Rumor detection method, device and system and medium

By constructing social network graph and graph convolution processing, combined with the weighted processing of Momentum parameters, the problem of poor detection capabilities of multi-information channels in the existing technology is solved, the accuracy and stability of rumor detection are improved, and the to-multi-modality is realized Effective identification of information.

CN119004281APending Publication Date: 2024-11-22PING AN TECH (SHENZHEN) CO LTD
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202411152586.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-21
Publication Date
2024-11-22

AI Technical Summary

Technical Problem

When faced with multiple information channels, the existing rumor detection methods have poor generalization capabilities, resulting in a decrease in accuracy and stability, and the inability to effectively identify and filter Internet rumors.

Method used

By collecting multimodal information, building a social network graph, using graph convolution neural network to fuse entity relationships, update node features, and weighted average of node vectors and edge vectors through Momentum parameters, combining with the trained rumor detection model for binary classification processing , output rumor detection results.

Benefits of technology

It improves the accuracy and stability of rumor detection, and can more comprehensively represent user behavior characteristics and information content, thereby achieving more accurate rumor detection in a multimodal information environment.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119004281A_ABST
    Figure CN119004281A_ABST
Patent Text Reader

Abstract

The invention discloses a rumor detection method, device and system and a medium, and the method comprises the steps: collecting to-be-detected multi-modal information, and constructing a corresponding social network graph according to the multi-modal information; performing entity relationship fusion on the social network graph through a graph convolutional neural network, and updating node features of the social network graph; inputting the updated social network graph into a trained rumor detection model, and performing weighted average on node vectors and edge vectors of nodes in the updated social network graph through Momentum parameters in the model to obtain corresponding weighted average vectors; and performing dichotomy processing on the weighted average vector through a classifier in the rumor detection model, and outputting a corresponding rumor detection result. Detection of multi-modal information is supported by constructing a social network graph and performing graph convolution processing, and Momentum parameters are introduced to perform weighting processing on node features, so that the rumor detection accuracy and stability are improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of financial technology, and in particular to a rumor detection method, device, system and medium. Background Art

[0002] The development of social media platforms has greatly promoted the spread of information. On the one hand, it has expanded the channels for Internet users to learn about information. On the other hand, it has led to the spread of unconfirmed events, namely Internet rumors. In the financial field, rumor detection is a relatively important technology. Through rumor detection, market fraud can be prevented, information transparency can be improved, and systemic risks can be reduced.

[0003] Existing rumor detection is often achieved through sentiment dictionaries, deep learning models, knowledge graphs, etc. However, existing methods often only use one source of information for training, and have poor generalization capabilities for multiple information channels. When faced with mixed information, the model appears relatively powerless, reducing the accuracy and stability of rumor detection. Summary of the invention

[0004] In view of the above-mentioned deficiencies in the prior art, the object of the present invention is to provide a rumor detection method, device, system and medium that can be applied to financial technology or other related fields, the main purpose of which is to improve the accuracy and stability of rumor detection.

[0005] The technical solution of the present invention is as follows:

[0006] A first aspect of the present invention provides a rumor detection method, comprising:

[0007] Collecting multimodal information to be detected, and constructing a corresponding social network graph based on the multimodal data;

[0008] Performing entity relationship fusion on the social network graph through a graph convolutional neural network to update node features of the social network graph;

[0009] The updated social network graph is input into the trained rumor detection model, and the node vector and edge vector of each node in the updated social network graph are weighted averaged through the Momentum parameter in the model to obtain the corresponding weighted average vector;

[0010] The weighted average vector is subjected to binary classification processing by the classifier in the rumor detection model, and a corresponding rumor detection result is output.

[0011] In one embodiment, the collecting of the multimodal information to be detected and constructing a corresponding social network graph according to the multimodal data includes:

[0012] Collecting multimodal information to be detected, wherein the multimodal information includes user information, propagation mode information and propagation content information;

[0013] According to the user information, the propagation mode information and the propagation content information, a user feature vector, a propagation mode vector and an information content vector are respectively extracted;

[0014] The user feature vector is used as a node vector, and the corresponding edge vector is calculated according to the propagation mode vector and the information content vector, and the corresponding social network graph is constructed through a graph neural network.

[0015] In one embodiment, the calculating the corresponding edge vector according to the propagation mode vector and the information content vector specifically includes:

[0016] A weighted sum is performed on the propagation mode vector and the information content vector to obtain a corresponding edge vector.

