Early warning method, device, equipment, medium and product

Abstract

The invention provides an early warning method and device, equipment, a medium and a product, and relates to the technical field of data processing. According to the specific technical scheme, the method comprises the steps of obtaining a mechanism relation matrix and a mechanism feature matrix of a to-be-processed target service mechanism in a supply chain in an early warning service; the mechanism relation matrix and the mechanism feature matrix are input into a graph neural network model, the graph neural network model comprises a self-coding module, a random walk module and a self-training clustering module, and in the training process of the graph neural network model, context prediction loss output by the random walk module is used for constraining model parameters. A mechanism relation matrix and a mechanism feature matrix are converted into potential feature vectors of a target business mechanism through model parameters of a self-encoding module and a self-encoder module, and massive and sparse matrixes possibly containing noise are compressed into dense and low-dimensional vectors through the self-encoding module, so that redundant information is removed, the data quality is improved, and the service quality of the target business mechanism is improved. And the accuracy of subsequent early warning is improved.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to an early warning method, apparatus, device, medium, and product. Background Technology

[0002] With the development of economy and technology, the number of business entities experiencing operational anomalies is increasing. In order to reduce the occurrence of anomalies in business entities, it is necessary to provide early warnings for a large number of business entities.

[0003] Currently used technologies employ traditional early warning models to provide warnings to institutions, but these models suffer from insufficient accuracy. Summary of the Invention

[0004] This application provides an early warning method, apparatus, device, medium, and product that can improve the accuracy of early warnings for target business entities.

[0005] Firstly, an early warning method is provided. This method includes: obtaining the organizational relationship matrix and organizational feature matrix of the target business entity to be processed in the early warning business within the supply chain. The supply chain includes interconnected target business entities and multiple other business entities. The organizational relationship matrix characterizes the relationships between each target business entity and each of the other business entities. The organizational feature matrix includes the organizational attribute characteristics of each target business entity across multiple health dimensions. Directly obtaining the matrix corresponding to the target business entity eliminates the need for further matrix processing before inputting it into the subsequent model, saving data preprocessing time and improving the efficiency of early warning.

[0006] The organizational relationship matrix and organizational feature matrix are input into a graph neural network (GNN) model, which includes an autoencoder module, a random walk module, and a self-training clustering module. During the training process of the GNN model, the context prediction loss output by the random walk module is used to constrain the model parameters in the autoencoder and self-training clustering modules. Firstly, the input of the GNN combines the relationships between the target business organization and other business organizations, as well as the organizational attribute characteristics of the target business organization across multiple health dimensions. This provides more comprehensive input data, which helps improve the accuracy of the model output and thus the accuracy of early warnings. Secondly, the context prediction loss output by the random walk module constrains the model parameters in the autoencoder and self-training clustering modules, making the model parameters more reasonable and thus improving the model's deterministic accuracy.

[0007] By utilizing the model parameters of the autoencoder and auto-encoder modules, the organizational relationship matrix and organizational feature matrix are transformed into latent feature vectors for the target business organizations. The dimensionality of these latent feature vectors is smaller than that of either the organizational relationship matrix or the organizational feature matrix. Firstly, the autoencoder module compresses massive, sparse, and potentially noisy matrices into dense, low-dimensional vectors, removing redundant information and improving data quality. Secondly, the latent feature vectors reflect the deep fusion result of the organizational relationship matrix and the organizational feature matrix, providing data support for subsequent processing.

[0008] By using a self-trained clustering module and its model parameters, the latent feature vectors are transformed into the probability of a target business entity corresponding to a reference group of institutions. Firstly, the output is no longer a binary judgment of whether something is abnormal; instead, the probability of belonging is quantified, making the early warning more interpretable. Secondly, since the probability of belonging is the probability of a target business entity corresponding to a reference group of institutions, the probability output using the reference group as a reference target is more accurate, thus improving the accuracy of subsequent early warnings.

[0009] When the probability of attribution is greater than or equal to a preset probability, an early warning is issued to the target business unit. Firstly, the preset probability is flexibly configurable, ensuring controllable warning sensitivity. Secondly, issuing a timely warning to the target business unit before any actual anomalies occur can prevent losses.

[0010] This method, on the one hand, combines the relationships between the target business unit and other business units, as well as the organizational attribute characteristics of the target business unit in multiple health dimensions, into the input of the graph neural network. The input data is more comprehensive, which helps to improve the accuracy of the model output and thus the accuracy of the early warning. On the other hand, the context prediction loss output by the random walk module constrains the model parameters in the autoencoder module and the self-training clustering module, making the model parameters corresponding to the modules more reasonable, thereby improving the determination accuracy of the model and the accuracy of the early warning for the target business unit.

[0011] In one possible implementation of the first aspect, the model parameters of the autoencoder module include: shared weights of the first layer of the graph neural network, mean weights and offset weights of the second layer of the graph neural network; through the autoencoder module and the model parameters of the autoencoder module, the organizational relationship matrix and the organizational feature matrix are transformed into latent feature vectors of the target business organization, including: through the autoencoder module, adding self-loops and normalizing the organizational relationship matrix to obtain the normalized target matrix of the organizational relationship matrix; transforming the target matrix and the organizational feature matrix through shared weights and mean weights to obtain the mean feature vector of the target business organization; transforming the target matrix and the organizational feature matrix through shared weights and offset weights to obtain the offset feature vector of the target business organization; sampling Gaussian data containing the mean feature vector and the offset feature vector to obtain the latent feature vector of the target business organization.

[0012] In this embodiment, firstly, the relation matrix is ​​subjected to self-loop addition and normalization to enhance model stability and improve information transmission. Secondly, shared weights and mean / offset weights are used to generate mean and offset feature vectors respectively, efficiently achieving variational inference, learning the probability distribution of the data, and providing a basis for subsequent data sampling of Gaussian data. Thirdly, latent feature vectors are obtained by sampling the Gaussian distribution, introducing randomness. By randomly sampling the learned distribution, slightly different latent vectors can be obtained each time. This introduces beneficial noise into the model, equivalent to a kind of regularization, forcing the latent space to be smoother and more continuous, and improving the model's robustness to noise and small perturbations.

[0013] In one possible implementation of the first aspect, the model parameters of the self-trained clustering module include: a target organization center vector, multiple reference organization center vectors including the target organization center vector, and organization degrees of freedom, wherein the organization center vector and each reference organization center vector belong to a reference organization group; through the self-trained clustering module and its model parameters, the latent feature vector is transformed into the probability of the target business organization belonging to the reference organization group, including: inputting the latent feature vector into the self-trained clustering module, and determining the difference between the latent feature vector and the target organization center vector as a first difference; determining the difference between the latent feature vector and each reference organization center vector as a second difference; and transforming the first difference and each second difference according to the organization degrees of freedom to obtain the probability of the target business organization belonging to the reference organization group.

[0014] In this embodiment, firstly, the first and second differences can reflect the degree of matching between the target institution and all other possible clusters, facilitating the subsequent conversion of the second difference into a quantified attribution probability. Secondly, based on the statistical distribution of degrees of freedom, all the first and second differences are transformed into a probability distribution. Compared to hard clustering with either / or approaches, this provides a more refined probability distribution, and the probability value itself represents the confidence level, making the early warning decision more interpretable.

[0015] In one possible implementation of the first aspect, the training method for the graph neural network model includes: during the training process, obtaining the total loss of the graph neural network model; the total loss includes the reconstruction loss of the autoencoder module, the context prediction loss of the random walk module, and the clustering loss of the self-trained clustering module; if the total loss satisfies the model iteration condition, backpropagating and updating the model parameters of the autoencoder module, the random walk module, and the self-trained clustering module respectively, until the total loss no longer satisfies the model iteration condition, thereby obtaining the trained graph neural network model and model parameters.

[0016] In this embodiment, firstly, the reconstruction loss, context prediction loss, and clustering loss are summed into a single total loss, enabling the model to optimize three key objectives simultaneously in a single training iteration. This avoids objective conflicts or suboptimal solutions that may arise from phased training, ensuring higher accuracy in the model's output attribution probabilities. Secondly, a clear training stopping criterion is provided when the total loss is less than a preset loss or the corresponding number of model iterations reaches a preset number. This process is fully automated, ultimately outputting a well-trained, parameter-fixed optimal model, thus ensuring the efficiency of model training to a certain extent.

[0017] In one possible implementation of the first aspect, the method further includes: acquiring training samples for early warning services and model parameters of the autoencoder module; wherein the training samples include an adjacency matrix and a node feature matrix, the adjacency matrix includes multiple business units and the node relationships between them, the node feature matrix includes the node attribute features of each business unit in multiple health dimensions, and the model parameters of the autoencoder module include the first weight of the first layer of the graph neural network and the second mean weight and the second offset weight of the second layer of the graph neural network; the adjacency matrix is ​​subjected to self-loop addition processing and normalization processing to obtain a candidate matrix of normalized self-loops corresponding to the adjacency matrix; the candidate matrix and the node feature matrix are transformed according to the first weight and the second mean weight to obtain the mean node matrix corresponding to each business unit; the candidate matrix and the node feature matrix are transformed according to the first weight and the second offset weight to obtain the offset node matrix corresponding to each business unit; Gaussian data containing the mean node matrix and the offset node matrix are sampled to obtain the potential node vector of each business unit; and the reconstruction loss of the autoencoder module is obtained based on the potential node vectors and node relationships.

[0018] In this embodiment, firstly, self-loops are added to the adjacency matrix and normalized to enhance model stability and improve information transmission. Secondly, the first weight and the second mean / offset weights are used to generate mean and offset feature matrices respectively, efficiently achieving variational inference and learning the probability distribution of the data, providing a basis for subsequent data sampling of Gaussian data. Thirdly, latent node vectors are obtained by sampling the Gaussian distribution, introducing randomness. By randomly sampling the learned distribution, slightly different latent node vectors can be obtained each time. This introduces beneficial noise into the model, equivalent to a regularization, forcing the latent space to be smoother and more continuous, improving the model's robustness to noise and small perturbations. Fourthly, the autoencoding process of the graph neural network is placed within the variational Bayesian framework. Through a Bernoulli distribution-based loss function aimed at edge reconstruction, the model learns a high-quality latent feature representation that reflects the node's own attributes, accurately encodes its network structure role, and is probabilistically robust. This provides information-rich, structure-aware, and robust feature inputs for subsequent clustering and anomaly warning tasks.

