A fraud detection system for supply chain finance

Abstract

The application provides a fraud detection system for supply chain finance, taking the state of the collateral involved in supply chain finance as the core clue, taking enterprise credit as the auxiliary, effectively identifying fraud based on the collateral and enterprise knowledge graph, making up for the lack of fraud identification model in heterogeneous and heterogeneous information, and perfecting the relationship of collateral, enterprise and its associated data. The system includes four modules of data acquisition, graph construction, fraud identification and fraud interpretation. Different data sources can be used to obtain various data objects, including enterprises and collaterals, by different acquisition methods through the data acquisition module. The graph construction module generates the corresponding knowledge graph based on ontology and relationship modeling from the data. The fraud identification module uses fraud detection and identification methods based on the knowledge graph to obtain fraud judgments for different objects. Through the fraud interpretation module, the characteristics and graph relationship of the fraud object are explained and visualized, providing assistance for subsequent decision-making by risk control personnel.

Description

Technical Field

[0001] This invention relates to a fraud detection system for supply chain finance. Background Technology

[0002] Current fraud detection in supply chain finance primarily targets enterprises, utilizing information such as their industry, fundamental strengths, management practices, financial indicators, asset size, sales revenue, and external support. However, enterprises seeking supply chain finance financing are often small and medium-sized enterprises (SMEs) within the industrial chain, whose fundamentals are relatively weak. If only enterprise-level ratings are applied, the results may be low, leading to difficulties and high costs in obtaining financing.

[0003] Currently, fraud detection methods are widely used in e-commerce, credit cards, credit loans, insurance, and other fields. These belong to the same or similar technical fields as this invention, sharing similar basic principles and frameworks. Both consist of two main parts: hardware and software for collecting data on the target object, and hardware and software for data processing. The data processing software comprises basic modules such as data preprocessing, fraud detection, and result output. The information processing technologies used, such as text analysis, statistical analysis, data mining, machine / deep learning, and software engineering techniques, are also general-purpose technologies. Therefore, it is relatively difficult to form proprietary technical features that require protection in terms of the basic data collection and processing flow, technology, and software structure.

[0004] The proprietary technical features of fraud detection methods are mainly reflected in the proprietary technologies formed by the relationship division and definition methods of the graph and the algorithm fusion methods. After analyzing these two aspects of the existing technologies, it was found that the current fraud identification technology still has the following limitations: (1) It only provides fraud identification for a single object, lacks correlation analysis of multi-source, heterogeneous and heterogeneous data, and the constructed graph only focuses on a single main object; (2) The fraud identification model is single-object oriented and does not utilize the information of different objects to identify the complex relationships between different objects; (3) It cannot provide interpretability on multiple graphs. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a fraud detection solution for supply chain finance, employing the following technical solution:

[0006] This invention provides a fraud detection system for supply chain finance, used to detect fraud risks from multi-source information in supply chain finance. It is characterized by comprising: a data acquisition module, a graph construction module, a fraud identification module, and a fraud interpretation module. The data acquisition module collects data objects from different data sources and performs data preprocessing. The data objects include at least collateral and enterprises. The graph construction module, communicating with the data acquisition module, models the preprocessed data objects based on their corresponding knowledge ontology and relationships, generating a knowledge graph corresponding to the data objects. The fraud identification module, communicating with the graph construction module, learns the representations of data objects within their respective graphs and shares the complex relationships between graphs to determine whether different data objects, i.e., collateral and enterprises, are involved in fraud. The fraud interpretation module, communicating with the fraud identification module, visualizes the node characteristics and graph relationships of collateral and enterprises suspected of fraud, providing explanations for fraud determinations to risk control personnel.

[0007] The fraud detection system for supply chain finance provided by this invention may also have the following technical features: the data acquisition module has a data acquisition unit and a preprocessing unit. The data acquisition unit is used to collect the main data corresponding to the data objects of different natures through different acquisition methods. The acquisition methods include at least manual input, calling internal business systems, calling third-party interfaces, and web data crawling. The preprocessing unit performs data preprocessing on the data objects and their corresponding main data collected by the data acquisition unit. The data preprocessing includes at least data parsing, data deduplication, data fusion, and tag processing.

[0008] The fraud detection system for supply chain finance provided by this invention may also have the following technical features: when the data object is an enterprise, its corresponding subject data includes at least the enterprise's registration time, registered capital, paid-in capital, industry to which the enterprise belongs, personnel, enterprise email, telephone, asset size, sales revenue, invoices, orders, contracts, and logistics; personnel include at least the enterprise's shareholders, senior executives, and legal representatives; when the nature of the data object is collateral, its corresponding subject data includes at least the type of collateral, the amount of collateral, the enterprise involved in the collateral, the type of collateral, and warehousing facilities.

