False news detection method and system based on multi-user network behavior, terminal and medium

Abstract

The application relates to the technical field of network fake news detection, and discloses a fake news detection method, system, terminal and medium based on multi-user network behavior, which respectively constructs global structure learning from a user-user perspective and a user-news perspective to respectively obtain a user global representation vector based on social preference and a user global representation vector simultaneously reflecting social and content preference, improves stability in the diversity and correlation of user behavior, learns a user behavior mode under specific news propagation from a local propagation sequence perspective, highlights the difference between true and false news in propagation, comprehensively utilizes propagation characteristics to combine a multi-head self-attention network to capture key propagation behavior characteristics and interactions from multiple different subspaces in parallel, and generates a comprehensive local propagation representation vector; the comprehensive local propagation representation vector is detected through a full connection network to obtain a probability of taking a target news as false, and the difference between true and false news propagation modes is improved.

Description

Technical Field

[0001] This invention relates to the field of online fake news detection technology, specifically to a method, system, terminal, and medium for detecting fake news based on multi-user network behavior. Background Technology

[0002] The rise of social media has changed the way people access information. However, its open communication environment and free participation have also accelerated the generation and spread of misinformation, especially malicious misinformation that is carefully designed to mislead the public. After spreading rapidly through online and offline channels, it may cause irreversible and serious consequences in the real world.

[0003] Classical detection methods primarily rely on the semantics or style of news. However, since fake news is deliberately fabricated to mislead consumers, these methods struggle to identify well-disguised fake news, especially in the context of large language models. Malicious interference can significantly degrade the performance of text-based detection models. Experiments have shown that fake news rewritten using large language models can cause some detection models to drop their F1 scores by more than 30%. Therefore, there is an urgent need to develop more stable detection methods. Recent research has attempted to supplement content features by incorporating features from multiple dimensions, such as related comments, participant attributes, and propagation structures. These strategies improve detection effectiveness, but they are inevitably susceptible to being deceived by the embellished identities of social bots and their fabricated comments. Therefore, more robust detection methods are needed. Inspired by graph representation learning techniques, a series of fake news detection methods based on propagation structure learning have been proposed recently. However, most methods mainly rely on neural networks to learn user forwarding behavior features under a single news item, rarely using deep learning from multiple perspectives to study the global connections between users and between news and users, thus limiting the mining of user identities and behavioral preferences in news dissemination. Summary of the Invention

[0004] In order to overcome the shortcomings of the existing technology, the present invention aims to provide a method, system, terminal and medium for detecting fake news based on multi-user network behavior, so as to solve the technical problems of weak stability and incomplete utilization of dissemination characteristics in the existing technology due to the diversity and correlation of user behavior.

[0005] This invention is achieved through the following technical solution:

[0006] In a first aspect, the present invention provides a method for detecting fake news based on multi-user network behavior, comprising:

[0007] Users with interactive behaviors are linked together to construct a global interaction network, and then aggregated and learned through a multi-layer convolutional graph neural network to obtain a global representation vector of users based on social preferences.

[0008] By learning users' interactive behavior at the news level, a multi-layer user behavior hypergraph network is obtained by explicitly modeling users' content preferences on the user global representation vector based on social preferences. The hypergraph convolutional neural network is then used to learn the user global representation vector that simultaneously reflects social and content preferences.

[0009] The corresponding local sequence initial representations are extracted from the user global representation vector based on social preferences and the user global representation vector that simultaneously reflects social and content preferences. A multi-head self-attention network is used to capture key propagation behavior features and their interactions in parallel from multiple different subspaces to generate a comprehensive local propagation representation vector.

[0010] The comprehensive local propagation representation vector is passed through a fully connected network for detection to obtain the probability that the target news is fake, thus completing the fake news detection work.

[0011] Preferably, the global interactive network uses a news set N = {n1, n2, ..., n} consisting of m news items. m In this context, the set of users participating in news dissemination and the set of edges connecting users are used to construct the global user-to-user interaction behavior, which is expressed as follows:

[0012] G = (U, E)

[0013] In the formula: U={u1,u2,…,u n} represents a set of n users; E = {(u i ,u j )|u i ,u j ∈U} represents the set of edges connecting users;

[0014] The global interactive behaviors include posting-forwarding, posting-commenting, forwarding-commenting, forwarding-forwarding, and commenting-commenting.

