Enterprise risk early warning method and system based on multi-source data fusion

Abstract

The invention relates to the field of data management, in particular to an enterprise risk early warning method and system based on multi-source data fusion. Comprising the following steps: receiving an internal document and a knowledge base, extracting an enterprise data set, and constructing a dynamic knowledge graph; in the influence propagation model based on the time-space diagram neural network, executing multi-layer iterative message passing based on a double-layer attention mechanism, and generating a health state vector; inputting the health state vector into a multi-task decoding framework, and purifying the features of each risk task through an advisory depolarization network to generate a risk specific vector; enabling the risk specificity vector to pass through a plurality of task solution terminals to obtain a risk estimation value; and the influence propagation model is updated according to the difference value between the risk estimation value and the real loss through a conflict resolution optimizer based on gradient projection. According to the method, systematic risk opening quantification is carried out through a double-layer attention mechanism, and pseudo-correlation interference in multi-source data is reduced through an adversarial depolarization network.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of data management, in particular to an enterprise risk early warning method and system based on multi-source data fusion. BACKGROUND

[0002] In computer science, graph structure is a general data model for representing entities and their complex relationships. The data among many real-world enterprises is more suitable to be modeled as heterogeneous graphs due to the inclusion of multiple types of nodes and edges. However, effective representation learning on heterogeneous graphs faces a series of recognized technical challenges. Existing methods usually rely on artificially predefined "meta-paths" or network patterns to guide the aggregation of information among different types of nodes, but are difficult to adapt to dynamically changing graph structures or new analysis tasks. Moreover, real-world graph data, such as financial networks or social networks, are continuously and dynamically changing, with new nodes and edges appearing constantly. An effective graph representation learning system must be able to handle this dynamic nature and support inductive learning, i.e., generating representations for new nodes that did not appear during training.

[0003] Graph Neural Network (GNN) is a standard paradigm for processing graph-structured data. However, to handle the heterogeneous graph resulting from the complex relationships among enterprises, when GNN expands the receptive field of nodes by stacking multiple layers, the representation vectors of nodes inevitably tend to homogenization, eventually converging to a value that cannot be distinguished, resulting in the loss of expression ability of the deep model.

[0004] In summary, there is a need to design a deep graph neural network architecture that can dynamically construct a unified representation from heterogeneous multi-modal data while maintaining node heterogeneity, enhancing the precision and stability of the early warning model.

[0005] To this end, an enterprise risk early warning method and system based on multi-source data fusion are proposed. SUMMARY

[0006] The present application aims to provide an enterprise risk early warning method and system based on multi-source data fusion to solve the problems raised in the background.

[0007] To achieve the above-mentioned purpose, the present application provides the following technical solution: an enterprise risk early warning method based on multi-source data fusion, comprising: Receiving internal documents and knowledge base, extracting data assets, risk events, enterprise relationships and health indicators to form an enterprise data group; using the enterprise data group to construct a dynamic knowledge graph; In the influence propagation model based on the spatio-temporal graph neural network, multi-layer iterative message passing based on a double-layer attention mechanism is performed on the dynamic knowledge graph, and the state information of each node is updated according to the state information of adjacent nodes and the attributes of directed edges to generate a health state vector; The health state vector is input into a multi-task decoding framework containing a pseudo-relevance suppression mechanism, a risk-specific vector is generated by purifying the features of each risk task through an adversarial debiasing network, and a risk estimate is obtained by passing the risk-specific vector through multiple task decoding heads; the difference between the risk estimate and the real loss is updated by a conflict resolution optimizer based on gradient projection to update the influence propagation model.

[0008] Preferably, the specific process of receiving internal documents and knowledge bases includes: For internal documents, a natural language processing pipeline is applied to identify data assets and risk events and capture dynamic semantic links between data assets and between data assets and risk events; the natural language processing pipeline integrates named entity recognition, event extraction, entity linking and document-level relationship extraction; for knowledge bases, a document parsing pipeline is used to extract health indicators through layout analysis and table parsing, and known corporate relationships containing holdings and guarantees are identified from the knowledge bases to form a corporate data group.

[0009] Preferably, the specific process of constructing a dynamic knowledge graph includes: Receiving a corporate data group, linking data assets and risk events to unique identifiers in a knowledge base, and instantiating a corresponding node in a graph database for each data asset and risk event; health indicators are attached to the corresponding node as key-value pairs; entity relationships, corporate relationships and relationships between data assets and risk events are converted into directed edges connecting corresponding node pairs, and the type of the directed edge is defined by the nature of the relationship.

[0010] Preferably, the specific generation process of the health state vector includes: For each data asset node in the dynamic knowledge graph, the initial state information at the zeroth layer is defined as a feature vector composed of a health state vector; for a preset number of propagation levels, multi-layer iterative message passing for perceiving spatio-temporal influence is performed for each data asset node at each layer; the state information of each data asset node at the last propagation level is designated as the health state vector, wherein the health state vector encodes systematic influences from the topology and node attributes within the neighborhood and quantifies potential chain reactions.

[0011] Preferably, the specific process of multi-layer iterative message passing includes: In the relationship layer, the adjacent nodes are grouped according to the types of the directed edges; for each group, a relationship-level attention weight is calculated based on the state information of the center node and preset attributes and the relationship type; in the entity layer, for each adjacent node in each group, an entity-level attention weight is calculated based on the current state information of the center node and the adjacent node; the state information of each adjacent node is weighted by the corresponding entity-level attention weight and then twice weighted by the relationship-level attention weight of the group to which the adjacent node belongs; all the message vectors after the double weighting are aggregated into a comprehensive message vector; the state information of the data asset node at the previous level is combined with the comprehensive message vector, and the combined result is input into a nonlinear update function to generate the state information of the data asset node at the current level.

