Enterprise data auditing methods, devices, electronic equipment and storage media

Abstract

This application relates to the field of data detection, specifically to a method, apparatus, electronic device, and storage medium for enterprise data auditing. The enterprise data auditing method includes: acquiring enterprise data to be audited from a target enterprise; extracting information from the enterprise data to obtain multiple key pieces of first-entity information; constructing an information relationship graph of the first-entity information based on the multiple key pieces of first-entity information, wherein the graph includes multiple first-entity nodes and edges connecting the nodes, each node corresponding to a key piece of first-entity information, and the edges representing the relationships between these nodes; inputting the information relationship graph into a target graph neural network to obtain feature representations of the first-entity nodes; and performing enterprise data anomaly detection based on the feature representations of the first-entity nodes to obtain the data auditing results for the target enterprise. This application can accurately audit enterprise data without increasing the enterprise's manpower burden.

Description

Technical Field

[0001] This application relates to the field of data inspection, specifically to a method, apparatus, electronic device, and storage medium for enterprise data auditing. Background Technology

[0002] In modern enterprise management, the operating data reported by multiple business units within an enterprise to the corporate headquarters or management may be distorted or biased, making it difficult for the enterprise as a whole to make an accurate judgment on the true operating status of each business unit, thereby affecting the enterprise's strategic decision-making and resource allocation.

[0003] Currently, manual auditing or manually defined audit rules can be used to audit enterprise data to determine whether there is information distortion or bias. However, as enterprises grow in size and their organizational structures become more complex, enterprise data becomes increasingly complex, and the above-mentioned data auditing methods will lead to an increase in the enterprise's human resource burden.

[0004] Therefore, how to accurately audit corporate data without increasing the company's manpower burden has become an urgent problem to be solved. Summary of the Invention

[0005] In view of the above, the embodiments of this application provide a method for auditing enterprise data, which can accurately audit enterprise data without increasing the enterprise's manpower burden.

[0006] In a first aspect, embodiments of this application provide a method for auditing enterprise data, including: Obtain data on the target company that is to be audited; Information was extracted from the data of the companies to be investigated, resulting in multiple key pieces of information about the first company. A first enterprise information relationship graph is constructed based on the multiple first enterprise key information. The first enterprise information relationship graph includes multiple first nodes and edges connecting the multiple first nodes. Each first node corresponds to a first enterprise key information, and the edges are used to represent the association relationship between the first enterprise key information. The first enterprise information relationship graph is input into the target graph neural network to obtain the feature representations of the plurality of first nodes; Based on the feature representations of the multiple first nodes, enterprise data anomaly detection is performed to obtain the data audit results of the target enterprise.

[0007] This application embodiment extracts information from the data of the target enterprise to be audited, eliminating interfering information. Then, based on the extracted key information of multiple first enterprises, a first enterprise information relationship graph is constructed, thereby transforming the enterprise data into structured graphical data, clearly presenting the inherent relationship between the key information of each first enterprise. Next, the first enterprise information relationship graph is input into a target graph neural network to mine the self-characteristics and correlation patterns of the key information of the first enterprises, thereby obtaining the feature representation of multiple first nodes. Based on the feature representation of these first nodes, enterprise data anomaly detection is performed. By combining the correlation between the key information of the first enterprises, anomalies in the data to be audited can be automatically and accurately identified. Compared with traditional manual auditing or rule verification, this not only enables automated auditing of enterprise data, reducing the manpower burden caused by enterprise data auditing and improving auditing efficiency, but also avoids the problem of strong subjectivity caused by manual auditing, ensuring the accuracy of audit results.

[0008] In some embodiments, enterprise data anomaly detection is performed based on the feature representations of the plurality of first nodes to obtain the data audit results of the target enterprise, including: The feature representations of the multiple first nodes are subjected to feature aggregation processing to obtain the feature representation of the target enterprise; Obtain a normal enterprise data distribution model, wherein the normal enterprise data distribution model is used to characterize the data distribution of normal enterprise features; Based on the normal enterprise data distribution model and the target enterprise feature representation, enterprise data anomaly detection is performed to obtain the data audit results.

[0009] By adopting this technical solution, the normal enterprise data distribution model is used as a reference benchmark. Anomaly detection is performed based on the overall feature representation of the data of the enterprise to be audited (i.e., the feature representation of the target enterprise) and the normal enterprise data distribution model. This enables comprehensive, objective, and standardized data auditing of the data of the enterprise to be audited. In addition, the embodiments of this application are based on normal enterprise data modeling, which can get rid of the dependence on a large number of labeled abnormal datasets in abnormal data monitoring. It can identify abnormal data in a large amount of enterprise operating data and reduce the cost of enterprise data auditing.

[0010] In some embodiments, the normal enterprise data distribution model is a target hypersphere; obtaining the normal enterprise data distribution model includes: Multiple first-enterprise data samples were obtained, and the first-enterprise data samples were normal enterprise data. Based on the target graph neural network and the multiple first enterprise data samples, multiple normal enterprise feature representations are obtained; The center and radius of the target hypersphere are determined based on the multiple normal enterprise feature representations and the preset objective function. The objective function aims to minimize the sum of the first expression and the second expression; the first expression represents the radius of the target hypersphere. The second expression is used to represent the average degree of deviation of the plurality of normal enterprise characteristics from the target hypersphere.

[0011] In this embodiment, the target hypersphere is determined by an objective function. The objective function is optimized by minimizing the average deviation between the objective function and multiple first enterprise data and the target hypersphere. This not only allows for clustered fitting of normal enterprise data features, but also accommodates the slight noise in normal enterprise data, avoiding model overfitting, thereby improving the generalization ability of the normal enterprise data distribution model.

[0012] In some embodiments, enterprise data anomaly detection is performed based on the normal enterprise data distribution model and the target enterprise feature representation to obtain the data audit results, including: If the target enterprise's characteristic indicates that the distance from the center of the target hypersphere exceeds the radius, then the data of the enterprise to be investigated is determined to be abnormal data. If the distance between the target enterprise characteristic and the center of the target hypersphere does not exceed the radius, then the data of the enterprise to be investigated is determined to be normal data.

[0013] In some embodiments, feature aggregation processing is performed on the feature representations of the plurality of first nodes to obtain the feature representation of the target enterprise, including: Determine the weights of the plurality of first nodes; The target enterprise feature representation is obtained by weighted summation based on the weights and feature representations of the multiple first nodes.