[0015] Furthermore, the user global representation vector based on social preferences is obtained through aggregation learning via a multi-layer convolutional graph neural network, where the hierarchical information transfer expression of the multi-layer convolutional graph neural network is as follows:

[0016]

[0017] in, It is randomly initialized from a normal distribution, where D is the embedding dimension, l is the number of network layers, σ represents the ReLU activation function, and W... G It is a learnable weight matrix. and The adjacency matrix and degree matrix of a graph G with self-loops are aggregated through L layers of GCN to obtain a user global representation vector X based on social preferences. G .

[0018] Preferably, the multi-layered user behavior hypergraph network explicitly models user content preferences by learning user interaction behaviors at the news level. These user interaction behaviors at the news level include posting, forwarding, and commenting. The posting behavior is then modeled by the hypergraph network HG. T =(U,HE) T The forwarding behavior is described in the Hypergraph Network (HG). R =(U,HE) R The commenting behavior is described in the Hypergraph Network (HG). C =(U,HE) C ), where posting behavior is represented by each hyperedge HE in the hypergraph network. Tj Connect all published news items j All users who shared or commented on the relevant posts; the hyperedges in the hypergraph networks for forwarding and commenting behaviors represent all users who shared or commented on the news-related posts, respectively.

[0019] Furthermore, the specific learning process of the hypergraph convolutional neural network is as follows:

[0020] S1, input the hypergraph network for posting behavior, the hypergraph network for forwarding behavior, and the hypergraph network for commenting behavior into the hypergraph convolutional neural network for aggregation;

[0021] S2, obtain the representation vector of the hyperedge by aggregating the representation vectors of all nodes connected by the hyperedge in the graph;

[0022] S3. After obtaining the representation vectors of the hyperedges, the representation vectors of all the hyperedges that the user nodes have participated in are aggregated in reverse to obtain the user feature representation vectors of the posting behavior hypergraph network, the forwarding behavior hypergraph network, and the commenting behavior hypergraph network, respectively.

[0023] S4. A bottom-up two-layer gating fusion mechanism is established. The first-layer gating fusion mechanism adjusts and merges the weights of the user feature representation vectors of the forwarding behavior hypergraph network and the commenting behavior hypergraph network respectively. The weights of the merged user feature representation vectors of the forwarding behavior hypergraph network and the commenting behavior hypergraph network are used as a whole and input into the second-layer gating fusion mechanism along with the user feature representation vector of the posting behavior hypergraph network to calculate the gating coefficients and complete the fusion to obtain a global user representation vector that reflects both social and content preferences.

[0024] Preferably, local propagation behavior sequence modeling is based on a self-attention mechanism, where given news n j and its propagation sequence s jThe system extracts corresponding local sequence initial representations from the user global representation vector based on social preferences and the user global representation vector that simultaneously reflects social and content preferences. It also generates a d-dimensional behavior representation for the user's behavior type. The input of the self-attention network is obtained by concatenating the local sequence initial representation and the d-dimensional behavior representation. It captures key propagation behavior features and their interactions in parallel from multiple different subspaces, generates a comprehensive local propagation representation vector, inputs the comprehensive local propagation representation vector into a two-layer fully connected neural network, and takes the average of the user representation vectors as the final news representation vector.

[0025] Preferably, the comprehensive local propagation representation vector is processed through a fully connected network and a softmax algorithm to target news n. j The probability of fake news is calculated by minimizing the cross-entropy error using training samples with real labels, thus obtaining the overall loss function, as shown in the following formula:

[0026]

[0027]

[0028] in, The news n predicted by the model j The probability of being false, W p It is a learnable feature transformation matrix, b p y represents the bias coefficients; θ represents all the parameters that need to be learned. j This serves as a true label for news.

[0029] Secondly, the present invention provides a fake news detection system based on multi-user network behavior, comprising:

[0030] The global user interaction learning module is used to connect users with interactive behaviors and construct a global interaction network. It then performs aggregation learning through a multi-layer convolutional graph neural network to obtain a global user representation vector based on social preferences.

[0031] The Global News Participation Learning Module is used to explicitly model users' content preferences by learning their interactive behaviors at the news level on the user global representation vector based on social preferences, thereby obtaining a multi-layer user behavior hypergraph network. This network is then learned through a hypergraph convolutional neural network to obtain a user global representation vector that simultaneously reflects social and content preferences.

