An information tracing method and device, a storage medium and an electronic device

Abstract

The specification discloses an information tracing method and device, a storage medium and an electronic device. Embodiments of the specification perform information simulation propagation on a graph network to obtain state information of each node in the propagated graph network. The state information and the basic network structure information of the graph network are input into a graph neural network model to be trained to predict a starting node of information propagation in the propagated graph network through the graph neural network model. The graph neural network model is trained with minimization of a difference between the predicted starting node and an actual starting node as an optimization target. In this method, the propagation source is predicted without relying on a propagation coefficient of information propagation, but through supervised training of the graph neural network model, the graph neural network model learns a corresponding relationship between the propagated graph network and the propagation source, so as to predict the propagation source in the graph network to be traced based on the trained graph neural network model, thereby improving the accuracy of the predicted propagation source.

Description

Technical Field

[0001] This specification relates to the field of information traceability technology, and in particular to an information traceability method, apparatus, storage medium and electronic device. Background Technology

[0002] Currently, information tracing has applications in various scenarios, such as tracing the source of rumors on social networks, locating the source of infectious diseases, and tracing the source of online viruses. In information tracing, propagation source prediction can be based on graph networks. A graph network (GN) is a set of functions organized in a graph structure within a topological space for relational reasoning.

[0003] In existing technologies, a node can be randomly selected multiple times from a graph network as a simulated propagation source. Using pre-set propagation coefficients in the propagation dynamics model, information propagation is simulated starting from this simulated source within the graph network. This process determines the node whose state most closely approximates the actual state, which is then used as the final predicted propagation source. The propagation coefficients can include factors such as infection rate, recovery rate, and propagation time. The node states can include infected, recovered, and uninfected states.

[0004] However, existing methods for predicting propagation sources rely on the values ​​of propagation coefficients in propagation dynamics models. In reality, the propagation coefficients are unknown during information propagation. Thus, assuming propagation coefficients may lead to inaccurate predictions of propagation sources. Summary of the Invention

[0005] This specification provides an information tracing method, apparatus, storage medium, and electronic device to partially solve the problems existing in the prior art.

[0006] The embodiments in this specification adopt the following technical solutions:

[0007] This specification provides an information tracing method, including:

[0008] Obtain a pre-built graph network;

[0009] A simulation instruction is generated for the graph network, and an information simulation propagation task for the graph network is executed through the simulation instruction to obtain the propagated graph network and the state information of each node in the propagated graph network.

[0010] The state information of each node and the basic network structure information of the graph network are input into the graph neural network model to be trained. Based on the state information of each node, the graph neural network model determines the initial features corresponding to each node. Based on the initial features corresponding to each node, the correlation between any two nodes is determined. Based on the correlation between any two nodes, the basic network structure information is updated to obtain updated information. Based on the updated information, the initial features corresponding to each node are processed to obtain processed features corresponding to each node. Based on the processed features corresponding to each node, the overall features of the propagated graph network are determined. Based on the overall features and the processed features corresponding to each node, the starting node of information propagation in the propagated graph network is predicted.

[0011] The optimization objective is to minimize the difference between the predicted starting node and the actual starting node. The graph neural network model is then trained to obtain the trained model.

[0012] After obtaining the trained model, an information tracing request is received. Based on the information tracing request, the state information of each target node in the target graph network to be traced and the basic network structure information of the target graph network are input into the trained model so as to predict the starting target node of information propagation in the target graph network through the trained model, and to execute the task based on the starting target node.

[0013] Optionally, before obtaining the pre-built graph network, the method further includes:

[0014] Obtain the target object group and determine the direct relationships between each target object in the target object group;

[0015] A graph network is constructed with each target object as a node and the direct relationships between target objects as edges.

[0016] Optionally, the simulation instructions are used to execute an information simulation propagation task for the graph network, obtaining the propagated graph network and the state information of each node in the propagated graph network, specifically including:

[0017] According to the simulation instructions, a node is randomly selected from the pre-constructed graph network to be determined as the starting node for information propagation;

[0018] The determined starting node and the pre-constructed graph network are input into the propagation dynamics model. Based on the preset propagation coefficients, the propagation dynamics model propagates information from the determined starting node into the graph network, resulting in the propagated graph network and the state information of each node in the propagated graph network.

[0019] Optionally, the basic network structure information includes: an adjacency matrix or a Laplace matrix.

[0020] Optionally, the graph neural network model includes: an input layer, an attention layer, a graph convolutional layer, a graph representation layer, and an output layer;

[0021] The state information of each node and the basic network structure information of the graph network are input into the graph neural network model. Based on the state information of each node, the graph neural network model determines the initial features corresponding to each node. Based on the initial features of each node, the correlation between any two nodes is determined. Based on the correlation between any two nodes, the basic network structure information is updated to obtain updated information. Based on the updated information, the initial features corresponding to each node are processed to obtain processed features corresponding to each node. Based on the processed features of each node, the overall features of the propagated graph network are determined. Based on the overall features and the processed features of each node, the starting node for information propagation in the propagated graph network is predicted. Specifically, this includes:

[0022] The state information of each node and the basic network structure information of the graph network are input into the graph neural network model to be trained, so that the state information of each node is encoded through the input layer to obtain the initial features corresponding to each node.

[0023] The initial features corresponding to each node are input into the attention layer, so that the correlation between any two nodes can be determined by the attention layer based on the initial features corresponding to each node.

[0024] The correlation between any two nodes and the basic network structure information are input into the graph convolutional layer. The graph convolutional layer updates the basic network structure information based on the correlation between any two nodes to obtain updated information. Based on the updated information, the initial features corresponding to each node are processed to obtain the processed features corresponding to each node.

[0025] The processed features corresponding to each node output by the graph convolutional layer are input into the graph representation layer, so that the overall features of the propagated graph network can be determined by the graph representation layer based on the processed features corresponding to each node.

[0026] The overall features and the processed features corresponding to each node are input into the output layer, so that the starting node of information propagation in the propagation graph network can be predicted by the output layer based on the overall features and the processed features corresponding to each node.

[0027] Optionally, the basic network structure information is updated based on the correlation between any two nodes to obtain updated information, specifically including:

[0028] Based on the correlation between any two nodes, determine the correlation matrix that includes the correlation between any two nodes;

[0029] The correlation matrix is ​​multiplied by the basic network structure information to obtain the dot product result, which is then used as the updated information after updating the basic network structure information.

[0030] Optionally, based on the updated information, the initial features corresponding to each node are processed to obtain the processed features corresponding to each node, specifically including:

[0031] The initial features corresponding to each node and the updated information are input into the graph convolutional network in the graph convolutional layer. The graph convolutional network performs graph convolution on the initial features corresponding to each node according to the updated information to obtain the graph convolutional features corresponding to each node. The graph convolutional features corresponding to each node are added to the initial features corresponding to each node to obtain the processed features corresponding to each node.

[0032] Optionally, the graph convolutional layer comprises multiple graph convolutional networks;

[0033] The initial features corresponding to each node and the updated information are input into the graph convolutional network in the graph convolutional layer. The graph convolutional network performs graph convolution on the initial features corresponding to each node based on the updated information to obtain the graph convolutional features corresponding to each node. The graph convolutional features corresponding to each node are then added to the initial features corresponding to each node to obtain the processed features corresponding to each node. Specifically, this includes:

[0034] For each graph convolutional network, the processed features corresponding to each node output by the previous graph convolutional network are redefined as the initial features corresponding to each node input to the current graph convolutional network.

