Credit risk assessment and transmission method based on multi-source data and knowledge spectrogram

Abstract

The invention discloses a credit risk assessment and transmission method based on multi-source data and a knowledge map, and the method comprises the steps: collecting credit event data from the multi-source data through an integrated crawler technology and a keyword-based event extraction technology, and carrying out the cleaning and conversion, so as to construct a credit knowledge map. The credit knowledge graph identifies each credit subject through a uniform resource identifier (URI), uses a credit evaluation index system as a data specification, and continuously updates through real-time data acquisition. In addition, the invention further provides a credit risk multi-path transmission composite influence probability calculation model, and credit risk management is optimized by analyzing influence factors of nodes in a credit subject association model.

Description

Technical Field

[0001] The present invention relates to the technical field of credit risk management, and in particular to a credit risk assessment and transmission method based on multi-source data and knowledge spectrum. Background Art

[0002] In recent years, as big data technology has been widely used, various platforms have used the data collected by their platforms to calculate credit indicators and build their own credit evaluation systems. However, there are many problems in this process, resulting in differences in information and evaluation results for the same object on different platforms, and there is no good coordination mechanism. The data is scattered, heterogeneous, and low-quality, making it difficult to directly apply it to judge the overall credit level of an object.

[0003] At present, my country has initially built a national credit information sharing platform to realize the collection and sharing of credit information of various departments and localities. The National Information Center has established the Credit China Network Platform to collect and provide personal credit information. However, in the process of collecting and sharing these credit information, there are credit information conflicts such as different credit measurement methods and inconsistent evaluation standards; information ownership and privacy protection in the process of information migration; inconsistent credit data structure in the process of information transmission, and inability to share data. In addition, the credit of different platforms still mainly relies on separate entry and maintenance, and re-rating based on the credibility of the entered information. Some platform credit subjects may forge false data, and some credit subjects may have data missing. Few studies focus on the conflicts between credit subject indicators and credit ratings caused by data missing and inconsistency under cross-platforms, and the models and methods of credit transmission are missing.

[0004] In response to the above problems, this study first collects credit events from multiple data sources, including cross-platform credit information sharing, financial news, and negative online public opinion events. Then, a credit knowledge graph is proposed and constructed to build a knowledge graph based on the existing database and update the knowledge graph through data obtained from the information collection part, providing path support for credit information transmission research. Finally, the modeling and measurement methods of influence are analyzed through network topology, credit event analysis, and credit subject composite impact factors. Summary of the invention

[0005] The purpose of this section is to summarize some aspects of embodiments of the present invention and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section and the specification abstract and the invention title of this application to avoid blurring the purpose of this section, the specification abstract and the invention title, and such simplifications or omissions cannot be used to limit the scope of the present invention.

[0006] In view of the differences in credit information on different platforms, the data is scattered, heterogeneous and low-quality, making it difficult to directly use it to judge an objective overall credit level. In addition, data sharing cannot be achieved, and the model and method of credit transmission are missing. Therefore, the present invention is proposed.

[0007] Therefore, the technical problem solved by the present invention is: to propose and construct a credit knowledge graph, to build a knowledge graph based on an existing database and to update the knowledge graph through data obtained through the information collection part, and to analyze the modeling and measurement methods of influence.

[0008] To solve the above technical problems, the present invention provides the following technical solutions: a credit risk assessment and transmission method based on multi-source data and knowledge spectrum, comprising the following steps: S1: using web crawler technology and event extraction technology to collect credit-related data from financial news websites, social media and other platforms, and clean and convert them; S2: based on the collected data, construct a credit knowledge spectrum, including entity linking, relationship extraction and knowledge fusion, to form a credit service network with wide coverage and full categories; S3: using social network analysis methods to study the correlation between credit subjects and construct a social network influence maximization problem model; S4: proposing a credit risk multi-path transmission compound impact probability calculation model to simulate the credit risk propagation process in the social network.