[0019] In another possible implementation of the first aspect, the method further includes: obtaining the model parameters of the random walk module, the model parameters of the random walk module including the window size; performing random walk processing on each business unit in the adjacency matrix according to the window size to obtain multiple node sequences, the node sequences including intermediate nodes as intermediate units, upper nodes of the preceding units as intermediate units, and lower nodes of the following units as intermediate units, the number of nodes contained in each node sequence corresponding to the window size; for each node sequence, obtaining the sequence prediction loss corresponding to the node sequence based on the node attribute features corresponding to the preceding units, the node attribute features corresponding to the following units, the node relationship between the preceding units and the intermediate units, and the node relationship between the following units and the intermediate units; and obtaining the context prediction loss of the random walk module based on each sequence prediction loss.

[0020] In this embodiment, firstly, the model can flexibly control the scope of context information capture during random walks. The window size determines how far away from the central node a node in the sequence will be considered as context. Secondly, node sequences simulating information or abnormal propagation paths are dynamically generated from the original static adjacency matrix. Multiple sequences provide rich samples of local network structure. Thirdly, by aggregating the individual losses generated by a large number of random walk sequences, an accurate context prediction loss is obtained.

[0021] In another possible implementation of the first aspect, the method further includes: determining multiple initial cluster centers in the potential node vectors of each business unit; if each initial cluster center satisfies the K-means clustering termination condition, determining the target organization center vector and at least one reference organization center vector containing the target organization center vector in the potential node vectors corresponding to each initial cluster center; generating the attribution probability distribution corresponding to each business unit based on each potential node vector, the target organization center vector, each reference organization center vector, and the initial degrees of freedom; determining the target distribution based on the attribution probability distribution; if the initial degrees of freedom are greater than 1, the concentration of the target distribution is greater than the concentration of the attribution probability distribution; and calculating the KL divergence of the target distribution and the attribution probability distribution to obtain the clustering loss of the self-trained clustering module.

[0022] In this embodiment, firstly, the KL divergence loss, as a self-generated supervisory signal, guides feature learning to optimize in a direction that makes the clustering results more distinct and deterministic, thereby greatly improving the cohesion and separability of the clusters. Secondly, the KL clustering loss is backpropagated, simultaneously optimizing the feature encoder parameters and the cluster center vector, thus achieving optimization of the self-trained clustering module.

[0023] Secondly, a warning device is provided, the device comprising: The matrix acquisition module is used to acquire the organizational relationship matrix and organizational feature matrix of the target business entities to be processed in the early warning business within the supply chain. The supply chain includes interconnected target business units and multiple other business units. The organizational relationship matrix is ​​used to characterize the relationships between the target business units and each of the other business units. The organizational characteristic matrix includes the organizational attribute characteristics of the target business units in multiple health dimensions. The model processing module is used to input the organizational relationship matrix and organizational feature matrix into the graph neural network model. The graph neural network model includes an autoencoder module, a random walk module, and a self-training clustering module. During the training of the graph neural network model, the context prediction loss output by the random walk module is used to constrain the model parameters in the autoencoder module and the self-training clustering module. The matrix processing module is used to transform the organizational relationship matrix and organizational feature matrix into potential feature vectors of the target business organization using the model parameters of the autoencoder module and the autoencoder module; wherein the dimension of the potential feature vector is smaller than the dimension of the organizational relationship matrix or the organizational feature matrix. The parameter processing module is used to transform the latent feature vector into the probability of the target business unit belonging to the reference institution group through the self-trained clustering module and the model parameters of the self-trained clustering module. The early warning module is used to issue early warnings to target business entities when the probability of attribution is greater than or equal to a preset probability.

[0024] Thirdly, an early warning device is provided, the method comprising: a memory and at least one processor. The memory is communicatively connected to the processor. The memory is used to store computer program code, which includes computer instructions. When the processor executes the computer instructions, it causes the electronic device to perform the method as described in the first aspect and any possible implementation thereof.

[0025] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer instructions. When these computer instructions are executed by a processor, they are used to implement the method as described in the first aspect and any possible implementation thereof.

[0026] Fifthly, embodiments of this application provide a computer program product that, when running on a computer / executed by the computer's processor, implements the method described in the first aspect and any possible design thereof. The computer may be the early warning device described in the third aspect and any possible implementation thereof.

[0027] Understandably, the beneficial effects that the warning device of the second aspect, the warning equipment of the third aspect, the computer-readable storage medium of the fourth aspect, and the computer program product of the fifth aspect can achieve can be referred to the beneficial effects of the first aspect and any possible implementation thereof, which will not be repeated here. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the structure of an early warning system provided in an embodiment of this application; Figure 2 A flowchart illustrating an early warning method provided in an embodiment of this application; Figure 3 A flowchart illustrating a method for generating potential feature vectors provided in an embodiment of this application; Figure 4 A flowchart illustrating a method for generating attribution probability provided in an embodiment of this application; Figure 5 A flowchart illustrating an optimization method for reconstruction loss provided in an embodiment of this application; Figure 6 A flowchart illustrating a clustering loss optimization method provided in an embodiment of this application; Figure 7 A training process data flow diagram provided in an embodiment of this application; Figure 8This is a schematic diagram of the structure of an early warning device provided in an embodiment of this application; Figure 9 This is a schematic diagram of the structure of an early warning device provided in an embodiment of this application. Detailed Implementation

[0029] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this embodiment, unless otherwise stated, "a plurality of" means two or more.

[0030] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0031] The technical solutions provided in this application, including the collection, storage, use, processing, transmission, provision, and disclosure of financial data or user data, comply with relevant laws and regulations and do not violate public order and good morals.

[0032] It should be noted that in the embodiments of this application, certain software, components, models and other existing solutions in the industry may be mentioned. These should be regarded as exemplary and are only intended to illustrate the feasibility of implementing the technical solution of this application. However, it does not mean that the applicant has used or necessarily used the solution.

[0033] With economic and technological development, the number of business entities experiencing operational anomalies is increasing. To prevent losses, it is necessary to provide early warnings for a vast number of business entities. For example, the increasing number of abnormal events occurring in micro and small enterprises (MSEs) is causing them significant losses. Providing early warnings to MSEs allows them to conduct timely self-checks and avoid further losses.

[0034] In related technologies, the first aspect is to use traditional early warning models to issue early warnings to institutions. The input of traditional early warning models is usually the characteristics of the institution itself, which has the problem of insufficient accuracy in early warning.

[0035] Secondly, while traditional machine learning methods and some data-driven techniques were employed, these methods have limitations in handling complex and nonlinear features, resulting in less than ideal predictive performance. Traditional machine learning methods often struggle to effectively extract the complex nonlinear features among various participants in supply chain finance. For example, supply chain finance involves multiple links and participants, with highly complex relationships and interactions. Early warning models based on GNNs focus on learning entity-level attention at nodes, often neglecting the fact that different feature dimensions encapsulate different aspects of anomalies in a company, thus still resulting in insufficient accuracy in early warnings.

[0036] In this embodiment, regarding the first aspect mentioned above, firstly, the input of the graph neural network combines the relationships between the target business unit and other business units, as well as the organizational attribute characteristics of the target business unit in multiple health dimensions. The input data is more comprehensive, which helps improve the accuracy of the model output and thus the accuracy of the early warning. Secondly, the context prediction loss output by the random walk module constrains the model parameters in the autoencoder module and the self-trained clustering module, making the model parameters corresponding to the modules more reasonable, thereby improving the model's deterministic accuracy.

[0037] Regarding the second aspect mentioned above, firstly, a graph neural network model was used to accurately determine the probability of a target business enterprise belonging to a reference institution group. The autoencoder module in the graph neural network model compresses massive, sparse, and potentially noisy matrices into dense, low-dimensional vectors, removing redundant information and improving data quality. Through the model parameters of the graph neural network, complex nonlinear features can be effectively extracted, resulting in more comprehensive nonlinear features and overcoming the limitations of traditional machine learning methods. Secondly, the input of the graph neural network model combines the relationships between the target business institution and other business institutions, as well as the institutional attribute characteristics of the target business institution in multiple health dimensions. The input data is more comprehensive, better extracting and analyzing complex nonlinear features in the supply chain finance environment, which helps improve the accuracy of the model output, thereby improving the accuracy of early warnings, and thus enhancing the ability to identify credit anomalies in SMEs.

[0038] The early warning method provided in this application embodiment can be applied to, for example, Figure 1 The warning system shown includes a warning device 101 and a terminal 102.

[0039] The early warning device 101 is used to acquire the organizational relationship matrix and organizational feature matrix of the target business unit to be processed in the early warning business within the supply chain. The supply chain includes interconnected target business units and multiple other business units. The organizational relationship matrix is ​​used to characterize the relationships between the target business unit and each of the other business units. The organizational feature matrix includes the organizational attribute characteristics of the target business unit in multiple health dimensions.

[0040] The early warning device 101 is used to input the mechanism relationship matrix and mechanism feature matrix into the graph neural network model; wherein, the graph neural network model includes an autoencoder module, a random walk module and a self-training clustering module; during the training process of the graph neural network model, the context prediction loss output by the random walk module is used to constrain the model parameters in the autoencoder module and the self-training clustering module.

[0041] The early warning device 101 is used to transform the organizational relationship matrix and organizational feature matrix into potential feature vectors of the target business organization through the model parameters of the self-encoding module and the self-encoder module; wherein the dimension of the potential feature vector is smaller than the dimension of the organizational relationship matrix or the organizational feature matrix.

[0042] The early warning device 101 is used to convert the potential feature vector into the probability of the target business unit belonging to the reference institution group through the self-trained clustering module and the model parameters of the self-trained clustering module.

[0043] The early warning device 101 is used to issue an early warning to the terminal 102 corresponding to the target business unit when the attribution probability is greater than or equal to the preset probability.

[0044] The early warning device 101 can be a server, which can be implemented by a standalone server or a server cluster composed of multiple servers.

