A knowledge graph-based multi-modal enterprise credit risk assessment method and device

Abstract

The application provides a kind of multi-modal enterprise credit risk assessment method based on knowledge graph, by constructing enterprise financial knowledge graph, integrating enterprise relationship, industry policy, supply chain information and other data, using multi-modal embedding technology to process entity static attributes and associated information, combined with heterogeneous graph neural network to capture associated information, dynamic spatio-temporal attention mechanism to mine time and space features of time series data, and based on attention weight to identify core risk transmission path, finally fuse graph-level features, dynamic spatio-temporal features and business rules, output structured assessment results. This method effectively integrates multi-modal data, solves the problem of missing association relationship modeling, realizes the deep integration of enterprise multi-source data and the accurate extraction of risk features, and can effectively improve the accuracy and explainability of enterprise credit risk assessment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and in particular to a multi-modal enterprise credit risk assessment method and device based on a knowledge graph, a storage medium and an electronic device. BACKGROUND

[0002] Small and micro enterprises play an important role in economic development, but their credit risk assessment faces multiple challenges. Traditional assessment methods mainly rely on limited structured data such as financial statements, which are difficult to fully depict the complex operating conditions and risk characteristics of enterprises. Traditional methods of processing unstructured data such as invoice images and transaction records only stop at the stage of single-modal independent analysis, lack a unified modeling representation, and result in the omission of key risk information. In addition, the complex relationship network between small and micro enterprises and their shareholders, upstream and downstream of the supply chain, industry policies, and regional economies is difficult to effectively capture through traditional models, making the existing credit assessment unable to reflect the real risk transmission path of the enterprise. At the same time, the existing risk assessment methods generally have a "black box" problem, lack of explainable analysis of the causes of risk, and are difficult to meet the requirements of financial regulation for transparent risk decision-making. With the development of artificial intelligence technology, how to integrate multi-modal data, mine deep correlation characteristics, and generate explainable assessment results has become a key technical problem in the field of small and micro enterprise credit risk assessment. SUMMARY

[0003] The present application aims to provide a multi-modal enterprise credit risk assessment method based on a knowledge graph, which integrates multi-source heterogeneous data, constructs a financial knowledge graph, fuses multi-modal embedding technology, applies heterogeneous graph neural networks and dynamic spatio-temporal attention mechanisms, and generates an explainable reasoning path, thereby improving the accuracy and transparency of small and micro enterprise credit risk assessment.

[0004] In a first aspect, the present application provides a data processing method based on a knowledge graph, comprising: a multi-modal small and micro enterprise credit risk assessment method based on a knowledge graph, characterized by comprising the following steps:

[0005] S1. Collecting multi-source enterprise-related structured data and unstructured data, wherein the unstructured data includes text data, image data and time series data, and preprocessing the data;

[0006] S2. Constructing an enterprise financial knowledge graph, taking structured data as entity attributes, mapping unstructured data to associated relationships through entity linking technology, generating triples and storing them in a database;

[0007] S3. Extracting entity static attribute data and entity association information data of the knowledge graph for embedding representation, and fusing to generate multi-modal initial feature vectors;

[0008] S4. Input the multimodal initial feature vector into the heterogeneous graph neural network to extract the graph-level feature vector;

[0009] S5. Extract the time-series attribute data of entities from the knowledge graph, and generate dynamic spatiotemporal feature vectors through time attention mechanism and spatial graph convolution;

[0010] S6. After fusing the graph-level feature vector and the dynamic spatiotemporal feature vector, input the fusion into the multilayer perceptron (MLP) classifier to output the credit risk probability.

[0011] S7. Identify the nodes and relationship paths in the knowledge graph that have the greatest impact on credit risk based on attention weights, and determine the core paths;

[0012] S8. Output the enterprise credit risk assessment results by combining the core path, risk probability and business rules.

[0013] In one embodiment, step S1 further includes step S1.1: collecting structured and unstructured data from enterprise databases, government open platforms, supply chain platforms, and news websites, wherein the unstructured data includes text data, image data, and time series data.

[0014] Step S1.2: The cleaning of unstructured data specifically includes: processing text data using word segmentation, stop word removal, and part-of-speech tagging techniques; extracting information from image data using OCR technology; aligning time series data by timestamps and using interpolation or forward padding to fill in missing values.

[0015] In one implementation, step S2 further includes step S2.1: defining the entity types of the knowledge graph as including enterprises, industries, policies, supply chain nodes, and related parties;

[0016] Step S2.2: Define the relationship types, including enterprise-industry, enterprise-policy, enterprise-supply chain, enterprise-affiliates, and the relationship between unstructured data and enterprises;

[0017] Step S2.3: Define attributes including company establishment time, registered capital, financial indicators, and policy effective date;

[0018] Step S2.4: Generate triples using entity linking technology;

[0019] Step S2.5: Use a graph database to store the knowledge graph.

[0020] In one implementation, step S3 further includes S3.1 normalizing the numerical structured data in the entity attributes of the knowledge graph and encoding the categorical structured data; S3.2 encoding the text data in the entity association information of the knowledge graph and generating a global embedding vector through pooling operations; extracting visual features from the image data and mapping them to the same-dimensional space.

[0021] S3.3 Generate a multimodal initial feature vector by weighted fusion of the entity attributes and entity association information data after the above processing.

[0022] In one embodiment, step S4 further includes step S4.1: inputting the knowledge graph into a heterogeneous graph neural network (HGNN) and defining a node feature matrix H and a heterogeneous adjacency matrix A;

[0023] Step S4.2: Aggregate neighbor features through a heterogeneous message passing mechanism: Where Wr is the weight matrix of relation r;

[0024] Step S4.3: Use a meta-path-guided aggregation strategy (such as the enterprise-industry-policy path) to capture cross-type dependencies;

[0025] Step S4.4: Generate graph-level feature vector G and node-level feature vectors through multi-layer graph convolution. .

[0026] In one embodiment, step S5 further includes step S5.1: dividing the time series attribute data according to time windows to generate a feature matrix T;

[0027] Step S5.2. Calculate the weights using a time attention mechanism:

[0028]

[0029] Among them, t t W is the feature vector at a given time point. t W is the weight matrix in the time attention mechanism. h Let b be the hidden layer weight matrix. h This is the hidden layer bias vector;

[0030] Step S5.3. Construct a spatial adjacency matrix S based on the enterprise's geographical location, and extract regional risk features through graph convolution;

[0031] Step S5.4. Concatenate the temporal features after time attention weighting with the spatial features to generate a dynamic risk feature vector ST.

[0032] In one embodiment, step S6 further includes step 6.1: aligning and concatenating the graph-level feature vector G output by the heterogeneous graph neural network with the dynamic risk feature vector ST to form the final input feature vector X = [G; ST];

[0033] Step 6.2. Standardize the concatenated feature vector to ensure that the contribution of each dimension of the feature to the classifier is balanced;

[0034] Step 6.3. Input the standardized feature vector into a multilayer perceptron (MLP) classifier, which includes an input layer, two hidden layers and an output layer;

[0035] Step 6.4. After calculation by the multi-layer neural network of the classifier, the credit risk probability P of the micro and small enterprises is output.

