Multi-value chain evolution prediction method, system and storage medium

The distributed cooperative time series graph neural network model addresses data sharing challenges by predicting entity relationships and states in multi-value chains, ensuring privacy and enhancing prediction accuracy.

JP2024167887AActive Publication Date: 2024-12-04CHENGDU UNIVERSITY OF TECHNOLOGY
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
JP2024076248
Authority / Receiving Office
JP · JP
Patent Type
Applications
Current Assignee / Owner
Priority Date
2023-05-22
Filing Date
2024-05-09
Publication Date
2024-12-04
Estimated Expiration
2044-05-09

AI Technical Summary

Technical Problem

Existing prediction methods for multi-value chains fail to consider dynamic time delays and rely on neural networks, leading to underutilization of data due to data compliance and security concerns, hindering effective data sharing and relationship understanding between entities in the chain.

Method used

A method using a distributed cooperative time series graph neural network model that allows clients to train locally without sharing original data, predicting entity relationships by calculating feature vectors and edge information based on local and global network parameters.

Benefits of technology

Enables dynamic prediction of entity states and relationships within multi-value chains without sharing original data, enhancing data utilization and privacy protection while improving prediction accuracy.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 2024167887000001_ABST
    Figure 2024167887000001_ABST
Patent Text Reader

Abstract

To provide a multi-value chain evolution prediction method, a system and a storage medium.SOLUTION: In a dispersion cooperation time-series graph neural network, on the assumption that each client as a subject in a value chain does not share local original data, a server mutually acts on each client participating in a dispersion cooperation time-series graph neural network model training on the basis of local original data of the client, and trains a dispersion cooperation time-series graph neural network model, and the client dynamically predicts a state of each of the subjects and a relation between each of the subjects on the basis of the model.SELECTED DRAWING: Figure 2
Need to check novelty before this filing date? Find Prior Art

Description

[Technical field]

[0001] The present invention belongs to the field of artificial intelligence, specifically, a method for predicting the evolution of multi-value chains; This invention relates to a system and a storage medium. [Background technology]

[0002] Significant improvements in machine hardware performance have made it possible to apply deep learning based on neural networks. Although the method has been very successful in areas such as intelligent recommendation and network performance prediction, The prediction methods in seldom take dynamic time delays into account, and rely on neural networks. Most predictions take into account topology, and output information that changes over time. It is not possible. Each actor in society must take responsibility for data compliance, trade secrets, information security, and business competition. As companies in the upstream and downstream of the multi-value chain are concerned with these issues, Companies and individuals are reluctant to disclose their original data, resulting in upstream, midstream and downstream The degree of data sharing is low, large amounts of data are not fully utilized, and the Companies, businesses and individuals who set up their own companies have difficulty understanding the relationships between the other entities. The challenge is to predict the relationships between subjects without obtaining the original data of the body. This has become an urgent issue for companies involved in the food supply chain. Summary of the Invention

[0003] The present invention has been made in consideration of the above-mentioned problems of the conventional technology, and Under the assumption that the subjects (e.g., each client) do not share local original data, A multi-value chain that can dynamically predict the state of actors and the relationships between each actor An evolutionary prediction method, system, and storage medium are provided. The present invention provides the following solutions: In one embodiment, the present invention provides a method for predicting the evolution of a multi-value chain, including: Clients who participate in the distributed collaborative time series graph neural network model training For each client in the The distribution time series graph data of each client and the neighboring clients of each client are Based on the feature vector of the node in which each client is located, The feature vector is a node where each of the clients is located and a It is used to characterize the information of edges between nodes where neighboring clients of an The distributed time series graph data of each client is used to node information of the client and the node where each client is located and the neighboring nodes of each client The information includes information on edges between nodes where the clients are located, and each of the clients Based on the feature vector of the node where each of the clients is located, The local network parameters of the distributed time series graph neural network model are generating said local network parameters and transmitting said local network parameters to a server; Based on the local network parameters transmitted from each client, The global network parameters of the distributed time series graph neural network model meter and broadcast the global network parameters to each of the clients. Finally, each of the clients broadcasts the global network parameters. Using the distributed time series graph neural network model of each client, Update the network parameters and display the current time series distribution graph data for each client. The data is input to the updated distributed and linked time series graph neural network model, and each of the classes is The nodes where the clients are located and the nodes where the neighboring clients of each client are located. It outputs the predicted values ​​of the edges between the nodes. In the method for predicting the evolution of a multi-value chain provided by the present invention, a distributed linked time series Each client participating in graph neural network model training is connected to itself and its neighbors. The node where the client is located (the node where the neighboring client of one client is located) The nodes that receive the data are also called the neighbors of the client. The server obtains only the information of the nodes where all clients are located, and the client The client uses the local original data (i.e. data generated locally on the client) If the data is not shared (distributed time series graph data), each client must interact with the server. A globally distributed and collaborative time-series graph neural network that can be applied to each client We train the model and then compare the nodes where each client is located in the multi-value chain. Dynamically completes prediction of edges between the nodes where the neighboring clients are located, and calculates the edge Based on the prediction results, the state of each agent and the relationships between the agents are dynamically predicted. In another embodiment, the local network of the distributed and linked time-series graph neural network is The network parameters include local weights, local feature transformation matrices and local losses. The global network parameters of a distributed time series graph neural network are These include global weights, global feature transformation matrix and global loss. In another embodiment, the distributed and linked time series graph neural network model is a feature learning neural network. a feature learning network and a label prediction network, , the distributed time series graph data of each of the clients at a historical point in time and Each client is assigned a feature vector based on the feature vector of the node where the neighboring client is located. is used to output the feature vector of the node where the label prediction network is located. The network determines whether each of the clients is located on the basis of the feature vector of the node in which the client is located. The nodes where the clients are located and the nodes where the neighboring clients of each client are located The feature learning network is used to output the predicted value of the edge. The label prediction network consists of a single layer and a fully connected layer. In another embodiment, the feature vector of the node where each client i is located is JPEG2024167887000002.jpg13162, where JPEG2024167887000003.jpg814 is JPEG2024167887000004.jpg The client at time 43 The client output from the l-th layer of the feature learning network corresponding to JPEG2024167887000005.jpg53 JPEG2024167887000006.jpg53 indicates the feature vector of the node where it is located, JPEG2024167887000007.jpg1225 is the client's Adjacent clients for JPEG2024167887000008.jpg53 The first step of the feature learning network corresponding to JPEG2024167887000009.jpg73 JPEG2024167887000010.jpg515 indicates the feature vector of the node where the adjacent client j is located, output from the layer, JPEG2024167887000011.jpg1025 is JPEG2024167887000012.jpg The client at time 43 The first step of the feature learning network corresponding to JPEG2024167887000013.jpg53 JPEG2024167887000014.jpg515th layer output from the client JPEG2024167887000015.jpg53 indicates the feature vector of the node where it is located, JPEG2024167887000016.jpg631 shows an aggregation operation, JPEG2024167887000017.jpg519 shows linear operations, JPEG2024167887000018.jpg717 shows the local feature transformation matrix. In another embodiment, the feature vector of the node where each client i is located is JPEG2024167887000019.jpg13170, where JPEG2024167887000020.jpg814 is JPEG2024167887000021.jpg43 The client output from the l-th layer of the feature learning network corresponding to JPEG2024167887000022.jpg53 JPEG2024167887000023.jpg53 indicates the feature vector of the node where it is located, JPEG2024167887000024.jpg1225 is JPEG2024167887000025.jpg43 Adjacent clients for JPEG2024167887000026.jpg53 The first step of the feature learning network corresponding to JPEG2024167887000027.jpg73 JPEG2024167887000028.jpg515 output from the adjacent client JPEG2024167887000029.jpg73 indicates the feature vector of the node where it is located, JPEG2024167887000030.jpg1025 is JPEG2024167887000031.jpg43 The first step of the feature learning network corresponding to JPEG2024167887000032.jpg53 JPEG2024167887000033.jpg515th layer output from the client JPEG2024167887000034.jpg53 indicates the feature vector of the node where it is located, JPEG2024167887000035.jpg631 shows an aggregation operation, JPEG2024167887000036.jpg519 shows linear operations, JPEG2024167887000037.jpg717 indicates the local feature transformation matrix, JPEG2024167887000038.jpg68 is the client The node where JPEG2024167887000039.jpg53 is located and the client Adjacent clients for JPEG2024167887000040.jpg53 This shows the correlation coefficient between the nodes where JPEG2024167887000041.jpg73 is located. In another embodiment, each of the clients may receive a feature vector of the node in which the client is located. Based on the vector, the distributed and linked time series graph neural network of each client is The generating of local network parameters of the model includes: The feature vector of the node where each client is located is assigned to the label corresponding to each client. The nodes where the clients are located and the nodes where the clients are located are input to the network. generating predictions of edges between nodes where neighboring clients of the same client are located; The client is connected to the node where the client is located and the neighboring nodes of the client. Based on the predicted value of the edge between the nodes where the clients are located and the true value of the edge, determining a local loss of the distributed, coordinated time-series graph neural network model; and each of the clients performs backpropagation based on the local loss, and each of the clients Local network of distributed time series graph neural network model of clients and generating a query parameter. In another embodiment, the method for predicting the evolution of a multi-value chain provided by the present invention includes: Each of the clients provides the server with the distributed, coordinated time-series graph neural network. The server then receives the local loss of the model from each client. Based on the local loss, the distributed time series graph neural network model Based on this, the server may further determine a global loss for each client. The distributed link is based on the local network parameters transmitted from the client. We obtain global network parameters of a time series graph neural network model. That is, the server receives the local network traffic sent from each client. Perform backpropagation based on the parameters and the global loss, Obtaining global network parameters of a rough neural network model Includes. In one embodiment, the present invention provides a client device, the client device being connected to a server. and each of the above-mentioned embodiments of the method for predicting the evolution of a multi-value chain. Realizes the actions to be performed by the client. In one embodiment, the present invention interacts with each of a plurality of clients, The method for predicting the evolution of a multi-value chain according to any embodiment of the present invention is performed by a server. The ISP provides the server that is used to implement the operations described above. In one embodiment, the present invention provides a system for predicting the evolution of a multi-value chain, comprising: and a plurality of clients, The server participates in training a neural network model among the plurality of clients. Interact with the client and use one of the above multi-value chain evolution forecasting methods By executing the above, the distributed and coordinated time series graph neural network model is trained. Based on the distributed and linked time series graph neural network model, Between the node where each client in the chain is located and the node where other clients are located Predict the edges of. In one embodiment, the present invention provides a computer readable storage medium, the computer readable storage medium comprising: A medium having computer program instructions stored thereon, said computer program instructions When executed by a multi-core processor, the multi-core processor is Implement a method for predicting value chain evolution. [Brief description of the drawings]

