A model splicing and fusion method based on graph database

Abstract

The present invention relates to the field of graph data processing technology, and in particular to a model splicing and fusion method based on a graph database. The method comprises the following steps: establishing a data model in the form of a graph by using the data of each model, and performing graph database construction and model graph structure retrieval processing to obtain a model graph structure dataset; performing graph shallow feature analysis on the model graph structure dataset to obtain graph node shallow feature data and graph edge shallow feature data of each model; performing feature pattern screening analysis and model splicing and reconstruction processing on the graph node shallow feature data and graph edge shallow feature data of each model to obtain a graph data splicing and reconstruction model; performing feature decomposition processing and model splicing loss analysis on each graph data splicing and reconstruction model, and performing model fusion processing to obtain a graph data fusion model. The present invention can realize the splicing and fusion of models, thereby improving splicing efficiency and accuracy.

Description

Technical Field

[0001] The present invention relates to the field of graph data processing technology, and in particular to a model splicing and fusion method based on a graph database. Background Art

[0002] In today's digital age, the explosive growth of data has posed significant challenges to information management and analysis. While traditional relational databases excel at storing structured data, they have limited processing capabilities for unstructured or semi-structured data. Graph databases, as an emerging database type, have attracted significant attention for their ability to efficiently store and query graph-structured data. They are suitable for a variety of scenarios, including social network analysis, recommender systems, and bioinformatics. Furthermore, in real-life applications, multiple models from different data sources exist, each containing distinct features and structures. Effectively combining and fusing these models into a unified model to improve the efficiency and accuracy of data processing and analysis has become a critical issue. However, traditional model concatenation methods often suffer from inaccurate model structure fusion and incomplete features. Summary of the Invention

[0003] Based on this, it is necessary for the present invention to provide a model splicing and fusion method based on a graph database to solve at least one of the above technical problems.

[0004] To achieve the above objectives, a model splicing and fusion method based on a graph database includes the following steps:

[0005] Step S1: establishing a data model by converting the data of each model into a graph to obtain graph data models of different modalities; constructing a graph database for the graph data models of different modalities to generate a model graph database;

[0006] Step S2: performing a model graph structure retrieval process on the model graph database to obtain a model graph structure dataset; performing a graph shallow feature analysis on the model graph structure dataset to obtain graph node shallow feature data and graph edge shallow feature data of each model;

[0007] Step S3: Performing feature pattern screening analysis on the shallow feature data of the graph nodes and the shallow feature data of the graph edges of each model to obtain low-frequency basic feature data and high-frequency detail feature data of each model; performing model splicing and reconstruction processing on the graph data model based on the low-frequency basic feature data and high-frequency detail feature data of each model to obtain a graph data splicing and reconstruction model;

[0008] Step S4: performing feature decomposition processing on each graph data splicing reconstruction model to obtain basic decomposition feature data and detail decomposition feature data of each reconstructed model; performing model splicing loss analysis on the basic decomposition feature data and detail decomposition feature data of each reconstructed model based on the low-frequency basic feature data and high-frequency detail feature data of each model to obtain the splicing basic feature loss and splicing detail feature loss of each reconstructed model; performing model fusion processing on the basic decomposition feature data and detail decomposition feature data of each reconstructed model based on the splicing basic feature loss and splicing detail feature loss of each reconstructed model to obtain a graph data fusion model.

[0009] The present invention first establishes a data model for the data of each model in the form of a graph, thereby constructing a corresponding graph data model for data of different modalities. This step converts the data into a graph structure, which can better capture the complex relationships and topological structures between the data, providing a more flexible and rich data representation for subsequent processing. The establishment of this graph data model enables the association and interaction between cross-modal data to be clearly expressed, providing a basis for subsequent processing. At the same time, by constructing a graph database for graph data models of different modalities, dynamic graph data can be stored and managed in an efficient and convenient manner, which can achieve effective management and query of large-scale graph data. It can also map graph data of different modalities into a unified data space, thereby achieving conversion and alignment between graph data of different modalities. Such a model graph database has good scalability and flexibility, thereby providing a strong basic support for the application and analysis of graph data. Secondly, by performing model graph structure retrieval processing on the model graph database, the model graphs stored in the database are retrieved and extracted to obtain a data set of model graph structures. The main purpose of this process is to screen out model graphs that meet specific conditions or requirements from the database to provide basic data for subsequent analysis and processing. By performing shallow feature analysis on the nodes of the model graph structure dataset, we extract features including node attributes, types, and importance. This analysis helps us understand the basic characteristics of nodes in the model graph, thus providing a foundation for subsequent model analysis and mining. Furthermore, by performing shallow feature analysis on the edges of the graph, we extract features such as edge type, weight, and direction. This analysis helps us gain a deeper understanding of the relationship characteristics between the various models in the model graph, thus providing a data foundation for subsequent model mining and graph analysis. Then, by performing feature pattern screening analysis on the shallow feature data of the graph nodes and edges of each model, we can distinguish the graph structure feature fusion data of the corresponding model into low-frequency basic feature data and high-frequency detailed feature data. The purpose of this step is to distinguish the importance and stability of the model structure features, providing guidance for subsequent processing. The graph data model is spliced ​​and reconstructed based on the low-frequency basic feature data and high-frequency detailed feature data of each model. This step, by comprehensively considering basic and detailed features, can reconstruct a more complete and accurate graph data model, which facilitates model understanding and application, thereby improving the integrity and accuracy of the feature data. Finally, the individual graph data splicing and reconstruction models are input into a preset graph data encoder for feature decomposition, aiming to decompose the complex graph data structure into basic features and detailed features. This step, through the processing of the graph data encoder, can convert the graph data into interpretable basic features and detailed features, thus providing a foundation for subsequent feature analysis and optimization.Based on the low-frequency basic feature data of each model, the basic decomposition feature data of each reconstructed model is analyzed for model splicing basic feature loss to assess the information loss of the model in low-frequency basic features. This step focuses on the similarities and differences in the overall structure of the models, providing guidance for subsequent feature optimization and ensuring the quality of basic feature reconstruction. Furthermore, based on the high-frequency detail feature data of each model, the detail decomposition feature data of each reconstructed model is analyzed for model splicing detail feature loss to assess the information loss of the model in high-frequency detail features. This step focuses on the similarities and differences in local details, helping to understand the detailed features of the model reconstruction and providing a reference for subsequent optimization. Furthermore, based on the splicing basic feature loss and splicing detail feature loss of each reconstructed model, the basic decomposition feature data and detail decomposition feature data of each reconstructed model are fused. This step fuses the optimized feature data and reassembles them into a complete graph data model. This step comprehensively considers basic and detail features, resulting in a more accurate and complete graph data model and improving the accuracy of graph data model fusion. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Other features, objects and advantages of the present invention will become more apparent upon reading the detailed description of non-limiting embodiments thereof made with reference to the following drawings:

[0011] Figure 1 Schematic diagram of the steps of the model splicing and fusion method based on the graph database of the present invention;

[0012] Figure 2 for Figure 1 Detailed step flow diagram of step S1;

[0013] Figure 3 for Figure 2 Detailed step flow chart of step S15. DETAILED DESCRIPTION

[0014] The following is a clear and complete description of the technical method of the present invention in conjunction with the accompanying drawings. Obviously, the embodiments described are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative work are within the scope of protection of the present invention.

[0015] In addition, the accompanying drawings are merely schematic illustrations of the present invention and are not necessarily drawn to scale. Identical reference numerals in the figures denote identical or similar parts, and thus repetitive descriptions thereof will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities that do not necessarily correspond to physically or logically separate entities. These functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor and / or microcontroller approaches.

[0016] It should be understood that although the terms "first," "second," and the like may be used herein to describe various elements, these elements should not be limited by these terms. These terms are used solely to distinguish one element from another. For example, a first element may be referred to as a second element, and similarly, a second element may be referred to as a first element, without departing from the scope of the exemplary embodiments. The term "and / or" as used herein includes any and all combinations of one or more of the listed associated items.

[0017] To achieve this, please refer to Figures 1 to 3 The present invention provides a model splicing and fusion method based on a graph database, the method comprising the following steps:

[0018] Step S1: establishing a data model by converting the data of each model into a graph to obtain graph data models of different modalities; constructing a graph database for the graph data models of different modalities to generate a model graph database;

[0019] Step S2: performing a model graph structure retrieval process on the model graph database to obtain a model graph structure dataset; performing a graph shallow feature analysis on the model graph structure dataset to obtain graph node shallow feature data and graph edge shallow feature data of each model;

[0020] Step S3: Performing feature pattern screening analysis on the shallow feature data of the graph nodes and the shallow feature data of the graph edges of each model to obtain low-frequency basic feature data and high-frequency detail feature data of each model; performing model splicing and reconstruction processing on the graph data model based on the low-frequency basic feature data and high-frequency detail feature data of each model to obtain a graph data splicing and reconstruction model;

[0021] Step S4: performing feature decomposition processing on each graph data splicing reconstruction model to obtain basic decomposition feature data and detail decomposition feature data of each reconstructed model; performing model splicing loss analysis on the basic decomposition feature data and detail decomposition feature data of each reconstructed model based on the low-frequency basic feature data and high-frequency detail feature data of each model to obtain the splicing basic feature loss and splicing detail feature loss of each reconstructed model; performing model fusion processing on the basic decomposition feature data and detail decomposition feature data of each reconstructed model based on the splicing basic feature loss and splicing detail feature loss of each reconstructed model to obtain a graph data fusion model.

