A dynamic intelligent risk control method and device, a storage medium and an electronic device

Abstract

This invention provides a dynamic intelligent risk control method, device, storage medium, and electronic device, applied in the field of artificial intelligence technology. The invention acquires multimodal data of target enterprises, performs feature extraction and alignment to obtain multimodal feature vectors, and dynamically updates the enterprise relationship network based on these vectors. It integrates multimodal features and the relationship network to identify risk associations between enterprises, predicts the transmission path and impact of risks within the network, and obtains risk transmission information. Finally, it combines multimodal features, enterprise relationship networks, and risk transmission information to generate interpretable risk analysis results. This invention constructs an end-to-end and logically transparent risk analysis framework, achieving accurate risk prediction while intuitively revealing the formation, transmission process, and evidence chain of risks, thus improving the accuracy and credibility of risk identification.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a dynamic intelligent risk control method, device, storage medium, and electronic device. Background Technology

[0002] Existing bank risk control systems suffer from three major pain points: First, information silos: data from transaction, credit, and public opinion systems are fragmented, making it difficult to correlate cross-modal risk clues (such as abnormal fund flows and related-party equity changes) and form a unified risk view. Second, static relationships: relying on static rules and equity registration, these systems cannot dynamically capture the real-time evolution of implicit relationship networks formed through multi-layered shareholding and cross-appointment, leading to delayed predictions of risk contagion paths. Third, black boxes in decision-making: while deep learning models provide accurate predictions, the process lacks transparency, failing to clearly explain the specific paths and key evidence of risk transmission, thus failing to meet regulatory requirements for interpretability.

[0003] Existing technologies attempt to fuse multimodal data through simple feature stitching, but lack in-depth processing of modal differences; they utilize static graph analysis of corporate relationships, but cannot achieve dynamic updates; and they employ ex-post explanation methods such as SHAP (Shapley Additive Explanations) or LIME (Local Interpretable Model-agnostic Explanations), but struggle to provide a clear and complete causal transmission path. These methods still suffer from shortcomings in addressing complex and dynamic financial risks, including coarse fusion, slow response, and insufficient explanation.

[0004] Therefore, how to enhance the interpretability of risk decisions and improve the accuracy of risk identification has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of the above problems, the present invention provides a dynamic intelligent risk control method, device, storage medium, and electronic device to overcome or at least partially solve the above problems. The technical solution is as follows:

[0006] A dynamic intelligent risk control method includes:

[0007] Obtain multimodal data of the target company;

[0008] Feature extraction and alignment are performed on the multimodal data to obtain multimodal feature vectors;

[0009] The enterprise relationship network of the target enterprise is dynamically updated using the multimodal feature vectors.

[0010] By fusing the multimodal feature vectors and the enterprise relationship network, risk associations between enterprises can be identified;

[0011] Based on the identified risk associations, the transmission path and impact of risks in the enterprise relationship network are predicted to obtain risk transmission information;

[0012] Using the multimodal feature vectors, the enterprise relationship network, and the risk transmission information, interpretable risk analysis results are generated.

[0013] Optionally, the multimodal data includes the target enterprise's time-series corporate behavior data, enterprise relationship diagram structure data, and enterprise market environment text data. The step of extracting and aligning features from the multimodal data to obtain a multimodal feature vector includes:

[0014] Differential encoding is performed on the enterprise behavior time series data, the enterprise relationship graph structure data, and the enterprise market environment text data respectively, and corresponding time series features, graph structure features, and text features are extracted.

[0015] The temporal features, graph structure features, and text features are mapped to a unified semantic space to generate aligned multimodal feature vectors.

[0016] Optionally, the step of differentially encoding the enterprise behavior time-series data, the enterprise relationship graph structure data, and the enterprise market environment text data, and extracting corresponding time-series features, graph structure features, and text features, includes:

[0017] The enterprise behavior time-series data is encoded using a preset temporal neural network architecture to extract time-series features that characterize the dynamic features of funds;

[0018] The enterprise relationship graph structure data is encoded using a dynamic hypergraph neural network to extract graph structure features that represent the multi-dimensional relationships between enterprises;

[0019] The enterprise market environment text data is encoded using a pre-trained language model to extract text semantic features.

[0020] Optionally, dynamically updating the enterprise relationship network of the target enterprise using the multimodal feature vector includes:

[0021] Based on the multimodal feature vectors within the sliding time window, the relationship strength between nodes in the enterprise relationship network of the target enterprise is dynamically updated;

[0022] The set of hyperedges in the enterprise relationship network is expanded by clustering methods to capture inter-enterprise meta-relationships.

[0023] Optionally, identifying risk associations between enterprises by fusing the multimodal feature vectors and the enterprise relationship network includes:

[0024] A variant attention mechanism is used to fuse modal differences in the multimodal feature vectors to generate a fused context vector.

[0025] Based on the context vector and the enterprise relationship network, the explicit relationships and implicit associations in the enterprise relationship network are identified.

[0026] Optionally, the variant attention mechanism achieves modality-differentiated information fusion by introducing a modality-specific attention mask corresponding to the data type and a function that adaptively learns the relationship strength based on the interaction of the query vector, key vector, and value vector.

[0027] Optionally, generating interpretable risk analysis results using the multimodal feature vectors, the enterprise relationship network, and the risk transmission information includes:

[0028] Using the multimodal feature vectors and the risk transmission information, an interpretable mathematical model of risk transmission is generated through neural symbolic regression fitting;

[0029] Extract key risk transmission paths from the aforementioned corporate relationship network;

[0030] Based on the aforementioned risk transmission mathematical model and the aforementioned key risk transmission paths, a natural language report containing risk descriptions and transmission logic is generated.

[0031] A dynamic intelligent risk control device includes: a multimodal data acquisition unit, a multimodal feature vector acquisition unit, an enterprise relationship network update unit, a risk association identification unit, a risk transmission information acquisition unit, and a risk analysis result generation unit;

[0032] The multimodal data acquisition unit is used to acquire multimodal data of the target enterprise;

[0033] The multimodal feature vector acquisition unit is used to extract and align features from the multimodal data to obtain multimodal feature vectors.

[0034] The enterprise relationship network update unit is used to dynamically update the enterprise relationship network of the target enterprise using the multimodal feature vector;

[0035] The risk association identification unit is used to identify risk associations between enterprises by fusing the multimodal feature vectors and the enterprise relationship network;

[0036] The risk transmission information acquisition unit is used to predict the transmission path and impact of risks in the enterprise relationship network based on the identified risk associations, and to obtain risk transmission information.

[0037] The risk analysis result generation unit is used to generate interpretable risk analysis results by utilizing the multimodal feature vector, the enterprise relationship network, and the risk transmission information.

[0038] A computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the dynamic intelligent risk control method described above.

[0039] An electronic device includes at least one processor, at least one memory connected to the processor, and a bus; wherein the processor and the memory communicate with each other via the bus; the processor is used to call program instructions in the memory to execute the dynamic intelligent risk control method.

