Rumor detection method and device based on graph neural diffusion network

Abstract

The invention discloses a rumor detection method and device based on a graph neural diffusion network, and belongs to the technical field of information detection. The method and the device are used for overcoming the defect that false texts cannot be accurately detected and recognized in the prior art. The method comprises the following steps: constructing a text propagation graph to obtain an original text propagation graph, performing data enhancement by using the original text propagation graph to obtain three enhanced views, and then extracting an echo room of the original text propagation graph through an edge betweenness-based algorithm. And training a global graph encoder model by using the enhanced graph and the original graph, training an echo room encoder model and an encoder model by using an echo room, and detecting the authenticity of the text by using the trained models. The defect that false texts cannot be accurately detected and recognized in the prior art is overcome.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of information detection, and more particularly to a rumor detection method and device based on a graph neural diffusion network. BACKGROUND

[0002] In the current digital era, social media platforms have rapidly developed and become the main way for people to obtain and disseminate information. However, the quality of information is uneven, including the growing virtual or rumor information. In order to automatically and accurately identify virtual or rumor information, researchers have proposed methods based on text content and propagation graphs. Relying only on information such as emotions and positions in text or pictures limits the recognition accuracy of the model in cross-domain and cross-event scenarios.

[0003] In recent years, many scholars have proposed methods for modeling text propagation graphs based on RNN (Recurrent Neural Network), CNN (Convolutional Neural Network), and GNN (Graph Neural Network). However, these methods face two key limitations: difficulty in effectively extracting long-distance dependencies and dense subgraph features formed by the echo chamber effect. They use general graph modeling networks to extract text propagation graph features without focusing on the unique characteristics of text propagation graphs, which limits the recognition efficiency and accuracy of the model. SUMMARY

[0004] The embodiments of the present application provide a rumor detection method and device based on a graph neural diffusion network to solve the defect that the prior art cannot accurately detect and identify false text.

[0005] The embodiments of the present application provide a rumor detection method based on a graph neural diffusion network, comprising:

[0006] determining the original node features of the original text propagation graph, performing data augmentation on the original text propagation graph to obtain an augmented view and a set number of echo chambers;

[0007] constructing an initial GDN-ODE network including a mapping module, a diffusion module, and a readout module; determining the initial dynamic features X dyn (0) corresponding to the original node features according to the mapping module; discretizing the time interval of the diffusion process into multiple time steps, at the first time step, determining the current dynamic features of the second time step based on the initial dynamic features, the diffusion module, and the adaptive Runge-Kutta method, and iteratively looping multiple time steps until the termination dynamic features are obtained, and obtaining the augmented view encoding corresponding to the termination dynamic features according to the readout module;

[0008] determine a global view encoding corresponding to the original text propagation graph according to the trained GDN-ODE network and the original node features; determine a graph-level encoding of each echo chamber subgraph and an echo chamber encoding set according to the trained GDN-ODE network and the set number of echo chambers;

[0009] The global view encoding and the echo chamber encoding are spliced in the node dimension to obtain a final propagation graph encoding, and the final propagation graph encoding is input into an MLP classifier to obtain a category label of the original text associated with the original text propagation graph.

[0010] An embodiment of the present application provides a rumor detection device based on a graph neural diffusion network, comprising:

[0011] A first determining unit is configured to determine original node features of an original text propagation graph, perform data enhancement on the original text propagation graph to obtain an enhanced view and a set number of echo chambers.

[0012] A first obtaining unit is configured to construct an initial GDN-ODE network comprising a mapping module, a diffusion module and a readout module; determine an initial dynamic feature X dyn (0) corresponding to the original node features according to the mapping module; discretize a time interval of a diffusion process into a plurality of time steps, determine a current dynamic feature of a second time step based on the initial dynamic feature, the diffusion module and an adaptive Runge-Kutta method in a first time step, and iteratively loop in a plurality of time steps for a plurality of rounds until a terminal dynamic feature is obtained, and obtain an enhanced view encoding corresponding to the terminal dynamic feature according to the readout module.

[0013] A second determining unit is configured to determine a global view encoding corresponding to the original text propagation graph according to the trained GDN-ODE network and the original node features; and determine a graph-level encoding of each echo chamber subgraph and an echo chamber encoding set according to the trained GDN-ODE network and the set number of echo chambers.

[0014] A second obtaining unit is configured to splice the global view encoding and the echo chamber encoding in the node dimension to obtain a final propagation graph encoding, and input the final propagation graph encoding into an MLP classifier to obtain a category label of the original text associated with the original text propagation graph.

[0015] An embodiment of the present application provides a computer device, which comprises a memory and a processor, the memory stores a computer program, and the computer program is executed by the processor to enable the processor to execute any one of the rumor detection methods based on the graph neural diffusion network.

[0016] The embodiment of the present application provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to make the processor execute the rumor detection method based on the graph neural diffusion network.

[0017] The embodiment of the present application provides a rumor detection method and device based on a graph neural diffusion network, which comprises the following steps: determining an original node feature of an original text propagation graph, performing data enhancement on the original text propagation graph to obtain an enhanced view and a set number of echo chambers, constructing an initial GDN-ODE network comprising a mapping module, a diffusion module and a readout module, determining an initial dynamic feature corresponding to the original node feature according to the mapping module, discretizing a time interval of a diffusion process into a plurality of time steps, determining a current dynamic feature of a second time step based on the initial dynamic feature, the diffusion module and an adaptive Runge-Kutta method in the first time step, and iteratively circulating a plurality of time steps in multiple rounds until a terminal dynamic feature is obtained, and obtaining an enhanced view code corresponding to the terminal dynamic feature according to the readout module; determining a global view code corresponding to the original text propagation graph according to the trained GDN-ODE network and the original node feature; determining a graph-level code of each echo chamber subgraph and an echo chamber code set according to the trained GDN-ODE network and the set number of echo chambers; the global view code and the echo chamber code are spliced according to the node dimension to obtain a final propagation graph code, the final propagation graph code is input into an MLP classifier to obtain a category label of the original text associated with the original text propagation graph. The method constructs a text propagation graph, obtains an original text propagation graph, performs data enhancement on the original text propagation graph to obtain three enhanced views, and then extracts echo chambers of the original text propagation graph based on an edge betweenness algorithm. Next, the enhanced graph and the original graph are used to train a global graph encoder model, the echo chambers are used to train an echo chamber encoder model, and the encoder model is used to detect the authenticity of the text. On the one hand, the method breaks through the limitations of traditional methods in modeling long-distance dependencies, can extract multi-scale features, and thus more accurately identifies false texts. On the other hand, the method overcomes the shortcomings of previous methods that ignore the dense subgraph features formed by echo chamber effects, so that the model is more suitable for false text detection and more suitable for actual application scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0019] Figure 1 A rumor detection method based on a graph neural diffusion network provided by an embodiment of the present application;

[0020] Figure 2 An information system II parameter analysis experimental result schematic view provided by an embodiment of the present application;

[0021] Figure 3 A rumor detection device structure schematic view provided by an embodiment of the present application. DETAILED DESCRIPTION

[0022] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative labor fall within the protection scope of the present application.

