Multi-modal enterprise credit risk assessment method and device based on knowledge graph

Abstract

The invention provides a multi-modal enterprise credit risk assessment method based on a knowledge graph, which integrates data such as enterprise relationships, industry policies and supply chain information by constructing an enterprise financial knowledge graph, processes entity static attributes and associated information by using a multi-modal embedding technology, captures the associated information in combination with a heterogeneous graph neural network, and evaluates the credit risk of an enterprise. And the dynamic space-time attention mechanism mines time and space features of the time series data, identifies a core risk conduction path based on an attention weight, and finally fuses graph-level features, dynamic space-time features and business rules to output a structured evaluation result. According to the method, multi-modal data is effectively integrated, the problem of incidence relation modeling deficiency is solved, deep fusion of enterprise multi-source data and accurate extraction of risk features are realized, and the accuracy and interpretability of enterprise credit risk assessment can be effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a multimodal enterprise credit risk assessment method and device based on a knowledge graph, as well as a storage medium and an electronic device. Background Art

[0002] Small and micro enterprises (SMEs) play a vital role in economic development, but their credit risk assessment faces multiple challenges. Traditional assessment methods primarily rely on limited structured data, such as financial statements, making it difficult to fully capture the complex operating conditions and risk profiles of enterprises. Traditional methods for processing unstructured data, such as invoice images and transaction flows, remain limited to single-modal, independent analysis, lacking a unified modeling representation, resulting in the omission of key risk information. Furthermore, traditional models struggle to effectively capture the complex network of relationships between SMEs and their shareholders, upstream and downstream supply chains, industry policies, and regional economies, rendering existing credit assessments incapable of reflecting the true risk transmission pathways of enterprises. Furthermore, existing risk assessment methods suffer from a widespread "black box" problem, lacking interpretable analysis of risk drivers, and thus failing to meet financial regulatory requirements for transparent risk decision-making. With the advancement of artificial intelligence (AI), integrating multimodal data, mining deep correlations, and generating interpretable assessment results have become key technical challenges in SME credit risk assessment. Summary of the Invention

[0003] The present invention aims to provide a multimodal enterprise credit risk assessment method based on knowledge graph, which improves the accuracy and transparency of credit risk assessment of small and micro enterprises by integrating multi-source heterogeneous data, constructing a financial knowledge graph, fusing multimodal embedding technology, applying heterogeneous graph neural networks and dynamic spatiotemporal attention mechanisms, and generating explainable reasoning paths.

[0004] In a first aspect, an embodiment of the present invention provides a data processing method based on a knowledge graph, including: a multimodal small and micro enterprise credit risk assessment method based on a knowledge graph, characterized by comprising the following steps:

[0005] S1. Collect enterprise-related structured data and unstructured data from multiple sources, where the unstructured data includes text data, image data, and time series data, and preprocess the data;

[0006] S2. Build an enterprise financial knowledge graph, treating structured data as entity attributes and mapping unstructured data into relationships using entity linking technology. Triples are generated and stored in a database.

[0007] S3. 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;

[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 temporal attention mechanism and spatial graph convolution;

[0010] S6. Fusion of the graph-level feature vector and the dynamic spatiotemporal feature vector is input into a multi-layer 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. Combine core paths, risk probabilities, and business rules to output enterprise credit risk assessment results.

[0013] In one embodiment, the step S1 further includes step S1.1: collecting structured data and unstructured data from enterprise databases, government public 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 unstructured data is cleaned, specifically including: 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 filling.

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

[0016] Step S2.2: Define relationship types including enterprise-industry, enterprise-policy, enterprise-supply chain, enterprise-related party, 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 time;

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

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

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

[0021] S3.3 generates 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, the step S4 further includes step S4.1: inputting the knowledge graph into a heterogeneous graph neural network (HGNN) to define a node feature matrix H and a heterogeneous adjacency matrix A;

[0023] Step S4.2: Aggregate neighbor features via heterogeneous message passing mechanism:

[0024] Where Wr is the weight matrix of relation r;

[0025] Step S4.3: Adopt a meta-path-guided aggregation strategy (e.g., enterprise-industry-policy path) to capture cross-type dependencies;

[0026] Step S4.4: Generate graph-level feature vector G and node-level feature vector H' through multi-layer graph convolution.

[0027] In one embodiment, the step S5 further includes step S5.1. dividing the time series attribute data into time windows to generate a feature matrix T;

[0028] Step S5.2. Calculate weights through the temporal attention mechanism:

[0029] α t =softmax(W t Tanh(W t t t +b h ))

[0030] Among them, t t is the time point feature vector, W t is the weight matrix in the temporal attention mechanism, W h is the hidden layer weight matrix, b h is the hidden layer bias vector;

[0031] Step S5.3. Construct a spatial adjacency matrix S based on the enterprise's geographic location and extract regional risk features through graph convolution.

[0032] Step S5.4. Concatenate the temporal features weighted by temporal attention with the spatial features to generate a dynamic risk feature vector ST.

[0033] In one embodiment, the step S6 further includes step 6.1. dimensional alignment and concatenation of the graph-level feature vector G output by the heterogeneous graph neural network and the dynamic risk feature vector ST to form a final input feature vector X=[G;ST];

[0034] Step 6.2. Normalize the concatenated feature vector to ensure that the contributions of features from each dimension to the classifier are balanced.

[0035] Step 6.3. Input the normalized feature vector into a multi-layer perceptron (MLP) classifier, which includes an input layer, two hidden layers, and an output layer.

[0036] Step 6.4. After calculation by the multi-layer neural network of the classifier, the credit risk probability P of small and micro enterprises is output.

[0037] In one embodiment, the step S7 further includes step S7.1: fusing the graph-level feature G and the dynamic spatiotemporal feature ST and inputting the result into an MLP classifier to output a credit risk probability P;

[0038] Step 7.2. Calibrate the prediction results based on business rules.

[0039] Step 7.3. Output comprehensive results including risk level, critical path, and explanation report.