[0022] In the embodiment of the present invention, please refer to Figure 1 FIG. 1 is a flow chart of the steps of the model splicing and fusion method based on a graph database of the present invention. In this example, the model splicing and fusion method based on a graph database includes the following steps:

[0023] Step S1: establishing a data model by converting the data of each model into a graph to obtain graph data models of different modalities; constructing a graph database for the graph data models of different modalities to generate a model graph database;

[0024] The embodiment of the present invention collects data from various models, including different types of data such as text, images, and numerical values, and models the data of each model in the form of a graph, where nodes represent data samples and edges represent associations or similarities between data, so as to convert the corresponding data into a graph structure and better capture the complex relationships and topological structures between model data, thereby obtaining graph data models of different modalities. At the same time, by using cross-modal learning technology to perform cross-modal embedding processing on graph data models of different modalities, the graph data of different modalities are mapped into a unified data space, and a corresponding graph neural network model is constructed to learn the representation of graph data in the unified graph data space model, thereby mapping the graph data into a low-dimensional continuous vector space and extracting feature vectors related to nodes and feature vectors related to edges. Then, by using a suitable graph database system (such as Neo4j, TigerGraph, etc.), the corresponding node representation feature vectors and edge representation feature vectors are constructed to construct a graph database, so as to store the feature vectors of nodes and edges of the dynamic graph in the graph database, and establish an index to support fast query and analysis operations, and finally construct a generated model graph database.

[0025] Step S2: performing a model graph structure retrieval process on the model graph database to obtain a model graph structure dataset; performing a graph shallow feature analysis on the model graph structure dataset to obtain graph node shallow feature data and graph edge shallow feature data of each model;

[0026] The embodiment of the present invention retrieves the corresponding model graph structure from the model graph database by using an appropriate query language or tool to retrieve the model graph structure containing the required information and organize it into the form of a data set, thereby obtaining a model graph structure data set. At the same time, each model graph in the model graph structure data set is processed by using a node structure analysis method to identify and extract the nodes in each model graph and determine the structural relationship between them, such as the hierarchical relationship or the connection method. The semantic relationship between each model graph node in the model graph structure data set is analyzed by using natural language processing technology to identify the semantic association between the nodes. The feature analysis is performed on each corresponding model graph node in the model graph structure data set by combining the graphic node structure of each model and the graphic node semantic relationship obtained by the analysis, so as to extract features including the attributes, type, importance, etc. of the node, and help understand the characteristics of each node. The basic characteristics of the nodes in the model graph (such as the degree distribution of the nodes, the average distance between the nodes, and other information), and then, by using graph theory and machine learning methods to identify and analyze the edge type of each model graph in the model graph structure dataset, in order to identify and analyze the graph edge connection type relationship between each model graph, and according to the graph edge type relationship between each model obtained by analysis, perform graph edge feature analysis on the corresponding model graph structure dataset, including feature extraction of edge type, weight, direction, etc., to extract the shallow features of the graph edges of each model, and help understand the features and attributes of different types of edges in the model graph, and finally obtain the shallow feature data of the graph edges of each model.

[0027] Step S3: Performing feature pattern screening analysis on the shallow feature data of the graph nodes and the shallow feature data of the graph edges of each model to obtain low-frequency basic feature data and high-frequency detail feature data of each model; performing model splicing and reconstruction processing on the graph data model based on the low-frequency basic feature data and high-frequency detail feature data of each model to obtain a graph data splicing and reconstruction model;

[0028] In an embodiment of the present invention, a graph feature fusion analysis method (such as feature concatenation, feature splicing, or a neural network model) is used to perform feature fusion on the shallow feature data of the graph nodes and the shallow feature data of the graph edges of each model obtained by analysis, so as to combine or aggregate the features of the nodes and edges to obtain the overall graph structure features. The graph structure fusion features of each model obtained after feature fusion are analyzed using a time series analysis method to identify the graph structure feature change patterns between different models, including the change patterns between node and edge features, such as changes in topological structure and evolution of attribute features. The fluctuation change rate of the corresponding graph structure features is quantitatively calculated from the features. At the same time, low-frequency basic features and high-frequency detail features are selected by discriminating based on the fluctuation frequency of the quantified feature pattern changes. Then, a preset graph data decoder is used to splice and reconstruct the corresponding graph data models by combining the low-frequency basic feature data and high-frequency detail feature data of each model obtained by analysis, so as to comprehensively consider the basic features and detail features to reconstruct a more complete and accurate graph data model, and ensure that the feature data of each model are reasonably combined to reconstruct the original graph data, ultimately obtaining a graph data splicing and reconstruction model.

[0029] Step S4: performing feature decomposition processing on each graph data splicing reconstruction model to obtain basic decomposition feature data and detail decomposition feature data of each reconstructed model; performing model splicing loss analysis on the basic decomposition feature data and detail decomposition feature data of each reconstructed model based on the low-frequency basic feature data and high-frequency detail feature data of each model to obtain the splicing basic feature loss and splicing detail feature loss of each reconstructed model; performing model fusion processing on the basic decomposition feature data and detail decomposition feature data of each reconstructed model based on the splicing basic feature loss and splicing detail feature loss of each reconstructed model to obtain a graph data fusion model.

[0030] The embodiment of the present invention inputs each graph data splicing reconstruction model into a preset graph data encoder. This encoder is a trained graph neural network model that is used to decompose and encode the input graph data model into a feature vector, and decomposes the complex graph data model structure into basic features and detail features by using the graph data encoder, wherein the basic decomposition features contain the global features of the graph data, and the detail decomposition features contain the local detail information of the graph data, thereby obtaining the basic decomposition feature data and detail decomposition feature data of each reconstruction model. At the same time, by using the low-frequency basic feature data of each model and the basic decomposition feature data of each reconstruction model, the degree of information loss of the reconstruction model on the low-frequency basic features is quantified, and by using the high-frequency detail feature data of each model and the detail decomposition feature data of each reconstruction model, the degree of information loss of the reconstruction model on the high-frequency detail features is quantified, and the difference in the detail features of the model reconstruction is evaluated, thereby obtaining the splicing basic feature loss and splicing detail feature loss of each reconstruction model. Then, by combining the analysis of the splicing basic feature loss and splicing detail feature loss of each reconstruction model, the basic decomposition feature data and detail decomposition feature data of the corresponding reconstruction models are optimized respectively to adjust the parameters of the encoder or decoder to improve the feature reconstruction quality of the model, and comprehensively consider the loss of basic features and detail features, adjust the feature representation to improve the performance of the model in basic features and details, and input the basic decomposition optimization features and detail decomposition optimization features of each reconstruction model into the preset graph data decoder for model fusion, which involves the decoder decoding and synthesizing the basic feature data and detail feature data, and recombining them into a complete graph data model, which can combine the optimized feature data of each reconstruction model to achieve the effect of image fusion, and finally obtain a graph data fusion model.

[0031] The present invention first establishes a data model for the data of each model in the form of a graph, thereby constructing a corresponding graph data model for data of different modalities. This step converts the data into a graph structure, which can better capture the complex relationships and topological structures between the data, providing a more flexible and rich data representation for subsequent processing. The establishment of this graph data model enables the association and interaction between cross-modal data to be clearly expressed, providing a basis for subsequent processing. At the same time, by constructing a graph database for graph data models of different modalities, dynamic graph data can be stored and managed in an efficient and convenient manner, which can achieve effective management and query of large-scale graph data. It can also map graph data of different modalities into a unified data space, thereby achieving conversion and alignment between graph data of different modalities. Such a model graph database has good scalability and flexibility, thereby providing a strong basic support for the application and analysis of graph data. Secondly, by performing model graph structure retrieval processing on the model graph database, the model graphs stored in the database are retrieved and extracted to obtain a data set of model graph structures. The main purpose of this process is to screen out model graphs that meet specific conditions or requirements from the database to provide basic data for subsequent analysis and processing. By performing shallow feature analysis on the nodes of the model graph structure dataset, we extract features including node attributes, types, and importance. This analysis helps us understand the basic characteristics of nodes in the model graph, thus providing a foundation for subsequent model analysis and mining. Furthermore, by performing shallow feature analysis on the edges of the graph, we extract features such as edge type, weight, and direction. This analysis helps us gain a deeper understanding of the relationship characteristics between the various models in the model graph, thus providing a data foundation for subsequent model mining and graph analysis. Then, by performing feature pattern screening analysis on the shallow feature data of the graph nodes and edges of each model, we can distinguish the graph structure feature fusion data of the corresponding model into low-frequency basic feature data and high-frequency detailed feature data. The purpose of this step is to distinguish the importance and stability of the model structure features, providing guidance for subsequent processing. The graph data model is spliced ​​and reconstructed based on the low-frequency basic feature data and high-frequency detailed feature data of each model. This step, by comprehensively considering basic and detailed features, can reconstruct a more complete and accurate graph data model, which facilitates model understanding and application, thereby improving the integrity and accuracy of the feature data. Finally, the individual graph data splicing and reconstruction models are input into a preset graph data encoder for feature decomposition, aiming to decompose the complex graph data structure into basic features and detailed features. This step, through the processing of the graph data encoder, can convert the graph data into interpretable basic features and detailed features, thus providing a foundation for subsequent feature analysis and optimization.Based on the low-frequency basic feature data of each model, the basic decomposition feature data of each reconstructed model is analyzed for model splicing basic feature loss to assess the information loss of the model in low-frequency basic features. This step focuses on the similarities and differences in the overall structure of the models, providing guidance for subsequent feature optimization and ensuring the quality of basic feature reconstruction. Furthermore, based on the high-frequency detail feature data of each model, the detail decomposition feature data of each reconstructed model is analyzed for model splicing detail feature loss to assess the information loss of the model in high-frequency detail features. This step focuses on the similarities and differences in local details, helping to understand the detailed features of the model reconstruction and providing a reference for subsequent optimization. Furthermore, based on the splicing basic feature loss and splicing detail feature loss of each reconstructed model, the basic decomposition feature data and detail decomposition feature data of each reconstructed model are fused. This step fuses the optimized feature data and reassembles them into a complete graph data model. This step comprehensively considers basic and detail features, resulting in a more accurate and complete graph data model and improving the accuracy of graph data model fusion.