[0009] As a preferred solution of the credit risk assessment and transmission method based on multi-source data and knowledge spectrum described in the present invention, web crawlers are divided into general web crawlers, focused web crawlers, incremental web crawlers and deep web crawlers.

[0010] As a preferred solution of the credit risk assessment and transmission method based on multi-source data and knowledge spectrum described in the present invention, the steps of constructing the credit knowledge spectrum are as follows:

[0011] (1) Unify the labeling of existing credit entities and use the credit evaluation index system as a standard;

[0012] (2) Align the labeled entities according to the existing credit knowledge base, connect the existing knowledge with various relationships, determine the SPO triples, and establish the relationship between the entities in the knowledge graph;

[0013] (3) Perform update operations such as adding, modifying, and deleting relationships on the knowledge graph using the data obtained from real-time data collection;

[0014] (4) Obtain the knowledge graph of credit subject relationships as support for the credit transmission relationship network.

[0015] As a preferred solution of the credit risk assessment and transmission method based on multi-source data and knowledge spectrum described in the present invention, the correlation between credit subjects mainly includes investment, supply relationship, and debt relationship.

[0016] As a preferred solution of the credit risk assessment and transmission method based on multi-source data and knowledge spectrum described in the present invention, the social network influence analysis step includes a measurement method based on topological network degree centrality, influence node analysis based on the classical theory of weak connections, and the three-degree influence principle.

[0017] As a preferred solution of the credit risk assessment and transmission method based on multi-source data and knowledge spectrum described in the present invention, the multi-path transmission of credit risk includes single-path continuous transmission and multi-path transmission.

[0018] Beneficial effects of the present invention:

[0019] 1. Through multi-source data collection, the comprehensiveness and accuracy of credit information are improved.

[0020] 2. The construction of the knowledge graph enhances the relevance and comprehensibility of credit information.

[0021] 3. Social network influence analysis and credit risk transmission model provide new perspectives and methods for credit risk assessment and management. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative labor. Among them:

[0023] Figure 1 This is a diagram of the crawler system architecture provided by the present invention.

[0024] Figure 2 This is a basic principle diagram of the network crawler provided by the present invention.

[0025] Figure 3 This is a screenshot of the crawled data provided by the present invention.

[0026] Figure 4 The technical roadmap provided for the present invention.

[0027] Figure 5 This is the annual report feature extraction diagram provided by the present invention.

[0028] Figure 6 This is a diagram of the BERT model provided by the present invention.

[0029] Figure 7 This is a structural diagram of the news sentiment analysis model provided by the present invention.

[0030] Figure 8 A flow chart is established for the knowledge graph provided by the present invention.

[0031] Fig. 9 This is an example diagram of the credit knowledge graph provided by the present invention.

[0032] Fig.10 This is a display diagram of the credit knowledge graph provided by the present invention.

[0033] Fig.11 This is the credit subject association network diagram provided by the present invention.

[0034] Fig.12 A single-path continuous transfer diagram provided by the present invention.

[0035] Fig.13 A multi-path composite transfer diagram provided by the present invention.

[0036] Fig.14 A processing diagram of a ring provided by the present invention. DETAILED DESCRIPTION

[0037] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0038] 1. Perception and acquisition of credit risk events

[0039] The present invention perceives and acquires credit risk events based on multi-source data, wherein financial news and public opinion events are acquired by means of web crawlers. For the crawled financial news, a central event extraction system based on a credit event library is designed to extract credit-related information from financial news; a sequence labeling method based on a hybrid model is adopted, combining BERT with dynamic word representation capability, a bidirectional LSTM that can add trigger word information encoding, and CRF with global feature constraint effect, to achieve credit event argument extraction for central event sentences. Natural language processing technology is used to perform operations such as word segmentation and keyword extraction on the crawled network public opinion text data, and sentiment prediction is performed by constructing a word importance index model; a credit-related text sentiment classification model is constructed based on attention mechanism and deep learning, and sentiment classification labels are automatically obtained to provide a reference for credit risk transmission. The overall architecture of the system is as follows: Figure 1 shown.