[0040] In a second aspect, the present invention provides a multimodal small and micro enterprise credit risk assessment system based on a knowledge graph, characterized by including:

[0041] The data collection and preprocessing module collects enterprise-related structured data and unstructured data from multiple sources, where the unstructured data includes text data, image data, and time series data, and preprocesses the data;

[0042] The knowledge graph construction module is used to build an enterprise financial knowledge graph, using structured data as entity attributes and unstructured data as association relationships through entity linking technology to generate triples and store them in the database;

[0043] 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;

[0044] Heterogeneous graph neural network module, used to input multimodal initial feature vectors into the heterogeneous graph neural network and extract graph-level feature vectors;

[0045] Dynamic spatiotemporal attention module, which is used to extract time series attribute data of entities from the knowledge graph and generate dynamic spatiotemporal feature vectors through temporal attention mechanism and spatial graph convolution;

[0046] The feature vector fusion module is used to fuse the graph-level feature vector and the dynamic spatiotemporal feature vector and input them into the multi-layer perceptron (MLP) classifier to output the credit risk probability;

[0047] The path determination module is used to identify the nodes and relationship paths that have the greatest impact on credit risk in the knowledge graph based on attention weights, and determine the core path;

[0048] The risk assessment module is used to output enterprise credit risk assessment results by combining core paths, risk probabilities and business rules.

[0049] In a third aspect, an embodiment of the present invention further provides an electronic device, comprising a processor and a memory, wherein the memory stores computer-executable instructions that can be executed by the processor, and the processor executes the computer-executable instructions to implement any one of the methods provided in the first aspect.

[0050] In a fourth aspect, an embodiment of the present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium stores computer-executable instructions. When the computer-executable instructions are called and executed by a processor, the computer-executable instructions prompt the processor to implement any one of the methods provided in the first aspect.

[0051] Through the above technical solution, compared with the existing technology, the present invention has the following beneficial effects: The present invention proposes a multimodal small business credit assessment method based on knowledge graphs. This model integrates structured data such as corporate relationships, industry policies, and supply chain information by constructing a financial knowledge graph, and combines multimodal embedding technology to process the knowledge graph. This method captures associations through heterogeneous graph neural networks and mines time series data features using a dynamic spatiotemporal attention mechanism. This method effectively integrates multimodal data, achieves deep fusion of multi-source data, and accurately extracts risk characteristics, thereby improving the accuracy and interpretability of corporate credit risk assessments.

[0052] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] The above and other objects, features, and advantages of the present invention will become more apparent through a more detailed description of the embodiments of the present invention in conjunction with the accompanying drawings. The accompanying drawings are provided to provide a further understanding of the embodiments of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and are not intended to limit the present invention. In the drawings, the same reference numerals generally represent the same components or steps.

[0054] Figure 1 It is a flowchart of a method provided by an exemplary embodiment of the present invention.

[0055] Figure 2 It is a schematic diagram of the enterprise financial knowledge graph architecture provided by an exemplary embodiment of the present invention.

[0056] Figure 3 It is a dynamic spatiotemporal feature generation framework diagram provided by an exemplary embodiment of the present invention.

[0057] Figure 4 It is a schematic structural diagram of a device provided by an exemplary embodiment of the present invention. DETAILED DESCRIPTION

[0058] Below, the exemplary embodiments according to the present invention will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments of the present invention, and it should be understood that the present invention is not limited to the exemplary embodiments described herein.

[0059] It should be noted that the relative arrangement of components and steps, the numerical expressions and numerical values set forth in these embodiments do not limit the scope of the present invention unless specifically stated otherwise.

[0060] Those skilled in the art will understand that the terms "first" and "second" in the embodiments of the present invention are only used to distinguish different steps, devices or modules, and neither represent any specific technical meaning nor indicate the necessary logical order between them.

[0061] It should also be understood that, in the embodiments of the present invention, “a plurality of” may refer to two or more than two, and “at least one” may refer to one, two or more than two.

[0062] 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 limited or otherwise indicated in the context.

[0063] To facilitate understanding of this embodiment, a data processing method based on knowledge graph disclosed in an embodiment of the present invention is first introduced in detail, as shown in the attached figure. Figure 1 As shown in FIG, a multimodal credit risk assessment method for small and micro enterprises based on knowledge graph includes the following steps:

[0064] S1. Collect enterprise-related structured data and unstructured data from multiple sources, where the unstructured data includes text data, image data, and time series data, and preprocess the data;

[0065] 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.

[0066] Furthermore, before constructing the knowledge graph, data preprocessing and text data cleaning are required. Specifically, use Jieba分词 and other tools to perform word segmentation on the text, remove stop words such as "的", "地", "得" and other words without practical meaning, and perform词性标注 to identify key information such as nouns and verbs, so as to perform preliminary cleaning and structured extraction of the text data.

[0067] 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.

[0068] Furthermore, for time series data, align transaction flow, credit records and other time series data according to time stamps, and use linear interpolation or forward filling methods to complete missing values to ensure the integrity and consistency of the data.

[0069] 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;

[0070] Among them, the construction of the financial knowledge graph specifically includes:

[0071] Step 2.1: Define the entity types of the knowledge graph including enterprises, industries, policies, supply chain nodes, related parties; among them, the related parties include shareholders / executives;

[0072] Step 2.2: Define the relationship types including 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), where the "individual" entity covers related roles such as shareholders and executives, and the relationship types include being affected by policies, upstream and downstream cooperation, equity association / employment relationship, etc.

[0073] Step 2.3: Define attributes including company establishment time, registered capital, financial indicators, and policy effective time; attribute information includes establishment time, registered capital, employee size, debt-to-asset ratio, current ratio, policy scope of application, effective time, etc.

[0074] Step 2.4: Generate triples (entity-relationship-entity / attribute) through entity linking technology such as DBPedia Spotlight;

[0075] Step 2.5: Use the Neo4j graph database to store the knowledge graph and support graph query and traversal.

