Knowledge graph construction method and device, equipment and storage medium

Abstract

The application relates to a knowledge graph construction method, which identifies and determines multiple entity types participating in activities according to the requirement of constructing a knowledge graph, and defines the basic relationship between the entity types. By constructing the entity types and the basic relationship into an initial graph structure in the form of nodes and edges, a preliminary knowledge graph is formed. According to the time characteristics of the basic relationship, a time sequence characteristic weight matrix is created, which is stored in the preliminary knowledge graph together with the attributes of the basic relationship. Then, the direction characteristics are extracted from the basic relationship to generate a direction characteristic matrix, which is stored together with other characteristics. The multi-dimensional characteristics of the nodes and edges in the preliminary knowledge graph are combined and optimized to finally generate an updated representation and form a target knowledge graph. The application can not only dynamically adapt to the knowledge requirements in different scenarios, but also significantly improve the performance of the knowledge graph in time sequence analysis and directional reasoning by introducing the time sequence characteristics and the direction characteristics.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, apparatus, device and storage medium for constructing a knowledge graph. Background Technology

[0002] In the financial sector, with the increasing complexity of financial transactions and the dramatic growth in data volume, effectively managing and analyzing this data has become crucial. Traditional financial data analysis methods often rely on rule-driven or statistical models, which face numerous challenges when dealing with complex financial relationship networks, especially when analyzing multi-relationship, multi-dimensional financial transaction data. This data typically involves interactions between multiple entities (such as customers, accounts, and enterprises), and these interactions exhibit highly dynamic temporal and directional characteristics.

[0003] To better understand and manage these complex financial relationship networks, knowledge graphs have been introduced as an emerging technological tool into financial data analysis. Knowledge graphs can represent entities and their relationships in a graph structure, forming a complete network graph by representing entities as nodes and relationships as edges. This graph can not only help identify and analyze direct relationships between entities, but also uncover potential complex relationships through reasoning and inference.

[0004] However, existing knowledge graph construction methods have limitations in handling dynamic features (such as temporal and directional features) in financial transactions. For example, traditional methods often struggle to effectively capture changes in fund flows and the evolution of relationships over time, which is crucial in applications such as anti-money laundering, risk management, and fund flow analysis. To overcome these challenges, more advanced feature processing and optimization techniques, such as graph neural networks (GNNs), need to be introduced into the knowledge graph construction process to enhance the effectiveness of knowledge graphs in complex financial environments. Summary of the Invention

[0005] The main objective of this invention is to provide a knowledge graph construction method, apparatus, device, and storage medium, aiming to solve the technical problem that existing technologies cannot effectively handle the temporal and directional characteristics of basic relationships between entities, resulting in deficiencies in dynamic change analysis and flow identification in complex networks.

[0006] To achieve the above objectives, the present invention provides a knowledge graph construction method, comprising:

[0007] Based on the need to construct a knowledge graph, multiple entity types participating in the activity are identified and determined, and the basic relationships between these entity types are defined.

[0008] The entity types and the basic relationships are constructed into an initial graph structure in the form of nodes and edges, forming a preliminary knowledge graph;

[0009] Based on the temporal characteristics of the basic relations, a temporal feature weight matrix is ​​created, and the temporal feature weight matrix is ​​stored together with the attributes of the basic relations in the graph structure of the preliminary knowledge graph.

[0010] Extract directional features from each basic relation, create a directional feature matrix for each basic relation based on the extracted directional features, and store the directional feature matrix together with the attributes of the basic relation in the graph structure of the preliminary knowledge graph;

[0011] The multi-dimensional features of nodes and edges in the preliminary knowledge graph are combined and optimized to generate an updated representation, which is then applied to the preliminary knowledge graph to form the target knowledge graph.

[0012] In one embodiment, the entity types and the basic relationships are constructed into an initial graph structure in the form of nodes and edges to form a preliminary knowledge graph, including:

[0013] The entity types and the basic relationships are constructed into an initial graph structure in the form of nodes and edges, forming a preliminary relationship network;

[0014] The preliminary relational network is expanded into a complete knowledge graph structure. The node types and edge types in the knowledge graph are determined, and feature vectors are assigned to each node and edge to form the basic representation of the knowledge graph.

[0015] Create a vector representation for each basic relation, containing the type information and temporal characteristics of the basic relation;

[0016] The attributes of the basic relationships are embedded into the node and edge representations of the knowledge graph, and joint learning is performed with the node embedding to form the preliminary knowledge graph.

[0017] In one embodiment, the attributes of the basic relationships are embedded into the node and edge representations of the knowledge graph, and jointly learned with the node embeddings to form the preliminary knowledge graph, including:

[0018] Relation type features are extracted from each basic relation. These relation type features describe the specific properties of the basic relation and are represented as a first embedding vector.

[0019] The event occurrence time feature is extracted from each basic relation. The event occurrence time feature is used to represent the time sequence or time point of the occurrence of the basic relation and is represented as a second Embedding vector.

[0020] After combining the first Embedding vector with the second Embedding vector, it is input together with the Embedding vector of the corresponding node into the CompGCN model for joint learning;

[0021] Through joint learning of the CompGCN model, updated feature representations of nodes and edges are generated, and these updated feature representations are integrated into the graph structure to generate the preliminary knowledge graph.

[0022] In one embodiment, the multi-dimensional features of nodes and edges in the preliminary knowledge graph are combined and optimized to generate an updated representation, which is then applied to the preliminary knowledge graph to form the target knowledge graph, including:

[0023] For each node and edge in the preliminary knowledge graph, multi-dimensional features are embedded, and the attribute features, time features, direction features and relation type features of the nodes and edges are vectorized to form an initial feature representation;

[0024] The initial feature representation is input into the CompGCN model, and the multi-dimensional features of nodes and edges are aggregated through multi-layer graph convolution operations to integrate the feature information of nodes, their neighboring nodes, and associated edges.

[0025] After each graph convolution operation, a non-linear activation function is applied to activate the aggregated features, thereby improving the feature representation of nodes and edges.

[0026] After multi-layer graph convolution and non-linear activation processing, the features output from each layer are combined to generate updated representations of nodes and edges.

[0027] The updated node and edge representations are reapplied to the preliminary knowledge graph, replacing the original representations, to ultimately form the target knowledge graph.

[0028] In one embodiment, multi-dimensional features of nodes and edges are aggregated through multi-layer graph convolution operations, including:

[0029] The aggregation function is determined based on the multi-dimensional features of nodes and edges in the preliminary knowledge graph;

[0030] In each layer of graph convolution operation, the node's own features are combined with the features of its neighboring nodes to generate an updated node representation, and the edge's own features are combined with the features of its associated nodes to generate an updated edge representation.

[0031] In multi-layer graph convolution operations, aggregation functions are applied layer by layer. Each layer performs new aggregation and updates based on the node and edge representations generated in the previous layer, gradually optimizing the node and edge representations to capture deeper feature information.

[0032] In one embodiment, the entity types and the basic relationships are constructed into an initial graph structure in the form of nodes and edges to form a preliminary knowledge graph, including:

[0033] The entity types are represented as nodes, and corresponding nodes are created for each entity type in the graph structure;

[0034] The basic relationships are represented as edges, and corresponding edges are created for each basic relationship in the graph structure;

[0035] Add a reverse relation to each basic relation, forming a reverse edge in the graph structure that corresponds to the original relation;

[0036] Add a self-circulating edge to each node in the graph structure to preserve the node's own information;

[0037] The reverse edges and self-loop edges are processed into feature vectors, which are then represented as vectors and combined with the feature vectors of the basic relations to construct the initial graph structure.