[0045] In other embodiments, the warning system includes a warning server for executing the method steps corresponding to the warning device 101 and the terminal 102.

[0046] The early warning method provided in this application can be applied to different business scenarios.

[0047] For example, in the financial industry, the target business entity can be a micro or small enterprise, while other business entities can be enterprises that are interconnected with the micro or small enterprise in the supply chain.

[0048] For example, in the healthcare industry, the target business entity can be a pharmacy store, while other business entities can be companies or stores that are interconnected with the pharmacy store in the supply chain.

[0049] For example, in a manufacturing scenario, the target business entity can be a manufacturing company, while other business entities can be companies that are interconnected with the manufacturing company in the supply chain.

[0050] Therefore, the embodiments of this application do not limit specific business scenarios.

[0051] In one embodiment, such as Figure 2 As shown, an early warning method is provided that can be applied to... Figure 1 The method can be implemented in early warning devices, or applied to servers with data processing capabilities, cloud computing platforms, edge computing devices, or chips, or devices with computing capabilities. This application does not limit the specific form of the device executing the method; taking the application of the method to a server as an example, the method specifically includes: S201, obtain the organizational relationship matrix and organizational characteristic matrix of the target business entities to be processed in the early warning business in the supply chain.

[0052] To reduce data preprocessing time, the organizational relationship matrix and organizational feature matrix of the target business units to be processed in the early warning business can be directly obtained. The supply chain includes interconnected target business units and multiple other business units. The organizational relationship matrix represents the relationships between each target business unit and each of the other business units, while the organizational feature matrix includes the organizational attribute characteristics of each target business unit across multiple health dimensions.

[0053] In some embodiments, the early warning service can be a specific task for abnormal alarms and special monitoring.

[0054] The target business entity can be a single entity being assessed and monitored in the current early warning operation. Specifically, the target business entity can be a small or micro enterprise or a pharmacy, etc.

[0055] A supply chain can be a network of transactions and collaborations formed by a target business entity and its upstream and downstream partners (other business entities).

[0056] An organizational relationship matrix can be a matrix describing the relationships between a target business entity and other business entities. For example, a target business entity could be a micro or small enterprise, and the relationships between the micro or small enterprise and other business entities could include customer relationships, investor relationships, and supplier relationships.

[0057] Other business entities can be entities in the supply chain that are related to the target business entity.

[0058] An organizational feature matrix can be a matrix that describes the organizational attribute characteristics of a target business organization across multiple health dimensions.

[0059] A health dimension can be a dimension describing the business health of a target organization. Organizational attributes of the health dimension can include registered capital, establishment date, number of legal cases, and administrative processing indicators.

[0060] Specifically, registered capital can be the total amount of capital planned to be invested by shareholders as declared and registered with the registration authority upon the establishment of the target business entity. Establishment date can be the date on which the target business entity completes its establishment registration with the industrial and commercial administration department and obtains legal person status. Number of legal cases can be the total number of legal cases in which the target business entity is a party within a certain statistical period. Administrative processing mark can be the objective rating of the target business entity by the administrative agency.

[0061] In one possible implementation, before obtaining the matrix, the server can acquire the network structure between the target business entity and multiple other business entities. The server can parse the network structure to obtain the organizational relationship matrix and organizational feature matrix of the target business entity to be processed in the early warning business within the supply chain. The supply chain includes the interconnected target business entity and multiple other business entities. The organizational relationship matrix characterizes the relationships between the target business entity and each of the other business entities, and the organizational feature matrix includes the organizational attribute characteristics of the target business entity across multiple health dimensions. Acquiring the network structure allows for more comprehensive data, providing strong data support for subsequent model determination.

[0062] For example, the network structure G(a,b,c) can be defined as follows: a can represent the combination of institutional nodes (including the target business entity and multiple other business entities); b can represent the set of edges connecting the institutions (relationships between institutions, including customer, investor, and supplier relationships); and c can represent the characteristics of the institutional nodes (each institution includes four attributes: registered capital, establishment time, number of legal events, and administrative processing markers). By analyzing the network structure, we obtain the institutional relationship matrix A and institutional characteristic matrix X of the target business entity to be processed in the early warning business within the supply chain.

[0063] In this embodiment, the matrix corresponding to the target business unit is directly obtained, and it can be input into the subsequent model without further processing of the matrix, which saves data preprocessing time and improves the efficiency of early warning.

[0064] S202, input the organizational relationship matrix and organizational feature matrix into the graph neural network model.

[0065] After obtaining the organizational relationship matrix and the organizational feature matrix, the organizational relationship matrix and the organizational feature matrix can be input into the graph neural network model so that the autoencoder module, random walk module and self-training clustering module in the graph neural network model can process them in sequence.

[0066] The graph neural network model includes an autoencoder module, a random walk module, and a self-training clustering module. During the training process of the graph neural network model, the context prediction loss output by the random walk module is used to constrain the model parameters in the autoencoder module and the self-training clustering module.

[0067] In some embodiments, the graph neural network model can be a deep learning model designed to process graph structure (network) data. It is understood that the graph neural network model can process matrices corresponding to graph structures.

[0068] An autoencoder module can be an unsupervised learning component in a model. Its core function is to compress and reconstruct the input data, with the aim of learning a more fundamental, lower-dimensional, dense representation of the data (i.e., latent feature vectors).

[0069] A random walk module can be a component in a model used to capture information about the network structure. This component learns higher-order proximity and contextual relationships between nodes by simulating the process of random walks in a relational network.

[0070] A self-trained clustering module can be a component in a model used to discover inherent groupings (clustering) in the data. During model training, it can autonomously perform soft classification on learned features and calculate the probability for each node that it belongs to a different cluster (such as an outlier).

[0071] Context prediction loss is a numerical metric (loss function) generated by the random walk module. It measures the difference between the node context relationships determined by the model and the actual network structure; the smaller the difference, the better the model's understanding of the network structure.

[0072] The model training process can be represented as an iterative optimization phase in which the model parameters of the autoencoder module, random walk module, and self-training clustering module are automatically adjusted by an algorithm to minimize the loss function and make the model output approach the expected target. It is the core process by which the model learns knowledge and rules.

[0073] Model parameters are variables such as weights and biases that need to be learned and adjusted during the training process.

[0074] In one possible implementation, before inputting the organizational relationship matrix and organizational feature matrix into the graph neural network model, the graph neural network model needs to be trained to obtain a trained graph neural network model. Then, the organizational relationship matrix and organizational feature matrix are input into the trained graph neural network model, and the matrices are processed according to the modules in the trained graph neural network model to obtain the model prediction results.

[0075] In one possible implementation, training the graph neural network model may specifically include: during training, the server obtains the total loss of the graph neural network model; the total loss includes the reconstruction loss of the autoencoder module, the context prediction loss of the random walk module, and the clustering loss of the self-trained clustering module; if the total loss satisfies the model iteration condition, the server backpropagates and updates the model parameters of the autoencoder module, the random walk module, and the self-trained clustering module respectively, until the total loss no longer satisfies the model iteration condition, thus obtaining the trained graph neural network model and its parameters. During the training of the graph neural network model, the context prediction loss output by the random walk module is used to constrain the model parameters in the autoencoder module and the self-trained clustering module.

[0076] In this embodiment, firstly, the input to the graph neural network combines the relationships between the target business unit and other business units, as well as the organizational attribute characteristics of the target business unit in multiple health dimensions. This results in more comprehensive input data, which helps improve the accuracy of the model output and thus the accuracy of the early warning. Secondly, the context prediction loss output by the random walk module constrains the model parameters in the autoencoder module and the self-trained clustering module, making the model parameters corresponding to the modules more reasonable, thereby improving the model's deterministic accuracy.

[0077] S203 transforms the organizational relationship matrix and organizational feature matrix into potential feature vectors of the target business organization through the model parameters of the autoencoder module and the autoencoder module.

[0078] After inputting the organizational relationship matrix and organizational feature matrix into the graph neural network model, the model parameters of the autoencoder module and the autoencoder module are used to transform the organizational relationship matrix and organizational feature matrix into potential feature vectors of the target business organization. This allows the subsequent self-training clustering module to process the potential feature vectors. The model parameters of the graph neural network model trained in the above manner can effectively extract complex nonlinear features, and the extracted nonlinear features are more comprehensive.

[0079] In some embodiments, the model parameters of the autoencoder module can be weights and biases that need to be trained and adjusted during the process of converting input data into latent feature vectors.

[0080] Latent eigenvectors are low-dimensional, dense numerical vectors obtained after input data (such as institutional feature matrices and institutional relationship matrices) has been compressed and abstracted by an autoencoder module. The dimension of the latent eigenvector is smaller than that of the institutional relationship matrix or the institutional feature matrix. Compressing massive, sparse, and potentially noisy matrices into dense, low-dimensional vectors removes redundant information and improves data quality.

[0081] For example, such as Figure 3The flowchart illustrating the method for generating latent feature vectors is shown. In step S203 above, the organizational relationship matrix and organizational feature matrix are transformed into latent feature vectors of the target business organization through the model parameters of the autoencoder module and the autoencoder module. Specifically, this includes: S301, through the self-encoding module, performs self-loop addition and normalization processing on the organizational relationship matrix to obtain the normalized self-loop target matrix corresponding to the organizational relationship matrix.

[0082] After inputting the organizational relationship matrix and organizational feature matrix into the graph neural network model, the autoencoder module performs self-loop addition and normalization processing on the organizational relationship matrix to obtain the normalized self-loop target matrix corresponding to the organizational relationship matrix.

[0083] The target matrix can be a matrix obtained by adding self-loops and normalizing the institutional relationship matrix.

[0084] For example, using the autoencoder module, self-loop processing is first added to the organizational relationship matrix A to obtain A+I (identity matrix), and then A+I is normalized to obtain the normalized target matrix of self-loops corresponding to the organizational relationship matrix. Target matrix The expression is as follows:

[0085] Where D represents a diagonal matrix.

[0086] In this embodiment, adding a self-loop A+I means that each node also considers its own features during information aggregation. Normalization prevents nodes with high degrees from having excessive weight in feature aggregation, making the model determination more reasonable.