[0035] The initial features corresponding to each node of the graph convolutional network and the updated information are input into the graph convolutional network. The graph convolutional network performs graph convolution on the initial features corresponding to each node according to the updated information to obtain the graph convolutional features corresponding to each node in the graph convolutional network. The graph convolutional features corresponding to each node in the graph convolutional network are added to the initial features corresponding to each node input into the graph convolutional network to obtain the processed features corresponding to each node output by the graph convolutional network.

[0036] Optionally, based on the processed features corresponding to each node, the overall features of the propagated graph network are determined, specifically including:

[0037] Based on the processed features corresponding to each node, the importance of each node in the post-propagation graph network is determined;

[0038] For each state, based on the importance of each node in the post-propagation graph network, a specified number of nodes are selected from the nodes in that state as important nodes;

[0039] Pooling is performed on the processed features corresponding to the important nodes to obtain the pooling result under this state;

[0040] The pooling results under each state are weighted and summed to obtain the overall features of the propagated graph network.

[0041] Optionally, based on the overall features and the processed features corresponding to each node, the starting node for information propagation in the post-propagation graph network is predicted, specifically including:

[0042] Based on the processed features corresponding to each node and the overall features, the similarity between each node and the propagated graph network is determined.

[0043] Based on the processed features of each node, the overall features, and the similarity between each node and the post-propagation graph network, the starting node for information propagation in the post-propagation graph network is predicted.

[0044] Optionally, based on the processed features corresponding to each node, the overall features, and the similarity between each node and the post-propagation graph network, the starting node for information propagation in the post-propagation graph network is predicted, specifically including:

[0045] For each node, based on the processed features corresponding to the node, the overall features, and the similarity between the node and the post-propagation graph network, the predicted probability that the node is the starting node for information propagation in the post-propagation graph network is predicted, and this probability is used as the predicted probability for the node.

[0046] The node with the highest predicted probability is selected from the predicted probabilities of each node and used as the starting node for information propagation in the post-propagation graph network.

[0047] Optionally, the graph neural network model is trained with the optimization objective of minimizing the difference between the predicted starting node and the actual starting node, specifically including:

[0048] The predicted probability output by the graph neural network model for each node as the starting node is determined as the predicted probability for each node.

[0049] The predicted probability for each node is normalized to obtain the normalized result for each node.

[0050] The graph neural network model is trained with the optimization objective of minimizing the difference between the predicted starting node and the actual starting node, and minimizing the difference between the normalized result corresponding to each node and the actual normalized result corresponding to each node. The actual normalized result corresponding to a node in a specified state is greater than the actual normalized result corresponding to a node in a non-specified state.

[0051] Optionally, according to the information tracing request, the state information of each target node in the target graph network to be traced and the basic network structure information of the target graph network are input into the trained model, so as to predict the starting target node of information propagation in the target graph network through the trained model, specifically including:

[0052] Based on the information tracing request, determine the target graph network to be traced, and determine the state information of each target node in the target graph network;

[0053] The state information of each target node and the basic network structure information of the target graph network are input into the trained model. The trained model determines the initial features corresponding to each target node based on the state information of each target node, determines the correlation between any two target nodes based on the initial features of each target node, updates the basic network structure information based on the correlation between any two target nodes to obtain updated information, processes the initial features corresponding to each target node based on the updated information to obtain processed features corresponding to each target node, determines the overall features of the target graph network based on the processed features of each target node, and predicts the starting target node for information propagation in the target graph network based on the overall features and the processed features of each target node.

[0054] This specification provides an information traceability device, comprising:

[0055] The acquisition module is used to acquire pre-built graph networks;

[0056] The information simulation propagation module is used to generate simulation instructions for the graph network, so as to execute the information simulation propagation task for the graph network through the simulation instructions, and obtain the propagated graph network and the state information of each node in the propagated graph network;

[0057] The prediction module is used to input the state information of each node and the basic network structure information of the graph network into the graph neural network model to be trained. Based on the state information of each node, the graph neural network model determines the initial features corresponding to each node. Based on the correlation between any two nodes, the basic network structure information is updated to obtain updated information. Based on the updated information, the initial features corresponding to each node are processed to obtain the processed features corresponding to each node. Based on the processed features corresponding to each node, the overall features of the propagated graph network are determined. Based on the overall features and the processed features corresponding to each node, the starting node of information propagation in the propagated graph network is predicted.

[0058] The training module is used to train the graph neural network model with the optimization objective of minimizing the difference between the predicted starting node and the actual starting node, so as to obtain the trained model.

[0059] The information tracing module is used to receive an information tracing request after obtaining the trained model, and input the state information of each target node in the target graph network to be traced and the basic network structure information of the target graph network into the trained model according to the information tracing request, so as to predict the starting target node of information propagation in the target graph network through the trained model, and execute the task according to the starting target node.

[0060] This specification provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned information tracing method.

[0061] This specification provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the aforementioned information tracing method.

[0062] The above-described at least one technical solution adopted in the embodiments of this specification can achieve the following beneficial effects:

[0063] In this embodiment, information propagation is first simulated based on the constructed graph network to obtain the state information of each node in the propagated graph network, which serves as training samples. The training samples and the basic network structure information of the graph network are input into the graph neural network model to be trained, so that the graph neural network model can predict the starting node of information propagation in the propagated graph network. The graph neural network model is trained with the optimization objective of minimizing the difference between the predicted starting node and the actual starting node. In this method, the propagation source is not predicted based on the propagation coefficient of information propagation. Instead, the graph neural network model is supervised to learn the correspondence between the propagated graph network and the propagation source of information propagation. Based on the trained graph neural network model, the prediction of the starting node in the graph network to be traced is improved, thereby enhancing the accuracy of the propagation source prediction. Attached Figure Description

[0064] The accompanying drawings, which are included to provide a further understanding of this specification and form part of this specification, illustrate exemplary embodiments and are used to explain this specification, but do not constitute an undue limitation thereof. In the drawings:

[0065] Figure 1 A flowchart illustrating the information tracing method provided in the embodiments of this specification;

[0066] Figure 2 A schematic diagram illustrating the state information of each node in the propagated graph network provided in the embodiments of this specification;

[0067] Figure 3 This is a schematic diagram of the model structure of the graph neural network model provided in the embodiments of this specification;

[0068] Figure 4 This is a schematic diagram of the information traceability device provided in the embodiments of this specification;

[0069] Figure 5 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this specification. Detailed Implementation

[0070] To make the objectives, technical solutions, and advantages of this specification clearer, the technical solutions of this specification will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this specification, and not all of them. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this specification.

[0071] The technical solutions provided in the various embodiments of this specification are described in detail below with reference to the accompanying drawings.

[0072] Figure 1 This is a flowchart illustrating the information tracing method provided in the embodiments of this specification. This information tracing method can be applied to a server and includes:

[0073] S100: Obtain a pre-built graph network.

[0074] S102: Generate simulation instructions for the graph network, and execute the information simulation propagation task for the graph network through the simulation instructions to obtain the propagated graph network and the state information of each node in the propagated graph network.

[0075] In the embodiments of this specification, a constructed graph network can be obtained first, and then information propagation can be simulated on the graph network to obtain training samples. Then, based on the training samples, the graph neural network model is subjected to supervised training, enabling the trained graph neural network to accurately predict the source of information propagation. The graph network can be a graph structure composed of several nodes and several edges, and the graph network can represent the relationships between various target objects.