[0040] By employing the above technical solutions, this invention provides a dynamic intelligent risk control method, device, storage medium, and electronic device that obtains multimodal data of the target enterprise; extracts and aligns features from the multimodal data to obtain multimodal feature vectors; dynamically updates the enterprise relationship network of the target enterprise using the multimodal feature vectors; identifies risk associations between enterprises by fusing the multimodal feature vectors and the enterprise relationship network; predicts the transmission path and impact of risks in the enterprise relationship network based on the identified risk associations, obtaining risk transmission information; and generates interpretable risk analysis results using the multimodal feature vectors, enterprise relationship network, and risk transmission information. This invention, through deep integration of multimodal features, dynamically updated enterprise relationship networks, and risk transmission path information, constructs an end-to-end, logically transparent risk analysis framework. While predicting risks, it intuitively reveals the formation, transmission, and evidence chain of risks, thereby achieving accurate risk identification and credible interpretation.

[0041] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0042] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0043] Figure 1 A flowchart illustrating one embodiment of the dynamic intelligent risk control method provided by this invention is shown.

[0044] Figure 2The diagram shows a specific implementation of step S110 in the dynamic intelligent risk control method provided by the present invention.

[0045] Figure 3 The diagram shows a specific implementation of step S120 in the dynamic intelligent risk control method provided by the present invention.

[0046] Figure 4 The diagram shows a specific implementation of step S130 in the dynamic intelligent risk control method provided by the present invention.

[0047] Figure 5 The diagram shows a specific implementation of step S150 in the dynamic intelligent risk control method provided by the present invention.

[0048] Figure 6 A schematic diagram of the structure of the dynamic intelligent risk control device provided in an embodiment of the present invention is shown;

[0049] Figure 7 A schematic diagram of the structure of an electronic device provided in an embodiment of the present invention is shown. Detailed Implementation

[0050] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this invention will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.

[0051] In the field of financial risk control, continuous monitoring and assessment of corporate credit and operational risks is a crucial core function for banks and other financial institutions. Existing bank risk control systems mainly rely on the analysis of structured data (such as financial statements and transaction records) and pre-set rule engines. However, traditional systems have significant limitations when facing increasingly complex business environments, hidden related-party transactions, and cross-market risk transmission.

[0052] Currently, existing bank risk control systems mainly face the following three core pain points:

[0053] First, the problem of information silos is severe. Traditional risk control systems typically employ a modular design, with transaction tracking systems, credit inquiry systems, and public opinion monitoring systems operating independently, resulting in a disconnect between data and processing logic. For example, an unusually high frequency of fund transfers by a company (time-series data) may be intrinsically linked to recent equity structure changes at another related company (graph-structure data) and negative news reports in the market (text data), both pointing to potential collaborative risks. However, existing systems lack effective mechanisms for deep integration and unified analysis of this heterogeneous, multi-source, and cross-modal data, leading to a fragmented risk view and hindering the formation of a comprehensive and multi-dimensional risk understanding.

[0054] Second, relationship inference is static, lacking the ability to dynamically capture relationships. Existing systems identify inter-enterprise relationships primarily based on static equity relationships and guarantee chains registered in business registration information, or rely on fixed rules and thresholds for judgment. This static map cannot reflect the dynamic evolution of enterprise relationship networks in real time. For example, enterprises may form substantial control or significant influence through complex methods such as multi-layered nested shareholding, cross-appointment, and hidden concerted action parties. These implicit connections change continuously over time. Traditional risk control models struggle to continuously and automatically discover and update these dynamically evolving complex relationship networks, leading to delayed or even ineffective predictions of the path of risk contagion among related parties.

[0055] Third, the decision-making logic is opaque, and the "black box" model fails to meet regulatory and internal control requirements. Although advanced models, such as deep learning, have improved the accuracy of risk signal identification and prediction, their internal decision-making processes are highly complex and opaque. When the system triggers a high-risk warning, it often cannot clearly and intuitively explain to risk managers and regulatory agencies how the risk is transmitted through which key paths in the enterprise network, or the specific contributions of different data sources (such as a particular transaction or a public opinion event) to the final risk assessment. This lack of interpretability not only hinders manual review and intervention in risk decisions but also makes it difficult to meet the increasingly stringent regulatory compliance requirements of the financial industry (such as "algorithmic auditing").

[0056] To address these challenges, existing technologies attempt improvements from different angles: In multimodal data fusion, simple feature concatenation or weighted averaging based on expert experience are commonly used to merge feature vectors from different sources. However, this approach lacks targeted processing of the inherent characteristics and differences of different modalities, such as time-series data (dynamic sequences), graph data (structural relationships), and text data (semantic information), resulting in coarse fusion effects and failing to achieve deep semantic alignment and complementarity. In enterprise relationship analysis, it mainly relies on graph databases based on static snapshots and traditional graph algorithms (such as community detection and centrality calculation). The relationship networks constructed by these methods are static or periodically updated, unable to achieve dynamic perception and incremental updates of enterprise relationships at the millisecond or minute level, and difficult to capture instantaneous changes in the relationship network driven by risk events. In terms of model interpretability, ex-post, model-agnostic explanation techniques are often used, such as SHAP (Shapley Additive Explanations) or LIME (Local Interpretable Model-agnostic Explanations). While these methods can assess feature importance, they are essentially local approximations of complex model decisions. They fail to provide a complete and logically clear explanation chain from the source of risk and its transmission path to its final impact, and cannot generate risk narrative reports that business personnel can directly understand.

[0057] Based on this, this invention provides a dynamic intelligent risk control method. It acquires multimodal data of target enterprises, performs feature extraction and alignment to obtain multimodal feature vectors, and dynamically updates the enterprise relationship network based on these vectors. By fusing multimodal features and the relationship network, it identifies risk associations between enterprises, predicts the transmission path and impact of risks within the network, and obtains risk transmission information. Finally, by combining multimodal features, enterprise relationship networks, and risk transmission information, it generates interpretable risk analysis results. Therefore, this invention constructs an end-to-end and logically transparent risk analysis framework, achieving both accurate risk prediction and intuitively revealing the formation, transmission process, and evidence chain of risks, thus improving the accuracy and credibility of risk identification.

[0058] like Figure 1 The diagram shows a flowchart of one embodiment of the dynamic intelligent risk control method provided by this invention. The method may include:

[0059] S100: Obtain multimodal data of the target company.

[0060] The target enterprise refers to the corporate entity that is the object of risk identification and analysis in the risk control system, which can be a corporate client that is subject to bank credit, regulatory or compliance attention.

[0061] Multimodal data, also known as multi-source heterogeneous data, refers to enterprise-related data from various heterogeneous channels and forms, including transaction records, equity and employment information, news and public opinion, and announcement texts, among other multi-dimensional information.

[0062] Specifically, embodiments of the present invention can collect multimodal data from target enterprises in real time through streaming data interfaces, batch data interfaces, and API gateways. This includes time-series data such as cash flow records, enterprise relationship diagram structure data such as equity structure and employment relationships, and enterprise market environment text data such as news and public opinion. The multimodal data can be encapsulated in standard formats such as JSON to ensure standardized access and dynamic updates, supporting subsequent real-time processing and analysis.