[0038] The feature vectors of the nodes and edges are stored in the initial graph structure. By integrating basic relations, reverse edges, and self-looping edges, a complete preliminary knowledge graph is finally formed.

[0039] In one embodiment, directional features are extracted from each basic relationship, and a directional feature matrix is ​​created based on the extracted directional features, including:

[0040] Extract directional features from each basic relationship, where the directional features represent the direction of fund flow within the relationship;

[0041] The extracted directional features are mapped to binary values, where 0 represents no directionality or a bidirectional relationship, and 1 represents a definite directionality or a unidirectional relationship.

[0042] The mapped directional features are quantized and converted into quantized values ​​that can represent the strength or importance of directionality. A directional feature matrix is ​​then created based on the quantized directional features.

[0043] Furthermore, to achieve the above objectives, the present invention also provides a knowledge graph construction apparatus, comprising:

[0044] The entity and relationship definition module, based on the need to build a knowledge graph, identifies and determines multiple entity types participating in the activity, and defines the basic relationships between these entity types;

[0045] The initial graph structure construction module constructs the entity types and basic relationships into an initial graph structure in the form of nodes and edges, forming a preliminary knowledge graph;

[0046] The temporal feature processing module creates a temporal feature weight matrix based on the temporal features of the basic relationships, and stores the temporal feature weight matrix together with the attributes of the basic relationships in the graph structure of the preliminary knowledge graph.

[0047] The directional feature processing module extracts directional features from each basic relation, creates a directional feature matrix for each basic relation based on the extracted directional features, and stores the directional feature matrix together with the attributes of the basic relation in the graph structure of the preliminary knowledge graph.

[0048] The feature combination and optimization module combines and optimizes the multi-dimensional features of nodes and edges in the preliminary knowledge graph to generate updated representations, which are then applied to the preliminary knowledge graph to form the target knowledge graph.

[0049] Furthermore, to achieve the above objectives, the present invention also provides a computer device, the computer device including a memory, a processor, and a knowledge graph construction program stored in the memory and executable on the processor, wherein when the knowledge graph construction program is executed by the processor, it implements the steps of the knowledge graph construction method as described above.

[0050] Furthermore, to achieve the above objectives, the present invention also provides a computer storage medium storing a knowledge graph construction program, which, when executed by a processor, implements the steps of the knowledge graph construction method described above.

[0051] Beneficial Effects: This invention relates to a knowledge graph construction method. Based on the requirements for constructing a knowledge graph, it identifies and determines multiple entity types participating in activities and defines the basic relationships between these entity types. By constructing the entity types and basic relationships into an initial graph structure in the form of nodes and edges, a preliminary knowledge graph is formed. Based on the temporal characteristics of the basic relationships, a temporal feature weight matrix is ​​created and stored in the preliminary knowledge graph along with the attributes of the basic relationships. Next, directional features are extracted from the basic relationships to generate a directional feature matrix, which is stored along with other features. The multi-dimensional features of the nodes and edges in the preliminary knowledge graph are combined and optimized to finally generate an updated representation and form the target knowledge graph. This invention not only dynamically adapts to the knowledge needs of different scenarios but also significantly improves the performance of the knowledge graph in temporal analysis and directional reasoning by introducing temporal and directional features. Attached Figure Description

[0052] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings:

[0053] Figure 1 This is a flowchart illustrating an embodiment of the knowledge graph construction method of the present invention;

[0054] Figure 2 This is a schematic diagram of the functional modules of a preferred embodiment of the knowledge graph construction device of the present invention;

[0055] Figure 3 This is a schematic diagram of the hardware operating environment of the computer device involved in the embodiment of the present invention. Detailed Implementation

[0056] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0057] In the financial sector, with the increasing complexity of financial transactions and the dramatic growth in data volume, effectively managing and analyzing this data has become crucial. Traditional financial data analysis methods often rely on rule-driven or statistical models, which face numerous challenges when dealing with complex financial relationship networks, especially when analyzing multi-relationship, multi-dimensional financial transaction data. This data typically involves interactions between multiple entities (such as customers, accounts, and enterprises), and these interactions exhibit highly dynamic temporal and directional characteristics.

[0058] To better understand and manage these complex financial relationship networks, knowledge graphs have been introduced as an emerging technological tool into financial data analysis. Knowledge graphs can represent entities and their relationships in a graph structure, forming a complete network graph by representing entities as nodes and relationships as edges. This graph can not only help identify and analyze direct relationships between entities, but also uncover potential complex relationships through reasoning and inference.

[0059] However, existing knowledge graph construction methods have limitations in handling dynamic features (such as temporal and directional features) in financial transactions. For example, traditional methods often struggle to effectively capture changes in fund flows and the evolution of relationships over time, which is crucial in applications such as anti-money laundering, risk management, and fund flow analysis. To overcome these challenges, more advanced feature processing and optimization techniques, such as graph neural networks (GNNs), need to be introduced into the knowledge graph construction process to enhance the effectiveness of knowledge graphs in complex financial environments.

[0060] Please see Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the knowledge graph construction method provided by the present invention. It should be noted that although a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than that shown here.

[0061] like Figure 1 As shown, the knowledge graph construction method proposed in this invention includes the following steps:

[0062] S10, Based on the need to construct a knowledge graph, identify and determine multiple entity types participating in the activity, and define the basic relationships between the entity types;

[0063] In this embodiment, "activity" refers to any interaction, process, or operation involving multiple entities within a knowledge graph. These activities can be widely applied in various fields, such as:

[0064] Commercial activities, such as transactions, collaborations, and contract signings, may involve multiple entities, including customers, companies, and suppliers.

[0065] Social activities: such as communication, sharing, and recommendations in social networks; entities may include users, posts, comments, etc.

[0066] Technical processes: such as data flow, communication, and processing in a technical system, involving entities such as devices, applications, and data flows.

[0067] Scientific research: such as experimental processes and scientific research collaborations, involving researchers, experimental equipment, experimental data, etc.

[0068] In this embodiment and subsequent embodiments, financial activities are used as an example for illustration. Entity types in financial activities refer to key participants playing different roles in financial transactions or financial relationship networks. These entities can be individuals, businesses, financial institutions, or other legal entities, interconnected through various financial activities. Specific entity types include, but are not limited to: Customers: Individuals or businesses, typically acting as the source or destination of funds, playing a key role in transactions. Accounts: Financial accounts associated with customers or businesses, through which operations such as transfers, deposits, and loans are performed. Businesses: Companies or legal entities participating in financial activities, potentially involved in complex financial activities such as lending, investment, and trading. Financial Institutions: Banks, insurance companies, investment funds, etc., which act as intermediaries, guarantors, or fund managers in financial activities.

[0069] Based on the goals of the knowledge graph construction (such as anti-money laundering, credit assessment, etc.), key participants in financial activities are identified. These participants are typically entity types directly related to core business operations; for example, in an anti-money laundering scenario, customers, accounts, trading platforms, and banks are all key entities. To ensure the knowledge graph can perform effective reasoning, diverse entity types are determined to cover various roles in financial activities. For example, in lending relationships, in addition to borrowers and lending institutions, entities such as guarantors and insurance companies may also be included. The selection of entity types also needs to consider how these entities are connected through relationships and whether these relationships facilitate reasoning. For example, the "transaction" relationship between banks and customers may play a crucial role in anti-money laundering reasoning. The definition of entity types should consider integration with existing data, ensuring that entities obtained from different data sources can be accurately represented in the knowledge graph and can be interconnected.

