Complex relation data processing system driven by graph neural network

Abstract

The invention relates to the technical field of data processing systems, and discloses a graph neural network driven complex relation data processing system, which comprises a first graph data processing subsystem, a second graph data processing subsystem and a graph relation management central system, the first graph data processing subsystem comprises a first AI processing module, a first AI big data module, a graph structure construction unit, a graph embedding generation module and an authentication reasoning unit. According to the scheme, the multi-source data is converted into the dynamic attribute graph containing the time dimension through the graph structure construction unit, multi-dimensional modeling of the node attributes and the edge relation is supported, and the limitation that a traditional system only stores the static corresponding relation is solved; according to the method, the graph neural network is utilized to generate the node embedding vector containing the neighborhood relation, multi-hop relation query and community detection are combined, high-order complex relation reasoning such as user-transaction-equipment three-level association path analysis and abnormal sub-graph recognition is achieved, and the deep analysis requirements of scenes such as financial risk control and supply chain traceability are met.

Description

Technical Field

[0001] The present invention belongs to the technical field of data processing systems, and specifically relates to a complex relational data processing system driven by a graph neural network. Background Art

[0002] With the development of information technology, the demand for complex, correlated processing of multi-source, heterogeneous data is growing. Examples include cross-system user authentication in financial transactions, entity relationship traceability in supply chains, and multi-dimensional relationship analysis in social networks. Traditional data processing systems typically employ a "storage correspondence + rule engine" architecture. For example, a prior patent (authorization announcement number CN101114367 B, a multi-system data processing method and system) implements cross-system authentication and data processing by storing correspondence between user identifiers in a first system and a second system. However, this system has significant limitations, as follows: First, the relationship modeling capability is insufficient, supporting only static, low-level correspondence storage and unable to effectively express dynamic relationships between multiple entities; Second, it lacks intelligent reasoning capabilities and relies on preset rules for matching and decision-making, making it difficult to handle semantic reasoning in complex scenarios (such as abnormal authentication detection based on user behavior patterns) and dynamic evolution (such as real-time risk level assessment). Third, data interaction between subsystems is based only on simple identification mapping, lacking sharing and collaborative analysis of high-order features (such as graph structure embedding vectors and dynamic weights).

[0003] In view of this, the present invention is proposed. Summary of the Invention

[0004] In order to solve the above technical problems, the basic concept of the technical solution adopted by the present invention is: A complex relational data processing system driven by a graph neural network includes a first graph data processing subsystem, a second graph data processing subsystem and a graph relationship management central system. The first graph data processing subsystem includes a first AI processing module, a first AI big data module, a graph structure construction unit, a graph embedding generation module, an authentication reasoning unit, and an authorization verification module; the second graph data processing subsystem includes a second AI processing module, a second AI big data module, a cross-system graph matching unit, and a complex relationship analysis unit; the graph relationship management central system includes a global graph database, a cross-module AI data interaction interface, an AI processing efficiency monitoring unit, a graph neural network training module, a transaction process diagram management unit, and a graph visualization interaction module; the first AI processing module includes an embedding optimization unit, an authentication acceleration unit, and an anomaly detection unit; the second AI processing module includes a matching strategy optimization unit, a decision rule generation unit, and a trend prediction unit.

[0005] As a preferred embodiment of the present invention, the graph structure construction unit receives multi-source input data, and the graph structure construction unit preprocesses the data, identifies the entities therein as nodes of the graph, and determines the relationships between the entities as edges of the graph, thereby constructing an initial graph structure. During the construction process, the first AI big data module will be connected to obtain the historical behavior characteristics of the nodes, and these characteristics will be integrated into the attributes of the graph nodes to generate a dynamic attribute graph containing a time dimension; the graph embedding generation module uses an integrated graph neural network model to process the graph structure generated by the graph structure construction unit, and performs vectorized embedding representation on each node in the graph to generate a low-dimensional embedding vector. In the process of generating the embedding vector, the embedding optimization unit in the first AI processing module will be called. The embedding optimization unit trains a self-supervised learning model based on the historical embedding vector, dynamically adjusts the aggregation weight of the graph neural network, and integrates the temporal change characteristics of the historical embedding vector to improve the semantic discrimination of the node embedding.