[0009] The fraud detection system for supply chain finance provided by this invention may also have the following technical features: the graph construction module has a knowledge modeling unit and a knowledge extraction unit. The knowledge modeling unit uses a top-down approach to perform knowledge modeling based on knowledge ontology and relationships from preprocessed data to generate a knowledge graph. The knowledge modeling includes domain ontology, domain attributes, and domain relationships. The knowledge extraction unit extracts knowledge from the modeling results and uses knowledge representation forms based on the knowledge extraction results to obtain knowledge representations. Knowledge extraction includes structured data knowledge extraction, and the knowledge representation forms include two forms: entity-relationship-attribute and entity-relationship-entity.

[0010] The fraud detection system for supply chain finance provided by this invention also has the following technical features: For knowledge modeling of data objects as collateral: the collateral domain ontology includes collateral, collateral type, mortgage type, and warehousing facilities; collateral domain relationships include the affiliation between collateral and collateral type, collateral and mortgage type, and collateral and warehousing facilities; collateral domain attributes include amount, collateral currency, occurrence time, and expiration time. For knowledge modeling of data objects as enterprises: the enterprise domain ontology includes enterprises, personnel, email addresses, telephone numbers, invoices, orders, contracts, and logistics; enterprise domain relationships include the affiliation between enterprises and personnel types, enterprises and email addresses, enterprises and telephone numbers, and invoice flow, order flow, contract flow, and logistics relationships between enterprises; enterprise domain attributes include enterprise registration time, registered capital, paid-in capital, and the industry to which the enterprise belongs; enterprise domain relationships also include the affiliation between collateral and enterprise type.

[0011] The fraud detection system for supply chain finance provided by this invention may also have the following technical features: the data acquisition module has a storage unit that stores data in a database manner, including a graph database and a relational database, which respectively store the knowledge graph and knowledge representation generated by the graph construction module.

[0012] The fraud detection system for supply chain finance provided by this invention may also have the following technical features, wherein the structured data knowledge extraction includes at least transformation extraction, which is a declarative language based on the relationship between relational database rules and knowledge representation forms, to transform structured data into collateral and enterprise knowledge for knowledge extraction.

[0013] The fraud detection system for supply chain finance provided by this invention may also have the following technical features: the fraud identification module includes a Metapath extraction unit, a GCN node embedding unit, a node embedding sharing unit, and a fraud determination unit. The Metapath extraction unit extracts multiple homogeneous graphs from the input heterogeneous knowledge graph, i.e., the collateral graph or the enterprise graph, for subsequent model learning. The GCN node embedding unit uses GCN to learn node embeddings for the multiple homogeneous graphs extracted by the Metapath extraction unit, and merges the embeddings obtained from different homogeneous graphs to generate collateral embeddings and enterprise embeddings. The node embedding sharing unit shares the collateral embeddings and enterprise embeddings, wherein the enterprise embeddings are merged into the collateral embeddings using the Concat method, and the collateral embeddings are processed by the Att-BLSTM model and then merged into the embeddings of the enterprises involved. The fraud determination unit classifies the collateral embeddings and enterprise embeddings using the Softmax method to determine whether each collateral and each enterprise is fraudulent.

[0014] The fraud detection system for supply chain finance provided by this invention may also have the following technical features: The fraud interpretation module includes a collateral fraud interpretation unit and a company fraud interpretation unit. The collateral fraud interpretation unit is used to visualize the characteristics and graph relationships of collateral determined to be fraudulent. The displayed content includes at least the importance of collateral node features, heterogeneous relationships in the collateral graph, companies involved in the collateral, and heterogeneous relationships in the company graphs of the companies involved in the collateral. The corresponding interpretation objects include at least collateral graph interpretation, querying related companies, company graph interpretation, and relationship merging. The company fraud interpretation unit is used to visualize the characteristics and graph relationships of companies determined to be fraudulent. The displayed content includes at least the importance of company node features, heterogeneous relationships in the company graph, collateral involved in the company, and heterogeneous relationships in the collateral graphs of the collateral involved in the company. The corresponding interpretation objects include at least company graph interpretation, selecting related collateral, collateral graph interpretation, and relationship merging.

[0015] The fraud detection system for supply chain finance provided by this invention may also have the following technical feature: the collateral involved by the enterprise is obtained by selecting the K collaterals with the highest scores from the Attention scores output by the Att-BLSTM model.

[0016] The invention's function and effects:

[0017] According to the present invention, a fraud detection system for supply chain finance includes a data acquisition module, a graph construction module, a fraud identification module, and a fraud interpretation module. First, the data acquisition module can acquire various data objects, including enterprises, personnel, collateral, invoices, orders, contracts, and logistics, from different data sources using different acquisition methods. Second, the graph construction module generates a knowledge graph corresponding to the data objects based on ontology and relationship modeling from the acquired data. Then, the fraud identification module uses the knowledge graph as input and employs fraud detection and identification methods to obtain fraud determinations for different objects. Finally, the fraud interpretation module provides a visual representation of the object characteristics and graph relationships for the fraud determinations, and provides explanations of the fraud determinations to risk control personnel to assist their subsequent decision-making.