[0070] In practical applications, entity type identification is not limited to static definitions but may also involve dynamic adjustments. For example, new entity types, such as "borrower," "guarantor," and "investor," may be introduced for different financial products or service scenarios. Accurately defining these entity types can lay the foundation for subsequent relationship building and knowledge graph formation.

[0071] Based on the identified entity types, the fundamental relationships between these entities are defined. Fundamental relationships refer to the most basic interactions and connections formed between these entities in financial activities. These relationships determine how funds flow between different entities and constitute the core content of the financial relationship network. Specific fundamental relationships include, but are not limited to: Transfer: The process of transferring funds from one account to another, typically involving interaction between a customer and a bank. Lending: An entity (such as an individual or business) borrows funds from a financial institution or other entity and promises to repay the principal and interest in the future. Shareholding: An individual or institution holds shares in a company, participating in the company's shareholder equity. Insurance: A protection service provided by a financial institution to a customer, paying compensation in the event of a specific risk event. Each relationship includes not only a static definition but may also involve relationship attributes such as "loan amount," "interest rate," and "guarantee period." In a complex financial network, the definition of fundamental relationships needs to fully consider the interaction patterns and business logic between entities to ensure the completeness and accuracy of the knowledge graph.

[0072] In one specific implementation, the financial transaction monitoring system first identifies and classifies various entity types participating in transactions. The system not only identifies standard financial entities, such as customers and bank accounts, but also expands the definition of entity types to include virtual currency wallets and cross-border payment platforms. The basic relationships between these entities are not limited to transfers and loans, but extend to include cross-border transactions and multi-currency exchanges. The system constructs a preliminary knowledge graph based on these extended definitions. When processing temporal features, the system creates a temporal feature weight matrix, considering not only the specific time the transaction occurred but also the impact of time zone differences and transaction delays. For directional features, the system assigns directional features to each financial relationship and further extends this to handle multi-directional relationships (such as multi-level fund transfer paths). Finally, by combining and optimizing multi-dimensional features, more complex node and edge representations are generated to cope with complex international financial transaction networks, ultimately forming the target knowledge graph.

[0073] In another specific implementation, the anti-fraud detection system identifies and classifies more granular entity types, such as high-frequency trading accounts, accounts associated with known fraudsters, and accounts registered using anonymous emails. When defining basic relationships, the system not only handles direct fund transfers but also extends to include indirect connections (such as multi-level transfers through intermediary accounts). When creating the temporal feature weight matrix, the system expands the scope of temporal features, considering the frequency and regularity of historical transactions, and uses machine learning models to predict future transaction time patterns. During the creation of the directional feature matrix, the system introduces adaptive directional feature processing, dynamically adjusting the directional feature matrix to adapt to constantly changing fraud patterns. Finally, through a more complex aggregation and optimization process, the system generates updated node and edge representations, enabling the knowledge graph to capture and identify signs of fraudulent behavior in real time.

[0074] By defining the entity types and fundamental relationships in financial activities, an accurate and comprehensive knowledge graph can be constructed, effectively reflecting the structure and dynamic changes of financial transaction networks. This enhances the accuracy and depth of financial data analysis, providing strong technical support for risk management, anti-money laundering, and transaction monitoring.

[0075] S20, construct the entity types and the basic relationships into an initial graph structure in the form of nodes and edges to form a preliminary knowledge graph;

[0076] In this embodiment, in financial activities, entity types and basic relationships are constructed into an initial graph structure in the form of nodes and edges to form a preliminary knowledge graph, in order to clearly represent different entities and their relationships in the graph.

[0077] Nodes are the basic units in a graph structure. Each node represents a financial entity type, such as a customer, account, or enterprise. By transforming entity types into nodes, the system can intuitively represent the position and role of these entities in the financial network.

[0078] An edge is a line connecting two nodes, representing the basic relationship between the two entities, such as transfer, loan, or shareholding. Each edge not only shows the existence of the relationship but can also include attribute information about the relationship, such as amount and timestamp, further enriching the data representation capabilities of the graph.

[0079] By integrating all identified nodes and edges into a unified graph structure, a preliminary knowledge graph is formed. This graph can provide a global view for subsequent data analysis and processing, helping to identify potential risks, discover emerging financial relationships, and provide a foundation for in-depth analysis.

[0080] In one specific implementation, the financial fraud detection system identifies the types of entities in financial transactions, such as individual customers, merchants, and bank accounts, representing them as nodes in a graph structure. The system also determines the financial relationships between these entities, such as payments, transfers, and refunds, and represents them as edges. Building upon this, the system extends the features of each edge to include information such as transaction frequency and time intervals, thereby constructing a preliminary knowledge graph. This knowledge graph is used to detect abnormal transaction patterns, such as money laundering through multiple small transfers.

[0081] By constructing a preliminary knowledge graph of financial entity types and basic relationships in the form of nodes and edges, the relational structure within the financial network can be comprehensively presented. This not only enhances the visualization and understanding of financial data but also provides a robust data foundation for subsequent risk management, anti-money laundering analysis, and credit assessment, significantly improving the accuracy and efficiency of financial data analysis.

[0082] S30, Based on the temporal characteristics of the basic relations, create a temporal feature weight matrix, and store the temporal feature weight matrix together with the attributes of the basic relations in the graph structure of the preliminary knowledge graph;

[0083] In this embodiment, a time-series feature weight matrix is ​​created based on the temporal characteristics of the underlying relationships in financial activities, in order to better capture and analyze the dynamic changes in financial transactions.

[0084] Temporal characteristics refer to the specific points in time or time periods during which transactions occur in financial relationships. These characteristics are crucial for understanding the timing of fund flows and identifying unusual transaction behavior. For example, frequent large transfers within a short period may be an indication of money laundering.

[0085] Based on the extracted temporal features, a temporal feature weight matrix is ​​constructed. The element values ​​of this matrix reflect the weight of the same type of underlying relationship at different points in time. This matrix can quantify the impact of temporal features on financial relationships and help identify the cumulative effect of funds over different periods.

[0086] The created time-series feature weight matrix is ​​stored together with other attributes of the basic relationships (such as amount, direction features, etc.) in the graph structure of the initial knowledge graph. By incorporating time features into the knowledge graph, the graph can more accurately reflect the dynamic changes in capital flows and provide richer data support for subsequent analysis.

[0087] In one specific implementation, the system extracts time features from cross-border transactions, creates a time-series feature weight matrix, and analyzes the spatiotemporal distribution characteristics of these transactions by combining the geographical location data of each transaction. Embedding this matrix into a preliminary knowledge graph enables the system to capture the time-series characteristics of cross-border money laundering activities, such as transactions conducted simultaneously at multiple locations. Ultimately, by analyzing these complex time patterns, the system can identify hidden money laundering networks and fund transfer chains.

[0088] By creating and applying a time-series feature weight matrix, the ability of the financial activity knowledge graph to process dynamic time features is significantly enhanced. Particularly in anti-money laundering scenarios, it enables more accurate identification and analysis of money laundering activities, improving financial institutions' ability to prevent illicit fund flows, thereby enhancing the security and compliance of the entire system.