[0032] Preferably, step S1 includes the following steps:

[0033] Step S11: Building a data model by converting the data of each model into a graph to obtain graph data models of different modalities;

[0034] Step S12: Perform cross-modal graph embedding processing on graph data models of different modalities to obtain a unified graph data space model;

[0035] Step S13: Using graph neural network technology to perform graph representation learning on the unified graph data space model to obtain effective representation feature vectors of dynamic graph data;

[0036] Step S14: extracting graph node and edge feature vectors from the dynamic graph data's effective representation feature vectors to obtain dynamic graph node representation feature vectors and dynamic graph edge representation feature vectors;

[0037] Step S15: constructing a graph database for the dynamic graph node representation feature vectors and the dynamic graph edge representation feature vectors based on the unified graph data space model to generate a model graph database.

[0038] As an embodiment of the present invention, refer to Figure 2 As shown, Figure 1 Detailed step flow diagram of step S1 in FIG. 1 , in this embodiment, step S1 includes the following steps:

[0039] Step S11: Building a data model by converting the data of each model into a graph to obtain graph data models of different modalities;

[0040] The embodiment of the present invention collects data from each model, including different types of data such as text, images, and numbers, and models the data of each model in the form of a graph, where nodes represent data samples and edges represent the association or similarity between data, so as to convert the corresponding data into a graph structure and better capture the complex relationships and topological structures between model data, and finally obtain graph data models of different modalities.

[0041] Step S12: Perform cross-modal graph embedding processing on graph data models of different modalities to obtain a unified graph data space model;

[0042] The embodiment of the present invention uses cross-modal learning technology to perform cross-modal embedding processing on graph data models of different modalities to map graph data of different modalities into a unified data space, and realizes the conversion and alignment between graph data of different modalities, and finally obtains a unified graph data space model.

[0043] Step S13: Using graph neural network technology to perform graph representation learning on the unified graph data space model to obtain effective representation feature vectors of dynamic graph data;

[0044] The embodiments of the present invention use graph neural network technology (such as graph convolutional network (GCN), graph attention network (GAT) and other methods) to build a corresponding graph neural network model to learn the representation of graph data in a unified graph data space model, so as to map the graph data into a low-dimensional continuous vector space, and extract feature vectors with rich semantic information from the unified graph data space, and finally obtain feature vectors that effectively represent dynamic graph data.

[0045] Step S14: extracting graph node and edge feature vectors from the dynamic graph data's effective representation feature vectors to obtain dynamic graph node representation feature vectors and dynamic graph edge representation feature vectors;

[0046] In this embodiment, pooling operations or node embedding techniques in graph neural networks are used to extract node-related feature vectors from the dynamic graph data's effective representation feature vectors, thereby obtaining a dynamic graph node representation feature vector. Subsequently, convolutional neural networks or attention mechanisms are used to extract edge-related feature vectors from the dynamic graph data's effective representation feature vectors, ultimately obtaining a dynamic graph edge representation feature vector.

[0047] Step S15: constructing a graph database for the dynamic graph node representation feature vectors and the dynamic graph edge representation feature vectors based on the unified graph data space model to generate a model graph database.

[0048] The embodiment of the present invention uses a suitable graph database system (such as Neo4j, TigerGraph, etc.) in combination with a unified graph data space model to construct a graph database for the corresponding dynamic graph node representation feature vectors and dynamic graph edge representation feature vectors, so as to store the feature vectors of the nodes and edges of the dynamic graph in the graph database, and establish an index to support fast query and analysis operations, and finally construct a generation model graph database.

[0049] The present invention first establishes a data model in the form of a graph for the data of each model, and can construct a corresponding graph data model for data of different modalities. This step can better capture the complex relationship and topological structure between the data by converting the data into a graph structure, and provide a more flexible and rich data representation for subsequent processing. The establishment of this graph data model enables the association and interaction between cross-modal data to be clearly expressed, providing a basis for subsequent processing. Secondly, by performing cross-modal graph embedding processing on graph data models of different modalities, graph data of different modalities can be mapped to a unified data space, thereby realizing conversion and alignment between modalities. Such a unified graph data space model has better consistency and comparability, and provides more convenient conditions for subsequent data processing and analysis. Then, by using graph neural network technology to perform graph representation learning on the unified graph data space model, feature vectors with rich semantic information can be extracted from the unified graph data space, realizing effective representation of dynamic graph data. These feature vectors can better capture the intrinsic characteristics and structural information of the data, thereby providing strong support for subsequent data analysis and application. Next, by extracting graph node and edge feature vectors from the effective representation feature vectors of dynamic graph data, the node and edge information in the graph data can be separated and represented, allowing each node and edge to be effectively described and understood. This feature vector extraction process provides more accurate and effective input for subsequent data processing and analysis. Finally, by constructing a graph database based on the dynamic graph node representation feature vectors and the dynamic graph edge representation feature vectors based on a unified graph data space model, dynamic graph data can be stored and managed in an efficient and convenient manner, achieving effective management and query of large-scale graph data. This model graph database has good scalability and flexibility, providing a strong foundation for the application and analysis of graph data.

[0050] Preferably, step S15 includes the following steps:

[0051] Step S151: Analyze the graph data characteristics of the unified graph data space model to obtain graph data model characteristics; analyze the real-time query requirements of the unified graph data space model to obtain graph data real-time query requirement data;

[0052] Step S152: performing query access pattern analysis on the unified graph data space model based on the graph data real-time query demand data to obtain graph data query access pattern data;

[0053] Step S153: performing graph index identification and analysis on the unified graph data space model according to the graph data model characteristics and the graph data query access pattern data to obtain the initial index structure of the graph database;

[0054] Step S154: performing graph data temporal change analysis on the dynamic graph node representation feature vector and the dynamic graph edge representation feature vector to obtain graph data temporal dynamic change data; performing adaptive index adjustment on the initial index structure of the graph database based on the graph data temporal dynamic change data to obtain a graph database adaptive index optimization structure;

[0055] Step S155: construct a graph database based on the dynamic graph node representation feature vector and the dynamic graph edge representation feature vector according to the graph database adaptive index optimization structure to generate a model graph database.

[0056] As an embodiment of the present invention, refer to Figure 3 As shown, Figure 2 Detailed step flow diagram of step S15 in the embodiment, step S15 includes the following steps:

[0057] Step S151: Analyze the graph data characteristics of the unified graph data space model to obtain graph data model characteristics; analyze the real-time query requirements of the unified graph data space model to obtain graph data real-time query requirement data;

[0058] The present invention collects various graph data samples from a unified graph data space model and analyzes the graph data's structure, node attributes, edge attributes, and topological structure, thereby obtaining graph data model characteristics. Then, using a query demand analysis method, real-time query demand analysis is performed on the corresponding graph data within the unified graph data space model to obtain specific requirements and conditions for real-time graph data queries, such as the frequency of read operations and the breadth of the query scope. Ultimately, the real-time graph data query demand data is obtained.

[0059] Step S152: performing query access pattern analysis on the unified graph data space model based on the graph data real-time query demand data to obtain graph data query access pattern data;

[0060] The embodiment of the present invention uses an access pattern recognition method to identify and analyze the corresponding unified graph data space model by combining the real-time query demand data of the graph data obtained through analysis, so as to identify the corresponding query patterns therefrom, including node association queries, path queries, graph traversal and other patterns, and finally obtain graph data query access pattern data.

[0061] Step S153: performing graph index identification and analysis on the unified graph data space model according to the graph data model characteristics and the graph data query access pattern data to obtain the initial index structure of the graph database;

[0062] The embodiment of the present invention identifies and analyzes the graph index structure of the corresponding graph data in the unified graph data space model by combining the analyzed graph data model characteristics and graph data query access pattern data, so as to identify the index type and structure suitable for the graph data characteristics and query requirements, such as adjacency list, adjacency matrix, index list, etc., and finally obtain the initial index structure of the graph database.

