CN114202399A

CN114202399A - Intelligent approval method and related device

Info

Publication number: CN114202399A
Application number: CN202111519441.9A
Authority: CN
Inventors: 姚望; 宁义双; 叶秋成; 方首宇; 杨晓东; 邱翔; 宁可; 邹茂桃
Original assignee: Kingdee Software China Co Ltd
Current assignee: Kingdee Software China Co Ltd
Priority date: 2021-12-13
Filing date: 2021-12-13
Publication date: 2022-03-18

Abstract

The embodiment of the application discloses an intelligent approval method and a related device, and the method can effectively relieve the negative influence on model training caused by data imbalance (for example, in an actual scene, the number of documents which cannot be approved in an initial training sample is small) by performing data imbalance processing on extracted data characteristics; in addition, the target training sample is from the historical document data preprocessed by the enterprise, but not from the historical approval data of the specific enterprise, so that the target approval model obtained by final training can adaptively and more accurately audit whether the bill of the enterprise passes the audit, and the generalization effect and the user experience of the audit model are improved.

Description

Intelligent approval method and related device

Technical Field

The embodiment of the application relates to the technical field of internet, in particular to an intelligent approval method and a related device.

Background

In the enterprise financial work, the reimbursement bill (also called as a bill of lading) provided by a bill of lading needs to be audited, wherein the key point of auditing comprises auditing the reimbursement amount, the expense type, the attachment content and other information on the original bill so as to judge whether the bill has error information.

The method for auditing the documents in the current financial field mainly comprises the following steps: using specific historical approval data (for example, specifying document data of a certain enterprise as a training sample), training a machine learning model offline to assist financial staff in auditing the bill picking task; or, performing manual review.

In the practical application process, however, the off-line trained machine learning model is not suitable for some other enterprises due to the fact that specific historical approval data is used for model training, namely, the model is difficult to generalize; on the other hand, although the accuracy of the manual review method is high, the manual review method has high labor cost and is easy to have efficiency problems.

Disclosure of Invention

The embodiment of the application provides an intelligent approval method and a related device, which are used for solving the problem that the existing document auditing model is difficult to adaptively and accurately audit document data of different merchants.

In a first aspect, an embodiment of the present application provides an intelligent approval method, including:

acquiring a plurality of pieces of historical document data preprocessed by an enterprise in a historical time period;

respectively extracting data characteristics in each piece of historical document data to serve as a training sample;

carrying out data unbalance processing on all the training samples;

performing machine learning training on the initial auditing model by using the training samples subjected to unbalance processing to construct a target auditing model;

uploading at least one piece of document data to be examined of the enterprise to the target examination and verification model so as to obtain a target predicted value of each document to be examined and approved through the target examination and verification model, wherein the target predicted value is used for indicating whether the document data to be examined and approved passes the examination and verification of the target examination and verification model, and the document data to be examined and approved is preprocessed and contains the data characteristics.

Optionally, the data characteristics include reimbursement amount, charge type, credit rating of the bill taker, number of times the bill is rejected, organization creation time, and whether the image upload is acceptable.

Optionally, the performing data imbalance processing on all the training samples includes:

calculating an initial predicted value of each training sample through a LightGBM algorithm in the initial auditing model, wherein the initial predicted value is used for representing whether a historical document comprising the current initial training sample passes the approval of the initial auditing model;

and setting weights for training samples which are not approved in the loss function of the initial auditing model, wherein one training sample corresponds to one loss function.

Optionally, the machine learning training the initial audit model by using the training samples after the imbalance processing includes:

adding regular terms in all the loss functions to obtain a target function for adjusting parameters of the initial auditing model;

and performing machine learning training on the parameter-adjusted initial auditing model by using the training sample subjected to imbalance processing to finally obtain the target auditing model.

A second aspect of the embodiments of the present application provides an intelligent approval apparatus, including:

the system comprises an acquisition unit, a processing unit and a processing unit, wherein the acquisition unit is used for acquiring a plurality of pieces of historical document data preprocessed by an enterprise in a historical time period;

the processing unit is used for respectively extracting the data characteristics in each piece of historical document data to serve as a training sample;

the processing unit is further used for performing data imbalance processing on all the training samples;

the processing unit is further used for performing machine learning training on the initial auditing model by using the training samples after the unbalance processing so as to construct a target auditing model;

and the transmission unit is used for uploading at least one piece of document data to be examined of the enterprise to the target auditing model so as to obtain a target predicted value of each document to be examined and approved through the target auditing model, wherein the target predicted value is used for indicating whether the document data to be examined and approved passes the approval of the target auditing model, and the document data to be examined and approved is preprocessed and contains the data characteristics.