[0032] The local propagation sequence learning module is used to extract the corresponding local sequence initial representation from the user global representation vector based on social preferences and the user global representation vector that simultaneously reflects social and content preferences. It also uses a multi-head self-attention network to capture key propagation behavior features and their interactions in parallel from multiple different subspaces to generate a comprehensive local propagation representation vector.

[0033] The detection module is used to process the comprehensive local propagation representation vector through a fully connected network to obtain the probability that the target news is fake, thus completing the fake news detection task.

[0034] Thirdly, the present invention provides a mobile terminal, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the fake news detection method based on multi-user network behavior as described above.

[0035] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the fake news detection method based on multi-user network behavior as described above.

[0036] Compared with the prior art, the present invention has the following beneficial technical effects:

[0037] This invention provides a method for detecting fake news based on multi-user network behavior. It constructs global structure learning from both user-user and user-news perspectives to obtain user global representation vectors based on social preferences and those reflecting both social and content preferences, improving stability in the face of diverse and interconnected user behaviors. It learns user behavior patterns under specific news dissemination from the perspective of local propagation sequences, highlighting the differences between true and false news in dissemination. It comprehensively utilizes propagation features combined with a multi-head self-attention network to capture key propagation behavior features and their interactions in parallel from multiple different subspaces, generating a comprehensive local propagation representation vector. This comprehensive local propagation representation vector is then passed through a fully connected network for detection to calculate the probability that the target news is fake, improving the ability to capture the differences in dissemination patterns between true and false news.

[0038] Furthermore, this invention constructs a multi-layered global hypergraph network from the perspectives of user posting behavior, forwarding behavior, and commenting behavior, and performs joint learning to deeply explore the diversity and correlation of user behavior, thereby accurately describing users' preferences in social media and news types, and helping to capture subtle falsification clues.

[0039] Furthermore, this invention learns user behavior patterns under specific news dissemination from the perspective of local propagation sequences. It inputs the comprehensive local propagation representation vector into a two-layer fully connected neural network and takes the average of the user representation vectors as the final news representation vector. It also combines a fully connected network and softmax to analyze the target news n. j The calculation of the probability of fake news highlights the differences in the dissemination of true and false news. Attached Figure Description

[0040] Figure 1 This is a flowchart of the fake news detection method based on multi-user network behavior in this invention;

[0041] Figure 2 This is an architecture diagram of the fake news detection system based on multi-user network behavior in this invention;

[0042] Figure 3 The graph shows the detection performance of this invention on the Politifact and Gossipcop datasets.

[0043] Figure 4 This is a graph showing the detection performance of this invention on the Fang dataset;

[0044] Figure 5 This is a comparison chart of the separation performance of different modules of the present invention across three datasets;

[0045] In the diagram: 1-Global user interaction learning module; 2-Global news participation learning module; 3-Local propagation sequence learning module; 4-Detection module. Detailed Implementation

[0046] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0047] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0048] The present invention will now be described in further detail with reference to the accompanying drawings:

[0049] The purpose of this invention is to provide a method, system, terminal, and medium for detecting fake news based on multi-user network behavior, so as to solve the technical problems of weak stability and incomplete utilization of dissemination characteristics in the existing technology due to the diversity and correlation of user behavior.

[0050] Example 1

[0051] See Figure 1 In one embodiment of the present invention, a method for detecting fake news based on multi-user network behavior is provided, comprising:

[0052] Step 1: Connect users with interactive behaviors and construct a global interaction network. Then, perform aggregation learning through a multi-layer convolutional graph neural network to obtain a global representation vector of users based on social preferences.

[0053] Specifically, the global interactive network uses a news set N = {n1, n2, ..., n} consisting of m news items. m In this context, the set of users participating in news dissemination and the set of edges connecting users are used to construct the global user-to-user interaction behavior, which is expressed as follows:

[0054] G = (U, E)

[0055] In the formula: U={u1,u2,…,u n} represents a set of n users; E = {(u i ,u j )|u i ,u j ∈U} represents the set of edges connecting users;

[0056] The global interactive behaviors include posting-forwarding, posting-commenting, forwarding-commenting, forwarding-forwarding, and commenting-commenting.

[0057] Specifically, based on the user interaction behavior network G, in order to avoid the loss of the influence of the posting user node on the child nodes after multiple orders, this invention first enhances the structure of the initial tree propagation graph by adding edges from the root user node to all child nodes.