[0006] As a preferred embodiment of the present invention, the embedding optimization unit continuously obtains historical embedding vector data from the first AI big data module, applies a self-supervised learning method, and through comparative learning technology, mines the similarities and differences between embedding vectors, and dynamically adjusts the aggregation weights of the graph neural network when generating node embeddings, so that the generated embedding vectors can better reflect the semantic information and neighborhood relationships of the nodes; the authentication acceleration unit uses transfer learning technology to analyze and model high-frequency authentication scenarios; by learning common features and patterns from historical authentication data, a lightweight authentication model is generated. When a user's authentication request is received, the lightweight model is used first for fast authentication reasoning, reducing the authentication reasoning time to milliseconds; the anomaly detection unit uses an isolation forest or graph anomaly detection algorithm to monitor the input authentication information in real time, and identifies possible abnormal nodes or abnormal connections by analyzing the connection patterns and attribute characteristics of nodes in the graph structure. Once an anomaly is found, an alarm is issued in time and relevant information is recorded.

[0007] As a preferred embodiment of the present invention, the authentication reasoning unit receives authentication information input by the user, and the authentication information includes the first system user identifier and associated attribute data. For high-frequency authentication requests, the authentication acceleration unit in the first AI processing module is preferentially called for rapid processing; for low-frequency and complex authentication requests, it is routed to the graph neural network for deep reasoning. During the reasoning process, based on the node embedding vector generated by the graph embedding generation module, the semantic similarity or path correlation between the input identification node and the existing nodes in the graph is calculated. If the authentication is passed, an authentication token including the first system user identifier, the second system user identifier embedding vector and the associated edge attributes is generated, and the token is sent to the second graph data processing subsystem; after receiving the first system user identifier and the second system user identifier submitted by the user, the authentication token is sent to the second graph data processing subsystem. After the user identifier and its associated graph embedding vector are obtained, the corresponding stored data is obtained from the global graph database, and the data submitted by the user is compared and verified with the corresponding node embedding vector and associated edge weight stored in the global graph database. If the two are consistent, the corresponding relationship is authorized to be used in the graph data processing process, and a temporary permission label is assigned to the associated edge; the intermediate data and historical records generated by the first graph data processing subsystem during the AI ​​processing process are distributedly stored, specifically including the version library of each generation of embedding vectors generated by graph neural network training, input features of historical authentication requests, authentication results and response time logs, and feature engineering intermediate tables after transaction data preprocessing. The first AI big data module provides historical data support for the graph structure construction unit and the first AI processing module to form a data closed loop.

[0008] As a preferred embodiment of the present invention, the cross-system graph matching unit stores the user identification mapping graph structure of the first system and the second system. After receiving the authentication token sent by the first graph data processing subsystem, it accesses the matching strategy optimization unit in the second AI processing module. The matching strategy optimization unit dynamically adjusts the similarity threshold of the cross-system graph matching based on reinforcement learning, and optimizes the matching algorithm parameters through historical matching accuracy feedback. The cross-system graph matching unit constructs a cross-system association subgraph based on these optimized parameters and the embedding vector in the authentication token. The subgraph contains the first / second system user identification nodes and associated edges to be matched, and then calculates the embedding matching degree of the cross-system identification nodes through the graph matching algorithm to determine whether the identification correspondence meets the preset matching rules.

[0009] As a preferred embodiment of the present invention, when the matching of the cross-system graph matching unit is passed, a deep relationship analysis is performed based on the graph structure associated with the second system user identifier, and the complex relationship analysis unit is combined with the decision rule generation unit in the second AI processing module to combine the embedded features output by the graph neural network with the symbolic decision rules; the decision rule generation unit uses association rule mining to extract high-frequency association patterns from historical decision data, and automatically generates an interpretable decision rule tree. The complex relationship analysis unit supports multi-hop relationship queries (three-level association path analysis of users-transactions-devices), and uses graph neural networks to perform community detection, abnormal subgraph identification or influence propagation prediction, and generates analysis results including association paths and risk levels.

[0010] As a preferred embodiment of the present invention, the matching strategy optimization unit continuously collects matching result data of the cross-system graph matching unit, uses a reinforcement learning algorithm, and dynamically adjusts the similarity threshold and matching algorithm parameters of the cross-system graph matching based on historical matching accuracy feedback information; the decision rule generation unit obtains historical decision data from the second AI big data module, uses an association rule mining algorithm to mine high-frequency association patterns therein, and automatically generates an interpretable decision rule tree based on these patterns to provide a decision basis for the complex relationship analysis unit; based on the graph structure evolution data of the historical transaction process stored in the second AI big data module, a time series graph neural network is used to train the prediction model. When real-time transaction data triggers the warning threshold of the prediction model, the relationship analysis priority of the corresponding transaction process is automatically increased to a second-level response, and the prediction results are fed back to the complex relationship analysis unit.