[0040] Since media websites have their own characteristics and web page structures are not the same, it is necessary to write crawler strategies that are suitable for the web page structure of specific media websites. In essence, a crawler is an Internet information collection tool. In early 1993, Matthew Gray implemented the first Internet crawler Wanders. More and more academic papers on web crawlers have appeared at Internet conferences. Web crawlers can be divided into the following types according to system structure and implementation technology: General Purpose Web Crawler, Focused Web Crawler, Incremental Web Crawler, Deep Web Crawler. Due to the complex and diverse web page structures of different media platforms, it is impossible to use a single type of web crawler. Therefore, this study combines the two types of technologies, general crawlers and deep web crawlers, to achieve data collection. This study uses a breadth-first traversal algorithm and designs the following: Figure 2 The web crawler shown.

[0041] In addition, in order to improve the efficiency of data collection, the crawler is designed with multi-threading. The Java language itself provides support for multi-threading. There are two ways of multi-threading depending on the application inheritance or implementation of the object: one is that the concurrently running objects directly inherit the Java thread class Thread; the other is to define the concurrent execution object to implement the Runnable interface. The first method is adopted in the specific implementation of the program to implement the crawler thread class CrawlerThread. The CrawlerThread class inherits from the ThreadController class for multi-thread control. When there are URLs in the queue to be crawled or the specified number of layers has not been reached, a new thread is created, and the number of layers of the crawled page and the maximum number of threads are limited by parameters. When there are no URLs to be crawled, the CrawlerController terminates itself and notifies the ThreadController through the message system, and the ThreadController performs the queue conversion work.

[0042] The basic process of web crawler operation is mainly divided into the following steps: (1) web page acquisition; (2) web page parsing; (3) data storage. The principle of web crawler is to simulate the browser to send HTTP request. The crawler program sends a request to the web server through HTTP request, obtains the response from the server, and then downloads the web page, and then completes the crawling work of the crawler. Web page parsing is the operation of denoising the web page. Most of the web pages today exist in HTML format. Web page denoising is to extract the required content from the web page. When the web crawler extracts the content of the web page, it needs to analyze the HTML result of the web page to extract useful information, such as Figure 3 shown.

[0043] like Figure 4 As shown, the main technical routes after data collection are as follows:

[0044] (1) In the empirical analysis of factors related to corporate credit risk, the first choice is to use literature review, expert consultation, research and other methods to build a corporate credit risk index system. Then, a hybrid feature selection method is used to build a corporate credit risk index system.

[0045] (2) In constructing a credit risk early warning model that integrates unstructured text information, we first use deep learning-based text analysis technology to extract risk-related text information features from relevant texts (such as board of directors' operating reviews, etc.). Then, we combine machine learning methods (such as support vector machine (SVM), random forest (Randomforest RF), neural network (Neural network NN), etc.) to build a credit risk assessment model.

[0046] First, the BERT model is used to extract the financial text features, and then the AE neural network is introduced to reduce the dimension of this text feature. Figure 5 The BERT model uses the encoder structure of the bidirectional transformer and discards the decoder module, but the model structure is deeper than the transformer, so it automatically has bidirectional encoding capabilities and powerful feature extraction capabilities. Figure 6 shown.

[0047] An autoencoder (AE) consists of an input layer, a hidden layer, and an output layer, and mainly includes two parts: an encoder and a decoder. The autoencoder attempts to learn a certain representation of the input data in the hidden layer to reconstruct the input in the output layer, so its output is basically the same as the input. It is a neural network that reproduces the input signal as much as possible. In addition, the autoencoder does not require labels for learning features and is widely used for feature extraction in an unsupervised manner. The autoencoder can map high-dimensional input data to low-dimensional feature encodings through encoding operations, thereby reducing the data dimension.