[0036] In one implementation, step S7 further includes step S7.1: fusing the graph-level feature G and the dynamic spatiotemporal feature ST and inputting them into an MLP classifier to output the credit risk probability P;

[0037] Step 7.2. Calibrate the prediction results in accordance with business rules;

[0038] Step 7.3. Output a comprehensive result including risk level, critical path, and explanatory report.

[0039] Secondly, the present invention provides a multimodal credit risk assessment system for micro and small enterprises based on knowledge graphs, characterized in that it includes:

[0040] The data acquisition and preprocessing module collects multi-source structured and unstructured data related to the enterprise, including text data, image data, and time series data, and performs preprocessing on the data.

[0041] The knowledge graph construction module is used to build enterprise financial knowledge graphs. It uses structured data as entity attributes and unstructured data as relationships through entity linking technology, generating triples and storing them in the database.

[0042] The multimodal embedding module is used to extract entity static attribute data and entity association information data from the knowledge graph, embed them separately, and fuse them to generate a multimodal initial feature vector;

[0043] The heterogeneous graph neural network module is used to input multimodal initial feature vectors into the heterogeneous graph neural network and extract graph-level feature vectors;

[0044] The dynamic spatiotemporal attention module is used to extract the time-series attribute data of entities from the knowledge graph and generate dynamic spatiotemporal feature vectors through the temporal attention mechanism and spatial graph convolution.

[0045] The feature vector fusion module is used to fuse graph-level feature vectors and dynamic spatiotemporal feature vectors and input them into the multilayer perceptron (MLP) classifier to output the credit risk probability.

[0046] The path determination module is used to identify the nodes and relationship paths in the knowledge graph that have the greatest impact on credit risk based on attention weights, and to determine the core path;

[0047] The risk assessment module is used to combine core paths, risk probabilities, and business rules to output enterprise credit risk assessment results.

[0048] Thirdly, embodiments of the present invention also provide an electronic device, including a processor and a memory, the memory storing computer-executable instructions executable by the processor, the processor executing the computer-executable instructions to implement the method described in any of the first aspects.

[0049] Fourthly, embodiments of the present invention also provide a computer-readable storage medium storing computer-executable instructions, which, when invoked and executed by a processor, cause the processor to implement the method described in any of the first aspects.

[0050] Compared with existing technologies, this invention offers the following advantages through the above technical solutions: This invention proposes a multimodal small business credit assessment method based on knowledge graphs. This model integrates structured data such as enterprise relationships, industry policies, and supply chain information by constructing a financial knowledge graph. It combines multimodal embedding technology to process the knowledge graph. This method captures relationships through heterogeneous graph neural networks and mines temporal data features using a dynamic spatiotemporal attention mechanism, effectively integrating multimodal data to achieve deep fusion of multi-source data and accurate extraction of risk features, thereby improving the accuracy and interpretability of enterprise credit risk assessment.

[0051] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0052] The above and other objects, features, and advantages of the present invention will become more apparent from the more detailed description of the embodiments of the invention in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same parts or steps.

[0053] Figure 1 This is a flowchart illustrating a method provided in an exemplary embodiment of the present invention.

[0054] Figure 2 This is a schematic diagram of an enterprise financial knowledge graph architecture provided by an exemplary embodiment of the present invention.

[0055] Figure 3 This is a diagram of a dynamic spatiotemporal feature generation framework provided by an exemplary embodiment of the present invention.

[0056] Figure 4 This is a schematic diagram of the structure of a device provided in an exemplary embodiment of the present invention. Detailed Implementation

[0057] Hereinafter, exemplary embodiments according to the present invention will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments of the present invention. It should be understood that the present invention is not limited to the exemplary embodiments described herein.

[0058] It should be noted that, unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention.

[0059] Those skilled in the art will understand that the terms "first," "second," etc., in the embodiments of the present invention are only used to distinguish different steps, devices, or modules, and do not represent any specific technical meaning, nor do they indicate a necessary logical order between them.

[0060] It should also be understood that in the embodiments of the present invention, "multiple" can refer to two or more, and "at least one" can refer to one, two or more.

[0061] It should also be understood that any component, data or structure mentioned in the embodiments of the present invention can generally be understood as one or more unless explicitly defined or given contrary instructions in the context.

[0062] To facilitate understanding of this embodiment, a data processing method based on knowledge graphs disclosed in this invention will first be described in detail, as shown in the attached figure. Figure 1 As shown, a multimodal credit risk assessment method for micro and small enterprises based on knowledge graphs includes the following steps:

[0063] S1. Collect multi-source enterprise-related structured and unstructured data, wherein the unstructured data includes text data, image data, and time series data, and preprocess the data;

[0064] The specific implementation of step S1 includes: collecting structured and unstructured data from multiple channels such as enterprise databases, government public platforms like the enterprise credit information publicity system, industry supply chain management systems, news websites, etc.; among them, the structured data covers basic enterprise information, including enterprise name, registered address, unified social credit code, financial statements, tax records, equity structure, etc. The unstructured data includes enterprise official website texts, such as enterprise introductions, product service descriptions, invoice images, transaction flow time series, news reports, etc.

[0065] Furthermore, before constructing the knowledge graph, data preprocessing and text data cleaning are required. Specifically, use Jieba分词 (Jieba word segmentation) etc. to perform word segmentation on the text, remove stop words such as "的" "地" "得" and other words without practical meaning, and perform word tagging to identify key information such as nouns and verbs, so as to initially clean and structurally extract the text data.

[0066] Furthermore, for image data processing, apply optical character recognition (OCR) technology. For example, accurately extract key information from invoice images, such as invoice number, amount, invoice date, etc., and convert it into a structured text format for subsequent analysis.

[0067] Furthermore, for time series data, align transaction flows, credit records and other time series data according to timestamps, and use linear interpolation or forward filling methods to complete missing values to ensure the integrity and consistency of the data.

[0068] S2. Construct an enterprise financial knowledge graph, use the structured data as entity attributes, map the unstructured data as association relationships through entity linking technology, generate triples and store them in the database;

[0069] Among them, the construction of the financial knowledge graph specifically includes:

[0070] Step 2.1: Define the entity types of the knowledge graph to include enterprises, industries, policies, supply chain nodes, related parties; among them, the related parties include shareholders / executives;

[0071] Step 2.2: Define the relationship types to include enterprise - industry, enterprise - policy, enterprise - supply chain, enterprise - related party, and the association relationship between unstructured data and enterprises; in terms of entity types, define "enterprise", "individual", "industry", "policy", "supply chain node" and "unstructured data entity" (such as text, image, time series data point), among which the "individual" entity covers related roles such as shareholders and executives, and among which the relationship types include being affected by policies, upstream and downstream cooperation, equity association / employment relationship, etc.

[0072] Note: "Jieba分词" is a specific term in Chinese for a word segmentation tool, and it's kept in Chinese here as there might not be an exact equivalent in English that conveys the same meaning precisely. If a more appropriate English equivalent needs to be found for this tool name in a more formal context, it should be adjusted accordingly.Step 2.3: Define attributes including company establishment time, registered capital, financial indicators, and policy effective date; among which attribute information includes establishment time, registered capital, employee size, debt-to-equity ratio, current ratio, policy scope of application, and effective date.