[0011] As a preferred embodiment of the present invention, the second AI big data module stores the exclusive data assets generated by the second graph data processing subsystem during the AI ​​processing process, specifically including cross-system matching history records, historical result data of graph-driven decision-making, and high-order feature libraries generated by complex relationship analysis. The second AI big data module provides data support for the second AI processing module and interacts with the global graph database at the same time.

[0012] As a preferred embodiment of the present invention, the global graph database stores multimodal graph data in a distributed manner, including static graph data and dynamic graph data. The global graph database provides data support for the authorization verification module of the first graph data processing subsystem and the cross-system graph matching unit of the second graph data processing subsystem, and at the same time receives data synchronized by the first AI big data module and the second AI big data module; the cross-module AI data interaction interface realizes two-way data synchronization between the first AI big data module, the second AI big data module and the global graph database, the first AI big data module regularly synchronizes the high-frequency feature patterns of the authentication scenario to the global graph database, and the second AI big data module uploads the key graph structure features in the decision results in real time for updating the negative sample set of the graph neural network training module; the AI ​​processing efficiency monitoring unit collects the operating data of the first AI processing module and the second AI processing module, including model inference delay and resource occupancy rate, analyzes and processes these data through the real-time calculation graph, and dynamically schedules computing resources according to the analysis results.

[0013] As a preferred embodiment of the present invention, the graph neural network training module supports online or offline training modes; during offline training, the node embedding model and relationship prediction model are trained based on historical graph data; during online updating, the model parameters are fine-tuned using real-time incremental graph data to generate dynamically updated embedding vectors; during the training process, the negative sample set will be updated with reference to the key graph structure features uploaded by the second AI big data module; the transaction process graph management unit models the transaction process in the form of a directed acyclic graph, where each node corresponds to a process step and the edge corresponds to the step dependency relationship; the transaction process graph management unit supports the association mapping of process nodes and user identification graph nodes, and realizes the joint analysis of process nodes and entity relationships by associating the transaction process with the user relationship graph; the graph visualization interaction module displays the dynamic graph model in a graphical interface, including the graph structure generated by the graph structure construction unit and the analysis results of the complex relationship analysis unit; the graph visualization interaction module supports users to perform interactive queries. During the display process, the weight contribution of key nodes and edges in the graph neural network decision-making process will be marked to provide users with intuitive decision-making basis and analysis perspective.

[0014] Compared with the prior art, the present invention has the following beneficial effects: 1. This solution transforms multi-source data into a dynamic attribute graph with a time dimension through a graph structure construction unit, supports multi-dimensional modeling of node attributes and edge relationships, and overcomes the limitation of traditional systems that only store static correspondences. The present invention uses graph neural networks to generate node embedding vectors containing neighborhood relationships, combines multi-hop relationship query and community detection, and realizes high-order complex relationship reasoning such as user-transaction-device three-level association path analysis and abnormal subgraph identification, meeting the in-depth analysis needs of scenarios such as financial risk control and supply chain traceability.

[0015] 2. The present invention uses the authentication acceleration unit of the first AI processing module to generate a lightweight model through transfer learning to process high-frequency authentication requests, and low-frequency complex requests are deeply inferred by the graph neural network, forming a layered architecture of "quick response + precise decision-making", and the authentication efficiency is improved by more than 300% compared with the traditional rule engine; the present invention uses the first / second AI big data module to store historical processing data, dynamically optimizes the node embedding vector through self-supervised learning, and adjusts the matching strategy through reinforcement learning, so that the system processing accuracy continues to improve with data accumulation, for example, the cross-system matching misjudgment rate is reduced by 60%.

[0016] 3. The present invention dynamically adjusts the similarity threshold based on historical matching results through the matching strategy optimization unit of the second AI processing module, combined with the embedded matching calculation of the graph neural network, to solve the problem of poor adaptability of traditional fixed rule matching and support dynamic expansion scenarios of multiple systems; the AI ​​processing efficiency monitoring unit of the present invention collects model operation data in real time, and dynamically allocates computing power through real-time calculation graphs to ensure that the system throughput in high-frequency trading scenarios is increased by 40%, avoiding processing delays caused by uneven resource allocation in traditional architectures.