[0012] Preferably, the specific generation process of the risk-specific vector comprises: Pseudo-correlations between different risk task labels are identified through statistical analysis; for each main risk task, a main predictor and an adversarial network are constructed; the main predictor receives a health state vector and generates an intermediate feature representation; the adversarial network receives the intermediate feature representation and predicts a confusion task that has pseudo-correlation with the main risk; through adversarial training, the parameters of the main predictor are updated to minimize the prediction loss of itself while maximizing the prediction loss of the adversarial network, thereby generating a risk-specific vector that has removed pseudo-correlation information.

[0013] Preferably, the specific generation process of the risk-specific vector comprises: The health state vector is classified to generate a context feature; the risk-specific vector and the context feature are spliced to generate an augmented feature vector; the augmented feature vector is input in parallel to multiple task decoding heads, each of which corresponds to a risk dimension, and outputs a risk estimate for the risk dimension.

[0014] Preferably, the specific process of updating the influence propagation model comprises: The risk estimate is compared with the corresponding true loss to generate a loss value; for each loss value, a set of gradient vectors about the shared parameters of the influence propagation model is generated by calculating the gradient through a backpropagation algorithm; when a conflict is detected between the gradient vectors of any two tasks, at least one of the conflicting gradient vectors is adjusted to reduce the conflict component with the other gradient vector, thereby generating a modified gradient vector; the shared parameters of the influence propagation model are updated using all the modified parameter update vectors.

[0015] An enterprise risk early warning system based on multi-source data fusion, comprising: Multi-source data module: receive internal documents and knowledge base, extract data assets, risk events, enterprise relationships and health indicators, and form enterprise data groups; use the enterprise data groups to build a dynamic knowledge graph; Influence propagation module: in the influence propagation model based on the space-time graph neural network, multi-layer iterative message passing based on a double-layer attention mechanism is performed on the dynamic knowledge graph, and the state information of each node is updated according to the state information of adjacent nodes and the attributes of directed edges to generate a health state vector; Multi-task module: the health state vector is input into a multi-task decoding framework containing a pseudo-correlation suppression mechanism, the features of each risk task are purified through an adversarial debiasing network to generate a risk-specific vector; the risk-specific vector is input into multiple task decoding heads to obtain a risk estimate; and the difference between the risk estimate and the real loss is updated by a conflict resolution optimizer based on gradient projection to update the influence propagation model.

[0016] Compared with the prior art, the beneficial effects of the present application are: 1. In the message passing process of the space-time graph neural network, the weight differences of different business relationships in risk transmission are learned and distinguished in the relationship layer, and the influence of key neighbor nodes is identified in the entity layer. This design makes the model no longer only identify the existence of risks, but can dynamically learn and quantify the strength of risk transmission from data. These dynamically weighted influences are encoded into the health state vector, thereby converting complex risk interactions and transmission strengths into measurable numerical values. Thus, a node representation with higher information density and better resistance to "over-smoothing" effect is generated. At the same time, the two separable attention weights provide a technical interpretability for the model, allowing to trace back and understand the specific contributions of different connection types and neighbor nodes to the final vector representation of a node.

[0017] 2. By constructing a main predictor and an adversarial network and performing adversarial training, the model is forced to generate a risk-specific vector that eliminates most pseudo-correlated information, and the pseudo-correlation interference between different risk categories in multi-source data is reduced, ensuring that the evaluation of a single risk is based on its core driving factors rather than being contaminated by signals from other risk dimensions. Thus, independent and high-precision quantification of each risk category can be achieved, improving the robustness and generalization ability of the model.

[0018] 3. By combining the adversarial debiasing mechanism with the gradient projection-based conflict resolution optimizer, the optimization framework ensures that the model can stably and efficiently learn how to independently assess risks and how to quantify the dynamic transmission between risks. The adversarial network ensures the independence of each risk assessment at the feature level, and during the model training process, the gradient projection optimizer also adjusts the conflicting gradient vectors to solve the negative interference that may occur when updating the shared parameters of these risk tasks. This dual mechanism of decoupling at the feature level and collaboration at the optimization level creates a more stable training environment in terms of computation, improves the stability of model training, accelerates the convergence speed, and improves the utilization efficiency of computing resources. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 A flowchart of an enterprise risk early warning method based on multi-source data fusion is provided for the embodiments of the present application. Figure 2 A transmission schematic diagram of an influence propagation model is provided for the embodiments of the present application. Figure 3 A structural diagram of an enterprise risk early warning system based on multi-source data fusion is provided for the embodiments of the present application. DETAILED DESCRIPTION

[0020] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0021] Please refer to Figures 1-3 The enterprise risk early warning method based on multi-source data fusion provided by the present application has the following specific steps: Internal documents and knowledge bases are received, data assets, risk events, enterprise relationships and health indicators are extracted, and an enterprise data group is formed; a dynamic knowledge graph is constructed using the enterprise data group; In the influence propagation model based on the spatiotemporal graph neural network, a multi-layer iterative message passing based on a double-layer attention mechanism is performed on the dynamic knowledge graph, the state information of each node is updated according to the state information of adjacent nodes and the attributes of directed edges, and a health state vector is generated; The health state vector is input into a multi-task decoding framework containing a pseudo-correlation inhibition mechanism, a risk-specific vector is generated by purifying the features of each risk task through an adversarial debiasing network, the risk-specific vector is input into multiple task decoding heads to obtain a risk estimate, and the difference between the risk estimate and the actual loss is used to update the influence propagation model through a conflict resolution optimizer based on gradient projection.

[0022] The technical solutions of the present application will be further described in detail below in combination with specific embodiments.