[0076] Depending on the specific application scenario of information tracing, the target objects constituting the graph network can be different. For example, in the application scenario of tracing the source of rumors on social networks, the target object can be a social media account, and the relationship between the target objects can be mutual following between social media accounts. In the application scenario of locating the source of infectious diseases, the target object can be an individual user, and the relationship between the target objects can be a direct connection between users, such as a parent-child relationship or a spousal relationship. In the application scenario of tracing the source of online viruses, the target object can be a network address, and the relationship between the target objects can be the communication links between network addresses.

[0077] In the embodiments of this specification, a graph network must be constructed before obtaining the graph network.

[0078] Specifically, the process involves obtaining a group of target objects and determining the direct relationships between them. Then, using each target object as a node and the direct relationships between them as edges, a graph network is constructed. After constructing the graph network, the adjacency matrix or the Laplacian matrix corresponding to the adjacency matrix can be determined. When any two target objects have a direct relationship, the adjacency value between their nodes is 1; when no direct relationship exists between them, the adjacency value is 0.

[0079] The graph network can be G = (V, E). V represents the nodes in the graph network, i.e., the target objects, and E represents the edges in the graph network. If there is a direct relationship between two target objects, there is an edge between the nodes of these two target objects. The adjacency matrix of the graph network can be represented as: A ∈ R n×n Let n represent n nodes, where a in the adjacency matrix A is a number if and only if there is a direct association between node i and node j. ij =1. Furthermore, the Laplace matrix of the adjacency matrix is ​​denoted by S. Where A is the adjacency matrix of the graph network, I is the identity matrix, and d() represents the degree calculation.

[0080] After constructing the graph network, the constructed graph network can be obtained, and simulation instructions can be generated for the graph network. These instructions are then used to execute a simulated information propagation task on the graph network, obtaining the propagated graph network and the state information of each node within it. The simulation instructions carry the propagation dynamics model used for the simulated information propagation, the number of simulated propagations, and an indication of randomly selecting a node as the propagation source. The propagation dynamics model can include: SI model, SIR model, SEIR model, Independent Cascade (IC) model, Linear Threshold (LT) model, etc. The number of simulated propagations represents the number of training samples; the state information of each node in the propagated network obtained from one simulated propagation is considered one training sample.

[0081] The state information of a node can be used to indicate whether it is active or not. Depending on the application scenario, the activation status will represent different situations. For example, in the application scenario of tracing the source of rumors on social networks, activation indicates that a rumor has been spread, while inactivation indicates that no rumor has been spread. The corresponding state information can include: active state, inactive state. In the application scenario of locating the source of infectious diseases, activation indicates that the node has been infected, inactivation indicates that the node has not been infected, and a change from active to inactive indicates that the node has recovered after infection. The corresponding state information can include: infected state, uninfected state, and recovered state. In the application scenario of tracing the source of network viruses, activation indicates that the network virus has been spread, while inactivation indicates that the network virus has not been spread. The corresponding state information can include: active state, inactive state.

[0082] Specifically, by using the propagation dynamics model carried in the simulation instructions, an information propagation simulation task can be performed on the graph network to obtain the propagated graph network and the state information of each node in the propagated graph network.

[0083] Furthermore, according to the simulation instructions, a node is randomly selected from the constructed graph network and designated as the starting node for information propagation. This selected starting node is actually the real starting node during the training of the graph neural network model. Then, the selected starting node and the graph network are input into the propagation dynamics model. Based on a preset propagation coefficient, the propagation dynamics model initiates information propagation within the graph network from the selected starting node until the propagation time is reached, yielding the post-propagation graph network and the state information of each node within it. The propagation coefficient can include factors such as infection rate, recovery rate, and propagation time.

[0084] Taking the SIR model as an example, the propagation dynamics model categorizes node states into three types: infected (I), uninfected (S), and recovered (R). Before using the propagation dynamics model, the infection rate, recovery rate, and propagation time can be set. When using the propagation dynamics model, information can be propagated from the starting node in the graph network to other nodes in the graph network according to the propagation coefficient.

[0085] Taking an information dissemination simulation as an example, Figure 2 This is a schematic diagram of the state information of each node in the propagation graph network provided in the embodiments of this specification. Figure 2 In the diagram, infected nodes are black, recovered nodes are covered by a diagonal line, and uninfected nodes are white.

[0086] In addition, when multiple training samples are needed, multiple information propagation simulations can be performed using a propagation dynamics model, and the state information of each node in the graph network after multiple information propagation simulations can be used as multiple training samples.

[0087] For each simulated information propagation, a node can be randomly selected from the graph network as the starting node for information propagation. Then, the randomly selected starting node and the pre-constructed graph network are input into the propagation dynamics model. Based on the preset propagation coefficients, the propagation dynamics model propagates information in the graph network starting from the randomly selected starting node, and obtains the state information of the graph network after propagation and each node in the graph network after propagation, which serves as the training sample obtained after this simulated information propagation.

[0088] S104: Input the state information of each node and the basic network structure information of the graph network into the graph neural network model to be trained, so that the graph neural network model determines the initial features corresponding to each node based on the state information of each node, determines the correlation between any two nodes based on the initial features corresponding to each node, updates the basic network structure information based on the correlation between any two nodes to obtain updated information, processes the initial features corresponding to each node based on the updated information to obtain processed features corresponding to each node, determines the overall features of the propagated graph network based on the processed features corresponding to each node, and predicts the starting node of information propagation in the propagated graph network based on the overall features and the processed features corresponding to each node.

[0089] In the embodiments of this specification, after obtaining training samples, supervised training can be performed on the graph neural network model based on the training samples and labels, enabling the graph neural network model to learn the correspondence between the information propagation source and the propagated graph network. The graph neural network model structure includes: an input layer, an attention layer, a graph convolutional layer, a graph representation layer, and an output layer.

[0090] Specifically, the state information of each node in the propagated graph network and the basic network structure information of the graph network (or the propagated graph network) can be input into the graph neural network model to be trained. Based on the state information of each node, the graph neural network model determines the initial features corresponding to each node. Based on the initial features of each node, the correlation between any two nodes is determined. Then, based on the correlation between any two nodes, the basic network structure information is updated to obtain updated information. Based on the updated information, the initial features corresponding to each node are processed to obtain processed features corresponding to each node. Then, based on the processed features of each node, the overall features of the propagated graph network are determined. Finally, based on the overall features of the propagated graph network and the processed features of each node, the starting node of information propagation in the propagated graph network is predicted. Here, the starting node of information propagation can refer to the source of information propagation, and the basic network structure information can refer to the adjacency matrix or Laplacian matrix representing the connection relationships between nodes in the graph network.

[0091] Furthermore, the state information of each node and the basic network structure information of the graph network can be input into the graph neural network model to be trained, so that the state information of each node can be encoded through the input layer of the graph neural network model to obtain the features corresponding to each node as initial features.

[0092] Although the basic network structure information of a graph network can reflect whether there is a direct connection between any two nodes, for any two nodes that are not directly connected, these two nodes may be indirectly connected through other nodes. Therefore, it is necessary to calculate the correlation degree between any two nodes and update the basic network structure information based on the calculated correlation degree so that graph convolution can be performed using the updated network structure information.