[0076] Specifically, attribute definitions cover both static features and dynamic time-series information: static attributes of corporate entities include founding date, registered capital, employee size, and registered address (including geographic coordinates). Dynamic attributes store transaction flows and credit records in a time series format (e.g., quarterly transaction amounts are recorded in the format of "2023Q1: 5 million yuan"). Financial attributes are broken down into quantitative indicators such as debt-to-asset ratio and current ratio. Policy entities include attributes such as effective date, applicable industry scope, and support intensity. Text entities store summary content and sentiment, while image entities store key information extracted through optical character recognition (OCR) such as invoice amount and invoice date, as well as the file storage path. In the data processing phase, structured data is directly mapped to entity attributes, while unstructured data is semantically linked through technical tools. Text uses entity linking technologies such as DBPedia Spotlight to identify company names and establish a "description" relationship. Images use OCR to extract company identification information and establish an "attribution" relationship. Time-series data is embedded in the knowledge graph as dynamic attributes of corporate entities in the form of "timestamp: value" key-value pairs.

[0077] For example, as attached Figure 2 The enterprise financial knowledge graph architecture diagram shown in the figure specifically creates a knowledge graph containing multiple entities and their relationships based on the collected enterprise data. For example, there is a "belonging industry" relationship between "Enterprise A" and "Industry B", and there is a "affected by policy" relationship between "Enterprise A" and "Policy C".

[0078] Furthermore, the cleaned data is mapped to the ontology through the entity linking technology DBPedia Spotlight to generate triples (entity-relationship-entity / attribute); the Neo4j graph database is used to store the knowledge graph, and the entity types are clearly distinguished through node labels. The relationship edge annotations clearly indicate the type, supporting efficient graph query and complex relationship traversal, so that the relationship path between any two entities can be quickly queried subsequently.

[0079] S3. Extract entity static attributes and entity association information from the knowledge graph, embed them separately, and fuse them to generate a multimodal initial feature vector. The multimodal data embedding representation learning specifically includes:

[0080] 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 debt-to-asset ratio and operating income in financial indicators, normalization methods are used to eliminate the influence of different indicator dimensions.

[0081] 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, mapping category information into a continuous vector space to better capture the potential associations between categories.

[0082] Furthermore, entity association information primarily comes from unstructured data, and different embedding strategies are employed for different types of unstructured data. Specifically, for text data, unstructured data embedding involves inputting data such as corporate website text and news into a pre-trained language model, which outputs a sequence of contextually relevant word vectors. This sequence of word vectors is then aggregated to produce a global embedding vector that represents the overall semantic information of the text.

[0083] For image data embedding, we input image data such as invoice images, business licenses, and video frames into a convolutional neural network. The convolutional and pooling layers extract the image's visual features. Finally, a fully connected layer maps the extracted features to a vector space with the same dimensions as the text embedding, achieving a unified representation of image and text data.

[0084] For embedding dynamic data such as time series, we process transaction flow data using a long short-term memory network (LSTM) or a Transformer encoder. LSTM effectively captures long-term dependencies in time series data, while the Transformer encoder uses a self-attention mechanism to parallelize the importance of time points, generating an embedding vector that contains temporal context.

[0085] Further multimodal fusion specifically 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 data embedding vectors with unstructured data embedding vectors; or utilizing an attention mechanism to calculate the importance weights of data from different modalities and performing weighted fusion based on these weights to obtain the initial multimodal feature vectors for small and micro enterprises. For example, if text data is more critical to corporate credit risk assessment, a higher weight will be assigned to the text embedding vector during the fusion process, so that the fused feature vector can more accurately reflect the comprehensive characteristics of the enterprise.

[0086] Credit risk assessments for small and micro enterprises cannot rely solely on a single type of data. Multimodal feature vectors integrate structured data such as basic company information and financial statements with unstructured data such as official website text and invoice images, providing a comprehensive and integrated picture of a company's actual condition. For example, financial statements can reveal a company's financial health, while official website text may include information such as its development strategy and market positioning, and invoice images can reflect its business transactions. By converting these different types of data into a unified feature vector, information loss can be avoided, providing a richer and more accurate basis for subsequent risk assessments.

[0087] S4. Input the multimodal initial feature vector into the heterogeneous graph neural network to extract the graph-level feature vector;

[0088] The specific implementation of step 4 includes: Step 4.1. Input the multimodal initial feature vector into the heterogeneous graph neural network (HGNN), where 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 Represents the embedding vector of the i-th node, i = 1, 2, ... n. These embedding vectors are provided by the multimodal initial feature vector obtained in step 3, providing an initial feature representation for each node;

[0089] The heterogeneous adjacency matrix A is represented by A = {Ar} r∈R , where R is the set of all relation types in the knowledge graph. For each relation r∈R, A r Is the adjacency matrix corresponding to the relationship r. If there is a relationship r between node i and node j, then (A r (i,j)=1; if the relationship does not exist, then (A r (i,j)=0.

[0090] Step 4.2. For each node, aggregate the features of different types of neighboring nodes through the heterogeneous message passing mechanism. The specific calculation formula is:

[0091] in, is the neighbor set of node i under relation r, represents the attention score between node i and node j under relationship r, which is used to measure the influence of neighbor node j on node i; W r is the weight matrix corresponding to the relationship r, which is used to linearly transform the neighbor node features to adapt to the feature aggregation requirements under different relationships, h j Represents the eigenvector of neighbor node j, where neighbor node j is the neighbor set of node i under relationship r A node in the data set; for an enterprise node, it is connected to the industry node under the "industry" relationship and to other enterprise nodes under the "upstream and downstream cooperation" relationship. The above formula aggregates the feature information from the industry node and the upstream and downstream enterprise nodes respectively.

[0092] Step 4.3. Use a meta-path-guided aggregation strategy (e.g., enterprise-industry-policy path) to capture cross-type dependencies. A meta-path is a composite path of specific semantic relationships, used to capture long-distance dependencies between nodes across different types. This step uses a meta-path-guided aggregation strategy to fuse feature information from different relationship paths. The specific implementation method and parameters are as follows:

[0093] 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 the enterprise, linking to the industry through the industry relationship, and then linking to the policy through the industry's policy-affected relationship. For each meta-path, feature aggregation is performed in sequence according to the order of the relationships on the path. Taking the "enterprise-industry-policy" meta-path as an example, the enterprise node first aggregates the features of the industry node to which it belongs, then the industry node aggregates the features of the relevant policy nodes, and finally these features are fused into the enterprise node. The specific fusion method uses a weighted summation method, and the weight can be set according to the importance of the meta-path.