Optionally, the processing unit is specifically configured to calculate an initial predicted value of each training sample through a LightGBM algorithm in the initial audit model, where the initial predicted value is used to indicate whether a history document including a current initial training sample passes an approval of the initial audit model;

the processing unit is specifically configured to set a weight for each training sample that fails to pass the examination and approval in the loss function of the initial examination and approval model, where one training sample corresponds to one loss function.

Optionally, the processing unit is specifically configured to add a regular term to all the loss functions to obtain a target function for parameter adjustment of the initial audit model;

the processing unit is specifically configured to perform machine learning training on the parameter-adjusted initial auditing model by using the training samples subjected to the imbalance processing, so as to finally obtain the target auditing model.

A third aspect of the embodiments of the present application provides an intelligent approval apparatus, including:

the system comprises a central processing unit, a memory and an input/output interface;

the memory is a transient memory or a persistent memory;

the central processing unit is configured to communicate with the memory and execute the instructions in the memory to perform the method described in the first aspect of the embodiments of the present application or any specific implementation manner of the first aspect.

A fourth aspect of embodiments of the present application provides a computer-readable storage medium, including instructions that, when executed on a computer, cause the computer to perform a method as described in the first aspect of embodiments of the present application or any specific implementation manner of the first aspect.

A fifth aspect of embodiments of the present application provides a computer program product comprising instructions that, when run on a computer, cause the computer to perform a method as described in the first aspect of embodiments of the present application or any implementation manner of the first aspect.

According to the technical scheme, the embodiment of the application has the following advantages:

according to the intelligent approval method, the extracted data features are subjected to data imbalance processing, so that negative effects of data imbalance (for example, the number of documents which cannot be approved in an initial training sample is small in actual scene) on model training can be effectively relieved; in addition, the target training sample is from the historical document data preprocessed by the enterprise, but not from the historical approval data of the specific enterprise, so that the target approval model obtained by final training can adaptively and more accurately audit whether the bill of the enterprise passes the audit, and the generalization effect and the user experience of the audit model are improved.

Drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the description of the embodiments are briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments described in the present application, and other drawings can be obtained by those skilled in the art according to the drawings.

FIG. 1 is a schematic flow chart of an intelligent approval method according to an embodiment of the present application;

FIG. 2 is a schematic structural diagram of an intelligent approval apparatus according to an embodiment of the present application;

fig. 3 is another schematic structural diagram of the intelligent approval apparatus according to the embodiment of the present application.

Detailed Description

The initial audit model and the target audit model in the embodiments of the present application may be understood as LightGBM models. The LightGBM model belongs to a decision tree model, which is called a decision tree because such decision branches are drawn to form branches much like a tree. In practical applications, on the basis of the known occurrence probability of various situations, the probability that the expected value of the net current value is greater than or equal to zero is obtained by forming a decision tree, so that the project risk can be predicted and evaluated, and the feasibility of the decision can be analyzed, so that the decision tree analysis method can be understood as a graphical method which intuitively uses probability analysis.

In the embodiment of the application, the meaning of intelligent approval is that aiming at reimbursement documents provided by different enterprises, whether the documents pass the approval or not is accurately predicted by using a machine learning model adaptive to each enterprise. In addition, the true value is used to indicate whether a certain document passes the review (i.e. whether the information in the document is wrong or not) under the true condition, and specifically may be a result of the review that can be manually reviewed and the accuracy can be ensured, for example, a value of 0 is used to indicate that the document passes the manual review, and a value of 1 is used to indicate that the manual review does not pass.

Referring to fig. 1, a first aspect of the present application provides an embodiment of an intelligent approval method, including:

101. acquiring a plurality of pieces of historical document data preprocessed by an enterprise.