[0017] In one embodiment, the Momentum parameter is trained by the following steps:

[0018] Calculate the user rumor rate and initialize the Momentum parameter, where the Momentum parameter is used to adjust the gradient update direction and size of the model;

[0019] When updating the gradient during model training, the learning rate is updated according to the current Momentum parameter and the user's rumor-making rate, and the Momentum parameter is adjusted by the updated learning rate until the model training is completed.

[0020] In one embodiment, calculating the user rumor rate and initializing the Momentum parameter specifically includes:

[0021] The user rumor rate is calculated by the formula Urr = (1 + 2^2) * (Nrx + Ntx) / (2^2 * Nrx + Ntx), where Urr is the user rumor rate, Nrx is the number of received rumors, and Ntx is the number of sent rumors;

[0022] Initialize the Momentum parameter to the user's rumor rate.

[0023] In one embodiment, updating the learning rate according to the current Momentum parameter and the user rumor spreading rate specifically includes:

[0024] Through the formula learning_rate'=learning_rate*Momentum weight coefficient

[0025]

[0026] Update the learning rate, where learning_rate' is the updated learning rate, learning_rate is the learning rate before the update, Momentum is the current Momentum parameter, and Urr is the user rumor rate.

[0027] In one embodiment, the classifier includes a fully connected layer and a softmax layer, and the classifier in the rumor detection model performs binary classification processing on the weighted average vector and outputs a corresponding rumor detection result, including:

[0028] Inputting the weighted average vector into the fully connected layer, and outputting the original prediction values ​​of the rumor category and the non-rumor category;

[0029] The original prediction values ​​of the rumor category and the non-rumor category are input into the softmax layer, and the original prediction values ​​are converted into corresponding probabilities through the softmax function, and the category with the highest probability is selected as the rumor detection result.

[0030] A second aspect of the present invention provides a rumor detection device, comprising:

[0031] A graph construction module, used to collect multimodal information to be detected and construct a corresponding social network graph based on the multimodal data;

[0032] A graph convolution module, used to perform entity relationship fusion on the social network graph through a graph convolutional neural network, and update node features of the social network graph;

[0033] A weighted processing module is used to input the updated social network graph into the trained rumor detection model, and perform weighted averaging of the node vectors and edge vectors of each node in the updated social network graph through the Momentum parameter in the model to obtain a corresponding weighted average vector;

[0034] The detection and classification module is used to perform binary classification processing on the weighted average vector through the classifier in the rumor detection model, and output the corresponding rumor detection result.

[0035] A third aspect of the present invention provides a rumor detection system, the system comprising at least one processor; and

[0036] a memory communicatively connected to the at least one processor; wherein,

[0037] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can perform the above-mentioned rumor detection method.

[0038] A fourth aspect of the present invention provides a non-volatile computer-readable storage medium, which stores computer-executable instructions. When the computer-executable instructions are executed by one or more processors, the one or more processors can execute the above-mentioned rumor detection method.

[0039] Beneficial effects: The present invention discloses a rumor detection method, device, system and medium. Compared with the prior art, the embodiments of the present invention collect multimodal information to be detected, and construct a corresponding social network graph according to the multimodal data; perform entity relationship fusion on the social network graph through a graph convolutional neural network, and update the node features of the social network graph; input the updated social network graph into a trained rumor detection model, and perform weighted averaging of the node vectors and edge vectors of each node in the updated social network graph through the Momentum parameter in the model to obtain a corresponding weighted average vector; perform binary classification processing on the weighted average vector through the classifier in the rumor detection model, and output the corresponding rumor detection result. The detection of multimodal information is supported by constructing a social network graph and graph convolution processing, and the Momentum parameter is introduced to perform weighted processing on node features, thereby increasing the accuracy and stability of rumor detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:

[0041] Figure 1 A flow chart of a rumor detection method provided by an embodiment of the present invention;

[0042] Figure 2 A flow chart of step S101 in the rumor detection method provided in an embodiment of the present invention;

[0043] Figure 3 A flow chart of training and obtaining Momentum parameters in the rumor detection method provided in an embodiment of the present invention;

[0044] Figure 4 A flow chart of step S104 in the rumor detection method provided by an embodiment of the present invention;

[0045] Figure 5 A schematic diagram of the functional modules of a rumor detection device provided by an embodiment of the present invention;

[0046] Figure 6 A schematic diagram of the hardware structure of a rumor detection system provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0047] In order to make the purpose, technical solution and effect of the present invention clearer and more specific, the present invention is further described in detail below. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. The embodiments of the present invention are described below in conjunction with the accompanying drawings.