[0063] Step S154: performing graph data temporal change analysis on the dynamic graph node representation feature vector and the dynamic graph edge representation feature vector to obtain graph data temporal dynamic change data; performing adaptive index adjustment on the initial index structure of the graph database based on the graph data temporal dynamic change data to obtain a graph database adaptive index optimization structure;

[0064] The embodiments of the present invention use a temporal change analysis method to analyze the changes in the feature vectors representing dynamic graph nodes and edges, thereby observing the evolution of graph data over time and identifying the dynamic change patterns of nodes and edges, such as adding or deleting nodes, updating edges, etc., thereby obtaining temporal dynamic change data of the graph data. Then, by combining the temporal dynamic change data of the graph data obtained through analysis, an adaptive adjustment algorithm is used to adjust the corresponding initial index structure of the graph database. This dynamically optimizes the index storage and query strategies based on the trend and frequency of data changes, adapting them to the dynamic changes of the graph data, and ultimately obtaining an adaptive index optimization structure for the graph database.

[0065] Step S155: construct a graph database based on the dynamic graph node representation feature vector and the dynamic graph edge representation feature vector according to the graph database adaptive index optimization structure to generate a model graph database.

[0066] The embodiment of the present invention uses a suitable graph database system (such as Neo4j, TigerGraph, etc.) to construct a graph database by combining the graph database adaptive index optimization structure obtained through analysis to represent the corresponding dynamic graph node feature vectors and dynamic graph edge feature vectors, so as to establish corresponding index and storage structures to ensure that they can adapt to the temporal changes of graph data and changes in query requirements, and finally construct a generation model graph database.

[0067] The present invention first analyzes the characteristics of the graph data on the unified graph data space model to reveal its inherent structure and properties. This step, by analyzing the characteristics of the graph data model, can provide a deep understanding of the organization of the graph data, the relationship between nodes and edges, and the topological structure of the graph data. At the same time, by performing a real-time query demand analysis on the unified graph data space model, the specific requirements and conditions for real-time queries on the graph data are obtained. Such analysis helps to provide a basis for subsequent query optimization and index design. Secondly, based on the real-time query demand data of the graph data, the query access pattern analysis of the unified graph data space model is performed. This step aims to understand the query pattern of the graph data, including the characteristics of the query frequency, type and complexity. By analyzing the query access pattern data, it can provide guidance and basis for designing efficient query processing and index structures, thereby improving the query performance and efficiency of the graph database. Then, based on the characteristics of the graph data model and the graph data query access pattern data, the unified graph data space model is subjected to graph index identification analysis. The goal of this step is to identify the index type and structure that is suitable for the characteristics of the graph data and the query demand, so as to support efficient graph data query and retrieval operations. A properly designed initial index structure can improve the query efficiency and response speed of a graph database, thereby enhancing system performance. Next, a temporal change analysis is performed on the representation feature vectors of dynamic graph nodes and edges to obtain temporal dynamic change data of the graph data. This temporal change analysis can reveal the patterns and trends of graph data evolution over time, providing an important basis for subsequent index optimization and data management. Furthermore, adaptive index adjustments are performed on the initial index structure of the graph database based on the temporal dynamic change data of the graph data, effectively responding to graph data changes and dynamically adjusting the index structure, thus providing basic data support for subsequent processing. Finally, the graph database is constructed by applying the adaptive index optimization structure to the dynamic graph node representation feature vectors and dynamic graph edge representation feature vectors. This step applies the adaptive index optimization structure to the graph database construction process, ensuring that the database index structure can adapt to the temporal changes of the graph data and changes in query requirements. The resulting model graph database has higher query efficiency and data management capabilities, providing stronger support for graph data analysis and applications.

[0068] Preferably, step S2 includes the following steps:

[0069] Step S21: performing a model graph structure retrieval process on the model graph database to obtain a model graph structure dataset;

[0070] The embodiment of the present invention retrieves the corresponding model graph structure from the model graph database by using an appropriate query language or tool to retrieve the model graph structure containing the required information, and organizes it into a data set, ultimately obtaining a model graph structure data set.

[0071] Step S22: extracting the node structure and node semantic relationship of the model graph structure dataset to obtain the graph node structure and the contextual semantic relationship of the graph nodes of each model;

[0072] The present invention uses a node structure analysis method to process each model graph within a model graph structure dataset to identify and extract the nodes within each model graph and determine the structural relationships between them, such as hierarchical relationships or connection methods, thereby obtaining the graph node structure of each model. Natural language processing techniques are then used to analyze the semantic relationships between the nodes in each model graph within the model graph structure dataset to identify semantic associations between the nodes and determine the contextual semantic relationships between the nodes, ultimately obtaining the contextual semantic relationships between the graph nodes of each model.

[0073] Step S23: performing node shallow feature analysis on the model graph structure dataset based on the graph node structure of each model and the contextual semantic relationship of the graph nodes to obtain shallow feature data of the graph nodes of each model;

[0074] The embodiment of the present invention performs feature analysis on each model graph node corresponding to the model graph structure data set by combining the graphic node structure of each model obtained by analysis and the contextual semantic relationship of the graphic nodes, so as to extract features including the attributes, types, importance, etc. of the nodes, and help understand the basic features of the nodes in each model graph (such as the degree distribution of the nodes, the average distance between nodes, and other information), and finally obtain the shallow feature data of the graphic nodes of each model.

[0075] Step S24: performing graph edge type identification analysis on the model graph structure dataset to obtain graph edge type relationships between various models;

[0076] The embodiment of the present invention uses graph theory and machine learning methods to identify and analyze the edge type of each model graph in the model graph structure dataset, so as to identify and analyze the graph edge connection type relationship between each model graph, and reveal the association relationship and connection relationship between different models in the model graph, and finally obtain the graph edge type relationship between each model.

[0077] Step S25: performing graph edge shallow feature analysis on the model graph structure dataset based on the graph edge type relationship between the various models to obtain graph edge shallow feature data of the various models.

[0078] The embodiment of the present invention performs feature analysis of the graph edges of the corresponding model graph structure data set by combining the graph edge type relationships between the various models obtained through analysis, including feature extraction of edge type, weight, direction, etc., to extract the shallow features of the graph edges of each model, and help understand the features and attributes of different types of edges in the model graph, and finally obtain the shallow feature data of the graph edges of each model.

[0079] The present invention first performs a model graph structure retrieval process on the model graph database, and retrieves and extracts the model graphs stored in the database to obtain a data set of model graph structures. The main purpose of this process is to screen out model graphs that meet specific conditions or requirements from the database, and provide basic data for subsequent analysis and processing. Secondly, by performing a node structure and node semantic relationship extraction process on the model graph structure data set, the node structure of each model graph will be extracted, and the semantic relationship between the nodes will be analyzed. The purpose of this step is to gain an in-depth understanding of the organization of the nodes in the model graph and the semantic association between the nodes, thereby providing data support for subsequent feature analysis and association mining. Then, the model graph structure data set is subjected to a shallow feature analysis of the nodes based on the graphic node structure of each model and the upper and lower semantic relationships of the graphic nodes, so as to realize feature extraction of aspects including the attributes, types, importance, etc. of the nodes. Such analysis helps to understand the basic characteristics of the nodes in the model graph, thereby providing a basis for subsequent model analysis and mining. Next, by performing graph edge type recognition analysis on the model graph structure dataset, the graph edge type relationships between the various models in the model graph are identified and analyzed. The purpose of this step is to reveal the associations and connection methods between different models in the model graph, thereby providing data support for subsequent edge feature analysis and graph relationship mining. Finally, by performing shallow graph edge feature analysis on the model graph structure dataset based on the graph edge type relationships between the various models, shallow feature analysis of the edges in the model graph is performed, including feature extraction of edge type, weight, direction, and other aspects. This analysis helps to deeply understand the relationship characteristics between the various models in the model graph, thereby providing a data foundation for subsequent model mining and graph analysis.

[0080] Preferably, step S25 includes the following steps:

[0081] Step S251: performing vector space embedding processing on the graph edge type relationships between the various models to obtain edge type relationship embedding vectors of the various models;

[0082] The embodiments of the present invention use technologies such as Word2Vec and GloVe to represent the edge type relationship of each model's graph as an embedding in a vector space, so as to convert the edge type relationship in the model graph into a high-dimensional vector representation, and ensure that the edge type relationship between different models can be effectively mapped into a unified vector space, and finally obtain the edge type relationship embedding vector of each model.

[0083] Step S252: performing edge type structure analysis on the edge type relationship embedding vectors of each model according to the graph edge type relationships between the models to obtain different types of relationship edge structures of each model;

[0084] The embodiment of the present invention performs edge type structure identification analysis on the corresponding edge type relationship embedding vectors in each model by combining the graphical edge type relationships between the various models obtained through analysis, so as to extract the structural information of the edge type, such as the weight, direction, connectivity and other information of the edge, from the embedding vector, and reveal the structure of different types of relationship edges in the model graph, and finally obtain different types of relationship edge structures of each model.

[0085] Step S253: performing semantic relationship recognition analysis on the edge type relationship embedding vectors of each model to obtain different types of edge semantic relationships of each model;

[0086] The embodiment of the present invention uses a semantic relationship recognition method to identify and analyze the semantic relationship between the edge type relationship embedding vectors of each model to identify and determine the semantic relationship represented by the different types of edges of each model, thereby better understanding the meaning of the model graph and ultimately obtaining the semantic relationship of the different types of edges of each model.

[0087] Step S254: performing graph edge shallow feature analysis on the model graph structure dataset according to different types of relationship edge structures of each model and different types of edge semantic relationships of each model to obtain graph edge shallow feature data of each model.