[0023] Embodiment one The embodiment one of the present application discloses an enterprise risk early warning method and system based on multi-source data fusion, referring to Figure 1 The specific steps include: S1, receiving internal documents and knowledge bases, extracting data assets, risk events, enterprise relationships and health indicators to form an enterprise data group; using the enterprise data group, a dynamic knowledge graph is constructed; S2, in an influence propagation model based on a spatiotemporal graph neural network, a multi-layer iterative message passing based on a double-layer attention mechanism is performed on the dynamic knowledge graph to generate a health state vector; S3, the health state vector is input into a multi-task decoding framework containing a pseudo-correlation inhibition mechanism, a risk-specific vector is generated by purifying the features of each risk task through an adversarial debiasing network; S4, the risk-specific vector is input into multiple task decoding heads to obtain a risk estimate; S5, the difference between the risk estimate and the actual loss is used to update the influence propagation model through a conflict resolution optimizer based on gradient projection.

[0024] Further, internal documents are received to extract data assets, risk events and entity relationships, knowledge bases are received to extract health indicators and enterprise relationships to form an enterprise data group; using the enterprise data group, a dynamic knowledge graph is constructed; corresponding to the above S1 step; the specific implementation process includes: This embodiment adopts a double-pipeline data processing architecture to fuse multi-source heterogeneous data. Pipeline A, i.e. the natural language processing (NLP) pipeline, is responsible for processing unstructured internal documents such as market analysis and research reports. This pipeline is composed of multiple operation units connected in series. A named entity recognition (NER) unit is applied to identify and classify data assets and risk events (such as "credit downgrade" and "takeover offer") from text based on a Transformer model (such as FinBERT) fine-tuned on a financial corpus. An event extraction unit further identifies the attributes of the event, such as the participants, time and location. A relation extraction (RE) module captures the dynamic semantic relationships between identified entities, for example, extracting the relationship triple (Company A)-[acquires]->(Company B) from "Company A announces acquisition of Company B". An entity linking module links the extracted entities (such as International Business Machines Corporation) to unique identifiers in the knowledge base to eliminate ambiguity.

[0025] Pipeline B, i.e., the document parsing pipeline, is responsible for processing semi-structured knowledge bases, especially financial statements (e.g., annual reports) in PDF format. The pipeline first applies a document layout analysis (DLA) algorithm to segment the PDF pages into logical regions, such as text blocks, titles, and tables. A table detection and extraction module (e.g., using the pdfplumber library) locates and parses the financial statements, extracting health indicators (e.g., asset-liability ratio, current ratio, net profit) and their corresponding numerical values. Meanwhile, the pipeline identifies declared corporate relationships (e.g., holding, guarantee) from the report's footnotes or specific chapters.

[0026] All the information extracted by the two pipelines, including data assets, risk events, entity relationships, health indicators, and corporate relationships, collectively form the corporate data group. This data group is injected into a native graph database (e.g., Neo4j or TigerGraph) to construct a dynamic knowledge graph. In this process, each unique data asset and risk event is instantiated as a node. The extracted health indicators are attached to the corresponding company nodes as key-value pairs. All relationships (including semantic relationships and corporate relationships) are converted into typed directed edges connecting the corresponding node pairs (e.g., IS_SUPPLIER_TO, HAS_GUARANTEE). The knowledge graph is dynamically updated as new data is continuously ingested, reflecting changes in the financial market in real time.

[0027] By adopting a dual-pipeline data processing architecture, the present invention can collaboratively extract high-frequency, event-driven qualitative risk signals (e.g., risk events, semantic relationships between entities) from internal documents and high-precision, quantitative health indicators from knowledge bases. Subsequently, by uniformly instantiating these heterogeneous information as nodes, properties, and typed edges in a dynamic knowledge graph, the present invention transforms scattered, disordered data into a structured, machine-readable knowledge network that reflects the complex corporate ecosystem. This approach not only greatly enriches the data dimensions of risk analysis, but also lays a comprehensive and solid data foundation for subsequent systemic risk propagation analysis.

[0028] Further, in the influence propagation model based on the spatio-temporal graph neural network, a multi-layer iterative message passing based on a double-layer attention mechanism is performed on the dynamic knowledge graph, and state information of each node is updated according to state information of adjacent nodes and properties of directed edges to generate a health state vector; corresponding to the S2 step; refer to Figure 2 ; the specific process includes: The embodiment adopts a space-time graph neural network (STGNN) as the core of the influence propagation model, and the theoretical basis is that financial risk is not a static and isolated attribute, but a dynamic propagation process in a complex network composed of enterprises. The STGNN architecture can capture the spatial dimension (defined by the network topology of the knowledge graph) and the time dimension (defined by the time series characteristics of the enterprise operating indicators) of risk transmission. Specifically, a space-time graph neural network architecture coupled with a graph convolution network (GCN) and a gated recurrent unit (GRU) is adopted, which realizes accurate capture of the space-time dynamics of risk by working together in each propagation level (or time step). To achieve this coupling, the model follows the calculation process of “spatial aggregation first, then time update” in each propagation level. Specifically, first, the spatial information of all neighbor nodes of the center node is aggregated by the graph convolution network module to generate a temporary aggregation message vector; then, the aggregation message vector and the state vector (i.e., the hidden state) of the center node at the previous level are input into the gated recurrent unit module. The GRU module fuses the newly transmitted spatial risk information with the historical state information of the node itself through its internal update gate and reset gate mechanism, calculates and outputs the new state vector of the node at the current level. This structure ensures that the model can simultaneously consider the influence of network topology structure and the time evolution of risk state at each iteration step.