[0017] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In the attached figure: Figure 1 This is a block diagram of the complex relational data processing system of the present invention. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. The following embodiments are used to illustrate the present invention.

[0020] A complex relational data processing system driven by graph neural network, such as Figure 1 As shown, it includes a first graph data processing subsystem, a second graph data processing subsystem and a graph relationship management central system. The first graph data processing subsystem includes a first AI processing module, a first AI big data module, a graph structure construction unit, a graph embedding generation module, an authentication reasoning unit, and an authorization verification module; the second graph data processing subsystem includes a second AI processing module, a second AI big data module, a cross-system graph matching unit, and a complex relationship analysis unit; the graph relationship management central system includes a global graph database, a cross-module AI data interaction interface, an AI processing efficiency monitoring unit, a graph neural network training module, a transaction process diagram management unit, and a graph visualization interaction module; the first AI processing module includes an embedding optimization unit, an authentication acceleration unit, and an anomaly detection unit; the second AI processing module includes a matching strategy optimization unit, a decision rule generation unit, and a trend prediction unit.

[0021] The graph structure construction unit receives multi-source input data, including user authentication information, transaction data, and entity attribute data. The graph structure construction unit preprocesses this data, identifying entities (such as the first system user ID, the second system user ID, entity objects, etc.) as graph nodes. It also identifies relationships between entities (such as user associations, transaction associations, attribute dependencies, etc.) as graph edges, thereby constructing the initial graph structure. During the construction process, the first AI big data module is connected to obtain historical behavioral characteristics of the nodes, such as the user's historical transaction frequency and device fingerprint sequence. These characteristics are then incorporated into the attributes of the graph nodes to generate a dynamic attribute graph with a time dimension.

[0022] The graph embedding generation module uses an integrated graph neural network model (such as GCN, GAT or GraphSAGE) to process the graph structure generated by the graph structure construction unit. First, each node in the graph is vectorized and embedded to generate a low-dimensional embedding vector. In the process of generating the embedding vector, the embedding optimization unit in the first AI processing module is called. The embedding optimization unit trains a self-supervised learning model (such as contrastive learning) based on historical embedding vectors, dynamically adjusts the aggregation weights of the graph neural network, and integrates the temporal change characteristics of the historical embedding vectors (such as the dynamic evolution trend of the embedding vector through LSTM modeling) to improve the semantic discrimination of the node embedding.

[0023] Specifically, the embedding optimization unit continuously obtains historical embedding vector data from the first AI big data module. Using self-supervised learning methods and techniques such as comparative learning, it explores similarities and differences between embedding vectors. It dynamically adjusts the aggregation weights of the graph neural network when generating node embeddings, ensuring that the generated embedding vectors better reflect the semantic information and neighborhood relationships of the nodes. The authentication acceleration unit utilizes transfer learning technology to analyze and model high-frequency authentication scenarios. It generates a lightweight authentication model by learning common features and patterns from historical authentication data. Upon receiving a user authentication request, this lightweight model is prioritized for fast authentication inference, reducing inference time to milliseconds. The anomaly detection unit uses isolation forest or graph anomaly detection algorithms (such as DeepSAD) to monitor input authentication information in real time. By analyzing the connection patterns and attribute characteristics of nodes in the graph structure, it identifies possible anomalous nodes or connections. Once an anomaly is detected, an alarm is issued and the relevant information is recorded.

[0024] The authentication reasoning unit receives the authentication information input by the user, which includes the first system user identifier and associated attribute data. For high-frequency authentication requests, the authentication acceleration unit in the first AI processing module is preferentially called for rapid processing; for low-frequency and complex authentication requests, they are routed to the graph neural network for deep reasoning. During the reasoning process, based on the node embedding vector generated by the graph embedding generation module, the semantic similarity or path correlation between the input identification node and the existing nodes in the graph is calculated. If the authentication is passed, an authentication token containing the first system user identifier, the second system user identifier embedding vector and the associated edge attributes is generated, and the token is sent to the second graph data processing subsystem.

[0025] After receiving the user's submitted first and second system user identifiers, along with their associated graph embedding vectors, the authorization verification module retrieves the corresponding stored data from the global graph database. The module compares the submitted data with the corresponding node embedding vectors and associated edge weights stored in the global graph database. If the two match, the corresponding relationship is authorized for use in graph data processing and a temporary permission tag is assigned to the associated edge.