[0088] The embodiment of the present invention uses corresponding feature analysis methods (including importance, impact assessment, association mining and other methods) to perform feature analysis on the corresponding model graph structure data set by combining the different types of relationship edge structures of each model and the different types of edge semantic relationships of each model to comprehensively analyze the characteristics of the edges in the model graph, such as the importance of the edges, the scope of influence, the degree of correlation, etc., which helps to understand the characteristics and attributes of the model graph structure, and finally obtain the shallow feature data of the graph edges of each model.

[0089] The present invention first performs vector space embedding processing on the graphical edge type relationships between each model, which can convert the edge type relationships in the model graph into high-dimensional vector representations, so that the edge type relationships of each model can be represented as a vector. Through vector space embedding, the edge type relationships between different models can be effectively mapped into a unified vector space, thereby providing a comparable basis for subsequent analysis and processing. Secondly, the edge type relationship embedding vectors of each model are subjected to edge type structure analysis based on the graphical edge type relationships between each model. The purpose of this step is to extract the structural information of the edge type, such as the weight, direction, connectivity, etc. of the edge, from the embedding vector, thereby revealing the structure of the edges of different types of relationships in the model graph. Then, semantic relationship recognition analysis is performed on the edge type relationship embedding vectors of each model to analyze features such as similarity and distance in the vector space. In this way, the semantic relationships between different types of edges, such as similarity, dependency, etc., can be inferred, thereby deeply understanding the association between each model in the model graph. Finally, we conduct a shallow feature analysis of the graph edges of the model graph structure dataset based on the different types of relational edge structures and the different types of edge semantic relationships of each model. This step combines the structural characteristics and semantic relationships of the edges to comprehensively analyze the characteristics of the edges in the model graph, such as the importance, scope of influence, and correlation of the edges, thereby providing deeper insights and data support for the understanding and application of the model graph.

[0090] Preferably, step S3 includes the following steps:

[0091] Step S31: performing graphic feature fusion analysis on the shallow feature data of the graphic nodes and the shallow feature data of the graphic edges of each model to obtain the graphic structure feature fusion data of each model;

[0092] The embodiment of the present invention uses a graph feature fusion analysis method (such as feature concatenation, feature splicing, or the use of a neural network model) to perform feature fusion on the shallow feature data of the graph nodes and the shallow feature data of the graph edges of each model obtained by analysis, so as to combine or aggregate the features of the nodes and edges to obtain the overall graph structure features, and finally obtain the graph structure feature fusion data of each model.

[0093] Step S32: performing feature pattern change analysis on the graphic structure feature fusion data of each model to obtain graphic structure feature change pattern data of each model;

[0094] The embodiment of the present invention uses a time series analysis method to analyze the graph structure feature fusion data of each model obtained after feature fusion to identify the graph structure feature change patterns between different models, which includes the change patterns between node and edge features, such as changes in topological structure, evolution of attribute features, etc., so as to further analyze the change laws of graph structure features and finally obtain the graph structure feature change pattern data of each model.

[0095] Step S33: performing fluctuation frequency calculation on the graphic structure characteristic change pattern data of each model using a characteristic pattern change fluctuation frequency calculation formula to obtain the graphic structure characteristic pattern change fluctuation frequency of each model;

[0096] The embodiment of the present invention combines the model quantity parameter, the graphic structure characteristic mode amplitude parameter, the graphic structure characteristic mode attenuation coefficient, the graphic structure characteristic mode change influencing factor, the graphic structure characteristic mode bandwidth parameter, the exponential function, the characteristic mode fluctuation phase angle, the characteristic mode fluctuation angle distribution mean, the characteristic mode fluctuation angle distribution standard deviation and related parameters to form a suitable characteristic mode change fluctuation frequency calculation formula to perform fluctuation frequency calculation on the graphic structure characteristic change pattern data of each model to quantify the fluctuation change rate of the corresponding graphic structure feature, and finally obtain the graphic structure characteristic mode change fluctuation frequency of each model.

[0097] The calculation formula for the characteristic mode change fluctuation frequency is as follows:

[0098]

[0099] Where n is the total number of models, i is the item index parameter of the model, f is the frequency of fluctuation of the graphical structure characteristic mode of the i-th model, α i is the amplitude parameter of the graphical structural characteristic mode of the i-th model, β i is the graph structure characteristic mode attenuation coefficient of the i-th model, γ i is the influence factor of the graph structure characteristic pattern change of the i-th model, δ i is the bandwidth parameter of the graphical structure characteristic mode of the i-th model, exp is the exponential function, ω i is the characteristic mode fluctuation phase angle of the i-th model, μ i is the mean value of the characteristic mode fluctuation angle distribution of the i-th model, σ i is the standard deviation of the characteristic mode fluctuation angle distribution of the i-th model, η i is the fluctuation frequency correction coefficient of the i-th model;

[0100] The present invention obtains a characteristic pattern change fluctuation frequency calculation formula by using a specific mathematical model and verifying it, which is used to calculate the fluctuation frequency of the graphic structure feature change pattern data of each model. The goal of the characteristic pattern change fluctuation frequency calculation formula is to describe the change of the graphic structure characteristics of each model in the frequency domain. By calculating the fluctuation frequency, the frequency distribution of the graphic structure feature changes in the model can be understood, thereby revealing the characteristic pattern change characteristics of the model. The integral part in the formula integrates the frequency change of the characteristic pattern to calculate the fluctuation frequency. This part can obtain the frequency domain characteristics of the pattern by comprehensively analyzing factors such as the amplitude, attenuation, and influencing factors of the pattern. In addition, the fluctuation frequency calculated by this formula describes the fluctuation frequency of the characteristic pattern change in the model. This fluctuation frequency can be used to measure the rate of change of the graphic structure characteristics in the model, that is, at what frequency the graphic structure characteristics change, thereby further understanding and processing the change pattern of the graphic structure characteristics. Therefore, the formula can fully consider the total number of models n, the item index parameter i of the model, and the fluctuation frequency f of the graphic structure feature pattern change of the i-th model. i , the amplitude parameter α of the graphical structural characteristic mode of the i-th model i , the graph structure characteristic mode attenuation coefficient β of the i-th model i , the influence factor of the change of the graphical structure characteristic pattern of the i-th model γ i , the bandwidth parameter δ of the graphical structural characteristic mode of the i-th model i , exponential function exp, characteristic mode fluctuation phase angle ω of the i-th model i , the mean value of the characteristic mode fluctuation angle distribution μ of the i-th model i , the standard deviation of the characteristic mode fluctuation angle distribution of the i-th model σ i , the fluctuation frequency correction coefficient η of the i-th model i , according to the graphical structural characteristic pattern of the i-th model, the fluctuation frequency f i The mutual correlation between the above parameters constitutes a functional relationship:

[0101]

[0102] This formula can realize the calculation process of the fluctuation frequency of the graphical structure feature change pattern data of each model, and at the same time, the fluctuation frequency correction coefficient η of the i-th model is used. i The introduction of can be adjusted according to the error conditions that occur during the calculation process, thereby improving the accuracy and applicability of the calculation formula for the characteristic mode change fluctuation frequency.

[0103] Step S34: comparing and judging the graphic structure feature pattern change fluctuation frequency of each model according to a preset feature pattern change fluctuation threshold; if the graphic structure feature pattern change fluctuation frequency of each model is less than the preset feature pattern change fluctuation threshold, the graphic structure feature fusion data of the corresponding model is judged as low-frequency basic feature data to obtain the low-frequency basic feature data of each model; if the graphic structure feature pattern change fluctuation frequency of each model is greater than or equal to the preset feature pattern change fluctuation threshold, the graphic structure feature fusion data of the corresponding model is judged as high-frequency detail feature data to obtain the high-frequency detail feature data of each model;

[0104] The embodiment of the present invention compares and judges the graphic structure feature pattern change fluctuation frequency of each model obtained by quantitative calculation by using a preset feature pattern change fluctuation threshold. If there is a feature pattern change fluctuation frequency of each model that is less than the preset feature pattern change fluctuation threshold, it means that the feature pattern fluctuation of the graphic structure feature fusion data of its corresponding model is small, and the graphic structure feature fusion data of its corresponding model is judged as a low-frequency basic feature, thereby obtaining the low-frequency basic feature data of each model; if there is a feature pattern change fluctuation frequency of each model that is greater than or equal to the preset feature pattern change fluctuation threshold, it means that the feature pattern fluctuation of the graphic structure feature fusion data of its corresponding model is large and more detailed, and the graphic structure feature fusion data of its corresponding model is judged as a high-frequency detail feature, and finally the high-frequency detail feature data of each model is obtained.

[0105] Step S35: performing model splicing and reconstruction processing on the graph data model based on the low-frequency basic feature data and high-frequency detail feature data of each model to obtain a graph data splicing and reconstruction model.

[0106] The embodiment of the present invention uses a preset graph data decoder to perform model splicing and reconstruction on the corresponding graph data model by combining the low-frequency basic feature data and high-frequency detail feature data of each model obtained through analysis, so as to comprehensively consider the basic features and detail features to reconstruct a more complete and accurate graph data model, and ensure that the feature data of each model are reasonably combined to reconstruct the original graph data, and finally obtain a graph data splicing and reconstruction model.