[0048] From the input layer to the hidden layer corresponds to the encoding function, which maps the input to the potential representation space, as shown in Formula 1:

[0049] h=f(x)=f(Wx+b) (1)

[0050] Where f is a nonlinear activation function, usually Relu, W and b are the weights and biases of the encoder, respectively. Subsequently, the latent representation is mapped to the output And reconstruct it using the decoder. The decoder can be expressed by formula 2:

[0051]

[0052] Where g is the activation function of the decoder, W′ is the weight matrix, and b′ is the bias vector. Consistent with the input x, the corresponding loss function is shown in Formula 3:

[0053]

[0054] The credit warning benchmark model is as follows:

[0055] (1) Logistic regression

[0056] Logistic regression uses linear regression as its theoretical support. It can map the regression results to between 0 and 1 through the sigmoid function. Because Logistic regression is easy to implement and has efficient training, it is widely used in two types of classification tasks. Its model is shown in Formula 4:

[0057]

[0058] Among them, p xi represents the probability of an event occurring, X i represents the explanatory variables.

[0059] (2) XGBoost

[0060] XGBoost is an ensemble algorithm based on gradient boosting trees. It builds several weak evaluators by introducing a regularized loss function on the data, and integrates these weak classifiers with low accuracy into a strong classifier with high accuracy, which not only reduces the risk of model overfitting, but also makes its classification performance better than a single model. Since XGBoost uses pre-sorting, weighted quantiles, sparse matrix recognition, and cache recognition technologies, it has the advantages of parallel computing, controllable algorithm complexity, and strong generalization ability. Its objective function is shown in Formula 5:

[0061]

[0062] (3) SVM

[0063] SVM finds a hyperplane that divides the samples into two categories with the largest interval. SVM can perform linear or nonlinear classification, regression, and even outlier detection tasks. SVM is particularly suitable for the classification of small and medium-sized complex data sets. Given a training set Input Vector By solving the following quadratic programming problem, as shown in Formula 6, the solution for approximate linear classification of the SVM classifier can be obtained.

[0064]

[0065] In the financial field, many news are daily work reports, such as "the first company announcement" or the same news appears multiple times in the news list. Most online news comes from relevant financial news websites. Duplicate news may appear because: (1) the author submits the same manuscript to multiple websites in order to enhance the influence of the news; (2) the news website crawls news from other websites for display in order to increase the richness of the content. These duplicate news have the characteristics of high similarity, even the same title, and similar release time (usually within one to three days). These news need to be eliminated as much as possible. In order to eliminate duplicate news, we can use the characteristics of the news and use the similarity measurement method to identify whether the news content is repeated.

[0066] For the crawled financial news, the sentiment analysis model used can be divided into three parts: BERT layer, bidirectional LSTM layer and CRF layer. The overall architecture of the model is as follows Figure 7 As shown in Figure 1, the main goal of the BETR layer is to extract rich semantic information from news text. For a text sequence with a central event sentence as input After being embedded in the BERT layer, it will be converted into the corresponding embedding result sequence Embedding_result = E1, E2, ..., En, including word vector information and position information. In the bidirectional LSTM layer, the input part is composed of the encoding of the obtained trigger word position information and the word representation result of BERT. After that, the bidirectional LSTM learns the text with the trigger word position information added from the forward and reverse directions respectively, and passes the corresponding predicted label information to the CRF layer. As a global feature constraint method for the central event sentence, the CRF layer mainly forms feature constraints on the entire input sequence based on the semantic and syntactic features extracted by BERT and bidirectional LSTM, and uses this feature constraint to determine the most reasonable set of signature prediction values ​​output by the bidirectional LSTM.