[0023] Figure 1 A rumor detection method based on a graph neural diffusion network provided by an embodiment of the present application; the following takes Figure 1 as an example to introduce in detail a rumor detection method based on a graph neural diffusion network provided by an embodiment of the present application, as shown in Figure 1 , the method comprises the following steps:

[0024] Step 101, determining an original node feature of an original text propagation graph, performing data enhancement on the original text propagation graph to obtain an enhanced view and a set number of echo chambers;

[0025] Step 102, constructing an initial GDN-ODE network comprising a mapping module, a diffusion module and a readout module; determining an initial dynamic feature corresponding to the original node feature according to the mapping module; discretizing a time interval of a diffusion process into a plurality of time steps, in the first time step, determining a current dynamic feature of the second time step based on the initial dynamic feature, the diffusion module and an adaptive Runge-Kutta method, iteratively circulating in a plurality of time steps for a plurality of rounds until a terminal dynamic feature is obtained, and obtaining an enhanced view code corresponding to the terminal dynamic feature according to the readout module;

[0026] Step 103, determining a global view code corresponding to the original text propagation graph according to the trained GDN-ODE network and the original node feature; determining a graph-level code of each echo chamber subgraph and an echo chamber code set according to the trained GDN-ODE network and the set number of echo chambers;

[0027] In step 104, the global view code and the echo chamber code are spliced in the node dimension to obtain a final propagation graph code, and the final propagation graph code is input into an MLP classifier to obtain a category label of the original text associated with the original text propagation graph.

[0028] In the embodiment of the application, a text propagation graph is constructed to obtain an original text propagation graph, data augmentation is performed on the original text propagation graph to obtain an enhanced view, and then an echo chamber of the original text propagation graph is extracted based on an edge betweenness algorithm. Next, an initial GDN-ODE network is trained using original node features, and the trained GDN-ODE network is used to detect the authenticity of the text. This method breaks through the limitations of traditional methods in modeling long-distance dependencies, can extract multi-scale features, and thus more accurately identifies untruthful texts. On the other hand, it overcomes the shortcomings of previous methods that ignore the dense subgraph features formed by the echo chamber effect, making the model more suitable for untruthful text detection and more suitable for actual application scenarios.

[0029] In the false information detection method based on the graph neural diffusion network, the original text propagation graph is a core data carrier, which abstracts the propagation process of the text on social media, forums and other platforms into a structured graph model. Through the triple form of "node-edge-feature", the propagation participants, propagation relationships and associated attributes of the text are quantitatively presented, providing a basis for subsequent extraction of long-distance dependencies, echo chambers and other false information propagation features.

[0030] In the embodiment of the application, the original text propagation graph is a structured representation of "participants-relationships" formed in the propagation process of text information. The essence is to use graph theory tools to mathematically model unstructured propagation behavior - converting "entities" (such as users, text nodes) in the propagation chain into "nodes" of the graph, "interaction behaviors between entities" (such as forwarding, commenting, quoting) into "edges" of the graph, and "attribute information of entities" (such as user credibility, text semantics) into "node features", finally forming structured data that can be processed by graph neural networks (such as GDN-ODE). For example, the propagation process of a "heavy rain warning" text is abstracted into a 4-node propagation graph (G=(V,E,X raw ).

[0031] In step 101, for the obtained original text propagation graph, the node set, edge set and original node matrix (original node features) included in the original text propagation graph need to be determined.

[0032] For example, an original text propagation graph G=(V,E,X raw ) includes a node set, an edge set and original node features as follows:

[0033] Node set V: V = {A, B, C, D} (A is an information publisher, and B, C, and D are propagators).

[0034] Edge set E: E = {(A, B), (A, C), (B, C), (B, D), (C, D)} (an edge represents an information forwarding / comment relationship).

[0035] Original node feature X raw : X raw (A) = [1.0, 0.8], X raw (B) = [0.6, 0.9], X raw (C) = [0.7, 0.7], X raw (D) = [0.2, 0.3].

[0036] In the embodiment of the application, after the original text propagation graph is obtained, data enhancement processing can be performed on the original text propagation graph to obtain an enhanced view. Since the enhanced view includes a feature view, a node view, and an edge view, three different data enhancement processes need to be performed on the original text propagation graph, specifically as follows:

[0037] Randomly mask the node features of the original text propagation graph, that is, randomly select nodes with a proportion of p f , and replace the feature vectors thereof with learnable mask labels x [MASK] to obtain a feature view.

[0038] Randomly discard a node set V n with a proportion of p drop from the original text propagation graph, and retain the edges between the remaining nodes to obtain a node view.

[0039] Randomly discard an edge set E e with a proportion of p drop from the original text propagation graph, and keep the node set complete to obtain an edge view.

[0040] In the embodiment of the application, the feature view, the node view, and the edge view are represented by the following formulas:

[0041] G f = (V f ,E,X fv ) (1-1)

[0042] G n = (V n ,E n ,X nv ) (1-2)

[0043] G e = (V,E e ,X ev(1-3)

[0044] wherein G f denotes a feature view, V f denotes a node set of the feature view, E denotes an edge set, X fv denotes a node feature matrix of the feature view, x [MASK] denotes a learnable mask label vector, p f denotes a feature mask ratio; G n denotes a node view, V n denotes a node set of the node view, E n denotes an edge set of the node view, X nv denotes a node feature matrix of the node view, p n denotes a node dropping ratio, V drop denotes a node set to be dropped; G e denotes an edge view, X ev denotes a node feature matrix of the edge view, V, E e , p e denotes an edge dropping ratio, E drop denotes an edge set to be dropped.

[0045] In actual application, the original text propagation graph is affected by the echo chamber effect, and forms a structure feature of dense subgraph, in order to mine unique structure features, an algorithm based on edge betweenness is adopted to extract the echo chamber in the embodiment of the application.

[0046] Firstly, the edge betweenness of each edge is determined according to the shortest path between all node pairs included in the original text propagation graph.