[0107] The present invention first performs a graph feature fusion analysis on the graph node shallow feature data and graph edge shallow feature data of each model, and can combine the graph node shallow feature data and graph edge shallow feature data of each model to obtain more comprehensive and integrated graph structure feature fusion data. The key to this step is to merge the features of the nodes and edges and consider the association between them, so as to capture the overall characteristics of the model graph. Secondly, by performing a feature pattern change analysis on the graph structure feature fusion data of each model, the graph structure feature change pattern between different models is identified, which includes the change pattern between node and edge features, such as the change of topological structure, the evolution of attribute features, etc., which helps to understand the evolution process and structural changes of the model. Then, by using the feature pattern change fluctuation frequency calculation formula, the graph structure feature change pattern data of each model is calculated for fluctuation frequency to quantify the rate of change of the graph structure features, which helps to identify the frequency and regularity of model structure changes and provide data support for subsequent pattern discrimination. Next, the frequency of fluctuations in the graph structure feature patterns of each model is compared and determined based on a preset threshold for feature pattern fluctuations. If the fluctuation frequency is below the threshold, the graph structure feature fusion data of the corresponding model is identified as low-frequency basic feature data; if it is above or equal to the threshold, it is identified as high-frequency detail feature data. This step aims to distinguish the importance and stability of the model's structural features and provide guidance for subsequent processing. Finally, the graph data model is spliced ​​and reconstructed based on the low-frequency basic feature data and high-frequency detail feature data of each model. This step, by comprehensively considering both basic and detail features, can reconstruct a more complete and accurate graph data model, which facilitates model understanding and application.

[0108] Preferably, step S35 includes the following steps:

[0109] Step S351: performing low-frequency feature global sampling processing on the low-frequency basic feature data of each model to obtain low-frequency feature global information data of each model;

[0110] The embodiment of the present invention performs feature sampling processing on the low-frequency basic feature data of each model by using a global sampling method (for example, low-pass filtering in the frequency domain or mean sampling in the spatial domain) to extract global information from the low-frequency basic feature data of each model, and capture the main features and trends of the model graph in the low-frequency domain, and finally obtain the low-frequency feature global information data of each model.

[0111] Step S352: performing high-frequency feature local sampling processing on the high-frequency detail feature data of each model to obtain high-frequency feature local information data of each model;

[0112] The embodiments of the present invention perform local sampling of high-frequency features on the high-frequency detail feature data of each model by using an edge detection algorithm or a texture detection method, so as to extract local information from the high-frequency detail feature data of each model, including information such as edges or textures, and capture subtle changes and special structures of the model graph in the high-frequency domain, and finally obtain the high-frequency feature local information data of each model.

[0113] Step S353: performing feature fusion analysis on the low-frequency feature global information data and the high-frequency feature local information data of each model to obtain effective fusion feature data of the graph data of each model;

[0114] The embodiment of the present invention uses corresponding feature fusion analysis technologies (such as weighted summation, feature cascade or neural network fusion model, etc.) to fuse and analyze the low-frequency feature global information data and high-frequency feature local information data of each model obtained by sampling, so as to combine the global information of low-frequency features and the local information of high-frequency detail features, and obtain a more complete image feature representation, thereby taking into account the overall features and local details of the model graph, and finally obtaining the effective fusion feature data of the graph data of each model.

[0115] Step S354: effectively fuse the feature data of the graph data of each model and input it into a preset graph data decoder to perform model splicing and reconstruction processing on the corresponding graph data model to obtain a graph data splicing and reconstruction model.

[0116] The embodiment of the present invention effectively fuses the feature data of the graph data of each model obtained by fusion and inputs it into a preset graph data decoder to decode and splice the corresponding graph data models, so as to convert the fused feature data into a visual graphic structure and realize the splicing reconstruction of the original graph data model, thereby reconstructing the structure and properties of the original model graph, and finally obtaining a graph data splicing reconstruction model.

[0117] The present invention first performs global sampling of low-frequency features on the low-frequency basic feature data of each model, aiming to extract global information from the low-frequency basic feature data of each model, and captures the main features and trends of the model graph in the low-frequency domain by sampling the overall features. The key to this step is to grasp the low-frequency features globally to ensure that the extracted features can represent the overall structure and properties of the model graph. Through global sampling, it is possible to grasp the overall features of the model, rather than just focusing on local details, so as to better understand the overall morphology and change trends of the model. Secondly, by performing local sampling of high-frequency features on the high-frequency detail feature data of each model, it is aimed to extract local information from the high-frequency detail feature data of each model, and capture the subtle changes and special structures of the model graph in the high-frequency domain by sampling the local features. The key to this step is to pay local attention to the high-frequency features to ensure that the extracted features can represent the local details and special features of the model graph. Through local sampling, some important but limited features in the model graph can be captured, so as to better understand the local properties and subtle structures of the model. Next, feature fusion analysis is performed on the low-frequency global information data and high-frequency local information data of each model. This aims to comprehensively consider the characteristics of the model graph in different frequency domains to obtain more comprehensive and accurate graph data effective fusion feature data. The key to this step is to rationally fuse global and local information, thereby taking into account both the overall characteristics and local details of the model graph, providing a more complete and accurate feature representation for subsequent model splicing and reconstruction. It can also fully utilize the advantages of low-frequency and high-frequency features to better understand and characterize the structure and properties of the model. Finally, the graph data effective fusion feature data of each model is input into a preset graph data decoder to perform model splicing and reconstruction on the corresponding graph data model, aiming to generate a complete and accurate graph data splicing and reconstruction model. The key to this step is to use the fused feature data to reconstruct the model to restore the structure and properties of the original model graph. Through the decoder processing, the fused feature data can be converted into a visual graphical structure, thereby achieving model splicing and reconstruction.

[0118] Preferably, step S4 includes the following steps:

[0119] Step S41: inputting the reconstructed model of each graph data splicing into a preset graph data encoder for feature decomposition processing to obtain basic decomposition feature data and detail decomposition feature data of each reconstructed model;

[0120] The embodiment of the present invention inputs each graph data splicing reconstruction model into a preset graph data encoder. This encoder is a trained graph neural network model, which is used to decompose and encode the input graph data model into a feature vector, and decompose the complex graph data model structure into basic features and detail features by using the graph data encoder, wherein the basic decomposition features contain the global features of the graph data, and the detail decomposition features contain the local detail information of the graph data, and finally obtain the basic decomposition feature data and detail decomposition feature data of each reconstruction model.

[0121] Step S42: performing model splicing basic feature loss analysis on the basic decomposition feature data of each reconstructed model based on the low-frequency basic feature data of each model to obtain the splicing basic feature loss of each reconstructed model;

[0122] The embodiment of the present invention uses a feature vector mapping method to map and transform the low-frequency basic feature data of each model and the basic decomposition feature data of each reconstructed model, so as to convert the original feature representation into a more representative and comparable vector space representation, and aligns the feature space by using a corresponding feature alignment algorithm (such as the least squares method or the iterative nearest point algorithm, etc.) to adjust the feature space of different models to a unified coordinate system so that they have consistent direction and scale, and quantifies the degree of information loss of the reconstructed model in the low-frequency basic features, thereby evaluating the differences in the basic features of the model reconstruction, and finally obtaining the splicing basic feature loss of each reconstructed model.

[0123] Step S43: performing model splicing detail feature loss analysis on the detail decomposition feature data of each reconstructed model based on the high-frequency detail feature data of each model to obtain the splicing detail feature loss of each reconstructed model;

[0124] The present invention similarly performs quantitative calculation of model reconstruction loss by utilizing the high-frequency detail feature data of each model and the detail decomposition feature data of each reconstructed model, so as to quantitatively calculate the degree of information loss of the reconstructed model in the high-frequency detail features, and evaluate the differences in the detail features of the model reconstruction, and finally obtain the splicing detail feature loss of each reconstructed model.

[0125] Step S44: performing feature optimization processing on the basic decomposition feature data and the detail decomposition feature data of each reconstructed model based on the splicing basic feature loss and the splicing detail feature loss of each reconstructed model to obtain the basic decomposition optimized feature data and the detail decomposition optimized feature data of each reconstructed model;

[0126] The embodiment of the present invention optimizes the basic decomposition feature data and the detail decomposition feature data of each corresponding reconstruction model by combining the splicing basic feature loss and the splicing detail feature loss of each reconstruction model obtained by analysis, so as to adjust the parameters of the encoder or decoder to improve the feature reconstruction quality of the model, and comprehensively considers the loss of basic features and detail features, adjusts the feature representation to improve the performance of the model in basic features and details, and finally obtains the basic decomposition optimized feature data and the detail decomposition optimized feature data of each reconstruction model.

[0127] Step S45: Inputting the basic decomposition optimization feature data and the detail decomposition optimization feature data of each reconstructed model into a preset graph data decoder for model fusion processing to obtain a graph data fusion model.

[0128] The embodiment of the present invention performs model fusion by inputting the basic decomposition optimization feature data and the detail decomposition optimization feature data of each reconstruction model into a preset graph data decoder. This involves the decoder decoding and synthesizing the basic feature data and the detail feature data, and recombining them into a complete graph data model. The model can combine the optimized feature data of each reconstruction model to achieve the effect of image fusion, and finally obtain a graph data fusion model.