[0067] 2. Credit Subject Knowledge Graph

[0068] At present, the data of various credit platforms have the characteristics of diverse sources, heterogeneity and heterogeneity. Credit data has not formed a unified standard. Information between different organizations, institutions and platforms is not interoperable, forming "information islands" and incompatibility. In response to the above problems with credit data, this project uses knowledge graph mining to analyze the potential connections between credit data, predicts and evaluates credit subjects, explores the feasibility of integrating and comprehensively mining multiple credit indicators, and improves the accuracy of credit prediction and evaluation for different enterprises and users.

[0069] Knowledge graph is a structured and semantic knowledge base that describes things, attributes and their connections in the objective world. The basic building block of commonly used knowledge graphs is the SPO triple (subject, predicate, object). Entities are connected to form a network through relationships. The entities or attributes represented by the subject and object in the triple are the nodes of the network, and the relationships represented by the predicate are the edges in the network. The focus of knowledge graph construction is entity linking and relationship extraction. Entity linking refers to finding the entities in the knowledge graph represented by the words representing entities from external data sources. Generally, similarity is used to judge the relevance between words and entities in the graph. For example, Huang Meigen et al. used semantic similarity calculation based on Bert to align the knowledge base index. Relationship extraction refers to the work of obtaining events from structured and unstructured data and abstracting them into triples. Due to the different degrees of data structuring, they are generally handled separately in work. By setting mapping rules and analyzing the correlation between entities, structured data can be used to extract events simply. In the research of Shi Yanping and others, the existing book metadata ontology was used to judge the relationship between various attributes in various resource platform management systems, which can quickly build and expand entity associations. Unstructured data is generally text. After entity extraction, the means of time extraction are divided into domain knowledge-driven and data-driven. Domain knowledge-driven is to set a series of templates and rules by domain experts and add them into the extraction, which can achieve rapid acquisition of events, but the scope of application is narrow, and it is only suitable for fields with small data scale, unified expression, and small amount of knowledge. Data-driven refers to the extraction model based on machine learning to learn domain knowledge from pre-labeled data and perform relationship identification.

[0070] The basic idea of ​​constructing a credit knowledge graph is: first, uniformly label the existing credit subjects and use the credit evaluation index system as a specification; second, align the labeled entities according to the existing credit knowledge base, connect the existing knowledge with various relationships, determine the SPO triples, and establish the relationship between the entities in the knowledge graph; then, use the data obtained from real-time data collection to update the knowledge graph by adding, modifying, deleting and other relationship operations; finally, obtain the credit subject relationship knowledge graph as the support for the credit transmission relationship network.

[0071] By consulting existing literature, we know that the relationship between subjects in the credit subject credit model mainly includes:

[0072] (a) Investment, especially controlling investment, means a strong connection between the shareholder entity and the corporate entity. The controlling enterprise can influence the operating decisions of the invested enterprise. The capital flow of the controlled enterprise has an important impact on the capital health of the holding enterprise.

[0073] (b) Supply relationship. The credit risks of suppliers and buyers are related to a certain extent. The financial risk of suppliers may cause buyers to be unable to complete production and affect profitability; the financial risk of buyers may affect the capital chain of suppliers and cause credit risk.

[0074] (c) Debt relationship. Creditor-liability relationship also has a transmission effect on credit risk. If a debtor with poor credit status fails to perform its obligations on time, it will weaken the creditor's ability to perform its obligations. However, if the debtor has a healthy cash flow, debt certificates such as bonds can also be used as a way for the creditor to perform its obligations, thereby enhancing the creditor's credit.

[0075] By analyzing the association model between the above credit entities, this study proposes a method to construct a credit knowledge graph, such as Figure 8 As shown. First, the credit data is preprocessed and all existing credit subjects are uniformly labeled to create a uniform resource identifier (URI) to determine that each entity corresponds to a unique credit subject; secondly, the credit evaluation index system in Topic 1 is used as the data specification, and the existing data is used to describe each entity, and the basic entity relationship is constructed as the relationship between enterprises, the relationship between enterprises and individuals, and the relationship between individuals; then the structured data and unstructured data obtained in the data collection are processed separately, and the unstructured data uses knowledge extraction technology to obtain the entity relationship and then uses entity alignment and knowledge fusion technology to integrate the data. The entities in the database are linked to create relationships between entities. The structured data extracts the relationship between concepts to establish a knowledge system, and the knowledge graph is updated and verified after obtaining the new SPO triples; finally, a dynamically updated knowledge graph is provided as a relationship network support for credit transmission.