[0093] Therefore, the initial features corresponding to each node can be input into the attention layer of the graph neural network model, so that the correlation between any two nodes can be determined by the attention layer based on the initial features corresponding to each node.

[0094] The formula for calculating the correlation between any two nodes is: a i,j =V T g(W s h i +W t h j +b s )+b v V T W s W t b s b v Let g be the attention coefficient matrix, g() be the activation function, and h be the activation function. i h is the initial feature corresponding to node i. j Let a be the initial feature corresponding to node j. i,j This represents the correlation between node i and node j. V T W s W t b s b v It is the matrix that needs to be learned during the training of a graph neural network model.

[0095] Then, the correlation between any two nodes and the basic network structure information can be input into the graph convolutional layer of the graph neural network model. The graph convolutional layer updates the basic network structure information based on the correlation between any two nodes, resulting in updated basic network structure information, which serves as the updated information. Then, based on the updated information, the initial features corresponding to each node are processed to obtain the processed features for each node.

[0096] When updating the basic network structure information, a correlation matrix containing the correlation between any two nodes can be determined based on their correlation. This correlation matrix is ​​then multiplied by the basic network structure information to obtain the updated information. The basic network structure information can be the adjacency matrix or Laplacian matrix of the graph network.

[0097] Taking the Laplace matrix as an example, the formula for updating the basic network structure information is: S = S·α. S can refer to the Laplace matrix, and α can refer to the relevance matrix.

[0098] When processing the initial features corresponding to each node based on the updated information, the updated information and the initial features corresponding to each node can be input into a graph convolutional network. The graph convolutional network then performs graph convolution on the initial features corresponding to each node based on the updated information, obtaining the convolutionally processed features for each node. Finally, the convolutionally processed features for each node are added to the initial features for each node to obtain the processed features for each node.

[0099] Introducing residual connections into graph convolutional networks with graph convolutional layers can preserve previous feature information and improve the training performance of the model.

[0100] When a convolutional layer contains multiple Graph Convolutional Networks (GCNs), for each GCN, the processed features corresponding to each node output by the previous GCN are redefined as the initial features input to each node of that GCN. Then, the initial features and updated information of each node in the previous GCN are input into the GCN. The GCN performs graph convolution on the initial features of each node based on the updated information, resulting in the convolutional features for each node. These convolutional features are then added to the initial features input to the previous GCN to obtain the processed features for each node. Alternatively, the processed features output by the previous GCN can be used as the initial features input to the next GCN.

[0101] The feature transfer between each graph convolutional network is as follows: H l+1 =σ(SH) l W l )+H l Among them, H l W represents the processed features corresponding to each node output by the l-th graph convolutional network. lLet S be the weight matrix of the l-th graph convolutional network. S represents the updated information. H l+1 W represents the processed features corresponding to each node output by the (l+1)th graph convolutional network. l It is the matrix that needs to be learned during the training of a graph neural network model.

[0102] When there is only one graph convolutional network in the graph convolutional layer, the processed features corresponding to each node output by the graph convolutional network can be directly input into the graph representation layer in the graph neural network model; when there are multiple graph convolutional networks in the graph convolutional layer, the processed features corresponding to each node output by the last graph convolutional network can be input into the graph representation layer in the graph neural network model.

[0103] The graph representation layer of a graph neural network model can be used to represent the overall features of the graph network after propagation.

[0104] The processed features corresponding to each node output by the graph convolutional layer are input into the graph representation layer. Based on the processed features corresponding to each node, the graph representation layer determines the overall features of the propagated graph network.

[0105] When determining the overall characteristics of the propagated graph network based on the processed features corresponding to each node, the importance of each node in the propagated graph network can be determined first based on the processed features corresponding to each node. Specifically, if a node is in a non-specified state, its importance in the propagated graph network is relatively small; or, if a node is located at the edge of the propagation range in the propagated graph network, its importance in the propagated graph network is relatively small.

[0106] The formula for calculating the importance of each node is: Z = σ(H) L Q), H L It is a matrix containing the processed features corresponding to each node of the output of the last graph convolutional layer, where Q is the weight matrix representing the importance of each node, and σ() is the activation function. Here, Q is the matrix that needs to be learned during the training of the graph neural network model.

[0107] Then, for each state, based on the importance of each node in the propagated graph network, a specified number of nodes in that state are selected as important nodes. Next, the processed features of the other nodes in that state (excluding the important nodes) are masked. Pooling is then performed on the processed features of the important nodes to obtain the pooling result for that state. Finally, the pooling results for each state are weighted and summed to obtain the overall features of the propagated graph network.

[0108] When filtering important nodes, the nodes in this state can be sorted from most important to least important, and then a specified number of nodes at the top of the sorted positions can be selected as important nodes.

[0109] When performing pooling operations on the processed features corresponding to important nodes, we can perform max pooling on the processed features corresponding to important nodes to obtain the first pooling result, and simultaneously perform average pooling on the processed features corresponding to important nodes to obtain the second pooling result. Summing the first pooling result and the second pooling result yields the pooling result for this state.

[0110] The formula for determining the overall characteristics of the graph network after propagation is: state represents different states, a represents state a, and W a Let M be the weight matrix for state 'a', where max() is the max pooling operation and mean() is the average pooling operation. a is the mask matrix for state 'a'. 'readout' represents the overall feature.

[0111] The overall features of the propagated graph network and the processed features of each node are input into the output layer of the graph neural network model. The output layer then predicts the starting node of information propagation in the propagated graph network based on the overall features and the processed features of each node.

[0112] When predicting the starting node for information propagation in the post-propagation graph network based on the overall features and the processed features corresponding to each node, the similarity between each node and the post-propagation graph network can be determined based on the overall features and the processed features corresponding to each node. Then, based on the processed features corresponding to each node, the overall features, and the similarity between each node and the post-propagation graph network, the starting node for information propagation in the post-propagation graph network is predicted.

[0113] When predicting the starting node for information propagation in the propagated graph network based on the processed features, overall features, and similarity between each node and the propagated graph network, for each node, a fully connected layer is applied to its processed features and the overall features of the propagated graph network to obtain its fully connected features. Then, based on the similarity between the node and the propagated graph network and its fully connected features, the predicted probability that the node is the starting node for information propagation in the propagated graph network is predicted, and this probability is used as the predicted probability for that node. Finally, the node with the highest predicted probability is selected as the starting node for information propagation in the propagated graph network.

[0114] The formula for calculating the predicted probability is: out=σ([HL ,readout]*P+cos(H L ,readout)). P is the weight matrix of the fully connected layer, [H L `P` represents the fully connected features of each node obtained by fully connecting the processed features of each node to the overall features of the propagated graph network through the fully connected layers contained in the output layer. `cos()` is used to calculate the similarity between the processed features of a node and the overall features. `Out` can be the predicted probability of each node as the starting node.

[0115] Based on the above description of the graph neural network model, this specification provides a schematic diagram of the model structure of a graph neural network model, as shown below. Figure 3 As shown. In Figure 3 In this graph neural network model, the input layer receives the state information of each node and the basic network structure information of the graph network, and outputs the initial features corresponding to each node. The attention layer receives the initial features corresponding to each node and outputs the correlation between any two nodes. The graph convolutional layer receives the correlation between any two nodes and the basic network structure information of the graph network, and outputs the processed features corresponding to each node. The graph representation layer receives the processed features corresponding to each node and outputs the overall features of the graph network after propagation. The output layer receives the overall features of the graph network after propagation and the processed features corresponding to each node, and outputs the predicted probability of each node being the starting node. The entire graph neural network model outputs the node with the highest predicted probability as the starting node.