[0089] S40, extract directional features from each basic relation, create a directional feature matrix for the basic relation based on the extracted directional features, and store the directional feature matrix together with the attributes of the basic relation in the graph structure of the preliminary knowledge graph;

[0090] In this embodiment, directional features are extracted from each basic relationship in financial activities, and a directional feature matrix is ​​created based on these directional features in order to better represent and analyze the direction of fund flows in the financial network.

[0091] Directional features refer to the direction in which funds flow within a financial relationship, such as from one account to another, or from one company to its subsidiary. Extracting these directional features clarifies the path of fund flows, which is particularly important in identifying complex financial activities such as money laundering and money transfers.

[0092] Based on the extracted directional features, a directional feature matrix is ​​created. This matrix represents the directionality of different basic relationships; for example, 0 indicates no directionality or a two-way relationship, while 1 indicates a clear directionality or a one-way relationship. Through this matrix representation, the system can better identify and analyze patterns in fund flows.

[0093] The directional feature matrix is ​​stored together with other basic relation attributes (such as amount and time features) in the graph structure of the initial knowledge graph. The purpose of this is to integrate directional features into the graph representation, so that the entire knowledge graph can not only show the existence of financial relationships, but also intuitively reflect the direction of fund flow, supporting subsequent analysis and reasoning.

[0094] In one specific implementation, the system analyzes the directional characteristics of cross-border fund transfers, focusing particularly on the path of funds moving from one country to another. The extracted directional characteristics are used to create a matrix for tracking the direction of international fund flows. By storing this information in a knowledge graph, the system can identify cross-border money laundering networks, especially in complex fund transfers involving multiple countries, and trace the final destination of funds, thereby enhancing the monitoring of illicit financial flows.

[0095] By extracting and applying directional features, the visualization and analysis capabilities of fund flows in financial networks can be significantly improved. Particularly in anti-money laundering scenarios, the use of directional feature matrices enables the system to more accurately identify complex fund flow paths and potential money laundering activities, thereby enhancing the security and risk management capabilities of the financial system.

[0096] S50 combines and optimizes the multi-dimensional features of nodes and edges in the preliminary knowledge graph to generate an updated representation, which is then applied to the preliminary knowledge graph to form the target knowledge graph.

[0097] In this embodiment, multi-dimensional features include node attributes (such as type and importance) and edge attributes (such as relationship type, weight, time characteristics, and direction characteristics). The combination of these features integrates information from different dimensions through algorithms to more comprehensively represent the roles and functions of nodes and edges in the financial network.

[0098] Optimization refers to adjusting the representations of nodes and edges using algorithms to make their roles in the graph structure more consistent with the dynamic changes of the actual financial transaction network. The optimization process may involve machine learning techniques such as Graph Convolutional Networks (GCNs), which iteratively calculate the graph structure to gradually approximate the optimal solution for the representation of each node and edge.

[0099] The combined and optimized nodes and edges generate new feature representations that more accurately reflect the state of entities and relationships within the knowledge graph. The generation of these new representations aims to improve reasoning and analysis within the graph.

[0100] The updated node and edge representations are then reapplied to the initial knowledge graph to form the final target knowledge graph. This target knowledge graph is the core data structure used by the system for financial data analysis, prediction, and reasoning.

[0101] In one specific implementation, firstly, each node and edge in the preliminary knowledge graph is embedded with multi-dimensional features. These features include the attribute features, temporal features, directional features, and relation type features of the nodes and edges.

[0102] The embedded multi-dimensional features are input into the CompGCN model, and feature aggregation and updating are performed using the following formula:

[0103]

[0104] in, Let f represent the new feature representation of node v in the (k+1)th layer; f represents the activation function, used to introduce nonlinearity and make the model more expressive; ∑ (u,r∈N(v)) This represents the summation of the neighbor nodes u and edges r of node v, where N(v) represents the set of neighbor nodes of node v. This represents the projection matrix of relation type r at the k-th level, used to transform the features of the relation type; The projection matrix of the temporal features of relation type r at the k-th level is used to map the temporal features to a new representation space; The projection matrix of the directional features of relation type r at the k-th level maps the directional features to a suitable representation space; Symbols represent node features Sum of edge features Specific operations between them. These can be Hadamard products (element-wise multiplication) or other forms of feature combination operations.

[0105] In this process, the system updates the feature representations of each node and edge layer by layer through multi-layer graph convolution operations. Each convolution operation integrates the feature information of neighboring nodes and edges into the target node. Through a non-linear activation function f, the system can enhance the expressive power of the features. After multiple layers of convolution operations, updated node and edge representations are generated. These updated representations can better reflect the position and relationship of nodes and edges in the entire knowledge graph. The updated node and edge representations are then reapplied to the initial knowledge graph, replacing the original representations, and finally generating the target knowledge graph. This target knowledge graph has stronger expressive power and can be used for more accurate reasoning and analysis.

[0106] In another specific implementation, the system expands the edges and nodes in the initial knowledge graph by adding geographical location features to analyze cross-border capital flows. The system generates updated node and edge representations by combining these multi-dimensional features (such as transaction amount, direction, and geographical location). Then, the system uses a clustering algorithm to optimize these representations, identifying patterns of capital flows between different countries and regions. The optimized representations are further integrated back into the initial knowledge graph to form a target knowledge graph that reflects international capital flow trends. This graph helps the system identify key paths in international money laundering networks.

[0107] In anti-money laundering systems, reasoning based on target knowledge graphs can identify potential money laundering activities. Target knowledge graphs, through prior multi-dimensional feature combination and optimization, contain rich information about financial entities and their relationships, such as transaction amounts, directional characteristics, time characteristics, and geographical locations. By utilizing this information, the system can perform deep reasoning to identify and analyze complex fund flow paths. Specific steps may include:

[0108] Preliminary Analysis and Node Screening: The system first screens high-risk nodes in the target knowledge graph. High-risk nodes are typically marked by features such as historical transaction records, geographical location, and fund flow direction. These nodes may represent suspicious accounts, entities that frequently conduct large transactions, or accounts that have participated in multiple cross-border transactions.

[0109] Path reasoning: Based on the selected high-risk nodes, the system initiates a path reasoning algorithm. Path reasoning focuses on all intermediate nodes and edges that the funds flow through from a high-risk node, ultimately reaching the target node. The system utilizes relational data in the knowledge graph to infer possible flow paths of funds and identify suspicious behaviors, such as circular transfers and tiered transfers.

[0110] Pattern matching: The system matches the inferred fund flow paths with known money laundering patterns. These patterns can include typical money laundering techniques such as splitting transactions, circular transactions, and cross-regional transactions. Through pattern matching, the system can quickly identify which paths meet the characteristics of money laundering and mark these paths as high-risk.

[0111] Analysis of Reasoning Results: Once the system identifies high-risk paths, further reasoning and analysis will focus on key nodes and relationships. The system will analyze the transaction frequency, transaction amount changes, and directional characteristics between these nodes to further verify whether these relationships are highly correlated with money laundering activities.

[0112] Report Generation and Early Warning: Finally, based on the reasoning results, the system generates a detailed analysis report, including identified high-risk fund flow paths, high-risk nodes involved, and potential money laundering patterns. The system also issues early warning signals to relevant departments, recommending further investigation and action.