[0087] S302 transforms the target matrix and the institutional feature matrix by using shared weights and mean weights to obtain the mean feature vector of the target business institution.

[0088] After obtaining the target matrix, the target matrix and the institutional feature matrix can be transformed by shared weights and mean weights to obtain the mean feature vector of the target business institution, thus achieving efficient variational inference and learning the probability distribution of the data.

[0089] The model parameters of the autoencoder module include: the shared weights of the first layer of the graph neural network, and the mean weights and offset weights of the second layer of the graph neural network.

[0090] In this context, the shared weights of the first layer of the graph neural network can be a set of general parameters (weight matrix) used in the first layer of the model. The role of the shared weights is to perform a unified linear transformation and information extraction on the target matrix and the mechanism feature matrix.

[0091] The mean weights of the second layer in a graph neural network are parameters specifically used to calculate the mean of the latent feature vector. These mean weights determine how the features from the first layer are mapped to the center of the latent space.

[0092] For example, the target matrix is ​​obtained. Then, the target matrix can be adjusted using shared weights and mean weights. The mean eigenvector of the target business entity is obtained by transforming the entity feature matrix X. Mean eigenvector The expression is as follows:

[0093] in, These are the shared weights of the first layer of the graph neural network. The mean weights of the second layer of the graph neural network. It is the hyperbolic tangent function.

[0094] In this embodiment, basic information is efficiently extracted from the original features by sharing weights, reducing the number of parameters and preventing overfitting.

[0095] S303 transforms the target matrix and the institutional feature matrix by sharing weights and offset weights to obtain the offset feature vector of the target business institution.

[0096] The offset weights of the second layer of a graph neural network are parameters specifically used in the second layer to calculate the latent feature vector (variance or uncertainty).

[0097] For example, the target matrix is ​​obtained. Then, by sharing weights and shifting weights, the target matrix can be adjusted. The offset feature vector of the target business unit is obtained by transforming the feature matrix X of the organization. Offset feature vector The expression is as follows:

[0098] in, These are the shared weights of the first layer of the graph neural network. These are the offset weights of the second layer of the graph neural network. It is the hyperbolic tangent function.

[0099] In this embodiment, firstly, by sharing weights, basic information is efficiently extracted from the original features, reducing the number of parameters and preventing overfitting. Secondly, instead of outputting a fixed vector, it outputs parameters of a Gaussian distribution (logarithm of mean and variance). This allows the model to learn the probability distribution of the data, rather than a single fixed value, enhancing its generalization ability.

[0100] S304, sample the Gaussian data containing the mean eigenvector and the offset eigenvector to obtain the potential eigenvector of the target business unit.

[0101] After obtaining the mean eigenvector and the offset eigenvector, the Gaussian data containing the mean eigenvector and the offset eigenvector are sampled to obtain the potential eigenvector of the target business unit.

[0102] Among them, Gaussian data can be data corresponding to a Gaussian distribution constructed based on the mean eigenvector and the offset eigenvector.

[0103] In one possible implementation, the server can construct a Gaussian distribution based on the mean eigenvector and the offset eigenvector. The server then samples from the Gaussian data corresponding to the Gaussian distribution to obtain the latent feature vector of the target business entity.

[0104] For example, after obtaining the mean eigenvector and the offset eigenvector, a Gaussian distribution is constructed based on the mean eigenvector and the offset eigenvector. The server then samples from the Gaussian data corresponding to the Gaussian distribution to obtain the latent feature vector of the target business entity. The expression for the Gaussian distribution is as follows:

[0105] in, It can represent a Gaussian distribution. It can represent the mean eigenvector. It can represent an offset eigenvector, and diag can represent a diagonal matrix. This can represent the latent feature vector of the target business unit. The latent feature vector of the target business unit is obtained by sampling Gaussian data corresponding to a Gaussian distribution. .

[0106] In this embodiment, firstly, the relation matrix is ​​subjected to self-loop addition and normalization to enhance model stability and improve information transmission. Secondly, shared weights and mean / offset weights are used to generate mean and offset feature vectors respectively, efficiently achieving variational inference, learning the probability distribution of the data, and providing a basis for subsequent data sampling of Gaussian data. Thirdly, latent feature vectors are obtained by sampling the Gaussian distribution, introducing randomness. By randomly sampling the learned distribution, slightly different latent vectors can be obtained each time. This introduces beneficial noise into the model, equivalent to a kind of regularization, forcing the latent space to be smoother and more continuous, and improving the model's robustness to noise and small perturbations.

[0107] S204, through the self-trained clustering module and the model parameters of the self-trained clustering module, transforms the latent feature vector into the probability of the target business unit belonging to the reference institution group.

[0108] After obtaining the latent feature vector, the latent feature vector is input into the self-training clustering module. Through the self-training clustering module and its model parameters, the latent feature vector is transformed into the probability of the target business unit corresponding to the reference group. This probability is then used for early warning. The model parameters of the graph neural network model trained in the above manner can effectively extract complex nonlinear features, and the extracted nonlinear features are more comprehensive.

[0109] In some embodiments, the model parameters of the self-training clustering module can be weights and biases that need to be trained and adjusted in the process of transforming input data (such as latent feature vectors) into the probability of the target business entity belonging to the reference group of institutions.

[0110] A reference group of institutions can be a collection of institutions that share certain common anomalous characteristics or business patterns, formed through model learning or business definition.

[0111] The attribution probability is the likelihood that a target business entity belongs to a specific group of reference entities, represented by a value between 0 and 1. Specifically, a higher attribution probability for a target business entity to a group of reference entities indicates a higher probability that the target business entity is abnormal, while a lower attribution probability indicates a lower probability that the target business entity is abnormal.

[0112] For example, such as Figure 4 The flowchart illustrating the method for generating attribution probabilities is shown above. In step S204, the latent feature vector is transformed into the attribution probability of the target business entity corresponding to the reference entity group through the self-trained clustering module and its model parameters. Specifically, this includes: S401, input the latent feature vector into the self-training clustering module, and determine the difference between the latent feature vector and the target organization center vector as the first difference.

[0113] After obtaining the latent feature vector, the latent feature vector is input into the self-training clustering module, and the difference between the latent feature vector and the target organization center vector is determined as the first difference.

[0114] The model parameters of the self-training clustering module include: the target mechanism center vector, multiple reference mechanism center vectors including the target mechanism center vector, and the mechanism degrees of freedom. The mechanism center vector and each reference mechanism center vector belong to the reference mechanism group. The target organization center vector can represent the vector corresponding to the probability of the organization having a high level. The closer the latent feature vector is to the target organization center vector, the higher the probability that the target business organization corresponding to the latent feature vector has a high level.

[0115] For example, the server obtains the latent feature vector. Then, the server can generate latent feature vectors. Input into the self-trained clustering module and the latent feature vector With the center vector of the target institution The difference between them was determined to be the first difference ( ).

[0116] In this embodiment of the application, the first difference can measure the probability of the target business unit having a high level, which makes it easier to convert the first difference into a quantitative attribution probability in the future.

[0117] S402, the difference between the latent eigenvector and the center vector of each reference mechanism is determined as the second difference.

[0118] After obtaining the latent eigenvectors, the differences between the latent eigenvectors and the center vectors of each reference mechanism are determined as the second differences.

[0119] The multiple reference institution center vectors can include vectors corresponding to high-level, medium-level, low-level, and no-level institutions. Similarly, the closer the latent feature vector is to the reference institution center vector, the higher the probability that the target business institution corresponding to the latent feature vector possesses a certain level (high-level, medium-level, low-level, or no-level). The multiple reference institution center vectors include the target institution center vector.

[0120] For example, the server obtains the latent feature vector Then, the latent feature vectors Each reference mechanism center vector The differences between them are defined as the second difference ( ).

[0121] In this embodiment of the application, multiple second differences can measure the degree of matching between the target organization and all other possible clusters, that is, the probability of the target business organization's existence at a certain level (high level, medium level, low level and no level), which facilitates the subsequent conversion of the second differences into a quantitative attribution probability.

[0122] S403, according to the institutional degrees of freedom, transform the first difference and each of the second differences to obtain the institutional affiliation probability of the target business institution corresponding to the reference institutional group.

[0123] After obtaining the first and second differences, the first and second differences are transformed according to the institutional degrees of freedom to obtain the institutional affiliation probability of the target business institution corresponding to the reference institutional group.

[0124] The degrees of freedom of a mechanism are key parameters that represent the shape of a cluster distribution. The degrees of freedom determine how tightly the features of the mechanism are distributed around the central vector.

[0125] For example, the server obtains the first difference ( ) and the second difference ( After that, the server operates according to the degree of institutional freedom. Regarding the first difference ( ) and each second difference ( The transformation is performed to obtain the probability of the target business unit belonging to the reference group of institutions. Institutional affiliation probability The expression is as follows:

[0126] Among them, degrees of freedom The size can be determined based on the actual training of the model.

[0127] In this embodiment, firstly, the first and second differences can reflect the degree of matching between the target institution and all other possible clusters, facilitating the subsequent conversion of the second difference into a quantified attribution probability. Secondly, based on the statistical distribution of degrees of freedom, all the first and second differences are transformed into a probability distribution. Compared to hard clustering with either / or approaches, this provides a more refined probability distribution, and the probability value itself represents the confidence level, making the early warning decision more interpretable.

[0128] S205: When the probability of attribution is greater than or equal to the preset probability, an early warning is issued to the target business unit.

[0129] Once the attribution probability is obtained, an early warning can be issued to the target business unit if the attribution probability is greater than or equal to the preset probability.

[0130] In some embodiments, the preset probability may be a threshold configured for the attribution probability.

[0131] In one possible implementation, the preset probability can be adjusted based on the tolerance for business anomalies, thus achieving controllability of the warning sensitivity. The server can obtain the tolerance for business anomalies. When the tolerance for business anomalies is high, the server lowers the preset probability to expand the monitoring range; when the tolerance for business anomalies is low, the server increases the preset probability to reduce false alarm interference.