[0014] The feature aggregation method in this application adopts a weighted summation method, which can assign different weights to different first nodes according to the enterprise data auditing needs, highlight the feature contribution of auditing information, and thus make the feature representation of the target enterprise fit the audit focus, which is conducive to improving the accuracy of subsequent enterprise data auditing.

[0015] In some embodiments, the training steps of the target graph neural network include: Obtain a second enterprise data sample, which includes multiple key pieces of information about the second enterprise; A second enterprise information relationship graph is determined based on the multiple second enterprise key information, wherein the second enterprise information relationship graph includes multiple second nodes, and each second node corresponds to an original feature vector of a second enterprise key information. Select at least one second node as a mask node in the second enterprise information relationship diagram; Based on the preset initial graph neural network and the unmasked nodes in the second enterprise information relationship graph, the feature vector of the masked node is predicted to obtain the predicted feature vector; The initial graph neural network is updated based on the original feature vector and the predicted feature vector to obtain the updated initial graph neural network. If the updated initial graph neural network is trained, the updated initial graph neural network will be used as the target graph neural network.

[0016] This application employs a graph self-prediction task to train a graph neural network, enabling the graph neural network to learn the reasoning ability between key enterprise information. This allows for better discovery of the key information behind the data of the enterprise to be audited and the interrelationships between them, thereby improving the accuracy of enterprise data auditing.

[0017] In some embodiments, obtaining the target company's data to be audited includes: In response to a request for enterprise data auditing, the system retrieves the target enterprise's data to be audited from the enterprise knowledge vector database. The process of extracting information from the data of the company to be investigated yields several key pieces of information about the first company, including: Information extraction prompts are determined based on a preset information extraction prompt template and the data of the enterprise to be investigated; The extracted prompts are input into a large language model to obtain the multiple key information of the first enterprise.

[0018] The embodiment of this application uses an enterprise knowledge vector database, which is beneficial for integrating multi-source heterogeneous data of enterprises. Then, using the data of the enterprise to be audited and the prompt word template is beneficial for guiding the large language model to perform accurate information extraction and improving the accuracy of information extraction.

[0019] Secondly, embodiments of this application also provide an enterprise data auditing device, comprising: The information extraction module is used to acquire data of the target company to be audited, and to extract information from the data of the target company to obtain multiple key information of the first company. A relationship graph generation module is used to construct a first enterprise information relationship graph based on the plurality of first enterprise key information. The first enterprise information relationship graph includes a plurality of first nodes and edges connecting the plurality of first nodes. Each first node corresponds to a first enterprise key information, and the edges are used to represent the association relationship between the first enterprise key information. The feature representation module is used to input the first enterprise information relationship graph into the target graph neural network to obtain the feature representations of the plurality of first nodes; The data audit module is used to detect enterprise data anomalies based on the feature representations of the multiple first nodes, and obtain the data audit results of the target enterprise.

[0020] Thirdly, embodiments of this application also provide an electronic device, the electronic device including a processor and a memory, the memory being used to store instructions, and the processor being used to call the instructions in the memory, causing the electronic device to execute the enterprise data auditing method described in the first aspect.

[0021] Fourthly, embodiments of this application also provide a computer-readable storage medium that stores computer instructions that, when executed on an electronic device, cause the electronic device to perform the enterprise data auditing method as described in the first aspect.

[0022] Understandably, the enterprise data auditing device, electronic equipment and storage medium provided above correspond to the enterprise data auditing method of the first aspect above. Therefore, the beneficial effects that can be achieved can be referred to the beneficial effects of the corresponding enterprise data auditing method provided above, and will not be repeated here. Attached Figure Description

[0023] Figure 1 This is a flowchart of the steps of an enterprise data auditing method provided according to an embodiment of this application.

[0024] Figure 2 This is an architecture diagram of an enterprise data auditing method provided according to an embodiment of this application.

[0025] Figure 3 This is a flowchart of the training steps for a target graph neural network according to an embodiment of this application.

[0026] Figure 4 This is a schematic diagram of the device structure of an enterprise data auditing device provided according to an embodiment of this application.

[0027] Figure 5 This is a schematic diagram of the structure of an electronic device provided according to an embodiment of this application. Detailed Implementation

[0028] To better understand the above-mentioned objectives, features, and advantages of this application, the application will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0029] The following description sets forth many specific details to provide a full understanding of this application. The described embodiments are only some, not all, of the embodiments of this application.

[0030] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein in the specification of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application.

[0031] It should be further noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0032] In this application, "at least one" means one or more, and "more than one" means two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone, where A and B can be singular or plural. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and drawings of this application are used to distinguish similar objects, not to describe a specific order or sequence.

[0033] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design that is described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design. Specifically, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0034] This application provides a method, apparatus, electronic device, and computer-readable storage medium for enterprise data auditing.

[0035] The enterprise data auditing method may include: acquiring enterprise data of the target enterprise to be audited; extracting information from the enterprise data to be audited to obtain multiple first enterprise key information; constructing a first enterprise information relationship graph based on the multiple first enterprise key information, wherein the first enterprise information relationship graph includes multiple first nodes and edges connecting the multiple first nodes, wherein each first node corresponds to one first enterprise key information, and the edges are used to represent the association relationship between the first enterprise key information; inputting the first enterprise information relationship graph into a target graph neural network to obtain feature representations of the multiple first nodes; and performing enterprise data anomaly detection based on the feature representations of the multiple first nodes to obtain the data auditing result of the target enterprise.

[0036] This application embodiment extracts information from the data of the target enterprise to be audited, which can remove interfering information. Then, based on the extracted key information of multiple first enterprises, a first enterprise information relationship graph is constructed, thereby transforming the enterprise data into structured graphical data, clearly presenting the inherent relationship between the key information of each first enterprise. Next, the first enterprise information relationship graph is input into the target graph neural network to mine the self-characteristics and correlation patterns of the key information of the first enterprise, thereby obtaining the feature representation of multiple first nodes. Based on the feature representation of these first nodes, enterprise data anomaly detection is performed, which can consider the correlation between the key information of the first enterprise, thereby automatically and accurately identifying the anomalies in the data to be audited.

[0037] Compared to traditional manual audits or rule verification, the embodiments of this application can not only automate the audit of enterprise data, reduce the manpower burden caused by enterprise data audits, and improve audit efficiency, but also avoid the problem of strong subjectivity caused by manual audits, automatically mine the self-characteristics and correlation patterns of the first enterprise's key information for data auditing, and ensure the accuracy of audit results.