[0113] By combining and optimizing the nodes and edges in the initial knowledge graph with multi-dimensional features, a more accurate and dynamic target knowledge graph can be generated. This significantly enhances the system's ability to analyze the flow of funds and changes in relationships within complex financial networks. Especially in anti-money laundering scenarios, it can more effectively identify hidden money laundering paths and abnormal fund flows, thereby improving the risk prevention and control capabilities of the financial system.

[0114] This invention relates to a knowledge graph construction method. Based on the requirements for constructing a knowledge graph, it identifies and determines multiple entity types participating in activities and defines the basic relationships between these entity types. An initial knowledge graph is formed by constructing the entity types and basic relationships as nodes and edges. A temporal feature weight matrix is ​​created based on the temporal characteristics of the basic relationships, and its attributes, along with those of the basic relationships, are stored in the initial knowledge graph. Next, directional features are extracted from the basic relationships to generate a directional feature matrix, which is stored along with other features. The multi-dimensional features of the nodes and edges in the initial knowledge graph are combined and optimized to generate an updated representation, forming the target knowledge graph. This invention not only dynamically adapts to the knowledge needs of different scenarios but also significantly improves the performance of the knowledge graph in temporal analysis and directional reasoning through the introduction of temporal and directional features.

[0115] In one embodiment, S20 includes:

[0116] S201, construct the entity types and the basic relationships into an initial graph structure in the form of nodes and edges to form a preliminary relationship network;

[0117] S202, the preliminary relationship network is expanded into a complete knowledge graph structure, the node types and edge types in the knowledge graph are determined, and feature vectors are assigned to each node and edge to form the basic representation of the knowledge graph;

[0118] S203, create a vector representation for each basic relation, containing the type information and temporal characteristics of the basic relation;

[0119] S204, embed the attributes of the basic relations into the node and edge representations of the knowledge graph, and perform joint learning with the node embedding to form the preliminary knowledge graph.

[0120] In this embodiment, all financial entities and the relationships between them are mapped to nodes and edges in a graph structure. Nodes represent financial entities (such as customers, accounts, companies, etc.), while edges represent basic relationships between entities (such as transfers, loans, etc.). In this way, a preliminary relationship network is formed, demonstrating the basic financial network structure.

[0121] The expansion builds upon the initial relational network, further enriching the graph's structure and content. Specifically, by adding more node and edge types (such as classifying customers as high-risk and low-risk, or transactions as normal and suspicious), the knowledge graph can more comprehensively represent complex financial networks.

[0122] A feature vector is a set of numerical vectors used to describe the attributes of a node or edge. For example, a node's feature vector might include the entity type, risk score, transaction frequency, etc.; while an edge's feature vector might include transaction amount, direction features, time features, etc. These feature vectors endow each node and edge in a knowledge graph with unique attributes, making the graph more expressive.

[0123] Detailed vector representations are created for each basic relationship, containing relationship-related type information (such as transfers and loans) and temporal characteristics (such as the time or frequency of the transaction). These vector representations help to accurately capture the characteristics of relationships within the graph and provide foundational data for subsequent analysis.

[0124] Embedding refers to integrating the attributes of basic relationships (such as time features, transaction amounts, etc.) into the representations of nodes and edges. This step, through joint learning, allows the representations of nodes and edges to influence and optimize each other, thereby generating a graph structure that more closely reflects reality. Through this joint learning process, the system ultimately generates a preliminary knowledge graph that can more accurately reflect the complexity and dynamism of financial relationships.

[0125] Through these steps, the system in this embodiment can expand from the initial relationship network to generate a more comprehensive and accurate preliminary knowledge graph. This graph not only displays the basic financial network structure but also captures more complex financial relationships and their characteristics. Particularly in anti-money laundering scenarios, the system significantly enhances its ability to identify complex fund flows and potential risks through joint learning and feature embedding, effectively improving the depth and accuracy of financial data analysis.

[0126] In one embodiment, S204 above includes:

[0127] S2041, Extract relation type features from each basic relation, the relation type features being used to describe the specific properties of the basic relation and being represented as a first Embedding vector;

[0128] S2042, Extract event occurrence time features from each basic relation. The event occurrence time features are used to represent the time sequence or time point of the occurrence of the basic relation and are represented as a second Embedding vector.

[0129] S2043, After combining the first Embedding vector with the second Embedding vector, input them together with the Embedding vector of the corresponding node into the CompGCN model for joint learning;

[0130] S2044, through joint learning of the CompGCN model, updated feature representations of nodes and edges are generated, and the updated feature representations of nodes and edges are integrated into the graph structure to generate the preliminary knowledge graph.

[0131] In this embodiment, relationship type features refer to information used to describe the nature of financial relationships, such as transfers, loans, and investments. These features are represented as vectors using specific algorithms (such as embedding methods), and this vector is the first embedding vector. This vector can express the nature of the relationship in numerical form, facilitating calculation and analysis. Event occurrence time features are used to record the time information of financial relationships, such as the time point or frequency of transactions. These time features are transformed into second embedding vectors using embedding methods, represented as numerical vectors to capture information in the time dimension. This is crucial for analyzing the temporality of fund flows and identifying abnormal transaction behavior. Combining the relationship type features (first embedding vector) with the time features (second embedding vector) forms a comprehensive feature vector. This feature vector, along with the embedding vector of the corresponding node, is input into the CompGCN (Graph Convolutional Network) model. Through joint learning, the model can capture complex financial relationships between nodes, especially in dynamic and temporally sequential scenarios. In the CompGCN model, the feature representations of nodes and edges are processed through multiple layers of convolution and activation functions to generate updated representations. These updated representations more accurately reflect the roles and relationships of nodes and edges within the financial network. Ultimately, these features are integrated into a graph structure, forming a preliminary knowledge graph. This knowledge graph not only encompasses static information about financial relationships but also dynamically reflects the complex paths of capital flows.

[0132] Through the above steps, this embodiment can generate a dynamic and rich preliminary knowledge graph, which can effectively analyze and identify complex capital flow patterns.

[0133] In one embodiment, the above S50 includes:

[0134] S501, embed multi-dimensional features into each node and edge in the preliminary knowledge graph, and vectorize the attribute features, time features, direction features and relationship type features of the nodes and edges to form an initial feature representation;

[0135] S502, the initial feature representation is input into the CompGCN model, and the multi-dimensional features of nodes and edges are aggregated through multi-layer graph convolution operation to integrate the feature information of nodes, their adjacent nodes and associated edges;

[0136] S503 applies a non-linear activation function after each graph convolution operation to activate the aggregated features, thereby improving the feature representation of nodes and edges.

[0137] S504 combines the features output from each layer after multi-layer graph convolution and non-linear activation processing to generate updated representations of nodes and edges.

[0138] S505, the updated node and edge representations are reapplied to the preliminary knowledge graph, replacing the original representations, and finally forming the target knowledge graph.

[0139] In this embodiment, multi-dimensional feature embedding refers to converting various attributes of nodes and edges (such as node type, edge relationship type, transaction time, directionality, etc.) into vector representations. These feature vectors provide rich descriptions for each node and edge, making them not just simple connections in the graph, but carrying specific information and capable of representing more complex relationships.

[0140] The initial feature representations are input into the CompGCN model, which processes these features through multi-layer graph convolutional network (GCN) operations. In each convolutional layer, the model aggregates the features of a node with the features of its neighboring nodes and their associated edges. This process, through layer-by-layer convolution, gradually integrates more contextual information, enabling the representation of each node and edge to contain a wider range of network information.