[0076] The specific steps are as follows:

[0077] S1: Give each credit subject a permanent identifier URI to distinguish and retrieve the entity, such as Fig. 9 and Fig.10 As shown;

[0078] S2: Use the credit evaluation index system as a specification and describe the basic information of each entity with RDF properties;

[0079] S3: Entity alignment is performed. First, a block method based on a hash function is used to select potential matching record pairs from all entity pairs in the existing credit knowledge base as candidates, and the size of the candidate items is reduced as much as possible. In this process, attribute information is incorporated and bidirectional alignment is used in the alignment prediction stage to improve the accuracy.

[0080] S4: Through the edit distance method, the cost of the editing operation in the model is assigned different weights. The similarity is defined as the conversion of the string into a standard using the least weighted editing operation. The attribute similarity is calculated. The similarity between data points of different categories is calculated based on hierarchical clustering to divide the data at different levels. Finally, a tree-like clustering structure is formed to obtain entity similarity. The attribute similarity and entity similarity obtained above are used to link records. The SPO triples are established based on the relationship between the obtained entity pairs to construct the relationship between entities.

[0081] S5: Depending on the degree of data structuring, when using the data obtained in data collection to update the knowledge graph, it is divided into two forms: building entity relationships through structured data and unstructured data.

[0082] When constructing relationships in structured data, entities and relationships between entities are identified by analyzing the field content, primary key, foreign key and other information of the data table. When extracting information, the concept of the entity (i.e., ontology type) and entity attributes can be extracted from the table describing the entity, and the entity alignment can be performed using steps 3 to 4. The relationship between concepts can be extracted from the table describing the relationship, and a preliminary knowledge system can be established, and evaluation and correction can be performed to generate the final knowledge relationship system.

[0083] For unstructured data, we first use natural language processing tools to perform preprocessing operations such as word segmentation, semantic analysis, dependency syntax analysis, and named entity recognition on the credit-related texts crawled by the information collection crawler. We use steps 3 to 4 to perform entity alignment and entity disambiguation. Then, we use template matching, statistical learning, and other methods to extract important information from the text, mainly including domain concepts, instances, and the relationship between concepts.

[0084] S6. The extracted new SPO triples are subjected to steps such as knowledge fusion and knowledge processing to ensure the reliability and validity of the data before forming the updated part of the knowledge graph, and undergoing manual and internal data learning verification; and the extracted credit information is used to continuously update the knowledge graph through step 5 for dynamic maintenance;

[0085] S7. Provide various credit relationship knowledge graphs as path support for the next step of credit information transmission.

[0086] 3. Social Network Influence Model

[0087] 3.1 Maximizing the influence of social networks

[0088] The definition of the social network influence maximization problem is: given a weighted directed graph G(V,E), it is required to find k nodes in G so that these k nodes have the greatest influence on the graph. This problem is an NP-complete problem.

[0089] Credit subject association network Fig.11 Among them, the factors of node influence can be modeled from the following aspects:

[0090] (1) Measurement method based on the degree centrality of topological networks. The centralization problem of nodes studies the degree to which the nodes in the graph are in the center of the graph, that is:

[0091] a. The closer the node is to the edge of the graph, the lower the centralization index.

[0092] b. The more frequent and densely connected a node is to other nodes, the higher the centralization index is.

[0093] c. Nodes that are close to other nodes with higher indicators generally have higher centralization indicators.

[0094] d. When the weight of an edge in the graph increases, the centralization index of the nodes associated with the edge will not decrease.