[0029] The importance of information in the knowledge graph decays over time. Before message passing, a time decay function, i.e., an exponential decay function, is applied to the feature vectors of all adjacent nodes and the attributes of the relationship edges where is the time difference from the record time point to the current calculation time point, is a learnable decay rate parameter, and its initial value is randomly selected in the range of [0.01, 0.1]; and for specific types of risk event nodes in the knowledge graph, when created and updated, a temporary impact factor I is generated according to the preset risk event table within a short time, temporarily increasing the weight of the directly associated data asset node in the model, and gradually decreasing over time. This design simulates the immediate and amplified response of the financial market to major negative news, enabling the model to capture short-term and dramatic risk fluctuations triggered by sudden events, rather than just long-term and gradual risk transmission. In the embodiment, the risk event type lookup table can be used to obtain the factor (for example, default event I=5.0, regulatory investigation I=3.0), which acts on the features of the associated nodes within T time steps (for example, T=7 days) after the event occurs. After the above time and event dimension enhancement processing, the “message” transmitted by each node to its neighbors not only contains its own operating state, but also contains the timeliness of its information and the immediate risk brought by sudden events, providing the space-time graph neural network with more rich and dynamic inputs.

[0030] The core of the model is a multi-layer iterative message passing mechanism. For each data asset node in the unprocessed financial network D in the knowledge graph, its initial state (0th layer) is defined as a feature vector composed of its health state vector. Subsequently, the model iteratively updates the state of each node through a preset number (e.g., L = 3) of propagation layers. At each layer in the spatial dimension, the state information of a central node is a function of its own state in the previous layer and the "messages" aggregated from its adjacent nodes. This process allows the model to capture indirect influences from within its L-hop neighborhood, effectively simulating the cascading effects of risk. The typical value range of the propagation layer L is 2 to 5 layers, and too few layers cannot capture long-distance influences, and too many layers are prone to over-smoothing problems.

[0031] The embodiment uses a double-layer attention mechanism to guide the weighting and aggregation of messages. For relationship-level attention, before aggregating neighbor information, the model first groups neighbors according to the type of edges, concatenates the feature representation of the central node with the embedding vector of the relationship type; then, the concatenated vector is linearly transformed through a learnable weight matrix and a LeakyReLU activation function is applied to obtain an original attention score; finally, a normalization function called "Softmax" is used to process the original attention scores of all relationship types of the node, obtaining the final relationship-level attention weight. This weight allows the model to learn the importance differences of different business relationships (such as "guarantee" and "supply") in risk transmission.

[0032] For entity-level attention, within each relationship group, the model further calculates the attention weight for each independent neighbor node, concatenates the feature representation of the central node with that of its neighbor nodes; then, the concatenated vector is linearly transformed through a learnable weight matrix and a LeakyReLU activation function is applied; then, the transformed result is dot multiplied with a learnable context vector to obtain an original entity attention score; finally, the Softmax normalization function is also used to process the original scores of all neighbor nodes within the relationship group, obtaining the final entity-level attention weight. This weight allows the model to dynamically focus on those neighbor nodes with worse operating status and higher "infectivity".

[0033] In the message aggregation and update process, the state information of each neighbor node is first weighted by its corresponding entity-level attention weight, then the weighted information of all nodes in the same relationship group is aggregated, and then the aggregated information is weighted again by the relationship-level attention weight of the group. All the double-weighted message vectors are finally integrated to update the state of the center node. The update process of the node can be summarized as follows: first, the state of each neighbor node at the last layer is linearly transformed, and the two-layer attention weights calculated above are used for double weighting to form the aggregated neighbor message; second, the state of the center node itself at the last layer is independently linearly transformed; finally, the model updates the node state using a gated recurrent unit (GRU). The neighbor message and the state vector of the center node itself at the last layer are input into a GRU unit. The GRU unit contains a reset gate and an update gate, which can intelligently decide how much historical information to retain and how much new external risk impact to integrate.

[0034] Specifically, the update of the node state follows the following calculation logic: The reset gate determines the degree of reception of the previous state to the current new information; the update gate controls the proportion of the previous state information brought into the current state; the candidate state calculates the new information that should be paid attention to at the current time; and the final output new state combines the historical information and the current new information. This update process generates the new state of the current layer by weighting and combining the state information of the last layer and the candidate state information of the current layer under the precise control of the update gate. Specifically, the update gate determines how much proportion of the past information should be retained and how much proportion of the current new information should be integrated, thereby realizing the dynamic update of the information.

[0035] After L-layer propagation, the state information of each data asset node at the last layer is designated as the "health state vector". This vector is a dense, learned low-dimensional representation that not only contains the financial status of the node itself, but more importantly, encodes the systemic impact determined by the topological structure and node attributes from its network neighborhood, thereby quantifying the potential domino effect risk of the entity. For example, as shown in Figure 2 not only focuses on the most critical risk transmission path to the center enterprise A, but also quantifies that node B is a more critical risk transmission source than node C.

[0036] The influence propagation model based on the spatio-temporal graph neural network adopted by the application can highly realistically simulate the conduction process of financial risks in a complex network. The relationship level attention mechanism enables the model to learn and distinguish the importance differences of different types of financial relationships in risk conduction; and the entity level attention mechanism enables the model to dynamically focus on neighbor nodes in the network that have worse risk conditions and are more "contagious". Through this refined message passing, the health status vector generated by the model is no longer a simple aggregation of the characteristics of the enterprise itself, but a quantification of the accumulated systemic risk exposure, so as to more accurately predict potential crises triggered by chain reactions.

[0037] Further, the health status vector is input into a multi-task decoding framework containing a pseudo-correlation inhibition mechanism, and a risk-specific vector is generated by an adversarial debiasing network for purifying the features of each risk task; corresponding to the S3 step; the specific implementation process includes: After generating the health status vector containing systemic influence, the embodiment inputs it into an advanced multi-task learning (MTL) decoding framework. The goal of the framework is to predict multiple different but possibly related financial risks (such as credit risk, market risk, operational risk) simultaneously using this shared, high-dimensional feature representation. The embodiment adopts a pseudo-correlation inhibition mechanism based on an adversarial debiasing network (Adversarial Debiasing Network), the core idea of which is to actively "purify" a dedicated feature representation for each risk task at the feature level without interference from other tasks.