In the practical application process, in order to improve the applicability and the prediction accuracy of the auditing model for different enterprise bill withdrawals, the modeling data for constructing the target auditing model is from the preprocessed historical document data of the enterprise; in addition, in the historical documents generally selected as training samples, the number of documents that pass the audit (true value) exceeds the number of documents that do not pass the audit (true value). Illustratively, 13 ten thousand pieces of historical document data of a company for one year can be selected, wherein the ratio of the number of documents passing and failing is 12: 1, and each piece of historical document data is preprocessed.

102. And extracting data features in the historical document data to serve as training samples.

The information of the bill drawing is checked, mainly to determine whether the bill drawing contains error information, for example, to determine whether the data features such as the filled reimbursement amount are wrong, so that the data features in each piece of historical document data need to be extracted as a training sample; in one embodiment, the data characteristics reviewed by the review model may specifically be reimbursement amount, fee type, credit rating of the bill taker, number of times the bill is rejected, organization creation time, and whether the image upload is qualified:

1) and (6) reimbursing the amount. According to data analysis, the larger the reimbursement amount is, the higher the probability that the document cannot be audited (predicted value) in the target audit model is.

2) The fee types can specifically comprise transportation fees, travel reimbursement fees and telephone fees, and documents containing certain fee types (such as telephone fees) are not approved in a high proportion.

3) The bill of lading credit rating. In an actual working scene, the enterprise scores credit for each elevator, specifically, the credit can be divided into five grades from 1 (lowest) to 5 (highest), and the lower the credit grade is, the higher the probability that the elevator is not approved is.

4) And the number of unqualified bills is obtained. In an actual working scene, each department in an enterprise can carry out department-level approval on each document to check whether the information about the department in the bill of lading is qualified (for example, the attendance checking department checks whether the rate charge in the bill of lading is within a preset reimbursement cost range), and the more times that the bill of lading is not qualified, the higher the probability that the bill of lading cannot be approved is.

5) The creation time is organized. The bill submitter of a newly-built organization is easy to make mistakes due to unfamiliarity with bill submission requirements, so that the probability that the bill submission audit cannot pass is increased.

6) And whether the image uploading is qualified or not. Because part of documents require to submit images, if the specified images are not uploaded or the uploaded images do not meet the preset image requirements, the documents can be easily checked in the target checking model without passing the checking.

103. And carrying out data imbalance processing on all the training samples.

In an actual scene, in the historical document data serving as a training sample (which may be referred to as training data), the number of documents that cannot be audited is relatively small (for example, the number of documents that pass audition: the number of documents that do not pass audition is 12: 1), and therefore, sample data needs to be processed in an unbalanced manner. In a specific embodiment, the performing the data imbalance processing on the model level for each training sample may specifically include:

1031. an initial prediction value is calculated for each training sample.

Calculating an initial predicted value of each training sample through a LightGBM algorithm in the initial auditing model, wherein the initial predicted value is used for indicating whether a historical document comprising the current initial training sample passes the approval of the initial auditing model; the LightGBM algorithm may fit the residuals by constructing multiple decision trees, and the formula for calculating the initial prediction value is as follows:

wherein, y_iFor representing the ith training sample x_iIs used for representing the ith training sample x_iWhether the examination and approval of the initial examination and approval model is passed; f. of_tFor representing the t-th residual tree (CART tree); f represents a decision tree; t is a hyperparameter, which can be set empirically, and can be set to 100 for example; f. of_t(x_i) For representing the t-th residual tree pair x_iThe predicted value of the t-th round residue of (1). What is the training data for each round of residual tree? If, y_tSum for representing t CART residual trees(i.e., the final predicted value), Y represents the true value of a training sample x, then the training data of t +1 tree is (x, Y-Y)^(t)). In machine learning, a decision tree is a prediction model and represents a mapping relation between object attributes and object values, and each node in the tree represents an object; in the intelligent approval scenario of the embodiment of the present application, a leaf node corresponds to a field object (i.e., a data feature, such as an amount of reimbursement), each branch path represents a possible attribute value, so that the value of each leaf node corresponds to the value of the field object represented by the path traveled from the root node to the leaf node, which can be specifically understood as that the value of a leaf node represents the predicted value of a field in a training sample, i.e., the predicted result of whether the field passes the approval.