[0073] Step 2.4: Generate triples (entity-relationship-entity / attribute) using entity linking technologies such as DBPedia Spotlight;

[0074] Step 2.5: Use the Neo4j graph database to store the knowledge graph, supporting graph query and traversal.

[0075] Specifically, the attribute definitions cover both static features and dynamic time-series information: static attributes of enterprise entities include establishment time, registered capital, employee size, and registered address (including geographical coordinates); dynamic attributes store transaction flows and credit records in time-series format (e.g., quarterly transaction amounts recorded in the format "2023Q1: 5 million yuan"); financial attributes are refined into quantitative indicators such as asset-liability ratio and current ratio; policy entities include attributes such as effective date, applicable industry scope, and support level; text entities store summary content and sentiment; image entities store key information extracted by OCR, such as invoice amount and invoice date, as well as file storage path. In the data processing stage, structured data is directly mapped to entity attributes, while unstructured data achieves semantic linking through technical tools: text uses entity linking technologies such as DBPedia Spotlight to identify enterprise names and establish "description" relationships; images use OCR to extract enterprise identification information and establish "attribution" relationships; and time-series data is embedded into the knowledge graph as dynamic attributes of enterprise entities in the form of "timestamp: value" key-value pairs.

[0076] For example, as shown in the appendix Figure 2 The diagram shows the enterprise financial knowledge graph architecture. Specifically, based on the collected enterprise data, a knowledge graph containing multiple entities and their relationships is created. For example, there is a "belonging industry" relationship between "enterprise A" and "industry B", and a "benefited by policy" relationship between "enterprise A" and "policy C".

[0077] Furthermore, the cleaned data is mapped to the ontology using the entity linking technology DBPedia Spotlight, generating triples (entity-relationship-entity / attribute); the knowledge graph is stored using the Neo4j graph database, which clearly distinguishes entity types through node labels and clearly marks the types of relationship edges, supporting efficient graph queries and complex relationship traversal, enabling subsequent quick queries of the relationship path between any two entities.

[0078] S3. Extract the static attributes and entity association information of the knowledge graph, embed them into representations, and fuse them to generate a multimodal initial feature vector; the multimodal data embedding representation learning specifically includes:

[0079] Furthermore, for the static attributes of entities in the knowledge graph, different embedding methods are adopted according to their data types. For numerical structured data such as asset-liability ratio and operating income in financial indicators, a normalization method is used for standardization to eliminate the influence of different indicator dimensions.

[0080] For categorical structured data such as industry classification and enterprise type, embedding learning methods are used, such as using the Word2Vec algorithm to generate industry embedding vectors, which map category information into a continuous vector space to better capture the potential relationships between categories.

[0081] Furthermore, entity association information primarily originates from unstructured data, and different embedding strategies are employed for different types of unstructured data. Specifically, unstructured data embedding includes, for text data embedding, inputting data such as corporate website text and news articles into a pre-trained language model, outputting context-related word vector sequences. Then, the word vector sequences are aggregated to obtain a global embedding vector representing the overall semantic information of the text.

[0082] For image data embedding, image data such as invoice images, business licenses, and video frame images are input into a convolutional neural network. Visual features of the images are extracted through structures such as convolutional layers and pooling layers. Finally, the extracted features are mapped to a vector space of the same dimension as the text embedding through a fully connected layer, achieving a unified representation of image data and text data.

[0083] For embedding dynamic data such as time series data, long short-term memory networks (LSTM) or Transformer encoders are used to process time series data such as transaction logs. LSTM can effectively capture long-term dependencies in time series data, while Transformer encoders use a self-attention mechanism to compute the importance between time points in parallel, thereby generating embedding vectors that contain temporal context information.

[0084] Further multimodal fusion includes: fusing entity static attribute data, such as structured data embedding vectors (including numerical and categorical data), with unstructured data embedding vectors (including text and images); concatenating structured and unstructured data embedding vectors; or using attention mechanisms to calculate the importance weights of different modalities and performing weighted fusion to obtain the initial multimodal feature vector for micro and small enterprises. For example, if text data is more critical for enterprise credit risk assessment, then higher weights are given to text embedding vectors during the fusion process, so that the fused feature vector can more accurately reflect the comprehensive characteristics of the enterprise.

[0085] Credit risk assessment for micro and small enterprises cannot rely solely on a single type of data. Multimodal feature vectors integrate structured data such as basic enterprise information and financial statements with unstructured data such as enterprise website text and invoice images, providing a comprehensive and integrated reflection of the enterprise's actual situation. For example, financial statements reflect the enterprise's financial health, while enterprise website text may contain information such as the enterprise's development strategy and market positioning, and invoice images can reflect the enterprise's business transaction details. By transforming these different types of data into a unified feature vector, information gaps can be avoided, providing a richer and more accurate basis for subsequent risk assessment.

[0086] S4. Input the multimodal initial feature vector into the heterogeneous graph neural network to extract the graph-level feature vector;

[0087] The specific implementation of step 4 includes: Step 4.1. Inputting the multimodal initial feature vector into the heterogeneous graph neural network (HGNN), wherein the heterogeneous graph neural network includes a node feature matrix H and a heterogeneous adjacency matrix A; the node feature matrix H is defined as H = [h1, h2, ... h... n ], where n is the number of nodes in the knowledge graph, h i Let represent the embedding vector of the i-th node, i = 1, 2, ..., n. These embedding vectors are provided by the multimodal initial feature vectors obtained in step 3, providing an initial feature representation for each node;

[0088] The heterogeneous adjacency matrix A is represented as Where R is the set of all relation types in the knowledge graph. For each relation... , It is the adjacency matrix corresponding to relation r. If there is a relation r between node i and node j, then (A r (i, j) = 1; if the relation does not exist, then (A r (i, j) = 0.

[0089] Step 4.2. For each node, aggregate the features of different types of neighboring nodes using a heterogeneous message passing mechanism. The specific calculation formula is as follows: ,in, Let i be the set of neighbors of node i under relation r. W represents the attention score between node i and node j under relation r, used to measure the influence of neighbor node j on node i; r h is the weight matrix corresponding to relation r, used to perform linear transformation on the features of neighboring nodes to adapt to the feature aggregation requirements under different relations. j This represents the feature vector of neighbor node j, where neighbor node j is the set of neighbors of node i under relation r. A node is defined as one of the nodes in a chain; for enterprise nodes, they are connected to industry nodes under the "industry" relationship and to other enterprise nodes under the "upstream and downstream cooperation" relationship. The above formula aggregates feature information from industry nodes and upstream and downstream enterprise nodes respectively.