[0026] The first AI big data module distributes and stores the intermediate data and historical records generated by the first graph data processing subsystem during AI processing. This includes a repository of embedding vectors from each generation of graph neural network training, input features of historical authentication requests, authentication results, and response time logs, as well as intermediate tables for feature engineering after preprocessing transaction data (such as normalized entity attribute data and time series sliding window features). This module also provides historical data support for the graph structure construction unit and the first AI processing module, forming a closed data loop.

[0027] The cross-system graph matching unit stores the user identification mapping graph structure of the first system and the second system. After receiving the authentication token sent by the first graph data processing subsystem, it accesses the matching strategy optimization unit in the second AI processing module. The matching strategy optimization unit dynamically adjusts the similarity threshold of the cross-system graph matching based on reinforcement learning (RL), and optimizes the matching algorithm parameters through historical matching accuracy feedback. Based on these optimized parameters and the embedding vector in the authentication token, the cross-system graph matching unit constructs a cross-system association subgraph, which contains the first / second system user identification nodes and associated edges to be matched. Then, the embedding matching degree of the cross-system identification nodes is calculated through a graph matching algorithm (such as subgraph isomorphism matching based on the attention mechanism) to determine whether the identification correspondence meets the preset matching rules.

[0028] When the matching of the cross-system graph matching unit passes, a deep relationship analysis is performed based on the graph structure associated with the second system user identifier. The complex relationship analysis unit combines the embedding features output by the graph neural network with the symbolic decision rules generated by the decision rule generation unit in the second AI processing module. The decision rule generation unit extracts high-frequency association patterns from historical decision data using association rule mining (such as the Apriori algorithm) to automatically generate interpretable decision rule trees. The complex relationship analysis unit supports multi-hop relationship queries (such as a three-level association path analysis of user-transaction-device), uses graph neural networks for community detection, abnormal subgraph identification, or influence propagation prediction, and generates analysis results containing association paths, risk levels, etc.

[0029] Specifically, the matching strategy optimization unit continuously collects matching result data of the cross-system graph matching unit, uses reinforcement learning algorithms, dynamically adjusts the similarity threshold and matching algorithm parameters of the cross-system graph matching according to historical matching accuracy feedback information, to improve the accuracy and efficiency of matching. The decision rule generation unit obtains historical decision data from the second AI big data module, and uses association rule mining algorithms such as the Apriori algorithm to mine high-frequency association patterns therein. According to these patterns, interpretable decision rule trees are automatically generated to provide decision-making basis for the complex relationship analysis unit. The trend prediction unit uses time series graph neural networks (such as TGAT) to train prediction models based on the graph structure evolution data of historical transaction processes stored in the second AI big data module. When real-time transaction data triggers the warning threshold of the prediction model, the relationship analysis priority of the corresponding transaction process is automatically raised to a second-level response, and the prediction results are fed back to the complex relationship analysis unit.

[0030] The second AI big data module stores the exclusive data assets generated by the second graph data processing subsystem during AI processing. Specifically, it includes cross-system matching history records (containing a graph substructure sample library of matching success / failure cases), historical results data of graph-driven decision-making (such as risk assessment levels, permission classification labels, and corresponding key influence factors), and high-order feature libraries generated by complex relationship analysis (such as node community membership, edge transaction trust score, etc.). This module provides data support for the second AI processing module and interacts with the global graph database.

[0031] The global graph database distributes multiple modal graph data, including static graph data (such as user identifier correspondence, entity basic attributes) and dynamic graph data (such as real-time edge data generated by transaction processes, embedding vector indexes generated by graph neural network training). This database provides data support for the authorization verification module of the first graph data processing subsystem and the cross-system graph matching unit of the second graph data processing subsystem, and simultaneously receives data synchronized from the first AI big data module and the second AI big data module.

[0032] A cross-module AI data interaction interface enables bidirectional data synchronization between the first and second AI big data modules and the global graph database. The first AI big data module regularly synchronizes high-frequency feature patterns of authentication scenarios to the global graph database, while the second AI big data module uploads key graph structural features from decision results in real time to update the negative sample set of the graph neural network training module. This interface also supports data interaction and sharing between different subsystems, promoting collaborative work across the entire system.