[0095] Based on the above key factors, the closeness centrality of a node in the credit indicator association network is defined by the difficulty of the node to reach other nodes in the current community. The closeness centrality reflects the influence of the current node in the community of the entire credit subject network.

[0096] (2) Influence nodes based on the classical theory of weak links. According to the definition of a community in a social network, it is a derived subgraph of an undirected weighted graph. If a generated subgraph of a graph satisfies the following conditions, this subgraph is called a community:

[0097] a. The connections within this subgraph are dense.

[0098] b. The connections between the internal nodes and external nodes of this subgraph are relatively sparse.

[0099] From this, we can see that there are relatively few edges that cross communities in the graph, which is a performance indicator when dividing communities. If a weak connection crosses communities, then although the weight of this edge is low, if this edge is deleted from the graph, there is almost no other path to replace this edge. Such weak connection nodes that cross communities have a heavier impact. They act as a bridge between different groups, and weak nodes cannot be ignored in the process of credit transmission.

[0100] (3) The three-degree influence principle. The three-degree influence principle proposed by Fowler et al. states that a node can not only affect its neighboring nodes (first degree), but also its neighboring nodes (second degree), and even its neighboring nodes (third degree). As long as the nodes are within three degrees of strong connection, they have the possibility of inducing behavior. If the degree exceeds three, the influence between nodes will disappear.

[0101] 3.2 Paths and influence of credit risk transmission

[0102] In order to study the path and influence of credit risk transmission in the credit subject association structure, the N nodes in the association structure are first converted into the adjacency matrix of a directed graph, that is, the node information in the credit subject association model is preprocessed as follows and converted into an adjacency matrix: Suppose there is a directed graph G = (V, E), V represents the set of nodes, E represents the set of edges, each node represents a credit individual or credit organization, and each edge represents the relationship between credit subjects (parent-child relationship, equity relationship, supply relationship, etc.), as shown in Formula 7.

[0103]

[0104] G is an N-order square matrix composed of elements as shown in Formula 1, as shown in Formula 8:

[0105]

[0106] In this N-order square matrix, the weight w of each edge is i,j (0 <w i,j <1) represents the strength of association between credit entities, that is, node i to its adjacent associated node j, The weight of the impact. Each credit subject obtains its credit value based on the credit evaluation model.

[0107] 3.3 Credit transfer model

[0108] When a credit subject i encounters credit risk, its transmission mode has the following situations:

[0109] (1) Single-path continuous transmission, such as Fig.12 The weights of credit risk are multiplied to obtain the compound indirect impact, as shown in Formula 9. This ensures that the credit risk propagated along the path continues to decrease, reflecting the risk buffering of each credit entity. At the same time, it also satisfies the positive correlation between the impact on the end node and the initial impact on the same transmission path.

[0110]

[0111] exist Fig.12 In the figure, when node 1 encounters credit risk, its influence on node 2 is 1*0.3=0.3, its influence on node 3 is 1*0.3*0.4=0.12, and its influence on node 4 is 1*0.3*0.4*0.3=0.036.

[0112] (2) Multi-path transmission. For the superimposed influences a and b, we calculate the impacts produced by their continuous transmission and propose the following composite influence calculation model: 1-(1-a)*(1-b), which ensures that the impact after superposition is not less than any impact before superposition, while at the same time not allowing this impact to exceed 1. 1-a and 1-b are understood as the degree of irrelevance between the two source nodes and the affected nodes respectively. Multiplying them together gives the composite irrelevance, thereby calculating the credit risk probability.

[0113] exist Fig.13 In the figure, assuming that credit risk node i transmits risk through the credit subject association network, its influence on node 3 is 0.4, and its influence on node 5 is 0.24. Then the composite influence of node 6 under multiple paths is calculated as: 1-(1-0.4*0.2)*(1-0.24*0.8)=0.25664.