[0038] Before constructing the adversarial network, it is necessary to identify the pseudo-correlation between different risk task labels. The application quantifies the correlation by calculating the Pearson correlation coefficient between each risk task label (such as credit risk, market risk) in the training data set. When the absolute value of the correlation coefficient of any two task labels is greater than a preset threshold (for example, the embodiment sets it to 0.5), it is determined that there is a pseudo-correlation between them that needs to be decoupled.

[0039] For each primary risk task (e.g., credit risk prediction), the system constructs two subnetworks, a primary predictor and an adversarial network. Specifically, the primary predictor is a three-layer multi-layer perceptron (MLP) with network structure [256, 128, 64] and ReLU activation function, which takes a 256-dimensional health state vector as input and outputs a 64-dimensional intermediate feature representation. The adversarial network is a two-layer MLP with network structure [64, 32] and ReLU activation function, which takes the 64-dimensional intermediate feature generated by the primary predictor as input. The primary predictor receives the health state vector and generates an intermediate feature representation, whose goal is to accurately predict the primary risk using this representation. The adversarial network receives the same intermediate feature representation, but its goal is to predict a confounding task (e.g., market risk) that is spuriously correlated with the primary risk. The entire training process is designed as a "zero-sum game". This is achieved by inserting a gradient reversal layer between the primary predictor and the adversarial network. The overall loss function of the model is composed of the prediction loss of the primary task and the prediction loss of the adversarial task. This layer behaves as an identity transformation in forward propagation, but in backward propagation, it multiplies the gradient from the adversarial network by a negative constant -γ. Specifically, the overall loss is equal to the prediction loss (e.g., cross-entropy loss) of the primary risk task minus the prediction loss (e.g., cross-entropy loss) of the adversarial task weighted by the hyperparameter γ, which is set to 0.5 in this embodiment. By optimizing this overall loss function, the model minimizes the prediction error of the primary task while maximizing the prediction error of the adversarial task. The optimization goal of the primary predictor is dual, minimizing its own loss of predicting the primary risk (credit risk) and maximizing the loss of the adversarial network predicting the confounding risk (market risk). Through this adversarial training, the primary predictor is forced to learn a feature representation that is highly predictive of the primary risk but contains as little confounding risk information as possible. In other words, in order to "fool" the adversarial network and make it unable to accurately predict the market risk, the primary predictor must eliminate all market risk-related signals from the intermediate feature it generates.

[0040] This purified feature vector, which has removed the spurious correlation information, is the risk-specific vector defined in the present invention. This process is repeated for each risk dimension that needs to be predicted, ultimately generating a set of mutually orthogonal and decoupled risk-specific vectors.

[0041] All risk-specific vectors can also be input into a risk synergistic fusion network; the core of the risk synergistic fusion network is a fusion model based on bilinear pooling, the parameters of the fusion model are specially trained to learn the interaction mode between different risk vectors; by learning the complex nonlinear interaction between different risk types, the impact of mutual amplification and offsetting risk is completed. The network will be used to enhance the input of the subsequent decoding head, so that the final risk assessment not only considers each independent risk dimension, but also quantifies the additional risk exposure generated by their interaction.

[0042] In this way, the embodiment ensures that the subsequent risk valuation is based on the most pure and direct driving factors, greatly improving the robustness and accuracy of the model, and reducing the misjudgment caused by accidental correlation in the data.

[0043] Further, the risk-specific vector is obtained by a plurality of task decoding heads to obtain a risk estimate; corresponding to the S4 step; the specific implementation process includes: The risk-specific vector is converted into the final specific risk estimate available for business decision-making. In order to utilize the specific information of the risk in the decoding stage and not completely lose the overall systemic risk background of the enterprise, the embodiment adopts a feature enhancement strategy. Specifically, the system inputs the 256-dimensional "health state vector" generated by the spatio-temporal graph neural network without adversarial debiasing into an auxiliary feedforward neural network composed of two fully connected layers to generate a set of high-level "context features". Specifically, the structure of the auxiliary network is [256, 128], that is, the input layer has 256 neurons, the hidden layer has 128 neurons, and ReLU is used as the activation function. The 128-dimensional vector output by the network is the "context feature", which can be understood as a highly condensed summary of the overall health status of the enterprise. Subsequently, for each risk task, the system splices (Concatenation) the corresponding risk-specific vector and the shared "context feature" to form an enhanced feature vector.

[0044] These enhanced feature vectors are input in parallel into a plurality of task-specific decoding heads (Task-Specific Decoder Heads). Each decoding head is an independent, relatively simple feedforward neural network, and its parameters are specially trained for the prediction target of a specific risk dimension. Since each decoding head is independent, it can learn the optimal nonlinear transformation of the enhanced feature vector to its specific output space without being disturbed by the learning process of other tasks.

[0045] In this embodiment, the credit risk decoding head is used to predict the default probability of an enterprise in the next year, which is a binary classification problem. The decoding head is composed of a two-layer fully connected network (e.g., input dimension from 64 to 1), and the last layer adopts a Sigmoid activation function to ensure that the output value is between 0 and 1, representing the default probability.

[0046] The market risk decoding head is used to predict the expected volatility of the stock price of an enterprise in the future, which is a regression problem. The decoding head can also be composed of a two-layer fully connected network, and the last layer adopts a ReLU or Softplus activation function to ensure that the output volatility prediction value is non-negative.

[0047] The operational risk decoding head is used to predict the level of operational risk (e.g., "low", "medium", "high"), which is a multi-classification problem. The decoding head is composed of a two-layer fully connected network (e.g., input dimension from 64 to 3), and the last layer adopts a Softmax activation function to output a probability distribution of three risk levels.

[0048] Through this parallel and dedicated decoding mechanism, this embodiment can efficiently and accurately convert the abstract and high-dimensional systemic risk representation into a multi-dimensional and operable enterprise risk portrait, thereby providing direct data support for the business activities of financial institutions.