[0063] As examples, embodiments of the present invention can access multi-source heterogeneous data related to a target enterprise in real time or near real time through streaming and batch data interfaces and API gateways in the data access layer. This multi-source heterogeneous data may include time-series transaction flow data from core banking systems and payment institutions, enterprise relationship diagram structure data (such as equity relationships and senior management appointment information) from business registration information platforms and equity penetration tools, and enterprise market environment text data (such as news reports and company announcements) from news media and public opinion monitoring platforms. All data is encapsulated and transmitted in a unified JSON format, providing standardized input for subsequent processing.

[0064] S110. Perform feature extraction and alignment on the multimodal data to obtain multimodal feature vectors.

[0065] Among them, multimodal feature vectors refer to a unified-dimensional numerical representation obtained by cross-modal alignment and fusion after special encoding and feature extraction of multimodal data, which is used to describe the comprehensive characteristics of the target enterprise under different data modalities.

[0066] Specifically, embodiments of the present invention can extract modality-specific feature representations from various heterogeneous data through differential coding methods, and then transform and integrate these heterogeneous features into a unified semantic space through feature mapping and fusion techniques, thereby generating semantically aligned and structurally consistent multimodal feature vectors.

[0067] Furthermore, embodiments of the present invention can employ temporal data encoders, graph structure encoders, and text semantic encoders respectively for deep feature extraction of different modal data. The extracted temporal, graph structure, and text features are then aligned in a unified semantic space through a multi-head attention mechanism in the cross-modal alignment module, ultimately mapping each modal feature to a unified semantic space and generating a fused and aligned multimodal feature vector.

[0068] As examples, embodiments of the present invention can perform multimodal differential feature extraction in parallel at the feature engineering layer: a time-series data encoder based on the Mamba architecture processes fund flows to extract dynamic patterns; a graph structure encoder based on a Dynamic Hypergraph Neural Network (DHGNN) processes relationships such as equity and guarantees to generate node and edge embeddings; and a text semantic encoder fine-tuned with financial domain knowledge processes public opinion text to generate semantic vectors. Embodiments of the present invention can employ a multi-head attention mechanism to project heterogeneous features from different encoders into a unified 256-dimensional semantic space, achieving semantic alignment and information complementarity between modalities, ultimately forming a consistent multimodal feature vector.

[0069] S120. Utilize multimodal feature vectors to dynamically update the enterprise relationship network of the target enterprise.

[0070] Among them, the enterprise relationship network refers to a dynamic graph structure based on multimodal feature vectors that reflects various types of relationships between enterprises (such as equity control, transaction associations, guarantee relationships, and industry linkages), and is used to reveal explicit and implicit connections between enterprises.

[0071] Specifically, embodiments of the present invention can dynamically adjust the connection strength between enterprise relationship network nodes according to the temporal changes of multimodal feature vectors, and use the clustering information of multimodal feature vectors to identify and construct new high-order association structures, thereby achieving continuous optimization and expansion of network topology and relationship composition.

[0072] Furthermore, embodiments of the present invention can dynamically construct and maintain enterprise relationship networks based on multimodal feature vectors, using temporal graph neural networks combined with a sliding window mechanism, and expand hyperedge sets through clustering algorithms to capture diverse and complex relationships. Community detection algorithms automatically identify and track the formation and evolution of associated enterprise groups, thereby achieving real-time and incremental dynamic updates of the enterprise relationship network.

[0073] As examples, embodiments of the present invention can receive aligned multimodal feature vectors through a dynamic graph layer and maintain and update the enterprise relationship network using an incremental dynamic hypergraph learning method. The temporal graph neural network dynamically adjusts the connection strength between network nodes (enterprises) using feature vectors within a 30-day sliding window. Simultaneously, based on clustering information from feature vectors (such as through the K-means algorithm), new higher-order associations (hyperedges) are identified and constructed to capture complex multi-dimensional relationships. The dynamic community detection algorithm then discovers and tracks the formation and evolution of associated enterprise groups in real time, thereby achieving continuous optimization and expansion of the network topology and relationship structure.

[0074] S130. By fusing multimodal feature vectors and enterprise relationship networks, risk associations between enterprises are identified.

[0075] Risk association refers to the potential risk transmission relationships between enterprises identified by integrating multimodal features and enterprise relationship networks. This includes explicit and implicit relationships, reflecting the path and intensity of risk transmission from one enterprise to another.

[0076] Specifically, embodiments of the present invention can integrate multimodal features into a unified contextual representation, and then combine it with the structural information of the enterprise relationship network to comprehensively analyze and identify explicit and potential risk associations existing in the enterprise relationship network.

[0077] Furthermore, embodiments of the present invention may employ a variant attention mechanism, designing differentiated attention computation strategies for temporal, structural, and semantic modalities, integrating multimodal features and dynamic graph information to identify various explicit and implicit risk associations, including guarantee relationships, transaction associations, equity control, industry linkages, and risk contagion.

[0078] As examples, embodiments of the present invention can perform modal-differentiated information fusion on multimodal feature vectors through a variant attention mechanism in the multimodal inference layer (e.g., applying different attention masks to different types of data such as equity control and abnormal transactions), generating a context-rich fusion vector. Then, a dynamic relational inference engine is used to combine this contextual vector with the current enterprise relationship network structure to comprehensively analyze and identify various relationships existing in the network. This includes both explicit relationships directly reflected in data such as equity and transactions, and implicit associations discovered through cross-modal inference and pattern analysis that are not registered in explicit data (such as potential implicit control and risk contagion associations), thereby comprehensively characterizing the risk relationships between enterprises.

[0079] S140. Based on the identified risk associations, predict the transmission path and impact of risks in the enterprise relationship network, and obtain risk transmission information.

[0080] Among them, risk transmission information refers to the dynamic characteristics of risk propagation in corporate relationship networks based on identified risk associations, such as the path, intensity, and timeliness of transmission, reflecting the process and scope of risk diffusion from the source to other related enterprises.

[0081] Specifically, embodiments of the present invention can utilize risk association, combined with a diffusion coefficient learning mechanism, to simulate the propagation process of risk events in enterprise relationship networks. It supports path prediction with a preset number of hops, estimates the risk transmission path, transmission intensity, and transmission time, forming complete dynamic information on risk transmission, and helping to predict the scope of risk diffusion and the timeliness of its impact.

[0082] As examples, embodiments of the present invention can perform simulations and deductions in the multimodal inference layer using a risk transmission path predictor based on an updated dynamic graph and identified risk associations (especially implicit risk associations). It supports multi-hop path searching (e.g., up to 5 hops) and has diffusion coefficient learning capabilities, enabling it to predict the most likely transmission path of a specific risk event (e.g., a company's default) in the relationship network, the transmission strength of each path, and the expected impact range and transmission time. Simultaneously, it analyzes causal relationships from time-series data to further verify and enrich the transmission logic. Finally, it outputs structured risk transmission information, including critical paths, intensity coefficients, and time predictions.

[0083] S150: Utilize multimodal feature vectors, corporate relationship networks, and risk transmission information to generate interpretable risk analysis results.

[0084] The risk analysis results refer to interpretable risk assessment reports generated through neural symbolic regression and causal reasoning by combining multimodal feature vectors, corporate relationship networks, and risk transmission information, which are used to support business decisions and regulatory compliance.