[0132] In this embodiment, the organizational relationship matrix and organizational feature matrix of the target business entity to be processed in the early warning operation are obtained within the supply chain. The supply chain includes interconnected target business entities and multiple other business entities. The organizational relationship matrix characterizes the relationships between each target business entity and each of the other business entities. The organizational feature matrix includes the organizational attribute characteristics of each target business entity across multiple health dimensions. Directly obtaining the matrix corresponding to the target business entity eliminates the need for further matrix processing before inputting it into the subsequent model, saving data preprocessing time and improving the efficiency of early warning.

[0133] The organizational relationship matrix and organizational feature matrix are input into a graph neural network (GNN) model, which includes an autoencoder module, a random walk module, and a self-training clustering module. During the training process of the GNN model, the context prediction loss output by the random walk module is used to constrain the model parameters in the autoencoder and self-training clustering modules. Firstly, the input of the GNN combines the relationships between the target business organization and other business organizations, as well as the organizational attribute characteristics of the target business organization across multiple health dimensions. This provides more comprehensive input data, which helps improve the accuracy of the model output and thus the accuracy of early warnings. Secondly, the context prediction loss output by the random walk module constrains the model parameters in the autoencoder and self-training clustering modules, making the model parameters more reasonable and thus improving the model's deterministic accuracy.

[0134] By utilizing the model parameters of the autoencoder and auto-encoder modules, the organizational relationship matrix and organizational feature matrix are transformed into latent feature vectors for the target business organizations. The dimensionality of these latent feature vectors is smaller than that of either the organizational relationship matrix or the organizational feature matrix. Firstly, the autoencoder module compresses massive, sparse, and potentially noisy matrices into dense, low-dimensional vectors, removing redundant information and improving data quality. Secondly, the latent feature vectors reflect the deep fusion result of the organizational relationship matrix and the organizational feature matrix, providing data support for subsequent processing.

[0135] By using a self-trained clustering module and its model parameters, the latent feature vectors are transformed into the probability of a target business entity corresponding to a reference group of institutions. Firstly, the output is no longer a binary judgment of whether something is abnormal; instead, the probability of belonging is quantified, making the early warning more interpretable. Secondly, since the probability of belonging is the probability of a target business entity corresponding to a reference group of institutions, the probability output using the reference group as a reference target is more accurate, thus improving the accuracy of subsequent early warnings.

[0136] When the probability of attribution is greater than or equal to a preset probability, an early warning is issued to the target business unit. Firstly, the preset probability is flexibly configurable, ensuring controllable warning sensitivity. Secondly, issuing a timely warning to the target business unit before any actual anomalies occur can prevent losses.

[0137] This method, on the one hand, combines the relationships between the target business unit and other business units, as well as the organizational attribute characteristics of the target business unit in multiple health dimensions, into the input of the graph neural network. The input data is more comprehensive, which helps to improve the accuracy of the model output and thus the accuracy of the early warning. On the other hand, the context prediction loss output by the random walk module constrains the model parameters in the autoencoder module and the self-training clustering module, making the model parameters corresponding to the modules more reasonable, thereby improving the determination accuracy of the model and the accuracy of the early warning for the target business unit.

[0138] Before inputting the organizational relationship matrix and organizational feature matrix into the graph neural network model, the graph neural network model needs to be trained. In one embodiment, the training method of the graph neural network model includes: During training, the total loss of the graph neural network model is obtained. If the total loss satisfies the model iteration condition, backpropagation is performed to update the model parameters of the autoencoder module, random walk module, and self-training clustering module until the total loss no longer satisfies the model iteration condition, thus obtaining the trained graph neural network model and its parameters.

[0139] In some embodiments, the total loss may be the sum of the losses generated in a single iteration of the graph neural network model. The total loss includes the reconstruction loss of the autoencoder module, the context prediction loss of the random walk module, and the clustering loss of the self-trained clustering module.

[0140] The reconstruction loss of the autoencoder module can represent the error in decoding the latent feature vectors back to the original input data (mechanical features).

[0141] The context prediction loss of the random walk module can represent the error between the contextual relationships determined by the model and the actual supply chain network.

[0142] The clustering loss of a self-trained clustering module can represent the distribution quality of latent feature vectors in the clustering space.

[0143] In one possible implementation, the model iteration condition can be that the total loss is greater than the preset loss. When the total loss is greater than the preset loss, backpropagation is performed and the model parameters of the autoencoder module, the random walk module, and the self-trained clustering module are updated until the total loss is less than or equal to the preset loss, thus obtaining the trained graph neural network model and model parameters.

[0144] In one possible implementation, the model iteration condition can be that the number of model iterations is less than a preset number, and the number of model iterations corresponding to the total loss is less than a preset number. In this case, backpropagation is performed to update the model parameters of the autoencoder module, the random walk module, and the self-trained clustering module until the number of model iterations corresponding to the total loss is greater than or equal to the preset number, thus obtaining the trained graph neural network model and model parameters.

[0145] In this embodiment, firstly, the reconstruction loss, context prediction loss, and clustering loss are summed into a single total loss, enabling the model to optimize three key objectives simultaneously in a single training iteration. This avoids objective conflicts or suboptimal solutions that may arise from phased training, ensuring higher accuracy in the model's output attribution probabilities. Secondly, a clear training stopping criterion is provided when the total loss is less than a preset loss or the corresponding number of model iterations reaches a preset number. This process is fully automated, ultimately outputting a well-trained, parameter-fixed optimal model, thus ensuring the efficiency of model training to a certain extent.

[0146] During the training of a graph neural network model, it is necessary to train the autoencoder module, which involves optimizing the reconstruction loss. In one embodiment, such as... Figure 5 The flowchart shown illustrates the optimization method for reconstruction loss. The graph neural network model method also includes: S501, obtain training samples for the early warning service and model parameters of the autoencoder module.

[0147] In some embodiments, training samples can be used as samples in the early warning system for training graph neural network models.

[0148] An adjacency matrix can be a matrix that describes the relationships between nodes in each business unit.

[0149] Node relationships can include customer relationships, investor relationships, and supplier relationships.

[0150] A node feature matrix can be a matrix that describes the node attribute characteristics of each business unit across multiple health dimensions.

[0151] Node attribute features include registered capital, establishment time, number of legal events, and administrative processing markers.

[0152] The model parameters of the autoencoder module include the first weight of the first layer of the graph neural network, the second mean weight of the second layer of the graph neural network, and the second offset weight of the second layer of the graph neural network.

[0153] The first weight of the first layer of a graph neural network can be a set of general parameters used in the first layer of the model during the training phase.

[0154] The second mean weight of the second layer of the graph neural network can be a parameter in the second layer specifically used to calculate the latent feature vector (mean).

[0155] The second offset weights of the second layer of a graph neural network can be parameters in the second layer specifically used to calculate the latent feature vector (offset).

[0156] For example, training samples for early warning services are obtained, including an adjacency matrix A and a node feature matrix X, and the first weights of the first layer of the graph neural network are obtained. The second mean weight of the second layer of the graph neural network The second offset weight of the second layer of the graph neural network .

[0157] In this embodiment, the input of the graph neural network combines the node relationships between business units and the node attribute features of each business unit in multiple health dimensions, resulting in higher quality input data and higher accuracy of the trained model and subsequent early warning.

[0158] S502, perform self-loop addition and normalization on the adjacency matrix to obtain the candidate matrix of normalized self-loops corresponding to the adjacency matrix.

[0159] After obtaining the adjacency matrix, the adjacency matrix is ​​processed by adding self-loops and normalizing to obtain the candidate matrix of normalized self-loops corresponding to the adjacency matrix, so that the candidate matrix can be transformed in the future.

[0160] For example, using an autoencoder module, self-loop processing is first added to the adjacency matrix A to obtain A+I (identity matrix), and then A+I is normalized to obtain the normalized candidate matrix of self-loops corresponding to the adjacency matrix A. Candidate matrix The expression is as follows:

[0161] Where D represents a diagonal matrix.

[0162] In this embodiment, adding a self-loop A+I means that each node also considers its own features during information aggregation. Normalization prevents nodes with high degrees from having excessive weight in feature aggregation, making the model prediction more reasonable.

[0163] S503, according to the first weight and the second mean weight, transform the candidate matrix and the node feature matrix to obtain the mean node matrix corresponding to each business unit.

[0164] After obtaining the first weight and the second mean weight, the candidate matrix and the node feature matrix are transformed according to the first weight and the second mean weight to obtain the mean node matrix corresponding to each business unit.

[0165] For example, the candidate matrix is ​​obtained. Then, the candidate matrix can be evaluated using the first weight and the second mean weight. By transforming the node feature matrix X, we obtain the mean node matrix corresponding to each business unit. Mean node matrix The expression is as follows:

[0166] in, This is the first weight of the first layer of the graph neural network. The second mean weight of the second layer of the graph neural network. It is the hyperbolic tangent function.

[0167] In this embodiment, basic information is efficiently extracted from the original features by sharing weights, reducing the number of parameters and preventing overfitting.

[0168] S504. Based on the first weight and the second offset weight, the candidate matrix and the node feature matrix are transformed to obtain the offset node matrix corresponding to each business unit.

[0169] After obtaining the first weight and the second offset weight, the candidate matrix and the node feature matrix are transformed according to the first weight and the second offset weight to obtain the offset node matrix corresponding to each business unit.

[0170] For example, the candidate matrix is ​​obtained. Then, the candidate matrix can be adjusted using the first weight and the second offset weight. The node feature matrix X is transformed to obtain the offset node matrix corresponding to each business unit. Offset node matrix The expression is as follows:

[0171] in, This is the first weight of the first layer of the graph neural network. This refers to the second offset weight of the second layer of the graph neural network. It is the hyperbolic tangent function.

[0172] In this embodiment, firstly, by sharing weights, basic information is efficiently extracted from the original features, reducing the number of parameters and preventing overfitting. Secondly, instead of outputting a fixed vector, it outputs parameters of a Gaussian distribution (logarithm of mean and variance). This allows the model to learn the probability distribution of the data, rather than a single fixed value, enhancing its generalization ability.

[0173] S505 samples the Gaussian data containing the mean node matrix and the offset node matrix to obtain the potential node vectors for each business unit.

[0174] After obtaining the mean node matrix and the offset node matrix, the Gaussian data containing the mean node matrix and the offset node matrix are sampled to obtain the potential node vectors of each business unit.