[0129] The present invention first inputs each graph data splicing reconstruction model into a preset graph data encoder for feature decomposition processing, aiming to decompose the complex graph data structure into basic features and detail features. This step, through the processing of the graph data encoder, can convert the graph data into interpretable basic features and detail features, thereby providing a basis for subsequent feature analysis and optimization. Secondly, based on the low-frequency basic feature data of each model, the basic decomposition feature data of each reconstructed model is subjected to model splicing basic feature loss analysis to evaluate the information loss of the model on the low-frequency basic features. This step focuses on the similarity and difference of the model in the overall structure, thereby providing guidance for subsequent feature optimization and ensuring the reconstruction quality of the basic features. Then, based on the high-frequency detail feature data of each model, the detail decomposition feature data of each reconstructed model is subjected to model splicing detail feature loss analysis to evaluate the information loss of the model on the high-frequency detail features. This step focuses on the similarity and difference of the model in local details, helps to understand the detail features of the model reconstruction, thereby providing a reference for subsequent optimization. Next, feature optimization processing is performed on the basic decomposition feature data and detailed decomposition feature data of each reconstructed model based on the splicing basic feature loss and splicing detailed feature loss of each reconstructed model. This aims to optimize the basic and detailed features of the model and improve the quality and accuracy of model reconstruction. This step comprehensively considers the loss of basic and detailed features and adjusts the feature representation to improve the effect of model reconstruction. Finally, model fusion processing is performed by inputting the basic decomposition optimized feature data and detailed decomposition optimized feature data of each reconstructed model into a preset graph data decoder. This step fuses the optimized feature data and reassembles it into a complete graph data model. It also comprehensively considers the basic and detailed features, thereby obtaining a more accurate and complete graph data model.

[0130] Preferably, step S42 includes the following steps:

[0131] Step S421: performing feature vector space mapping conversion on the low-frequency basic feature data of each model and the basic decomposition feature data of each reconstructed model, respectively, to obtain the low-frequency basic feature vector space of each model and the basic decomposition feature vector space of each reconstructed model;

[0132] The embodiment of the present invention uses a feature vector mapping method to map and transform the low-frequency basic feature data of each model and the basic decomposition feature data of each reconstructed model, so as to convert the original feature representation into a more representative and comparable vector space representation, so that the features between different models can be compared and analyzed in the same space, and finally the low-frequency basic feature vector space of each model and the basic decomposition feature vector space of each reconstructed model are obtained.

[0133] Step S422: performing feature space alignment processing on the low-frequency basic feature vector space of each model and the basic decomposition feature vector space of each reconstructed model to obtain the low-frequency basic feature alignment space of each model and the basic decomposition feature alignment space of each reconstructed model;

[0134] In an embodiment of the present invention, the low-frequency basic feature vector space of each model and the basic decomposition feature vector space of each reconstructed model are aligned by using a corresponding feature alignment algorithm (such as the least squares method or the iterative nearest point algorithm, etc.) to adjust the feature spaces of different models to a unified coordinate system so that they have consistent directions and scales, and ensure that the feature spaces of each model and the reconstructed model can be aligned, and finally obtain the low-frequency basic feature alignment space of each model and the basic decomposition feature alignment space of each reconstructed model.

[0135] Step S423: using the calculation formula for the loss of the stitching basic feature, the loss of the low-frequency basic feature alignment space of each model and the basic decomposition feature alignment space of each reconstructed model are calculated to obtain the stitching basic feature loss of each reconstructed model.

[0136] The embodiment of the present invention combines the basic feature loss weight parameter, the number parameter of the basic features, the number parameter of the basic decomposition features, the vector value of the basic features, the vector value of the basic decomposition features, the correlation coefficient operator between the basic features in the low-frequency basic feature alignment space and the basic decomposition features in the basic decomposition feature alignment space, and related parameters to form a suitable splicing basic feature loss calculation formula to perform loss calculation on the low-frequency basic feature alignment space of each model and the basic decomposition feature alignment space of each reconstructed model to quantify the degree of information loss of the reconstructed model on the low-frequency basic features, and finally obtain the splicing basic feature loss of each reconstructed model.

[0137] The present invention first performs feature vector space mapping conversion on the low-frequency basic feature data of each model and the basic decomposition feature data of each reconstruction model. This process aims to convert the original feature representation into a more representative and comparable vector space representation, so that the features between different models can be compared and analyzed in the same space, thereby providing a basis for subsequent feature alignment and loss calculation. Then, by performing feature space alignment processing on the low-frequency basic feature vector space of each model and the basic decomposition feature vector space of each reconstruction model, the feature spaces of different models can be adjusted to a unified coordinate system so that they have consistent directions and scales, thereby facilitating subsequent comparison and analysis. This process obtains the low-frequency basic feature alignment space of each model and the basic decomposition feature alignment space of each reconstruction model, thereby providing a unified benchmark for further calculation of loss. Finally, by using the splicing basic feature loss calculation formula to calculate the loss of the low-frequency basic feature alignment space of each model and the basic decomposition feature alignment space of each reconstructed model, the loss of each model in basic features after feature alignment can be evaluated, that is, the splicing basic feature loss. The purpose of this step is to quantify the degree of information loss of the model in low-frequency basic features, thereby providing guidance and basis for subsequent feature optimization and model fusion.

[0138] Preferably, the calculation formula for the splicing basic feature loss in step S423 is specifically:

[0139]

[0140] Where, is the concatenation basic feature loss, λ is the basic feature loss weight parameter, N is the total number of basic features in the low-frequency basic feature vector space, j is the item index parameter of the basic feature, M is the total number of basic decomposition features in the basic decomposition feature alignment space, k is the item index parameter of the basic decomposition feature, f(x j ) is the vector value of the jth basic feature in the low-frequency basic feature alignment space, g(y k ) is the vector value of the kth basic decomposition feature in the basic decomposition feature alignment space, ∈ jk is the correlation coefficient operator between the jth basic feature in the low-frequency basic feature alignment space and the kth basic decomposition feature in the basic decomposition feature alignment space.

[0141] The present invention uses a specific mathematical model and has been verified to obtain a splicing basic feature loss calculation formula, which is used to calculate the loss of the low-frequency basic feature alignment space of each model and the basic decomposition feature alignment space of each reconstructed model. The calculation goal of the splicing basic feature loss calculation formula is to measure the performance loss after the feature space alignment process. This loss is calculated by comparing the difference between the basic feature data reconstructed by the model and the original low-frequency basic feature data. The basic feature loss weight parameter in the formula is used to adjust the proportion of the basic feature loss in the overall loss. It can be adjusted according to the task requirements and model performance requirements to better balance the relationship between feature alignment and performance loss. In addition, the formula measures the difference between the model reconstruction features and the original features based on the mean square error (MSE), and uses the correlation coefficient operator between the basic features and the basic decomposition features to adjust the loss difference between them. At the same time, the overall splicing basic feature loss is obtained by weighted summing the differences of all basic features. Through this formula, the performance loss of the model after the feature space alignment process can be effectively quantitatively evaluated, thereby providing important guidance for further model improvement and optimization. Therefore, the formula can fully consider the splicing basic feature loss. The basic feature loss weight parameter λ, the total number of basic features N in the low-frequency basic feature vector space, the item index parameter j of the basic feature, the total number of basic decomposition features M in the basic decomposition feature alignment space, the item index parameter k of the basic decomposition feature, the vector value f(x) of the jth basic feature in the low-frequency basic feature alignment space j ), the vector value g(y of the kth basic decomposition feature in the basic decomposition feature alignment space k ) and the correlation coefficient operator ∈ between the jth basic feature in the low-frequency basic feature alignment space and the kth basic decomposition feature in the basic decomposition feature alignment space jk , according to the loss of basic features of splicing The mutual correlation between the above parameters constitutes a functional relationship This formula can realize the loss calculation process of the low-frequency basic feature alignment space of each model and the basic decomposition feature alignment space of each reconstructed model, thereby improving the accuracy and applicability of the splicing basic feature loss calculation formula.

[0142] The present invention is therefore intended to be illustrative and non-restrictive in all respects, with the scope of the invention being defined by the appended claims rather than the foregoing description, and all changes that come within the meaning and range of equivalents of the application documents are intended to be embraced therein.

[0143] The foregoing description is intended only to provide specific embodiments of the present invention, which will enable those skilled in the art to understand and implement the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not intended to be limited to the embodiments shown herein, but is to be construed in the widest possible manner consistent with the principles and novel features disclosed herein.