[0004] [Figure 1] 1 is a schematic diagram of a system architecture provided by the present invention; [Diagram 2] FIG. 1 is a schematic diagram of the architecture of a distributed, collaborative time-series graph neural network provided by the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0005] In order to clearly describe the embodiments of the present application, some of the following embodiments may appear first. Introduce some concepts. Original data refers to data with characteristics and characteristics acquired by each actor in the multi-value chain during a specific period. The original data is actual data including samples, and the original data is defined by a person skilled in the art. The original data is collected according to the rules and methods set forth above, and the original data is in the form of a tensor or table. The original data of the client is recorded in the form of a formula, and the characteristics and samples are For example, feature 1 is the amount of car sales, and the samples are for different clients. The corresponding January car sales data, February car sales data, March car sales data The data includes,feature 2 is the vehicle inventory, and the samples are,samples of different clients. The corresponding January vehicle inventory data, February vehicle inventory data, and March vehicle inventory data are The inventory data includes..., feature 3 is accessory sales volume, and the samples are The accessory sales volume data for January, February, March, and The accessory sales data includes the accessory inventory data, and feature 4 includes the accessory inventory data, and so on. be. The features and samples referred to in the present invention are not limited to those mentioned above, but include the following features: characterizes some or all of the data actually collected or recorded at each client The samples may be data actually collected or recorded at each client. It will be apparent to those skilled in the art that the input data may be a part or the whole of the data. The characteristics, dimensions, types, etc. of the original data of different clients are different. The characteristics, dimensions, types, etc. of the original data between different groups may also be different. Instead, the client may collect original data on a monthly basis and original data on a weekly basis. Data may be collected, and original data may be collected or recorded in real time; Those skilled in the art can arbitrarily set this according to actual needs. Metadata is data that describes an object, such as an information resource or data. The purpose of use is to identify resources, evaluate resources, track changes in resource usage, and Achieve simple and efficient management of network data, effectively discover information resources, The realization of search, integrated organization, and effective management of resource usage. Metadata is publicly available data. It includes information about the characteristics of the input data, Size, dimensions, attributes, generation time, type, shape, client flags, variable names, dimensions, etc. Metadata is data that describes other data or that is relevant to a particular resource. It is structured data used to provide information about each client's input data. generated by the client or server side based on the original data The metadata of the present invention is data that describes the original data of the client. Data that describes the attributes or characteristics of the original data. I want to be understood. Graph Neural Network (GNN) is a method for processing graph data (graph data is data that has graph structure characteristics). It is a neural network for learning from graph data and extracting features and Extract and discover patterns and perform graph learning tasks such as convergence, classification, prediction, segmentation, and generation. In the present invention, each client interacts with each other and the graphics It is possible to generate some data with the characteristics of the data structure, where each client Each client in the instance corresponds to a node (also called a graph node), The nodes corresponding to the clients form a graph. According to the method for predicting the evolution of a multi-value chain proposed by the present invention, a client No original data is shared, only metadata is shared, which allows the Privacy and security are protected, and no original data is shared. It enables secure collaboration between companies in the value chain. become. This invention provides a distributed and coordinated time series graph neural network (model), and — Adopts a “multi-client” distributed architecture, where the server is mainly responsible for parameters. It manages the data and coordinates clients to form a distributed, coordinated time-series graph neural network. By adopting a distributed and coordinated time series graph neural network, each client The distributed and linked time series graph neural network corresponding to the It is a component of the entire neural network and is trained locally on each client. The parameters of the obtained distributed linked time series graph neural networks are different. Optionally, a distributed, federated time-series graph neural network that is local to each client can be homogeneous or heterogeneous, i.e., each client’s local distributed federated time system The structures of the column graph neural networks may be the same or different. In addition, the distributed and linked time series graph neural network corresponds to the graph at each point in time. ,The nodes in the figure refer to clients or companies, and a client corresponds to a company, and The nodes and the edges between them are connected to each other through cooperation, trading, supply and demand relationships, etc. It shows the relationships between clients or between companies, and changes in the graph over time. The topological relationships change, and the edges in the graph change. Without sharing null data, predictions about future edges in the graph, or when a company will go under, We build a connected time series graph neural network that can predict whether or not a The present invention will now be described in detail with reference to the accompanying drawings in conjunction with embodiments. Referring first to FIG. 1, a system architecture in which exemplary embodiments according to the present disclosure may be used is shown. FIG. 1 is a schematic diagram showing a structure 100. FIG. 1 is a schematic diagram illustrating an example of a system architecture 100 according to an embodiment of the present disclosure. Figure 1 shows the prediction method for the evolution of multi-value chains and the hardware architecture. It should be noted that this is also referred to as an architectural diagram of the hardware operating environment of the present invention. The embodiment is a client device, which may be a terminal device such as a PC, a portable computer, etc. It is based on. The system architecture shown in Figure 1 is a distributed "server-multiple clients" architecture. The architecture shown in FIG. 1 is only one of the clients. In a distributed "server-multiple-client" architecture, The client then uses the original data to locally identify the current nodes and connected edges. A sub-model that characterizes the message (i.e., a local distributed time series graph neural network) It is used to train the training model. Preferably, each client has an original data module, a local parameter module, and module, data collection module, data labeling module, message passing module It has a local training module and a blockchain. The data module is used to store the original data of the client, The data collection module is used to collect and record the original data of the client. The local parameter module is a distributed federated time series graph module for clients. It is used to store all the parameters of the neural network and is used for data labeling. The training module uses data to train a distributed, coordinated time-series graph neural network. The message passing module handles client and server communication. Supports multiple modules for parallel communication and uploading information (i.e. blocking messages) The local training module is used to Ant is used to train a distributed federated time series graph neural network, The chain ensures traceability of all communications between the server and the client, and of information leaks. The server is used to record the transaction and non-repudiation. It is used to manage global parameters