[0018] In summary, the fraud detection system for supply chain finance of this invention effectively utilizes various risk factors related to data fraud in supply chain finance, including enterprises, personnel, emails, telephones, invoices, orders, contracts, and logistics, targeting heterogeneous collateral and enterprise-related data. It constructs a supply chain finance fraud detection system based on the status of collateral involved in supply chain finance as the core clue and enterprise credit as a secondary factor. The system proposes a process for constructing a knowledge graph of collateral and enterprises, and can effectively identify fraud based on this knowledge graph. It compensates for the deficiencies of fraud identification models in terms of heterogeneous and heterogeneous information, improves the relationship between collateral, enterprises, and their related data, provides visualized fraud explanations, enhances the accuracy of the indicators, and provides interpretability of the model. Attached Figure Description

[0019] Figure 1 This is a structural block diagram of a fraud detection system for supply chain finance in an embodiment of the present invention.

[0020] Figure 2 This is a structural block diagram of the fraud detection module in an embodiment of the present invention.

[0021] Figure 3 This is a flowchart illustrating the fraud detection process based on a fraud detection system for supply chain finance in this embodiment of the invention. Detailed Implementation

[0022] The experimental environment of this invention is set as supply chain accounts receivable financing fraud. Accounts receivable financing is a method by which financial institutions provide financing to SMEs based on actual collateral between core enterprises and upstream and downstream SMEs in the supply chain, helping to alleviate the financing difficulties and capital shortages of SMEs. The business model of accounts receivable financing involves three main roles: financial institutions, core enterprises, and SMEs. Common types of supply chain accounts receivable fraud are false collateral, self-financing, and duplicate financing. False collateral refers to the falsification of information such as trade and logistics. Self-financing refers to a close relationship between the borrower and the guarantor, such as relatives or friends, or actual control by the borrower. Duplicate financing refers to issuing multiple warehouse receipts for the same batch of goods, obtaining loans from multiple financing institutions through duplicate accounts receivable.

[0023] The above-mentioned experimental environment is a typical supply chain finance fraud scenario. Current fraud detection technologies, solutions, and systems for this scenario have the following limitations:

[0024] The data sources used are limited in variety and type, failing to consider the rich corporate information involved in the collateral. There is a lack of correlation analysis of multi-source, heterogeneous, and heterogeneous data, and no knowledge graph to support deeper correlation analysis. Current fraud detection models are primarily rule-based, relying too heavily on expert knowledge for fraud identification. Methods based on machine learning and deep learning only focus on individual collateral or corporate entities, neglecting the complex relationships between them and failing to provide adequate interpretability.

[0025] The fraud detection system in this embodiment can collect multi-source heterogeneous information about the collateral and enterprises involved in the scenario, construct various relationships between the collateral and enterprise objects through graphs, effectively identify fraudulent collateral and enterprises, and provide timely warnings to assist risk control personnel.

[0026] To make the technical means, creative features, objectives and effects of this invention easy to understand, the following describes in detail a fraud detection system for supply chain finance based on the present invention, in conjunction with embodiments and accompanying drawings.

[0027] Example

[0028] Figure 1 This is a structural block diagram of a fraud detection system for supply chain finance in an embodiment of the present invention.

[0029] like Figure 1 As shown, the fraud detection system for supply chain finance (hereinafter referred to as the fraud detection system) in this embodiment has a data acquisition module 10, a graph construction module 20, a fraud identification module 30, and a fraud interpretation module 40. These four modules are connected in a communication manner.

[0030] The data acquisition module 10 collects data objects from different data sources and performs data preprocessing. Data sources include manual input, internal business systems, third-party interfaces, and the network.

[0031] The data acquisition module 10 has a data acquisition unit, a preprocessing unit, and a storage unit.

[0032] The data acquisition unit is used to collect the main data corresponding to data objects of different natures using different acquisition methods. Acquisition methods include manual input, calling internal business systems, calling third-party interfaces, and web crawling. Specifically, the third-party interface calling method includes using fields such as the company's full name and personnel ID numbers to extract related data objects from search results; the web crawling method includes using company keywords to call web search engines and extract related data objects from search results.

[0033] The preprocessing unit is used to preprocess the data objects and their corresponding main data collected by the data acquisition unit. The data preprocessing method includes data parsing, data deduplication, data fusion and label processing.

[0034] The storage unit stores data in a database manner, including graph databases and relational databases, which are used to store knowledge graphs and knowledge representations, respectively.