[0141] Nonlinear activation functions (such as ReLU and Sigmoid) are used to process aggregated features after each convolutional operation. By applying these activation functions, the model can introduce nonlinearity, making the feature representations of nodes and edges more complex and refined.

[0142] After processing through multiple convolutional layers and activation functions, the model generates updated node and edge representations. These updated representations incorporate multi-dimensional information, enabling them to more accurately reflect the roles and relationships of nodes and edges within the graph.

[0143] Finally, the system reapplies the updated node and edge representations to the initial knowledge graph, replacing the previous initial representations to form the target knowledge graph. This target knowledge graph has higher expressive power and analytical accuracy, and can support complex reasoning and decision-making processes, especially in financial network analysis and anti-money laundering scenarios.

[0144] In one specific implementation, the system not only extracts the basic features of nodes and edges (such as transaction amount, time features, and direction features), but also adds features such as risk scores and geographical location. These newly added multi-dimensional features further enrich the initial feature representation, enabling the system to more comprehensively analyze the flow paths of funds and potential money laundering risks in cross-border transactions.

[0145] This embodiment, through the steps described above, can generate an optimized target knowledge graph. Particularly in anti-money laundering scenarios, the system can more accurately analyze and identify complex fund flow paths. By combining multi-dimensional features and graph convolutional networks, the system enhances its ability to express financial relationships, enabling a deeper understanding and revelation of abnormal behaviors hidden within financial networks, thereby effectively strengthening the security and compliance of the financial system.

[0146] In one embodiment, in S502 above, the multi-dimensional features of nodes and edges are aggregated through multi-layer graph convolution operations, including:

[0147] S5021, Determine the aggregation function based on the multi-dimensional features of nodes and edges in the preliminary knowledge graph;

[0148] S5022, in each layer of graph convolution operation, the node's own features are combined with the features of its neighboring nodes to generate an updated node representation, and the edge's own features are combined with the features of its associated nodes to generate an updated edge representation.

[0149] S5023 applies aggregation functions layer by layer in multi-layer graph convolution operations. Each layer performs new aggregation and updates based on the node and edge representations generated in the previous layer, gradually optimizing the node and edge representations to capture deeper feature information.

[0150] In this embodiment, the system first analyzes the multi-dimensional features of each node and edge in the preliminary knowledge graph. These features may include attribute features, temporal features, directional features, relation type features, etc. Based on these features, the system selects an appropriate aggregation function, which will be used in subsequent graph convolution operations. The choice of aggregation function is crucial to the final node and edge representation, as it determines how to integrate feature information from different dimensions.

[0151] The choice of aggregation function can be based on the type of node and edge features. For example, for numerical features (such as transaction amount or account balance), a weighted average or summation function can be used to balance the importance of different features. For categorical features (such as account type or transaction type), a mode or voting mechanism can be used to determine the most representative category.

[0152] The aggregation function can also be chosen based on the ultimate goal. For example, if the goal is to detect abnormal trading behavior, a maximum value function might be preferred to highlight extreme values. If the goal is to capture the overall trend of the network, a weighted average might better reflect the overall characteristics.

[0153] When choosing an aggregation function, consider the complexity of the connections between nodes and edges in the network: In high-density networks (with many nodes and edges that are closely interconnected), a weighted average can be used to balance the influence of different nodes and edges. In sparse networks (with fewer nodes and edges that are loosely connected), a summation function can be used to enhance the influence of fewer connection features.

[0154] Some aggregation functions (such as weighted averages) are easy to interpret, while other complex aggregation methods may be difficult to understand. In certain application scenarios, interpretability may be an important criterion for selecting aggregation functions.

[0155] In graph convolution operations, the system combines the features of each node with the features of its neighbors. This combination process allows each node's representation to include more information about its surrounding environment. Similarly, the system combines the features of an edge with the features of its associated nodes to generate an updated edge representation. This combination process can capture deeper relational information in the network structure, making the representations of nodes and edges more accurate and context-sensitive.

[0156] The system optimizes the representations of nodes and edges layer by layer through multi-layer graph convolution operations. In each layer, the system applies a pre-defined aggregation function to perform new aggregations and updates based on the node and edge representations generated in the previous layer. Each convolution layer refines the representations of nodes and edges, gradually capturing deeper-level feature information. This layer-by-layer aggregation and optimization process ensures that the final node and edge representations fully reflect the complex relationships of the entire network.

[0157] In one specific implementation, the system incorporates geographic location features when processing cross-border capital flows to better analyze the paths and risks of cross-border transactions. A weighted average is chosen as the aggregation function to comprehensively consider the impact of transaction amount and geographic location. In the multi-layer graph convolution, the system uses four convolutional layers to capture deep-seated patterns in cross-border capital flows and employs the ReLU activation function to enhance non-linear expressiveness.

[0158] In another specific implementation, the system simplifies the feature dimensions for high-frequency trading scenarios, focusing only on transaction frequency and amount. The system selects a maximum value function as the aggregation function to highlight outliers in high-frequency trading. In graph convolution operations, the system uses only two convolutional layers to quickly detect abnormal patterns in high-frequency trading, and uses a sigmoid activation function to process the probability output.

[0159] This embodiment combines and optimizes the multi-dimensional features of nodes and edges in the preliminary knowledge graph, thereby generating more representative and accurate node and edge representations and producing a more precise target knowledge graph.

[0160] In one embodiment, S20 includes:

[0161] S205, Represent the entity type as a node, and create a corresponding node for each entity type in the graph structure;

[0162] S206, Represent the basic relationships as edges, and create corresponding edges for each basic relationship in the graph structure;

[0163] S207, add a reverse relation to each basic relation, forming a reverse edge in the graph structure that corresponds to the original relation;

[0164] S208, add a self-circulating edge to each node in the graph structure to preserve the node's own information;

[0165] S209, perform eigenvectorization on reverse edges and self-loop edges, represent reverse edges and self-loop edges as vectors, and combine them with the eigenvectors of the basic relations to construct the initial graph structure;

[0166] S210, the feature vectors of the nodes and edges are stored in the initial graph structure, and by integrating basic relations, reverse edges and self-looping edges, a complete preliminary knowledge graph is finally formed.

[0167] In this embodiment, the system first identifies entity types in all financial activities, such as customers, accounts, and companies. Each entity type is mapped to a node in a graph structure. These nodes serve as basic elements in the initial knowledge graph, representing key participants in the financial network. Basic relationships, such as transfers, loans, and investments, are represented as edges between nodes. Each relationship type is mapped to a corresponding edge in the graph structure, connecting different nodes and demonstrating the direct relationships and interactions between entities.

[0168] To capture bidirectional relationships in the financial network, the system adds a reverse edge to each basic relationship. The reverse edge corresponds to the original edge but in the opposite direction, thus providing a more comprehensive representation of fund flows and the reciprocity of relationships. Self-circulating edges are edges that point from a node to itself, typically used to preserve the node's own information, such as account balance and transaction frequency. This design ensures that in subsequent analysis, the node's own characteristics are not ignored due to its relationships with other nodes.

[0169] After creating reverse edges and self-looping edges, the system performs feature vectorization on these edges, representing them as vectors. These vectors are then combined with the feature vectors of the basic relations to form a more complex and precise graph structure. This step ensures that the graph structure accurately reflects the multidimensional relationships in the financial network. All generated feature vectors are stored in the initial graph structure. By integrating the basic relations, reverse edges, and self-looping edges, the system ultimately forms a complete preliminary knowledge graph. This knowledge graph contains the basic structure and key relationships of the financial network, serving as the foundation for subsequent reasoning and analysis.