[0085] Specifically, embodiments of the present invention can construct an interpretable risk model based on multimodal feature vectors and risk transmission information, extract key transmission paths from enterprise relationship networks, and generate a natural language risk analysis report that combines numerical models and network path logic.

[0086] Furthermore, embodiments of the present invention can combine a neural symbol library and a complexity penalty mechanism to automatically derive an interpretable risk transmission mathematical model from multimodal features and dynamic graph data, extract key risk transmission paths and their confidence levels, and combine them with a natural language report generator to transform symbolic logic and causal paths into risk analysis reports that meet the requirements of financial professionals, thereby achieving transparency and business understanding in the risk assessment process and supporting intelligent risk control decisions and regulatory compliance.

[0087] As examples, embodiments of the present invention can use an interpretability layer to transform complex model reasoning into transparent and credible business insights. A neurosymbolic regression engine, taking multimodal feature vectors and risk transmission information as input, searches a predefined library of mathematical functions to automatically fit an interpretable mathematical model of risk transmission (such as an equation describing how risk changes with association strength). A causal path extractor then extracts the most critical risk transmission paths from the enterprise relationship network based on transmission strength and confidence thresholds. Finally, a natural language report generator integrates the mathematical model obtained from symbolic regression, the extracted key transmission paths, and other risk evidence, and generates a natural language risk analysis report containing quantitative risk descriptions, transmission logic analysis, and response recommendations, following a financial professional template.

[0088] This invention provides a dynamic intelligent risk control method, which includes: obtaining multimodal data of a target enterprise; extracting and aligning features from the multimodal data to obtain multimodal feature vectors; dynamically updating the enterprise relationship network of the target enterprise using the multimodal feature vectors; identifying risk associations between enterprises by fusing the multimodal feature vectors and the enterprise relationship network; predicting the transmission path and impact of risks in the enterprise relationship network based on the identified risk associations to obtain risk transmission information; and generating interpretable risk analysis results using the multimodal feature vectors, the enterprise relationship network, and the risk transmission information. This invention constructs an end-to-end, logically transparent risk analysis framework by deeply fusing multimodal features, dynamically updated enterprise relationship networks, and risk transmission path information. While predicting risks, it intuitively reveals the formation, transmission, and evidence chain of risks, thereby achieving accurate risk identification and credible interpretation.

[0089] Optionally, in the above Figure 1 Based on one or more corresponding embodiments, in another optional embodiment provided by the present invention, the multimodal data may include time-series data of the target enterprise's corporate behavior, corporate relationship diagram structure data, and textual data of the enterprise's market environment.

[0090] Corporate behavior time series data refers to the time series data of various behaviors and activities of an enterprise within a certain time range, which may include time series information reflecting the dynamic operation of the enterprise, such as cash flow, transaction records, business changes and payment behavior.

[0091] Among them, enterprise relationship graph structure data refers to graphical structure data that describes various explicit and implicit relationships between enterprises. It can include multi-dimensional connection information of enterprise nodes and edges, such as equity structure, controlling relationship, employment relationship, transaction relationship, guarantee relationship and industry alliance.

[0092] Among them, enterprise market environment text data refers to unstructured data derived from textual information such as news reports, announcements, public opinion monitoring, and policy documents, reflecting the market environment, industry dynamics, public opinion risks, and policy changes of enterprises.

[0093] Optional, based on Figure 1 The method shown is as follows: Figure 2 The diagram shows a specific implementation of step S110 in the dynamic intelligent risk control method provided by this invention. Step S110 may specifically include:

[0094] S200. Differentiated coding is performed on the time-series data of enterprise behavior, the structural data of enterprise relationship diagrams, and the textual data of enterprise market environment, respectively, and the corresponding time-series features, diagram structure features, and textual features are extracted.

[0095] Among them, time-series features refer to feature vectors extracted from enterprise behavior time-series data through a time-series data encoder. These features can capture the dynamic characteristics of behaviors such as cash flow, including fluctuation patterns, periodic changes, and trend information, providing a time-dimensional representation for dynamic risk analysis.

[0096] Among them, graph structure features refer to the high-dimensional features of nodes and edges extracted from enterprise relationship graph structure data through graph embedding. These features can encode the topological structure information and relationship strength of complex and diverse relationships between enterprises, reflecting the position of enterprises in the relationship network and their associated attributes.

[0097] Among them, text features refer to semantic vectors extracted from enterprise market environment text data through text vectorization, which are mapped to a unified semantic space. This can capture risk-related semantic information in the text and help reflect the external environment and potential risks faced by the enterprise.

[0098] Specifically, embodiments of the present invention can use a neural network model adapted to different types of enterprise data for feature encoding, respectively using a time series model to process sequence data, a graph neural network to process relational data, and a language model to process enterprise market environment text data, thereby extracting time series features, graph structure features, and text features that can accurately reflect the core information and inherent laws of each modality.

[0099] Furthermore, embodiments of the present invention can design specialized encoders for deep feature extraction based on different modalities of data. For time-series data of corporate behavior, a time-series data encoder based on the Mamba architecture is adopted, which captures the short-term fluctuations and long-term trends of time-series data such as cash flow through a 6-layer network structure and gating mechanism, generating 128-dimensional inputs mapped to 512-dimensional hidden layers of time-series features; for corporate relationship graph structure data, a dynamic hypergraph neural network is used, and the hyperedge set is expanded through the K-means algorithm to effectively capture complex and diverse corporate relationships and extract vertex features, edge features, and hyperedge features; for corporate market environment text data, a BERT model fine-tuned for the financial field is used to extract text semantic features, which are then mapped to a 256-dimensional semantic space through a projection layer.

[0100] S210. Map temporal features, graph structure features, and text features to a unified semantic space to generate aligned multimodal feature vectors.

[0101] Specifically, this embodiment of the invention employs a multi-head attention mechanism to semantically align features across three modalities: using 256 dimensions as a unified dimension, eight attention heads are set up, serving as query, key, and value inputs respectively. The relevance weights between different modalities are dynamically calculated, achieving dynamic fusion and alignment of the feature space. By introducing modality-specific attention masks and adjusting the attention distribution, the semantic consistency and information complementarity between modalities are enhanced, ultimately generating multimodal feature vectors with consistent dimensions and rich semantics, providing a unified and high-quality input representation for subsequent risk reasoning and dynamic relationship modeling.

[0102] This invention, through differentiated encoding of enterprise behavior time-series data, enterprise relationship diagram structure data, and enterprise market environment text data, and by mapping their features to a unified semantic space for alignment, can fully explore the unique value of various types of information, achieve deep integration of multi-dimensional data, and thus improve the accuracy of enterprise risk identification and the interpretability of analysis results.

[0103] Optionally, in the above Figure 2 Based on one or more corresponding embodiments, in another optional embodiment provided by the present invention, step S200 may specifically include:

[0104] By using a pre-defined temporal neural network architecture to encode time-series data of corporate behavior, time-series features representing dynamic characteristics of funds are extracted.