[0038] The enterprise data auditing method of this application can be applied to one or more electronic devices. The electronic device is a device capable of automatically performing numerical calculations and / or information processing according to pre-set or stored instructions. Its hardware includes, but is not limited to, processors, microprogrammed control units (MCUs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), digital signal processors (DSPs), embedded devices, etc.

[0039] Figure 1This is a flowchart illustrating the steps of one embodiment of the enterprise data auditing method of this application. Depending on different needs, the order of the steps in the flowchart can be changed, and some steps can be omitted.

[0040] See Figure 1 As shown, enterprise data auditing methods may include the following steps.

[0041] Step 101: Obtain the data of the target company to be audited.

[0042] The target company refers to the company that needs to be audited.

[0043] The data of the company to be audited can be any kind of data related to the production and operation of the target company. For example, the data of the company to be audited can include company registration information, industry classification, company tax data, company production resource consumption data, etc., but is not limited to these.

[0044] In some embodiments, the way an electronic device obtains data of a company to be audited may include: in response to a request for auditing company data, searching for the data of the target company in a corporate knowledge vector database (also known as a corporate information database).

[0045] For example, electronic devices can respond to a company’s data audit request by performing multi-round, multi-granular semantic similarity retrieval from the company’s knowledge vector database to accurately locate scattered information fragments related to the company’s output value assessment, thereby obtaining the company’s data to be audited.

[0046] The enterprise knowledge vector database is a pre-built vector database.

[0047] In some embodiments, the construction of an enterprise knowledge vector database may include the following steps a1 to a3, which are described below.

[0048] Step a1: Collect enterprise data from multiple data sources to obtain multi-source heterogeneous enterprise data.

[0049] For example, embodiments of this application provide a data collection module and a data preprocessing module.

[0050] This data collection module can collect enterprise data from multiple data sources, resulting in multi-source heterogeneous enterprise data.

[0051] Multi-source heterogeneous enterprise data can include static data and dynamic data.

[0052] Static data refers to information that remains relatively stable over a long period of time, such as company registration information and industry classifications.

[0053] Dynamic data refers to production and operation-related data that changes frequently over time, exhibiting strong timeliness and variability. For example, dynamic data may include: Corporate tax data, including value-added tax, income tax and other related tax information, mainly comes from tax authorities; Data on enterprise production resource consumption, such as energy consumption of water, electricity, gas, and steam, are obtained from relevant departments such as the water bureau, power company, and gas company. Other cost data for enterprises, such as personnel composition and salary expenditure, can be extracted from materials such as annual reports and statistical reports submitted by enterprises to government management departments. Enterprise asset, investment and loan data, covering financial information such as fixed assets, foreign investment and bank loans, mainly comes from the banking system or financial data reported by enterprises; Enterprise intellectual property data, including information on patents, trademarks, and software copyrights, can be obtained from publicly available data on the official website of the State Intellectual Property Office. Other supplementary data for enterprises, such as unstructured information like media reports, social reputation, and public evaluations, can be collected from news websites, social media, and other channels through web crawling technology.

[0054] The multi-source heterogeneous enterprise data listed above are merely examples. In practical applications, they can be set according to requirements, and this application embodiment does not limit them.

[0055] After the data collection module collects data from multiple heterogeneous enterprises, the data preprocessing module can perform the following steps a2 to a3.

[0056] Step a2 involves preprocessing the multi-source heterogeneous enterprise data to obtain preprocessed enterprise data.

[0057] For example, the data preprocessing module can interface with the output of the data collection module, which is used to transform multi-source heterogeneous enterprise data with wide range of sources and different formats into a structured and unified text data stream that is easy to analyze in depth later.

[0058] The data preprocessing module can preprocess non-electronic or non-text format data, such as scanned documents, images, and PDF files, as well as heterogeneous enterprise data from multiple sources.

[0059] refer to Figure 2 As shown, the data preprocessing module can integrate and call optical character recognition (OCR) and intelligent document processing technologies to achieve high-precision text recognition and extraction, ensuring that the original information is transformed into machine-readable text without loss.

[0060] Secondly, the data preprocessing module can further call large language model search tools (such as DeepSeek) to conduct online searches and automatically generate multi-dimensional in-depth research reports on the target company to make up for the limitations of internal data and build a more comprehensive corporate profile. This report, as a key supplementary material, greatly enriches the analytical context.

[0061] Step a3: The preprocessed enterprise data is converted into a vector representation using a text embedding model, and the vector representation is stored in the enterprise knowledge vector database (i.e., the enterprise information database).

[0062] The data preprocessing module can use a text embedding model to convert all the processed text data into high-dimensional vector representations and persist them to a dedicated vector database, thereby enabling subsequent semantic-based fast retrieval and vectorized computation, and providing efficient and unified data services for upstream reasoning and analysis modules.

[0063] The embodiments of this application can process heterogeneous enterprise data from multiple sources such as taxation, resource consumption, and cost, and unify their data formats, which can significantly improve the robustness and efficiency of data auditing.

[0064] Once the data of the company to be audited is retrieved from the enterprise knowledge vector database, step 102 can be executed.

[0065] Step 102: Extract information from the data of the companies to be audited to obtain key information about the leading companies.

[0066] The first key enterprise information refers to structured information that plays a crucial role in detecting anomalies in enterprise data. For example, the first key enterprise information may include, but is not limited to, value-added tax payment, monthly industrial electricity consumption, total salaries of core technical personnel, number of newly added patents, and operating profit.

[0067] Electronic devices can input the data of the company to be audited into the company information extraction model to obtain multiple key information about the company.

[0068] This enterprise information extraction model can be a large language model.

[0069] For example, step 102 may include: determining information extraction prompts based on a preset information extraction prompt template and the data of the enterprise to be investigated; and inputting the information extraction prompts into a large language model to obtain the key information of the multiple first enterprises.

[0070] The preset information extraction prompt template can be a standardized template designed according to the business needs of enterprise data auditing. The type of key enterprise information can be specified in the information extraction prompt template.

[0071] For example, information extraction prompt templates may include the following: "Please extract key information related to the enterprise's output value assessment from the following data of the enterprises to be audited, including but not limited to: value-added tax payment, monthly industrial electricity consumption, total salary of core technical personnel, number of invention patents, operating profit, R&D expense deduction, fixed asset depreciation, and bank loan interest."

[0072] The output format is [Enterprise Key Information Name: Enterprise Key Information Value]: {Data of the Enterprise to be Audited}.

[0073] Then, the data of the companies to be audited is filled into the template to generate information extraction prompts. The generated information extraction prompts are then input into the large language model to obtain multiple key information about the first company.