of the network. Preferably, the server comprises a metadata module, a parameter database module, Parameter management module, communication module and global training module and blog A metadata module manages and stores metadata. The parameter database module is used to perform distributed and linked time series graph neural networking. It is used to store all the parameters of the network. The model manages and updates the parameters of a distributed, coordinated time-series graph neural network. The communication module is a server and multi-client message passing module. It is used to support parallel communication and information uploading, and is a global training module. The protocol is used to fuse data sent by multiple clients and is used to The chain ensures all server-client communications, traceability and denial of data leakage. In one embodiment of the present invention, each distributed and linked time series graph is used to record the prevention of The local network performs training locally on the client, and the server mainly handles parameters. Data processing is performed (i.e., the distributed and coordinated time series graph obtained by training each client). processing the parameters of the neural network), said processing including averaging (i.e. The server uses the distributed time series graph neural network obtained from each client. In another embodiment of the present invention, each distributed and linked time series The graph neural network is trained locally on the client and the resulting distributed correlation Local network parameters of a mobile time series graph neural network are sent to the server The server then uses a neural network to analyze the local network The processing includes parameter fitting using a neural network. including forecasting and / or predictions. The hardware parts of the client and server are the processor, e.g., CPU, network It consists of a data interface, a user interface, a memory, and a communication bus. A communication bus is used to realize the connection communication between the different components. A face is a display and an input unit, e.g. a keyboard. board) and selectable user interface is standard wired interface , and a wireless interface. The network interface may be a selectable standard Wired interface, wireless interface (e.g. WI-FI interface, BLE The memory may include high-speed RAM memory. , or non-volatile memory, such as disk The memory may be a storage device that is selectively independent of the processor. This is also fine. The server and client structure shown in Figure 1 does not limit the server or client. is not intended to be limiting and servers and clients may have more or fewer components than those shown. It may consist of a combination of specific components or different arrangements of components. It will be understood by those skilled in the art that this may also be achieved. As shown in Figure 2, memory as a computer storage medium is used by the operating system program, network communication module program, user interface module program The program may include a program for a neural network of a time series graph. For multiple clients in a multi-value chain, a node is assigned to each client. The nodes and clients correspond one-to-one. The system manages the programs of the hardware and software resources of the client device. Manage and control distributed, coordinated time series graph neural network programs and other Used to support the operation of software or programs. In a graph neural network, if we assume that the nodes are constant, This corresponds to the topological structure of the graph, and the edges of the graph change over time. Figure 1 In the server shown in Figure 1, the communication module is mainly between the server and client A, and Client B, Client C, Client D, and Client E It is used to send and receive requests, data, etc., and is a message packet in multi-client. The Sing module is mainly used to send and receive requests and data between each client and the server. where there may be a greater or lesser number of clients. Preferably, the communication module and the message passing module are connected to respective networks. Data is communicated through the interface, and the processor reads the distributed coordinated time stored in the memory. Call the sequence graph neural network program and provide it according to the embodiment of the present invention. The method may be used to implement a method for predicting the evolution of a multi-value chain. The method for predicting the evolution of a multi-value chain provided by an embodiment of the present invention is as follows: Includes 01~S104. S101, the client establishes a distributed collaborative time series graph neural network model do. The distributed time series graph neural network model is a feature learning network and a learning The feature learning network mainly contains a message passing layer. The client receives messages from its neighbors and passes messages to the neighbors. Graph convolution and graph attention are applied via a filtering layer. (graph attention), graph transformation The label prediction network consists of a fully connected layer, and the nodes It is used to predict the edges between This invention is a distributed and linked time series graph neural network with a server-multi-client structure. It provides a network architecture and JPEG2024167887000042.jpg54 and JPEG2024167887000043.jpg44 clients JPEG2024167887000044.jpg840, and the company serves one client, with one node for each client. JPEG2024167887000045.jpg56 is established, and the nodes corresponding to the clients in the multi-value chain form a mesh Structure time series graph data JPEG2024167887000046.jpg729 is formed, and distributed collaborative graph neural network training is performed in the multi-value chain. All clients participating in the network are connected to a single distributed time series graph neural network. In the present invention, unless mutually contradictory, time series graph data JPEG2024167887000047.jpg729 is a figure It is abbreviated as JPEG2024167887000048.jpg54, and a collection of time-series graph data at multiple points in time is called multi-temporal graph data. The graph is called JPEG2024167887000049.jpg54 exists in the entire network space, or logically, and each client JPEG2024167887000050.jpg55 is a graph 1 node in JPEG2024167887000051.jpg54 Supports JPEG2024167887000052.jpg56. In this invention, a distributed and linked time series graph neural network is adopted, and each client Therefore, the established distributed and linked time series graph neural network is a distributed and linked time series graph It is a component of the entire neural network and is trained locally on each client. The parameters of the obtained distributed time series graph neural network are different and the complete The distributed time series graph data is not known to any entity in the network. contains any client or server, where each client JPEG2024167887000053.jpg55 is the client Adjacent clients for JPEG2024167887000054.jpg55 Only allow JPEG2024167887000055.jpg65 to communicate with the server, JPEG2024167887000056.jpg44 is a natural number, or JPEG2024167887000057.jpg44 is a positive integer, JPEG2024167887000058.jpg53 indicates the serial number of the client or node, JPEG2024167887000059.jpg53 is 0~ JPEG2024167887000060.jpg44 takes a value between JPEG2024167887000061.jpg73 is the neighboring client or the node where the neighboring client is located (or the neighboring node The serial number of the JPEG2024167887000062.jpg73 is 0~ In one embodiment, the neighboring client or neighboring node in the present invention is: It refers to a first-order neighboring client or a first-order neighboring node. The neighboring nodes include first-order neighboring nodes and / or second-order neighboring nodes. The neighboring nodes in the present invention may be first-order neighboring nodes and / or second-order neighboring nodes and / or other neighboring nodes. Any node, including higher-order neighbors For JPEG2024167887000064.jpg56, the adjacency list on the graph JPEG2024167887000065.jpg617 constitutes the first feature, i.e., the structural relationship of the graph. In an alternative embodiment, the server provides the client with a distributed, coordinated time series graph neural network. The request message includes a request for the data source required for training. Preferably, these data are metadata, e.g., value flows, demand flows, business flows, etc. Client If JPEG2024167887000066.jpg56 