[0048] The development of social media platforms has greatly promoted the spread of information. On the one hand, it has expanded the channels for Internet users to learn about information. On the other hand, it has led to the spread of unconfirmed events, namely Internet rumors. In the financial field, rumor detection is a relatively important technology. Through rumor detection, market fraud can be prevented, information transparency can be improved, and systemic risks can be reduced.

[0049] Existing rumor detection is often achieved through sentiment dictionaries, deep learning models, knowledge graphs, etc. However, existing methods often only use one source of information for training, and have poor generalization capabilities for multiple information channels. When faced with mixed information, the model appears relatively powerless, reducing the accuracy and stability of rumor detection.

[0050] In order to solve the above problems, the present invention proposes a rumor detection method. Figure 1 , Figure 1 A flowchart of an embodiment of the rumor detection method provided by the present invention. The system architecture of the rumor detection method provided by this embodiment may include a first terminal device, a second terminal device, a third terminal device, a network and a server. The network is used to provide a medium for a communication link between the first terminal device, the second terminal device, the third terminal device and the server. The network may include various connection types, such as wired and / or wireless communication links, etc.

[0051] The user may use the first terminal device, the second terminal device, and the third terminal device to interact with the server through the network to receive or send messages, etc. Various communication client applications may be installed on the first terminal device, the second terminal device, and the third terminal device, such as knowledge reading applications, web browser applications, search applications, instant messaging tools, email clients, and / or social platform software, etc. (only as examples).

[0052] The first terminal device, the second terminal device, and the third terminal device may be various electronic devices having display screens and supporting web browsing, including but not limited to smart phones, tablet computers, laptop computers, desktop computers, and the like.

[0053] The server may be a server that provides various services, such as a background management server that provides support for the content browsed by the user using the first terminal device, the second terminal device, and the third terminal device (for example only). The background management server may analyze and process the received data such as the user request, and feed back the processing results (such as web pages, information, or data obtained or generated according to the user request) to the terminal device. The server may be a cloud server, also known as a cloud computing server or a cloud host, which is a host product in the cloud computing service system to solve the defects of difficult management and weak business scalability in traditional physical hosts and VPS services ("Virtual Private Server", or "VPS" for short). The server may also be a server for a distributed system, or a server combined with a blockchain.

[0054] It should be noted that the rumor detection method provided in the embodiment of the present invention can generally be executed by the first terminal device, the second terminal device or the third terminal device. Correspondingly, the rumor detection device provided in the embodiment of the present invention can also be set in the first terminal device, the second terminal device or the third terminal device.

[0055] Alternatively, the rumor detection method provided by the embodiment of the present invention may also be generally executed by a server. Accordingly, the rumor detection device provided by the embodiment of the present invention may generally be arranged in a server. The rumor detection method provided by the embodiment of the present invention may also be executed by a server or a server cluster that is different from the server and can communicate with the first terminal device, the second terminal device, the third terminal device and / or the server. Accordingly, the rumor detection device provided by the embodiment of the present invention may also be arranged in a server or a server cluster that is different from the server and can communicate with the first terminal device, the second terminal device, the third terminal device and / or the server.

[0056] It should be understood that the numbers of the above terminal devices, networks and servers are only illustrative. Any number of terminal devices, networks and servers may be provided according to implementation requirements.

[0057] like Figure 1 As shown, the rumor detection method provided in this embodiment specifically includes the following steps:

[0058] S101, collecting multimodal information to be detected, and constructing a corresponding social network graph according to the multimodal data.

[0059] In this embodiment, multimodal information to be detected is collected from a social media platform. The multimodal information may include information dissemination content in different forms, such as text, images, videos, audio, etc.; different dissemination sources, such as dissemination on different social media platforms; and user interaction operations on the disseminated information, such as publishing, commenting, forwarding, etc., and rumor detection is performed by collecting multimodal information circulated on social media platforms, which is conducive to understanding the dissemination content from multiple perspectives. At the same time, a corresponding social network graph is constructed based on the multimodal information. The nodes and edges in the social network graph may contain rich features, such as user attributes, relationship types, interaction frequencies, etc., so that the multimodal dissemination data can be represented by the node features in the social network graph and the relationship between adjacent nodes. This can more deeply understand and analyze complex phenomena in social networks, provide support for rumor detection tasks, and improve the accuracy of rumor detection.