[0047] Then, the edge with the maximum edge betweenness is removed from the original text propagation graph, and the edge set included in the original text propagation graph is updated; if the original text propagation graph has a split new subgraph after being updated, the new subgraph is added to the echo chamber set;

[0048] Finally, when the number of echo chambers included in the echo chamber set reaches a set value, the updating of the edge set included in the original text propagation graph is stopped.

[0049] wherein the edge betweenness of each edge is determined by the following formula:

[0050]

[0051] wherein g(e) denotes the edge betweenness of edge e, reflecting the connectivity importance of the edge in the propagation graph, σ st denotes the total number of shortest paths from node s to node t, σ st (e) the number of shortest paths from node s to node t through edge e, s and t denote any two different nodes in the propagation graph.

[0052] For example, extract echo chamber (1st iteration: edge betweenness iteration split)

[0053] Iteration initialization: input original text propagation graph G, set target echo chamber number m = 2, G cur = G, echo chamber set C = 0.

[0054] 1st iteration:

[0055] 1) Calculate edge betweenness: according to g(A,B) = 3, g(A,C) = 3, g(B,C) = 2, g(B,D) = 1, g(C,D) = 1 (the larger the edge betweenness, the stronger the connectivity);

[0056] 2) Remove the largest edge (A,B) and (A,C) (the largest), update G cur Edge set {(B,C), (B,D), (C,D)};

[0057] 3) Check split: G cur No split, C is still empty.

[0058] 2nd iteration:

[0059] 1) Recalculate G cur Edge betweenness: g(B,C) = 2, g(B,D) = 1, g(C,D) = 1;

[0060] 2) Remove edge (B,C), update G cur Edge set {(B,D), (C,D)};

[0061] 3) Check split: G cur Split into subgraphs {B,D} and {C,D}, but {B,D,C} is still connected, continue iteration.

[0062] 3rd iteration:

[0063] 1) Calculate G cur Edge betweenness: g(B,D) = 1, g(C,D) = 1;

[0064] 2) Remove edge (B,D), G cur Split into {A} and {B,C,D}, add two dense subgraphs to C, |C| = 2 at this time;

[0065] 3) Iteration termination: output echo chamber set C = {C1 = {A}, C2 = {B,C,D}}.

[0066] In the embodiments of the present application, in order to distinguish the GDN-ODE network before training, the GDN-ODE network during training and the GDN-ODE network after training, in the text, the initially constructed GDN-ODE network is referred to as the initial GDN-ODE network, the GDN-ODE network during training is referred to as the first GDN-ODE network, the second GDN-ODE network and the like in turn, and the GDN-ODE network after training is referred to as the trained GDN-ODE network.

[0067] In step 102, the initial GDN-ODE network is constructed, and the initial GDN-ODE network includes a mapping module, a diffusion module and a reading module, each of which includes corresponding parameters.

[0068] The initial GDN-ODE network is trained, specifically including:

[0069] Firstly, the mapping module included in the initial GDN-ODE network is used to map the original node features included in the original text propagation graph to obtain initial dynamic features, and it should be noted that the mapping module is composed of a linear layer, a batch normalization layer, a ReLU activation function layer and a Dropout layer.

[0070]

[0071] Wherein, X raw represents the original node feature matrix, X dyn represents the dynamic node domain feature matrix, X dyn (0) represents the initial dynamic feature matrix, X dyn (t) represents the dynamic feature matrix at time t, X dyn (T) represents the terminal dynamic feature matrix, represents the node feature mapping function.

[0072] Then the node feature state is updated through the diffusion module, specifically, the time interval of the diffusion process is discretized into multiple time steps, and at the first time step, based on the initial dynamic features and the node domain and edge domain feature matrices included in the diffusion module, the initial gradient operator, the initial divergence operator and the feature change rate at the first time are determined in turn. Further, the feature change rate at the first time is based on the adaptive Runge-Kutta method to obtain the current dynamic features at the second time step.

[0073] In the above method, the initial gradient operator, the initial divergence operator and the feature change rate at the first time are determined by the following formulas respectively:

[0074]

[0075]

[0076]

[0077] G = diag(a(X raw,i ,X raw,j )) = diag(A ij ) ∈ R |E|×|E| (2-5)

[0078]

[0079] The update equation of node feature state is

[0080]

[0081] wherein, represents the edge domain feature matrix, describing the feature difference between nodes, represents the gradient operator element corresponding to the edge (i,j), X dyn (i) represents the dynamic feature vector of node i, X dyn (j) represents the dynamic feature vector of node j, i,j represents a pair of nodes in the original text propagation graph; div(X dyn ) represents the divergence operator result, (div(X dyn ) i represents the divergence value of node i, that is, the sum of the feature flow intensity of all adjacent edges of node i, A ij represents the weight of edge (i,j), ∑ j:(i,j)∈E represents the sum of all nodes j connected to node i through edges; represents the feature change rate at time t, that is, the evolution speed of dynamic feature with time, X dyn (t) represents the dynamic feature matrix at time t, G(X dyn (t),t) represents the diffusion rate matrix, which changes with time t and dynamic feature X dyn (t), represents the gradient operator at time t, G represents the diffusion rate matrix, diag(·) represents the diagonal matrix constructor function, taking the input vector as the diagonal element and the rest as 0, a(X raw,i ,X raw,j ) represents the dot product attention function, which calculates the original feature similarity between nodes i and j, X raw,i represents the original feature vector of the i-th node, X raw,j represents the original feature vector of the j-th node, d represents the dimension of the original node feature, represents the scaling factor, which is used to avoid the gradient instability caused by the too large dot product result when the feature dimension is too high, X dyn (T) represents the dynamic feature matrix at the termination time T, X dyn(0) denotes the dynamic feature matrix at the initial time t = 0, T denotes the termination time of the diffusion process, and t denotes the diffusion process time variable, denotes the gradient operator, denotes the real number field.

[0082] It should be noted that in the embodiments of the present application, the adaptive Runge-Kutta method is used to solve the ordinary differential equation, which can effectively improve the stability of the numerical value in the long-distance diffusion process.

[0083] It should be noted that the above method is for the first time step in the plurality of time steps, and the current dynamic feature of the second time step is obtained in the first time step according to the above algorithm, so in the second time step, the current dynamic feature of the third time step can be obtained based on the current dynamic feature of the second time step, the diffusion module and the adaptive Runge-Kutta method, and the same is true for the subsequent time steps, until the plurality of time steps included in a diffusion process are all executed.