is willing to participate, the server will feed a message indicating that it will receive the request. In another alternative embodiment, the client JPEG2024167887000067.jpg55 actively sends a request to the server to participate in graph neural network training. The server can receive the Choose whether or not you agree to participate in JPEG2024167887000068.jpg55, and the client If the client agrees to participate in JPEG2024167887000069.jpg55 JPEG2024167887000070.jpg55 establishes a distributed collaborative time series graph neural network model. In the present invention, a distributed and linked time series graph neural network model corresponding to a client is The local parameters of the network are expressed as the local feature change matrix , including local weights and local losses, and a distributed, coordinated time-series graph neural network The global parameters of the network model (global network parameters) are at least These include global weights, global feature transformation matrix and global loss. About distributed and coordinated time series graph neural network model, The global loss function of the model at time 515 is JPEG2024167887000072.jpg879, and the local loss function is It is represented as JPEG2024167887000073.jpg879, JPEG2024167887000074.jpg43 The global loss function of the model is JPEG2024167887000075.jpg751, and the local loss function is It may be represented as JPEG2024167887000076.jpg751. JPEG2024167887000077.jpg515 The global loss of the distributed collaborative time series graph neural network model is JPEG2024167887000078.jpg826 and the local loss is JPEG2024167887000079.jpg826, JPEG2024167887000080.jpg43 The global loss of the model is JPEG2024167887000081.jpg712 and the local loss function is The image is JPEG2024167887000082.jpg712. As shown in Figure 2, node A, node B, node C, node D, and node E are clusters. Client A, Client B, Client C, Client D, Client E And client A's JPEG2024167887000083.jpg515 The local network parameters at time 515 are at least the dashed box next to client A. Local feature transformation matrix in JPEG2024167887000084.jpg830, local weights JPEG2024167887000085.jpg831 and local loss JPEG2024167887000086.jpg837, and in the present invention, local loss JPEG2024167887000087.jpg837 is the loss of the distributed time series graph neural network model corresponding to client A. The loss is calculated by the loss function of the model, and the JPEG2024167887000088.jpg43 The local network parameters at time 43 are at least the dashed box next to client A. Local feature transformation matrix in JPEG2024167887000089.jpg717, local weights JPEG2024167887000090.jpg717 and local loss JPEG2024167887000091.jpg723, and correspondingly, each of the other nodes such as node B, node C, node D, node E, etc. are the local network parameters at the corresponding time (Fig. (not shown). As of JPEG2024167887000092.jpg515, the global network parameters in the server are at least Feature transformation matrix JPEG2024167887000093.jpg734, global weights JPEG2024167887000094.jpg831 and global loss Contains JPEG2024167887000095.jpg826, JPEG2024167887000096.jpg43 Global feature transformation matrix in the server JPEG2024167887000097.jpg720, global weights JPEG2024167887000098.jpg717 and global loss JPEG2024167887000099.jpg712, where JPEG2024167887000100.jpg43 indicates a point in time, time or timing. Referring to FIG. 2, node B is a server, node Node D can communicate with only three servers: node A and node C. Node D can communicate with only three servers: node A, node C, and node C. Node E can communicate with only four nodes, and similarly, other nodes are described in the present specification and The explanation will be omitted below. S102, Data for training distributed, time-series graph neural network models Get the. The server is the client If JPEG2024167887000101.jpg55 agrees to participate in training the distributed federated time series graph neural network, Server is a client The consent token is sent to JPEG2024167887000102.jpg55, and the consent token is sent to the client. JPEG2024167887000103.jpg55 is used to participate in training the graph neural network. Now, the client JPEG2024167887000104.jpg55 is modeled by adjacent clients with related data, i.e., multi-value Invite them to train a distributed, collaborative time series graph neural network model on the chain. , the list of neighbors willing to participate JPEG2024167887000105.jpg713, where the neighbor list is Exclude JPEG2024167887000106.jpg55 itself, JPEG2024167887000107.jpg53 is the client's serial number, JPEG2024167887000108.jpg53 is a natural number, or JPEG2024167887000109.jpg53 is a positive integer. Each client stores the node information of the node where the client is located and the Between the node where the client is located and the node where the client's neighboring clients are located The edge information can be obtained, and thus the distributed time series graph data of the client can be It should be understood that the distributed time series graph data of each client is formed. The data includes the node information of the node where each client is located and the The information on the edge between the node and the node where the adjacent client of each client is located is include. Optionally, the node information of the node where the client is located in the distributed time series graph data is The node identification information may include other information, for example the node identification information may include the serial number of the node. In different application scenarios, the specific content of the node information may be The node where the client is located and the neighboring nodes of the client may be different. The information about the edges between the nodes where the ants are located may include edge identification information or other information. For example, the edge identification information may be the name or number of the edge. In different application scenarios, the specific content of the edge information may be different. Logically aggregate the distributed time series graph data of all clients to generate the entire distributed time series graph. Rough Data JPEG2024167887000110.jpg725 can be formed, where: JPEG2024167887000111.jpg640 indicates a point in time, time or timing, JPEG2024167887000112.jpg55 shows a set of nodes, and each node JPEG2024167887000113.jpg56 represents one client, i.e. one company, JPEG2024167887000114.jpg67 shows a set of edges at time t, and each edge JPEG2024167887000115.jpg67 is one label vector JPEG2024167887000116.jpg89 (where the edge label vector is also called the edge label), JPEG2024167887000117.jpg89 is a node JPEG2024167887000118.jpg56 and This indicates whether there is a cooperative relationship, demand flow, business flow, and value flow between JPEG2024167887000119.jpg65. In addition, the entire above JPEG2024167887000120.jpg725 is a logical distributed time series graph data, and the server is the complete distributed time series graph data JPEG2024167887000121.jpg725 is not recognized by the server, JPEG2024167887000122.jpg55 can be obtained, but the set of edges JPEG2024167887000123.jpg67, edge label vector Unable to obtain JPEG2024167887000124.jpg89. Based on the above, in the present invention, a distributed and linked time series graph neural network model is The feature learning network in the model is based on the distributed time series graph data of each client at each historical point. Based on the data of each client and the feature vectors of the nodes where the neighboring clients of each client are located, Based on this, it is used to output the feature vector of the node where each client is located. The label prediction network in the distributed time series graph neural network model is Based on the feature vector of the node where the client is located, Outputs the predicted edge of the node where each client's neighbors are located and the node where each client's neighbors are located. is used to. The functions of each client and server are as follows: Each client JPEG2024167887000125.jpg55 is a graph The client in JPEG2024167887000126.jpg54 Node corresponding to JPEG2024167887000127.jpg55 Related calculations for JPEG2024167887000128.jpg56, used to process JPEG2024167887000129.jpg617, the associated calculations including: 1) Node Calculate the feature vector of JPEG2024167887000130.jpg56 and JPEG2024167887000131.jpg56 and its adjacent nodes Weight of connected edges of JPEG2024167887000132.jpg65 Predict JPEG2024167887000133.jpg69, where JPEG2024167887000134.jpg53 is the node JPEG2024167887000135.jpg56, JPEG2024167887000136.jpg73 is the adjacent node Indicates the serial number of JPEG2024167887000137.jpg65. 