[0035] In this embodiment, the data objects collected include enterprises, personnel, collateral, invoices, orders, contracts, and logistics. The data structure of these data objects is structured data. The process of collecting and processing the above-mentioned data objects of different natures through the data acquisition module 10 is as follows:

[0036] Taking enterprise data as an example: the collected content, i.e., the main data, includes the enterprise's registration date, registered capital, paid-in capital, industry, personnel, email address, telephone number, asset size, sales revenue, invoices, orders, contracts, and logistics. Personnel include shareholders, senior management, and legal representative. Data sources include manual input, third-party interfaces, and the internet. Collection methods include manual input and calling third-party interfaces. Third-party interfaces are those that can extract current enterprise information, including third-party enterprise information systems (such as Tianyancha and Qichacha) and government information systems (Credit China). Data preprocessing methods include content extraction and tagging, and the data is stored in a database.

[0037] Taking data objects whose nature is collateral as an example: the collected content, i.e., the main data, includes the type of collateral, the amount of collateral, the companies involved in the collateral, the type of collateral, and warehousing facilities. The data sources include manual input and internal business systems. Collection methods include manual input and accessing internal business systems. Internal business systems are information systems capable of extracting current collateral, including the information systems of the institution (e.g., supply chain finance platforms). Data preprocessing methods include content extraction and tagging, and the data is stored in a database.

[0038] Given the diverse types of objects and complex information involved in supply chain finance scenarios, the consideration of fraud detection in supply chain finance needs to take into account different aspects such as the status of collateral and corporate credit. Different objects themselves contain rich heterogeneous information and have different graph structure characteristics. At the same time, there are rich relationships between different objects.

[0039] Therefore, this embodiment takes the status of collateral involved in supply chain finance as the core clue and corporate credit as an auxiliary factor, and constructs different graphs to maintain the differences in the characteristics and complex relationships of different objects. The ontology, attributes, and relationships of the graph are designed separately for different objects. For supply chain finance scenarios, the graph relationship dimensions include trade flow (orders or contracts), capital flow, logistics, invoice flow, information flow, supply chain relationships, and social relationships.

[0040] The knowledge graph construction module 20, based on the preprocessed data objects, models them according to the corresponding knowledge ontology and relationships, generating a knowledge graph corresponding to the data objects. This knowledge graph construction module 20 includes a knowledge modeling unit and a knowledge extraction unit.

[0041] The knowledge modeling unit employs a top-down approach to perform knowledge modeling based on knowledge ontology and relationships from preprocessed data, generating a knowledge graph. Knowledge modeling includes domain ontology, domain relationships, and domain attributes.

[0042] Since the collateral and enterprise domain knowledge graph addresses specific issues such as anomaly and risk analysis in supply chain finance, and the definitions of ontology, relations, and attributes all serve specific risk problems, the primary approach is to manually (by domain experts) construct the ontology. This involves a top-down approach to building the knowledge graph, starting with the design of the ontology, ontology library, and data schema. Then, entity information is injected into this framework. For the analysis of accounts receivable financing fraud cases in supply chain finance, a seven-step method is used to construct the ontology.

[0043] This embodiment employs the seven-step method developed by Stanford University School of Medicine. This seven-step method is an ontology development approach based on the ontology development tool Protégé, and is quite practical, primarily used for constructing domain ontologies. The seven steps of this method are: determining the domain and scope of the ontology; examining the possibility of reusing existing ontologies; listing the important terms in the ontology; defining classes and class hierarchies; defining class properties; defining property facets; and creating instances.

[0044] The development tool Protégé supports the expression of various knowledge elements and the definition of various knowledge rules. More importantly, it offers multiple extension interfaces, supports various language rules for ontology representation, and provides conversion and export functions between different formats. This tool is used to construct collateral and enterprise knowledge graph ontology. Specifically:

[0045] In this embodiment, the knowledge modeling for data objects that are collateral is as follows:

[0046] The subject of the collateral domain includes the collateral, the type of collateral, the type of mortgage, and warehousing facilities;

[0047] The relationships within the collateral domain include the ownership relationships between collateral and collateral types, the ownership relationships between collateral and mortgage types, and the ownership relationships between collateral and warehousing facilities;

[0048] The attributes of collateral include the amount, currency of the collateral, time of occurrence, and expiration date.

[0049] The following content pertains to knowledge modeling for enterprise-level data objects:

[0050] Enterprise domain ontology includes enterprises, personnel, email addresses, telephone numbers, invoices, orders, contracts, and logistics;

[0051] Enterprise domain relationships include the affiliation between enterprises and personnel types, the affiliation between enterprises and email addresses, the affiliation between enterprises and telephone numbers, and the relationships between enterprises in terms of invoice flow, order flow, contract flow, and logistics.