[0090] Step 4.3. Employ a meta-path-guided aggregation strategy (e.g., enterprise-industry-policy path) to capture cross-type dependencies. A meta-path is a combination of paths with specific semantic relationships, used to capture long-distance dependencies between cross-type nodes. This step uses a meta-path-guided aggregation strategy to fuse feature information from different relationship paths. The specific implementation and parameters are as follows:

[0091] Based on business needs and the characteristics of the knowledge graph, representative meta-paths are selected. For example, the "Enterprise-Industry-Policy" meta-path represents starting from an enterprise, connecting it to the industry through its industry affiliation, and then connecting it to the policy through the industry's influence on the policy. For each meta-path, features are aggregated sequentially according to the order of relationships along the path. Taking the "Enterprise-Industry-Policy" meta-path as an example, firstly, the enterprise node aggregates features from its industry nodes; then, the industry nodes aggregate features from relevant policy nodes; finally, these features are merged into the enterprise node. The specific fusion method uses a weighted summation approach, and the weights can be set according to the importance of the meta-path.

[0092] Step 4.4. Generate graph-level feature vector G and node-level feature vectors through multi-layer graph convolution. Through multi-layer graph convolution operations, the embedding vectors of nodes are updated layer by layer, enabling each node to aggregate feature information from its neighboring nodes and their neighbors, as well as other higher-order related nodes. This ultimately generates a graph-level feature vector G containing rich higher-order relationships, used to represent the macroscopic features of the entire knowledge graph, and a node-level feature vector H′, used to represent the comprehensive features of each node within the graph. This heterogeneous graph neural network employs an L-layer graph convolutional network, with each layer's update formula using a symmetric normalized graph convolution method. ,in Let l be the node feature matrix of the l-th layer. The degree matrix corresponding to relation r, with diagonal elements ,

[0093] Let σ be the weight matrix of the l-th layer relation r; σ is the activation function.

[0094] After L layers of graph convolution, the final graph-level feature vector G is then processed by global average pooling.

[0095] We obtain the result, where n is the total number of nodes in the knowledge graph, such as the total number of various entities including enterprises, industries, policies, and supply chain nodes. Let G be the feature vector of the i-th node after the convolution of the L-th layer graph, containing the high-order association features of the node and its L-hop neighbors, such as the comprehensive features aggregated by enterprises through paths such as shareholders, supply chains, and industry policies. Then G represents the global risk transmission pattern of the knowledge graph, such as the impact of industry policies on regional enterprise clusters; the node-level feature vector H' = It includes multi-hop neighbor information of nodes, such as the risk transmission characteristics of enterprises through shareholders and related parties and supply chain nodes, which provide high-order correlation characteristics for subsequent dynamic spatiotemporal feature fusion.

[0096] By using multimodal feature vectors as input to a heterogeneous graph neural network (HBR), this method models nodes and relationships in a knowledge graph, capturing complex connections. The multimodal feature vectors provide the initial feature representation for each node in the HBR. Through the message passing mechanism of the HBR, nodes can aggregate feature information from their neighbors, thereby updating their own feature representation. In this process, various information from the multimodal feature vectors can propagate and fuse within the graph structure, further uncovering potential relationships between nodes and generating graph-level and node-level feature vectors containing higher-order relationships.

[0097] S5. Extract time-series attributes from the knowledge graph, including transaction records, credit records, etc., and generate dynamic spatiotemporal feature vectors through time attention mechanism and spatial graph convolution.

[0098] As attached Figure 3 The dynamic spatiotemporal feature generation framework diagram shown includes the following specific implementation steps in step S5:

[0099] Step 5.1. Divide the time series data into time windows (quarterly / annual) and generate a feature matrix T; sort the time series data by timestamp, divide the data into windows based on natural quarters (e.g., 2023Q1, 2023Q2), with a sliding step of one quarter to ensure no window overlap. Generate a time step feature vector t by aggregating the data within each window using the mean. t If data for a certain quarter is missing, interpolation is used: the average data from companies in the same industry and region during the same period is used to fill the gap, ensuring the integrity of the feature matrix T.

[0100] Step 5.2. Calculate the weights using the time attention mechanism:

[0101] Among them, t t W is the feature vector at each time step. t W is the weight matrix in the time attention mechanism. h Let b be the hidden layer weight matrix. h This is the bias vector for the hidden layer. Through this mechanism, the model can automatically learn and focus on key time points that have a significant impact on credit risk, highlighting the feature weights of abnormal time steps, such as the quarter in which defaults occur. For example, if a company experiences large-scale transaction anomalies or credit defaults in a specific quarter, the weight of that time step will be assigned a higher value.

[0102] Step 5.3. Construct a spatial adjacency matrix S based on the geographical location of enterprises, and extract regional risk features through graph convolution; the construction of the adjacency matrix includes a geographical distance metric: Euclidean distance is calculated based on the latitude and longitude of the enterprise's registered location, and enterprises with a distance of less than 10 kilometers are marked as spatially adjacent S(i,j)=1; if two enterprises belong to the same industry and the distance is less than 5 kilometers, the adjacency weight is enhanced to S(i,j)=2, which reflects the risk transmission effect of industrial clusters through industry agglomeration enhancement.

[0103] Spatial graph convolution operation, specifically using GCN graph convolutional network layers to extract region features, ultimately outputs a region risk feature vector X. S ;

[0104] Step 5.4. Concatenate the temporal features (weighted by time attention) with the spatial features to generate the dynamic risk feature vector ST. Apply attention weighting to the time series feature matrix T: Where T is the total number of time steps in the time series, which is the number of time steps after dividing the data into time windows. For example, if the data is divided into quarters for 5 years, then T = 20 (4 quarters per year, 20 quarters in 5 years). The attention weight at time step t is a scalar value. It is calculated using the time attention mechanism in step 5.2 and reflects the importance of the feature vector at time step t in the overall time series, satisfying the following condition: . t t Let be the feature vector of each time step. The feature vector of the t-th time step is the feature representation of each time step obtained after dividing by time window and preprocessing in step 5.1. The comprehensive feature of the time dimension is obtained through the formula. It focuses on risk signals at key time steps (such as cash flow disruptions for two consecutive quarters), and obtains comprehensive time series features after weighting through a time attention mechanism, focusing on key time step information that has a significant impact on credit risk.

[0105] Convolution of spatial graph output X S Global average pooling is used to obtain the overall risk characteristic S of the region. reg Assuming the spatial graph convolution output has features from n enterprise nodes, the global average pooling formula can be expressed as:

[0106] Time characteristics Spatial features Concatenate along dimensions to generate a dynamic spatiotemporal feature vector ST=[T] att ;S reg ];

[0107] Furthermore, time-series data (such as transaction records and credit history) in the multimodal feature vector serves as input to the dynamic spatiotemporal attention mechanism. After being divided into time windows, the time-series data in the multimodal feature vector is used to calculate the importance weights of different time points through the time attention mechanism, thereby highlighting information from key time points. Simultaneously, a spatial adjacency matrix constructed using the enterprise's geographic information, along with other information from the multimodal feature vector, is fused through spatiotemporal feature fusion to generate a dynamic risk feature vector containing spatiotemporal context. This allows for better mining of risk features in both temporal and spatial dimensions.

[0108] Furthermore, the temporal feature mining of the dynamic spatiotemporal attention mechanism also includes: processing the time dimension in the dynamic spatiotemporal attention mechanism, including: dividing the time-series data (such as enterprise historical transaction records, credit records, etc.) according to time windows (such as each quarter or year as a window), and taking the data in each time window as a feature vector of a time step, thereby generating a time series feature matrix T=[t1,t2,...,t t ].