2) Send and / or receive feature vectors of nodes, e.g., a client JPEG2024167887000138.jpg55 adjacent client to client Send the feature vector of the node where JPEG2024167887000139.jpg55 is located, or send it to its neighbors from its neighboring clients. The feature vector of the node where the adjacent client is located is received. 3) Perform backpropagation based on the local loss (backpropagation is a prior art in this field). be). client JPEG2024167887000140.jpg55 is the node where the neighboring client is located. JPEG2024167887000141.jpg65 can be obtained and the client The edge between the node where JPEG2024167887000142.jpg55 is located and the node where its adjacent client is located JPEG2024167887000143.jpg728, and edge labels JPEG2024167887000144.jpg917 can be obtained. JPEG2024167887000145.jpg56 and adjacent nodes The edge of JPEG2024167887000146.jpg65 is JPEG2024167887000147.jpg43 weight JPEG2024167887000148.jpg69 is the edge label The image is JPEG2024167887000149.jpg917. The server function collects the losses of each client and calculates a global loss (i.e. Each client calculates the local loss of the distributed time series graph neural network model. The Adam algorithm is used to update the global parameters and A distributed, collaborative time-series graph neural network that starts backpropagation on all clients. As shown in Figure 2, the server is responsible for coordinating the entire training process of the network model. It may be a third-party server, and the server may be one or more. The associated functionality may be implemented by a distributed server cluster. In one embodiment, the neural network may reside on a server. In one embodiment, the neural network on the server is connected to the neural network of the client. Unlike the network, the neural network on the server is preferably a recurrent neural network. Neural Network (RNN), LSTM (Long Short Term Memory ory, i.e. long short-term memory networks, a special kind of recurrent neural network Graph Neural Network (GNN), Graph Attention (GAT) or Graph Any neural network such as Graph Transformer In some embodiments, the sub- When the server generates the current parameters (global network parameters), The parameter data in the server at the historical point in time is taken into account, and the generated Broadcast the parameters to all clients or to selected clients. can be done. In another embodiment, the neural network on the server is connected to the neural network of the client. It is part of a distributed neural network, similar to a distributed network, and the server is It corresponds to one node in the network. In another embodiment, the neural network does not reside on a server and all the queries are processed at any one time. It simply processes the data uploaded by the client. For security purposes, third-party servers store We do not know the global topology structure information of JPEG2024167887000150.jpg54, i.e., the nodes of the graph JPEG2024167887000151.jpg55, but the edges of the graph Since JPEG2024167887000152.jpg67 cannot be known, privacy protection for each node is achieved. S103, each client interacts with the server to create a distributed time series graph neural network. Train the network model. (1) The client updates the feature vector of the node based on the feature learning network. The process in which the client updates the feature vectors of the nodes is a distributed, coordinated time-series graph neural network. For each client participating in the network model training, Ianto is at a historical point (e.g. 0~ JPEG2024167887000153.jpg517) Based on the feature vector of the node where the client is located, generating a feature vector for each node, where the feature vector for each node is Between the node where each client is located and the node where each client's neighboring clients are located is used to characterize the edge information of First, based on the feature learning network, we use the graph snapshots at each point in time to We perform node feature learning using the graph snapshots at each point in time. We first use the graph data, and then use the multi-temporal graph data to learn the temporal evolution characteristics of node features. The feature vectors of the nodes obtained by training are used to predict the edges of the graph. Note that the feature learning network includes multiple layers (or feature learning layers), e.g. JPEG2024167887000154.jpg contains 54 layers, where the first JPEG2024167887000155.jpg522 layer is the distributed time series graph data local to the client (i.e. Generate a feature vector of the node where the client is located based on the original data. The remaining layers are used to JPEG2024167887000156.jpg is used to update the feature vector obtained from the 523 layer. Specifically, we consider multiple clusters participating in distributed time series graph neural network model training. Client serial number JPEG2024167887000157.jpg53's client (Client i or Client JPEG2024167887000158.jpg714 is For example, take JPEG2024167887000159.jpg527. JPEG2024167887000160.jpg53 is an arbitrary client among multiple clients. Client JPEG2024167887000161.jpg53 is also Adjacent clients for JPEG2024167887000162.jpg53 JPEG2024167887000163.jpg73 from adjacent clients Receive the feature vector generated by JPEG2024167887000164.jpg73, e.g. JPEG2024167887000165.jpg639 is sent to the client via the Internet. Send a request to JPEG2024167887000166.jpg73 Get the feature vector of the node where JPEG2024167887000167.jpg740 is located, JPEG2024167887000168.jpg639 is sent to all neighboring clients Receive the feature vector of the node where JPEG2024167887000169.jpg73 is located. For example, during the process of training a distributed time series graph neural network model, JPEG2024167887000170.jpgIn 44 rounds of iterations, JPEG2024167887000171.jpg43 The first distributed collaborative time series graph neural network JPEG2024167887000172.jpg53 layers, client JPEG2024167887000173.jpg53 is the client A kind of update rule for updating the feature vector of the node where JPEG2024167887000174.jpg53 is located is (i.e., the client nt The feature vector of the node where JPEG2024167887000175.jpg53 is located is: JPEG2024167887000176.jpg13162, where JPEG2024167887000177.jpg814 is JPEG2024167887000178.jpgClient at time 43 The first step of the feature learning network corresponding to JPEG2024167887000179.jpg53 JPEG2024167887000180.jpg53 Layer output from client JPEG2024167887000181.jpg53 indicates the feature vector of the node where it is located, JPEG2024167887000182.jpg1225 is JPEG2024167887000183.jpg43 Client Adjacent clients for JPEG2024167887000184.jpg53 The first step of the feature learning network corresponding to JPEG2024167887000185.jpg73 JPEG2024167887000186.jpg515 adjacent clients output from layer The feature vector of the node where JPEG2024167887000187.jpg73 is located is shown ( JPEG2024167887000188.jpg1225 is the client JPEG2024167887000189.jpg53 is the client JPEG2024167887000190.jpg73), JPEG2024167887000191.jpg1025 is JPEG2024167887000192.jpgClient at time 43 The first step of the feature learning network corresponding to JPEG2024167887000193.jpg53 JPEG2024167887000194.jpg515 layers output from the client JPEG2024167887000195.jpg53 indicates the feature vector of the node where it is located, JPEG2024167887000196.jpg631 shows an aggregation operation, JPEG2024167887000197.jpg519 shows linear operations, JPEG2024167887000198.jpg717 shows the local feature transformation matrix. Selectable, JPEG2024167887000199.jpg631 Aggregation methods include summation, averaging, maximization, etc., or any combination of these. . Client i: JPEG2024167887000200.jpg43 global feature transformation matrix Send a request to the server to get JPEG2024167887000201.jpg717, where the server: Global feature transformation matrix at time 858 of JPEG2024167887000202.jpg JPEG2024167887000203.jpg717, i.e., the timing effect of multi-temporal graph neural network learning by weight matrix Calculate. Using JPEG2024167887000204.jpg717 Linearly convert JPEG2024167887000205.jpg814 and send it to the client The node where JPEG2024167887000206.jpg53 is located JPEG2024167887000207.jpgFinal feature vector in 53 layers Get JPEG2024167887000208.jpg814. By repeating the above steps, the feature vectors of the nodes in each layer of the feature learning network are Calculate. JPEG2024167887000209.jpg43: The first distributed collaborative