[0049] Further, the difference between the risk valuation and the actual loss is updated by a conflict resolution optimizer based on gradient projection to update the influence propagation model; corresponding to the above S5 step; the specific implementation process includes: Compare the output of multi-task learning (MTL) with the real-world results and effectively back-propagate the resulting error signals to update and optimize the shared influence propagation model (i.e., the space-time graph neural network) as the core of the system.

[0050] After obtaining the risk estimates of each risk dimension (e.g. credit risk, market risk), they are compared with the corresponding real historical data (e.g. whether a default has occurred in history, the actual stock price volatility) to calculate an independent loss value for each task. In the backpropagation phase of model training, the loss of each task will calculate a gradient vector about the shared parameters that affect the propagation model. However, in standard multi-task learning, a serious challenge is the Conflicting Gradients. Conflicts occur when the gradients of different tasks point in different or even opposite directions. Mathematically, if the dot product (or cosine similarity) of two task gradient vectors g_i and g_j is negative, it is determined that they are in conflict. If these conflicting gradients are directly averaged and used to update the model parameters, it may lead to the performance of one task improving at the expense of the performance of another task, a phenomenon known as "negative transfer", which ultimately hinders the overall convergence and performance of the model.

[0051] To systematically solve this problem, the embodiment adopts a conflict resolution optimizer based on gradient projection. The optimizer performs a "gradient surgery" before each parameter update, and for any two task gradient vectors g_i and g_j, the optimizer calculates their dot product. If g_i · g_j < 0, it is determined that they are in conflict. Once a conflict is detected, a projection operation is performed to eliminate the conflicting component. Taking the modified gradient g_i as an example, first calculate the projection vector of g_i in the direction of g_j, which is calculated by the dot product of g_i and g_j divided by the square of the modulus of g_j, and then multiplied by the vector g_j. Then, subtract this calculated projection vector from the original gradient g_i to get the modified gradient g_i. This operation is geometrically equivalent to projecting g_i onto the normal plane of g_j, thereby removing its conflicting part opposite to g_j. The same operation is performed on the gradient g_j to eliminate its conflict with g_i. The modified gradient vector retains the tendency to advance in the optimal direction of each task while eliminating direct conflicts with each other. The sum of all conflict resolution modified gradient vectors is used to update the shared parameters that affect the propagation model.

[0052] To achieve more intelligent optimization, the embodiment estimates the uncertainty of the model by using the Monte Carlo Dropout technique in the last layer of each task decoding head. That is, during the model inference phase, the Dropout layer (for example, p = 0.5) is still activated, and N (for example, N = 20) forward propagations are performed on the same input sample to obtain N different prediction results. Then, the variance of the N prediction results is calculated as a quantitative measure of the uncertainty of the task prediction. When detecting that the gradients of any two tasks (for example, task i and task j) conflict, the optimizer dynamically adjusts the projection strength according to the uncertainty of the two tasks. The core operation of this adjustment is gradient projection, that is, the corrected gradient is calculated by subtracting the projection component of the original gradient in the direction of the conflicting gradient. In the embodiment, the size of the projection component is dynamically scaled by a weight factor, which is the uncertainty variance of task j divided by the sum of the uncertainty variances of task i and task j. The weight factor makes the correction amplitude of the gradient smaller when the uncertainty of task i is much higher than that of task j; otherwise, when the uncertainty of task i is lower, the gradient is more fully corrected. This adaptive mechanism allows the model to dynamically allocate more optimization resources to the most difficult and insufficiently learned risk task, thereby accelerating the convergence of the overall model and improving the prediction accuracy of key risks in complex scenarios.

[0053] By adopting this conflict resolution optimizer based on gradient projection, the embodiment ensures that each update of the model during multi-task learning does not directly interfere with the optimization direction of all tasks, thereby achieving a more stable and efficient training process and improving the final performance and reliability of the entire risk early warning system.

[0054] The present application proposes an end-to-end enterprise risk early warning framework, which fuses multi-source heterogeneous data into a dynamic knowledge graph, simulates risk propagation using a spatio-temporal graph neural network, and performs risk assessment and model optimization using a multi-task learning framework that includes adversarial debiasing and gradient conflict resolution. This method constructs a systematic risk early warning system that can dynamically perceive, quantify, and continuously self-optimize, improving the accuracy and timeliness of risk prediction and enabling the transformation from raw unstructured data to structured and actionable risk scores. The present application models the transmission path of risks between enterprises using a graph neural network, which can reveal and quantify systemic risks and cascading effects that traditional isolated models cannot perceive. Moreover, its multi-task learning mechanism, through adversarial debiasing and gradient conflict resolution, ensures that the prediction of different risk dimensions is robust and decoupled, reducing false positives caused by pseudo-correlation.

[0055] Embodiment Two This example two demonstrates the application of the system using the method of the present application in the specific planning process in financial institution A; see Figure 3 The specific implementation of the multi-source data module is as follows: The system first collects data from multiple preset data sources in real time or regularly through API interface. These data sources are divided into two categories, unstructured and semi-structured text data, which mainly include market analysis reports and industry research reports obtained in real time through data agreements signed with financial information service providers; at the same time, the system also deploys automated programs to regularly crawl the PDF documents of enterprise annual and quarterly financial reports from the public databases of relevant securities regulatory agencies. Structured data is usually accessed through an API to an authoritative enterprise information data provider to obtain verified enterprise association relationships (such as parent company, subsidiary) and standardized enterprise unique identifiers.