[0084] In the embodiments of the present application, if a GDN-ODE network is to be obtained, multiple rounds of circulation are required, and each round of circulation corresponds to the plurality of time steps included in a diffusion process, that is, if a termination dynamic feature is to be obtained, multiple rounds of circulation are required.

[0085] In the embodiments of the present application, the plurality of time steps included in the first round of circulation are iterated, and the first termination dynamic feature is obtained, and further, according to the first termination dynamic feature and the readout module, the first enhanced view encoding corresponding to the first termination dynamic feature is obtained.

[0086] For example:

[0087] Step 1: Construct an initial GDN-ODE network:

[0088] 1. Mapping module : Initialize linear layer weights The batch normalization parameters γ and β are randomly initialized.

[0089] 2. Diffusion module: preset gradient operator: : divergence operator

[0090] 3. Readout module : Initialize linear layer weights The Dropout probability is set to 0.5.

[0091] Step 2: Preprocessing and initial encoding verification:

[0092] 1. The mapping module outputs the initial dynamic feature:

[0093]

[0094] 2. The diffusion rate matrix G is calculated:

[0095] 1) The dot product attention formula is: When the original node feature vectors of node A and node B are respectively: X raw (A) = [1.0, 0.8], X raw (B) = [0.6, 0.9], and d = 2, then: raw,i · X raw,j = (1.0 x 0.6) + (0.8 x 0.9) = 1.32. Substituting the above values into the dot product attention formula, we get: A AB = 0.93.

[0096] Similarly, we can get in turn: A AC = 0.89, A BC = 0.74, A BD = 0.27, A CD = 0.25

[0097] 2) The diffusion rate matrix is: G = diag(a(X raw,i , X raw,j )), and the diffusion rate matrix is a diagonal matrix, whose diagonal elements correspond to the dot product attention weight of each edge (i.e. A ij ):

[0098] 3. Iterative solution: using the adaptive Runge-Kutta method, discretize [0, 1.0] into 5 time steps (t = 0, 0.2, 0.4, 0.8, 1.0), and iteratively calculate the current dynamic feature X dyn (T) at each time step.

[0099] At t = 0.2,

[0100] At t = 0.4,

[0101] Iterate in turn to t = 1.0, output

[0102] 4. The readout module: after the batch normalization layer, the ReLU activation function layer, the linear layer and the batch normalization layer, the first enhanced view encoding is obtained:

[0103] Further, after obtaining the first enhanced view encoding, it is also necessary to determine the latent space encoding of the three views, the reconstruction loss and the global multi-view contrast loss according to the initial GDN-ODE network and the three enhanced views determined in step 101.

[0104] Specifically, the initial GDN-ODE network is used to perform an encoding operation on the feature view, the node view and the edge view, to obtain the latent space encoding of the feature view, the latent space encoding of the node view and the latent space encoding of the edge view. In the embodiment of the application, the node features of the feature view, the node view and the edge view are obtained through the mapping module, the diffusion module and the readout module of the initial GDN-ODE network, to obtain the latent space encoding of the feature view, the latent space encoding of the node view and the latent space encoding of the edge view. The features can fully reflect the overall propagation graph, including long-distance dependencies.

[0105] Further, the initial GDN-ODE network is used to perform a decoding operation on the latent space encoding of the feature view, to obtain a reconstructed feature matrix. Then, based on the original node features and the reconstructed feature vectors included in the reconstructed feature matrix, the reconstruction loss shown below is determined:

[0106]

[0107] Wherein, L rec (g) represents the reconstruction loss, the index γ increases the punishment degree for difficult samples, x rec represents the reconstructed feature matrix, x rec,i represents the reconstructed feature vector of the i-th node, v i ∈V f , v i represents the i-th node in the feature view, x rec,i represents the reconstructed feature vector of the node v i , x raw,i represents the feature vector of the node v i .

[0108] Further, the latent space encoding of the node view, the latent space encoding of the edge view and the latent space encoding of the feature view are normalized to obtain the normalized encoding vector of the node view, the normalized encoding vector of the edge view and the normalized encoding vector of the feature view:

[0109] H n,v = Norm (Z n ) (2-9)

[0110] H e,v = Norm (Z e ) (2-10)

[0111] H f,v= Norm(Z f ) (2-11)

[0112] where H n,v denotes the normalized encoding vector of the node view, H e,v denotes the normalized encoding vector of the edge view, and H f,v denotes the normalized encoding vector of the feature view, Z n denotes the latent space encoding matrix of the node view, Z e denotes the latent space encoding matrix of the edge view, and Z f denotes the latent space encoding matrix of the feature view.

[0113] Further, a global multi-view contrastive loss is constructed according to the normalized encoding vector of the node view, the normalized encoding vector of the edge view, and the normalized encoding vector of the feature view:

[0114]

[0115] wherein, denotes the global multi-view contrastive loss, the encoding distance of the same view is pulled closer, and the encoding distance of the different view is pushed farther away, S intra denotes the intra-view negative sample similarity sum, i.e., the similarity sum between the node view encodings of different graphs, S inter denotes the inter-view negative sample similarity sum, i.e., the similarity sum between the node view and feature view encodings of different graphs, denotes the negative sample encoding vector of the node view, g - denotes the negative sample propagation graph set, i.e., the propagation graph set that is not currently calculated, denotes the sum of the node view encodings of all negative sample propagation graphs, denotes the negative sample encoding vector of the node view, denotes the negative sample encoding vector of the feature view, denotes the sum of the feature view encodings of all negative sample propagation graphs. In the embodiments of the present application, the intra-view negative sample represents the similarity between the node views of different graphs, and the inter-view negative sample represents the similarity between the node view and feature view of different graphs.

[0116] In the above example, if the core parameters of the initial GDN-ODE network are:

[0117] 1) Mapping module : γ = 1, β = 0.

[0118] 2) Diffusion module: preset gradient operator: divergence operator

[0119] 3) readout module : initialize linear layer weights Dropout probability is set to 0.5.