[0074] Large language models possess deep semantic understanding, context learning, and complex reasoning capabilities. Large language models can be GPT series models, DeepSeek models, etc., but are not limited to these.

[0075] After receiving information extraction prompts, the large language model parses the data of the companies to be audited, extracts key information according to the prompts, and outputs multiple structured key information for the first company. For example, if the data of the companies to be audited includes "the number of invention patents is 10", the large language model will extract the key information for the first company: [number of invention patents: 10].

[0076] This application embodiment utilizes the deep semantic understanding, context learning, and complex reasoning capabilities of a large-scale language model to deeply fuse and analyze the retrieved information, guiding the large-scale language model to perform advanced tasks such as entity recognition and relation extraction, and dynamically extracting key structured elements that constitute the enterprise operation profile from unstructured text.

[0077] This embodiment effectively combines the information assurance capabilities of retrieval technology with the generalization reasoning capabilities of large-scale language models, ensuring the comprehensiveness, accuracy, and interpretability of element extraction, and providing high-quality, structured input for subsequent graph neural network modeling and anomaly detection.

[0078] Step 103: Construct a first enterprise information relationship diagram based on multiple first enterprise key information.

[0079] The First Enterprise Information Relationship Diagram is a graph structure data used to represent the relationships between key information of the First Enterprise.

[0080] The first enterprise information relationship graph includes multiple first nodes and edges connecting the multiple first nodes. Each first node corresponds to a key piece of information about the first enterprise. The edges are used to represent the relationships between the key pieces of information about the first enterprise.

[0081] In some embodiments, step 103 may include mapping each key information of the first enterprise to a first node, for example, value-added tax payment corresponds to first node 1, monthly industrial electricity consumption corresponds to first node 2, operating profit corresponds to first node 3, etc.

[0082] Then, based on the business logic or data-driven relationships between the key information of the first enterprise, edges are constructed between nodes.

[0083] For example, monthly industrial electricity consumption and value-added tax payment are related at the production and operation level. Therefore, an edge is established between the first node 1 and the first node 2. The weight of the edge can be determined by the correlation calculation between the key information of the first enterprise (such as cosine similarity).

[0084] By constructing the first enterprise information relationship graph, scattered key information can be transformed into structured graph data, providing a foundation for subsequent graph neural network mining of deep information relationships.

[0085] Step 104: Input the first enterprise information relationship graph into the target graph neural network to obtain the feature representations of multiple first nodes.

[0086] Target graph neural networks are pre-trained graph neural network models with the ability to mine information associations. For example, target graph neural networks can be models built based on graph convolutional networks, graph attention networks, or graph sampling aggregation networks.

[0087] The target graph neural network has a spatiotemporally consistent network structure. In the target graph neural network, spatiotemporal consistency is reflected in the connectivity and dependency between nodes. The target graph neural network can capture the temporal dependency and interaction of multiple first nodes in the first enterprise information relationship graph, thereby obtaining the feature representation of multiple first nodes.

[0088] After receiving the first enterprise information relationship graph, the target graph neural network encodes the features of each first node through multi-layer information propagation and aggregation operations.

[0089] When generating the feature representation of the first node, the target graph neural network can combine the feature information of the current neighboring node and its own original features to generate a low-dimensional dense feature representation (i.e., the feature representation of the first node) that reflects the attributes of the current first node and its relationship with other first nodes.

[0090] For example, the feature representation of the value-added tax payment amount node not only includes the numerical features of its payment amount, but also integrates the feature information of the first node, such as the monthly industrial electricity consumption and operating profit associated with it, thereby realizing the association and fusion of key information.

[0091] The embodiments of this application employ target graph neural networks to uncover key information behind enterprise data and the interrelationships between them, thereby improving the accuracy of enterprise data auditing.

[0092] Step 105: Perform enterprise data anomaly detection based on the feature representations of multiple first nodes to obtain the data audit results of the target enterprise.

[0093] In some embodiments, step 105 can be implemented by the following steps b1 to b3: Step b1: Perform feature aggregation processing on the feature representations of the multiple first nodes to obtain the feature representation of the target enterprise.

[0094] The purpose of feature aggregation processing is to merge the feature representations of multiple nodes into a feature vector that can characterize the overall operational status of the target enterprise.

[0095] In some embodiments, the electronic device may determine the weights of the plurality of first nodes; The target enterprise feature representation is obtained by weighted summation based on the weights of the plurality of first nodes and the feature representations of the plurality of first nodes.

[0096] For example, an electronic device can perform a weighted summation of the feature representations of the multiple first nodes using a preset feature aggregation formula to obtain the feature representation of the target enterprise; The feature aggregation formula is shown below: ; Among them, the The target enterprise features are represented by the following: The number of the plurality of first nodes; For the first The feature representation of the first node; the For the first The weight of the first node.

[0097] It can be preset or automatically optimized during the target graph neural network training process. This can reflect the importance of the first node in the enterprise's anomaly detection.

[0098] The above feature aggregation formula can comprehensively consider the characteristics and importance of all key information nodes to generate a comprehensive and accurate overall enterprise feature representation.

[0099] Step b2: Obtain the normal enterprise data distribution model, which is used to characterize the data distribution of normal enterprise features.

[0100] Normal enterprise characteristics are represented as the characteristics of normal enterprise data.

[0101] The normal enterprise data distribution model is constructed based on the characteristic representation of normal enterprises and is used to define the characteristic distribution range of normal enterprise data.

[0102] In some embodiments, a single-class anomaly detection method (Deep Support Vector Data Description, DeepSVDD) can be used to model normal enterprise data, thereby obtaining a normal enterprise data distribution model.

[0103] For example, a normal enterprise data distribution model can be a target hypersphere, and the construction of the target hypersphere can include the following steps b2.1 to b2.3: Step b2.1: Obtain the first enterprise data sample, which is normal enterprise data.

[0104] The first enterprise data sample can come from the target enterprise's historical enterprise data or from the historical enterprise data of enterprises similar to the target enterprise. This makes the normal enterprise data distribution model conform to the distribution characteristics of normal data of this type of enterprise, thereby improving the accuracy of auditing enterprise data for this type of enterprise.

[0105] The first enterprise data sample can also come from historical enterprise data of other types of enterprises, thereby covering enterprises of different industries and sizes to ensure the generalization ability of the model.

[0106] The data type of the first enterprise data sample can be consistent with the data of the enterprise to be audited, and it can include various related data such as static data and dynamic data.

[0107] Step b2.2: Based on the target graph neural network and the multiple first enterprise data samples, multiple normal enterprise feature representations are obtained.