[0116] S106: The graph neural network model is trained with the goal of minimizing the difference between the predicted starting node and the actual starting node, and the trained model is obtained.

[0117] In the embodiments of this specification, after the starting node is predicted by the graph neural network model, the model parameters of the graph neural network model can be updated through backpropagation, that is, the graph neural network model is trained.

[0118] When training a graph neural network model, the optimization objective can be to minimize the difference between the predicted starting node and the actual starting node. This can be used to train the graph neural network model and obtain the trained graph neural network model as the post-trained model.

[0119] Specifically, the predicted probability output by the graph neural network model for each node as the starting node can be determined as the predicted probability for each node. Based on the difference between the predicted probability of each node and its label, the loss for the graph neural network model is determined as the first loss. The graph neural network model is trained with the goal of minimizing the first loss. Here, the first loss can refer to the cross-entropy loss. The label can be a 0 / 1 value. The closer the predicted probability of a node is to 1, the more likely that node is to be the starting node.

[0120] The formula for calculating the first loss is as follows: y i Let be the label of the i-th node. Let be the predicted probability corresponding to the i-th node.

[0121] Furthermore, taking the SIR model as an example, nodes in an infected state and nodes in a recovered state are more likely to be sources of information propagation than nodes in a vulnerable state. Therefore, this specification adds a first label to nodes in an infected state and nodes in a recovered state, and adds a second label to nodes in a vulnerable state. The value of the first label is greater than the value of the second label, indicating that nodes in an infected state and nodes in a recovered state are more likely to be sources of information propagation than nodes in a vulnerable state.

[0122] The value of the first label can be the normalized result of normalizing a first preset number, and the value of the second label can be the normalized result of normalizing a second preset number. The first preset number can be 1, and the second preset number can be 0.

[0123] Therefore, when training a graph neural network model, after determining the predicted probability for each node, the predicted probability for each node can be normalized to obtain a normalized result for each node. Then, the graph neural network model is trained with the optimization objective of minimizing the difference between the normalized result for each node and the true normalized result for each node, as well as minimizing the difference between the predicted starting node and the true starting node. Specifically, the true normalized result for a node in a specified state is greater than the true normalized result for a node in an unspecified state. A node in a specified state can refer to a node in an infected state or a node in a recovered state, while a node in an unspecified state can refer to a node in a vulnerable state.

[0124] With the optimization objectives of minimizing the difference between the predicted starting node and the true starting node, and minimizing the difference between the normalized result corresponding to each node and the true normalized result corresponding to each node, a second loss can be determined for the graph neural network model during training, based on the difference between the normalized result corresponding to each node and the true normalized result corresponding to each node. This second loss can be the KL divergence loss. Then, the first and second losses are summed to obtain the total loss. The graph neural network model is trained with the goal of minimizing the total loss.

[0125] The formula for calculating the second loss is as follows: p i This represents the true normalized result for the i-th node.

[0126] In addition to the first and second losses mentioned above, in order to reduce the risk of model overfitting, an L2 regularization term for the model parameters is introduced into the loss function during the training process of the graph neural network model.

[0127] Specifically, the parameters of the graph neural network model are determined, and L2 regularization is applied to each parameter to obtain the third loss. The first, second, and third losses are summed to obtain the comprehensive loss. The graph neural network model is trained with the goal of minimizing the comprehensive loss.

[0128] The formula for calculating the third loss is: l3 = γ‖W‖2. γ is the hyperparameter, and W is the model parameter. The overall loss is: l 综 = l1 + l2 + l3.

[0129] S108: After obtaining the trained model, receive an information tracing request, and according to the information tracing request, input the state information of each target node in the target graph network to be traced and the basic network structure information of the target graph network into the trained model, so as to predict the starting target node of information propagation in the target graph network through the trained model, and execute the task according to the starting target node.

[0130] In the embodiments of this specification, after training the graph neural network model to obtain the trained model, the propagation source of information in the target graph network to be traced can be predicted based on the trained model, and the task can be executed based on the predicted propagation source. The target graph network is the graph network after information propagation, and each node in the target graph network can serve as a target node.

[0131] Specifically, the system first receives an information tracing request, which carries the target graph network to be traced. Then, based on the tracing request, the state information of each target node in the target graph network and the basic network structure information are input into the trained model. The trained model then predicts the starting target node for information propagation in the target graph network, and executes the task based on this starting target node. The starting target node can refer to the source of information propagation in the target graph network.

[0132] When, based on an information tracing request, the state information of each target node in the target graph network to be traced, along with the basic network structure information of the target graph network, is input into a trained model to predict the starting target node for information propagation in the target graph network, the target graph network to be traced and the state information of each target node in the target graph network can be determined according to the information tracing request. Then, the state information of each target node and the basic network structure information of the target graph network are input into the trained model. Based on the state information of each target node, the trained model determines the initial features corresponding to each target node. Based on the initial features corresponding to each target node, the correlation between any two target nodes is determined. Based on the correlation between any two target nodes, the basic network structure information is updated to obtain updated information. Based on the updated information, the initial features corresponding to each target node are processed to obtain processed features corresponding to each target node. Based on the processed features corresponding to each target node, the overall features of the target graph network are determined. Based on the overall features and the processed features corresponding to each target node, the starting node for information propagation in the target graph network is predicted.

[0133] When predicting the starting node for information propagation in the target graph network based on the overall features and the processed features corresponding to each target node, the similarity between each target node and the target graph network can be determined. Then, for each target node, based on its processed features, the overall features of the target graph network, and the similarity between the target node and the target graph network, the predicted probability that the target node is the starting target node for information propagation in the target graph network is predicted, and this prediction is taken as the predicted probability of the target node. Finally, the target node with the highest predicted probability is selected from the predicted probabilities of each target node and is taken as the starting target node for information propagation in the target graph network.

[0134] When executing tasks based on predicted starting target nodes, the tasks performed vary depending on the application scenario. For example, in the application scenario of tracing the source of rumors on social networks, the task might be to delete or block the social media account corresponding to the starting target node. In the application scenario of locating the source of infectious diseases, the task might be to display the information of the user corresponding to the predicted starting target node on the terminal device of epidemic prevention workers, or to send an alarm message to the terminal device of the user corresponding to the starting target node to warn the user not to go out. In the application scenario of tracing the source of network viruses, the task might be to add the network address corresponding to the starting target node to the blacklist to prevent the continued spread of network viruses through that network address.

[0135] It should be noted that all actions involving the acquisition of signals, information, or data in this application are carried out in compliance with the relevant data protection laws and policies of the country where the application is located, and with the authorization granted by the owner of the relevant device.

[0136] Through the above Figure 1 As shown in the method, this specification first simulates information propagation based on the constructed graph network to obtain the state information of each node in the propagated graph network, which serves as training samples. The training samples and the basic network structure information of the graph network are input into the graph neural network model to be trained, so that the graph neural network model can predict the starting node of information propagation in the propagated graph network. The graph neural network model is trained with the optimization objective of minimizing the difference between the predicted starting node and the actual starting node. In this method, the propagation source is not predicted based on the propagation coefficient of information propagation. Instead, the graph neural network model is supervised to learn the correspondence between the propagated graph network and the propagation source of information propagation. Based on the trained graph neural network model, the prediction of the starting node in the graph network to be traced is improved, thereby enhancing the accuracy of the propagation source prediction.