[0170] In one specific implementation, when analyzing high-risk capital flows, the system expands the node representation to include risk scores and historical transaction behavior. The generation rules for reverse edges are adjusted to generate only for large-scale cross-border transactions, and their weights are increased to highlight their importance. Self-looping edges dynamically update the node state based on real-time market data, ensuring the graph reflects current market conditions. By vectorizing features using a deep learning model, the resulting preliminary knowledge graph exhibits higher timeliness and accuracy.

[0171] In another specific implementation, when processing internal enterprise transactions, self-looping edges are used to capture time-series data for each node to analyze trends in transaction behavior. The weights of reverse edges are dynamically adjusted based on historical data to capture abnormal behavior related to reverse fund flows. Feature vectorization employs a linear embedding method for rapid processing of large-scale low-dimensional data. The generated preliminary knowledge graph focuses on analyzing internal enterprise fund flows, optimizing the ability to detect transaction risks.

[0172] Through the steps described above, this embodiment enables the system to construct a preliminary knowledge graph containing rich information. The addition of reverse edges allows the network to capture bidirectional relationships, while the introduction of self-looping edges preserves the inherent information of each node. Through feature vectorization, the graph structure can more accurately represent complex financial relationships. The resulting preliminary knowledge graph provides a solid foundation for further analysis and reasoning, enhancing the comprehensiveness and accuracy of financial network analysis.

[0173] In one embodiment, in S40 above, directional features are extracted from each basic relationship, and a directional feature matrix is ​​created based on the extracted directional features, including:

[0174] S401, Extract directional features from each basic relationship, where the directional features represent the direction of fund flow within the relationship;

[0175] S402, map the extracted directional features into binary values, where 0 represents no directionality or bidirectional relationship, and 1 represents a clear directionality or unidirectional relationship;

[0176] S403 quantizes the mapped directional features, converting them into quantized values ​​that can represent the strength or importance of directionality, and creates a directional feature matrix based on the quantized directional features.

[0177] In this embodiment, the system extracts directional features from each basic relationship. These features represent the specific direction of fund flow within that relationship. For example, whether funds flow from one account to another or in both directions. This feature is crucial for accurately modeling and analyzing fund flow paths, especially when analyzing fund flow networks, as it helps identify the inflow and outflow paths of funds.

[0178] The extracted directional features are mapped to binary values, where 0 represents no directionality (such as two-way capital flow) or multi-directional relationships, and 1 represents a unidirectional relationship with a clear direction. This binary mapping simplifies the representation of directional features and facilitates subsequent processing and calculation.

[0179] The system further quantifies the mapped directional features, transforming them into numerical values ​​that represent the strength or importance of directionality. These quantified values ​​are used to construct a directional feature matrix, which can describe the directional characteristics of fund flows in financial relationships in more detail. This matrix provides foundational data for subsequent analysis, playing a particularly important role in the reasoning and detection of complex financial networks.

[0180] In one specific implementation, when analyzing high-risk cross-border transactions, the system extracts directional features by not only considering the direction of capital flow, but also combining transaction amount and time characteristics. By dynamically adjusting the mapping method, it achieves a more accurate quantification of directional intensity.

[0181] In another specific implementation, when the system uses the directional feature matrix in the enterprise's capital flow network, the influence of historical transaction data is added. The generated matrix can reflect long-term trends and provide stronger support for high-risk transactions in the analysis.

[0182] In other specific implementations, in real-time monitoring scenarios, the system dynamically updates the directional feature matrix to ensure that the analysis results reflect the latest changes in the network, helping to promptly detect potential abnormal transaction behaviors.

[0183] Through the steps described above, this embodiment enables the system to accurately extract and process directional features in financial relationships. The resulting directional feature matrix provides crucial foundational data for subsequent financial network analysis. The quantification of directional features not only simplifies data representation but also enhances the model's ability to analyze the paths and intensity of capital flows, helping to more accurately identify potential risks and abnormal behaviors in complex financial relationships.

[0184] The present invention also provides a knowledge graph construction apparatus, with reference to Figure 2 , Figure 2 This is a schematic diagram of the functional modules of a preferred embodiment of the knowledge graph construction device of the present invention. The knowledge graph construction device includes:

[0185] The entity and relationship definition module, based on the need to build a knowledge graph, identifies and determines multiple entity types participating in the activity, and defines the basic relationships between these entity types;

[0186] The initial graph structure construction module constructs the entity types and basic relationships into an initial graph structure in the form of nodes and edges, forming a preliminary knowledge graph;

[0187] The temporal feature processing module creates a temporal feature weight matrix based on the temporal features of the basic relationships, and stores the temporal feature weight matrix together with the attributes of the basic relationships in the graph structure of the preliminary knowledge graph.

[0188] The directional feature processing module extracts directional features from each basic relation, creates a directional feature matrix for each basic relation based on the extracted directional features, and stores the directional feature matrix together with the attributes of the basic relation in the graph structure of the preliminary knowledge graph.

[0189] The feature combination and optimization module combines and optimizes the multi-dimensional features of nodes and edges in the preliminary knowledge graph to generate updated representations, which are then applied to the preliminary knowledge graph to form the target knowledge graph.

[0190] The specific implementation of the knowledge graph construction device of the present invention is basically the same as the embodiments of the knowledge graph construction method described above, and will not be repeated here.

[0191] The present invention also provides a computer device, such as Figure 3As shown, the computer device may include: a processor 1001, such as a CPU; a communication bus 1002; a user interface 1003; a network interface 1004; and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen and an input unit such as a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface). The memory 1005 may be high-speed RAM or non-volatile memory, such as a disk drive. Optionally, the memory 1005 may also be a storage device independent of the aforementioned processor 1001.

[0192] Those skilled in the art will understand that Figure 3 The hardware structure of the computer device shown does not constitute a limitation on the computer device and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0193] like Figure 3 As shown, the memory 1005, which serves as a storage medium, may include an operating system, a network communication module, a user interface module, and a knowledge graph construction program. The operating system is a program that manages and controls computer equipment and software resources, supporting the operation of the network communication module, the user interface module, the knowledge graph construction program, and other programs or software. The network communication module manages and controls the network interface 1004; the user interface module manages and controls the user interface 1003.

[0194] exist Figure 3 In the computer device hardware structure shown, the network interface 1004 is mainly used to connect to the backend server and communicate data with the backend server; the user interface 1003 is mainly used to connect to the client and communicate data with the client; the processor 1001 can call the knowledge graph construction program stored in the memory 1005 and perform the same operation as the knowledge graph construction method.

[0195] The specific implementation of the computer device of the present invention is basically the same as the embodiments of the knowledge graph construction method described above, and will not be repeated here.

[0196] Furthermore, this embodiment of the invention also proposes a computer storage medium storing a knowledge graph construction program, which, when executed by a processor, implements the steps of the knowledge graph construction method described above.

[0197] The specific implementation of the computer storage medium of the present invention is basically the same as the embodiments of the knowledge graph construction method described above, and will not be repeated here.