[0094] Step 4.4. Generate graph-level feature vector G and node-level feature vector H' through multi-layer graph convolution. Through multi-layer graph convolution operations, the embedding vector of the node is updated layer by layer, so that each node can aggregate the feature information of its neighbor nodes and the neighbors of the neighbor nodes and other high-order related nodes, and finally generate a graph-level feature vector G containing rich high-order association relationships, which is used to represent the macro characteristics of the entire knowledge graph, and a node-level feature vector H', which is used to represent the comprehensive characteristics of each node in the graph. This heterogeneous graph neural network adopts an L-layer graph convolution network, and the update formula of each layer adopts a symmetric normalized graph convolution method. Among them H l is the node feature matrix of the lth layer, is the degree matrix corresponding to relation r, and the diagonal elements

[0095] is the weight matrix of the l-th layer relation r; σ is the activation function.

[0096] After L layers of graph convolution, the final generated graph-level feature vector G is pooled by global average pooling.

[0097] Where n is the total number of nodes in the knowledge graph, such as the total number of various entities such as enterprises, industries, policies, and supply chain nodes. is the feature vector of the i-th node after the L-th layer of graph convolution, which contains the high-order correlation features of the node and its L-hop neighbors, such as the comprehensive features of the enterprise aggregation through shareholders, supply chains, industry policies, etc. Then G represents the global risk transmission model of the knowledge graph, such as the impact of industry policies on regional enterprise clusters; the node-level feature vector H'=H l ,Including the multi-hop neighbor information of the node, such as the risk transmission ,features of the enterprise through shareholder related parties and supply chain nodes, ,provides high-order correlation features for the subsequent dynamic spatiotemporal feature fusion.

[0098] By using multimodal feature vectors as input to a heterogeneous graph neural network, the network is used to model nodes and relationships in the knowledge graph and capture complex associations. The multimodal feature vectors provide the heterogeneous graph neural network with an initial feature representation of each node. Through the heterogeneous graph neural network's message passing mechanism, nodes can aggregate feature information from their neighbors to update their own feature representations. In this process, the various information in the multimodal feature vectors can be propagated and fused within the graph structure, further exploring potential relationships between nodes and generating graph-level and node-level feature vectors that incorporate high-order associations.

[0099] S5. Extract time series attributes from the knowledge graph, including transaction flows and credit records, and generate dynamic spatiotemporal feature vectors using a temporal attention mechanism and spatial graph convolution.

[0100] As attached Figure 3 The dynamic spatiotemporal feature generation framework shown in FIG. 5 includes the following steps:

[0101] Step 5.1. Divide the time series data by time window (quarter / year) to generate the feature matrix T; sort the time series data by timestamp, divide the windows into natural quarters (such as 2023Q1, 2023Q2), and slide the steps to 1 quarter to ensure that the windows do not overlap. The data in each window is aggregated by the mean to generate the time step feature vector t tIf the data for a quarter are missing, the interpolation method is used: the mean data of the same industry and region enterprises in the same period are used to fill in the missing data to ensure the integrity of the feature matrix T.

[0102] Step 5.2. Calculate weights through the temporal attention mechanism:

[0103] α t =softmax(W t Tanh(W t t t +b h ))

[0104] Among them, t t is the time step feature vector, W t is the weight matrix in the temporal attention mechanism, W h is the hidden layer weight matrix, b h is the hidden layer bias vector; through this mechanism, the model can automatically learn and focus on key time points that have a greater 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 this time step will be assigned a higher value.

[0105] Step 5.3. Construct a spatial adjacency matrix S based on the geographical location of the enterprises, and extract regional risk characteristics through graph convolution. The construction of the adjacency matrix includes a geographic distance metric: the Euclidean distance is calculated based on the longitude and latitude of the enterprise registration location, and enterprises with a distance <10 kilometers are marked as spatially adjacent S(i,j)=1. If two enterprises belong to the same industry and the distance is <5 kilometers, the adjacency weight is enhanced to S(i,j)=2, which is enhanced through industry agglomeration and reflects the risk transmission effect of industrial clusters.

[0106] Spatial graph convolution operation, specifically using the GCN graph convolution network layer to extract regional features, and finally output the regional risk feature vector X S ;

[0107] Step 5.4. Concatenate the time series features and spatial features after time attention weighting to generate the dynamic risk feature vector ST. Perform attention weighting on the time series feature matrix T: Where T is the total number of time steps in the time series, that is, the number of time steps after the time window is divided. For example, if 5 years of data are divided by quarter, then T = 20 (4 quarters a year, 20 quarters in 5 years). t : The attention weight of time step t is a scalar value. It is calculated by the temporal attention mechanism in step 5.2 and reflects the importance of the feature vector of the t-th time step in the overall time series and satisfies t tis the time step feature vector, the feature vector of the t-th time step is the feature representation of each time step obtained after the time window is divided and preprocessed in step 5.1. The time dimension comprehensive feature T is obtained by formula att , focusing on risk signals at key time steps (such as cash flow interruption for two consecutive quarters), and the comprehensive features of the time series obtained after weighting by the time attention mechanism, focusing on the key time step information that has a greater impact on credit risk.

[0108] Convolution of the spatial graph output X S Perform global average pooling to obtain the overall risk characteristics of the region S reg Assuming that the spatial graph convolution output has the features of n enterprise nodes, the global average pooling formula can be expressed as

[0109] The time feature T att With spatial feature X s (i) Splicing by dimension to generate dynamic spatiotemporal feature vector ST = [T att ;S reg ];

[0110] Furthermore, the 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 the time series data in the multimodal feature vector is divided into time windows, the temporal attention mechanism calculates the importance weights of different time points, thereby highlighting information at key time points. Simultaneously, a spatial adjacency matrix constructed using the enterprise's geographic information is combined with other information in the multimodal feature vector to generate a dynamic risk feature vector containing spatiotemporal context through spatiotemporal feature fusion. This allows for better exploration of risk characteristics in both temporal and spatial dimensions.