[0056] The collected data is automatically distributed to two parallel processing pipelines according to its format, all unstructured news and research report texts are sent to pipeline A (NLP pipeline). This pipeline first uses a language model based on the Transformer architecture fine-tuned on a financial corpus for named entity recognition, identifying company names and key risk events (such as "default" "regulatory investigation" "supply chain disruption") in the text, and the relationship extraction module captures the semantic connections between these entities, such as extracting the relationship triples (Company A)-[owns supplier]->(Company B) and (Company B)-[experiences]->(production stop) from "Company A announced that its supplier Company B has stopped production due to reasons". The financial report files downloaded from the official disclosure system are sent to pipeline B (document parsing pipeline), which applies document layout analysis techniques to deconstruct the financial report files into logical areas such as text blocks and tables, and the table extraction module based on libraries such as pdfplumber locates and parses the balance sheet and income statement, extracting key health indicators such as "total assets" "total liabilities" "revenue" and "net profit", and also scans the report footnotes to identify and extract explicitly stated guarantee or holding relationships.

[0057] The processing results of the two pipelines, together with the data obtained from the structured API, constitute a complete enterprise data set, which is continuously injected into a knowledge graph driven by a native graph database.

[0058] The system creates a financial node for each company using the standardized enterprise unique identifier obtained from the enterprise information data provider, which ensures that the "Company A" mentioned in the news can be accurately linked to the "Company A" in its financial reports, and creates a RiskEvent node for each identified risk event. Health indicators extracted from pipeline B (such as net_income: 50000000) and indicators obtained from market data APIs (such as stock_volatility: 0.35) are added as attributes of the corresponding financial node with financial report period or date timestamp. Relationships extracted from all sources are converted into directed edges with types connecting corresponding node pairs, such as HAS_SUPPLIER, HAS_GUARANTEE, IS_EXPOSED_TO.

[0059] Embodiment Three This embodiment three shows the specific planning process of the system using the method of the present application in financial institution A; the specific implementation of the impact propagation module is as follows: Receive the dynamic knowledge graph and generate a health state vector for the core enterprise "Company A" in the graph through the spatio-temporal graph neural network (STGNN) to quantify its systemic risk. First, the system needs to construct an initial state vector for each enterprise node. This process includes the following steps: first, feature selection, the system selects a total of 40 key health indicators covering solvency, profitability, operational capacity and growth capacity from financial statements, such as asset-liability ratio, quick ratio, net profit growth rate, etc. Second, data preprocessing, all selected indicators are standardized by Z-score to eliminate dimensional influence, and for missing values, the annual median of the corresponding industry is filled. Third, feature embedding and dimensionality reduction, since the original indicator dimension is low and the features are heterogeneous, the system inputs the preprocessed 40-dimensional indicator vector into a multi-layer perceptron (MLP) composed of two fully connected networks for feature embedding. The network structure of this multi-layer perceptron is [40, 80, 128], i.e. 40 neurons in the input layer, 80 neurons in the hidden layer, and 128 neurons in the output layer, with ReLU as the activation function. The 128-dimensional dense vector output by this multi-layer perceptron is defined as the initial state vector of the node, which serves as the starting point for subsequent model calculations. Subsequently, the model performs iterative message passing in a total of 3 propagation levels. At each level, the state of Company A is updated according to its associated enterprise network.

[0060] This updating process is guided by a sophisticated two-layer attention mechanism: first, a relation-level attention mechanism assigns weights to different types of relations, e.g. the model might learn that the risk transmission weight for the “provides guarantee” relation is 0.65, much higher than the “owns loan” (0.25) or “is parent company” (0.10) relations; then, within each relation grouping, an entity-level attention mechanism computes the importance of each neighbor node according to its own state, a neighbor with higher own risk (e.g. another company that is guaranteed by company A) will get a higher attention weight. The information of all neighbor nodes is aggregated into a comprehensive message vector after being doubly weighted by the two layers of attention. This vector is input into a gated recurrent unit (GRU) together with the state vector of company A in the previous level, which intelligently fuses the externally transmitted risk shock with the inherent state of company A. After 3 rounds of iterative computation, the system finally generates a 256-dimensional health state vector for company A, which densely encodes all its intrinsic and systemic risk exposures within a three-hop neighborhood.

[0061] Embodiment Four This embodiment four demonstrates the specific planning process of the system using the method of the present application in financial institution A; the specific implementation of the multi-task module is as follows: This module receives the 256-dimensional health state vector generated by the impact propagation module for “company A” (e.g. [-1.57, 2.03,...]) and decodes it into a plurality of specific risk estimates. First, to strip the pseudo-correlation between different risks, the system uses an adversarial debiasing network to purify the vector. Taking the generation of independent credit risk features as an example, a main predictor network receives the health state vector to predict the default probability of company A, while an adversarial network attempts to predict a confusion task (e.g. market risk) from the intermediate features of the main predictor.

[0062] Through the mechanism of the gradient inversion layer, the main predictor maximizes the market risk prediction error of the adversarial network while minimizing its own credit risk prediction error, thereby being forced to generate a 64-dimensional “credit risk specificity vector” that does not contain market risk information. This process will be performed in parallel for market risk, operational risk, etc. to generate a set of mutually risk-specific vectors. Subsequently, these specificity vectors are sent to the respective task decoding heads.

[0063] During the backpropagation training phase of the whole model, the system adopts a conflict resolution optimizer based on gradient projection. If the cosine similarity between the loss gradient of the credit risk task and the loss gradient of the market risk task is negative in a certain iteration, the optimizer will automatically project one gradient onto the normal plane of the other gradient to eliminate its conflict component. Using the corrected gradient to update the shared parameters can ensure the stability of multi-task learning and the optimization of overall performance, generating a set of independent risk portraits for Company A While embodiments of the present application have been shown and described with reference to particular embodiments thereof, it will be understood by those of ordinary skill in the art that various changes in form and details can be made therein without departing from the spirit and scope of the application which is defined by the following claims and their equivalents.