[0060] In one embodiment, Figure 2 As shown, the collecting of multimodal information to be detected and constructing a corresponding social network graph according to the multimodal data include:

[0061] S201, collecting multimodal information to be detected, where the multimodal information includes user information, propagation mode information, and propagation content information;

[0062] S202, extracting a user feature vector, a propagation mode vector and an information content vector according to the user information, the propagation mode information and the propagation content information;

[0063] S203: Use the user feature vector as a node vector, calculate the corresponding edge vector according to the propagation mode vector and the information content vector, and construct a corresponding social network graph through a graph neural network.

[0064] In this embodiment, a social network graph is constructed for the collected multimodal information through a graph neural network, where the collected multimodal information includes user information, dissemination mode information, and dissemination content information. For example, if there is a message circulating on a social media platform about "XYZ Company is about to go bankrupt", the publisher of this information, text information, image information about this message disseminated through other video channels, etc. can be collected to perform multi-angle rumor detection and analysis.

[0065] The collected user information, propagation mode information and propagation content information are respectively feature extracted to obtain the corresponding user feature vector, propagation mode vector and information content vector, wherein the user information can be randomly initialized or initialized based on the user's historical behavior data, such as the user's past forwarding behavior, comment content, etc. The extracted single user feature vector is represented by (1, hidden_size); the propagation mode vector is extracted based on the message source in the propagation mode information, and the propagation mode vector form_tensor is a tensor of (1, hidden_size), indicating different information sources. For example, if there are two sources of multimodal data (image and text), there are 2 propagation mode vectors in total (2, hidden_size); the information content vector info_tensor is also of (1, hidden_size) size, wherein the text information in the propagation content information can be obtained by passing the text through the embedding layer and the textcnn model, and the result is maxpooled to obtain (1, hidden_size), and the image information in the propagation content information can be obtained by pre-training the image VIT model to obtain a (1, hidden_size) vector. This enables feature extraction from different angles and methods of multimodal information.

[0066] A social network graph is constructed based on the extracted multimodal vectors. In specific implementation, social media users are taken as nodes, and interactions between users (such as forwarding and commenting) are taken as edges, that is, the user feature vector of (1, hidden_size) is taken as the node vector, and the corresponding edge vector is calculated according to the propagation mode vector and the information content vector. The corresponding social network graph is constructed through a graph neural network. When calculating the edge vector, the propagation mode vector and the information content vector are weightedly summed to obtain the corresponding edge vector. For example, the edge of the graph network is edge=0.3*form_tensor+0.7*info_tensor. The specific weights can be flexibly set according to needs, and this is not limited in this embodiment. The graph construction of multimodal information allows the model to simultaneously consider information from different modes, such as text, images or videos, when detecting rumors, providing a comprehensive feature representation for each edge. The edge vector is represented by a weighted sum of the propagation mode vector and the information content vector, so that the edge vector contains not only the characteristics of the information content, but also the characteristics of the propagation mode, such as how the information is forwarded or commented on and the source of the propagation. This helps the rumor detection model capture the rich characteristics of the information content and enhance the model's ability to express the characteristics of information propagation. Taking users as node vectors and defining edge vectors as a combination of propagation mode and information content can provide richer and more comprehensive information for social network analysis and rumor detection, thereby improving the reliability of rumor detection.

[0067] S102: Perform entity relationship fusion on the social network graph through a graph convolutional neural network to update node features of the social network graph.

[0068] In this embodiment, after constructing the social network graph, the constructed graph network is used as the input of the graph convolution neural network (GCN) to complete the fusion of entity relationships. The fusion of entity relationships specifically refers to integrating the relationship (edge) information between different entities (nodes) together to better understand the interactions and dependencies between entities. The graph convolution operation combines the features of the node with the features of its neighboring nodes, so that the original attributes of the node and the relationship information transmitted through the edge are fused together to form a richer node representation. Then update the node features of the social network graph. The specific graph convolution formula is: H^(l+1)=σ(D^-1 / 2*A*D^-1 / 2*H^l*W^l), where H^l represents the node feature matrix of the lth layer, W^l represents the weight matrix of the lth layer, A represents the adjacency matrix, D is the degree matrix, and σ represents the activation function. In rumor detection, the graph convolution operation can be used to merge the personal characteristics of user nodes in the social network graph and the edges that combine the propagation pattern and information content, and then update the user's feature representation to include the influence of other users in the social network. This can more comprehensively understand the information flow and user behavior in the social network, thereby more accurately identifying and filtering rumors.