[0175] Among them, Gaussian data can be data corresponding to a Gaussian distribution constructed based on the mean node matrix and the offset node matrix.

[0176] In one possible implementation, the server can construct a Gaussian distribution based on the mean node matrix and the offset node matrix. The server then samples from the Gaussian data corresponding to the Gaussian distribution to obtain the potential node vectors for each business unit.

[0177] For example, after obtaining the mean node matrix and the offset node matrix, a Gaussian distribution is constructed based on the mean node matrix and the offset node matrix. The server then samples from the Gaussian data corresponding to the Gaussian distribution to obtain the potential node vectors of the target business organization. The expression for the Gaussian distribution is as follows:

[0178] in, It can represent a Gaussian distribution. It can represent a mean node matrix. It can represent an offset node matrix, and diag can represent a diagonal matrix. This can represent the potential node vector of the target business unit. The potential node vector of the target business unit is obtained by sampling Gaussian data corresponding to a Gaussian distribution. .

[0179] In this embodiment, potential node vectors are obtained by sampling from a Gaussian distribution, introducing randomness. By randomly sampling the learned distribution, slightly different potential vectors can be obtained each time. This introduces beneficial noise into the model, which is equivalent to a form of regularization, forcing the latent space to be smoother and more continuous, thus improving the model's robustness to noise and small perturbations.

[0180] S506, based on the vectors of each potential node and the relationships between nodes, obtains the reconstruction loss of the autoencoder module.

[0181] After obtaining the potential node vectors and node relationships, the reconstruction loss of the autoencoder module can be obtained based on each potential node vector and node relationship.

[0182] For example, the server can base its actions on the vectors of each potential node. Relationship with nodes The reconstruction loss of the autoencoder module is constructed. Reconstruction loss The expression is as follows:

[0183] in, For the expectation operator, the expectation is based on the distribution of the encoder output. It's calculated. We're not using a fixed... Instead, consider All possible values ​​of , and according to their probability of occurrence (by (Given) a weighted average.

[0184] It is understandable that the encoder output distribution can be calculated based on the Gaussian distribution. After obtaining the Gaussian distribution, the variational approximation probability distribution of the nodes can be determined based on the Gaussian distribution. Nodal variational approximation probability distribution The expression is as follows:

[0185] Then obtain the nodal variational approximation probability distribution Then, the probability distribution can be approximated based on the variational distribution of the nodes. Approximate representation of the encoder output distribution The distribution of encoder output The expression is as follows:

[0186] in, This represents the decoder likelihood; the higher the value, the stronger the decoder's ability to reconstruct the real image. (The last part, "decoder likelihood," appears to be a typo and can be left as is.) The factorization expression is as follows:

[0187] Specifically, the expansion of the decoder's likelihood after factorization. It can be used express, Let be the conditional probability distribution of an edge between a specific pair of nodes (i,j). In other words, Given an edge between two specific firms, the conditional probability distribution is given. The expression is as follows:

[0188] in, It can represent the conditional probability that there is an edge between a specific pair of nodes (i,j). It can be determined using the likelihood function (re-weight) formula of the Bernoulli distribution:

[0189] Among them, node relationships This can represent the node relationship between node i and node j, where there is an edge between node i and node j. =1, does not exist It is 0. Let be the potential node vector of node i. Let be the potential node vector of node j, and T be the transpose operation. This is the activation function.

[0190] In this embodiment, firstly, self-loops are added to the adjacency matrix and normalized to enhance model stability and improve information transmission. Secondly, the first weight and the second mean / offset weights are used to generate mean and offset feature matrices respectively, efficiently achieving variational inference and learning the probability distribution of the data, providing a basis for subsequent data sampling of Gaussian data. Thirdly, latent node vectors are obtained by sampling the Gaussian distribution, introducing randomness. By randomly sampling the learned distribution, slightly different latent node vectors can be obtained each time. This introduces beneficial noise into the model, equivalent to a regularization, forcing the latent space to be smoother and more continuous, improving the model's robustness to noise and small perturbations. Fourthly, the autoencoding process of the graph neural network is placed within the variational Bayesian framework. Through a Bernoulli distribution-based loss function aimed at edge reconstruction, the model learns a high-quality latent feature representation that reflects the node's own attributes, accurately encodes its network structure role, and is probabilistically robust. This provides information-rich, structure-aware, and robust feature inputs for subsequent clustering and anomaly warning tasks.

[0191] During the training of a graph neural network model, the random walk module needs to be trained, which involves optimizing the context loss. In one embodiment, the method for the graph neural network model further includes: obtaining the model parameters of the random walk module, including the window size; performing a random walk on each business unit in the adjacency matrix according to the window size to obtain multiple node sequences; for each node sequence, obtaining the sequence prediction loss based on the node attribute features corresponding to the preceding unit, the node attribute features corresponding to the following unit, the node relationship between the preceding unit and the intermediate unit, and the node relationship between the following unit and the intermediate unit; and obtaining the context prediction loss of the random walk module based on each sequence prediction loss.

[0192] In some embodiments, the window size can represent the length of the sequence output by the random walk module. A larger window size results in a longer output sequence and more nodes. Conversely, a smaller window size results in a shorter output sequence and fewer nodes.

[0193] A node sequence can be a sequence containing several business units as nodes. A node sequence includes intermediate nodes (intermediate units), preceding nodes (preceding units), and following nodes (following units). The number of nodes in each node sequence corresponds to the window size. For example, if the window size is 3, the output node sequence will contain one intermediate node, one preceding node, and one following node.

[0194] In one possible implementation, for each node sequence, the sequence prediction loss corresponding to the node sequence is obtained based on the node attribute features corresponding to the preceding organization, the node attribute features corresponding to the following organization, the node relationship between the preceding organization and the intermediate organization, and the node relationship between the following organization and the intermediate organization.

[0195] For example, after obtaining the node sequence, it can be treated as a sentence and input into the SkipGram language model. This model can transform the node attribute features corresponding to the preceding and following structures, the node relationships between the preceding and following structures, and the node relationships between the following and intermediate structures into the sequence prediction loss corresponding to the node sequence. After obtaining the sequence prediction loss corresponding to the node sequence, the sequence prediction losses corresponding to multiple node sequences can be added, averaged, or weighted averaged to obtain the context prediction loss of the random walk module. .

[0196] In this embodiment, firstly, the model can flexibly control the scope of context information capture during random walks. The window size determines how far away from the central node a node in the sequence will be considered as context. Secondly, node sequences simulating information or abnormal propagation paths are dynamically generated from the original static adjacency matrix. Multiple sequences provide rich samples of local network structure. Thirdly, by aggregating the individual losses generated by a large number of random walk sequences, an accurate context prediction loss is obtained.

[0197] During the training of a graph neural network model, it is necessary to train the self-training clustering module, which is to optimize the clustering. In one embodiment, such as... Figure 6 The flowchart shown illustrates the optimization method for clustering loss. The training method for the graph neural network model also includes: S601, determine multiple initial cluster centers from the potential node vectors of each business unit.

[0198] After obtaining the potential node vectors of each business unit, the potential node vectors can be clustered to determine multiple initial cluster centers.

[0199] In some embodiments, the initial cluster center can be In one possible implementation, after the server obtains the potential node vectors of each business unit, it can perform K-means clustering on the potential node vectors to determine multiple initial cluster centers.

[0200] In this embodiment, instead of randomly assigning or manually setting centers, the initial centers are found directly based on the low-dimensional latent features learned by the model that integrate attributes and structure. This ensures that the initial centers are located in dense areas of the true data distribution, providing a high-quality starting point for subsequent probabilistic clustering.

[0201] S602, if each initial cluster center satisfies the K-means clustering termination condition, determine the target organization center vector and at least one reference organization center vector containing the target organization center vector from the potential node vectors corresponding to each initial cluster center.

[0202] After obtaining multiple initial cluster centers, it can be determined whether each initial cluster center satisfies the K-means clustering termination condition. If each initial cluster center satisfies the K-means clustering termination condition, the target organization center vector and at least one reference organization center vector containing the target organization center vector can be determined from the potential node vectors corresponding to each initial cluster center.

[0203] In some embodiments, In one possible implementation, if the initial cluster centers do not meet the K-means clustering termination condition, the server can re-perform K-means clustering based on the initial cluster centers and the potential node vectors of each business unit to obtain new cluster centers. This process continues until the new cluster centers meet the K-means clustering termination condition or the number of clustering iterations reaches a preset number. From the potential node vectors corresponding to the cluster centers of the last clustering, the target organization center vector and at least one reference organization center vector containing the target organization center vector are determined.

[0204] For example, at each initial cluster center Given that the K-means clustering termination condition is met, the latent node vectors corresponding to each initial cluster center are... In the process, determine the center vector of the target mechanism. and at least one reference mechanism center vector containing the target mechanism center vector. .

[0205] In this embodiment of the application, in each training iteration, the model will run K-means clustering once on the current latent feature space to dynamically update the target institution center vector (the cluster center to which the current target belongs) and all reference institution center vectors (the centers of all clusters).

[0206] S603 generates the attribution probability distribution for each business unit based on the vectors of each potential node, the center vector of the target institution, the center vectors of each reference institution, and the initial degrees of freedom.

[0207] After obtaining the vectors of each potential node, the center vector of the target mechanism, the center vectors of each reference mechanism, and the initial degrees of freedom, we can proceed based on the vectors of each potential node. Target organization center vector , center vectors of each reference mechanism and initial degrees of freedom Generate the attribution probability distribution (also known as the student t-distribution) for each business unit.

[0208] For example, the expression for the attribution probability distribution corresponding to each business unit is as follows:

[0209] In this embodiment, firstly, the first and second differences can reflect the degree of matching between the target institution and all other possible clusters, facilitating the subsequent conversion of the second difference into a quantified attribution probability distribution. Secondly, based on the statistical distribution of degrees of freedom, all the first and second differences are transformed into a single probability distribution. Compared to hard clustering with either / or approaches, this provides a more refined probability distribution, and the probability value itself represents the confidence level, making the early warning decision more interpretable.

[0210] S604, Determine the target distribution based on the attribution probability distribution.