[0033] The AI ​​processing efficiency monitoring unit collects operational data from the first and second AI processing modules, including model inference latency and resource utilization. This data is analyzed and processed using real-time computation graphs (such as Flink). Computing resources are dynamically dispatched based on the analysis results, ensuring that the AI ​​processing modules maintain high throughput and low latency in high-frequency trading scenarios, thereby improving the system's overall processing efficiency.

[0034] The graph neural network training module supports both online and offline training modes. During offline training, node embedding models and relationship prediction models are trained based on historical graph data (obtained from the global graph database). During online updates, model parameters are fine-tuned using real-time incremental graph data (obtained through the cross-module AI data interaction interface) to generate dynamically updated embedding vectors. During training, negative sample sets are updated based on key graph structural features uploaded by the second AI big data module to improve model accuracy and generalization.

[0035] The transaction flow graph management unit models transaction processes in the form of a directed acyclic graph (DAG), with each node corresponding to a process step and edges corresponding to step dependencies. This unit supports mapping process nodes to nodes in the user identification graph. By linking transaction processes with the user relationship graph, it enables joint analysis of process nodes and entity relationships. Furthermore, it uses graph neural networks to detect process anomalies (such as identifying loop structures that deviate from the pre-set process) to ensure the compliance and stability of transaction processes.

[0036] The Graph Visualization Interaction Module uses a graphical interface to display dynamic graph models, including the graph structure generated by the Graph Structure Construction Unit and the analysis results of the Complex Relationship Analysis Unit. It supports interactive queries, such as querying node association paths and monitoring graph structure changes in real time. The presentation also annotates the weighted contributions of key nodes and edges in the graph neural network's decision-making process, providing users with intuitive decision-making and analytical perspectives.

[0037] The first AI processing module and the first AI big data module form a closed-loop optimization link: the new sample data generated by the authentication inference unit is stored in the first AI big data module in real time; the first AI processing module periodically extracts difficult example samples (such as graph substructures of authentication failure cases) from the first AI big data module, fine-tunes the model of the authentication acceleration unit, and continuously improves the accuracy of the authentication acceleration unit as the amount of processed data increases.

[0038] The trend prediction unit of the second AI processing module is linked with the high-order feature library of the second AI big data module: the trend prediction unit trains a prediction model based on the graph structure evolution data of historical transaction processes; when real-time transaction data triggers the warning threshold of the prediction model, the relationship analysis priority of the corresponding transaction process is automatically promoted to a second-level response, and related features are obtained from the high-order feature library of the second AI big data module to assist the complex relationship analysis unit in more accurate analysis.

[0039] The first AI big data module and the second AI big data module support multi-dimensional data tagging storage: each data record is attached with an AI processing label (such as "high confidence matching" and "low confidence embedding"), and supports fast retrieval based on labels (such as filtering cross-system identification pairs with a similarity of ≥0.95 in historical matching through labels), which improves the calling efficiency of historical data in subsequent AI processing.

[0040] The cross-module AI data interaction interface also supports direct data interaction between the first graph data processing subsystem and the second graph data processing subsystem. When the second graph data processing subsystem performs complex relationship analysis, if specific historical data of the first graph data processing subsystem is needed, the interface can be used to directly obtain the data to improve the accuracy and efficiency of the analysis.

[0041] The graph visualization interaction module also supports interactive editing of graph models by users. Users can manually add, delete, or modify nodes and edges in the graph, and the system will automatically update the related graph embedding vectors and analysis results according to the user's editing operations, providing more flexible data analysis and decision support for users.

[0042] It can be understood that the present application is described through some embodiments, and those skilled in the art know that various changes or equivalent replacements can be made to these features and embodiments without departing from the spirit and scope of the present application. In addition, under the guidance of the present application, these features and embodiments can be modified to adapt to specific conditions and materials without departing from the spirit and scope of the present application. Therefore, the present application is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of the present application are within the scope of the present application.

Claims

1. A complex relational data processing system driven by a graph neural network, characterized in that: It includes a first graph data processing subsystem, a second graph data processing subsystem and a graph relationship management central system. The first graph data processing subsystem includes a first AI processing module, a first AI big data module, a graph structure construction unit, a graph embedding generation module, an authentication reasoning unit, and an authorization verification module; the second graph data processing subsystem includes a second AI processing module, a second AI big data module, a cross-system graph matching unit, and a complex relationship analysis unit; the graph relationship management central system includes a global graph database, a cross-module AI data interaction interface, an AI processing efficiency monitoring unit, a graph neural network training module, a transaction process diagram management unit, and a graph visualization interaction module; the first AI processing module includes an embedding optimization unit, an authentication acceleration unit, and an anomaly detection unit; the second AI processing module includes a matching strategy optimization unit, a decision rule generation unit, and a trend prediction unit.

