off amount anomaly detection model, enterprises can automatically identify and mark write-off documents that may have problems, ensure the compliance of the write-off process, and prevent improper write-off behavior. The automated anomaly detection process can significantly reduce the workload of manual review, improve decision-making efficiency, and process a large number of write-off documents more quickly while maintaining sensitivity to abnormal situations.
[0048] Through the above technical solution, by matching the write-off type and the write-off subject in the write-off amount prediction model library, it can be ensured that the most appropriate prediction model is selected for each write-off document to be tested. The targeted prediction method can more accurately reflect the write-off amount characteristics under different write-off types and write-off subjects, thereby improving the accuracy of the prediction and adapting to the detection needs of different write-off scenarios; the entire process from the acquisition of document information, matching of prediction models, prediction of the write-off amount to the detection of the write-off amount has been automated, reducing manual intervention, improving processing efficiency, and improving the write-off process in an intelligent way. To ensure the accuracy and reliability of the process, based on the write-off amount forecast data set and the write-off amount reference data determined by business rules, the write-off amount of the document can be checked for write-off amount, and abnormal write-off documents that do not conform to the forecast value or business rules can be discovered in a timely manner. The write-off amount reference data is determined in combination with the forecast amount, ensuring that the write-off amount reference data conforms to the write-off characteristic rules. Automatic anomaly detection can be used to avoid the workload generated by comparing a single write-off document, and can meet the detection needs of a large number of write-off documents. Through the detection of write-off amounts and the identification of abnormal write-off documents, enterprises can ensure the compliance of the write-off process.
[0049] The present specification also provides a machine learning-based verification document detection device, such as Figure 2 As shown, the device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the above method.
[0050] The embodiments of the present specification also provide a non-volatile computer storage medium storing computer executable instructions, wherein the computer executable instructions are configured to execute the above method.
[0051] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device, equipment, and non-volatile computer storage medium embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.
[0052] The above is a description of a specific embodiment of the specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.
[0053] The devices and media provided in the embodiments of this specification correspond one-to-one to the methods. Therefore, the devices and media also have similar beneficial technical effects as the corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the devices and media will not be repeated here.
[0054] Those skilled in the art will appreciate that the embodiments of this specification may be provided as methods, systems, or computer program products. Therefore, this specification may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, this specification may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.
[0055] This specification is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of this specification. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.
[0056] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.
[0057] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.
[0058] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.
[0059] The memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. The memory is an example of a computer-readable medium.
[0060] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.
[0061] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.
[0062] The above description is only one or more embodiments of this specification and is not intended to limit this specification. For those skilled in the art, one or more embodiments of this specification may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of one or more embodiments of this specification shall be included in the scope of the claims of this specification.
Claims
1. A method for detecting cancellation documents based on machine learning, characterized in that: The method comprises: Acquire multiple documents to be checked and written off, and determine the document information of each document to be checked and written off, wherein the document information includes the type of writing off, the amount of the document to be written off, and the writing off subject; According to the write-off type and the write-off subject, matching is performed in a pre-built write-off amount prediction model library to determine at least one corresponding write-off amount prediction model; By using each of the write-off amount prediction models, each of the to-be-tested write-off documents is predicted respectively, and a write-off amount prediction data set corresponding to each of the to-be-tested write-off documents is determined, so as to determine the write-off amount reference data based on the write-off amount prediction data set and the pre-acquired business rules; Based on the write-off amount reference data, a write-off amount detection is performed on the document to be detected for write-off, so as to identify abnormal write-off documents.
2. The method for detecting cancellation documents based on machine learning according to claim 1, characterized in that: According to the write-off type and the write-off subject, matching is performed in a pre-built write-off amount prediction model library to determine at least one corresponding write-off amount prediction model, specifically including: Model identification is performed in the write-off amount prediction model library respectively based on the write-off type and the write-off subject, so as to determine whether there is a first write-off amount prediction model corresponding to the write-off type, and to determine whether there is a second write-off amount prediction model corresponding to the write-off subject; When the first write-off amount prediction model exists and the second write-off amount prediction model exists, determining at least one corresponding write-off amount prediction model based on the first write-off amount prediction model and the second write-off amount prediction model; When the first write-off amount prediction model does not exist and the second write-off amount prediction model exists, obtaining a general write-off amount prediction model, and determining at least one corresponding write-off amount prediction model based on the general write-off amount prediction model and the second write-off amount prediction model; When the second write-off amount prediction model does not exist and the first write-off amount prediction model exists, a general write-off amount prediction model is obtained, and at least one corresponding write-off amount prediction model is determined based on the general write-off amount prediction model and the first write-off amount prediction model.