[0211] After obtaining the attribution probability distribution, the target distribution is determined based on the attribution probability distribution and the probabilities of different k values.

[0212] In some embodiments, the probabilities of different k values ​​can represent the probabilities of a business entity when the probability index is k, i.e. .

[0213] Example, target distribution The expression is as follows:

[0214] Among them, when the initial degrees of freedom are greater than 1, the concentration of the target distribution is greater than the concentration of the belonging probability distribution.

[0215] In this embodiment, firstly, by combining the initial degrees of freedom to generate the target distribution corresponding to the membership probability distribution, richer membership information is provided (e.g., "70% belong to high-level groups, 30% belong to unlevel groups"), which is more valuable for early warning decisions. Secondly, the initial degrees of freedom are parameters that adjust the degree of distribution concentration. The lower the degrees of freedom, the more dispersed the target distribution, that is, the thicker the tail of the target distribution, the less sensitive it is to outliers, and the more robust the clustering; conversely, it is more stringent. This provides the model with noise resistance and flexibility.

[0216] S605 calculates the KL divergence of the target distribution and the attribution probability distribution to obtain the clustering loss of the self-trained clustering module.

[0217] In some embodiments, KL divergence can be a metric that measures the degree of difference between two probability distributions. It describes the amount of information lost when one of the distributions (such as the approximate distribution Q) is used to approximate the true distribution P.

[0218] For example, the target distribution is obtained. and Attribution Probability Distribution Then, the clustering loss of the self-trained clustering module The corresponding expression for calculating the KL divergence is shown below:

[0219] In one example, such as Figure 7 The training process data flow diagram shown illustrates the involvement of an autoencoder module, a random walk module, and a self-training clustering module. The autoencoder module generates the mean node matrix based on the node feature matrix X and the adjacency matrix A. and offset node matrix Then, based on the mean node matrix... and offset node matrix The process involves generating a latent node vector Z, followed by a reconstruction loss based on this vector. A random walk module generates a context loss based on the node feature matrix X and adjacency matrix A. A self-training clustering module generates a clustering loss based on the latent node vector Z. The total loss includes the reconstruction loss from the autoencoder module, the context prediction loss from the random walk module, and the clustering loss from the self-training clustering module. Backpropagation is performed to update the model parameters of the autoencoder, random walk, and self-training clustering modules, provided the total loss meets the model iteration conditions, until the total loss no longer meets the iteration conditions, resulting in a trained graph neural network model and its parameters. This training process can specifically target the training of a graph neural network model for micro and small enterprises. Specifically, during training, the autoencoder module embeds the topological structure G(a,b,c) and attribute information of the micro and small enterprise network structure into a low-dimensional representation (Z). The random walk module then extracts the context information from the network structure, fully exploring the multi-layered, deep implicit information within the network structure. Finally, early warnings are issued based on the embedded feature results for the micro and small enterprises.

[0220] In this embodiment, firstly, the KL divergence loss, as a self-generated supervisory signal, guides feature learning to optimize in a direction that makes the clustering results more distinct and deterministic, thereby greatly improving the cohesion and separability of the clusters. Secondly, the KL clustering loss is backpropagated, simultaneously optimizing the feature encoder parameters and the cluster center vector, thus achieving optimization of the self-trained clustering module.

[0221] Figure 8 This is a schematic diagram of the structure of an early warning device provided in an embodiment of this application. Figure 8 As shown, the early warning device includes: The system includes a matrix acquisition module 801, a model processing module 802, a matrix processing module 803, a parameter processing module 804, and an early warning module 805.

[0222] The matrix acquisition module 801 is used to acquire the organizational relationship matrix and organizational feature matrix of the target business entities to be processed in the early warning business in the supply chain. The supply chain includes interconnected target business units and multiple other business units. The organizational relationship matrix is ​​used to characterize the relationships between the target business units and each of the other business units. The organizational characteristic matrix includes the organizational attribute characteristics of the target business units in multiple health dimensions. The model processing module 802 is used to input the organizational relationship matrix and the organizational feature matrix into the graph neural network model; wherein, the graph neural network model includes an autoencoder module, a random walk module and a self-training clustering module; during the training process of the graph neural network model, the context prediction loss output by the random walk module is used to constrain the model parameters in the autoencoder module and the self-training clustering module; The matrix processing module 803 is used to transform the organizational relationship matrix and the organizational feature matrix into potential feature vectors of the target business organization through the model parameters of the autoencoder module and the autoencoder module; wherein the dimension of the potential feature vector is smaller than the dimension of the organizational relationship matrix or the organizational feature matrix. The parameter processing module 804 is used to convert the latent feature vector into the probability of the target business unit belonging to the reference institution group through the self-trained clustering module and the model parameters of the self-trained clustering module. The early warning module 805 is used to issue an early warning to the target business unit when the probability of attribution is greater than or equal to the preset probability.

[0223] In other embodiments, the model parameters of the autoencoder module include: shared weights of the first layer of the graph neural network, mean weights and offset weights of the second layer of the graph neural network; the model processing module 802 is further configured to perform self-loop addition processing and normalization processing on the organizational relationship matrix through the autoencoder module to obtain a normalized target matrix of self-loops corresponding to the organizational relationship matrix; transform the target matrix and the organizational feature matrix through shared weights and mean weights to obtain the mean feature vector of the target business organization; transform the target matrix and the organizational feature matrix through shared weights and offset weights to obtain the offset feature vector of the target business organization; and sample Gaussian data containing the mean feature vector and the offset feature vector to obtain the latent feature vector of the target business organization.

[0224] In other embodiments, the model parameters of the self-training clustering module include: a target organization center vector, multiple reference organization center vectors including the target organization center vector, and organization degrees of freedom, wherein the organization center vector and each reference organization center vector belong to a reference organization group; the matrix processing module 803 is further used to input the latent feature vector into the self-training clustering module, and determine the difference between the latent feature vector and the target organization center vector as a first difference; determine the difference between the latent feature vector and each reference organization center vector as a second difference; and transform the first difference and each second difference according to the organization degrees of freedom to obtain the organization affiliation probability of the target business organization corresponding to the reference organization group.

[0225] In other embodiments, the aforementioned early warning device may further include a model training module, used to obtain the total loss of the graph neural network model during the training process; the total loss includes the reconstruction loss of the autoencoder module, the context prediction loss of the random walk module, and the clustering loss of the self-trained clustering module; when the total loss satisfies the model iteration condition, backpropagation is performed and the model parameters of the autoencoder module, the random walk module, and the self-trained clustering module are updated until the total loss no longer satisfies the model iteration condition, thereby obtaining the trained graph neural network model and model parameters.

[0226] In other embodiments, the model training module is further configured to acquire training samples for the early warning service and model parameters of the autoencoder module. The training samples include an adjacency matrix and a node feature matrix. The adjacency matrix includes multiple business units and the node relationships between them. The node feature matrix includes the node attribute features of each business unit across multiple health dimensions. The model parameters of the autoencoder module include the first weight of the first layer of the graph neural network and the second mean weight and second offset weight of the second layer of the graph neural network. The adjacency matrix is ​​processed by adding self-loops and normalizing to obtain a candidate matrix of normalized self-loops corresponding to the adjacency matrix. The candidate matrix and node feature matrix are transformed according to the first weight and the second mean weight to obtain the mean node matrix corresponding to each business unit. The candidate matrix and node feature matrix are transformed according to the first weight and the second offset weight to obtain the offset node matrix corresponding to each business unit. Gaussian data containing the mean node matrix and the offset node matrix are sampled to obtain the potential node vectors for each business unit. Based on the potential node vectors and node relationships, the reconstruction loss of the autoencoder module is obtained.

[0227] In other embodiments, the model training module is further configured to obtain model parameters of the random walk module, including the window size; perform random walk processing on each business unit in the adjacency matrix according to the window size to obtain multiple node sequences, each node sequence including an intermediate node as an intermediate unit, an upper node of the preceding unit as an intermediate unit, and a lower node of the following unit as an intermediate unit, the number of nodes in each node sequence corresponding to the window size; for each node sequence, obtain the sequence prediction loss corresponding to the node sequence based on the node attribute features corresponding to the preceding unit, the node attribute features corresponding to the following unit, the node relationship between the preceding unit and the intermediate unit, and the node relationship between the following unit and the intermediate unit; and obtain the context prediction loss of the random walk module based on each sequence prediction loss.

[0228] In other embodiments, the model training module is further configured to determine multiple initial cluster centers in the potential node vectors of each business unit; if each initial cluster center satisfies the K-means clustering termination condition, determine the target organization center vector and at least one reference organization center vector containing the target organization center vector in the potential node vectors corresponding to each initial cluster center; generate the belonging probability distribution corresponding to each business unit based on each potential node vector, the target organization center vector, each reference organization center vector, and the initial degrees of freedom; determine the target distribution based on the belonging probability distribution; if the initial degrees of freedom are greater than 1, the concentration of the target distribution is greater than the concentration of the belonging probability distribution; and perform KL divergence calculation on the target distribution and the belonging probability distribution to obtain the clustering loss of the self-trained clustering module.

[0229] The early warning device provided in this application embodiment can execute the method shown in the above method embodiment. Its implementation principle and beneficial effects can be referred to the relevant description in the method embodiment, and will not be repeated here.

[0230] Figure 9 This is a schematic diagram of the structure of an early warning device provided in an embodiment of this application. Figure 9 As shown, the early warning device includes: a memory 901, a transceiver 902, and at least one processor 903.

[0231] The transceiver 902 is used to interact with other devices to send and receive data. For example, in this embodiment, the transceiver 902 can specifically be used to send and receive mechanism relationship matrices and mechanism feature matrices.

[0232] The memory 901 is used to store computer program code, which includes computer instructions. These computer instructions run in the aforementioned warning device to implement the method shown in the above-described method embodiments. For example, the memory may include high-speed random access memory (RAM), and may also include non-volatile memory (NVM), such as at least one disk storage device, or a USB flash drive, portable hard drive, read-only memory, disk, or optical disc, etc.

[0233] Processor 903 can be a general-purpose processor, including a Central Processing Unit (CPU), a network processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. Processor 903 can also be other general-purpose processors. The general-purpose processor can be a microprocessor or any conventional processor.