[0114] Through the above analysis, the calculation formula of the influence of multi-path transmission is obtained, as shown in Formula 10.

[0115]

[0116] (3) Circular transmission. Fig.14 In a complex credit-related network, there are cyclic credit associations between many closely related entities, which are reflected as loops in the network structure. When calculating the influence of multi-path credit risk transmission, the following preprocessing strategy is adopted to avoid the appearance of loops: according to the direction of influence transmission, the successor node adj of the current node i is expanded step by step according to the hierarchical structure. During the expansion process, if the node adj has already appeared in the upper layer, it means that the node has already been affected at a higher level. According to the principle of three-degree influence, the lower the level, the weaker the influence on the node, or even disappears. Therefore, the influence of the current node i on its successor node adj can be ignored. Repeat the above layer-by-layer expansion operation until all nodes are traversed. Applying this strategy can avoid the appearance of loops on the transmission path in advance, rather than deleting the loops after discovering that there are loops on the transmission path.

[0117] In the actual credit risk transmission network, if a credit node has a credit risk, the risk will be transmitted to its associated nodes, and this impact cannot be transmitted back, otherwise a loop will be generated. At this time, the successor node that the node can affect is its subordinate node. At the same time, after processing the loop, the transmission has the following characteristics:

[0118] (1) The impact of credit risk is transmitted in one direction, that is, the upper-level nodes can affect the lower-level nodes, but not vice versa.

[0119] (2) A lower-level node may be affected by one or more upper-level nodes.

[0120] Technical effects: Through multi-source data collection, the comprehensiveness and accuracy of credit information are improved; the construction of knowledge spectrum enhances the relevance and comprehensibility of credit information; social network influence analysis and credit risk transmission model provide new perspectives and methods for credit risk assessment and management.

[0121] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A credit risk assessment and transmission method based on multi-source data and knowledge spectrum, characterized in that: The following steps are involved: S1: Use web crawler technology and event extraction technology to collect credit-related data from financial news websites, social media and other platforms, and clean and convert them; S2: Based on the collected data, a credit knowledge graph is constructed, including entity linking, relationship extraction, and knowledge fusion, to form a credit service network with wide coverage and full categories; S3: Using social network analysis methods, we study the correlation between credit entities and build a model for maximizing social network influence; S4: A probability calculation model for the compound impact of multi-path transmission of credit risk is proposed to simulate the propagation process of credit risk in social networks.

2. The credit risk assessment and transmission method based on multi-source data and knowledge spectrum according to claim 1 is characterized by: The web crawlers are divided into general web crawlers, focused web crawlers, incremental web crawlers and deep web crawlers.

3. The credit risk assessment and transmission method based on multi-source data and knowledge spectrum according to claim 2 is characterized by: The steps of constructing the credit knowledge graph are as follows: (1) Unify the labeling of existing credit entities and use the credit evaluation index system as a standard; (2) Align the labeled entities according to the existing credit knowledge base, connect the existing knowledge with various relationships, determine the SPO triples, and establish the relationship between the entities in the knowledge graph; (3) Perform update operations such as adding, modifying, and deleting relationships on the knowledge graph using the data obtained from real-time data collection; (4) Obtain the knowledge graph of credit subject relationships as support for the credit transmission relationship network.

4. The credit risk assessment and transmission method based on multi-source data and knowledge spectrum according to claim 3 is characterized by: The correlations between the credit entities mainly include investment, supply relationship and debt relationship.

5. The credit risk assessment and transmission method based on multi-source data and knowledge spectrum according to claim 4 is characterized in that: The social network influence analysis step includes a measurement method based on topological network degree centrality, influence node analysis based on the classical theory of weak connections, and the three-degree influence principle.

6. The credit risk assessment and transmission method based on multi-source data and knowledge spectrum according to claim 5 is characterized by: The multi-path transmission of credit risk includes single-path continuous transmission, multi-path transmission, and ring transmission.

Images