[0069] S103, inputting the updated social network graph into the trained rumor detection model, performing weighted averaging on the node vectors and edge vectors of each node in the updated social network graph using the Momentum parameter in the model, and obtaining a corresponding weighted average vector.

[0070] In this embodiment, after obtaining updated node features through the graph convolutional neural network, the updated social network graph is input into the trained rumor detection model, and the node vectors and edge vectors of each node in the social network graph are weighted averaged by the Momentum parameter in the model. The Momentum parameter is a trainable parameter in the rumor detection model. By combining the Momentum parameter with the user's node vector and edge vector for weighted averaging, the obtained weighted average vector can more comprehensively characterize the user's behavioral characteristics and information content, thereby making more accurate judgments during rumor detection classification and improving the accuracy of rumor detection.

[0071] In one embodiment, Figure 3 As shown, the Momentum parameters are trained by the following steps:

[0072] S301, calculating the user rumor rate and initializing the Momentum parameter, where the Momentum parameter is used to adjust the gradient update direction and size of the model;

[0073] S302: When updating the gradient during model training, the learning rate is updated according to the current Momentum parameter and the user's rumor-making rate, and the Momentum parameter is adjusted by the updated learning rate until the model training is completed.

[0074] In this embodiment, the Momentum parameter is a trainable optimization parameter, specifically a parameter used to optimize the accelerated convergence of stochastic gradient descent, which is used to adjust the gradient update direction and size of the model. By introducing a momentum term, the update of the model parameters is accelerated. This momentum term predicts the gradient direction of the next step based on the gradient direction of the previous step, thereby making the parameter update smoother, reducing oscillation and accelerating convergence.

[0075] When the parameter is obtained during training, the Momentum parameter is initialized first. Specifically, the user rumor rate is calculated by the formula Urr = (1 + 2^2) * (Nrx + Ntx) / (2^2 * Nrx + Ntx), where Urr is the user rumor rate, Nrx is the number of received rumors, and Ntx is the number of sent rumors. The Momentum parameter is initialized to the user rumor rate. When the gradient is updated during training, the learning rate is updated according to the current Momentum parameter and the user rumor rate, and the Momentum parameter is adjusted by the updated learning rate until the model completes the training, that is, a variable learning rate is used for the trainable Momentum parameter, and the learning rate is updated specifically by the following formula:

[0076] learning_rate' = learning_rate * Momentum weight coefficient

[0077]

[0078] Among them, learning_rate' is the updated learning rate, learning_rate is the learning rate before the update, Momentum is the current Momentum parameter, and Urr is the user's rumor rate. During the training process, the update speed of the Momentum parameter can be adjusted according to the difference between it and the user's rumor rate, so as to obtain the Momentum parameter through more efficient training. By introducing the Momentum mechanism, the accuracy and stability of the model prediction are increased.

[0079] In specific implementation, the optimization of Momentum parameters can be performed at different stages, depending on the design of the model, available data, expected performance and resource constraints. For example, in the model training stage, the Momentum parameters are optimized for each user using available historical data. If enough user behavior data is accessed, the Momentum parameters learned by the model are designed to minimize the loss function of the rumor detection task by performing multiple iterations on the training set, so as to obtain the user's Momentum parameters before the model is deployed; for example, after the model is deployed, all users use the same initial Momentum parameters at the time of initial deployment. This initial value may be fixed or initialized based on the user's average rumor rate, etc.; once the model starts to receive real-time data, the Momentum parameters of each user are gradually adjusted according to the detection results of the model and the actual behavior of the user, so as to allow the model to learn and adapt in actual applications, closer to the current user behavior pattern. The specific optimization method can be flexibly selected according to factors such as data availability, real-time performance requirements, user behavior changes, etc., and of course, it can also be combined with pre-optimization and post-deployment optimization, that is, using historical data for preliminary optimization, and then further adjustment based on real-time feedback after the model is deployed. This embodiment does not limit this.

[0080] S104. Perform binary classification processing on the weighted average vector through the classifier in the rumor detection model, and output a corresponding rumor detection result.

[0081] In this embodiment, the node vectors and edge vectors of each node in the updated social network graph are weighted averaged by the Momentum parameter to obtain a weighted average vector of (1, hidden_size), and the weighted average vector is further classified by the classifier in the model to output the corresponding rumor detection result, that is, the detection result of whether the multimodal information to be detected is a rumor or a non-rumor. The specific classifier can be, for example, a fully connected layer plus a softmax activation function as a binary classification neural network structure; or a linear layer of a single output unit is used, without an activation function, and then a threshold (usually 0.5) is applied to determine the category; or a logistic regression model is used to convert the output into a probability using a sigmoid function, and then used for a binary classification problem, etc. The specific binary classification implementation method can be flexibly set according to demand, and this embodiment does not limit this. The detection of multimodal information is supported by constructing a social network graph and graph convolution processing, and the Momentum parameter is introduced to weight node features, thereby increasing the accuracy and stability of rumor detection.