[0137] The above are the information tracing methods provided in the embodiments of this specification. Based on the same idea, this specification also provides corresponding devices, storage media and electronic devices.

[0138] Figure 4 This is a schematic diagram of an information tracing device provided in an embodiment of this specification. The device includes:

[0139] Acquisition module 401 is used to acquire a pre-built graph network;

[0140] The information simulation propagation module 402 is used to generate simulation instructions for the graph network, so as to execute the information simulation propagation task for the graph network through the simulation instructions, and obtain the propagated graph network and the state information of each node in the propagated graph network.

[0141] The prediction module 403 is used to input the state information of each node and the basic network structure information of the graph network into the graph neural network model to be trained, so that the graph neural network model can determine the initial features corresponding to each node based on the state information of each node, update the basic network structure information based on the correlation between any two nodes to obtain updated information, process the initial features corresponding to each node according to the updated information to obtain processed features corresponding to each node, determine the overall features of the propagated graph network based on the processed features corresponding to each node, and predict the starting node of information propagation in the propagated graph network according to the overall features and the processed features corresponding to each node.

[0142] Training module 404 is used to train the graph neural network model with the optimization objective of minimizing the difference between the predicted starting node and the actual starting node, so as to obtain the trained model.

[0143] The information tracing module 405 is used to receive an information tracing request after obtaining the trained model, and input the state information of each target node in the target graph network to be traced and the basic network structure information of the target graph network into the trained model according to the information tracing request, so as to predict the starting target node of information propagation in the target graph network through the trained model, and execute the task according to the starting target node.

[0144] Optionally, before obtaining the pre-constructed graph network, the acquisition module 401 is further configured to: acquire a target object group and determine the direct association between each target object in the target object group; and construct a graph network with each target object as a node and the direct association between each target object as an edge.

[0145] Optionally, the information simulation propagation module 402 is specifically used to: randomly select a node from the pre-constructed graph network as the starting node for information propagation according to the simulation instruction; input the selected starting node and the pre-constructed graph network into the propagation dynamics model, so that the propagation dynamics model, based on a preset propagation coefficient, propagates information in the graph network starting from the selected starting node, and obtains the propagated graph network and the state information of each node in the propagated graph network.

[0146] Optionally, the basic network structure information includes: an adjacency matrix or a Laplace matrix.

[0147] Optionally, the graph neural network model includes: an input layer, an attention layer, a graph convolutional layer, a graph representation layer, and an output layer; the prediction module 403 is specifically used to: input the state information of each node and the basic network structure information of the graph network into the graph neural network model to be trained, so as to encode the state information of each node through the input layer to obtain the initial features corresponding to each node; input the initial features corresponding to each node into the attention layer, so as to determine the correlation between any two nodes based on the initial features corresponding to each node through the attention layer; input the correlation between any two nodes and the basic network structure information into the graph convolutional layer, so as to determine the correlation between any two nodes and the basic network structure information based on the initial features corresponding to each node through the graph convolutional layer. Based on the correlation between any two nodes, the basic network structure information is updated to obtain updated information. According to the updated information, the initial features corresponding to each node are processed to obtain processed features corresponding to each node. The processed features corresponding to each node output from the graph convolutional layer are input into the graph representation layer. Through the graph representation layer, based on the processed features corresponding to each node, the overall features of the propagated graph network are determined. The overall features and the processed features corresponding to each node are input into the output layer. Through the output layer, based on the overall features and the processed features corresponding to each node, the starting node for information propagation in the propagated graph network is predicted.

[0148] Optionally, the prediction module 403 is specifically used to: determine a correlation matrix containing the correlation between any two nodes based on the correlation between any two nodes; perform a dot product between the correlation matrix and the basic network structure information to obtain the dot product result; and use the dot product result as the updated information after updating the basic network structure information.

[0149] Optionally, the prediction module 403 is specifically used to input the initial features corresponding to each node and the updated information into the graph convolutional network in the graph convolutional layer, so that the graph convolutional network performs graph convolution on the initial features corresponding to each node according to the updated information to obtain the graph convolutional features corresponding to each node, and adds the graph convolutional features corresponding to each node to the initial features corresponding to each node to obtain the processed features corresponding to each node.

[0150] Optionally, the prediction module 403 is specifically configured to: the graph convolutional layer comprises multiple graph convolutional networks; sequentially, for each graph convolutional network, redetermine the processed features corresponding to each node output by the previous graph convolutional network as the initial features corresponding to each node input to the graph convolutional network; input the initial features corresponding to each node input to the graph convolutional network and the updated information into the graph convolutional network, so that the graph convolutional network performs graph convolution on the initial features corresponding to each node according to the updated information to obtain the graph convolutional features corresponding to each node in the graph convolutional network; add the graph convolutional features corresponding to each node in the graph convolutional network to the initial features corresponding to each node input to the graph convolutional network to obtain the processed features corresponding to each node output by the graph convolutional network.

[0151] Optionally, the prediction module 403 is specifically used to: determine the importance of each node in the propagation graph network based on the processed features corresponding to each node; for each state, select a specified number of nodes as important nodes from the nodes in that state according to the importance of each node in the propagation graph network; perform pooling operation on the processed features corresponding to the important nodes to obtain the pooling result in that state; and perform weighted summation of the pooling results in each state to obtain the overall features of the propagation graph network.

[0152] Optionally, the prediction module 403 is specifically used to: determine the similarity between each node and the propagation graph network based on the processed features corresponding to each node and the overall features; and predict the starting node of information propagation in the propagation graph network based on the processed features corresponding to each node, the overall features, and the similarity between each node and the propagation graph network.

[0153] Optionally, the prediction module 403 is specifically used to, for each node, predict the probability that the node is the starting node for information propagation in the propagation graph network based on the processed features corresponding to the node, the overall features, and the similarity between the node and the propagation graph network, and use this probability as the prediction probability corresponding to the node; and select the node with the highest prediction probability from the prediction probabilities corresponding to each node as the starting node for information propagation in the propagation graph network.

[0154] Optionally, the training module 404 is specifically used to: determine the predicted probability output by the graph neural network model for each node as the starting node, as the predicted probability corresponding to each node; normalize the predicted probability corresponding to each node to obtain the normalized result corresponding to each node; and train the graph neural network model with the optimization objective of minimizing the difference between the predicted starting node and the true starting node, and minimizing the difference between the normalized result corresponding to each node and the true normalized result corresponding to each node; wherein, the true normalized result corresponding to a node in a specified state is greater than the true normalized result corresponding to a node in a non-specified state.

[0155] Optionally, the information tracing module 405 is specifically used to: determine the target graph network to be traced according to the information tracing request, and determine the state information of each target node in the target graph network; input the state information of each target node and the basic network structure information of the target graph network into the trained model, so that the trained model can determine the initial features corresponding to each target node according to the state information of each target node, determine the correlation between any two target nodes according to the initial features corresponding to each target node, update the basic network structure information based on the correlation between any two target nodes to obtain updated information, process the initial features corresponding to each target node according to the updated information to obtain processed features corresponding to each target node, determine the overall features of the target graph network based on the processed features corresponding to each target node, and predict the starting target node of information propagation in the target graph network based on the overall features and the processed features corresponding to each target node.