[0111] Furthermore, the dynamic spatiotemporal attention mechanism time series feature mining also includes: the time dimension processing in the dynamic spatiotemporal attention mechanism includes: dividing the time series data (such as historical transaction flow of enterprises, credit records, etc.) according to time windows (such as one window per quarter or year), 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 ].

[0112] Furthermore, the spatial dimension processing within the dynamic spatiotemporal attention mechanism also includes constructing a spatial adjacency matrix S, representing the spatial adjacency relationships between enterprises, by combining geographic information such as the enterprise's registered location and location of operation. Graph convolution is then used to process the spatial adjacency matrix and enterprise feature vectors to extract regional risk characteristics.

[0113] The time series features weighted by time attention are concatenated and fused with the spatial region features to generate a dynamic risk feature vector (ST) containing spatiotemporal contextual information. For example, for multiple small and micro enterprises in a region, if the recent economic environment deteriorates, indicating an increase in spatial risk, and the enterprises themselves experience abnormal fluctuations in recent transaction flows, indicating an increase in temporal risk, the fused dynamic risk feature vector will comprehensively reflect these two risk factors.

[0114] S6. Fusion of the graph-level feature vector and the dynamic spatiotemporal feature vector is input into a multi-layer perceptron (MLP) classifier to output the credit risk probability.

[0115] The specific implementation of step 6 includes: 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, the following steps are included:

[0116] S6.1. Transform the graph-level feature vector G (dimension d) output by the heterogeneous graph neural network into G , representing the global correlation risk of the enterprise) and the dynamic risk feature vector ST (dimension d ST , including time-space context features) are spliced, and dimension alignment and splicing are performed to form the final input feature vector X = [G; ST], with a dimension of d G +d ST .

[0117] The Z-score normalization method is used to normalize the concatenated feature vector to ensure that the contribution of each dimension feature to the classifier is balanced. The multi-layer perceptron (MLP) classifier architecture specifically includes

[0118] Input layer: The number of neurons is equal to the concatenated feature dimension d G +d ST , directly receive the standardized feature vector X'.

[0119] Hidden layer: Set up 2 fully connected hidden layers (the number of layers can be adjusted according to actual needs):

[0120] First hidden layer: The number of neurons n1 can be set to 256, and the activation function uses ReLU (Rectified Linear Unit), the formula is: h1 = max(0, W1X'+b1)

[0121] Among them, W1 is the weight matrix from the input layer to the first hidden layer, and b1 is the bias vector.

[0122] The second hidden layer: the number of neurons n2 = 128, the activation function uses Leaky ReLU to alleviate the problem of neuron "death", the formula is: h2 = max(αh,h) where α = 0.01, h2 is the output after linear transformation, and α is the slope parameter.

[0123] Output layer: A single neuron uses the Sigmoid activation function to output the credit risk probability P∈[0,1], the formula is: Among them, h2 is the output vector of the second hidden layer, W2 is the weight matrix from the second hidden layer to the output layer, which is used to map the hidden layer features to the probability space, b2 is the bias vector, and P represents the probability of default of small and micro enterprises. The larger the value, the higher the risk. Specifically, the multimodal feature vector is processed by the heterogeneous graph neural network and the dynamic spatiotemporal attention mechanism to obtain the graph-level feature vector and the dynamic spatiotemporal feature vector, which are fused as the input of the multi-layer perceptron classifier. The multi-layer perceptron classifier outputs the credit risk probability of small and micro enterprises based on these feature vectors. The graph-level feature vector G and the dynamic spatiotemporal feature ST are input into the multi-layer perceptron MLP classifier, and after calculation by the multi-layer neural network of the classifier, the credit risk probability P of the small and micro enterprises is output. This probability can be expressed as a default probability (such as the possibility of default within the next year) or a credit rating (such as three levels of high risk, medium risk, and low risk). The feature preprocessing, network architecture parameters, training methods, etc. of the MLP can be set as needed by the technology in this field.

[0124] Step 6.2. Calibrate the forecast results based on business rules; for example, if the debt-to-asset ratio is >70%, trigger an alert.

[0125] S7. Identify the nodes and relationship paths in the knowledge graph that have the greatest impact on credit risk based on the attention weights, and determine the core paths. Specifically, step S7 also includes: S7.1. Fusing 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;

[0126] S7.2. Based on the comprehensive weight matrix, identify key nodes whose credit risk impact exceeds a preset threshold;

[0127] S7.3. Starting from key nodes, use a graph traversal algorithm combined with meta-paths to extract core risk transmission paths with high frequency and high weight.

[0128] S7.4. Combined with business logic, screen out risk transmission paths including defaults by related parties, impacts of industry policies, and regional risks, and identify them as core paths.

[0129] 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's credit risk is high, a BFS algorithm can be used to traverse the edges of the knowledge graph and, combined with paths with high attention weights, identify paths that may be caused by an operational crisis at its upstream supplier (connected through supply chain relationships), which in turn affects the company's raw material supply and production, ultimately leading to increased credit risk.

[0130] The extracted paths are screened, and core paths with a frequency above a set threshold (e.g., 50%) are retained. For example, if the path "enterprise → upstream supply chain → industry policy → risk indicator" appears frequently during multiple assessments and has a significant impact on risk prediction, it is retained as a key path.

[0131] Furthermore, the visual explanation generation includes mapping the core reasoning path to the visual interface of the knowledge graph and annotating the attribute information of key nodes. For example, the path is displayed in the graphical interface as node-to-node connections. For nodes related to highly indebted enterprises, key attributes such as their debt ratio are annotated. For nodes related to related party guarantee defaults, detailed information such as the default time and the amount involved are annotated, intuitively demonstrating the path and key factors of risk transmission.

[0132] Furthermore, a risk explanation report can be generated, which combines visual graphs (such as force-directed graphs to display knowledge graph structure and reasoning paths) with textual descriptions to demonstrate risk transmission paths and key influencing factors. For example, the report not only graphically displays the risk paths associated with enterprises, supply chains, policies, etc., but also details the impact of each key node and relationship on credit risk, such as how the tightening of a certain policy leads to rising costs for enterprises and tight capital chains, thereby providing decision makers such as financial institutions with a clear and intuitive basis for risk assessment.