[0082] In one embodiment, the classifier includes a fully connected layer and a softmax layer, such as Figure 4As shown, the weighted average vector is subjected to binary classification processing by the classifier in the rumor detection model, and the corresponding rumor detection result is output, including:

[0083] S401, inputting the weighted average vector into the fully connected layer, and outputting the original prediction values ​​of the rumor category and the non-rumor category;

[0084] S402. Input the original prediction values ​​of the rumor category and the non-rumor category into the softmax layer, convert the original prediction values ​​into corresponding probabilities through the softmax function, and select the category with the highest probability as the rumor detection result.

[0085] In this embodiment, the rumor detection model uses a fully connected layer and a softmax layer as a classifier, wherein the output of the fully connected layer is used as the input of the softmax layer, and the output of the softmax layer is used as the final prediction result of the model. First, the weighted average vector is input into the fully connected layer, the input tensor is multiplied by the weight matrix in the fully connected layer, and a bias term is added, and then an activation function is applied to obtain the original prediction value of the rumor category and the non-rumor category, namely the Logits value, which is the confidence or original prediction score of the model for each category, and then an activation function is applied to the original prediction value, that is, it is input into the softmax layer, and the Logits is converted into a probability distribution through the softmax function. The softmax function can map Logits to a real value between [0,1], and the sum of the probabilities of all categories is 1, and the category with the highest probability is selected as the final prediction result of the model. In rumor detection, the probability of the rumor category and the probability of the non-rumor category are compared, and the category with a larger probability is used as the rumor detection result of the current multimodal information, so as to achieve reliable and accurate rumor detection.

[0086] It should be noted that there is not necessarily a certain order between the above-mentioned steps. A person skilled in the art can understand, based on the description of the embodiments of the present invention, that in different embodiments, the above-mentioned steps may have different execution orders, that is, they may be executed in parallel, may be executed interchangeably, and so on.

[0087] Another embodiment of the present invention provides a rumor detection device, such as Figure 5 As shown, the device 1 comprises:

[0088] A graph construction module 11, used to collect multimodal information to be detected and construct a corresponding social network graph according to the multimodal data;

[0089] A graph convolution module 12, configured to perform entity relationship fusion on the social network graph through a graph convolution neural network, and update node features of the social network graph;

[0090] The weighted processing module 13 is used to input the updated social network graph into the trained rumor detection model, and perform weighted averaging on the node vectors and edge vectors of each node in the updated social network graph through the Momentum parameter in the model to obtain a corresponding weighted average vector;

[0091] The detection and classification module 14 is used to perform binary classification processing on the weighted average vector through the classifier in the rumor detection model, and output a corresponding rumor detection result.

[0092] The module referred to in the present invention refers to a series of computer program instruction segments that can perform specific functions, which is more suitable for describing the rumor detection execution process than a program. For the specific implementation of each module, please refer to the corresponding method embodiment above, which will not be repeated here.

[0093] In one embodiment, the graph construction module 11 includes:

[0094] A collection unit, used to collect multimodal information to be detected, wherein the multimodal information includes user information, propagation mode information and propagation content information;

[0095] A feature extraction unit, configured to extract a user feature vector, a propagation mode vector and an information content vector according to the user information, the propagation mode information and the propagation content information;

[0096] A graph construction unit is used to use the user feature vector as a node vector, calculate the corresponding edge vector according to the propagation mode vector and the information content vector, and construct a corresponding social network graph through a graph neural network.

[0097] In one embodiment, the graph construction unit is specifically used to:

[0098] A weighted sum is performed on the propagation mode vector and the information content vector to obtain a corresponding edge vector.

[0099] In one embodiment, the device 1 further comprises:

[0100] An initialization module, used to calculate the user rumor-mongering rate and initialize the Momentum parameter, wherein the Momentum parameter is used to adjust the gradient update direction and size of the model;

[0101] The learning rate update module is used to update the learning rate according to the current Momentum parameter and the user's rumor-making rate when updating the gradient during model training, and adjust the Momentum parameter by the updated learning rate until the model training is completed.