[0105] Specifically, embodiments of the present invention can employ a time-series data encoder based on the Mamba architecture. This encoder consists of a 6-layer deep neural network with an input dimension of 128 and a hidden layer dimension of 512. It incorporates a gating mechanism to efficiently capture short-term fluctuations and long-term trends in corporate behavioral time-series data such as cash flow. By progressively encoding the financial data within a sliding time window, a time-series feature representation with time dependence and dynamic changes is generated, providing accurate behavioral time-series information for subsequent dynamic risk analysis.

[0106] Dynamic hypergraph neural networks are used to encode enterprise relationship graph structure data and extract graph structure features that represent multiple relationships between enterprises.

[0107] Specifically, this embodiment of the invention can employ a dynamic hypergraph neural network to encode the enterprise relationship graph structure data, with a node feature dimension of 256, an edge feature dimension of 128, and a hyperedge feature dimension of 192. The hyperedge set is expanded using the K-means algorithm to effectively capture diverse and complex relationships between enterprises, including equity control, transactions, and guarantee relationships. This encoder supports dynamic updates within a 30-day sliding window, adjusting the feature vectors of nodes and edges in real time to reflect the time-varying and diverse nature of the enterprise relationship network, providing rich structural information for risk transmission path prediction and implicit association identification.

[0108] We use pre-trained language models to encode enterprise market environment text data and extract semantic features from the text.

[0109] Specifically, embodiments of the present invention can encode enterprise market environment text data based on a BERT pre-trained language model fine-tuned in the financial field. First, unstructured texts such as news reports, announcements, and public opinion are segmented and preprocessed, then input into the BERT model to obtain deep semantic vector representations. Subsequently, a projection layer maps the text features to a unified 256-dimensional semantic space, consistent with the dimensions of temporal and graph structure features. These text semantic features can reflect changes in the market environment, public opinion risks, and policy dynamics, assisting in improving the multimodal panoramic perception of enterprise risks.

[0110] This invention employs pre-defined temporal neural networks, dynamic hypergraph neural networks, and pre-trained language models to perform differentiated encoding on enterprise behavior time-series data, enterprise relationship graph structure data, and enterprise market environment text data, respectively. This approach can fully mine and characterize the dynamic features, structural relationship features, and semantic information of different types of data, thereby providing a more accurate and richer feature foundation for subsequent multimodal feature alignment and enterprise risk identification, and improving the comprehensiveness and accuracy of enterprise risk analysis.

[0111] Optional, based on Figure 1 The method shown is as follows: Figure 3 The diagram shows a specific implementation of step S120 in the dynamic intelligent risk control method provided by this invention. Step S120 may specifically include:

[0112] S300: Based on the multimodal feature vectors within the sliding time window, dynamically update the relationship strength between nodes in the enterprise relationship network of the target enterprise.

[0113] In dynamic data processing, a sliding time window refers to selecting a fixed-length time interval as the analysis scope and continuously "sliding" forward as time progresses, covering the latest data within that time period in real time. For example, in this embodiment of the invention, a 30-day sliding window can be used to continuously analyze the time-series data of corporate behavior and relationship data within the most recent 30 days, dynamically capturing the changing trends of cash flow and corporate relationships, thereby achieving real-time updates of relationship strength and risk propagation, and avoiding inaccurate judgments due to outdated historical data.

[0114] Relationship strength refers to a numerical indicator used to quantify the importance and closeness of various relationships between enterprises (such as equity control, transactions, and guarantee relationships). Relationship strength comprehensively considers factors such as transaction frequency, capital amount, equity ratio, and degree of association, reflecting the actual influence of inter-enterprise connections. The dynamic relationship strength learning mechanism enables the system to adaptively adjust the weights of various relationships, improving its ability to identify key links in the risk transmission path.

[0115] Specifically, embodiments of the present invention can set a fixed-length sliding time window, within which multimodal data of the target enterprise and its related enterprises are continuously collected, including time-series data of capital flows, transaction and equity relationship graph structure data of enterprises, and textual data of market sentiment and enterprise market environment. Through preset time-series data encoders, graph structure encoders, and text semantic encoders, each type of data is transformed into a multimodal feature vector of a unified dimension. At the dynamic graph layer, a time-series graph neural network is used to fuse and encode the multimodal features within the time window. A relationship strength learning function is used to calculate the relationship weights between nodes, dynamically adjusting the relationship strength between nodes in the enterprise relationship network to achieve real-time reflection of the closeness of connections and risk transmission potential between enterprises. As the time window slides, the network relationship strength is continuously updated, ensuring the timeliness and accuracy of the relationship network.

[0116] S310. Expand the set of hyperedges in the enterprise relationship network through clustering methods to capture diverse relationships between enterprises.

[0117] In graph theory, a hyperedge represents an edge that connects multiple nodes simultaneously, and the set of hyperedges refers to the collection of all such multi-node edges. Unlike edges representing traditional binary relationships, hyperedges can depict diverse and complex connections between enterprises, such as guarantee groups involving multiple enterprises or cross-shareholding structures. This invention expands the hyperedge set through clustering algorithms, automatically identifying and constructing high-order relationship structures containing multiple enterprise nodes, enriching the expressive power of enterprise relationship networks, and improving the capture of implicit risk associations.

[0118] Specifically, embodiments of the present invention can perform clustering analysis on the high-dimensional feature vectors of enterprise nodes using clustering algorithms. During the clustering process, groups of multiple enterprise nodes are identified. These groups correspond to diverse and complex relationships between enterprises, such as joint guarantee circles, cross-shareholding structures, or multi-party transaction alliances. The clustering results are used as an extended set of hyperedges, dynamically constructing hyperedges containing multiple nodes, enriching the expressive power of enterprise relationship networks. This set of hyperedges is continuously updated to reflect the latest changes in diverse relationships, improving the ability to capture and analyze implicit association risks.

[0119] This invention utilizes multimodal feature vectors based on sliding time windows to dynamically update the relationship strength between nodes in an enterprise relationship network, and expands the hyperedge set through clustering methods to capture complex and diverse relationships between enterprises. This enables real-time and accurate characterization of enterprise relationships and effective mining of implicit associations, thereby improving the accuracy of risk association identification and the comprehensiveness of risk transmission path prediction.

[0120] Optional, based on Figure 1 The method shown is as follows: Figure 4 The diagram shows a specific implementation of step S130 in the dynamic intelligent risk control method provided by this invention. Step S130 may specifically include:

[0121] S400: Employ a variant attention mechanism to fuse modal differences in information from multimodal feature vectors, generating a fused context vector.

[0122] The variant attention mechanism refers to a specialized attention calculation method designed for different modalities of data (such as time-series data, enterprise relationship graph structure data, and enterprise market environment text data) based on the standard attention mechanism. This mechanism introduces a modality-specific attention mask (variant mask) to adjust the calculation method of attention weights, achieving differentiated processing and weighted fusion of various data types. Furthermore, combined with a dynamic relationship strength learning function, the variant attention mechanism can adaptively adjust the weight distribution of different types of relationships between enterprises, accurately capturing key factors in risk transmission paths and improving the expressive power of multimodal information fusion and the accuracy of risk identification.

[0123] Optionally, the variant attention mechanism provided in this embodiment of the invention achieves modality-differentiated information fusion by introducing a modality-specific attention mask corresponding to the data type and an adaptive learning function of relation strength based on the interaction of query vector, key vector and value vector.