[0108] In some embodiments, for each first enterprise data sample, the electronic device can extract information from the first enterprise data sample to obtain multiple first enterprise key information samples; construct a first enterprise information relationship graph sample based on the multiple first enterprise key information samples, the first enterprise information relationship graph including multiple first node samples and edges connecting the multiple first node samples, wherein each first node sample corresponds to one first enterprise key information sample, and the edges are used to represent the association relationship between the first enterprise key information samples; input the first enterprise information relationship graph sample into a target graph neural network to obtain the feature representation of the multiple first node samples; perform feature aggregation processing on the feature representation of the multiple first node samples to obtain a normal enterprise feature representation.

[0109] The implementation details of the above steps are roughly the same as those of steps 102 and 104, and will not be repeated here.

[0110] After obtaining the normal enterprise characteristic representation of each first enterprise data sample, step b2.3 can be executed.

[0111] Step b2.3: ​​Determine the center and radius of the target hypersphere based on the multiple normal enterprise feature representations and the preset objective function.

[0112] In this context, the unknown to be solved in the objective function is the center of the target hypersphere. and radius The electronic device substitutes multiple first normal enterprise feature representations into the objective function and solves for them to obtain the central... and radius .

[0113] The objective function can be defined as minimizing the sum of the first and second expressions.

[0114] The first expression is used to represent the radius of the target hypersphere. For example, the first expression can be the square of the radius of the target hypersphere (i.e., (but not limited to this).

[0115] The second expression is used to represent the average degree of deviation of the multiple normal enterprise characteristics from the target hypersphere.

[0116] In some embodiments, the electronic device may determine the degree of deviation of each normal enterprise feature from the target hypersphere to obtain multiple degrees of deviation; then, the average expression of the multiple degrees of deviation is used as a second expression.

[0117] Among them, when the normal enterprise feature indicates that it is located inside or on the boundary of the target hypersphere, the degree of deviation of the normal enterprise feature from the target hypersphere is no deviation, that is, the degree of deviation is 0.

[0118] When a normal enterprise feature represents a feature outside the target hypersphere, that is, when a normal enterprise feature represents a feature outside the target hypersphere and not located at the boundary of the target hypersphere, the degree of deviation can be determined based on the distance of the normal enterprise feature representation from the center of the target hypersphere and the radius of the target hypersphere.

[0119] For example, representing the characteristics of a normal enterprise. It can calculate the normal enterprise characteristic representation to the center of the target hypersphere. square Euclidean distance Then subtract the square of the target hypersphere's radius from the square of the Euclidean distance. (Right now, Thus, the degree of deviation of the normal enterprise characteristics from the target hypersphere is obtained.

[0120] For example, the objective function is as follows: ; Among them, the The center of the target hypersphere; The radius of the target hypersphere; The quantity represented by the plurality of normal enterprise characteristics; For the first First Enterprise Data Sample This represents the normal characteristics of a business; This is the first expression; This is the second expression.

[0121] By solving the above objective function using optimization algorithms such as gradient descent, a target hypersphere with a minimized volume can be obtained. This target hypersphere can contain all normal enterprise feature representations to the greatest extent, thereby accurately representing the feature distribution of normal enterprises.

[0122] Step b3: Based on the normal enterprise data distribution model and the target enterprise feature representation, perform enterprise data anomaly detection to obtain the data audit results.

[0123] The results of data audits are used to indicate whether there are any anomalies in the data of the company being audited (i.e., whether there is any information distortion or deviation).

[0124] In some embodiments, the electronic device can calculate a target enterprise feature representation. The distance dist (e.g., Euclidean distance) from the center c of the target hypersphere; If the distance between the target enterprise feature representation and the center of the target hypersphere does not exceed the radius of the target hypersphere (i.e., dist≤R), it indicates that the target enterprise feature representation falls within the feature distribution range of normal enterprise data, and the data of the enterprise to be investigated is determined to be normal data.

[0125] If the distance between the target enterprise feature representation and the center of the target hypersphere exceeds the radius of the target hypersphere (i.e., dist>R), it indicates that the target enterprise feature representation deviates from the normal feature distribution range of enterprise data, and the data of the enterprise to be investigated is determined to be abnormal data, that is, the data of the enterprise to be investigated may have information distortion or deviation.

[0126] After detecting abnormal data, electronic devices can issue risk warnings, informing users that the company's data is abnormal.

[0127] This application's embodiments treat enterprise data auditing as an anomaly detection task. By processing heterogeneous enterprise data from multiple sources such as taxation, resource consumption, cost, assets, and intellectual property, it utilizes optical character recognition (OCR) technology to unify the data format and uses a large language model to extract key information of the first enterprise. It can simultaneously process heterogeneous enterprise data from multiple sources such as taxation, resource consumption, and cost, and unify their data formats, significantly improving the robustness and efficiency of data auditing.

[0128] Secondly, compared to manually defined rules or traditional machine learning's neglect of the implicit correlation features behind a large amount of enterprise data, the embodiments of this application construct a target graph neural network to process the information relationship graph of the first enterprise, which can capture the implicit correlations between the key information of the first enterprise, thereby promoting the understanding of the data of the enterprise to be audited. That is, the embodiments of this application can mine the key information behind the enterprise data and the interrelationships between them, thereby improving the accuracy of enterprise data audit.

[0129] Compared to traditional supervised learning methods, this application's embodiments use normal enterprise data samples to construct a normal enterprise data distribution model for anomaly detection. Data that deviates from this distribution is identified as anomalies. This can overcome the dependence on a large number of labeled anomaly datasets and solve the problem of insufficient abnormal enterprise data samples in enterprise data samples. Abnormal data can be identified in a large amount of enterprise operating data through normal enterprise data samples, which can effectively reduce the cost of enterprise protective gear inspection.

[0130] In other words, by transforming the enterprise data auditing task into an anomaly detection task and integrating the semantic understanding capability of large-scale language models with the relational modeling capability of graph neural networks, this application embodiment can deeply mine information from multi-source heterogeneous enterprise data. This application embodiment constructs a self-supervised detection framework that does not require a large amount of anomaly labeled data, which can get rid of excessive dependence on rules and labels in the process of enterprise data auditing.

[0131] The target graph neural network used in this application embodiment can be trained using a graph self-prediction task to improve the model's ability to understand data features. The training steps of the target graph neural network are described below.

[0132] Combination Figure 3 As shown, the training steps for the target graph neural network include: Step 301: Obtain a second enterprise data sample, which includes multiple key pieces of information about the second enterprise.

[0133] The second enterprise data sample is enterprise data used to train the graph neural network.