[0102] In one embodiment, the initialization module is specifically used to:

[0103] The user rumor rate is calculated by the formula Urr = (1 + 2^2) * (Nrx + Ntx) / (2^2 * Nrx + Ntx), where Urr is the user rumor rate, Nrx is the number of received rumors, and Ntx is the number of sent rumors;

[0104] Initialize the Momentum parameter to the user's rumor rate.

[0105] In one embodiment, the learning rate updating module is specifically used to:

[0106] Through the formula learning_rate'=learning_rate*Momentum weight coefficient

[0107]

[0108] Update the learning rate, where learning_rate' is the updated learning rate, learning_rate is the learning rate before the update, Momentum is the current Momentum parameter, and Urr is the user rumor rate.

[0109] In one embodiment, the detection and classification module 14 includes:

[0110] A fully connected unit, used for inputting the weighted average vector into the fully connected layer, and outputting original prediction values ​​of rumor categories and non-rumor categories;

[0111] The classification output unit is used to input the original prediction values ​​of the rumor category and the non-rumor category into the softmax layer, convert the original prediction values ​​into corresponding probabilities through the softmax function, and select the category with the highest probability as the rumor detection result.

[0112] Another embodiment of the present invention provides a rumor detection system. Figure 6 As shown, the system 10 includes:

[0113] One or more processors 110 and memory 120, Figure 6 A processor 110 is used as an example for description. The processor 110 and the memory 120 may be connected via a bus or other means. Figure 6 The example of connecting through bus is taken in the following.

[0114] The processor 110 is used to complete various control logics of the system 10, and it can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA), a single-chip microcomputer, an ARM (Acorn RISC Machine) or other programmable logic device, a discrete gate or transistor logic, a discrete hardware component, or any combination of these components. In addition, the processor 110 can also be any traditional processor, microprocessor or state machine. The processor 110 can also be implemented as a combination of computing devices, for example, a combination of a DSP and a microprocessor, a plurality of microprocessors, one or more microprocessors combined with a DSP and / or any other such configuration.

[0115] The memory 120 is a non-volatile computer-readable storage medium that can be used to store non-volatile software programs, non-volatile computer executable programs and modules, such as program instructions corresponding to the rumor detection method in the embodiment of the present invention. The processor 110 executes various functional applications and data processing of the system 10 by running the non-volatile software programs, instructions and units stored in the memory 120, that is, implementing the rumor detection method in the above method embodiment.

[0116] The memory 120 may include a program storage area and a data storage area, wherein the program storage area may store an operating system and an application required by at least one function; the data storage area may store data created according to the use of the system 10, etc. In addition, the memory 120 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some embodiments, the memory 120 may optionally include a memory remotely arranged relative to the processor 110, and these remote memories may be connected to the system 10 via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0117] One or more units are stored in the memory 120, and when executed by one or more processors 110, the rumor detection method in any of the above method embodiments is executed, for example, the following steps are executed:

[0118] Collecting multimodal information to be detected, and constructing a corresponding social network graph based on the multimodal data;

[0119] Performing entity relationship fusion on the social network graph through a graph convolutional neural network to update node features of the social network graph;

[0120] The updated social network graph is input into the trained rumor detection model, and the node vector and edge vector of each node in the updated social network graph are weighted averaged through the Momentum parameter in the model to obtain the corresponding weighted average vector;

[0121] The weighted average vector is subjected to binary classification processing by the classifier in the rumor detection model, and a corresponding rumor detection result is output.

[0122] In one embodiment, the collecting of the multimodal information to be detected and constructing a corresponding social network graph according to the multimodal data includes:

[0123] Collecting multimodal information to be detected, wherein the multimodal information includes user information, propagation mode information and propagation content information;

[0124] According to the user information, the propagation mode information and the propagation content information, a user feature vector, a propagation mode vector and an information content vector are respectively extracted;

[0125] The user feature vector is used as a node vector, and the corresponding edge vector is calculated according to the propagation mode vector and the information content vector, and the corresponding social network graph is constructed through a graph neural network.

[0126] In one embodiment, the calculating the corresponding edge vector according to the propagation mode vector and the information content vector specifically includes:

[0127] A weighted sum is performed on the propagation mode vector and the information content vector to obtain a corresponding edge vector.