[0124] The context vector refers to the fused feature vector obtained by weighting and normalizing the query vector, key vector, and value vector through a variant attention mechanism. The context vector comprehensively reflects the correlation and importance information among the input multimodal features, serving as a unified representation after multimodal fusion for subsequent dynamic relationship reasoning and risk identification.

[0125] Specifically, in this embodiment of the invention, a variant attention mechanism can receive feature vectors (e.g., abnormal funding features, node embeddings, negative public opinion semantic vectors) from temporal, graph structure, and textual modalities as input in the multimodal inference layer. The variant attention mechanism contains 12 attention heads, and the model dimension is 512. Its core lies in introducing modality-specific variant mask matrices for different modal data (temporal, structural, semantic), thereby enabling the standard attention computation (e.g., ...) to be performed correctly. ,in, , , These are the query vector, key vector, and value vector, respectively. For feature dimension, Based on the scaling factor, the algorithm introduces variant masks for different modal data. Adjust attention calculation: For example, for structural relationships like "equity control," the mask guides attention to assign them higher weights; for temporal patterns like "abnormal transactions," a different attention calculation strategy is used. In this way, the mechanism adaptively integrates information from different sources, ultimately outputting a context vector that comprehensively reflects multimodal association information, providing a unified and semantically rich representation for subsequent reasoning.

[0126] S410. Based on context vectors and enterprise relationship networks, identify explicit relationships and implicit associations in enterprise relationship networks.

[0127] Among them, explicit relationships refer to the relationships between enterprises that are clearly proven by legal and business documents such as business registration, contracts, and transaction records in the enterprise relationship network, such as equity control, guarantee relationships, and transaction exchanges.

[0128] Implicit connections refer to potential inter-enterprise relationships that are not covered by traditional registration or explicit recording systems but are discovered through multimodal data fusion and inference analysis. Examples include implicit control or influence links inferred through abnormal fund flows, cross-appointment of key natural persons, and semantic correlations in public opinion.

[0129] Specifically, embodiments of the present invention can use a dynamic relationship inferencer with a corporate relationship network constructed from context vectors and a dynamic graph layer as input. First, explicit relationships are identified and verified based on known connections in the corporate relationship network. Utilizing the cross-modal fusion information contained in the context vectors, implicit relationships not directly reflected in business registration or traditional graphs are inferred by analyzing potential association patterns between nodes, such as "implicit actual control formed through key natural persons" or "undisclosed potential risk contagion paths".

[0130] This invention employs a variant attention mechanism to achieve differentiated fusion of multimodal features and combines it with enterprise relationship networks. This enables the accurate identification of explicit and implicit relationships between enterprises, thereby comprehensively revealing complex risk associations and improving the accuracy and depth of risk identification.

[0131] Optional, based on Figure 1 The method shown is as follows: Figure 5The diagram shows a specific implementation of step S150 in the dynamic intelligent risk control method provided by this invention. Step S150 may specifically include:

[0132] S500 utilizes multimodal feature vectors and risk transmission information to generate an interpretable mathematical model of risk transmission through neural symbolic regression fitting.

[0133] Among them, the risk transmission mathematical model refers to a mathematical expression automatically fitted from multimodal feature vectors and risk transmission information using a neural symbolic regression algorithm. This expression is used to quantitatively describe the propagation patterns and impact mechanisms of financial risk in corporate relationship networks. The risk transmission mathematical model combines basic mathematical functions with specific functions in the financial field to clarify the weights and interactions of various risk factors (such as equity risk, transaction risk, and public opinion risk), exhibiting good predictive accuracy and business interpretability.

[0134] Specifically, embodiments of the present invention can integrate multimodal feature vectors, including time-series cash flow features, graph-structured enterprise relationship features, and textual semantic features, and combine them with identified risk transmission information to construct an input dataset for symbolic regression. Based on a predefined mathematical function library, the neural symbolic regression engine automatically derives the risk transmission dynamic equation by searching for the best-fitting model in the mathematical expression space, while considering both fitting accuracy and expression complexity. This quantifies the influence weights of different risk factors and their interaction relationships, achieving accurate modeling of the risk transmission mechanism with good business interpretability, supporting subsequent risk assessment and decision analysis.

[0135] As examples, in the interpretability layer, the neural symbolic regression engine receives feature vectors (such as the temporal anomaly degree and correlation strength of enterprise nodes) and historical risk transmission labels as input from the multimodal inference layer. The engine has a built-in symbol library containing basic mathematical functions such as addition, multiplication, exponentiation, and logarithms, as well as financial domain-specific functions (such as the risk contagion coefficient function), and sets a complexity penalty coefficient of 0.01. The neural symbolic regression engine employs a neural-guided optimization process to search the mathematical expression space for a mathematical model that best fits the relationship between the input data and the risk labels, for example: ,in, Risk value; , and These are quantitative indicators for equity risk, transaction risk, and public opinion risk, respectively. , and An interpretable equation is generated for each weight. This process balances goodness of fit with model complexity, ensuring that the final mathematical model is both predictive and provides business personnel with a clear explanation of risk drivers and weights.

[0136] S510. Extract key risk transmission paths from corporate relationship networks.

[0137] Among them, the key risk transmission path refers to the specific sequence of paths through which risks are transmitted from one enterprise to another in the enterprise relationship network, identified by causal path extraction algorithms. This includes the enterprise nodes involved and the explicit or implicit relationships between them. This path reflects the actual route of risk transmission, and the strength and confidence of each relationship in the path have been rigorously screened. It can reveal the key links and transmission chains of risk diffusion, providing intuitive evidence for risk early warning and control.

[0138] Specifically, embodiments of the present invention can identify risk transmission paths based on a dynamic enterprise relationship network constructed from a dynamic graph layer using a causal path extraction module. First, edge relationships that meet the minimum confidence threshold are selected, limiting the maximum path depth. Then, combining the dynamic relationship strength and risk transmission prediction results, a shortest path algorithm or heuristic search method is used to extract high-confidence critical paths from the source enterprise to the target enterprise. Each path includes detailed information on each node enterprise and its related relationships, as well as risk transmission strength indicators, intuitively reflecting the risk propagation chain and providing precise evidence for risk control.

[0139] As examples, embodiments of the present invention can operate using a causal path extractor based on a dynamic graph layer-maintained enterprise relationship network (including nodes, edges, superedges, and their dynamic weights) and identified risk sources and target enterprises. The causal path extractor sets a minimum confidence threshold (e.g., 0.7) and a maximum path depth (e.g., 4 hops), and employs a path search algorithm (e.g., Dijkstra's algorithm) to find the connection sequence in the network with the highest cumulative relationship strength and meeting the confidence requirements between the risk source (e.g., the high-risk enterprise "E Investment") and the core target enterprise (e.g., "Group A"). This process identifies and outputs key risk transmission chains, such as "Wang Moumou (key natural person) → Company B (abnormal transaction) → Subsidiaries C / D (equity control) → Group A," and attaches evidence for each step in the path (e.g., specific transaction records, equity relationships), thereby clearly revealing the specific propagation channels of risk in the complex network.