time series graph neural network JPEG2024167887000210.jpg53 layers, client JPEG2024167887000211.jpg53 is the client Another update rule for updating the feature vector of the node where JPEG2024167887000212.jpg53 is located is (i.e., the client The feature vector of the node where JPEG2024167887000213.jpg53 is located is: JPEG2024167887000214.jpg13170, where JPEG2024167887000215.jpg814 is JPEG2024167887000216.jpgClient at time 43 The first step of the feature learning network corresponding to JPEG2024167887000217.jpg53 JPEG2024167887000218.jpg53 Layer output from client JPEG2024167887000219.jpg53 indicates the feature vector of the node where it is located, JPEG2024167887000220.jpg1225 is JPEG2024167887000221.jpg43 Client Adjacent clients for JPEG2024167887000222.jpg53 The first step of the feature learning network corresponding to JPEG2024167887000223.jpg73 JPEG2024167887000224.jpg515 indicates the feature vector of the node where the neighboring client j is located, output from the layer, JPEG2024167887000225.jpg1025 is JPEG2024167887000226.jpgClient at time 43 The first step of the feature learning network corresponding to JPEG2024167887000227.jpg53 JPEG2024167887000228.jpg515 layers output from the client JPEG2024167887000229.jpg53 indicates the feature vector of the node where it is located, JPEG2024167887000230.jpg631 shows an aggregation operation, JPEG2024167887000231.jpg519 shows linear operations, JPEG2024167887000232.jpg717 indicates the local feature transformation matrix, JPEG2024167887000233.jpg68 is the client The node and client where JPEG2024167887000234.jpg53 is located Adjacent clients for JPEG2024167887000235.jpg53 This shows the correlation coefficient between the nodes where JPEG2024167887000236.jpg73 is located. In addition, during the training process of the distributed linked time series graph neural network, It performs message passing and backpropagation using the graph, which is communication intensive and time consuming. Nodes in a graph have a degree, which indicates the number of edges connected to a node. However, in some cases, there is a large difference between the degrees of the nodes in the graph, e.g., scale-free The degrees of network nodes have an exponential distribution, with a small number of nodes having very large degrees and most The degree of each node is very small. The number of messages sent / received by a node is small compared to the degree of the node. For example, a node with a high degree is the overall distributed time series graph neural network model. This is a performance bottleneck. To address this issue, the client When aggregating the feature vectors of the nodes where the ant is located, the client JPEG2024167887000237.jpg53 is each adjacent node Correlation coefficient of JPEG2024167887000238.jpg825 We train JPEG2024167887000239.jpg68, and when aggregating feature vectors, we combine correlation coefficients to perform aggregation. In one embodiment, the final global loss function By regularizing JPEG2024167887000240.jpg717, the correlation coefficients of most neighboring nodes become 0, and a small number of neighboring The absolute value of the correlation coefficient of a node tends to be much larger than zero. Preferably, the regularization method is JPEG2024167887000241.jpg76 is regularized. JPEG2024167887000242.jpg1843 or JPEG2024167887000243.jpg812 regularization. JPEG2024167887000244.jpg2253By regularizing, the correlation coefficients of most neighboring nodes become 0, and the correlation coefficients of a few neighboring nodes become The absolute value of is much larger than 0, and thus, the feature vectors of adjacent nodes are aggregated. When a client selects a neighbor, the client can select all neighbors. Instead of aggregating the feature vectors of each node, we use the feature vectors of some of the most relevant neighboring nodes. The vector is then aggregated, i.e., the set of adjacent nodes. JPEG2024167887000245.jpg713 most relevant JPEG2024167887000246.jpg Update the 54 neighbors, keep a few important neighbors, and remove unimportant neighbors It is possible to perform forward and backward propagation during the training process of a distributed linked time series graph neural network. This can reduce the number of communication messages generated by the do. Optionally, in this embodiment, each client in the plurality of clients selects a neighboring node. Or some of the clients can select neighboring nodes. However, the present invention is not limited to the embodiments. In one embodiment, the server records the response time of each client until the forward propagation is completed, and Sort the clients in descending order according to response time and select the one with the longest response time. JPEG2024167887000247.jpg 54 nodes were selected for regularization, with the longest response time JPEG2024167887000248.jpg54 clients may be notified to reduce the number of neighbors. (2) The client is a local network of the distributed time series graph neural network model. Generate network parameters. Each client selects each cluster based on the feature vector of the node where the client is located. Local network of distributed time series graph neural network model of clients Generate parameters and send the local network parameters to the server. Each client selects a node based on the feature vector of the node where the client is located. A local network of distributed and coordinated time series graph neural network models for clients The specific process of generating the network parameters is as follows: The feature vector of the node is input to the label prediction network corresponding to each client, and each The node where the client is located and the nodes where the neighboring clients of each client are located Each client generates a prediction of the edge between the nodes where the client is located and each The predicted value of the edge between the nodes where the neighboring clients of the client are located and the edge Based on the true value of the metric, the local Then, each client performs backpropagation based on its local loss. Local network of distributed time series graph neural network model of clients This includes generating parameters. The above clients For example, if we take JPEG2024167887000249.jpg53, the client JPEG2024167887000250.jpg53 is a client-side prediction network using a fully connected layer. The node and client where JPEG2024167887000251.jpg53 is located Adjacent clients for JPEG2024167887000252.jpg53 Edge between the nodes where JPEG2024167887000253.jpg73 is located Calculate the predicted value of the label of JPEG2024167887000254.jpg67 (i.e. edge (We predict the label vector of JPEG2024167887000255.jpg67). Specifically, we use the client output from the feature learning network. Ant The feature vector of the node where JPEG2024167887000256.jpg53 is located is input to the label prediction network, and the client The node and client where JPEG2024167887000257.jpg53 is located Adjacent clients for JPEG2024167887000258.jpg53 Output the predicted value of the edge of the node where JPEG2024167887000259.jpg73 is located. Client JPEG2024167887000260.jpg53 is the global weight of the edge prediction fully connected layer from the server. Take JPEG2024167887000261.jpg719 and global weights Edge based on JPEG2024167887000262.jpg719 Calculate the predicted value of the label of JPEG2024167887000263.jpg67, preferably the edge The predicted value of the label of JPEG2024167887000264.jpg67 is the label vector JPEG2024167887000265.jpg920 and label vector JPEG2024167887000266.jpg920 is determined by the following formula: JPEG2024167887000267.jpg989, where JPEG2024167887000268.jpg941 is the client The node and client where JPEG2024167887000269.jpg53 is located This is a distance function that indicates the distance between the nodes where JPEG2024167887000270.jpg73 is located. The distance function can be selected from cosine distance, Alternatively, the Karlian distance or the Mahalanobis distance may be employed. Furthermore, client i computes the local loss function (also called the local edge loss function) The distributed and linked time series graph neural network model corresponding to client i is Determine the local loss (also called the local edge loss) of the delta and use this local loss function teeth: JPEG2024167887000271.jpg1583, where: JPEG2024167887000272.jpg920 shows the local loss, JPEG2024167887000273.jpg917 is edge shows the true value of the label for JPEG2024167887000274.jpg67, JPEG2024167887000275.jpg920 is edge Shows the predicted label value for