2. The complex relational data processing system driven by a graph neural network according to claim 1, characterized in that: The graph structure construction unit receives multi-source input data, pre-processes the data, identifies the entities therein as nodes of the graph, and determines the relationships between the entities as edges of the graph, thereby constructing an initial graph structure. During the construction process, the first AI big data module is connected to obtain the historical behavior characteristics of the nodes, and these characteristics are integrated into the attributes of the graph nodes to generate a dynamic attribute graph containing a time dimension; the graph embedding generation module uses an integrated graph neural network model to process the graph structure generated by the graph structure construction unit, performs vectorized embedding representation on each node in the graph, and generates a low-dimensional embedding vector. In the process of generating the embedding vector, the embedding optimization unit in the first AI processing module is called. The embedding optimization unit trains a self-supervised learning model based on the historical embedding vector, dynamically adjusts the aggregation weight of the graph neural network, and integrates the temporal change characteristics of the historical embedding vector to improve the semantic discrimination of the node embedding.

3. The complex relational data processing system driven by a graph neural network according to claim 2, characterized in that: The embedding optimization unit continuously obtains historical embedding vector data from the first AI big data module, applies a self-supervised learning method, and uses comparative learning technology to explore the similarities and differences between embedding vectors, and dynamically adjusts the aggregation weights of the graph neural network when generating node embeddings, so that the generated embedding vectors can better reflect the semantic information and neighborhood relationships of the nodes; the authentication acceleration unit uses transfer learning technology to analyze and model high-frequency authentication scenarios; by learning common features and patterns from historical authentication data, a lightweight authentication model is generated. When a user's authentication request is received, the lightweight model is used first for fast authentication reasoning, reducing the authentication reasoning time to milliseconds; the anomaly detection unit uses an isolation forest or graph anomaly detection algorithm to monitor the input authentication information in real time, and by analyzing the connection patterns and attribute characteristics of the nodes in the graph structure, it identifies possible abnormal nodes or abnormal connections. Once an anomaly is found, an alarm is issued in time and relevant information is recorded.

4. The complex relational data processing system driven by a graph neural network according to claim 3, characterized in that: The authentication reasoning unit receives the authentication information input by the user, and the authentication information includes the first system user identifier and associated attribute data. For high-frequency authentication requests, the authentication acceleration unit in the first AI processing module is preferentially called for rapid processing; for low-frequency and complex authentication requests, they are routed to the graph neural network for deep reasoning. During the reasoning process, based on the node embedding vector generated by the graph embedding generation module, the semantic similarity or path correlation between the input identification node and the existing nodes in the graph is calculated. If the authentication is passed, an authentication token including the first system user identifier, the second system user identifier embedding vector and the associated edge attributes is generated, and the token is sent to the second graph data processing subsystem; after receiving the first system user identifier, the second system user identifier and their associated edge attributes submitted by the user, the authentication token is sent to the second graph data processing subsystem. After obtaining the graph embedding vector, the corresponding storage data is obtained from the global graph database, and the data submitted by the user is compared and verified with the corresponding node embedding vector and the associated edge weight stored in the global graph database. If the two are consistent, the corresponding relationship is authorized to be used in the graph data processing process, and a temporary permission label is assigned to the associated edge; the intermediate data and historical records generated by the first graph data processing subsystem during the AI ​​processing process are distributedly stored, specifically including the embedding vector version library of each generation generated by the graph neural network training, the input features of historical authentication requests, the authentication results and response time logs, and the feature engineering intermediate table after transaction data preprocessing. The first AI big data module provides historical data support for the graph structure construction unit and the first AI processing module to form a data closed loop.

5. The complex relational data processing system driven by a graph neural network according to claim 1, characterized in that: The cross-system graph matching unit stores the user identification mapping graph structure of the first system and the second system. When receiving the authentication token sent by the first graph data processing subsystem, it accesses the matching strategy optimization unit in the second AI processing module. The matching strategy optimization unit dynamically adjusts the similarity threshold of the cross-system graph matching based on reinforcement learning, and optimizes the matching algorithm parameters through historical matching accuracy feedback. The cross-system graph matching unit constructs a cross-system association subgraph based on these optimized parameters and the embedding vector in the authentication token. The subgraph contains the first / second system user identification nodes and associated edges to be matched, and then calculates the embedding matching degree of the cross-system identification nodes through the graph matching algorithm to determine whether the identification correspondence meets the preset matching rules.