[0052] The enterprise's industry attributes include the enterprise's registration date, registered capital, paid-in capital, and the industry to which the enterprise belongs.

[0053] In this embodiment, the aforementioned enterprise domain relationship also includes the ownership relationship between collateral and enterprise type.

[0054] The knowledge extraction unit extracts knowledge from the modeling results described above, and uses knowledge representation forms based on the extracted knowledge to obtain a knowledge representation. These knowledge representation forms include two types: entity-relationship-attribute and entity-relationship-entity.

[0055] This embodiment employs structured data knowledge extraction, which is a transformational extraction method. That is, based on a declarative language describing the relationship between relational database rules and knowledge representation forms, structured data is transformed into collateral and enterprise knowledge for knowledge extraction.

[0056] The multiple knowledge graphs generated in this embodiment can express the multi-dimensional relationships between heterogeneous and heterogeneous data, enabling penetrating analysis and cross-authentication, and improving fraud identification capabilities and accuracy. This is something that existing inventions cannot achieve through rules or single graphs.

[0057] Considering that different objects in the supply chain finance scenario, including enterprises, personnel, collateral, invoices, orders, contracts, logistics, etc., all contain their own information characteristics and rich heterogeneous relationships, and there are also complex relationships between different objects, fraud detection for supply chain finance needs to utilize the heterogeneous and heterogeneous information of different objects at the same time.

[0058] Based on the multiple knowledge graphs generated by the graph construction module 20, this embodiment employs a multi-task architecture for algorithm fusion to clearly and accurately determine the internal relationships and connections between different graphs. First, it learns the representations of different objects within their respective graphs. Through representation sharing, it learns the complex relationships between graphs, determines the fraudulent nature of different objects, and provides interpretability across multiple graphs after fraud determination, including explanations of node features and relationships. The implementation process is as follows:

[0059] The fraud detection module 30 is used to learn the representations of data objects in the knowledge graph within their respective graphs and to share the complex relationships between the graphs, in order to determine whether different data objects, namely collateral and enterprises, are involved in fraud. This fraud detection module 30 includes a Metapath extraction unit, a GCN node embedding unit, a node embedding sharing unit, and a fraud determination unit.

[0060] The Metapath extraction unit is used to extract multiple homogeneous graphs from the input heterogeneous knowledge graph, i.e., the collateral graph or the enterprise graph, for subsequent model learning. In this embodiment, the homogeneous graphs extracted by the Metapath extraction unit for the collateral graph include collateral-collateral type-collateral, collateral-collateral type-collateral, and collateral-warehousing facility-collateral; the homogeneous graphs extracted for the enterprise graph include enterprise-shareholder-enterprise, enterprise-executive-enterprise, enterprise-legal person-enterprise, enterprise-enterprise, enterprise-email-enterprise, enterprise-telephone-enterprise, enterprise-invoice-enterprise, enterprise-order-enterprise, enterprise-contract-enterprise, and enterprise-logistics-enterprise.

[0061] The GCN node embedding unit is used to learn node embeddings from multiple homogeneous graphs extracted by the Metapath extraction unit using GCN, and then merges the embeddings from different homogeneous graphs to generate collateral embeddings and enterprise embeddings. The node embedding learning process includes three steps: feature preprocessing, GCN, and feature postprocessing. In this embodiment, both feature preprocessing and feature postprocessing use a single-layer feedforward neural network, and the GCN is two-layered. The different embeddings are merged by directly concatenating the feature vectors.

[0062] The node embedding sharing unit shares the collateral embedding and the enterprise embedding. The enterprise embedding is merged into the collateral embedding through the Concat method, and the collateral embedding is processed by the Att-BLSTM model and then merged into the embedding of the enterprise involved.

[0063] In this embodiment, the process of obtaining the embedded collateral is as follows:

[0064] First, obtain relevant company information: search for the companies involved in the collateral, including the company that initiated the collateral and the company that received the collateral.

[0065] Next, obtain the enterprise embedding: query the embeddings of the collateral initiating enterprise and the collateral receiving enterprise that have been calculated in the aforementioned enterprise graph.

[0066] Finally, the collateral embedding is calculated: the obtained enterprise embedding is merged into the collateral embedding by directly concatenating the feature vectors.

[0067] Similarly, the process for obtaining the enterprise embedding in this embodiment is as follows:

[0068] First, obtain the relevant collateral: query all the collateral that the company has participated in throughout its history.

[0069] Next, the collateral embedding is calculated: the final embedding of historical collateral is calculated using the Att-BLSTM method, which includes a bidirectional LSTM and an Attention model.

[0070] Finally, the enterprise embedding is calculated: the collateral embedding output by the Att-BLSTM method is merged into the enterprise embedding by directly concatenating the feature vectors.