Claims

1. A method for enterprise risk early warning based on multi-source data fusion, characterized in that, include: Receive internal documents and knowledge bases, extract data assets, risk events, corporate relationships and health indicators, and form corporate data groups; By leveraging enterprise data sets, a dynamic knowledge graph can be constructed. In the influence propagation model based on spatiotemporal graph neural network, multi-layer iterative message passing based on a two-layer attention mechanism is performed on the dynamic knowledge graph. The state information of each node is updated in a weighted manner according to the state information of neighboring nodes and the attributes of directed edges to generate a health state vector. The health status vector is input into a multi-task decoding framework that includes a pseudo-correlation suppression mechanism. The features of each risk task are purified by an adversarial debiasing network to generate a risk-specific vector. The risk-specific vector is then passed through multiple task decoding heads to obtain a risk estimate. The difference between the risk estimate and the actual loss is used to update the impact propagation model through a gradient projection-based conflict resolution optimizer.

2. The enterprise risk early warning method based on multi-source data fusion according to claim 1, characterized in that, The specific process of receiving internal documents and the knowledge base includes: for internal documents, applying a natural language processing pipeline to identify data assets and risk events and capture the dynamic semantic relationships between the data assets and between data assets and risk events; the natural language processing pipeline integrates named entity recognition, event extraction, entity linking, and document-level relationship extraction; for the knowledge base, using a document parsing pipeline, through layout analysis and table parsing, extracting health indicators, and identifying known enterprise relationships including holding and guarantee from the knowledge base to form enterprise data groups.

3. The enterprise risk early warning method based on multi-source data fusion according to claim 1, characterized in that, The specific process of constructing a dynamic knowledge graph includes: receiving enterprise data groups, linking data assets and risk events to unique identifiers in the knowledge base, and instantiating a corresponding node for each data asset and risk event in the graph database; attaching health indicators as key-value pair attributes to the corresponding nodes; and transforming entity relationships, enterprise relationships, and relationships between data assets and risk events into directed edges connecting the corresponding node pairs, wherein the type of the directed edges is defined by the properties of the relationships.

4. The enterprise risk early warning method based on multi-source data fusion according to claim 1, characterized in that, The specific generation process of the health state vector includes: for each data asset node in the dynamic knowledge graph, the initial state information at layer zero is defined as the feature vector composed of the health state vector; for a preset number of propagation layers, each layer performs multi-layer iterative message passing to perceive the spatiotemporal influence for each data asset node; the state information of each data asset node at the last propagation layer is designated as the health state vector, wherein the health state vector encodes the systematic influence of the topology and node attributes in the neighborhood and quantifies the potential chain reaction.

5. The enterprise risk early warning method based on multi-source data fusion according to claim 4, characterized in that, The specific process of multi-layer iterative message passing includes: in the relation layer, adjacent nodes are grouped according to the type of directed edges; for each group, a relation-level attention weight is calculated based on the state information of the central node, preset attributes, and relation type; in the entity layer, within each group, for each adjacent node within the group, an entity-level attention weight is calculated based on the current state information of the central node and adjacent nodes; the state information of each adjacent node is weighted by the corresponding entity-level attention weight, and then weighted a second time by the relation-level attention weight of its group; all message vectors that have undergone double weighting are aggregated into a comprehensive message vector; by combining the comprehensive message vector with the state information of the data asset node in the previous level, and inputting the combined result into a nonlinear update function, the state information of the data asset node in the current level is generated.

6. The enterprise risk early warning method based on multi-source data fusion according to claim 1, characterized in that, The specific generation process of the risk-specific vector includes: identifying spurious correlations between different risk task labels through statistical analysis; constructing a master predictor and an adversarial network for each main risk task; the master predictor receiving a health state vector and generating an intermediate feature representation; the adversarial network receiving the intermediate feature representation and predicting confused tasks that are spuriously correlated with the main risk; and updating the parameters of the master predictor through adversarial training, minimizing its own prediction loss while maximizing the prediction loss of the adversarial network, thereby generating a risk-specific vector that has had spurious-related information removed.

7. The enterprise risk early warning method based on multi-source data fusion according to claim 1, characterized in that, The specific process of generating risk valuation includes: classifying the health status vector to generate contextual features; concatenating the risk-specific vector with the contextual features to generate an augmented feature vector; inputting the augmented feature vector in parallel to multiple task decoders, each task decoder corresponding to a risk dimension, and outputting a risk valuation for the risk dimension.

8. The enterprise risk early warning method based on multi-source data fusion according to claim 1, characterized in that, The specific process of updating the influence propagation model includes: comparing the risk estimate with the corresponding actual loss to generate a loss value; for each loss value, calculating the gradient with respect to the shared parameters of the influence propagation model using the backpropagation algorithm to generate a set of gradient vectors; when a conflict is detected between the gradient vectors of any two tasks, adjusting at least one conflicting gradient vector to reduce the conflict component with the other gradient vector to generate a corrected gradient vector; and updating the shared parameters of the influence propagation model using all the corrected parameter update vectors.

9. An enterprise risk early warning system based on multi-source data fusion, characterized in that, include: Multi-source data module: Receives internal documents and knowledge bases, extracts data assets, risk events, corporate relationships and health indicators to form corporate data groups; By leveraging enterprise data sets, a dynamic knowledge graph can be constructed. Influence propagation module: In the influence propagation model based on spatiotemporal graph neural network, multi-layer iterative message passing based on two-layer attention mechanism is performed on the dynamic knowledge graph. The state information of each node is updated in a weighted manner according to the state information of neighboring nodes and the attributes of directed edges to generate a health state vector. Multi-task module: The health status vector is input into a multi-task decoding framework that includes a pseudo-correlation suppression mechanism. The features of each risk task are purified by an adversarial debiasing network to generate a risk-specific vector. The risk-specific vector is then passed through multiple task decoding heads to obtain a risk estimate. The difference between the risk estimate and the actual loss is used to update the impact propagation model through a gradient projection-based conflict resolution optimizer.

Images