[0133] S8. Combine core paths, risk probabilities, and business rules to output enterprise credit risk assessment results.

[0134] Further, for the identified core risk conduction paths in step S7, assign basic weights 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 risk probabilities. 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, for related party default, W = 0.8; for policy impact, W = 0.7. Combining with the path length attenuation coefficient L = 1 / (1 + n), where n is the number of relationship edges, calculate the comprehensive weight α = L×W; correct the risk probability P output in step S6 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;

[0135] Compare the adjusted risk probability with the dynamic threshold set according to industry and financial indicators, and further calibrate the risk level in combination with business rules to determine the final risk probability for evaluation. Among them, set the dynamic decision threshold 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, divide the industry into three levels of risk categories and set the basic default probability threshold: for high-risk industries such as real estate and chemical industries with high policy sensitivity, complex capital chains and significant cyclical risks, the basic threshold T = 0.5;

[0136] Determine 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 as medium-risk industries, with a basic threshold T = 0.4;

[0137] Determine industries such as public utilities, life services, and scientific research and development with stable cash flows, strong policy support, and low default rates as low-risk industries, with a basic threshold T = 0.3; When the corrected risk probability P > T, trigger the corresponding risk level. Example: P > 0.6 is an extremely high risk; 0.5 < P < 0.6 is a high risk; other situations are judged according to the medium and low risk level rules.

[0138] For high-risk industries such as real estate, set the basic threshold T = 0.5 and the regional default rate R. If the target region R = 30% and the national average is 20%, then the dynamic threshold T' = T×(1 + 0.2×(R - average) / average) = 0.5×1.1 = 0.55. If the corrected probability P' = 0.84 exceeds T', trigger the 'extremely high risk' level. The above settings are only examples, and those skilled in the art can adjust according to the actual situation.

[0139] The rules engine uses business rules to conduct multi-level conditional judgments. Based on factors such as risk probability and core path weights, risk levels and corresponding action recommendations are generated, translating business rules into actual risk assessment decisions. Business rules include basic rules and extended rules. Basic rules further include risk level judgments and corresponding action recommendations based on risk probability. Basic rules based on core paths include: if the risk level is high based on the risk probability, the risk level will be directly increased by one level. For example, if the risk level is originally determined to be medium based on risk probability, it will be increased to medium-high due to the high weight of the related party default path. Action recommendations can be set to require the company to provide detailed financial information and repayment plans of the related party, increasing credit monitoring of the related party. If the related party default is serious, the company may 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 based on the policy's impact. If the policy is favorable, the risk level can be appropriately reduced by one level; if the policy is unfavorable, the risk level can be increased by one level. Action recommendations can be set, for example, to consider offering the company preferential policies, such as extending the repayment period, in case of favorable policies. If the policy is unfavorable, the company may be required to adjust its business strategy to adapt to the policy changes and strengthen supervision of its operating conditions. If the weight of the regional economic risk path is high, the risk level will be increased by one level; the disposal suggestion can be set to require the enterprise to increase the margin or collateral, strengthen the monitoring of the enterprise's capital flow, and prevent capital flight.

[0140] The extended rule settings include extended rules for the company's operating years, market share, and innovation capabilities. For example, the extended rules based on the company's operating years can be set to: when the company's operating years are <2 years, the risk level will be increased by one level based on the original risk level. The disposal suggestion can be set to require the company to provide a detailed business plan and future development plan, and increase the assessment of the company's management team. At the same time, reduce the credit limit and shorten the repayment period. When the company's operating years are ≥10 years and it has been profitable for nearly three consecutive years, the risk level will be reduced by one level based on the original risk level. The disposal suggestion is to appropriately increase the credit limit and grant a certain interest rate discount. At the same time, reduce the frequency of credit assessments, and conduct assessments every six months.

[0141] By setting these basic and extended rules, we can implement multi-level conditional judgments based on business rules. Based on factors such as risk probability and core paths, we can generate reasonable risk levels and corresponding treatment recommendations, thus transforming business rules into actual risk assessment decisions. The above settings are only examples, and those skilled in the art can adjust them according to their actual circumstances.

[0142] Furthermore, a comprehensive report is output that includes risk levels, key influencing factors, and reasoning pathways. For example, the report clearly states that a company's credit risk level is medium, with key influencing factors including poor performance of upstream companies in the supply chain and recent industry policy adjustments. The report also details the reasoning pathways and risk transmission mechanisms, providing strong support for financial institutions in making financing decisions.

[0143] Furthermore, users can use natural language input, such as "Query the main reasons for a company's default." The system will perform search and analysis within the knowledge graph and reasoning paths based on the query content, returning corresponding explainable reasoning results. For example, the system displays the paths and nodes related to the reasons for the company's default in a visual interface, and provides a detailed textual explanation of the reasons, such as the default of a related-party guarantee resulting in joint and several liability for the company, the withdrawal of funds, and other issues, further enhancing the convenience and accuracy of decision support.

[0144] In order to verify the effectiveness and accuracy of the method of the present invention, the present invention uses a real small and micro enterprise data set for experiments. A data set containing basic enterprise information, financial data, unstructured text and image data, and historical credit records is selected and processed and analyzed according to the above method steps. By comparing with traditional credit assessment methods such as statistical models based only on financial data, it is found that the method of the present invention has a significant improvement in the accuracy of credit risk prediction and can more accurately identify high-risk enterprises. At the same time, by evaluating the generated explainable reasoning path, it can reasonably and clearly explain the causes of credit risk, provide a reliable basis for actual financial decision-making, and effectively solve the "black box" problem of traditional models.

[0145] In order to verify the effectiveness and accuracy of the method of the present invention, the following experiments were conducted:

[0146] This dataset uses a real-world dataset of small and micro enterprises from the Yangtze River Delta region. It contains detailed information on over 500 small and micro enterprises, including basic information, financial data, official website text, news reports, images of invoices and business licenses, and historical credit records. Structured data accounts for 60%, while unstructured data accounts for 40%. The small and micro enterprises in the dataset span multiple industries, including manufacturing, services, wholesale and retail, and are highly representative.