[0071] The fraud detection unit classifies collateral embeddings and enterprise embeddings using the Softmax method to determine whether each collateral and each enterprise is fraudulent. The model's Loss function includes collateral fraud detection Loss and enterprise fraud detection Loss, calculated using cross-entropy loss. The weights for collateral fraud detection Loss and enterprise fraud detection Loss are the same, both being 0.5.

[0072] In this embodiment, the fraud detection module 30 performs the aforementioned fraud detection on the collateral map and enterprise map as follows: Figure 2 As shown.

[0073] Furthermore, addressing the high demand for interpretability of graphs, this embodiment also provides a fraud explanation module to explain and visualize object features and relationships across multiple graphs, a feature not offered by other existing deep learning-based methods. (See below:)

[0074] The fraud explanation module 40 includes a collateral fraud explanation unit and an enterprise fraud explanation unit, which are used to visualize the node characteristics and graph relationships of collateral and enterprises that are involved in fraud, and provide explanations for fraud determination for risk control personnel.

[0075] The collateral fraud interpretation unit is used to visualize the characteristics and graph relationships of collateral determined to be fraudulent. The interpretation objects include the importance of collateral node features, heterogeneous relationships in the collateral graph, the companies involved in the collateral, and heterogeneous relationships in the company graph of the companies involved in the collateral. The corresponding collateral fraud interpretation includes collateral graph interpretation, querying related companies, company graph interpretation, and relationship merging. Specifically:

[0076] Collateral Graph Interpretation: The GNNExplainer model is used to interpret the target collateral in the collateral graph, outputting collateral node features and heterogeneous correlations. Related Enterprise Query: This function queries the enterprises involved in the collateral, including the collateral initiating enterprise and the collateral receiving enterprise. Enterprise Graph Interpretation: The GNNExplainer model is used to interpret multiple target enterprises separately in the enterprise graph, outputting enterprise node features and heterogeneous correlations. Relationship Merging: This refers to merging the heterogeneous relationships of the target collateral and the heterogeneous relationships between two target enterprises into a single graph, adding an edge between the target collateral and each target enterprise to represent the collateral relationship.

[0077] The Enterprise Fraud Explanation Unit is used to visualize the characteristics and graph relationships of enterprises identified as fraudulent. The displayed content includes the importance of enterprise node features, heterogeneous relationships in the enterprise graph, the collateral involved in the enterprise, and heterogeneous relationships in the collateral graph of the collateral involved in the enterprise. The corresponding Enterprise Fraud Explanation objects include enterprise graph explanation, querying related collateral, collateral graph explanation, and relationship merging. Specifically:

[0078] Enterprise Graph Interpretation: The GNNExplainer model is used to interpret the target enterprise in the enterprise graph, outputting enterprise node features and heterogeneous correlations. Related Collateral Query: Based on the Attention score output by the Att-BLSTM model, the weights of the target enterprise's historical collaterals are used, and the K collaterals with the highest Attention scores are selected. Collateral Graph Interpretation: The GNNExplainer model is used to interpret multiple target collaterals separately in the collateral graph, outputting collateral node features and heterogeneous correlations. Relationship Merging: This refers to merging the heterogeneous relationships of the target enterprise and the K target collaterals into a single graph, adding an edge between the target enterprise and each target collateral to represent the collateral relationship.

[0079] In this embodiment, a WeChat mini-program / server system architecture is adopted. The WeChat mini-program provides users with an interface displaying collateral and corporate risks. The display method is not limited to this.

[0080] Figure 3 This is a flowchart illustrating the fraud detection process based on a fraud detection system for supply chain finance in this embodiment of the invention.

[0081] like Figure 3 As shown, the fraud detection process of a fraud detection system for supply chain finance scenarios is as follows:

[0082] First, the data acquisition module 10 collects and preprocesses data objects from different data sources using corresponding acquisition methods. In this embodiment, the data objects include collateral and enterprises. Second, the collected and processed data objects are used as input to the graph construction module 20. Based on the knowledge ontology and relationships corresponding to the input data objects, a knowledge graph is generated, namely, a collateral graph and an enterprise graph. Then, the fraud detection module 30 learns the representations of the data objects in their respective graphs and shares the complex relationships between the graphs to predict and determine whether the collateral and enterprises are fraudulent. Finally, the fraud explanation module 40 visualizes the characteristics of collateral nodes and enterprise nodes identified as fraudulent, as well as the graph relationships between collateral and enterprises, providing risk control personnel with an explanation of the relationship between fraudulent collateral and enterprises.

[0083] In summary, this embodiment provides a fraud detection system for supply chain finance. Each component has relatively complete functions and procedures, facilitating independent implementation and upgrades of each module. Furthermore, it is applicable to fraud detection in supply chain finance scenarios, addressing the shortcomings of existing detection systems in considering enterprise, collateral, invoices, orders, contracts, logistics, and their related data and relationships, while improving the accuracy of the indicators.