[0140] S520: Based on the risk transmission mathematical model and key risk transmission paths, generate a natural language report containing risk descriptions and transmission logic.

[0141] Natural language reports, in particular, refer to risk analysis texts that are automatically generated using a natural language generation module based on risk transmission mathematical models and key risk transmission paths, and that meet the professional requirements of financial business. The report content may include descriptions of risk events, causal relationships of risks, explanations of transmission paths, and early warning recommendations. By using easily understandable language, complex symbolic logic and reasoning processes are transformed into decision-making reference materials that business personnel and regulatory agencies can directly utilize, thus achieving transparency and actionability in risk assessment.

[0142] Specifically, embodiments of the present invention can invoke a natural language report generator, inputting a risk transmission mathematical model and key risk transmission path data. Combining this with a pre-set risk description template library and professional language specifications for the business domain, the generator automatically transforms complex symbolic logic and path relationships into easily understandable text content. The report details the background of the risk event, key enterprises along the transmission path and their relationships, transmission intensity, and expected impact time, clearly defining the transmission mechanism and risk diffusion logic. Simultaneously, the report can provide risk warning levels and control recommendations, ensuring that business personnel and regulatory agencies can quickly and accurately understand the risk situation, supporting scientific risk management and decision-making.

[0143] As examples, embodiments of the present invention can receive risk transmission mathematical models and key risk transmission paths as core inputs through a natural language report generator. The natural language report generator has a built-in risk description template library and style consistency control rules that meet financial professional requirements. It transforms symbolic mathematical equations (such as risk increase = 0.6 × (transaction anomaly degree) + 0.3 × (related cluster risk)) into business language such as "risk is mainly affected by transaction anomaly degree and related cluster risk, with weights of 0.6 and 0.3 respectively." Simultaneously, it transforms structured transmission paths into narrative logic such as "risk is transmitted through key individual Wang Moumou, via Company B, subsidiaries C / D, to Group A." Finally, by integrating these elements, a professional natural language report with a complete structure, clear logic, and including qualitative risk descriptions, quantitative analysis, transmission path diagrams, and response recommendations is automatically generated, directly serving risk management decisions.

[0144] This invention combines multimodal feature vectors, enterprise relationship networks, and risk transmission information, uses neural symbolic regression to generate a mathematical model of risk transmission and extract key risk transmission paths, and then outputs risk descriptions and transmission logic in natural language. This achieves accurate, interpretable, and easy-to-understand analysis results on the causes and transmission mechanisms of enterprise risks, significantly improving the scientific nature of risk identification and the transparency of decision-making.

[0145] Although the operations are described in a specific order, this should not be construed as requiring these operations to be performed in the specific order shown or in a sequential order. In certain environments, multitasking and parallel processing may be advantageous.

[0146] It should be understood that the various steps described in the method embodiments of the present invention may be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of the present invention is not limited in this respect.

[0147] Corresponding to the above method embodiments, this invention also provides a dynamic intelligent risk control device, the structure of which is as follows: Figure 6 As shown, it may include: a multimodal data acquisition unit 10, a multimodal feature vector acquisition unit 20, an enterprise relationship network update unit 30, a risk association identification unit 40, a risk transmission information acquisition unit 50, and a risk analysis result generation unit 60.

[0148] The multimodal data acquisition unit 10 is used to acquire multimodal data of the target enterprise.

[0149] The multimodal feature vector acquisition unit 20 is used to extract and align features from multimodal data to obtain multimodal feature vectors.

[0150] The enterprise relationship network update unit 30 is used to dynamically update the enterprise relationship network of the target enterprise using multimodal feature vectors.

[0151] Risk association identification unit 40 is used to identify risk associations between enterprises by fusing multimodal feature vectors and enterprise relationship networks.

[0152] The risk transmission information acquisition unit 50 is used to predict the transmission path and impact of risks in the enterprise relationship network based on the identified risk associations, and to obtain risk transmission information.

[0153] The risk analysis result generation unit 60 is used to generate interpretable risk analysis results by utilizing multimodal feature vectors, corporate relationship networks, and risk transmission information.

[0154] Optionally, multimodal data includes time-series data on the target company's corporate behavior, corporate relationship diagram structure data, and textual data on the company's market environment.

[0155] Optionally, the multimodal feature vector acquisition unit 20 may include a feature extraction subunit and a feature vector generation subunit.

[0156] The feature extraction subunit is used to perform differentiated encoding on enterprise behavior time series data, enterprise relationship graph structure data and enterprise market environment text data, and extract the corresponding time series features, graph structure features and text features.

[0157] The feature vector generation subunit is used to map temporal features, graph structure features, and text features to a unified semantic space, generating aligned multimodal feature vectors.

[0158] Optionally, the feature extraction subunit can be used to encode corporate behavior time-series data using a pre-defined temporal neural network architecture to extract time-series features representing dynamic capital characteristics; to encode corporate relationship graph structure data using a dynamic hypergraph neural network to extract graph structure features representing multiple relationships between enterprises; and to encode corporate market environment text data using a pre-trained language model to extract text semantic features.

[0159] Optionally, the enterprise relationship network update unit 30 can be used to dynamically update the relationship strength between nodes in the enterprise relationship network of the target enterprise based on the multimodal feature vector within the sliding time window; and to expand the set of hyperedges in the enterprise relationship network through clustering methods to capture inter-enterprise meta-relationships.

[0160] Optionally, the risk association identification unit 40 can be used to perform modal differentiation information fusion on multimodal feature vectors using a variant attention mechanism to generate a fused context vector; based on the context vector and the enterprise relationship network, it can identify explicit relationships and implicit associations in the enterprise relationship network.

[0161] Optionally, the variant attention mechanism achieves modality-differentiated information fusion by introducing a modality-specific attention mask corresponding to the data type and a function that adaptively learns the relation strength based on the interaction of the query vector, key vector, and value vector.

[0162] Optionally, the risk analysis result generation unit 60 can be used to generate an interpretable risk transmission mathematical model by using multimodal feature vectors and risk transmission information through neural symbolic regression fitting; extract key risk transmission paths from the enterprise relationship network; and generate a natural language report containing risk description and transmission logic based on the risk transmission mathematical model and key risk transmission paths.

[0163] This invention provides a dynamic intelligent risk control device, which is used to: acquire multimodal data of a target enterprise; extract and align features from the multimodal data to obtain multimodal feature vectors; dynamically update the enterprise relationship network of the target enterprise using the multimodal feature vectors; identify risk associations between enterprises by fusing the multimodal feature vectors and the enterprise relationship network; predict the transmission path and impact of risks in the enterprise relationship network based on the identified risk associations, and obtain risk transmission information; and generate interpretable risk analysis results using the multimodal feature vectors, the enterprise relationship network, and the risk transmission information. This invention constructs an end-to-end, logically transparent risk analysis framework by deeply fusing multimodal features, dynamically updated enterprise relationship networks, and risk transmission path information. While predicting risks, it intuitively reveals the formation, transmission, and evidence chain of risks, thereby achieving accurate risk identification and credible interpretation.