[0120] Specifically, the initial GDN-ODE is performed on the feature view, node view and edge view:

[0121] Step 1. Taking the feature view, node view and edge view as an example, the processing flow of the initial GDN-ODE for each view is consistent (mapping → diffusion → readout):

[0122] 1) feature view processing

[0123] Original features: (node C features are masked);

[0124] Mapping module: converted into initial dynamic features

[0125] Diffusion module: ODE numerical solution

[0126] Readout module: generate latent space encoding of the feature view

[0127] 2) After processing the node view and the edge view, the latent space encoding of the node view and the latent space encoding of the edge view are obtained respectively:

[0128] Step 2. Reconstruction loss L rec (g) calculation:

[0129] Formula:

[0130] Step 3. Calculation of global multi-view contrastive loss

[0131] Formula:

[0132] Example substitution:

[0133] Similarity of homographic view encoding:

[0134] exp(1.3)≈3.6693;

[0135] S intra (view intra-negative sample and): take other node encoding similarity and, ≈2.0;

[0136] S inter ​(inter-view negative samples and): take other graph encoding similarities, ≈ 1.5;

[0137] Substitute the formula:

[0138]

[0139] Further, the graph-level encoding of the echo chamber subgraph corresponding to the initial GDN-ODE network and the set number of echo chambers determined in step 101 also needs to be determined.

[0140] The initial GDN-ODE network is used to encode the echo chambers determined in step 101 to reach the set value, and then the graph-level encoding of each echo chamber font is obtained through the pooling operation. Then, the contrast loss of all echo chambers included in each original text propagation graph is calculated, that is, the similarity between the echo chamber pairs included in each original text propagation graph is calculated to measure:

[0141]

[0142] Wherein, represents the echo chamber contrast loss, which enhances the feature difference of different echo chambers in the same propagation graph, n represents the number of propagation graphs, m is the number of echo chambers, and the difference between each echo chamber is enhanced through the echo chamber contrast loss, highlighting its unique features, i represents the index of the propagation graph, and j1 represents the index of the echo chamber, represents the graph-level encoding vector of the j1th echo chamber in the ith propagation graph.

[0143] According to the above example, the initial GDN-ODE network encoding is performed on the echo chamber subgraph:

[0144] Taking the echo chambers C1={A} and C2={B, C, D} in the embodiment as an example, each step is explained in detail:

[0145] Step 1: Obtain the original features of the echo chamber subgraph. The echo chamber is a subgraph of the original text propagation graph, and its node features directly reuse the node features of the original text propagation graph (i.e. the input features of the enhanced view encoding, not the output of the enhanced view encoding):

[0146] 1) Echo chamber C1={A}, containing only node A, node feature X raw (A)=[1.0,0.9];

[0147] 2) Echo chamber C2={B, C, D}, containing nodes B, C, and D; node features: X raw (B)=[0.6,0.9], X raw (C)=[0.7,0.7], X raw (D)=[0.2,0.3];

[0148] Step 2: Perform initial GDN-ODE network on echo chamber subgraph:

[0149] For each echo chamber subgraph, perform the encoding process of "Mapping Module → Diffusion Module (time step iteration) → Readout Module" separately to obtain the node-level dynamic features of the echo chamber subgraph:

[0150] 1) Echo chamber C1 = {A}

[0151] Mapping Module: Convert the node feature X raw (A) = [1.0, 0.9] into initial dynamic feature X dyn,C1 (0) = [1.0, 0.9, 0.7] (in the example, the output of the mapping module is [1.2, 0.9, 0.7]);

[0152] Diffusion Module: Perform diffusion on the C1 subgraph for "time steps (t = 0 → 0.2 →... → 1.0)" to obtain the dynamic feature X dyn,C1 (1.0) = [1.2, 0.9, 0.7] (since C1 has only one node, there are no diffusion edges, and the feature is consistent with the initial dynamic feature);

[0153] Readout Module: Perform "batch normalization + ReLU + linear layer" on X dyn,C1 (1.0) to output the node-level encoding (since there is only one node, the view encoding is consistent with X dyn,C1 (1.0)).

[0154] 2) Echo chamber C2 = {B, C, D}

[0155] Mapping Module: Convert the original features of nodes B, C, and D into initial dynamic features to obtain (Each row corresponds to node B, node C, and node D);

[0156] Diffusion Module: Perform time step iteration on the C2 subgraph to obtain the dynamic feature X (node B, node C feature convergence, reflecting the aggregation of internal features of the echo chamber);

[0157] Readout Module: Perform "batch normalization + ReLU + linear layer" on X dyn,C2 (1.0) to output the node-level encoding (consistent with X dyn,C2 (1.0)).

[0158] Step 3: Pooling operation - compress node-level encoding into graph-level encoding

[0159] Echo chamber is a "subgraph", and the encoding of all its nodes needs to be compressed into a single vector (graph-level encoding) representing the features of the entire echo chamber. In the example, "mean pooling" is used (taking the average of node features dimension by dimension):

[0160] 1) Pooling of echo chamber C1 = {A}, since C1 has only 1 node, directly take its node-level encoding as graph-level encoding: z 1,1 = [1.2, 0.9, 0.7].

[0161] 2) Pooling of echo chamber C2 = {B, C, D}, take average of the 3 node encodings of C2 dimension by dimension:

[0162] 1st dimension: (1.0 + 1.0 + 0.6) ÷ 3 ≈ 0.87;

[0163] 2nd dimension: (0.85 + 0.85 + 0.7) ÷ 3 ≈ 0.8;

[0164] 3rd dimension: (0.65 + 0.65 + 0.5) ÷ 3 ≈ 0.6;

[0165] Finally, get the graph-level encoding: z 1,2 = [0.87, 0.8, 0.6].

[0166] Step 4: Constructing the echo chamber encoding set

[0167] Combine the graph-level encodings of all echo chambers to get the echo chamber encoding set Z_c:

[0168] Z c = [[1.2, 0.9, 0.7], [0.87, 0.8, 0.6]] (corresponding to the graph-level encodings of C1 and C2).

[0169] Step 5: Example of determining the echo chamber contrast loss and feature smoothing loss

[0170] 1) Calculation of echo chamber contrast loss Formula:

[0171] (n is the number of propagation graphs, m is the number of echo chambers, sim is the cosine similarity)

[0172] n = 1 (only 1 propagation graph), m = 2 (echo chambers m = 2, m = 2);

[0173] Echo chamber encoding: z 1,1 = [1.2, 0.9, 0.7], z 1,2 = [0.87, 0.8, 0.6];

[0174] Calculate the cosine similarity:

[0175]

[0176]

[0177] Substitute the formula:​​

[0178]

[0179] 2) Feature smoothing loss L smo Calculation

[0180] formula: (|E| represents the number of edges, h) i (for node encoding)

[0181] Example Substitution: Edge set E = {(A,B),(A,C),(B,C),(B,D),(C,D)}, |E| = 5; Node encoding: h A =[1.1,0.8]、h B =h C =[0.9,0.7]、h D =[0.7,0.6].