[0058] After the network is constructed, multilayer convolutional graph neural networks (GCNs) are used to learn global user interaction behavior. Given a network G = (U, E), the hierarchical information transfer expression of the multilayer convolutional graph neural network is as follows:

[0059]

[0060] in, It is randomly initialized from a normal distribution, where D is the embedding dimension, l is the number of network layers, σ represents the ReLU activation function, and W... G It is a learnable weight matrix. and The adjacency matrix and degree matrix of a graph G with self-loops are aggregated through L layers of GCN to obtain a user global representation vector X based on social preferences. G .

[0061] Step 2: On the user global representation vector based on social preferences, the user's content preferences are explicitly modeled by learning the user's interactive behavior at the news level to obtain a multi-layer user behavior hypergraph network. The hypergraph convolutional neural network is then used to learn the user global representation vector that simultaneously reflects social and content preferences.

[0062] Specifically, the multi-layered user behavior hypergraph network explicitly models user content preferences by learning user interaction behaviors at the news level. These user interaction behaviors at the news level include posting, forwarding, and commenting. The posting behavior is then modeled by the hypergraph network HG. T =(U,HE) T The forwarding behavior is described in the Hypergraph Network (HG). R =(U,HE) R The commenting behavior is described in the Hypergraph Network (HG). C =(U,HE) C ), where posting behavior is represented by each hyperedge HE in the hypergraph network. Tj Connect all published news items j All users who shared or commented on the relevant posts; the hyperedges in the hypergraph networks for forwarding and commenting behaviors represent all users who shared or commented on the news-related posts, respectively.

[0063] The specific learning process of the hypergraph convolutional neural network is as follows:

[0064] S1, input the hypergraph network for posting behavior, the hypergraph network for forwarding behavior, and the hypergraph network for commenting behavior into the hypergraph convolutional neural network for aggregation;

[0065] S2, obtain the representation vector of the hyperedge by aggregating the representation vectors of all nodes connected by the hyperedge in the graph;

[0066] S3. After obtaining the representation vectors of the hyperedges, the representation vectors of all the hyperedges that the user nodes have participated in are aggregated in reverse to obtain the user feature representation vectors of the posting behavior hypergraph network, the forwarding behavior hypergraph network, and the commenting behavior hypergraph network, respectively.

[0067] Specifically, in S31, Hypergraph Network HG targets posting behavior. T The first step in aggregation is to obtain the representation vector of the hyperedge by aggregating the representation vectors of all nodes connected by the hyperedge in the graph. The expression is as follows:

[0068]

[0069] Where D T HG represents the supergraph. T The degree matrix of the nodes, B T H represents the degree matrix of the hyperedge. T The hypergraph's association matrix represents the initial user features input. σ represents the ReLU activation function, l H This represents the current layer number in the network.

[0070] S32, after obtaining the representation vector of the hyperedge, reversely aggregate the representation vectors of all hyperedges that the user node has participated in, in order to update the node's representation vector. The expression is as follows:

[0071]

[0072] Where σ is the activation function ReLU, W T This is a trainable weight matrix. After passing through L... H After layered hypergraph neural network aggregation learning, the output is a posting hypergraph HG. T The final user node representation vector is

[0073] S33, targeting forwarding behavior in Hypergraph Network HG R And comment behavior Hypergraph Network HG C Repeat steps S31 and S32 respectively to obtain the user feature representation X. R and X C .

[0074] S4. User behavior representations learned from multiple perspectives show differences in importance. Posting behavior, as an active form of information dissemination, is highly representative, while reposting and commenting behavior, as driven dissemination methods, have slightly lower importance. A bottom-up two-layer gating fusion mechanism is established. The first-layer gating fusion mechanism adjusts and merges the weights of the user feature representation vectors of the reposting behavior hypergraph network and the commenting behavior hypergraph network, respectively. The weights of the merged user feature representation vectors of the reposting behavior hypergraph network and the commenting behavior hypergraph network are then used as the whole and input into the second-layer gating fusion mechanism along with the user feature representation vector of the posting behavior hypergraph network to calculate the gating coefficients and complete the fusion, resulting in a global user representation vector that simultaneously reflects social and content preferences.