[0134] The second enterprise data sample can be derived from historical enterprise data of various types of enterprises, and the data type can be consistent with the data of the enterprise to be audited.

[0135] The key information of the second enterprise is structured information extracted from the data sample of the second enterprise.

[0136] The method of extracting key information of the second enterprise from the data sample of the second enterprise is roughly the same as the method of extracting key information of the first enterprise from the data of the enterprise to be investigated, and will not be described again in the embodiments of this application.

[0137] Step 302: Determine the information relationship diagram of the second enterprise based on multiple key information of the second enterprise.

[0138] The second enterprise information relationship graph includes multiple second nodes and edges connecting the multiple second nodes.

[0139] Each key piece of information about a second enterprise corresponds to a second node, and the edges between nodes represent the relationships between the key pieces of information about second enterprises.

[0140] Each second node corresponds to an original feature vector of key information of the second enterprise, which is obtained by vectorizing the key information of the second enterprise.

[0141] In some embodiments, step 302 may include: vectorizing each second enterprise key information and mapping it to a second node; constructing edges between second nodes based on business logic associations or data-driven associations between the second enterprise key information, wherein the weight of the edges can be determined by calculating the correlation between the second enterprise key information (such as cosine similarity).

[0142] For example, the first The second key enterprise information (corresponding to the first) The original feature vector of the second node) It can be as follows: ; in, Represents the original feature vector. for A 1-dimensional real vector This refers to the quantity of key information for the second enterprise.

[0143] As the initial graph neural network is trained and iterated, the initial graph neural network... It will be continuously adjusted and optimized until... It can accurately reflect the first The association between key information of a second enterprise and key information of other second enterprises.

[0144] Step 303: Select at least one second node as a mask node in the second enterprise information relationship graph.

[0145] For example, an electronic device can select at least one second node as a mask node in the second enterprise information relationship graph according to a preset masking strategy. For example, some second nodes can be randomly selected as mask nodes with a probability of 15%.

[0146] In this process, the original feature vector of the mask node is replaced with a special, learnable [mask] embedding vector or zero vector to simulate a scenario where key enterprise information is missing.

[0147] Step 304: Based on the preset initial graph neural network and the unmasked nodes in the second enterprise information relationship graph, predict the feature vector of the masked node to obtain the predicted feature vector.

[0148] The initial graph neural network is an untrained graph neural network, and its network structure is consistent with the target graph neural network.

[0149] Electronic devices can input a second enterprise information relationship graph with masked nodes into an initial graph neural network. The initial graph neural network then propagates and aggregates information through the feature information of unmasked nodes and the relationships between nodes, predicting the original feature vector of the masked node and obtaining the predicted feature vector. For example, the initial graph neural network can determine the neighboring nodes of a masked node, propagate and aggregate information based on the feature vectors of the unmasked nodes in those neighboring nodes and the relationships between nodes, predicting the original feature vector of the masked node and obtaining the predicted feature vector.

[0150] Step 305: Update the initial graph neural network based on the original feature vector and the predicted feature vector to obtain the updated initial graph neural network.

[0151] For example, an electronic device can minimize the error between the original feature vector and the predicted feature vector as the training objective, and use the gradient descent algorithm to backpropagate the error to update the network parameters of the initial graph neural network (such as node embedding vectors, network layer weights, etc.).

[0152] If the updated initial graph neural network training is not completed, at least one second node can be selected again as a mask node in the second enterprise information relationship graph, and steps 304 to 305 can continue to be executed.

[0153] Alternatively, if the updated initial graph neural network training is not completed, the electronic device can acquire new second enterprise sample data and continue to execute steps 302 to 305 until the updated initial graph neural network training is completed, and then execute step 306.

[0154] Step 306: If the updated initial graph neural network has been trained, then the updated initial graph neural network is used as the target graph neural network.

[0155] The criteria for determining whether model training is complete may include: the number of training iterations reaching the preset maximum number of iterations, or the loss function value converging to a preset threshold, etc. This application does not limit these criteria.

[0156] After training, the updated initial graph neural network has powerful information association mining and feature prediction capabilities, and can accurately capture the business logic and implicit relationships between key enterprise information. It can be used as a target graph neural network for subsequent enterprise data auditing.

[0157] The embodiments of this application enable graph neural networks to not only focus on isolated numerical values ​​of key information of a single enterprise, but also to deeply understand and learn the complex collaborative relationships and logical consistency between key information of an enterprise.

[0158] For example, in order to successfully predict the obscured "corporate income tax" node, the graph neural network needs to comprehensively analyze information from its neighboring nodes such as "operating profit", "R&D expense deduction", "fixed asset depreciation" and "bank loan interest". In essence, it is simulating the logical reasoning process of a senior auditor when reviewing financial statements.

[0159] After multiple rounds of large-scale self-supervised pre-training on massive amounts of historical enterprise data covering numerous companies and spanning multiple fiscal years, the parameters of the graph neural network are fully optimized, enabling it to encode each key element into a low-dimensional, dense hidden state vector that is rich in context and deeply integrates the overall operational logic of the enterprise.

[0160] The self-supervised training paradigm described above can greatly reduce the reliance on massive amounts of manually labeled data. It can autonomously extract universally applicable and in-depth business rules and patterns from unlabeled enterprise data for use in judging enterprise data audit tasks.

[0161] In other words, through the above training process, the target graph neural network can learn the inherent laws of enterprise operation without relying on labeled abnormal samples, and can learn them through self-supervised training alone. This greatly reduces the dependence on labeled data and improves the model's generalization ability and business adaptability.

[0162] Based on the same concept as the enterprise data auditing method in the above embodiments, this application also provides an enterprise data auditing device, which can be used to execute the above enterprise data auditing method. For ease of explanation, the structural schematic diagram of the enterprise data auditing device embodiment only shows the parts related to the embodiments of this application. Those skilled in the art will understand that the illustrated structure does not constitute a limitation on the device, and may include more or fewer components than illustrated, or combine certain components, or have different component arrangements.

[0163] like Figure 4As shown, the enterprise data auditing device includes an information extraction module 401, a relationship diagram generation module 402, a feature representation module 403, and a data auditing module 404. In some embodiments, the above modules can be programmable software instructions stored in memory and executable by a processor. It is understood that in other embodiments, the above modules can also be program instructions or firmware embedded in the processor.