[0182] Calculate the L2 norm of the node encoding for each edge:

[0183] (A, B):

[0184] (A, C): ||h A -h C || = 0.2236 (same as (A, B));

[0185] (B, C): ||h B -h C || =;

[0186] (B, D):

[0187] (C, D): ||h C -h D || = 0.2236;

[0188] Substitute into the formula:

[0189] In this embodiment of the invention, based on the above method, while obtaining the first enhanced view encoding, the latent space encoding of the three views, the graph-level encoding of the echo chamber subgraph, and the echo chamber encoding set based on the initial GDN-ODE network, the reconstruction loss, the global multi-view contrast loss, and the echo chamber contrast loss are also determined.

[0190] Finally, the global graph encoder and echo chamber encoder models are trained by combining the feature smoothing loss function and the weight decay loss function. The feature smoothing regularization term penalizes large differences between adjacent nodes.

[0191]

[0192] wherein, L smo represents a feature smoothness loss, which penalizes large differences in the encoding of neighboring nodes and avoids overfitting, E represents an edge in the original text propagation graph, h i represents the final encoding vector of node i, h j represents the final encoding vector of node j.

[0193] Finally, the total loss is determined by the following formula:

[0194]

[0195] wherein, L represents the total loss, L rec represents a reconstruction loss, represents a multi-view contrastive loss, represents a reverberation contrastive loss, L smo represents a feature smoothness loss, L wdr represents a weight-decay regularization loss, λ1, λ2, λ3, λ4 represent hyperparameters that control the balance of each loss function.

[0196] In the embodiment of the application, after the multiple time step iterations included in the first round of iterations end, the first enhanced view encoding, the latent space encoding of the three views, the reverberation encoding set reconstruction loss, the global multi-view contrastive loss, the reverberation contrastive loss and the feature smoothness loss can be obtained in turn. At the same time, the total loss can be determined for the first time.

[0197] It should be noted that the sign of the end of the first round of iterations in the multiple rounds of iterations included in the embodiment of the application is the first total loss. After the end of the first round of iterations, the first GDN-ODE network can also be obtained. Here, the first GDN-ODE network represents the result obtained after the initial GDN-ODE network undergoes the first round of iterations. In the embodiment of the application, after the first total loss is obtained, it is also necessary to determine whether the cycle can be stopped according to the first total loss and the convergence threshold.

[0198] If the first total loss does not satisfy the convergence threshold, the second round of iterations needs to be performed, that is, the above step 102 is repeatedly executed until the Nth total loss obtained after the Nth round of iterations satisfies the convergence threshold. When the Nth total loss satisfies the convergence threshold, the Nth GDN-ODE network obtained in this round of iterations can be determined as the trained GDN-ODE network.

[0199] In step 103, according to the trained GDN-ODE network and the original node features, the global view encoding corresponding to the original text propagation graph is determined, and according to the trained GDN-ODE network and a set number of reverberations, the graph-level encoding of each reverberation subgraph and the reverberation encoding set are determined.

[0200] The part can refer to the above step 102, based on the initial GDN-ODE network, the first enhanced view corresponding to the original text propagation graph is obtained, and the graph-level coding of each echo chamber subgraph and the echo chamber coding set are determined based on the initial GDN-ODE network, which will not be repeated here.

[0201] In step 104, the final propagation graph coding is obtained by splicing the global view coding and the echo chamber coding in the node dimension, the final propagation graph coding is input into the MLP classifier, and the category label of the original text associated with the original text propagation graph is obtained.

[0202] Specifically, the text structure graph to be detected, the three enhanced views and the echo chamber are input into the rumor detection model, the global view coding and the echo chamber coding set are obtained, and the final propagation graph coding vector is obtained by splicing the global view coding and the echo chamber coding set:

[0203] H = Cat(H g ,H c )

[0204] Wherein, H represents the final propagation graph coding vector, which fuses global features and echo chamber features, H g is the global graph coding representation of the original text propagation graph to be detected, H c is the corresponding echo chamber coding representation; Cat(·) represents a splicing function.

[0205] S32: input the propagation graph coding vector into the MLP classifier, so as to obtain the true or false category y of the text corresponding to the text structure graph to be detected:

[0206] y = MLP(H)

[0207] Wherein, y represents the category label, i.e. the authenticity judgment result of the original text, MLP(·) represents a multi-layer perception machine classifier, receives the final propagation graph coding H, outputs a category probability, and H represents the final propagation graph coding vector.

[0208] Table 1 is the detailed results of the evaluation indicators of the method (EC-GDN) provided by the embodiment of the present application on two real data sets and other comparative methods for detecting false texts. Among them, Table 1(a) is a comparison chart of the detection accuracy, precision, recall and F1 score of information systems one and two on the English data set and other five unsupervised methods; Table 1(b) is a comparison chart of the detection accuracy, precision, recall and F1 score of information system three on the Chinese data set and other four unsupervised methods; Table 2 is a comparison chart of the detection accuracy and F1 score of the two information systems and other five supervised methods. It can be seen that on the English and Chinese data sets, EC-GDN has significant improvement in the four indicators. Compared with the model GAMC which only uses the traditional GIN network to encode the whole propagation graph, the detection accuracy of the three information systems is improved by 7.88%, 3.07% and 9.86% respectively. This shows that EC-GDN can aggregate multi-scale features by using continuous-time diffusion, and the explicit echo chamber extraction and echo chamber contrast learning highlight the subgraph differences, effectively improving the accuracy of rumor identification. In addition, compared with the existing methods based on text content, the method based on the original text propagation graph can achieve better detection effect, indicating that the propagation structure contains more features that can be used to identify false texts. At the same time, supervised methods are usually superior to unsupervised methods, because label information can guide the model to learn more discriminative false text features. While the EC-GDN method proposed in the present application maintains a high detection accuracy, it eliminates the dependence on labeled data, thereby having higher flexibility and applicability in actual application scenarios.

[0209] Table 1(a)

[0210]

[0211] Table 1(b)

[0212]

[0213] Table 2

[0214]

[0215] Table 3 is a comparison diagram of the results of ablation experiments of EC-GDN on information system two. The performance of the overall model is studied by removing each component. EC-GDN-w / o-glo represents removing the global graph encoder; EC-GDN-w / o-cha represents removing the echo chamber encoder; EC-GDN-w / o-GDN represents removing the graph neural diffusion network; EC-GDN-w / o-reg, EC-GDN-w / o-rec, and EC-GDN-w / o-con represent removing the regularization loss term, the reconstruction loss term, and the contrastive loss term, respectively. It can be seen that removing any module will cause the model performance to decline to varying degrees. The graph neural diffusion network has the greatest impact on the model, and the global graph encoder also plays a key role. Both parts play a crucial role in the extraction of long-distance dependencies. The remaining parts also have a certain effect on the improvement of the model performance.