[0156] This specification also provides a computer-readable storage medium storing a computer program that, when executed by a processor, can be used to perform the above-described actions. Figure 1 The provided information tracing method.

[0157] based on Figure 1 The information tracing method shown in this specification, in addition to the embodiments provided, also provides Figure 5 The diagram shows the structure of the electronic device. Figure 5 At the hardware level, this electronic device includes a processor, internal bus, network interface, memory, and non-volatile memory, and may also include other hardware required for business operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then executes it to achieve the above. Figure 1 The information tracing method described above.

[0158] Of course, in addition to software implementation, this specification does not exclude other implementation methods, such as logic devices or a combination of hardware and software. In other words, the execution subject of the following processing flow is not limited to each logic unit, but can also be hardware or logic devices.

[0159] In the 1990s, improvements to a technology could be clearly distinguished as either hardware improvements (e.g., improvements to the circuit structure of diodes, transistors, switches, etc.) or software improvements (improvements to the methodology). However, with technological advancements, many methodological improvements today can be considered direct improvements to the hardware circuit structure. Designers almost always obtain the corresponding hardware circuit structure by programming the improved methodology into the hardware circuit. Therefore, it cannot be said that a methodological improvement cannot be implemented using hardware physical modules. For example, a Programmable Logic Device (PLD) (such as a Field Programmable Gate Array (FPGA)) is such an integrated circuit whose logic function is determined by the user programming the device. Designers can program and "integrate" a digital system onto a PLD themselves, without needing chip manufacturers to design and manufacture dedicated integrated circuit chips. Furthermore, nowadays, instead of manually manufacturing integrated circuit chips, this programming is mostly implemented using "logic compiler" software. Similar to the software compiler used in program development, the original code before compilation must be written in a specific programming language, called a Hardware Description Language (HDL). There are many HDLs, such as ABEL (Advanced Boolean Expression Language), AHDL (Altera Hardware Description Language), Confluence, CUPL (Cornell University Programming Language), HDCal, JHDL (Java Hardware Description Language), Lava, Lola, MyHDL, PALASM, and RHDL (Ruby Hardware Description Language). Currently, the most commonly used are VHDL (Very-High-Speed ​​Integrated Circuit Hardware Description Language) and Verilog. Those skilled in the art should understand that by simply performing some logic programming on the method flow using one of these hardware description languages ​​and programming it into an integrated circuit, the hardware circuit implementing the logical method flow can be easily obtained.

[0160] The controller can be implemented in any suitable manner. For example, it can take the form of a microprocessor or processor and a computer-readable medium storing computer-readable program code (e.g., software or firmware) executable by the (micro)processor, logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers, and embedded microcontrollers. Examples of controllers include, but are not limited to, the following microcontrollers: ARC 625D, Atmel AT91SAM, Microchip PIC18F26K20, and Silicon Labs C8051F320. A memory controller can also be implemented as part of the control logic of the memory. Those skilled in the art will also recognize that, in addition to implementing the controller in purely computer-readable program code form, the same functionality can be achieved by logically programming the method steps to make the controller take the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, such a controller can be considered a hardware component, and the means included therein for implementing various functions can also be considered as structures within the hardware component. Alternatively, the means for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0161] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, a computer can be, for example, a personal computer, laptop computer, cellular phone, camera phone, smartphone, personal digital assistant, media player, navigation device, email device, game console, tablet computer, wearable device, or any combination of these devices.

[0162] For ease of description, the above devices are described in terms of function, divided into various units. Of course, in implementing this specification, the functions of each unit can be implemented in one or more software and / or hardware components.

[0163] Those skilled in the art will understand that embodiments of this specification can be provided as methods, systems, or computer program products. Therefore, this specification may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this specification may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0164] This specification is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this specification. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0165] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0166] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0167] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0168] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0169] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0170] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0171] Those skilled in the art will understand that the embodiments of this specification can be provided as methods, systems, or computer program products. Therefore, this specification may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this specification may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0172] This specification can be described in the general context of computer-executable instructions that are executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. This specification can also be practiced in distributed computing environments, where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0173] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0174] The above description is merely an embodiment of this specification and is not intended to limit this specification. Various modifications and variations can be made to this specification by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this specification should be included within the scope of the claims of this specification.

Claims

1. An information tracing method, characterized in that, include: Obtain a pre-built graph network; A simulation instruction is generated for the graph network, and an information simulation propagation task for the graph network is executed through the simulation instruction to obtain the propagated graph network and the state information of each node in the propagated graph network. The state information of each node and the basic network structure information of the graph network are input into the graph neural network model to be trained. Based on the state information of each node, the graph neural network model determines the initial features corresponding to each node. Based on the initial features corresponding to each node, the correlation between any two nodes is determined. Based on the correlation between any two nodes, the basic network structure information is updated to obtain updated information. Based on the updated information, the initial features corresponding to each node are processed to obtain processed features corresponding to each node. Based on the processed features corresponding to each node, the overall features of the propagated graph network are determined. Based on the overall features and the processed features corresponding to each node, the starting node of information propagation in the propagated graph network is predicted. The optimization objective is to minimize the difference between the predicted starting node and the actual starting node. The graph neural network model is then trained to obtain the trained model. After obtaining the trained model, an information tracing request is received. According to the information tracing request, the state information of each target node in the target graph network to be traced and the basic network structure information of the target graph network are input into the trained model so as to predict the starting target node of information propagation in the target graph network through the trained model, and execute the task according to the starting target node. Specifically, determining the overall features of the propagated graph network based on the processed features corresponding to each node includes: Based on the processed features corresponding to each node, the importance of each node in the post-propagation graph network is determined; For each state, based on the importance of each node in the post-propagation graph network, a specified number of nodes are selected from the nodes in that state as important nodes; Pooling is performed on the processed features corresponding to the important nodes to obtain the pooling result under this state; The pooling results under each state are weighted and summed to obtain the overall features of the propagated graph network.

2. The method as described in claim 1, characterized in that, Before obtaining the pre-built graph network, the method further includes: Obtain the target object group and determine the direct relationships between each target object in the target object group; A graph network is constructed with each target object as a node and the direct relationships between target objects as edges.

3. The method as described in claim 1, characterized in that, The simulation instructions are used to execute a simulated information propagation task for the graph network, obtaining the propagated graph network and the state information of each node in the propagated graph network, specifically including: According to the simulation instructions, a node is randomly selected from the pre-constructed graph network to be determined as the starting node for information propagation; The determined starting node and the pre-constructed graph network are input into the propagation dynamics model. Based on the preset propagation coefficients, the propagation dynamics model propagates information from the determined starting node into the graph network, resulting in the propagated graph network and the state information of each node in the propagated graph network.

4. The method as described in claim 1, characterized in that, The basic network structure information includes: adjacency matrix or Laplace matrix.