JPEG2024167887000276.jpg67. Finally, the client JPEG2024167887000277.jpg53 is the local network parameters and local loss Send JPEG2024167887000278.jpg920 to the server. (3) The server updates the global network parameters. The server determines the number of clients based on the local network parameters sent by each client. ,Global network parameters of distributed time series graph neural network model and broadcasts global network parameters to each client. do. The server Global weight matrix at time 515 of JPEG2024167887000279.jpg Global weight matrix at time 765 of JPEG2024167887000280.jpg It is used to compute JPEG2024167887000281.jpg719, selectable using RNN, GRU, LSTM or Transf ormer to obtain the global weight matrix Optionally, the server may compute the image data via the internet. JPEG2024167887000283.jpg719 is broadcast to all clients, and the clients are able to use the global network. This reduces the waiting time for receiving the First, the server calculates the distributed federation loss based on the local loss sent by each client. A global loss for the sequence graph neural network model is determined, and the global loss is is the sum of the local loss and regularization term for each client, i.e., JPEG2024167887000284.jpg1878 where, JPEG2024167887000285.jpg9115 is the sum of the local losses of each client, JPEG2024167887000286.jpg725 is the regularization term. The server then checks the local network parameters and Based on the global loss and the backpropagation, a distributed time series graph neural network is developed. The server obtains the global network parameters of the network model. Using the dam algorithm to estimate the parameters of a distributed time series graph neural network model meter is optimized, the server updates global network parameters: JPEG2024167887000287.jpg976, JPEG2024167887000288.jpg719 and JPEG2024167887000289.jpg721 shows the number of global network parameters. In one embodiment, the server broadcasts global network parameters to the client. broadcast to initiate backpropagation to all clients. In another embodiment, the server may broadcast global network parameters to the client. After the broadcast, the client decides whether to start backpropagation. The update of the local network parameters and the correlation coefficient matrix Responsible for updating JPEG2024167887000290.jpg54. In an embodiment of the present invention, each client receives the global Using global network parameters, we perform distributed collaboration of time series graph neural networks Update the network parameters of the network model. Optionally, in an embodiment of the present invention, one blockchain is established and a client-client Feature vectors, parameters and gradients of the client and server-client nodes Such communication traffic is stored on the blockchain to prevent repudiation and ensure the integrity of the server and client. The system may also support auditing of clients for privacy violations. S104, the client predicts the labels of the edges. In the embodiment of the present invention, the distributed and linked time series graph neural network is constructed by the above steps S101 to S103. After completing the training of the network model, each client The distributed time series graph neural network model after updating the distributed time series graph data of the Enter the node on which each client is located and the neighboring clients of each client into the The function outputs the predicted value of the edge between the nodes where the target is located. In one embodiment, Edge at time 54 of JPEG2024167887000291.jpg To predict the label of JPEG2024167887000292.jpg67, we use 0~ JPEG2024167887000293.jpg516 Distributed time series graph data Using JPEG2024167887000294.jpg1798, the distributed and linked time series graph neural network model of S101 to S103 above is We trained a distributed time series graph neural network model JPEG2024167887000295.jpgDistribution time series graph data at time 54, i.e. Inform the client that JPEG2024167887000296.jpg837 will be input. JPEG2024167887000297.jpg837 is the edge to be predicted from the distributed time series graph data of the client. JPEG2024167887000298.jpg67 information is removed, and finally the edge This represents outputting the predicted value of the label for JPEG2024167887000299.jpg67. Similarly, Edge at time 516 of JPEG2024167887000300.jpg To predict the label of JPEG2024167887000301.jpg67, we use 0~ JPEG2024167887000302.jpg54 Distribution graph data JPEG2024167887000303.jpg1786 is used to train a distributed collaborative time series graph neural network model. JPEG2024167887000304.jpg516, which is the distributed time series graph data. Inform the client that JPEG2024167887000305.jpg1761 will be input, and finally the edge Output the predicted value of the label of JPEG2024167887000306.jpg67. In summary, in the method for predicting the evolution of a multi-value chain provided by the present invention, Each client participating in the distributed collaborative time series graph neural network model training Obtain only the information of edges between the node where itself and its neighboring clients are located. The server can obtain only the information of the node where all clients are located. and the client can access the local original data (i.e. the client's local The distributed time series graph data generated by the distributed time series graph is the node information and the end If the client does not share the same address (including the client's address), each client must interact with the server to A globally distributed time series graph neural network model applicable to all We train the model and then identify the nodes where each client in the multi-value chain is located and their Dynamically predict edges between nodes where neighboring clients are located. In the embodiment of the present invention, a distributed and coordinated time series graph neural network model is used. We obtained the prediction results of the edges between the nodes where each client is located in the multi-value chain. Then, the state of each agent and the relationships between agents are dynamically predicted based on the edge prediction results. This can be done. In one embodiment, the method predicts future value flows between companies based on the edge prediction results. Dynamically predicting network traffic such as demand flow, business flow, and technology flow For example, it is possible to predict the sales volume of a product. When forecasting a company's sales in a system, accurate forecast results cannot be obtained using only the company's internal data. It is not possible to obtain comprehensive data on related companies upstream and downstream of the company’s value chain. For example, clients such as retailers, distributors, and manufacturing plants need to utilize local Data (sales volume data) is distributed and linked to a time series graph neural network model The input is the feature vector of the node, which is predicted through the feature learning network, and the label prediction network is used. The network is used to obtain the forecast of the product sales volume (i.e., the forecast of the edge of the model prediction). is the forecast value of the product sales volume). As can be seen from this, A distributed time series graph neural network model based on privacy-sensitive data Establish and forecast product sales volume, and target company's upstream and downstream related companies in the value chain By using data from various industries for distributed training, we can achieve more accurate sales volume forecasts. do. In another embodiment, the method determines the relationship between the entities and the Dynamically predict the state of each actor, i.e., the evolution of the enterprise nodes on the multi-value chain It is also possible to predict the state of any node. For JPEG2024167887000307.jpg56, The predicted value of an edge connected to JPEG2024167887000308.jpg56 is JPEG2024167887000309.jpg54, or JPEG2024167887000310.jpg54, the node This means that there is no cooperative relationship between JPEG2024167887000311.jpg56 and the other end node of the edge. For example, For JPEG2024167887000312.jpg56, The predicted values ​​of all edges connected to JPEG2024167887000313.jpg56 are all JPEG2024167887000314.jpg54, or It is predicted that there will be a continuous trend towards JPEG2024167887000315.jpg54. The evolutionary state of JPEG2024167887000316.jpg724 is "extinct", i.e., the corresponding client The evolutionary state of JPEG2024167887000317.jpg55 is "extinction".