[0075] In the first-layer gating fusion mechanism, the weights of reposting and commenting behaviors are adjusted and fused separately, with the specific formula as follows:

[0076]

[0077]

[0078] Among them, X RC To integrate the user's comments and reposts on the SuperMap, and These are the trainable attention vector and the feature transformation matrix, respectively, where σ represents the tanh activation function.

[0079] In the second layer of fusion, the combined features X of forwarding and commenting are used. RC As a whole, it is related to user characteristics X based on the posting hypergraph. T Together, they are input into the second gating mechanism, where gating coefficients are calculated and fusion is completed to obtain the final user feature representation vector X, as shown in the following expression:

[0080]

[0081]

[0082] in, and These are the trainable attention vector and the feature transformation matrix, respectively, where σ represents the tanh activation function.

[0083] Step 3: Extract the corresponding local sequence initial representation from the user global representation vector based on social preferences and the user global representation vector that simultaneously reflects social and content preferences, and use a multi-head self-attention network to capture key propagation behavior features and their interactions in parallel from multiple different subspaces to generate a comprehensive local propagation representation vector.

[0084] Specifically, modeling of local propagation behavior sequences based on self-attention mechanisms, where given news n j and its propagation sequence s j The system extracts corresponding local sequence initial representations from the user global representation vector based on social preferences and the user global representation vector that simultaneously reflects social and content preferences. It also generates a d-dimensional behavior representation for the user's behavior type. The input of the self-attention network is obtained by concatenating the local sequence initial representation and the d-dimensional behavior representation. The expression is as follows:

[0085]

[0086] h q,j =Att(h′) j W Q ,h′ j W K ,h′ j W V )

[0087] h j =[h 1,j h 2,j ;…;h H,j W O

[0088] Where H is the number of attention heads, d d Let K″ be the dimension of the vector representing the user node, and W be the transpose of matrix K. Q W K W V W O It is a learnable matrix.

[0089] Furthermore, key propagation behavior features and their interactions are captured in parallel from multiple different subspaces to generate a comprehensive local propagation representation vector. This comprehensive local propagation representation vector is then input into a two-layer fully connected neural network, and the average of the user representation vectors is taken as the final news representation vector, as shown below:

[0090]

[0091] Where σ is the activation function ReLU and b1 and b2 are trainable weight matrices.

[0092] In this invention, the propagation sequence generated by the dissemination of m news items is defined as S = {s1, s2, ..., s}. m}, where, for news n j The generated propagation sequence middle, For user ui Spreading news j The timestamp of the time, Indicates user u i Spreading news j The type of behavior at that time.

[0093] Step 4: The comprehensive local propagation representation vector is passed through a fully connected network for detection to obtain the probability that the target news is fake, thus completing the fake news detection work.

[0094] The cross-entropy error is minimized using training samples with true labels to obtain the overall loss function, as shown in the following formula:

[0095]

[0096]

[0097] in, The news n predicted by the model j The probability of being false, W p It is a learnable feature transformation matrix, b p y represents the bias coefficients; θ represents all the parameters that need to be learned. j This serves as a true label for news.

[0098] In this invention, each news item n j Having a label y j If the news is false, then y ∈{0,1}. j =1, otherwise y j =0;

[0099] according to Figure 3 and Figure 4 The experimental results shown demonstrate the excellent performance of this invention in the task of detecting fake news. Figure 5 The ablation experiment shown further confirms the rationality of the design of this invention and the effectiveness of each module.

[0100] In summary, this invention provides a fake news detection method based on multi-user network behavior. It simultaneously learns the propagation structure and sequence of news at both global and local levels, emphasizing the diversity and correlation of user behavior to better capture the differences in propagation characteristics between true and false news, overcoming the problems of weak stability and incomplete utilization of propagation features in existing methods. Specifically, this invention proposes a robust and effective fake news detection model that relies solely on user propagation behavior. It constructs global structure learning modules from both user-user and user-news perspectives, comprehensively mining user biases and identity clues. To further explore the diversity and correlation of user behavior, this invention constructs a multi-layered global hypergraph network from three behavioral perspectives: posting, reposting, and commenting, and performs joint learning to accurately describe user preferences in social interactions and news types, helping to capture subtle falsification clues. This invention not only focuses on global features but also learns user behavior patterns under specific news propagation from the perspective of local propagation sequences, further highlighting the differences in propagation between true and false news. Experimental results on three real datasets show that this invention is stable and effective in fake news detection tasks, especially under extreme conditions such as insufficient data and missing data labels, still achieving stable detection results.