3. The method for detecting cancellation documents based on machine learning according to claim 1, characterized in that: According to the write-off type and the write-off subject, before matching in a pre-built write-off amount prediction model library and determining at least one corresponding write-off amount prediction model, the method further includes: Acquire multiple historical write-off data, construct a historical write-off data set, and determine the write-off type and write-off subject information corresponding to each of the historical write-off data, wherein the write-off subject information includes any one of an external transaction subject and an internal transaction subject; Classifying the plurality of historical write-off data according to the write-off type corresponding to each of the historical write-off data, and determining a historical write-off data set of the same type corresponding to each of the write-off types; Classifying the plurality of historical write-off data according to the write-off subject information corresponding to each of the historical write-off data, and determining a historical write-off data set of the same subject corresponding to each of the write-off subject information; Using the historical write-off data set, each of the same type of historical write-off data sets, and each of the same subject historical write-off data sets, respectively, to perform model training, to obtain a general write-off amount prediction model, a first write-off amount prediction model corresponding to each of the write-off types, and a second write-off amount prediction model corresponding to each of the write-off subject information; A general identifier is set for the general write-off amount prediction model, and a write-off type identifier is set for each of the first write-off amount prediction models according to the write-off type, and a subject identifier is set for each of the second write-off amount prediction models according to the write-off subject information, so as to be stored in the write-off amount prediction model library.
4. The method for detecting cancellation documents based on machine learning according to claim 1, characterized in that: By using each of the write-off amount prediction models, each of the to-be-detected write-off documents is predicted respectively, and a write-off amount prediction data set corresponding to each of the to-be-detected write-off documents is determined, specifically including: According to the document characteristic information of each of the to-be-detected write-off documents, construct the write-off document characteristic data corresponding to each of the to-be-detected write-off documents, wherein the write-off document characteristic data includes time characteristic data, write-off type characteristic data, write-off subject behavior characteristic data and transaction amount data; The write-off document characteristic data is input into each of the write-off amount prediction models respectively, and a plurality of write-off amount prediction data corresponding to each of the to-be-detected write-off documents is determined to construct the write-off amount prediction data set.
5. The method for detecting cancellation documents based on machine learning according to claim 1, characterized in that: Based on the write-off amount prediction data set and the pre-acquired business rules, the write-off amount reference data is determined, specifically including: Determine the first write-off amount forecast data, the second write-off amount forecast data, and the model identifier corresponding to each write-off amount forecast data in the write-off amount forecast data set; Modifying the first write-off amount forecast data and the second write-off amount forecast data by using the model identifier corresponding to each of the write-off amount forecast data to determine modified write-off amount forecast data; Obtaining the business rules corresponding to each of the to-be-detected write-off documents, wherein the business rules include any one or more of the upper and lower limits of the write-off amount and the write-off ratio; Based on the revised write-off amount forecast data and pre-acquired business rules, write-off amount reference data is determined.
6. The method for detecting cancellation documents based on machine learning according to claim 5, characterized in that: The first write-off amount forecast data and the second write-off amount forecast data are modified according to the model identifier corresponding to each of the write-off amount forecast data to determine the modified write-off amount forecast data, specifically including: Performing a model deviation assessment on each of the write-off amount prediction models to determine a deviation adjustment factor corresponding to each of the write-off amount prediction models, so as to determine a corresponding designated deviation adjustment factor through a model identifier corresponding to each of the write-off amount prediction data; The first write-off amount forecast data and the second write-off amount forecast data are respectively revised by each of the specified deviation adjustment factors to determine first revised write-off amount forecast data and second revised write-off amount forecast data; The first revised write-off amount forecast data and the second revised write-off amount forecast data are integrated to determine revised write-off amount forecast data.
7. The method for detecting cancellation documents based on machine learning according to claim 1, characterized in that: Based on the write-off amount reference data, a write-off amount detection is performed on the document to be written off amount of the document to be written off to identify abnormal write-off documents, specifically including: The pre-built isolation forest model is trained using the write-off amount reference data to determine a write-off amount anomaly detection model; The document amount to be written off of each of the documents to be written off is input into the write-off amount anomaly detection model to determine the write-off document status corresponding to each of the documents to be written off, wherein the write-off document status includes any one of an abnormal write-off status and a normal write-off status.
8. The method for detecting cancellation documents based on machine learning according to claim 7, characterized in that: Inputting the document amount to be written off of each of the documents to be written off to be detected into the write-off amount anomaly detection model, and determining the document status of each of the documents to be written off corresponding to the documents to be written off, specifically includes: Inputting the amount of the document to be written off into the write-off amount anomaly detection model, calculating the number of edges from the root node to the leaf node to determine the path length corresponding to the amount of the document to be written off; Determine the abnormal score of the corresponding data point corresponding to the amount to be written off of each document according to the path length corresponding to the amount to be written off of each document; The write-off document status corresponding to the amount to be written off of each of the documents is determined based on the anomaly score of the data point.
9. A verification document detection device based on machine learning, characterized in that: The device comprises: at least one processor; and, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can perform the method according to any one of claims 1 to 8.
10. A non-volatile computer storage medium storing computer executable instructions, characterized in that: The computer executable instructions are configured to execute the method according to any one of claims 1 to 8.