5. The method as described in claim 1, characterized in that, The graph neural network model includes: an input layer, an attention layer, a graph convolutional layer, a graph representation layer, and an output layer; The state information of each node and the basic network structure information of the graph network are input into the graph neural network model. Based on the state information of each node, the graph neural network model determines the initial features corresponding to each node. Based on the initial features of each node, the correlation between any two nodes is determined. Based on the correlation between any two nodes, the basic network structure information is updated to obtain updated information. Based on the updated information, the initial features corresponding to each node are processed to obtain processed features corresponding to each node. Based on the processed features of each node, the overall features of the propagated graph network are determined. Based on the overall features and the processed features of each node, the starting node for information propagation in the propagated graph network is predicted. Specifically, this includes: The state information of each node and the basic network structure information of the graph network are input into the graph neural network model to be trained, so that the state information of each node is encoded through the input layer to obtain the initial features corresponding to each node. The initial features corresponding to each node are input into the attention layer, so that the correlation between any two nodes can be determined by the attention layer based on the initial features corresponding to each node. The correlation between any two nodes and the basic network structure information are input into the graph convolutional layer. The graph convolutional layer updates the basic network structure information based on the correlation between any two nodes to obtain updated information. Based on the updated information, the initial features corresponding to each node are processed to obtain the processed features corresponding to each node. The processed features corresponding to each node output by the graph convolutional layer are input into the graph representation layer, so that the overall features of the propagated graph network can be determined by the graph representation layer based on the processed features corresponding to each node. The overall features and the processed features corresponding to each node are input into the output layer, so that the starting node of information propagation in the propagation graph network can be predicted by the output layer based on the overall features and the processed features corresponding to each node.

6. The method as described in claim 1 or 5, characterized in that, The basic network structure information is updated based on the correlation between any two nodes to obtain the updated information, specifically including: Based on the correlation between any two nodes, determine the correlation matrix that includes the correlation between any two nodes; The correlation matrix is ​​multiplied by the basic network structure information to obtain the dot product result, which is then used as the updated information after updating the basic network structure information.

7. The method as described in claim 1 or 5, characterized in that, Based on the updated information, the initial features corresponding to each node are processed to obtain the processed features corresponding to each node, specifically including: The initial features corresponding to each node and the updated information are input into the graph convolutional network in the graph convolutional layer. The graph convolutional network performs graph convolution on the initial features corresponding to each node according to the updated information to obtain the graph convolutional features corresponding to each node. The graph convolutional features corresponding to each node are added to the initial features corresponding to each node to obtain the processed features corresponding to each node.

8. The method as described in claim 7, characterized in that, The graph convolutional layer contains multiple graph convolutional networks; The initial features corresponding to each node and the updated information are input into the graph convolutional network in the graph convolutional layer. The graph convolutional network performs graph convolution on the initial features corresponding to each node based on the updated information to obtain the graph convolutional features corresponding to each node. The graph convolutional features corresponding to each node are then added to the initial features corresponding to each node to obtain the processed features corresponding to each node. Specifically, this includes: For each graph convolutional network, the processed features corresponding to each node output by the previous graph convolutional network are redefined as the initial features corresponding to each node input to the current graph convolutional network. The initial features corresponding to each node of the graph convolutional network and the updated information are input into the graph convolutional network. The graph convolutional network performs graph convolution on the initial features corresponding to each node according to the updated information to obtain the graph convolutional features corresponding to each node in the graph convolutional network. The graph convolutional features corresponding to each node in the graph convolutional network are added to the initial features corresponding to each node input into the graph convolutional network to obtain the processed features corresponding to each node output by the graph convolutional network.

9. The method as described in claim 1 or 5, characterized in that, Based on the overall features and the processed features corresponding to each node, the starting node for information propagation in the post-propagation graph network is predicted, specifically including: Based on the processed features corresponding to each node and the overall features, the similarity between each node and the propagated graph network is determined. Based on the processed features of each node, the overall features, and the similarity between each node and the post-propagation graph network, the starting node for information propagation in the post-propagation graph network is predicted.

10. The method as described in claim 9, characterized in that, Based on the processed features corresponding to each node, the overall features, and the similarity between each node and the post-propagation graph network, the starting node for information propagation in the post-propagation graph network is predicted, specifically including: For each node, based on the processed features corresponding to the node, the overall features, and the similarity between the node and the post-propagation graph network, the predicted probability that the node is the starting node for information propagation in the post-propagation graph network is predicted, and this probability is used as the predicted probability for the node. The node with the highest predicted probability is selected from the predicted probabilities of each node and used as the starting node for information propagation in the post-propagation graph network.

11. The method as described in claim 1, characterized in that, The optimization objective is to minimize the difference between the predicted starting node and the actual starting node. The training of the graph neural network model specifically includes: The predicted probability output by the graph neural network model for each node as the starting node is determined as the predicted probability for each node. The predicted probability for each node is normalized to obtain the normalized result for each node. The graph neural network model is trained with the optimization objective of minimizing the difference between the predicted starting node and the actual starting node, and minimizing the difference between the normalized result corresponding to each node and the actual normalized result corresponding to each node. The actual normalized result corresponding to a node in a specified state is greater than the actual normalized result corresponding to a node in a non-specified state.

12. The method as described in claim 1, characterized in that, Based on the information tracing request, the state information of each target node in the target graph network to be traced, as well as the basic network structure information of the target graph network, are input into the trained model to predict the starting target node for information propagation in the target graph network through the trained model. Specifically, this includes: Based on the information tracing request, determine the target graph network to be traced, and determine the state information of each target node in the target graph network; The state information of each target node and the basic network structure information of the target graph network are input into the trained model. The trained model determines the initial features corresponding to each target node based on the state information of each target node, determines the correlation between any two target nodes based on the initial features of each target node, updates the basic network structure information based on the correlation between any two target nodes to obtain updated information, processes the initial features corresponding to each target node based on the updated information to obtain processed features corresponding to each target node, determines the overall features of the target graph network based on the processed features of each target node, and predicts the starting target node for information propagation in the target graph network based on the overall features and the processed features of each target node.

13. An information traceability device, characterized in that, include: The acquisition module is used to acquire pre-built graph networks; The information simulation propagation module is used to generate simulation instructions for the graph network, so as to execute the information simulation propagation task for the graph network through the simulation instructions, and obtain the propagated graph network and the state information of each node in the propagated graph network; The prediction module is used to input the state information of each node and the basic network structure information of the graph network into the graph neural network model to be trained. Based on the state information of each node, the graph neural network model determines the initial features corresponding to each node. Based on the initial features of each node, it determines the correlation between any two nodes. Based on the correlation between any two nodes, it updates the basic network structure information to obtain updated information. Based on the updated information, it processes the initial features corresponding to each node to obtain processed features corresponding to each node. Based on the processed features of each node, it determines the overall features of the propagated graph network. Based on the overall features... Based on the processed features corresponding to each node, the starting node for information propagation in the post-propagation graph network is predicted. The determination of the overall features of the post-propagation graph network based on the processed features corresponding to each node specifically includes: determining the importance of each node in the post-propagation graph network based on the processed features corresponding to each node; for each state, selecting a specified number of nodes as important nodes from the nodes in that state according to their importance in the post-propagation graph network; performing pooling operations on the processed features corresponding to the important nodes to obtain the pooling result for that state; and weighted summing the pooling results for each state to obtain the overall features of the post-propagation graph network. The training module is used to train the graph neural network model with the optimization objective of minimizing the difference between the predicted starting node and the actual starting node, so as to obtain the trained model. The information tracing module is used to receive an information tracing request after obtaining the trained model, and input the state information of each target node in the target graph network to be traced and the basic network structure information of the target graph network into the trained model according to the information tracing request, so as to predict the starting target node of information propagation in the target graph network through the trained model, and execute the task according to the starting target node.

14. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the method described in any one of claims 1-12.

15. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method described in any one of claims 1-12.

Images