[0101] The method provided by this invention is applicable to social network environments and information dissemination systems with frequent user interactions and certain patterns, such as microblogs, Twitter, and news websites.

[0102] Example 2

[0103] according to Figure 2 As shown, the present invention also provides a fake news detection system based on multi-user network behavior, including a global user interaction learning module 1, a global news participation learning module 2, a local propagation sequence learning module 3, and a detection module 4;

[0104] Global User Interaction Learning Module 1 is used to connect users with interactive behaviors and construct a global interaction network, and then perform aggregation learning through a multi-layer convolutional graph neural network to obtain a global user representation vector based on social preferences.

[0105] The Global News Participation Learning Module 2 is used to explicitly model users' content preferences by learning their interactive behaviors at the news level on the user global representation vector based on social preferences, and obtain a multi-layer user behavior hypergraph network. The hypergraph convolutional neural network is then used to learn the user global representation vector that simultaneously reflects social and content preferences.

[0106] The local propagation sequence learning module 3 is used to extract the corresponding initial representation of the local sequence from the user global representation vector based on social preferences and the user global representation vector that simultaneously reflects social and content preferences. It also uses a multi-head self-attention network to capture key propagation behavior features and their interactions in parallel from multiple different subspaces to generate a comprehensive local propagation representation vector.

[0107] Detection module 4 is used to process the comprehensive local propagation representation vector through a fully connected network to obtain the probability that the target news is fake, thereby completing the fake news detection work.

[0108] Example 3

[0109] The present invention also provides a mobile terminal, including a memory, a processor, and a computer program stored in the memory and executable on the processor, such as a fake news detection program based on multi-user network behavior.

[0110] When the processor executes the computer program, it implements the steps of the above-described method for detecting fake news based on multi-user network behavior, for example:

[0111] Users with interactive behaviors are linked together to construct a global interaction network, and then aggregated and learned through a multi-layer convolutional graph neural network to obtain a global representation vector of users based on social preferences.

[0112] By learning users' interactive behavior at the news level, a multi-layer user behavior hypergraph network is obtained by explicitly modeling users' content preferences on the user global representation vector based on social preferences. The hypergraph convolutional neural network is then used to learn the user global representation vector that simultaneously reflects social and content preferences.

[0113] The corresponding local sequence initial representations are extracted from the user global representation vector based on social preferences and the user global representation vector that simultaneously reflects social and content preferences. A multi-head self-attention network is used to capture key propagation behavior features and their interactions in parallel from multiple different subspaces to generate a comprehensive local propagation representation vector.

[0114] The comprehensive local propagation representation vector is passed through a fully connected network for detection to obtain the probability that the target news is fake, thus completing the fake news detection work.

[0115] Alternatively, when the processor executes the computer program, it implements the functions of each module in the above system, for example:

[0116] Global User Interaction Learning Module 1 is used to connect users with interactive behaviors and construct a global interaction network, and then perform aggregation learning through a multi-layer convolutional graph neural network to obtain a global user representation vector based on social preferences.

[0117] The Global News Participation Learning Module 2 is used to explicitly model users' content preferences by learning their interactive behaviors at the news level on the user global representation vector based on social preferences, and obtain a multi-layer user behavior hypergraph network. The hypergraph convolutional neural network is then used to learn the user global representation vector that simultaneously reflects social and content preferences.

[0118] The local propagation sequence learning module 3 is used to extract the corresponding initial representation of the local sequence from the user global representation vector based on social preferences and the user global representation vector that simultaneously reflects social and content preferences. It also uses a multi-head self-attention network to capture key propagation behavior features and their interactions in parallel from multiple different subspaces to generate a comprehensive local propagation representation vector.

[0119] Detection module 4 is used to process the comprehensive local propagation representation vector through a fully connected network to obtain the probability that the target news is fake, thereby completing the fake news detection work.

[0120] For example, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the mobile terminal.

[0121] For example, the computer program can be divided into a global user interaction learning module 1, a global news participation learning module 2, a local propagation sequence learning module 3, and a detection module 4;

[0122] The specific functions of each module are as follows:

[0123] Global User Interaction Learning Module 1 is used to connect users with interactive behaviors and construct a global interaction network, and then perform aggregation learning through a multi-layer convolutional graph neural network to obtain a global user representation vector based on social preferences.