[0164] Among them, the information extraction module 401 is used to obtain the data of the target enterprise to be audited, and to extract information from the data of the target enterprise to obtain multiple key information of the first enterprise. The relationship graph generation module 402 is used to construct a first enterprise information relationship graph based on the plurality of first enterprise key information. The first enterprise information relationship graph includes a plurality of first nodes and edges connecting the plurality of first nodes. Each first node corresponds to a first enterprise key information, and the edges are used to represent the association relationship between the first enterprise key information. Feature representation module 403 is used to input the first enterprise information relationship graph into the target graph neural network to obtain the feature representations of the plurality of first nodes; The data audit module 404 is used to perform enterprise data anomaly detection based on the feature representation of the multiple first nodes, and obtain the data audit results of the target enterprise.

[0165] In some embodiments, enterprise data anomaly detection is performed based on the feature representations of the plurality of first nodes to obtain the data audit results of the target enterprise, including: The feature representations of the multiple first nodes are subjected to feature aggregation processing to obtain the feature representation of the target enterprise; Obtain a normal enterprise data distribution model, wherein the normal enterprise data distribution model is used to characterize the data distribution of normal enterprise features; Based on the normal enterprise data distribution model and the target enterprise feature representation, enterprise data anomaly detection is performed to obtain the data audit results.

[0166] By adopting this technical solution, the normal enterprise data distribution model is used as a reference benchmark. Anomaly detection is performed based on the overall feature representation of the data of the enterprise to be audited (i.e., the feature representation of the target enterprise) and the normal enterprise data distribution model. This enables comprehensive, objective, and standardized data auditing of the data of the enterprise to be audited. In addition, the embodiments of this application are based on normal enterprise data modeling, which can get rid of the dependence on a large number of labeled abnormal datasets in abnormal data monitoring. It can identify abnormal data in a large amount of enterprise operating data and reduce the cost of enterprise data auditing.

[0167] In some embodiments, the normal enterprise data distribution model is a target hypersphere; obtaining the normal enterprise data distribution model includes: Multiple first-enterprise data samples were obtained, and the first-enterprise data samples were normal enterprise data. Based on the target graph neural network and the multiple first enterprise data samples, multiple normal enterprise feature representations are obtained; The center and radius of the target hypersphere are determined based on the multiple normal enterprise feature representations and the preset objective function. The objective function aims to minimize the sum of the first expression and the second expression; the first expression represents the radius of the target hypersphere. The second expression is used to represent the average degree of deviation of the plurality of normal enterprise characteristics from the target hypersphere.

[0168] In this embodiment, the target hypersphere is determined by an objective function. The objective function is optimized by minimizing the average deviation between the objective function and multiple first enterprise data and the target hypersphere. This not only allows for clustered fitting of normal enterprise data features, but also accommodates the slight noise in normal enterprise data, avoiding model overfitting, thereby improving the generalization ability of the normal enterprise data distribution model.

[0169] In some embodiments, enterprise data anomaly detection is performed based on the normal enterprise data distribution model and the target enterprise feature representation to obtain the data audit results, including: If the target enterprise's characteristic indicates that the distance from the center of the target hypersphere exceeds the radius, then the data of the enterprise to be investigated is determined to be abnormal data. If the distance between the target enterprise characteristic and the center of the target hypersphere does not exceed the radius, then the data of the enterprise to be investigated is determined to be normal data.

[0170] In some embodiments, feature aggregation processing is performed on the feature representations of the plurality of first nodes to obtain the feature representation of the target enterprise, including: Determine the weights of the plurality of first nodes; The target enterprise feature representation is obtained by weighted summation based on the weights and feature representations of the multiple first nodes.

[0171] The feature aggregation method in this application adopts a weighted summation method, which can assign different weights to different first nodes according to the enterprise data auditing needs, highlight the feature contribution of auditing information, and thus make the feature representation of the target enterprise fit the audit focus, which is conducive to improving the accuracy of subsequent enterprise data auditing.

[0172] In some embodiments, the training steps of the target graph neural network include: Obtain a second enterprise data sample, which includes multiple key pieces of information about the second enterprise; A second enterprise information relationship graph is determined based on the multiple second enterprise key information, wherein the second enterprise information relationship graph includes multiple second nodes, and each second node corresponds to an original feature vector of a second enterprise key information. Select at least one node as a mask node in the second enterprise information relationship diagram; Based on the preset initial graph neural network and the unmasked nodes in the second enterprise information relationship graph, the feature vector of the masked node is predicted to obtain the predicted feature vector; The initial graph neural network is updated based on the original feature vector and the predicted feature vector to obtain the updated initial graph neural network. If the updated initial graph neural network is trained, the updated initial graph neural network will be used as the target graph neural network.

[0173] This application employs a graph self-prediction task to train a graph neural network, enabling the graph neural network to learn the reasoning ability between key enterprise information. This allows for better discovery of the key information behind the data of the enterprise to be audited and the interrelationships between them, thereby improving the accuracy of enterprise data auditing.

[0174] In some embodiments, obtaining the target company's data to be audited includes: In response to a request for enterprise data auditing, the system retrieves the target enterprise's data to be audited from the enterprise knowledge vector database. The process of extracting information from the data of the company to be investigated yields several key pieces of information about the first company, including: Information extraction prompts are determined based on a preset information extraction prompt template and the data of the enterprise to be investigated; The extracted prompts are input into a large language model to obtain the multiple key information of the first enterprise.

[0175] The embodiment of this application uses an enterprise knowledge vector database, which is beneficial for integrating multi-source heterogeneous data of enterprises. Then, using the data of the enterprise to be audited and the prompt word template is beneficial for guiding the large language model to perform accurate information extraction and improving the accuracy of information extraction.

[0176] Figure 5 This is a schematic diagram of an embodiment of the electronic device of this application.

[0177] Electronic device 100 includes memory 20, processor 30, and computer program 40 stored in memory 20 and executable on processor 30. When processor 30 executes computer program 40, it implements the steps described in the above-described enterprise data auditing method embodiment, for example... Figure 1 Steps 101 to 105 are shown.

[0178] For example, computer program 40 can also be divided into one or more modules / units, which are stored in memory 20 and executed by processor 30. The one or more modules / units can be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of computer program 40 in electronic device 100.

[0179] Those skilled in the art will understand that the schematic diagram is merely an example of the electronic device 100 and does not constitute a limitation on the electronic device 100. It may include more or fewer components than shown in the diagram, or combine certain components, or different components. For example, the electronic device 100 may also include input / output devices, network access devices, buses, etc.

[0180] Processor 30 can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. General-purpose processors can be microprocessors, single-chip microcomputers, or any conventional processor.