[0147] 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. It predicts the probability of corporate default by establishing a logistic regression model. It is simple to understand and has high computational efficiency, but it has high requirements for the linear assumption of the data and has difficulty in handling complex nonlinear relationships. Support vector machines perform classification by finding the optimal classification hyperplane. It has certain advantages when dealing with small samples and nonlinear problems, but it is more sensitive to the choice of kernel function. The selection of these two methods for comparison can fully demonstrate the advantages of the method of the present invention over different types of traditional methods.

[0148] Accuracy, recall, F1 value, and area under the AUC-ROC curve are used as evaluation metrics. Accuracy measures the proportion of samples correctly predicted by the model to the total number of samples, reflecting the model's overall predictive ability. Recall measures the proportion of samples that are actually positive and correctly predicted as positive by the model, reflecting the model's ability to identify positive samples. The F1 value 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.

[0149] The dataset was divided into training, validation, and test sets in a ratio of 70%, 15%, and 15%. In the method presented here, the heterogeneous graph neural network was set to three 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. ReLU and Leaky ReLU were used as activation functions, respectively. During training, the Adam optimization algorithm was used, the learning rate was set to 0.001, and the number of training rounds was 100. For comparison, the default parameter settings were used for both logistic regression and support vector machines.

[0150] Evaluation Metrics Method of the present invention Logistic Regression Support Vector Machine Accuracy 85％ 70％ 75％ Recall 82％ 65％ 70％ F1 value 0.83 0.67 0.72 AUC-ROC 0.90 0.80 0.82

[0151] The present invention integrates multimodal data to comprehensively capture the risk characteristics of enterprises. The construction of a knowledge graph and the application of heterogeneous graph neural networks effectively mine the complex relationships between enterprises. The dynamic spatiotemporal attention mechanism better processes time series data, accurately capturing the temporal and spatial characteristics of risk changes. These advantages enable the present method to more accurately identify high-risk enterprises in actual financial risk control scenarios, providing financial institutions with more reliable decision-making basis and effectively reducing credit risk.

[0152] In summary, the present invention provides a comprehensive, accurate and interpretable credit risk assessment method for small and micro enterprises. By integrating and deeply mining multi-source heterogeneous data, it provides an innovative technical solution for the field of financial risk control, and has good application prospects and practical value.

[0153] Secondly, as attached Figure 2 As shown, a multimodal enterprise credit risk assessment system 200 based on knowledge graph is characterized by including:

[0154] The data collection and preprocessing module 201 collects enterprise-related structured data and unstructured data from multiple sources, wherein the unstructured data includes text data, image data, and time series data, and preprocesses the data;

[0155] The knowledge graph construction module 202 is used to construct an enterprise financial knowledge graph, using structured data as entity attributes and mapping unstructured data into association relationships through entity linking technology, generating triples and storing them in a database;

[0156] The multimodal embedding module 203 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;

[0157] 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;

[0158] Dynamic spatiotemporal attention module 205, used to extract time series attribute data of entities from the knowledge graph and generate dynamic spatiotemporal feature vectors through temporal attention mechanism and spatial graph convolution;

[0159] The feature vector fusion module 206 is used to fuse the graph-level feature vector and the dynamic spatiotemporal feature vector and input the result into the multi-layer perceptron (MLP) classifier to output the credit risk probability;

[0160] Path determination module 207, used to identify the nodes and relationship paths that have the greatest impact on credit risk in the knowledge graph based on attention weights, and determine the core path;

[0161] The risk assessment module 208 is used to output enterprise credit risk assessment results by combining core paths, risk probabilities and business rules.

[0162] In a third aspect, an embodiment of the present invention further provides an electronic device, comprising a processor and a memory, wherein the memory stores computer-executable instructions that can be executed by the processor, and the processor executes the computer-executable instructions to implement any one of the methods provided in the first aspect.

[0163] The core of the invention The present invention provides a data processing system, which includes a memory, a processor, and a computer program stored on the memory and runnable on the processor, wherein the processor executes the computer program and implements the data processing method based on the knowledge graph as described above.

[0164] A second aspect of the present invention provides a data processing device, which includes a memory, a processor, and a computer program stored in the memory and runnable on the processor, wherein the processor executes the computer program and implements the data processing method as described in the aforementioned embodiments.

[0165] In addition to the above-mentioned methods and devices, the embodiments herein are also computer program products, which include computer program instructions, which, when executed by a processor, cause the processor to perform the steps of the target detection method according to various embodiments herein described in the above-mentioned "Exemplary Method" section of this specification.

[0166] The basic principles of the present disclosure have been described above in conjunction with specific embodiments. However, it should be noted that the advantages, strengths, and effects mentioned in this disclosure are merely illustrative and not restrictive, and should not be construed as necessarily possessed by each embodiment of the present disclosure. Furthermore, the specific details disclosed above are provided for illustrative purposes and to facilitate understanding, rather than as limitations. These details do not limit the present disclosure to necessarily being implemented using these specific details.

[0167] Each embodiment in this specification is described in a progressive manner, with each embodiment focusing on its differences from the other embodiments. References to the same or similar parts between the various embodiments are sufficient. For system embodiments, since they largely correspond to method embodiments, their description is relatively simple. For relevant parts, references to the description of the method embodiments are sufficient.

[0168] The block diagrams of the devices, devices, equipment, and systems involved in this disclosure 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 will be appreciated by those skilled in the art, these devices, devices, equipment, and systems can be connected, arranged, or configured in any manner. Words such as "include," "comprise," "have," and the like are open-ended words, meaning "including but not limited to," and can be used interchangeably therewith. The words "or" and "and" used herein refer to the words "and / or" and can be used interchangeably therewith, unless the context clearly indicates otherwise. The word "such as" used herein refers to the phrase "such as but not limited to," and can be used interchangeably therewith.