[0109] Furthermore, the spatial dimension processing in the dynamic spatiotemporal attention mechanism also includes: constructing a spatial adjacency matrix S by combining geographical information such as the company's registered location and business location to represent the spatial adjacency relationships between companies. Graph convolution operations are then used to process the spatial adjacency matrix and company feature vectors to extract regional risk features.

[0110] By concatenating and fusing temporal features weighted by time attention with spatial regional features, a dynamic risk feature vector ST containing spatiotemporal context information is generated. For example, for multiple small and micro enterprises in a certain region, if the economic environment of the region has recently deteriorated, it indicates an increase in spatial risk. At the same time, if the enterprises themselves have recently experienced abnormal fluctuations in transaction volume, it indicates an increase in temporal risk. The fused dynamic risk feature vector will comprehensively reflect both of these risk factors.

[0111] S6. After fusing the graph-level feature vector and the dynamic spatiotemporal feature vector, input the fusion into the multilayer perceptron (MLP) classifier to output the credit risk probability.

[0112] Step 6 includes the following steps: Step 6.1: Input the graph-level features G and the dynamic spatiotemporal features ST into the MLP classifier and output the credit risk probability P; specifically, this includes the following steps:

[0113] S6.1. The graph-level feature vector G (dimension d) output by the heterogeneous graph neural network... G The dynamic risk feature vector ST (dimension d) generated by the dynamic spatiotemporal attention mechanism, which represents the overall risk of an enterprise, is a feature vector representing the risk of global interconnectedness within the enterprise. ST The features (including time-space context features) are concatenated, and dimension alignment and concatenation are performed to form the final input feature vector X = [G; ST], with dimension d. G +d ST .

[0114] Z-score normalization is used to standardize the concatenated feature vectors, ensuring a balanced contribution of each dimension of features to the classifier. The multilayer perceptron (MLP) classifier architecture specifically includes...

[0115] Input layer: The number of neurons equals the concatenated feature dimension d G + d ST Directly receive the standardized feature vector .

[0116] Hidden layers: Set 2 fully connected hidden layers (the number of layers can be adjusted according to actual needs):

[0117] First hidden layer: The number of neurons can be set to n1 = 256, and the activation function is ReLU (Rectified Linear Unit), with the following formula:

[0118] in, This is the weight matrix from the input layer to the first hidden layer. This is the bias vector.

[0119] The second hidden layer has 128 neurons (n² = 128). The activation function used is Leaky ReLU to mitigate neuron "death" issues. The formula is: in =0.01, The output after linear transformation. This is the slope parameter.

[0120] Output layer: A single neuron, using the sigmoid activation function to output the credit risk probability. The formula is: in, This is the output vector of the second hidden layer. b1 is the weight matrix from the second hidden layer to the output layer, used to map the hidden layer features to the probability space. b2 is the bias vector, and P represents the probability of default for micro and small enterprises (MSEs), with a higher value indicating higher risk. Specifically, the multimodal feature vectors are processed by a heterogeneous graph neural network and a dynamic spatiotemporal attention mechanism to obtain graph-level feature vectors and dynamic spatiotemporal feature vectors, which are then fused and used as input to the multilayer perceptron (MLP) classifier. The MLP classifier outputs the credit risk probability of MSEs based on these feature vectors. The graph-level feature vector G and the dynamic spatiotemporal feature ST are input to the MLP classifier, and after calculation by the classifier's multilayer neural network, the credit risk probability P of the MSE is output. This probability can be expressed as a default probability (e.g., the likelihood of default within the next year) or a credit rating (e.g., high-risk, medium-risk, and low-risk levels). The feature preprocessing, network architecture parameters, and training methods of the MLP can be set as needed using techniques in this field.

[0121] Step 6.2. Calibrate the forecast results in accordance with business rules; for example, trigger an early warning if the debt-to-asset ratio is >70%.

[0122] S7. Identify the nodes and relationship paths in the knowledge graph that have the greatest impact on credit risk based on attention weights, and determine the core path; specifically, step S7 also includes, S7.1. Integrate the heterogeneous relationship attention weights generated by the heterogeneous graph neural network and the spatiotemporal attention weights generated by the dynamic spatiotemporal attention mechanism to form a comprehensive weight matrix;

[0123] S7.2. Based on the comprehensive weight matrix, identify key nodes whose impact on credit risk exceeds a preset threshold;

[0124] S7.3. Starting from key nodes, use graph traversal algorithm combined with meta-path to extract core risk transmission paths that occur frequently and have large weight values;

[0125] S7.4. Based on business logic, identify the core paths that include defaults by related parties, the impact of industry policies, and the transmission paths of regional risks.

[0126] Furthermore, based on the attention weight matrix generated in steps 4 and 5, a graph traversal algorithm (such as breadth-first search, BFS) is used to identify the nodes and relationship paths in the knowledge graph that have the greatest impact on credit risk, starting from the target enterprise node. For example, if a company has a high credit risk, the BFS algorithm is used to traverse the edges of the knowledge graph and, combined with paths with high attention weights, it can be found that the risk may be due to an operational crisis in its upstream supplier (related through supply chain relationships), which in turn affects the company's raw material supply and production, ultimately leading to an increase in credit risk.

[0127] The extracted paths are filtered, retaining the core paths that appear more frequently than a set threshold (e.g., 50%). For example, if the path "enterprise → upstream of supply chain → industry policy → risk indicators" appears frequently during multiple evaluations and has a significant impact on risk prediction, it is retained as a critical path.

[0128] Furthermore, the visualization explanation generation includes mapping the core reasoning path to a knowledge graph visualization interface and annotating the attribute information of key nodes. For example, the path is displayed in the graphical interface as a node connection; for nodes of highly indebted enterprises, key attributes such as their debt ratio are annotated; for nodes of related party guarantee defaults, detailed information such as the default time and the amount involved is annotated, intuitively showing the path and key factors of risk transmission.

[0129] Furthermore, risk interpretation reports can be generated, showcasing risk transmission paths and key influencing factors through a combination of visual graphs (such as force-directed graphs displaying knowledge graph structures and reasoning paths) and textual explanations. For example, the report not only graphically illustrates the risk paths linking enterprises to supply chains and policies, but also details the impact mechanism of each key node and relationship on credit risk, such as how tightening policies lead to increased costs and strained cash flow for enterprises. This provides financial institutions and other decision-makers with clear and intuitive risk assessment criteria.

[0130] S8. Output the enterprise credit risk assessment results by combining the core path, risk probability and business rules.