[0124] The Global News Participation Learning Module 2 is used to explicitly model users' content preferences by learning their interactive behaviors at the news level on the user global representation vector based on social preferences, and obtain a multi-layer user behavior hypergraph network. The hypergraph convolutional neural network is then used to learn the user global representation vector that simultaneously reflects social and content preferences.

[0125] The local propagation sequence learning module 3 is used to extract the corresponding initial representation of the local sequence from the user global representation vector based on social preferences and the user global representation vector that simultaneously reflects social and content preferences. It also uses a multi-head self-attention network to capture key propagation behavior features and their interactions in parallel from multiple different subspaces to generate a comprehensive local propagation representation vector.

[0126] Detection module 4 is used to process the comprehensive local propagation representation vector through a fully connected network to obtain the probability that the target news is fake, thereby completing the fake news detection work.

[0127] The mobile terminal can be a computing device such as a desktop computer, laptop, handheld computer, or cloud server. The mobile terminal may include, but is not limited to, a processor and memory.

[0128] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the mobile terminal, connecting various parts of the mobile terminal via various interfaces and lines.

[0129] The memory can be used to store the computer program and / or module. The processor implements various functions of the mobile terminal by running or executing the computer program and / or module stored in the memory and calling the data stored in the memory.

[0130] The memory may primarily include a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a given function (such as sound playback or image playback). The data storage area may store data created based on the use of the phone (such as audio data or a phonebook). Furthermore, the memory may include high-speed random access memory (RAM) and non-volatile memory, such as hard disks, RAM, plug-in hard disks, SmartMediaCards (SMC), Secure Digital (SD) cards, flash cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0131] Example 4

[0132] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method for detecting fake news based on multi-user network behavior.

[0133] If the modules / units integrated in the mobile terminal are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium.

[0134] Based on this understanding, all or part of the processes in the above-described method can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium. When executed by a processor, the computer program can implement the steps of the above-described method for detecting fake news based on multi-user network behavior. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or some intermediate form.

[0135] The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium, etc.

[0136] It should be noted that the content contained in the computer-readable medium may be appropriately added to or subtracted from the content as required by the legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the computer-readable medium may not include electrical carrier signals and telecommunication signals.

[0137] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for detecting fake news based on multi-user network behavior, characterized in that, include: Users with interactive behaviors are linked together to construct a global interaction network, and then aggregated and learned through a multi-layer convolutional graph neural network to obtain a global representation vector of users based on social preferences. By learning users' interactive behavior at the news level, a multi-layer user behavior hypergraph network is obtained by explicitly modeling users' content preferences on the user global representation vector based on social preferences. The hypergraph convolutional neural network is then used to learn the user global representation vector that simultaneously reflects social and content preferences. The corresponding local sequence initial representations are extracted from the user global representation vector based on social preferences and the user global representation vector that simultaneously reflects social and content preferences. Then, a multi-head self-attention network is used to capture key propagation behavior features and their interactions in parallel from multiple different subspaces to generate a comprehensive local propagation representation vector. The comprehensive local propagation representation vector is passed through a fully connected network for detection to obtain the probability that the target news is fake, thus completing the fake news detection work; Among them, the multi-layer user behavior hypergraph network explicitly models user content preferences by learning user interaction behaviors at the news level. These user interaction behaviors at the news level include posting, forwarding, and commenting. The posting behavior hypergraph network... The forwarding behavior is described in the hypergraph network. The commenting behavior is a hypergraph network. The posting behavior is represented by each hyperedge in the hypergraph network. Connect all published news All users who shared or commented on the relevant posts; the hyperedges in the hypergraph networks for forwarding and commenting behaviors represent all users who shared or commented on the news-related posts, respectively. The specific learning process of the hypergraph convolutional neural network is as follows: S1, input the hypergraph network for posting behavior, the hypergraph network for forwarding behavior, and the hypergraph network for commenting behavior into the hypergraph convolutional neural network for aggregation; S2, obtain the representation vector of the hyperedge by aggregating the representation vectors of all nodes connected by the hyperedge in the graph; S3. After obtaining the representation vectors of the hyperedges, the representation vectors of all the hyperedges that the user nodes have participated in are aggregated in reverse to obtain the user feature representation vectors of the posting behavior hypergraph network, the forwarding behavior hypergraph network, and the commenting behavior hypergraph network, respectively. S4. A bottom-up two-layer gating fusion mechanism is established. The first-layer gating fusion mechanism adjusts and merges the weights of the user feature representation vectors of the forwarding behavior hypergraph network and the commenting behavior hypergraph network respectively. The weights of the merged user feature representation vectors of the forwarding behavior hypergraph network and the commenting behavior hypergraph network are used as a whole and input into the second-layer gating fusion mechanism along with the user feature representation vector of the posting behavior hypergraph network to calculate the gating coefficients and complete the fusion to obtain a global user representation vector that reflects both social and content preferences.