1032. And setting weight for the training sample which fails to pass the examination and approval.

By setting a larger weight for the failed training samples in the loss function of the initial auditing model, the negative influence caused by data imbalance (for example, the prediction of partial documents by the finally constructed target auditing model is inaccurate) can be effectively alleviated. Illustratively, the loss function (with weight n) of the ith training sample is as follows:

l(Y_i,y_i)＝-n·Y_i·logy_i-(1-Y_i)·log(1-y_i)；

where n is a weight set for a positive sample (true value is 1, i.e. a sample document that passes an audit in the true case), and y is_iThe method is used for representing the predicted value of the initial auditing model to the ith training sample (for example, when the predicted value is 0, the examination of the ith training sample is passed, and when the predicted value is 1, the examination is not passed), and Y_iThe real value is used for representing the ith training sample; the base of the logarithmic function in the above formula can be set empirically, and specifically can be set to 2, for example.

In the actual application process, other algorithms (e.g., Logistic Regression algorithm) other than the LightGBM may be selected to perform data imbalance processing on the training samples at the model level, and the specific details are not limited.

104. And training to obtain a target auditing model.

The basic principle of the embodiment of the application is that a machine learning model (LightGBM model) is adopted, and classification boundaries of whether a business document passes or does not pass are found in a training sample, namely a two-classification target auditing model used for examining and approving the business document is established; in the actual operation process, a machine learning library Scikit-leann (skleann) can be specifically adopted for modeling. In a specific implementation manner, the trained target audit model may specifically include:

1041. a regularization term is added to the loss function.

In practical application, a regularization term can be added to all the strip loss functions to effectively prevent overfitting as follows:

wherein obj represents a LightGBM objective function, and is used for adjusting model parameters of the initial audit model to construct a target audit model, which can be understood as a loss function; n is used to represent the number of training samples; t is used to represent the number of CART residual trees, and may specifically be set to 100, for example. l (Y)_i,y_i ^(t)) A loss function (with weights) for representing the ith training sample; omega (f)_t) The method is used for representing a regular term and can be used for solving the overfitting problem. In addition, the minimum depth of the CART tree in the LightGBM model in the embodiment of the present application may be 3, so as to avoid the problem of under-fitting due to too small depth of the tree.

1042. And training the initial auditing model after parameter adjustment.

And performing machine learning training on the parameter-adjusted initial auditing model by using the training sample subjected to imbalance processing to finally obtain a target auditing model.

As can be seen from the above description, the target auditing model in the embodiment of the present application is generated by using the historical document data of the enterprise as the training data for gradual online training, so that the method for predicting the to-be-audited document of the enterprise by using the historical document data has good adaptability and accuracy.

105. And uploading the data of the to-be-audited bill to a target auditing model for intelligent auditing.

At least one piece of document data to be examined and approved of an enterprise is pre-processed and then uploaded to a target examination and approval model, and a target predicted value of each document to be examined and approved can be calculated; the target predicted value is used for representing whether the to-be-examined document passes the examination and approval of the target examination and approval model. It should be noted that the target approval model mainly examines and approves whether the document passes the approval by examining and approving the data characteristics in each document data, so that each document data to be examined and approved should contain the data characteristics, and the high-efficiency and high-quality automatic approval effect of each enterprise document in the target approval model is realized.

Referring to fig. 2, a second aspect of the present application provides an embodiment of an intelligent approval apparatus, including:

the acquiring unit 201 is configured to acquire multiple pieces of historical document data preprocessed by an enterprise within a historical time period;

the processing unit 202 is configured to extract data features in each piece of historical document data as a training sample;

the processing unit 202 is further configured to perform data imbalance processing on all training samples;

the processing unit 202 is further configured to perform machine learning training on the initial audit model by using the training samples after the imbalance processing, so as to construct a target audit model;

the transmission unit 203 is configured to upload at least one piece of to-be-inspected receipt data of an enterprise to a target auditing model, so as to obtain a target predicted value of each to-be-inspected receipt through the target auditing model, where the target predicted value is used to indicate whether the to-be-inspected receipt data passes the approval of the target auditing model, and the to-be-inspected receipt data is preprocessed and includes data characteristics.

In this embodiment of the application, operations performed by each unit of the intelligent approval apparatus are similar to those described in the first aspect or any specific method embodiment of the first aspect, and details are not repeated here.