[0131] Furthermore, for the identified core risk conduction paths in step S7, basic weights are assigned according to risk types such as related party default and industry policy changes to quantify the impact degree of the core paths. The quantified results will be used for subsequent adjustment of the risk probability. The initial risk probability output in step S6 is corrected through the quantified weights of the core paths to obtain the adjusted risk probability. Specifically, basic weights W can be assigned to the core paths according to risk types. For example, the weight for related party default is 0.8, and the weight for policy impact is 0.7. Combining with the path length decay coefficient L = 1 / (1 + n), where n is the number of relationship edges, the comprehensive weight α = L×W is calculated. The risk probability P output in step S6 is corrected through P' = P×(1 + α). For example, the shareholder default path (n = 1, W = 0.8) corrects P = 0.6 to P' = 0.6×(1 + 0.5×0.8) = 0.84;

[0132] The adjusted risk probability is compared with the dynamic threshold set according to industry and financial indicators, and the risk level is further calibrated in combination with business rules to determine the final risk probability for evaluation. Among them, dynamic decision thresholds are set based on industry classification and regional economic indicators. For example, according to the historical default rate of the industry, the degree of regulatory strictness, and the characteristics of market fluctuations, the industry is divided into three levels of risk categories and basic default probability thresholds are set: For high-risk industries such as real estate, P2P finance, and chemical industry, which have high policy sensitivity, complex capital chains, and significant cyclical risks, the basic threshold T = 0.5;

[0133] Industries such as manufacturing, wholesale and retail, and logistics that are greatly affected by supply chain fluctuations and whose operating conditions can be quantified by financial indicators are determined as medium-risk industries, and the basic threshold T = 0.4;

[0134] Industries such as public utilities, life services, and scientific research and development, which have stable cash flows, strong policy support, and low default rates, are determined as low-risk industries, and the basic threshold T = 0.3. When the corrected risk probability P > T, the corresponding risk level is triggered. Example: P > 0.6 is extremely high risk; 0.5 < P < 0.6 is high risk; other situations are judged according to the medium and low risk level rules. ​​​​Through a rules engine, multi-level conditional judgments are made based on business rules. Risk levels and corresponding handling suggestions are generated based on factors such as risk probability and core path weights, transforming business rules into actual risk assessment decisions. These business rules include basic rules and extended rules. Basic rules further include risk level judgments based on risk probability and corresponding handling suggestions. Core path-based basic rules include: If the related party default path has a high weight, the risk level is directly increased by one level. For example, if the risk level was originally judged as medium risk based on risk probability, it is increased to medium-high risk due to the high weight of the related party default path. The handling suggestion is set to require the company to provide detailed financial information and repayment plans of related parties, and to increase credit monitoring of related parties. If the related party default is serious, the company can be required to repay part of the loan in advance. If the industry policy change path has a high weight, the risk level can be adjusted according to the degree of policy impact. If the policy is favorable, the risk level can be appropriately reduced by one level; if the policy is unfavorable, the risk level is increased by one level. The handling suggestion could be set as follows: when the policy is favorable, consider giving the company certain preferential policies, such as extending the repayment period; when the policy is unfavorable, require the company to adjust its business strategy to cope with the policy change and strengthen the supervision of the company's operating conditions. If the risk path of a region has a high weight, the risk level will be raised by one level. The proposed solution could be to require companies to increase their margin deposits or collateral, strengthen the monitoring of the company's cash flow, and prevent capital flight.

[0137] Extended rule settings include those based on factors such as the company's operating years, market share, and innovation capabilities. For example, an extended rule based on operating years could be set such that if a company has been operating for less than 2 years, its risk level is increased by one level. Recommendations could include requiring the company to provide a detailed business plan and future development strategy, and increasing the assessment of the company's management team. Simultaneously, the credit limit could be reduced, and the repayment period shortened. If a company has been operating for 10 years or more and has been profitable for the past 3 consecutive years, its risk level is reduced by one level. Recommendations could include appropriately increasing the credit limit and offering preferential interest rates. Simultaneously, the frequency of credit assessments could be reduced, ideally to once every six months.

[0138] By setting the above basic and extended rules, multi-level conditional judgments can be made based on business rules. Based on factors such as risk probability and core path, reasonable risk levels and corresponding handling suggestions can be generated, transforming business rules into actual risk assessment decisions. The above settings are merely examples; those skilled in the art can adjust them according to actual circumstances.

[0139] Furthermore, the report outputs a comprehensive report that includes risk level, key influencing factors, and reasoning path. For example, the report clearly indicates that the company's credit risk level is medium risk, and key influencing factors include poor performance of upstream companies in the supply chain and recent industry policy adjustments. It also details the reasoning path and risk transmission mechanism, providing strong support for financial institutions to make financing decisions.

[0140] Furthermore, users can input natural language queries such as "the main reasons for a company's default." The system then performs retrieval and analysis within the knowledge graph and reasoning path based on the input query, returning corresponding interpretable reasoning results. For example, the system can display the paths and nodes related to the reasons for the company's default in a visual interface, and provide a detailed explanation of the reasons in text form, such as related party guarantee defaults leading to joint liability for the company, or funds being diverted, thus increasing the risk of default. This further enhances the convenience and accuracy of decision support.

[0141] To verify the effectiveness and accuracy of the method of this invention, experiments were conducted using a real dataset of micro and small enterprises. A dataset containing basic enterprise information, financial data, unstructured text and image data, and historical credit records was selected and processed and analyzed according to the steps outlined above. By comparing it with traditional credit assessment methods, such as statistical models based solely on financial data, the method of this invention significantly improves the accuracy of credit risk prediction and can more accurately identify high-risk enterprises. Furthermore, by evaluating the generated interpretable reasoning paths, it was found that they can reasonably and clearly explain the causes of credit risk, providing a reliable basis for actual financial decision-making and effectively solving the "black box" problem of traditional models.

[0142] To verify the effectiveness and accuracy of the method of the present invention, the following experiments were conducted:

[0143] This dataset utilizes real-world data from micro and small enterprises in the Yangtze River Delta region. It covers detailed information on over 500 micro and small enterprises, including basic enterprise information, financial data, company website text, news reports, invoice images, business license images, and historical credit records. Structured data accounts for 60%, while unstructured data accounts for 40%. The micro and small enterprises in the dataset span multiple industries, including manufacturing, services, and wholesale and retail, demonstrating good representativeness.

[0144] Traditional credit assessment methods were selected for comparison, including Logistic Regression and Support Vector Machine (SVM). Logistic Regression is a commonly used credit risk assessment method that predicts the probability of corporate default by building a logistic regression model. It is simple to understand and computationally efficient, but it requires strong linearity assumptions about the data and struggles to handle complex nonlinear relationships. Support Vector Machine, on the other hand, classifies data by finding the optimal hyperplane, which has advantages in handling small samples and nonlinear problems, but it is sensitive to the choice of kernel function. Comparing these two methods comprehensively demonstrates the advantages of the proposed method over different types of traditional methods.

[0145] Accuracy, recall, F1 score, and area under the AUC-ROC curve are used as evaluation metrics. Accuracy measures the proportion of samples correctly predicted by the model out of the total samples, reflecting the model's overall predictive ability. Recall measures the proportion of actual positive samples that are correctly predicted as positive samples by the model, reflecting the model's ability to identify positive samples. The F1 score is the harmonic mean of accuracy and recall, taking into account the performance of both. The area under the AUC-ROC curve is used to evaluate the model's performance at different classification thresholds; the closer the value is to 1, the better the model's performance.

[0146] The dataset was divided into training, validation, and test sets at a ratio of 70%, 15%, and 15%, respectively. In the method of this invention, the heterogeneous graph neural network was set to 3 layers, the number of neurons in the first hidden layer of the multilayer perceptron classifier was set to 256, and the number of neurons in the second hidden layer was set to 128. The activation functions used were ReLU and Leaky ReLU, respectively. During training, the Adam optimization algorithm was used with a learning rate of 0.001 and 100 training epochs. For the comparison methods, Logistic Regression and SVM were both set to default parameters.