[0164] Regarding the apparatus in the above embodiments, the specific manner in which each unit performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0165] The dynamic intelligent risk control device includes a processor and a memory. The multimodal data acquisition unit 10, multimodal feature vector acquisition unit 20, enterprise relationship network update unit 30, risk association identification unit 40, risk transmission information acquisition unit 50, and risk analysis result generation unit 60 are all stored as program units in the memory. The processor executes the above-mentioned program units stored in the memory to realize the corresponding functions.

[0166] The processor contains a kernel, which retrieves the corresponding program units from memory. One or more kernels can be configured. By adjusting kernel parameters, multimodal data of the target enterprise is acquired, features are extracted and aligned to obtain a multimodal feature vector, and the enterprise relationship network is dynamically updated based on this vector. Multimodal features and the relationship network are then fused to identify risk associations between enterprises, predict the transmission path and impact of risks within the network, and obtain risk transmission information. Finally, by combining multimodal features, enterprise relationship networks, and risk transmission information, interpretable risk analysis results are generated.

[0167] This invention provides a computer-readable storage medium storing a program thereon, which, when executed by a processor, implements the dynamic intelligent risk control method.

[0168] This invention provides a processor for running a program, wherein the program executes the dynamic intelligent risk control method during runtime.

[0169] like Figure 7As shown, this embodiment of the invention provides an electronic device 1000, which includes at least one processor 1001, at least one memory 1002 connected to the processor 1001, and a bus 1003. The processor 1001 and the memory 1002 communicate with each other via the bus 1003. The processor 1001 is used to call program instructions in the memory 1002 to execute the aforementioned dynamic intelligent risk control method. The electronic device in this document can be a server, PC, PAD, mobile phone, etc.

[0170] The present invention also provides a computer program product that, when executed on an electronic device, is suitable for executing a program that initializes a dynamic intelligent risk control method.

[0171] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatuses, electronic devices (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable device, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0172] In a typical configuration, an electronic device includes one or more processors (CPUs), memory, and a bus. The electronic device may also include input / output interfaces, network interfaces, etc.

[0173] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, like read-only memory (ROM) or flash RAM, and memory includes at least one memory chip. Memory is an example of computer-readable media.

[0174] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0175] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this invention are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0176] It is understood that before using the technical solutions disclosed in the various embodiments of this disclosure, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in this disclosure in an appropriate manner in accordance with relevant laws and regulations, and user authorization should be obtained.

[0177] In the description of this invention, it should be understood that if the terms "upper", "lower", "front", "rear", "left" and "right" are used to indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the position or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention.

[0178] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0179] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0180] The above are merely embodiments of the present invention and are not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the present invention.

Claims

1. A dynamic intelligent risk control method, characterized in that, include: Obtain multimodal data of the target company; Feature extraction and alignment are performed on the multimodal data to obtain multimodal feature vectors; The enterprise relationship network of the target enterprise is dynamically updated using the multimodal feature vectors. By fusing the multimodal feature vectors and the enterprise relationship network, risk associations between enterprises can be identified; Based on the identified risk associations, the transmission path and impact of risks in the enterprise relationship network are predicted to obtain risk transmission information; Using the multimodal feature vectors, the enterprise relationship network, and the risk transmission information, interpretable risk analysis results are generated.

2. The method of claim 1, wherein, The multimodal data includes the target enterprise's time-series corporate behavior data, enterprise relationship diagram structure data, and enterprise market environment text data. The step of extracting and aligning features from the multimodal data to obtain a multimodal feature vector includes: Differential encoding is performed on the enterprise behavior time series data, the enterprise relationship graph structure data, and the enterprise market environment text data respectively, and corresponding time series features, graph structure features, and text features are extracted. The temporal features, graph structure features, and text features are mapped to a unified semantic space to generate aligned multimodal feature vectors.

3. The method of claim 2, wherein, The process involves differentially encoding the enterprise behavior time-series data, the enterprise relationship graph structure data, and the enterprise market environment text data, respectively, and extracting corresponding time-series features, graph structure features, and text features, including: The enterprise behavior time-series data is encoded using a preset temporal neural network architecture to extract time-series features that characterize the dynamic features of funds; The enterprise relationship graph structure data is encoded using a dynamic hypergraph neural network to extract graph structure features that represent the multi-dimensional relationships between enterprises; The enterprise market environment text data is encoded using a pre-trained language model to extract text semantic features.

4. The method of claim 1, wherein, The step of dynamically updating the enterprise relationship network of the target enterprise using the multimodal feature vector includes: Based on the multimodal feature vectors within the sliding time window, the relationship strength between nodes in the enterprise relationship network of the target enterprise is dynamically updated; The set of hyperedges in the enterprise relationship network is expanded by clustering methods to capture inter-enterprise meta-relationships.

5. The method of claim 1, wherein, The method of identifying risk associations between enterprises by fusing the multimodal feature vectors and the enterprise relationship network includes: A variant attention mechanism is used to fuse modal differences in the multimodal feature vectors to generate a fused context vector. Based on the context vector and the enterprise relationship network, the explicit relationships and implicit associations in the enterprise relationship network are identified.

6. The method of claim 5, wherein, The variant attention mechanism achieves modality-differentiated information fusion by introducing a modality-specific attention mask corresponding to the data type and an adaptive learning function based on the interaction of query vector, key vector, and value vector to learn the strength of the relationship.

7. The method of claim 1, wherein, The process of generating interpretable risk analysis results using the multimodal feature vectors, the enterprise relationship network, and the risk transmission information includes: Using the multimodal feature vectors and the risk transmission information, an interpretable mathematical model of risk transmission is generated through neural symbolic regression fitting; Extract key risk transmission paths from the aforementioned corporate relationship network; Based on the aforementioned risk transmission mathematical model and the aforementioned key risk transmission paths, a natural language report containing risk descriptions and transmission logic is generated.

8. A dynamic intelligent risk control device, characterized in that, include: The system includes a multimodal data acquisition unit, a multimodal feature vector acquisition unit, an enterprise relationship network update unit, a risk correlation identification unit, a risk transmission information acquisition unit, and a risk analysis result generation unit. The multimodal data acquisition unit is used to acquire multimodal data of the target enterprise; The multimodal feature vector acquisition unit is used to extract and align features from the multimodal data to obtain multimodal feature vectors. The enterprise relationship network update unit is used to dynamically update the enterprise relationship network of the target enterprise using the multimodal feature vector; The risk association identification unit is used to identify risk associations between enterprises by fusing the multimodal feature vectors and the enterprise relationship network; The risk transmission information acquisition unit is used to predict the transmission path and impact of risks in the enterprise relationship network based on the identified risk associations, and to obtain risk transmission information. The risk analysis result generation unit is used to generate interpretable risk analysis results by utilizing the multimodal feature vector, the enterprise relationship network, and the risk transmission information.

9. A computer-readable storage medium having stored thereon a program, characterized in that, When the program is executed by the processor, it implements the dynamic intelligent risk control method as described in any one of claims 1 to 7.

10. An electronic device, comprising: The electronic device includes at least one processor, at least one memory connected to the processor, and a bus; wherein the processor and the memory communicate with each other through the bus; the processor is used to call program instructions in the memory to execute the dynamic intelligent risk control method as described in any one of claims 1 to 7.

Images