[0181] The memory 20 can be used to store computer programs 40 and / or modules / units. The processor 30 implements various functions of the electronic device 100 by running or executing the computer programs and / or modules / units stored in the memory 20 and by calling data stored in the memory 20. The memory 20 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, application programs required for at least one function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the electronic device 100 (such as audio data), etc. In addition, the memory 20 may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, RAM, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other non-volatile solid-state storage device.

[0182] If the modules / units integrated in the electronic device 100 are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0183] In the several embodiments provided in this application, it should be understood that the disclosed electronic devices and methods can be implemented in other ways. For example, the electronic device embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and other division methods may be used in actual implementation.

[0184] Furthermore, the functional units in the various embodiments of this application can be integrated into the same processing unit, or each unit can exist physically separately, or two or more units can be integrated into the same unit. The integrated units described above can be implemented in hardware or in the form of hardware plus software functional modules.

[0185] It will be apparent to those skilled in the art that this application is not limited to the details of the exemplary embodiments described above, and that this application can be implemented in other specific forms without departing from the spirit or essential characteristics of this application. Therefore, the embodiments should be considered exemplary and not restrictive in all respects. Furthermore, it is clear that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units or electronic devices recited in the electronic device claims may also be implemented by the same unit or electronic device through software or hardware. The terms "first," "second," etc., are used to indicate names and do not indicate any particular order.

[0186] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application and are not intended to limit it. Although this application has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of this application without departing from the spirit and scope of the technical solutions of this application.

Claims

1. A method for auditing enterprise data, characterized in that, include: Obtain data on the target company that is to be audited; Information was extracted from the data of the companies to be investigated, resulting in multiple key pieces of information about the first company. A first enterprise information relationship graph is constructed based on the multiple first enterprise key information. The first enterprise information relationship graph includes multiple first nodes and edges connecting the multiple first nodes. Each first node corresponds to a first enterprise key information, and the edges are used to represent the association relationship between the first enterprise key information. The first enterprise information relationship graph is input into the target graph neural network to obtain the feature representations of the plurality of first nodes; Based on the feature representations of the multiple first nodes, enterprise data anomaly detection is performed to obtain the data audit results of the target enterprise.

2. The enterprise data auditing method as described in claim 1, characterized in that, The process of detecting enterprise data anomalies based on the feature representations of the multiple first nodes to obtain the data audit results of the target enterprise includes: The feature representations of the multiple first nodes are subjected to feature aggregation processing to obtain the feature representation of the target enterprise; Obtain a normal enterprise data distribution model, wherein the normal enterprise data distribution model is used to characterize the data distribution of normal enterprise features; Based on the normal enterprise data distribution model and the target enterprise feature representation, enterprise data anomaly detection is performed to obtain the data audit results.

3. The enterprise data auditing method as described in claim 2, characterized in that, The normal enterprise data distribution model is a target hypersphere; obtaining the normal enterprise data distribution model includes: Multiple first-enterprise data samples were obtained, and the first-enterprise data samples were normal enterprise data. Based on the target graph neural network and the multiple first enterprise data samples, multiple normal enterprise feature representations are obtained; The center and radius of the target hypersphere are determined based on the multiple normal enterprise feature representations and the preset objective function. The objective function aims to minimize the sum of the first expression and the second expression; the first expression represents the radius of the target hypersphere. The second expression is used to represent the average degree of deviation of the plurality of normal enterprise characteristics from the target hypersphere.

4. The enterprise data auditing method as described in claim 3, characterized in that, The process of detecting enterprise data anomalies based on the normal enterprise data distribution model and the target enterprise feature representation, to obtain the data audit results, includes: If the target enterprise's characteristic indicates that the distance from the center of the target hypersphere exceeds the radius, then the data of the enterprise to be investigated is determined to be abnormal data. If the distance between the target enterprise characteristic and the center of the target hypersphere does not exceed the radius, then the data of the enterprise to be investigated is determined to be normal data.

5. The enterprise data auditing method as described in claim 2, characterized in that, The feature aggregation process performed on the feature representations of the plurality of first nodes to obtain the feature representation of the target enterprise includes: Determine the weights of the plurality of first nodes; The target enterprise feature representation is obtained by weighted summation based on the weights and feature representations of the multiple first nodes.

6. The enterprise data auditing method as described in any one of claims 1 to 5, characterized in that, The training steps of the target graph neural network include: Obtain a second enterprise data sample, which includes multiple key pieces of information about the second enterprise; A second enterprise information relationship graph is determined based on the multiple second enterprise key information, wherein the second enterprise information relationship graph includes multiple second nodes, and each second node corresponds to an original feature vector of a second enterprise key information. Select at least one second node as a mask node in the second enterprise information relationship diagram; Based on the preset initial graph neural network and the unmasked nodes in the second enterprise information relationship graph, the feature vector of the masked node is predicted to obtain the predicted feature vector; The initial graph neural network is updated based on the original feature vector and the predicted feature vector to obtain the updated initial graph neural network. If the updated initial graph neural network is trained, the updated initial graph neural network will be used as the target graph neural network.

7. The enterprise data auditing method as described in claim 1, characterized in that, The acquisition of the target company's data for the company to be audited includes: In response to a request for enterprise data auditing, the system retrieves the target enterprise's data to be audited from the enterprise knowledge vector database. The process of extracting information from the data of the company to be investigated yields several key pieces of information about the first company, including: Information extraction prompts are determined based on a preset information extraction prompt template and the data of the enterprise to be investigated; The extracted prompts are input into a large language model to obtain the multiple key information of the first enterprise.

8. A corporate data auditing device, characterized in that, include: The information extraction module is used to acquire data of the target company to be audited, and to extract information from the data of the target company to obtain multiple key information of the first company. A relationship graph generation module is used to construct a first enterprise information relationship graph based on the plurality of first enterprise key information. The first enterprise information relationship graph includes a plurality of first nodes and edges connecting the plurality of first nodes. Each first node corresponds to a first enterprise key information, and the edges are used to represent the association relationship between the first enterprise key information. The feature representation module is used to input the first enterprise information relationship graph into the target graph neural network to obtain the feature representations of the plurality of first nodes; The data audit module is used to detect enterprise data anomalies based on the feature representations of the multiple first nodes, and obtain the data audit results of the target enterprise.

9. An electronic device, the electronic device comprising a processor and a memory, characterized in that, The memory is used to store instructions, and the processor is used to invoke the instructions in the memory to cause the electronic device to execute the enterprise data auditing method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that, when executed on an electronic device, cause the electronic device to perform the enterprise data auditing method as described in any one of claims 1 to 7.

Images