[0084] Functions and effects of the examples:

[0085] This embodiment provides a fraud detection system for supply chain finance, comprising a data acquisition module, a knowledge graph construction module, a fraud identification module, and a fraud interpretation module. First, the data acquisition module acquires various data objects from different sources using different acquisition methods, including enterprises, personnel, collateral, invoices, orders, contracts, and logistics. Second, the knowledge graph construction module generates a knowledge graph corresponding to the data objects based on ontology and relationship modeling from the acquired data. Then, the fraud identification module uses the knowledge graph as input and employs fraud detection and identification methods to obtain fraud determinations for different objects. Finally, the fraud interpretation module visualizes the object characteristics and knowledge graph relationships of the fraud determinations and provides explanations of the fraud determinations to risk control personnel.

[0086] In this embodiment, because the data acquisition module can use corresponding methods to acquire, preprocess, and store associated data for data objects and content of different natures, it can achieve data acquisition and processing from different sources, thus having greater applicability.

[0087] In addition, the graph construction module forms a complete knowledge graph construction process through knowledge modeling and knowledge extraction, which retains rich node, relationship and attribute information. It uses domain association information to construct graphs for collateral and enterprises respectively, effectively utilizing the characteristics and structures of different domains, which has important and beneficial effects on subsequent applications.

[0088] Furthermore, the fraud detection module simplifies the complex relationships between different objects in supply chain finance by utilizing graph nodes, relationships, and attribute information to compensate for the shortcomings of fraud detection models in analyzing heterogeneous and heterogeneous information. It also employs a multi-task framework, which effectively utilizes and integrates information from different objects, thereby improving fraud detection accuracy.

[0089] Meanwhile, in response to the weak interpretability of deep learning models and the high demand for interpretability in financial fraud, this example also uses a fraud explanation module to explain the complex relationships between node features and multiple heterogeneous graphs, which meets the needs of risk control personnel, improves user experience, enhances interpretability, and provides assistance for subsequent decision-making by risk control personnel.

[0090] In summary, the fraud detection system for supply chain finance in this embodiment targets heterogeneous collateral and enterprise-related data, including enterprise, personnel, email, telephone, invoices, orders, contracts, logistics, etc. It constructs a supply chain finance fraud detection system based on the status of collateral involved in supply chain finance as the core clue and enterprise credit as a secondary factor. The system proposes a process for constructing a knowledge graph of collateral and enterprises, and can effectively identify fraud based on this knowledge graph. This compensates for the deficiencies of fraud identification models in terms of heterogeneous and heterogeneous information, improves the relationship between collateral, enterprises, and their related data, provides visual explanations of fraud, enhances the accuracy of the indicators, and provides interpretability of the model.

[0091] The above embodiments are only used to illustrate specific implementations of the present invention, and the present invention is not limited to the scope of the description of the above embodiments.

Claims

1. A fraud detection system for supply chain finance, used to detect fraud risks from multi-source information in supply chain finance, characterized in that, include: Data acquisition module, graph construction module, fraud detection module, and fraud interpretation module; The data acquisition module collects data objects from different data sources and performs data preprocessing. These data objects include, at a minimum, collateral and enterprises. The knowledge graph construction module is communicatively connected to the data acquisition module. Based on the preprocessed data objects, it models the knowledge ontology and relationships corresponding to the data objects to generate a knowledge graph corresponding to the data objects. The fraud detection module is communicatively connected to the knowledge graph construction module. It learns the representations of data objects in the knowledge graph within their respective graphs and shares the complex relationships between the graphs to determine whether different data objects, namely collateral and enterprises, are involved in fraud. The fraud interpretation module, communicatively connected to the fraud identification module, is used to visualize the node characteristics and graph relationships of collateral and enterprises suspected of fraud, providing risk control personnel with explanations for fraud determination. The fraud detection module includes a Metapath extraction unit, a GCN node embedding unit, a node embedding sharing unit, and a fraud determination unit. The Metapath extraction unit is used to extract multiple homogeneous graphs from the input heterogeneous knowledge graph, i.e., the collateral graph or the enterprise graph, for subsequent model learning. The GCN node embedding unit performs node embedding learning on multiple homogeneous graphs extracted by the Metapath extraction unit using GCN, and merges the embeddings obtained from different homogeneous graphs to generate collateral embeddings and enterprise embeddings. The node embedding sharing unit shares the collateral embedding and the enterprise embedding. The enterprise embedding is merged into the collateral embedding using the Concat method, and the collateral embedding is processed by the Att-BLSTM model and then merged into the embedding of the enterprises involved. The fraud determination unit uses the Softmax method to classify the collateral embedding and enterprise embedding respectively, and determines whether each collateral and each enterprise is fraudulent.