2. The method for detecting fake news based on multi-user network behavior according to claim 1, characterized in that, The global interactive network uses a news set consisting of m news items. The global user-to-user interaction behavior is constructed by the set of users participating in news dissemination and the set of edges connecting users, and its expression is as follows: In the formula: Represents a set of n users; This represents the set of edges connecting users. The global interactive behaviors include posting-forwarding, posting-commenting, forwarding-commenting, forwarding-forwarding, and commenting-commenting.

3. The method for detecting fake news based on multi-user network behavior according to claim 2, characterized in that, The user global representation vector based on social preferences is obtained by aggregation learning through a multi-layer convolutional graph neural network, where the hierarchical information transfer expression of the multi-layer convolutional graph neural network is as follows: in, It is randomly initialized from a normal distribution. It is the embedded dimension. It refers to the number of layers in the network. This represents the ReLU activation function. It is a learnable weight matrix. +I and It is a graph with self-loops. The adjacency matrix and degree matrix are aggregated through L layers of GCN to obtain a global user representation vector based on social preferences. .

4. The method for detecting fake news based on multi-user network behavior according to claim 1, characterized in that, Modeling of local propagation behavior sequences based on self-attention mechanism, where given news and its propagation sequence The system extracts corresponding local sequence initial representations from the user global representation vector based on social preferences and the user global representation vector that simultaneously reflects social and content preferences. It also generates a d-dimensional behavior representation for the user's behavior type. The input of the self-attention network is obtained by concatenating the local sequence initial representation and the d-dimensional behavior representation. It captures key propagation behavior features and their interactions in parallel from multiple different subspaces, generates a comprehensive local propagation representation vector, inputs the comprehensive local propagation representation vector into a two-layer fully connected neural network, and takes the average of the user representation vectors as the final news representation vector.

5. The method for detecting fake news based on multi-user network behavior according to claim 1, characterized in that, The comprehensive local propagation representation vector is processed through a fully connected network and a softmax algorithm to target the news. The probability of fake news is calculated by minimizing the cross-entropy error using training samples with real labels, thus obtaining the overall loss function, as shown in the following formula: in, This indicates the news predicted by the model. The probability of it being false. It is a learnable feature transformation matrix. This is the bias coefficient; This represents all the parameters that need to be learned. This serves as a true label for news.

6. A fake news detection system based on multi-user network behavior, characterized in that, A method for detecting fake news based on multi-user network behavior as described in any one of claims 1-5 includes: The global user interaction learning module (1) is used to connect users with interactive behaviors and construct a global interaction network, and to perform aggregation learning through a multi-layer convolutional graph neural network to obtain a global user representation vector based on social preferences. The global news participation learning module (2) is used to explicitly model the user's content preferences by learning the user's interactive behavior at the news level on the user global representation vector based on social preferences to obtain a multi-layer user behavior hypergraph network, and to learn through the hypergraph convolutional neural network to obtain the user global representation vector that reflects both social and content preferences. The local propagation sequence learning module (3) is used to extract the corresponding local sequence initial representation from the user global representation vector based on social preferences and the user global representation vector that simultaneously reflects social and content preferences, and to use a multi-head self-attention network to capture key propagation behavior features and their interactions in parallel from multiple different subspaces to generate a comprehensive local propagation representation vector. The detection module (4) is used to detect the comprehensive local propagation representation vector through a fully connected network to obtain the probability that the target news is fake, and to complete the fake news detection work.

7. A mobile terminal, 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 computer program, it implements the steps of the fake news detection method based on multi-user network behavior as described in any one of claims 1-5.

8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the fake news detection method based on multi-user network behavior as described in any one of claims 1-5.

Images