[0234] The memory 901, transceiver 902, and processor 903 are communicatively connected. For example, the memory 901 and transceiver 902 can be connected to the processor 903 via a system bus to communicate with each other. The system bus can be a peripheral component interconnect (PCI) bus, an extended industry standard architecture (EISA) bus, an industry standard architecture (ISA) bus, etc. The system bus can be divided into address bus, data bus, control bus, etc. For ease of representation, only one thick line is used in the figure, but this does not mean that there is only one bus or one type of bus.

[0235] Optionally, the memory 901 can be either standalone or integrated with the processor 903. When the memory 901 is set up independently, it is connected to the processor 903 via the system bus.

[0236] This application also provides a chip for executing instructions, which is used to execute the early warning method described in the above embodiments.

[0237] This application also provides a computer-readable storage medium storing computer instructions. When these computer instructions are executed by a processor, they are used to implement the technical solution of the early warning method described in the above embodiments. Specifically, when the computer instructions are executed by a processor, the early warning device can perform the technical solution of the early warning method described in the above embodiments.

[0238] This application also provides a computer program product, which includes a computer program stored in a computer-readable storage medium. At least one processor can read the computer program from the computer-readable storage medium, and when the at least one processor executes the computer program, it can implement the technical solution of the early warning method in the above embodiments.

[0239] The aforementioned computer-readable storage media can be implemented from any type of volatile or non-volatile storage device or a combination thereof, such as Static Random-Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The computer-readable storage media can be any available medium accessible to a general-purpose or special-purpose computer.

[0240] An exemplary computer-readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the storage medium. Of course, the computer-readable storage medium can also be a component of the processor. The processor and the computer-readable storage medium can reside in an application-specific integrated circuit (ASIC). Alternatively, the processor and the computer-readable storage medium can exist as discrete components in an electronic control unit or main control device; this application does not limit this.

[0241] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or modules, and may be electrical, mechanical, or other forms.

[0242] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to implement the solution of this embodiment according to actual needs.

[0243] Furthermore, the functional modules in the various embodiments of this application can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit. The unit composed of the above modules can be implemented in hardware or in the form of hardware plus software functional units.

[0244] The integrated modules described above, implemented as software functional modules, can be stored in a computer-readable storage medium. These software functional modules, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute some steps of the methods of the various embodiments of this application.

[0245] It should be understood that the steps of the method disclosed in the embodiments of this application can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules in the processor.

[0246] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0247] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. An early warning method, characterized in that, The method includes: Obtain the organizational relationship matrix and organizational characteristic matrix of the target business entities to be processed in the early warning business within the supply chain; The supply chain includes interconnected target business entities and multiple other business entities. The organizational relationship matrix is ​​used to characterize the relationship between each target business entity and each of the other business entities. The organizational feature matrix includes the organizational attribute features of each target business entity in multiple health dimensions. The organizational relationship matrix and the organizational feature matrix are input into a graph neural network model; wherein, the graph neural network model includes an autoencoder module, a random walk module, and a self-training clustering module; during the training process of the graph neural network model, the context prediction loss output by the random walk module is used to constrain the model parameters in the autoencoder module and the self-training clustering module; The autoencoder module and its model parameters are used to transform the organizational relationship matrix and the organizational feature matrix into potential feature vectors of the target business organization; wherein the dimension of the potential feature vectors is smaller than the dimension of either the organizational relationship matrix or the organizational feature matrix. The latent feature vector is transformed into the probability of the target business unit belonging to the reference group of institutions through the self-trained clustering module and the model parameters of the self-trained clustering module. If the probability of attribution is greater than or equal to a preset probability, an early warning will be issued to the target business unit.

2. The method according to claim 1, characterized in that, The model parameters of the autoencoder module include: the shared weights of the first layer of the graph neural network, and the mean weights and offset weights of the second layer of the graph neural network; The step of transforming the organizational relationship matrix and the organizational feature matrix into the latent feature vector of the target business organization through the model parameters of the autoencoder module and the autoencoder module includes: The self-encoding module is used to add self-loops and normalize the organizational relationship matrix to obtain the normalized target matrix of self-loops corresponding to the organizational relationship matrix. The target matrix and the institutional feature matrix are transformed using the shared weight and the mean weight to obtain the mean feature vector of the target business institution. The target matrix and the organization feature matrix are transformed using the shared weights and the offset weights to obtain the offset feature vector of the target business organization; The Gaussian data containing the mean feature vector and the offset feature vector are sampled to obtain the potential feature vector of the target business entity.

3. The method according to claim 1, characterized in that, The model parameters of the self-training clustering module include: the target mechanism center vector, multiple reference mechanism center vectors including the target mechanism center vector, and mechanism degrees of freedom, wherein the mechanism center vector and each of the reference mechanism center vectors belong to the reference mechanism group; The step of converting the latent feature vector into the probability of a target business entity belonging to a reference group of entities using the self-trained clustering module and its model parameters includes: The latent feature vector is input into the self-training clustering module, and the difference between the latent feature vector and the target organization center vector is determined as the first difference. The difference between the potential feature vector and the center vector of each of the reference mechanisms is defined as the second difference; Based on the institutional degrees of freedom, the first difference and each of the second differences are transformed to obtain the institutional affiliation probability of the target business institution corresponding to the reference institutional group.

4. The method according to claim 1, characterized in that, The training method for the graph neural network model includes: During training, the total loss of the graph neural network model is obtained; the total loss includes the reconstruction loss of the autoencoder module, the context prediction loss of the random walk module, and the clustering loss of the self-training clustering module. If the total loss satisfies the model iteration condition, backpropagation is performed to update the model parameters of the autoencoder module, the random walk module, and the self-training clustering module until the total loss no longer satisfies the model iteration condition, thus obtaining the trained graph neural network model and model parameters.

5. The method according to claim 4, characterized in that, The method further includes: Obtain training samples for the early warning service and model parameters of the autoencoder module; The training samples include an adjacency matrix and a node feature matrix. The adjacency matrix includes multiple business units and the node relationships between them. The node feature matrix includes the node attribute features of each business unit in multiple health dimensions. The model parameters of the autoencoder module include the first weight of the first layer of the graph neural network and the second mean weight and the second offset weight of the second layer of the graph neural network. The adjacency matrix is ​​subjected to self-loop addition and normalization processing to obtain the normalized candidate matrix of self-loops corresponding to the adjacency matrix; The candidate matrix and the node feature matrix are transformed according to the first weight and the second mean weight to obtain the mean node matrix corresponding to each business unit. According to the first weight and the second offset weight, the candidate matrix and the node feature matrix are transformed to obtain the offset node matrix corresponding to each business unit; The Gaussian data containing the mean node matrix and the offset node matrix are sampled to obtain the potential node vectors of each of the business units. Based on the potential node vectors and the node relationships, the reconstruction loss of the autoencoder module is obtained.

6. The method according to claim 5, characterized in that, The method further includes: Obtain the model parameters of the random walk module, including the window size; Random walk processing is performed on each of the business units in the adjacency matrix according to the window size to obtain multiple node sequences. The node sequence includes an intermediate node as an intermediate unit, an upper node as an upper-text unit of the intermediate unit, and a lower node as a lower-text unit of the intermediate unit. The number of nodes contained in each node sequence corresponds to the window size. For each node sequence, based on the node attribute features corresponding to the preceding organization, the node attribute features corresponding to the following organization, the node relationship between the preceding organization and the intermediate organization, and the node relationship between the following organization and the intermediate organization, the sequence prediction loss corresponding to the node sequence is obtained. Based on the sequence prediction loss, the context prediction loss of the random walk module is obtained.

7. The method according to claim 5, characterized in that, The method further includes: Multiple initial cluster centers are determined from the potential node vectors of each of the aforementioned business entities; If each of the initial cluster centers satisfies the K-means clustering termination condition, then from the potential node vectors corresponding to each of the initial cluster centers, determine the target organization center vector and at least one reference organization center vector containing the target organization center vector. Based on the potential node vectors, the target institution center vectors, the reference institution center vectors, and the initial degrees of freedom, the affiliation probability distributions corresponding to each of the business institutions are generated. The target distribution is determined based on the attribution probability distribution; when the initial degrees of freedom are greater than 1, the concentration of the target distribution is greater than the concentration of the attribution probability distribution. The KL divergence of the target distribution and the attribution probability distribution is calculated to obtain the clustering loss of the self-trained clustering module.

8. An early warning device, characterized in that, The device includes: The matrix acquisition module is used to acquire the organizational relationship matrix and organizational feature matrix of the target business entities to be processed in the early warning business within the supply chain. The supply chain includes interconnected target business entities and multiple other business entities. The organizational relationship matrix is ​​used to characterize the relationship between each target business entity and each of the other business entities. The organizational feature matrix includes the organizational attribute features of each target business entity in multiple health dimensions. A model processing module is used to input the organization relationship matrix and the organization feature matrix into a graph neural network model; wherein, the graph neural network model includes an autoencoder module, a random walk module, and a self-training clustering module; during the training process of the graph neural network model, the context prediction loss output by the random walk module is used to constrain the model parameters in the autoencoder module and the self-training clustering module; A matrix processing module is used to transform the organizational relationship matrix and the organizational feature matrix into potential feature vectors of the target business organization using the model parameters of the autoencoder module and the autoencoder module; wherein the dimension of the potential feature vectors is smaller than the dimension of the organizational relationship matrix or the organizational feature matrix. The parameter processing module is used to convert the latent feature vector into the probability of the target business unit corresponding to the reference institution group through the self-trained clustering module and the model parameters of the self-trained clustering module. The early warning module is used to issue an early warning to the target business unit when the attribution probability is greater than or equal to a preset probability.

9. An early warning device, characterized in that, include: A memory and at least one processor; the memory is communicatively connected to the processor; the memory is used to store computer program code, the computer program code including computer instructions; when the processor executes the computer instructions, it causes the early warning device to perform the method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that, when executed by a processor, are used to implement the method as described in any one of claims 1-7.

11. A computer program product, characterized in that, When the computer program product is run on a computer / executed by the computer's processor, it implements the method as described in any one of claims 1-7.

Images