[0216] Table 3

[0217]

[0218] Figure 2 The diagram shows the effect of the four weight coefficients of the loss function provided by the embodiments of the present application on the fake text recognition rate. When λ1=0.01, λ2=1.0, λ3=10 -5 , and λ4=10 -5 , the model performance is optimal. When the coefficients are too large or too small, the model performance will decline, because the balance between the components is destroyed. This also illustrates the importance of balancing the components, which can improve the stability of the training and the fake text recognition accuracy of the model.

[0219] In summary, the embodiment of the present application provides a rumor detection method and device based on a graph neural diffusion network, which constructs a text propagation graph to obtain an original text propagation graph, and uses the original text propagation graph to perform data augmentation to obtain three enhanced views, and then extracts an echo chamber of the original text propagation graph through an edge betweenness-based algorithm; next, the enhanced graph and the original graph are used to train a global graph encoder model, and the echo chamber is used to train an echo chamber encoder model, and the encoder model is used to detect the authenticity of the text; the method can automatically detect false text information in an information system without the need for a large labeled data set; the method considers the original text content and the propagation structure to detect false text information, improving the accuracy of text information detection; the method effectively extracts long-distance dependency relationships of the original text propagation graph by applying a graph neural diffusion network based on an ordinary differential equation, solving the oversmoothing problem of traditional graph neural networks; on the one hand, the method breaks through the limitations of traditional methods in modeling long-distance dependency relationships, can extract multi-scale features, and thus more accurately identifies false text; on the other hand, the method overcomes the shortcomings of previous methods that ignore the dense subgraph features formed by the echo chamber effect, so that the model is more suitable for false text detection and more suitable for actual application scenarios. The problem of not fully considering the uniqueness of the text propagation structure is solved, and the need for a large amount of labeled data set is eliminated, which makes the false text information detection method provided by the embodiment of the present application have practical guiding significance.

[0220] Based on the same inventive concept, the embodiment of the present application provides a rumor detection device based on a graph neural diffusion network. Since the principle of solving the technical problem of the device is similar to that of a rumor detection method based on a graph neural diffusion network, the implementation of the device can be referred to the implementation of the method, and the repeated parts will not be described again.

[0221] As shown in Figure 3 , the device includes a first determination unit 301, a first obtaining unit 302, a second determination unit 303, and a second obtaining unit 304.

[0222] The first determination unit 301 is configured to determine an original node feature of an original text propagation graph, perform data augmentation on the original text propagation graph to obtain an enhanced view, and set a number of echo chambers.

[0223] The first obtaining unit 302 is configured to construct an initial GDN-ODE network including a mapping module, a diffusion module, and a readout module; determine an initial dynamic feature X dyn(0); discretize a time interval of the diffusion process into a plurality of time steps, in a first time step, determine a current dynamic feature of a second time step based on the initial dynamic feature, the diffusion module and an adaptive Runge-Kutta method, and iteratively loop in a plurality of time steps until a terminal dynamic feature is obtained, and obtain an enhanced view code corresponding to the terminal dynamic feature according to the readout module;

[0224] The second determining unit 303 is configured to determine a global view code corresponding to the original text propagation graph according to the trained GDN-ODE network and the original node feature; and determine a graph-level code of each echo chamber subgraph and an echo chamber code set according to the trained GDN-ODE network and a set number of echo chambers.

[0225] The second obtaining unit 304 is configured to splice the global view code and the echo chamber code in a node dimension to obtain a final propagation graph code, and input the final propagation graph code into an MLP classifier to obtain a category label of the original text associated with the original text propagation graph.

[0226] It should be understood that the above-mentioned rumor detection device based on the graph neural diffusion network includes units only for logical division according to the functions implemented by the device, and in actual application, the above-mentioned units can be stacked or split. And the rumor detection device based on the graph neural diffusion network provided in the embodiment realizes the functions corresponding to the rumor detection method based on the graph neural diffusion network provided in the above-mentioned embodiment. For more detailed processing procedures realized by the device, the above-mentioned method embodiment one has been described in detail, and will not be described in detail here.

[0227] Another embodiment of the present application also provides a computer device, which comprises a processor and a scene database; the scene database is configured to store computer program codes, and the computer program codes comprise computer instructions; when the processor executes the computer instructions, the electronic device executes each step of the rumor detection method based on the graph neural diffusion network shown in the above-mentioned method embodiment.

[0228] Another embodiment of the present application also provides a computer readable storage medium, which stores computer instructions; when the computer instructions run on a computer device, the computer device executes each step of the rumor detection method based on the graph neural diffusion network shown in the above-mentioned method embodiment.

[0229] Obviously, those skilled in the art can make various modifications and variations to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application belong to the scope of the claims of the present application and their equivalent technologies, the present application also intends to include these modifications and variations.

Claims

1. A rumor detection method based on graph neural diffusion networks, characterized in that, include: The original node features of the original text propagation graph are determined, and the original text propagation graph is augmented to obtain an enhanced view and a set number of echo chambers; Construct an initial GDN-ODE network including a mapping module, a diffusion module, and a readout module; determine the initial dynamic features corresponding to the original node features based on the mapping module. ; The time interval of the diffusion process is discretized into multiple time steps. In the first time step, the current dynamic feature of the second time step is determined based on the initial dynamic feature, the diffusion module, and the adaptive Runge-Kutta method. The process is iterated in multiple rounds and multiple time steps until the terminating dynamic feature is obtained. The enhanced view encoding corresponding to the terminating dynamic feature is obtained by the readout module. Based on the trained GDN-ODE network and the original node features, determine the global view encoding corresponding to the original text propagation graph; The graph-level encoding of each echo chamber subgraph and the set of echo chamber encodings are determined based on the trained GDN-ODE network and a set number of echo chambers. The global view encoding and the echo chamber encoding are concatenated along the node dimension to obtain the final propagation graph encoding. The final propagation graph encoding is then input into an MLP classifier to obtain the category label of the original text associated with the original text propagation graph.