2. The fraud detection system for supply chain finance according to claim 1, characterized in that: in, The data acquisition module includes a data acquisition unit and a preprocessing unit. The data acquisition unit is used to acquire the main data corresponding to data objects of different natures using different acquisition methods. The data collection methods include at least manual input, calling internal business systems, calling third-party interfaces, and web data crawling. The preprocessing unit performs data preprocessing on the data objects and their corresponding main data collected by the data acquisition unit. The data preprocessing includes at least data parsing, data deduplication, data fusion, and label processing.

3. A fraud detection system for supply chain finance according to claim 2, characterized in that: in, When the data object is an enterprise, its corresponding main data must include at least the enterprise's registration date, registered capital, paid-in capital, industry, personnel, enterprise email, telephone number, asset size, sales revenue, invoices, orders, contracts, and logistics. The personnel must include at least the enterprise's shareholders, senior executives, and legal representative. When the nature of the data object is collateral, its corresponding subject data shall include at least the type of collateral, the amount of collateral, the enterprise involved in the collateral, the type of collateral, and the warehousing facilities.

4. A fraud detection system for supply chain finance according to claim 1, characterized in that: in, The knowledge graph construction module includes a knowledge modeling unit and a knowledge extraction unit. The knowledge modeling unit employs a top-down approach to perform knowledge modeling based on knowledge ontology and relationships from the preprocessed data, generating a knowledge graph. The knowledge modeling includes domain ontology, domain attributes, and domain relationships. The knowledge extraction unit extracts knowledge from the modeling results and uses a knowledge representation form based on the extracted knowledge to obtain a knowledge representation. The knowledge extraction includes structured data knowledge extraction. The knowledge representation includes two forms: entity-relationship-attribute and entity-relationship-entity.

5. A fraud detection system for supply chain finance as described in claim 4, Its features are: in, Knowledge modeling for the data object as collateral: The subject of collateral includes collateral, types of collateral, types of mortgage, and warehousing facilities. Collateral relationships include ownership relationships between collateral and collateral types, ownership relationships between collateral and mortgage types, and ownership relationships between collateral and warehousing facilities. The attributes of collateral include the amount, currency of the collateral, time of occurrence, and deadline. Knowledge modeling for the data object being an enterprise: Enterprise domain ontology includes enterprises, individuals, email addresses, telephone numbers, invoices, orders, contracts, and logistics. Enterprise domain relationships include affiliations between enterprises and personnel types, between enterprises and email addresses, between enterprises and telephone numbers, and relationships related to invoice flow, order flow, contract flow, and logistics between enterprises. Enterprise industry attributes include the enterprise's registration date, registered capital, paid-in capital, and the industry to which the enterprise belongs; The enterprise domain relationship also includes the affiliation relationship between collateral and enterprise type.

6. A fraud detection system for supply chain finance according to claim 4, characterized in that: in, The data acquisition module also has a storage unit, which stores data in a database manner. The database includes a graph database and a relational database, which respectively store the knowledge graph and the knowledge representation generated by the graph construction module.

7. A fraud detection system for supply chain finance according to claim 4, characterized in that: in, The structured data knowledge extraction includes at least transformation extraction. The transformation and extraction is based on a declarative language that describes the relationship between relational database rules and knowledge representation forms, transforming the structured data into collateral and enterprise knowledge for knowledge extraction.

8. A fraud detection system for supply chain finance according to claim 1, characterized in that: in, The fraud interpretation module includes a collateral fraud interpretation unit and a corporate fraud interpretation unit. The collateral fraud interpretation unit is used to visualize the characteristics and graph relationships of collateral determined to be fraudulent. The displayed content includes at least the importance of collateral node features, heterogeneous relationships in the collateral graph, enterprises involved in the collateral, and heterogeneous relationships in the enterprise graph of enterprises involved in the collateral. The corresponding interpretation objects include at least collateral graph interpretation, querying related enterprises, enterprise graph interpretation, and relationship merging. The enterprise fraud interpretation unit is used to visualize the characteristics and graph relationships of enterprises determined to be fraudulent. The displayed content includes at least the importance of enterprise node characteristics, heterogeneous relationships in the enterprise graph, the collateral involved in the enterprise, and heterogeneous relationships in the collateral graph of the collateral involved in the enterprise. The corresponding interpretation objects include at least enterprise graph interpretation, selection of relevant collateral, collateral graph interpretation, and relationship merging.

9. A fraud detection system for supply chain finance according to claim 8, characterized in that: in, The collateral involved by the enterprise is obtained by selecting the K collaterals with the highest scores from the Attention scores output by the Att-BLSTM model.

Images