[0147]

[0148] The multimodal data fusion employed in this invention comprehensively captures the risk characteristics of enterprises; the construction of knowledge graphs and the application of heterogeneous graph neural networks effectively uncover complex relationships between enterprises; and the dynamic spatiotemporal attention mechanism better processes time-series data, accurately capturing the temporal and spatial characteristics of risk changes. These advantages enable the method of this invention to more accurately identify high-risk enterprises in practical financial risk control scenarios, providing financial institutions with more reliable decision-making basis and effectively reducing credit risk.

[0149] In summary, this invention provides a comprehensive, accurate, and interpretable method for assessing the credit risk of micro and small enterprises. By integrating and deeply mining multi-source heterogeneous data, it offers an innovative technical solution for the financial risk control field, possessing promising application prospects and practical value.

[0150] Secondly, as attached Figure 2 As shown, a multimodal enterprise credit risk assessment system 200 based on knowledge graphs is characterized by comprising:

[0151] The data acquisition and preprocessing module 201 acquires multi-source enterprise-related structured and unstructured data, wherein the unstructured data includes text data, image data and time series data, and preprocesses the data.

[0152] The knowledge graph construction module 202 is used to construct an enterprise financial knowledge graph. It uses structured data as entity attributes and unstructured data as relationships through entity linking technology, generating triples and storing them in the database.

[0153] The multimodal embedding module 203 is used to extract the entity static attribute data and entity association information data of the knowledge graph, embed them separately, and fuse them to generate a multimodal initial feature vector;

[0154] The heterogeneous graph neural network module 204 is used to input the multimodal initial feature vector into the heterogeneous graph neural network and extract the graph-level feature vector.

[0155] The dynamic spatiotemporal attention module 205 is used to extract the time-series attribute data of entities from the knowledge graph and generate dynamic spatiotemporal feature vectors through the temporal attention mechanism and spatial graph convolution.

[0156] The feature vector fusion module 206 is used to fuse graph-level feature vectors and dynamic spatiotemporal feature vectors and input them into a multilayer perceptron (MLP) classifier to output credit risk probability.

[0157] The path determination module 207 is used to identify the nodes and relationship paths in the knowledge graph that have the greatest impact on credit risk based on attention weights, and to determine the core path;

[0158] Risk assessment module 208 is used to combine core paths, risk probabilities, and business rules to output enterprise credit risk assessment results.

[0159] Thirdly, embodiments of the present invention also provide an electronic device, including a processor and a memory, the memory storing computer-executable instructions executable by the processor, the processor executing the computer-executable instructions to implement the method described in any of the first aspects.

[0160] The core of this invention is to provide a data processing system, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program and implements the knowledge graph-based data processing method as described above.

[0161] A second aspect of the present invention provides a data processing apparatus, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program and implements the data processing method as described in the foregoing embodiments.

[0162] In addition to the methods and apparatus described above, the embodiments herein are also computer program products that include computer program instructions that, when executed by a processor, cause the processor to perform the steps of the target detection methods according to the various embodiments herein as described in the "Exemplary Methods" section above.

[0163] The basic principles of this disclosure have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in this disclosure are merely examples and not limitations, and should not be considered as essential features of each embodiment of this disclosure. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the scope of this disclosure to the necessity of employing the aforementioned specific details for implementation.

[0164] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For system embodiments, since they largely correspond to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0165] The block diagrams of devices, apparatuses, devices, and systems disclosed herein are merely illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As those skilled in the art will recognize, these devices, apparatuses, devices, and systems can be connected, arranged, and configured in any manner. Words such as “comprising,” “including,” “having,” etc., are open-ended terms meaning “including but not limited to,” and are used interchangeably with them. The terms “or” and “and” as used herein refer to the terms “and / or,” and are used interchangeably with them unless the context clearly indicates otherwise. The term “such as” as used herein refers to the phrase “such as but not limited to,” and is used interchangeably with it.

[0166] The methods and apparatus of this disclosure may be implemented in many ways. For example, they may be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware. The above-described order of steps for the methods is for illustrative purposes only, and the steps of the methods of this disclosure are not limited to the order specifically described above unless otherwise specifically stated. Furthermore, in some embodiments, this disclosure may also be implemented as a program recorded on a recording medium, the program including machine-readable instructions for implementing the methods according to this disclosure. Thus, this disclosure also covers recording media storing programs for performing the methods according to this disclosure.

[0167] It should also be noted that in the apparatus, devices, and methods of this disclosure, the components or steps are decomposable and / or recombinable. Such decomposition and / or recombination should be considered equivalent to the present disclosure. The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use this disclosure. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein can be applied to other aspects without departing from the scope of this disclosure. Therefore, this disclosure is not intended to be limited to the aspects shown herein, but rather to be carried out within the widest scope consistent with the principles and novel features disclosed herein.

[0168] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of this disclosure to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations therein.

Claims

1. A multimodal enterprise credit risk assessment method based on knowledge graphs, characterized in that, Includes the following steps: S1. Collect multi-source enterprise-related structured and unstructured data, wherein the unstructured data includes text data, image data, and time series data, and preprocess the data; S2. Construct an enterprise financial knowledge graph, using structured data as entity attributes and unstructured data as relationships through entity linking technology, generating triples and storing them in the database; S3. Extract the static attribute data and entity association information data of the knowledge graph, embed them separately, and fuse them to generate a multimodal initial feature vector; S4. Input the multimodal initial feature vector into the heterogeneous graph neural network to extract the graph-level feature vector; Step S4 further includes: Step S4.

1. Input the multimodal initial feature vector, arrange the multimodal initial feature vector in node order to construct the node feature matrix H of the heterogeneous graph neural network, where each row corresponds to the multimodal initial feature of a node; the heterogeneous adjacency matrix A of the heterogeneous graph neural network is used to represent the connection relationship between different types of nodes, and different types of connections correspond to different elements in the matrix; Step S4.

2. Aggregate the neighbor feature vectors using the node feature matrix H, heterogeneous adjacency matrix A, and multimodal initial feature vectors obtained in step S4.1 through a heterogeneous message passing mechanism: Among them, W r Let r be the weight matrix of relation r. Let be the neighbor feature vectors aggregated by node i under relation r. Let i be the set of neighbors of node i under relation r. The attention score between node i and its neighbor node j under relation r is calculated using node features and relation weights, and is used to measure the importance of neighbor nodes to the current node. j Let be the initial multimodal initial feature vector of neighbor node j; Step S4.

3. For the neighbor feature vector, a meta-path-guided aggregation strategy is adopted to capture cross-type dependencies. A predefined meta-path is used to guide the information aggregation process, thereby capturing node features containing cross-type dependency information. Step S4.