2. The method as described in claim 1, characterized in that, The process of data augmentation of the original text propagation graph to obtain an enhanced view and a set number of time-aware echo chambers specifically includes: Based on the shortest path between all node pairs included in the original text propagation graph, determine the edge betweenness number of each edge, remove the edge with the maximum edge betweenness number from the original text propagation graph, and update the edge set included in the original text propagation graph. If updating the original text propagation graph results in a new subgraph that splits, add the new subgraph to the echo chamber set; When the number of echo chambers in the echo chamber set reaches a set value, the updating of the edge set included in the original text propagation graph will stop. The edge betweenness number of each edge is determined by the following formula: in, Representing an edge The edge betweenness number reflects the connectivity importance of an edge in the propagation graph. Represents a node To the node The total number of shortest paths, node To the node Passing by The number of shortest paths, , This represents any two distinct nodes in the propagation graph.

3. The method as described in claim 1, characterized in that, Within the first time step, based on the initial dynamic features, the diffusion module, and the adaptive Runge-Kutta method, the current dynamic features for the second time step are determined, specifically including: Based on the initial dynamic features The diffusion module determines the initial gradient operator, the initial divergence operator, and the feature change rate at the first time step; the feature change rate at the first time step is used to obtain the current dynamic features at the second time step based on the adaptive Runge-Kutta method; The initial gradient operator, the initial divergence operator, and the characteristic change rate at the first time step are expressed by the following formulas. in, This represents the boundary feature matrix, describing the feature differences between nodes. Representing an edge The corresponding gradient operator element, Represents a node The dynamic feature vector, Represents a node The dynamic feature vector, This represents a pair of nodes in the original text propagation graph; This represents the result of the divergence operator. Represents a node The divergence value, Representing an edge The weight, This indicates that for all nodes... There are nodes connected by edges Summation; express The characteristic rate of change at time t, express The dynamic feature matrix at time step, Represents the diffusivity matrix, which varies with time. and dynamic features change, express Gradient operator at time step, This represents the diffusivity matrix.

4. The method as described in claim 1, characterized in that, Before iterating through multiple rounds and time steps until a terminating dynamic feature is obtained, and before the enhanced view encoding corresponding to the terminating dynamic feature obtained by the readout module, the method further includes: The enhanced view is decoded using the initial GDN-ODE network to obtain initial reconstruction features. The initial reconstruction loss is determined based on the initial reconstruction features and the original node features. The initial total loss is determined based on the initial reconstruction loss, the initial global multi-view contrast loss, and the initial echo chamber loss. If the initial total loss does not meet the convergence threshold, the time interval of the diffusion process is discretized into multiple time steps according to the first GDN-ODE network and the original node features, and the first termination dynamic feature, the first enhanced view encoding and the first total loss are obtained by iteratively in the second time step. When the first total loss meets the convergence threshold, the first GDN-ODE network corresponding to the first total loss is determined as the trained GDN-ODE network.

5. The method as described in claim 4, characterized in that, The process of decoding the enhanced view based on the initial GDN-ODE network to obtain initial reconstruction features, and determining the initial reconstruction loss based on the initial reconstruction features and the original node features, specifically includes: The enhanced view includes a feature view, a node view, and an edge view; the initial GDN-ODE network encodes the feature view, the node view, and the edge view to obtain the latent space encoding of the node view, the latent space encoding of the edge view, and the latent space encoding of the feature view in sequence; the latent space encoding of the feature view is decoded to obtain the initial reconstructed feature matrix, and the initial reconstruction loss is obtained based on the original node features and the initial reconstructed features. The initial GDN-ODE network normalizes the latent spatial encoding of node views, edge views, and feature views, and obtains the initial global multi-view contrast loss. The initial reconstruction loss and the initial global multi-view contrast loss are as follows: Indicating reconstruction losses, the index Increase the penalties for difficult samples. Represents the reconstructed feature matrix. This represents the reconstructed feature vector of the i-th node. , This represents the i-th node in the feature view. Represents a node The reconstructed feature vector, Represents a node eigenvectors, This represents the global multi-view contrast loss, narrowing the encoding distance between different views of the same image and widening the encoding distance between different images. This represents the sum of negative sample similarities within the view. This represents the negative sample similarity between views.

6. The method as described in claim 4, characterized in that, The determination of the initial total loss specifically includes: The initial GDN-ODE network encodes each echo chamber, obtains the initial graph-level encoding of each echo chamber subgraph through pooling operations, and concatenates the initial graph-level encodings of each echo chamber subgraph to obtain the initial echo chamber encoding set and the initial echo chamber contrast loss. The initial echo chamber contrast loss and the initial total loss are as follows: This represents echo chamber contrast loss, enhancing the characteristic differences between different echo chambers in the same propagation map. This represents the number of propagation maps, where m is the number of echo chambers. Echo chamber contrast loss is used to enhance the differences between each echo chamber, highlighting their unique characteristics. Indicates the index of the propagation graph. Indicates the index of the echo chamber. Indicates the first The propagation diagram of the first... Graph-level encoding vectors for each echo chamber subgraph Indicates the total loss. Indicates the losses incurred during reconstruction. Indicates multi-view contrast loss. Indicates contrast loss in the echo chamber. Indicates feature smoothing loss, This represents the weight decay regularization loss. , , , This represents the hyperparameters that control the balance of the various loss functions.

7. A rumor detection device based on graph neural diffusion networks, characterized in that, include: The first determining unit is used to determine the original node features of the original text propagation graph, and to perform data augmentation on the original text propagation graph to obtain an enhanced view and a set number of echo chambers; The first obtaining unit is used to construct an initial GDN-ODE network including a mapping module, a diffusion module, and a readout module; and to determine the initial dynamic features corresponding to the original node features based on the mapping module. ; The time interval of the diffusion process is discretized into multiple time steps. In the first time step, the current dynamic feature of the second time step is determined based on the initial dynamic feature, the diffusion module, and the adaptive Runge-Kutta method. The process is iterated in multiple rounds and multiple time steps until the terminating dynamic feature is obtained. The enhanced view encoding corresponding to the terminating dynamic feature is obtained by the readout module. The second determining unit is used to determine the global view encoding corresponding to the original text propagation graph based on the trained GDN-ODE network and the original node features; The graph-level encoding of each echo chamber subgraph and the set of echo chamber encodings are determined based on the trained GDN-ODE network and a set number of echo chambers. The second obtaining unit is used to concatenate the global view encoding and the echo chamber encoding according to the node dimension to obtain the final propagation graph encoding. The final propagation graph encoding is input into the MLP classifier to obtain the category label of the original text associated with the original text propagation graph.

8. A computer device, characterized in that, The computer device includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, causes the processor to perform the rumor detection method based on graph neural diffusion networks as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The device contains a computer program that, when executed by a processor, causes the processor to perform the rumor detection method based on a graph neural diffusion network as described in any one of claims 1-6.

Images