Claims

1. A model splicing and fusion method based on graph database, characterized in that: The following steps are involved: Step S1: establishing a data model by converting the data of each model into a graph to obtain graph data models of different modalities; constructing a graph database for the graph data models of different modalities to generate a model graph database; Step S2: performing a model graph structure retrieval process on the model graph database to obtain a model graph structure dataset; performing a graph shallow feature analysis on the model graph structure dataset to obtain graph node shallow feature data and graph edge shallow feature data of each model; Step S3: Performing feature pattern screening analysis on the shallow feature data of the graph nodes and the shallow feature data of the graph edges of each model to obtain low-frequency basic feature data and high-frequency detail feature data of each model; performing model splicing and reconstruction processing on the graph data model based on the low-frequency basic feature data and high-frequency detail feature data of each model to obtain a graph data splicing and reconstruction model; wherein step S3 includes the following steps: Step S31: performing graphic feature fusion analysis on the shallow feature data of the graphic nodes and the shallow feature data of the graphic edges of each model to obtain the graphic structure feature fusion data of each model; Step S32: performing feature pattern change analysis on the graphic structure feature fusion data of each model to obtain graphic structure feature change pattern data of each model; Step S33: performing fluctuation frequency calculation on the graphic structure characteristic change pattern data of each model using a characteristic pattern change fluctuation frequency calculation formula to obtain the graphic structure characteristic pattern change fluctuation frequency of each model; The calculation formula for the characteristic mode change fluctuation frequency is as follows: Where n is the total number of models, i is the item index parameter of the model, and f i is the fluctuation frequency of the graphical structural characteristic mode of the i-th model, α i is the amplitude parameter of the graphical structural characteristic mode of the i-th model, β i is the graph structure characteristic mode attenuation coefficient of the i-th model, γ i is the influence factor of the graph structure characteristic pattern change of the i-th model, δ i is the bandwidth parameter of the graphical structure characteristic mode of the i-th model, exp is the exponential function, ω i is the characteristic mode fluctuation phase angle of the i-th model, μ i is the mean value of the characteristic mode fluctuation angle distribution of the i-th model, σ i is the standard deviation of the characteristic mode fluctuation angle distribution of the i-th model, η i is the fluctuation frequency correction coefficient of the i-th model; Step S34: comparing and judging the graphic structure feature pattern change fluctuation frequency of each model according to a preset feature pattern change fluctuation threshold; if the graphic structure feature pattern change fluctuation frequency of each model is less than the preset feature pattern change fluctuation threshold, the graphic structure feature fusion data of the corresponding model is judged as low-frequency basic feature data to obtain the low-frequency basic feature data of each model; if the graphic structure feature pattern change fluctuation frequency of each model is greater than or equal to the preset feature pattern change fluctuation threshold, the graphic structure feature fusion data of the corresponding model is judged as high-frequency detail feature data to obtain the high-frequency detail feature data of each model; Step S35: performing model splicing and reconstruction processing on the graph data model based on the low-frequency basic feature data and high-frequency detail feature data of each model to obtain a graph data splicing and reconstruction model; Step S4: performing feature decomposition processing on each graph data splicing reconstruction model to obtain basic decomposition feature data and detail decomposition feature data of each reconstructed model; performing model splicing loss analysis on the basic decomposition feature data and detail decomposition feature data of each reconstructed model based on the low-frequency basic feature data and high-frequency detail feature data of each model to obtain the splicing basic feature loss and splicing detail feature loss of each reconstructed model; performing model fusion processing on the basic decomposition feature data and detail decomposition feature data of each reconstructed model based on the splicing basic feature loss and splicing detail feature loss of each reconstructed model to obtain a graph data fusion model.

2. The model splicing and fusion method based on graph database according to claim 1 is characterized in that: Step S1 includes the following steps: Step S11: Building a data model by converting the data of each model into a graph to obtain graph data models of different modalities; Step S12: Perform cross-modal graph embedding processing on graph data models of different modalities to obtain a unified graph data space model; Step S13: Using graph neural network technology to perform graph representation learning on the unified graph data space model to obtain effective representation feature vectors of dynamic graph data; Step S14: extracting graph node and edge feature vectors from the dynamic graph data's effective representation feature vectors to obtain dynamic graph node representation feature vectors and dynamic graph edge representation feature vectors; Step S15: constructing a graph database for the dynamic graph node representation feature vectors and the dynamic graph edge representation feature vectors based on the unified graph data space model to generate a model graph database.

3. The model splicing and fusion method based on graph database according to claim 2 is characterized in that: Step S15 includes the following steps: Step S151: Analyze the graph data characteristics of the unified graph data space model to obtain graph data model characteristics; analyze the real-time query requirements of the unified graph data space model to obtain graph data real-time query requirement data; Step S152: performing query access pattern analysis on the unified graph data space model based on the graph data real-time query demand data to obtain graph data query access pattern data; Step S153: performing graph index identification and analysis on the unified graph data space model according to the graph data model characteristics and the graph data query access pattern data to obtain the initial index structure of the graph database; Step S154: performing graph data temporal change analysis on the dynamic graph node representation feature vector and the dynamic graph edge representation feature vector to obtain graph data temporal dynamic change data; performing adaptive index adjustment on the initial index structure of the graph database based on the graph data temporal dynamic change data to obtain a graph database adaptive index optimization structure; Step S155: construct a graph database based on the dynamic graph node representation feature vector and the dynamic graph edge representation feature vector according to the graph database adaptive index optimization structure to generate a model graph database.

4. The model splicing and fusion method based on graph database according to claim 1 is characterized in that: Step S2 includes the following steps: Step S21: performing a model graph structure retrieval process on the model graph database to obtain a model graph structure dataset; Step S22: extracting the node structure and node semantic relationship of the model graph structure dataset to obtain the graph node structure and the contextual semantic relationship of the graph nodes of each model; Step S23: performing node shallow feature analysis on the model graph structure dataset based on the graph node structure of each model and the contextual semantic relationship of the graph nodes to obtain shallow feature data of the graph nodes of each model; Step S24: performing graph edge type identification analysis on the model graph structure dataset to obtain graph edge type relationships between various models; Step S25: performing graph edge shallow feature analysis on the model graph structure dataset based on the graph edge type relationship between the various models to obtain graph edge shallow feature data of the various models.

5. The model splicing and fusion method based on graph database according to claim 4 is characterized in that: Step S25 includes the following steps: Step S251: performing vector space embedding processing on the graph edge type relationships between the various models to obtain edge type relationship embedding vectors of the various models; Step S252: performing edge type structure analysis on the edge type relationship embedding vectors of each model according to the graph edge type relationships between the models to obtain different types of relationship edge structures of each model; Step S253: performing semantic relationship recognition analysis on the edge type relationship embedding vectors of each model to obtain different types of edge semantic relationships of each model; Step S254: performing graph edge shallow feature analysis on the model graph structure dataset according to different types of relationship edge structures of each model and different types of edge semantic relationships of each model to obtain graph edge shallow feature data of each model.

6. The model splicing and fusion method based on graph database according to claim 1 is characterized in that: Step S35 includes the following steps: Step S351: performing low-frequency feature global sampling processing on the low-frequency basic feature data of each model to obtain low-frequency feature global information data of each model; Step S352: performing high-frequency feature local sampling processing on the high-frequency detail feature data of each model to obtain high-frequency feature local information data of each model; Step S353: performing feature fusion analysis on the low-frequency feature global information data and the high-frequency feature local information data of each model to obtain effective fusion feature data of the graph data of each model; Step S354: effectively fuse the feature data of the graph data of each model and input it into a preset graph data decoder to perform model splicing and reconstruction processing on the corresponding graph data model to obtain a graph data splicing and reconstruction model.

7. The model splicing and fusion method based on graph database according to claim 1 is characterized in that: Step S4 includes the following steps: Step S41: inputting the reconstructed model of each graph data splicing into a preset graph data encoder for feature decomposition processing to obtain basic decomposition feature data and detail decomposition feature data of each reconstructed model; Step S42: performing model splicing basic feature loss analysis on the basic decomposition feature data of each reconstructed model based on the low-frequency basic feature data of each model to obtain the splicing basic feature loss of each reconstructed model; Step S43: performing model splicing detail feature loss analysis on the detail decomposition feature data of each reconstructed model based on the high-frequency detail feature data of each model to obtain the splicing detail feature loss of each reconstructed model; Step S44: performing feature optimization processing on the basic decomposition feature data and the detail decomposition feature data of each reconstructed model based on the splicing basic feature loss and the splicing detail feature loss of each reconstructed model to obtain the basic decomposition optimized feature data and the detail decomposition optimized feature data of each reconstructed model; Step S45: Inputting the basic decomposition optimization feature data and the detail decomposition optimization feature data of each reconstructed model into a preset graph data decoder for model fusion processing to obtain a graph data fusion model.

8. The model splicing and fusion method based on graph database according to claim 7 is characterized in that: Step S42 includes the following steps: Step S421: performing feature vector space mapping conversion on the low-frequency basic feature data of each model and the basic decomposition feature data of each reconstructed model, respectively, to obtain the low-frequency basic feature vector space of each model and the basic decomposition feature vector space of each reconstructed model; Step S422: performing feature space alignment processing on the low-frequency basic feature vector space of each model and the basic decomposition feature vector space of each reconstructed model to obtain the low-frequency basic feature alignment space of each model and the basic decomposition feature alignment space of each reconstructed model; Step S423: using the calculation formula for the loss of the stitching basic feature, the loss of the low-frequency basic feature alignment space of each model and the basic decomposition feature alignment space of each reconstructed model are calculated to obtain the stitching basic feature loss of each reconstructed model.

9. The model splicing and fusion method based on graph database according to claim 8 is characterized in that: The calculation formula for the splicing basic feature loss in step S423 is specifically: Where, is the concatenation basic feature loss, λ is the basic feature loss weight parameter, N is the total number of basic features in the low-frequency basic feature vector space, j is the item index parameter of the basic feature, M is the total number of basic decomposition features in the basic decomposition feature alignment space, k is the item index parameter of the basic decomposition feature, f(x j ) is the vector value of the jth basic feature in the low-frequency basic feature alignment space, g(y k ) is the vector value of the kth basic decomposition feature in the basic decomposition feature alignment space, ∈ jk is the correlation coefficient operator between the jth basic feature in the low-frequency basic feature alignment space and the kth basic decomposition feature in the basic decomposition feature alignment space.

Images