Claims

1. A method for constructing a knowledge graph, characterized in that, Includes the following steps: Based on the need to construct a knowledge graph, multiple entity types of activities are identified and determined, and the basic relationships between these entity types are defined. The entity types and the basic relationships are constructed into an initial graph structure in the form of nodes and edges, forming a preliminary knowledge graph; Based on the temporal characteristics of the basic relations, a temporal feature weight matrix is ​​created, and the temporal feature weight matrix is ​​stored together with the attributes of the basic relations in the graph structure of the preliminary knowledge graph. The process involves extracting directional features from each basic relationship, creating a directional feature matrix based on these features, and then: extracting directional features from each basic relationship, where each feature represents the direction of capital flow within the relationship; mapping the extracted directional features to binary values, where 0 indicates no directionality or a bidirectional relationship, and 1 indicates a definite directionality or a unidirectional relationship; quantifying the mapped directional features to convert them into quantifiable values ​​that represent the strength or importance of the directionality; creating a directional feature matrix based on the quantified directional features; and storing the directional feature matrix along with the attributes of the basic relationships in the graph structure of the preliminary knowledge graph. The process involves combining and optimizing the multi-dimensional features of nodes and edges in a preliminary knowledge graph to generate updated representations, which are then applied to the preliminary knowledge graph to form a target knowledge graph. This includes: embedding multi-dimensional features into each node and edge in the preliminary knowledge graph, vectorizing the attribute features, temporal features, directional features, and relational type features of nodes and edges to form initial feature representations; inputting these initial feature representations into a CompGCN model, aggregating the multi-dimensional features of nodes and edges through multi-layer graph convolution operations to integrate the feature information of nodes, their neighboring nodes, and associated edges; applying a non-linear activation function after each layer of graph convolution operations to activate the aggregated features, improving the feature representations of nodes and edges; combining the features output from each layer after multi-layer graph convolution and non-linear activation processing to generate updated representations of nodes and edges; and reapplying the updated representations of nodes and edges to the preliminary knowledge graph, replacing the original representations, ultimately forming the target knowledge graph.

2. The knowledge graph construction method as described in claim 1, characterized in that, The entity types and the basic relationships are constructed into an initial graph structure in the form of nodes and edges, forming a preliminary knowledge graph, including: The entity types and the basic relationships are constructed into an initial graph structure in the form of nodes and edges, forming a preliminary relationship network; The preliminary relational network is expanded into a complete knowledge graph structure. The node types and edge types in the knowledge graph are determined, and feature vectors are assigned to each node and edge to form the basic representation of the knowledge graph. Create a vector representation for each basic relation, containing the type information and temporal characteristics of the basic relation; The attributes of the basic relationships are embedded into the node and edge representations of the knowledge graph, and joint learning is performed with the node embedding to form the preliminary knowledge graph.

3. The knowledge graph construction method as described in claim 2, characterized in that, The attributes of basic relationships are embedded into the node and edge representations of the knowledge graph, and joint learning is performed with the node embeddings to form the preliminary knowledge graph, including: Relation type features are extracted from each basic relation. These relation type features describe the specific properties of the basic relation and are represented as a first embedding vector. The event occurrence time feature is extracted from each basic relation. The event occurrence time feature is used to represent the time sequence or time point of the occurrence of the basic relation and is represented as a second Embedding vector. After combining the first Embedding vector with the second Embedding vector, it is input together with the Embedding vector of the corresponding node into the CompGCN model for joint learning; Through joint learning of the CompGCN model, updated feature representations of nodes and edges are generated, and these updated feature representations are integrated into the graph structure to generate the preliminary knowledge graph.

4. The knowledge graph construction method as described in claim 1, characterized in that, Multi-layer graph convolution operations are used to aggregate multi-dimensional features of nodes and edges, including: The aggregation function is determined based on the multi-dimensional features of nodes and edges in the preliminary knowledge graph; In each layer of graph convolution operation, the node's own features are combined with the features of its neighboring nodes to generate an updated node representation, and the edge's own features are combined with the features of its associated nodes to generate an updated edge representation. In multi-layer graph convolution operations, aggregation functions are applied layer by layer. Each layer performs new aggregation and updates based on the node and edge representations generated in the previous layer, gradually optimizing the node and edge representations to capture deeper feature information.

5. The knowledge graph construction method as described in claim 1, characterized in that, The entity types and the basic relationships are constructed into an initial graph structure in the form of nodes and edges, forming a preliminary knowledge graph, including: The entity types are represented as nodes, and corresponding nodes are created for each entity type in the graph structure; The basic relationships are represented as edges, and corresponding edges are created for each basic relationship in the graph structure; Add a reverse relation to each basic relation, forming a reverse edge in the graph structure that corresponds to the original relation; Add a self-circulating edge to each node in the graph structure to preserve the node's own information; The reverse edges and self-loop edges are processed into feature vectors, which are then represented as vectors and combined with the feature vectors of the basic relations to construct the initial graph structure. The feature vectors of the nodes and edges are stored in the initial graph structure. By integrating basic relations, reverse edges, and self-looping edges, a complete preliminary knowledge graph is finally formed.

6. A knowledge graph construction device, characterized in that, The knowledge graph construction device includes: The entity and relationship definition module, based on the need to build a knowledge graph, identifies and determines multiple entity types participating in the activity, and defines the basic relationships between these entity types; The initial graph structure construction module constructs the entity types and basic relationships into an initial graph structure in the form of nodes and edges, forming a preliminary knowledge graph; The temporal feature processing module creates a temporal feature weight matrix based on the temporal features of the basic relationships, and stores the temporal feature weight matrix together with the attributes of the basic relationships in the graph structure of the preliminary knowledge graph. The directional feature processing module extracts directional features from each basic relationship and creates a directional feature matrix based on these extracted features. This includes: extracting directional features from each basic relationship, where each feature represents the direction of capital flow within the relationship; mapping the extracted directional features to binary values, where 0 indicates no directionality or a bidirectional relationship, and 1 indicates a definite directionality or a unidirectional relationship; quantifying the mapped directional features to convert them into quantifiable values ​​that represent the strength or importance of the directionality; creating a directional feature matrix based on the quantified directional features; and storing the directional feature matrix along with the attributes of the basic relationships in the graph structure of the preliminary knowledge graph. The feature combination and optimization module combines and optimizes the multi-dimensional features of nodes and edges in the preliminary knowledge graph to generate updated representations, which are then applied to the preliminary knowledge graph to form the target knowledge graph. This includes: embedding multi-dimensional features into each node and edge in the preliminary knowledge graph, vectorizing the attribute features, temporal features, directional features, and relational type features of nodes and edges to form initial feature representations; inputting the initial feature representations into the CompGCN model, aggregating the multi-dimensional features of nodes and edges through multi-layer graph convolution operations to integrate the feature information of nodes, their neighboring nodes, and associated edges; applying a non-linear activation function after each layer of graph convolution operations to activate the aggregated features, improving the feature representations of nodes and edges; combining the features output from each layer after multi-layer graph convolution and non-linear activation processing to generate updated representations of nodes and edges; and reapplying the updated representations of nodes and edges to the preliminary knowledge graph, replacing the original representations, ultimately forming the target knowledge graph.

7. A computer device, characterized in that, The computer device includes a memory, a processor, and a knowledge graph construction program stored in the memory and executable on the processor. When executed by the processor, the knowledge graph construction program implements the steps of the knowledge graph construction method as described in any one of claims 1-5.

8. A computer storage medium, characterized in that, The storage medium stores a knowledge graph construction program, which, when executed by a processor, implements the steps of the knowledge graph construction method as described in any one of claims 1-5.

Images