[0169] The methods and apparatus of the present disclosure may be implemented in many ways. For example, the methods and apparatus of the present disclosure may be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware. The above order of steps for the method is for illustration only, and the steps of the method of the present disclosure are not limited to the order specifically described above unless otherwise specified. In addition, in some embodiments, the present disclosure may also be implemented as programs recorded in a recording medium, which include machine-readable instructions for implementing the methods according to the present disclosure. Thus, the present disclosure also covers recording media that store programs for executing the methods according to the present disclosure.

[0170] It should also be noted that, in the apparatus, equipment and method of the present disclosure, each component or each step can be decomposed and / or recombined. These decompositions and / or recombinations should be regarded as equivalent schemes of the present disclosure. The above description of the disclosed aspects is provided to enable any technician in this field to make or use the present disclosure. Various modifications to these aspects will be very obvious to those skilled in the art, and the general principles defined herein can be applied to other aspects without departing from the scope of the present disclosure. Therefore, the present disclosure is not intended to be limited to the aspects shown here, but to the widest range consistent with the principles and novel features disclosed herein.

[0171] The above description has been provided for the purpose of illustration and description. In addition, this description is not intended to limit the embodiments of the present disclosure to the forms disclosed herein. Although a number of example aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations thereof.

Claims

1. A multimodal enterprise credit risk assessment method based on knowledge graph, characterized by: The following steps are involved: S1. Collect enterprise-related structured data and unstructured data from multiple sources, where the unstructured data includes text data, image data, and time series data, and preprocess the data; S2. Build an enterprise financial knowledge graph, treating structured data as entity attributes and mapping unstructured data into relationships using entity linking technology. Triples are generated and stored in a database. S3. 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; S4. Input the multimodal initial feature vector into the heterogeneous graph neural network to extract the graph-level feature vector; S5. Extract the time series attribute data of entities from the knowledge graph and generate dynamic spatiotemporal feature vectors through temporal attention mechanism and spatial graph convolution; S6. Fusion of the graph-level feature vector and the dynamic spatiotemporal feature vector is input into a multi-layer 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. Combine core paths, risk probabilities, and business rules to output enterprise credit risk assessment results.

2. The method according to claim 1, wherein: The 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 is preprocessed, including 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, using interpolation or forward filling to fill in missing values, and standardizing structured data.

3. The method according to claim 1, wherein: The 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-related party, and the relationship between unstructured data and enterprises; Step S2.

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

4. Generate triples of 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, wherein The step S3 further comprises: 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 in the knowledge graph and generate a global embedding vector through pooling operations; extract visual features from image data and map them to a space of the same dimension; 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, wherein: The step S4 further includes: Step S4.

1. Input the multimodal initial feature vectors and arrange them in node order to construct a node feature matrix H for 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 relationships 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 heterogeneous message passing mechanism using the node feature matrix H, heterogeneous adjacency matrix A, and multimodal initial feature vectors obtained in step S4.1: Where Wr is the weight matrix of relation r, is the neighbor feature vector of node i aggregated under relationship r, is the neighbor set of node i under relation r, is the attention score between node i and neighbor node j under relationship r, which is calculated by node features and relationship weights and is used to measure the importance of neighbor nodes to the current node. j is the initial multimodal initial feature vector of neighbor node j; Step S4.

3. For the neighbor feature vectors, a meta-path-guided aggregation strategy is used to capture cross-type dependencies. A pre-defined 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 the multi-layer graph convolutional network to generate a graph-level feature vector G and a node-level feature vector H', where 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.

6. The method according to claim 1, wherein: The step S5 further includes: Step S5.

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

2. Calculate the weight of time point t through the temporal attention mechanism based on the feature matrix T: α t =softmax(W t ·Tanh(W t t t +b h )) Among them, α t is the attention weight at time point t, which is used to dynamically adjust the importance of different time steps, t t is the time point feature vector, W t is the weight matrix in the temporal attention mechanism, W h is the hidden layer weight matrix, b h is the hidden layer bias vector; Step S5.

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

4. Concatenate the temporal features weighted by temporal attention with the spatial features to generate a dynamic risk feature vector ST.

7. The method according to claim 1, wherein: The 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. Normalize the concatenated feature vector to ensure that the contributions of features from each dimension to the classifier are balanced. Step 6.

3. Input the normalized feature vector into a multi-layer 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 small and micro enterprises is output.

8. The method according to claim 1, characterized in that The step S7 further comprises: S7.

1. Fuse the heterogeneous relational 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 credit risk impact exceeds a preset threshold; S7.

3. Starting from key nodes, use a graph traversal algorithm combined with meta-paths to extract core risk transmission paths with high frequency and high weight. S7.

4. Combined with business logic, screen out the paths of default by related parties, impact of industry policies, and regional risk transmission, and identify them as core paths.

9. A multimodal enterprise credit risk assessment system based on knowledge graph, characterized by: include: The data collection and preprocessing module is used to collect enterprise-related structured data and unstructured data from multiple sources, where the unstructured data includes text data, image data, and time series data, and preprocess the data; The knowledge graph construction module is used to build an enterprise financial knowledge graph, using structured data as entity attributes and unstructured data as association relationships through entity linking technology to generate triples and store 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; Heterogeneous graph neural network module, used to input multimodal initial feature vectors into the heterogeneous graph neural network and extract graph-level feature vectors; Dynamic spatiotemporal attention module, which is used to extract time series attribute data of entities from the knowledge graph and generate dynamic spatiotemporal feature vectors through temporal attention mechanism and spatial graph convolution; The feature vector fusion module is used to fuse the graph-level feature vector and the dynamic spatiotemporal feature vector and input them into the multi-layer perceptron (MLP) classifier to output the credit risk probability; The path determination module is used to identify the nodes and relationship paths that have the greatest impact on credit risk in the knowledge graph based on attention weights, and determine the core path; The risk assessment module is used to output enterprise credit risk assessment results by combining core paths, risk probabilities and business rules.

10. An electronic device, characterized in that: The electronic device includes: a memory and a processor, wherein the memory and the processor are coupled; the memory stores program instructions, and when the program instructions are executed by the processor, the electronic device executes the method according to any one of claims 1 to 8.

11. A computer-readable storage medium, characterized in that The method comprises a computer program, which, when executed on an electronic device, causes the electronic device to execute the method according to any one of claims 1 to 8.

Images