Referring to fig. 3, the intelligent approval apparatus 300 according to the embodiment of the present disclosure may include one or more Central Processing Units (CPUs) 301 and a memory 305, where the memory 305 stores one or more applications or data.

Memory 305 may be volatile storage or persistent storage, among other things. The program stored in memory 305 may include one or more modules, each of which may include a sequence of instructions operating on the smart approval device. Still further, central processor 301 may be configured to communicate with memory 305 to execute a series of instruction operations in memory 305 on intelligent approval device 300.

The smart approval device 300 may also include one or more power supplies 302, one or more wired or wireless network interfaces 303, one or more input-output interfaces 304, and/or one or more operating systems, such as Windows Server, Mac OS XTM, UnixTM, LinuxTM, FreeBSDTM, etc.

The central processing unit 301 may perform the operations performed by any of the foregoing first aspect or any of the specific method embodiments of the first aspect, which are not described in detail herein.

It should be understood that, in the various embodiments of the present application, the size of the serial number of each step does not mean the execution sequence, and the execution sequence of each step should be determined by its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

It is clear to those skilled in the art that, for convenience and brevity of description, the specific working processes of the above-described systems, apparatuses and units may refer to the corresponding processes in the foregoing method embodiments, and are not described herein again.

In the several embodiments provided in the present application, it should be understood that the disclosed apparatus and method may be implemented in other ways. For example, the above-described apparatus embodiments are merely illustrative, and for example, the division of the units is only one logical functional division, and there may be other divisions when actually implementing, for example, a plurality of units or components may be combined or integrated into another system or apparatus, or some features may be omitted, or not implemented. In addition, the shown or discussed mutual coupling or direct coupling or communication connection may be an indirect coupling or communication connection through some interfaces, devices or units, and may be in an electrical, mechanical or other form.

The units described as separate parts may or may not be physically separate, and parts displayed as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of the embodiment.

In addition, functional units in the embodiments of the present application may be integrated into one processing unit, or each unit may exist alone physically, or two or more units are integrated into one unit. The integrated unit can be realized in a form of hardware, and can also be realized in a form of a software functional unit.

The integrated unit, if implemented in the form of a software functional unit and sold or used as a stand-alone product, may be stored in a computer readable storage medium. Based on such understanding, the technical solution of the present application may be substantially implemented or contributed to by the prior art, or all or part of the technical solution may be embodied in a software product, which is stored in a storage medium and includes instructions for causing a computer device (which may be a personal computer, a service server, or a network device) to execute all or part of the steps of the method according to the embodiments of the present application. And the aforementioned storage medium includes: a U-disk, a removable hard disk, a read-only memory (ROM), a Random Access Memory (RAM), a magnetic disk or an optical disk, and the like.

Claims

1. An intelligent approval method, comprising:

carrying out data unbalance processing on all the training samples;

2. The smart approval method of claim 1, wherein the data characteristics comprise reimbursement amount, charge type, credit rating of the bill taker, number of times the bill is not approved, organization creation time, and whether the image upload is acceptable.

3. The smart approval method of claim 1, wherein the data imbalance processing of all of the training samples comprises:

4. The smart approval method of claim 3, wherein the machine learning training of the initial audit model using the imbalance processed training samples comprises:

5. An intelligent approval device, comprising:

6. The intelligent approval apparatus according to claim 5, wherein the processing unit is specifically configured to calculate an initial predicted value of each training sample through a LightGBM algorithm in the initial approval model, where the initial predicted value is used to indicate whether a history document including a current initial training sample passes the approval of the initial approval model;

7. The intelligent approval apparatus according to claim 6, wherein the processing unit is specifically configured to add a regular term to all the loss functions to obtain an objective function for performing parameter adjustment on the initial audit model;

8. An intelligent approval device, comprising:

the memory is a transient memory or a persistent memory;

the central processor is configured to communicate with the memory and execute the operations of the instructions in the memory to perform the method of any of claims 1 to 4.

9. A computer-readable storage medium comprising instructions which, when executed on a computer, cause the computer to perform the method of any one of claims 1 to 4.

10. A computer program product comprising instructions for causing a computer to perform the method according to any one of claims 1 to 4 when the computer program product is run on a computer.