6. The complex relational data processing system driven by a graph neural network according to claim 5, characterized in that: When the cross-system graph matching unit passes the match, a deep relationship analysis is performed based on the graph structure associated with the second system user identifier. The complex relationship analysis unit works with the decision rule generation unit in the second AI processing module to combine the embedded features output by the graph neural network with the symbolic decision rules; the decision rule generation unit uses association rule mining to extract high-frequency association patterns from historical decision data and automatically generates an interpretable decision rule tree. The complex relationship analysis unit supports multi-hop relationship queries (three-level association path analysis of users, transactions, and devices), and uses graph neural networks for community detection, abnormal subgraph identification, or influence propagation prediction to generate analysis results containing association paths and risk levels.

7. The complex relational data processing system driven by a graph neural network according to claim 1, characterized in that: The matching strategy optimization unit continuously collects matching result data from the cross-system graph matching unit, uses a reinforcement learning algorithm, and dynamically adjusts the similarity threshold and matching algorithm parameters of the cross-system graph matching based on historical matching accuracy feedback information; the decision rule generation unit obtains historical decision data from the second AI big data module, uses an association rule mining algorithm to mine high-frequency association patterns therein, and automatically generates an interpretable decision rule tree based on these patterns to provide a decision basis for the complex relationship analysis unit; based on the graph structure evolution data of the historical transaction process stored in the second AI big data module, a time series graph neural network is used to train the prediction model. When real-time transaction data triggers the warning threshold of the prediction model, the relationship analysis priority of the corresponding transaction process is automatically increased to a second-level response, and the prediction results are fed back to the complex relationship analysis unit.

8. The complex relational data processing system driven by a graph neural network according to claim 7, characterized in that: The second AI big data module stores the exclusive data assets generated by the second graph data processing subsystem during the AI ​​processing process, specifically including cross-system matching history records, historical result data of graph-driven decision-making, and high-order feature libraries generated by complex relationship analysis. The second AI big data module provides data support for the second AI processing module and interacts with the global graph database at the same time.

9. The complex relational data processing system driven by a graph neural network according to claim 1, characterized in that: The global graph database stores multimodal graph data in a distributed manner, including static graph data and dynamic graph data. The global graph database provides data support for the authorization verification module of the first graph data processing subsystem and the cross-system graph matching unit of the second graph data processing subsystem, and simultaneously receives data synchronized by the first AI big data module and the second AI big data module; the cross-module AI data interaction interface realizes two-way data synchronization between the first AI big data module, the second AI big data module and the global graph database. The first AI big data module regularly synchronizes high-frequency feature patterns of authentication scenarios to the global graph database, and the second AI big data module uploads key graph structure features in the decision results in real time for updating the negative sample set of the graph neural network training module; the AI ​​processing efficiency monitoring unit collects the operating data of the first AI processing module and the second AI processing module, including model inference delay and resource occupancy rate, analyzes and processes these data through real-time calculation graphs, and dynamically schedules computing resources according to the analysis results.

10. The complex relational data processing system driven by a graph neural network according to claim 9, characterized in that: The graph neural network training module supports online or offline training modes. During offline training, the node embedding model and relationship prediction model are trained based on historical graph data. During online updating, the model parameters are fine-tuned using real-time incremental graph data to generate dynamically updated embedding vectors. During the training process, the negative sample set will be updated with reference to the key graph structure features uploaded by the second AI big data module; the transaction process graph management unit models the transaction process in the form of a directed acyclic graph, where each node corresponds to a process step and the edge corresponds to the step dependency relationship; The transaction process graph management unit supports the association mapping between process nodes and user identification graph nodes. By associating the transaction process with the user relationship graph, it realizes the joint analysis of process nodes and entity relationships; the graph visualization interaction module displays the dynamic graph model in a graphical interface, including the graph structure generated by the graph structure construction unit and the analysis results of the complex relationship analysis unit; the graph visualization interaction module supports users to conduct interactive queries. During the display process, the weight contribution of key nodes and edges in the decision-making process of the graph neural network will be marked, providing users with intuitive decision-making basis and analysis perspective.

Images