Claims

1. Each client participating in the distributed time series graph neural network model training For each client, the client Distributed time series graph data and the nodes where adjacent clients of each client are located Based on the feature vector of the node, the feature vector of the node where each client is located is calculated. The feature vector is generated based on the node where each of the clients is located and the is used to characterize the edge information between the nodes where the neighboring clients of The distributed time series graph data of each client is Node information and the node where each client is located and the neighboring nodes of each client including information on edges between the nodes where the clients are located; Each of the clients is assigned a feature vector based on the feature vector of the node in which the client is located. , the local distributed and linked time series graph neural network model of each client Generate network parameters and send the local network parameters to a server. To believe, The server determines the local network parameters sent by each client. Based on this, the global network of the distributed linked time series graph neural network model and transmitting the global network parameters to each of the clients. broadcast to the Each of the clients uses the global network parameters to The network parameters of the distributed time series graph neural network model of the Update the distributed time series graph data of each client at the current time point, and update the distributed collaboration Input the time series graph into the neural network model, and the node where each client is located is The predicted edge between the nodes where the neighboring clients of each client are located and the and outputting the above-mentioned.

2. The local network parameters include at least local weights, local feature transformation rows, Includes columns and local losses, The global network parameters include at least global weights, global features, The method of claim 1 , comprising a transformation matrix and a global loss.

3. The distributed and linked time series graph neural network model is a feature learning network and Includes a label prediction network, Here, the feature learning network is a distributed time series graph of each of the clients at a historical point in time. The rough data and the feature vector of the node where the neighboring clients of each client are located are to output a feature vector of the node where each of the clients is located based on the vector. The label prediction network is used to predict the location of the node where each of the clients is located. Based on the feature vector, the node where each client is located and the neighbors of each client are It is used to output the predicted value of the edge of the node where the adjacent client is located, The feature learning network is composed of a message passing layer, and the label prediction network 3. The method of claim 1, wherein:

4. Each of the above clients is a client If so, the client The feature vector of the node where is located is where teeth The client at the time The first feature learning network corresponding to Denote the feature vector of the node where the client i is located, output from the layer, teeth The client at the time Neighboring clients of The first feature learning network corresponding to The adjacent clients output from the layer denotes the feature vector of the node in which teeth The client at the time The first feature learning network corresponding to The client output from the layer denotes the feature vector of the node in which indicates an aggregation operation, denotes linear operation, 4. The method of claim 3, wherein: denotes a local feature transformation matrix.

5. Each of the above clients The feature vector of the node where is located is where: teeth The client at the time The first feature learning network corresponding to The client output from the layer denotes the feature vector of the node in which teeth The client at the time Neighboring clients of The first feature learning network corresponding to The adjacent clients output from the layer denotes the feature vector of the node in which teeth The client at the time The first feature learning network corresponding to The client output from the layer denotes the feature vector of the node in which indicates an aggregation operation, denotes linear operation, denotes the local feature transformation matrix, The client The node where the client is located and Neighboring clients of 4. The method of claim 3, wherein: r indicates a correlation coefficient between the nodes in which r is located.

6. Each of the clients is assigned a feature vector based on the feature vector of the node in which the client is located. , the local distributed and linked time series graph neural network model of each client Generating network parameters includes: Each of the clients transmits the feature vector of the node in which the client is located to each of the clients. The label prediction network corresponding to each client is input to the label prediction network. Prediction of edges between the node and the nodes where the neighboring clients of each of the clients are located. Generating values; Each of the clients is connected to a node where the client is located and a neighbor of the client. Based on the predicted value of the edge between the nodes where the adjacent clients are located and the true value of the edge, Based on this, the local loss of the distributed time series graph neural network model is determined. To do, Each of the clients performs backpropagation based on the local loss, Local network parameters of distributed time series graph neural network model 4. The method of claim 3, further comprising: generating data.

7. The method comprises: Each of the clients provides the server with the distributed, coordinated time-series graph neural network. Sending the local loss of the model, The server determines the local loss based on the local loss sent from each client. determining a global loss for the sequence graph neural network model. fruit, The server determines whether or not a particular network is available based on the local network parameters sent by each client. Based on this, the global network of the distributed and linked time series graph neural network model is To obtain the search parameters, The server receives the local network parameters and and performing backpropagation based on the global loss. obtaining global network parameters for the network model.

3. The method according to claim 1 or 2, characterized in that

8. The server and a plurality of clients are connected to each other via a distributed time series graph. The server is connected to the plurality of clients.

8. The method according to claim 7, The distributed and linked time series graph neural network model is trained, and the distributed and linked time series graph is Based on the rough neural network model, each client in the multi-value chain Between the node where each client is located and the node where the adjacent client of each client is located A multi-value chain evolution prediction system, characterized by predicting edges of

9. A computer program instruction is stored, the computer program instruction being a multimedia program. When executed by a multi-core processor, the multi-core processor is A computer-readable storage medium for carrying out a method.

Citation Information

Patent Citations

  • Graph neural network federal recommendation method for privacy protection

    CN113420232A

  • Value chain analysis device and method

    JP2015194863A

  • Value chain rating system, method and program

    JP6748250B1

  • Semantic learning in federated learning system

    WO2022057694A1

  • Robot fleet management and additive manufacturing for value chain networks

    WO2022133330A1