4. Input it into a multi-layer graph convolutional network to generate graph-level feature vector G and node-level feature vector H'. The graph-level feature vector G can reflect the global feature information of the entire graph, while the node-level feature vector contains the local feature information of each node. S5. Extract the time-series attribute data of entities from the knowledge graph, and generate dynamic spatiotemporal feature vectors through time attention mechanism and spatial graph convolution; S6. After fusing the graph-level feature vector and the dynamic spatiotemporal feature vector, input the fusion into the multilayer perceptron (MLP) classifier to output the credit risk probability. S7. Identify the nodes and relationship paths in the knowledge graph that have the greatest impact on credit risk based on attention weights, and determine the core paths; S8. Output the enterprise credit risk assessment results by combining the core path, risk probability and business rules.

2. The method according to claim 1, characterized in that: Step S1 further includes: Step S1.

1. Collect structured and unstructured data from enterprise databases, enterprise information disclosure platforms, supply chain platforms, and news websites; Step S1.

2. The data preprocessing includes cleaning unstructured data: processing text data using word segmentation, stop word removal, and part-of-speech tagging techniques; extracting information from image data using OCR technology; aligning time series data by timestamps and filling in missing values ​​using interpolation or forward padding; and standardizing structured data.

3. The method according to claim 1, characterized in that: Step S2 further includes: Step S2.

1. Define the entity types of the enterprise financial knowledge graph, including enterprises, industries, policies, supply chain nodes, and related parties; Step S2.

2. Define relationship types including enterprise-industry, enterprise-policy, enterprise-supply chain, enterprise-affiliates, and the relationship between unstructured data and enterprises; Step S2.

3. Define attributes including company establishment time, registered capital, financial indicators, and policy effective date; Step S2.

4. Generate triples for the knowledge graph using entity linking technology; Step S2.

5. Use a graph database to store the knowledge graph.

4. The method according to claim 1, characterized in that, Step S3 further includes: S3.

1. Normalize the numerical structured data in the entity attributes of the knowledge graph, and encode the categorical structured data; S3.

2. Encode the text data in the entity association information of the knowledge graph and generate a global embedding vector through pooling operation; extract visual features from the image data and map them to the same dimension space; S3.

3. Generate a multimodal initial feature vector by weighted fusion of the entity attributes and entity association information data after the above processing.

5. The method according to claim 1, characterized in that: Step S5 further includes: Step S5.

1. Divide the time series attribute data according to time windows to generate a feature matrix T; Step S5.

2. Based on the feature matrix T, calculate the weight of time point t using the time attention mechanism: in, The attention weights at time point t are used to dynamically adjust the importance of different time steps. t W is the feature vector at a given time point. t W is the weight matrix in the time attention mechanism. h Let b be the hidden layer weight matrix. h This is the hidden layer bias vector; Step S5.

3. Construct a spatial adjacency matrix S based on the enterprise's geographical location, and extract regional risk features through graph convolution; Step S5.

4. Concatenate the temporal features after time attention weighting with the spatial features to generate a dynamic risk feature vector ST.

6. The method according to claim 1, characterized in that: Step S6 further includes: Step 6.

1. Align and concatenate the graph-level feature vector G output by the heterogeneous graph neural network with the dynamic risk feature vector ST to form the final input feature vector X = [G; ST]; Step 6.

2. Standardize the concatenated feature vector to ensure that the contribution of each dimension of the feature to the classifier is balanced; Step 6.

3. Input the standardized feature vector into a multilayer perceptron (MLP) classifier, which includes an input layer, two hidden layers, and an output layer; Step 6.

4. After calculation by the multi-layer neural network of the classifier, the credit risk probability P of the micro and small enterprises is output.

7. The method according to claim 1, characterized in that, Step S7 further includes: S7.

1. Integrate the heterogeneous relation attention weights generated by the heterogeneous graph neural network and the spatiotemporal attention weights generated by the dynamic spatiotemporal attention mechanism to form a comprehensive weight matrix; S7.

2. Based on the comprehensive weight matrix, identify key nodes whose impact on credit risk exceeds a preset threshold; S7.

3. Starting from key nodes, use graph traversal algorithm combined with meta-path to extract core risk transmission paths that occur frequently and have large weight values; S7.

4. Based on business logic, identify the core paths of corporate related-party defaults, industry policy impacts, and regional risk transmission.

8. A multimodal enterprise credit risk assessment system based on knowledge graphs, characterized in that, include: The data acquisition and preprocessing module is used to acquire multi-source enterprise-related structured and unstructured data, wherein the unstructured data includes text data, image data and time series data, and to preprocess the data. The knowledge graph construction module is used to build enterprise financial knowledge graphs. It uses structured data as entity attributes and unstructured data as relationships through entity linking technology, generating triples and storing them in the database. The multimodal embedding module is used to extract entity static attribute data and entity association information data from the knowledge graph, embed them separately, and fuse them to generate a multimodal initial feature vector; The heterogeneous graph neural network module is used to input multimodal initial feature vectors into the heterogeneous graph neural network and extract graph-level feature vectors; The extraction of graph-level feature vectors further includes: Step S4 further includes: Step S4.

1. Input the multimodal initial feature vector, arrange the multimodal initial feature vector in node order to construct the node feature matrix H of the heterogeneous graph neural network, where each row corresponds to the multimodal initial feature of a node; the heterogeneous adjacency matrix A of the heterogeneous graph neural network is used to represent the connection relationship between different types of nodes, and different types of connections correspond to different elements in the matrix; Step S4.

2. Aggregate the neighbor feature vectors using the node feature matrix H, heterogeneous adjacency matrix A, and multimodal initial feature vectors obtained in step S4.1 through a heterogeneous message passing mechanism: Among them, W r Let r be the weight matrix of relation r. Let be the neighbor feature vectors aggregated by node i under relation r. Let i be the set of neighbors of node i under relation r. The attention score between node i and its neighbor node j under relation r is calculated using node features and relation weights, and is used to measure the importance of neighbor nodes to the current node. j Let be the initial multimodal initial feature vector of neighbor node j; Step S4.

3. For the neighbor feature vector, a meta-path-guided aggregation strategy is adopted to capture cross-type dependencies. A predefined meta-path is used to guide the information aggregation process, thereby capturing node features containing cross-type dependency information. Step S4.

4. Input it into a multi-layer graph convolutional network to generate graph-level feature vector G and node-level feature vector H'. The graph-level feature vector G can reflect the global feature information of the entire graph, while the node-level feature vector contains the local feature information of each node. The dynamic spatiotemporal attention module is used to extract the time-series attribute data of entities from the knowledge graph and generate dynamic spatiotemporal feature vectors through the temporal attention mechanism and spatial graph convolution. The feature vector fusion module is used to fuse graph-level feature vectors and dynamic spatiotemporal feature vectors and input them into the multilayer perceptron (MLP) classifier to output the credit risk probability. The path determination module is used to identify the nodes and relationship paths in the knowledge graph that have the greatest impact on credit risk based on attention weights, and to determine the core path; The risk assessment module is used to combine core paths, risk probabilities, and business rules to output enterprise credit risk assessment results.

9. An electronic device, characterized in that, The electronic device includes: a memory and a processor, the memory and the processor being coupled; the memory stores program instructions that, when executed by the processor, cause the electronic device to perform the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, Includes a computer program that, when run on an electronic device, causes the electronic device to perform the method as described in any one of claims 1 to 7.

Images