[0128] In one embodiment, the Momentum parameter is trained by the following steps:

[0129] Calculate the user rumor rate and initialize the Momentum parameter, where the Momentum parameter is used to adjust the gradient update direction and size of the model;

[0130] When updating the gradient during model training, the learning rate is updated according to the current Momentum parameter and the user's rumor-making rate, and the Momentum parameter is adjusted by the updated learning rate until the model training is completed.

[0131] In one embodiment, calculating the user rumor rate and initializing the Momentum parameter specifically includes:

[0132] The user rumor rate is calculated by the formula Urr = (1 + 2^2) * (Nrx + Ntx) / (2^2 * Nrx + Ntx), where Urr is the user rumor rate, Nrx is the number of received rumors, and Ntx is the number of sent rumors;

[0133] Initialize the Momentum parameter to the user's rumor rate.

[0134] In one embodiment, updating the learning rate according to the current Momentum parameter and the user rumor spreading rate specifically includes:

[0135] Through the formula learning_rate'=learning_rate*Momentum weight coefficient

[0136]

[0137] Update the learning rate, where learning_rate' is the updated learning rate, learning_rate is the learning rate before the update, Momentum is the current Momentum parameter, and Urr is the user rumor rate.

[0138] In one embodiment, the classifier includes a fully connected layer and a softmax layer, and the classifier in the rumor detection model performs binary classification processing on the weighted average vector and outputs a corresponding rumor detection result, including:

[0139] Inputting the weighted average vector into the fully connected layer, and outputting the original prediction values ​​of the rumor category and the non-rumor category;

[0140] The original prediction values ​​of the rumor category and the non-rumor category are input into the softmax layer, and the original prediction values ​​are converted into corresponding probabilities through the softmax function, and the category with the highest probability is selected as the rumor detection result.

[0141] An embodiment of the present invention provides a non-volatile computer-readable storage medium, which stores computer-executable instructions. When the computer-executable instructions are executed by one or more processors, the rumor detection method in any of the above method embodiments is executed, for example, the following steps are performed:

[0142] Collecting multimodal information to be detected, and constructing a corresponding social network graph based on the multimodal data;

[0143] Performing entity relationship fusion on the social network graph through a graph convolutional neural network to update node features of the social network graph;

[0144] The updated social network graph is input into the trained rumor detection model, and the node vector and edge vector of each node in the updated social network graph are weighted averaged through the Momentum parameter in the model to obtain the corresponding weighted average vector;

[0145] The weighted average vector is subjected to binary classification processing by the classifier in the rumor detection model, and a corresponding rumor detection result is output.

[0146] In one embodiment, the collecting of the multimodal information to be detected and constructing a corresponding social network graph according to the multimodal data includes:

[0147] Collecting multimodal information to be detected, wherein the multimodal information includes user information, propagation mode information and propagation content information;

[0148] According to the user information, the propagation mode information and the propagation content information, a user feature vector, a propagation mode vector and an information content vector are respectively extracted;

[0149] The user feature vector is used as a node vector, and the corresponding edge vector is calculated according to the propagation mode vector and the information content vector, and the corresponding social network graph is constructed through a graph neural network.

[0150] In one embodiment, the calculating the corresponding edge vector according to the propagation mode vector and the information content vector specifically includes:

[0151] A weighted sum is performed on the propagation mode vector and the information content vector to obtain a corresponding edge vector.

[0152] In one embodiment, the Momentum parameter is trained by the following steps:

[0153] Calculate the user rumor rate and initialize the Momentum parameter, where the Momentum parameter is used to adjust the gradient update direction and size of the model;

[0154] When updating the gradient during model training, the learning rate is updated according to the current Momentum parameter and the user's rumor-making rate, and the Momentum parameter is adjusted by the updated learning rate until the model training is completed.

[0155] In one embodiment, calculating the user rumor rate and initializing the Momentum parameter specifically includes:

[0156] The user rumor rate is calculated by the formula Urr = (1 + 2^2) * (Nrx + Ntx) / (2^2 * Nrx + Ntx), where Urr is the user rumor rate, Nrx is the number of received rumors, and Ntx is the number of sent rumors;

[0157] Initialize the Momentum parameter to the user's rumor rate.

[0158] In one embodiment, updating the learning rate according to the current Momentum parameter and the user rumor spreading rate specifically includes:

[0159] Through the formula learning_rate'=learning_rate*Momentum weight coefficient

[0160]

[0161] Update the learning rate, where learning_rate' is the updated learning rate, learning_rate is the learning rate before the update, Momentum is the current Momentum parameter, and Urr is the user rumor rate.

[0162] In one embodiment, the classifier includes a fully connected layer and a softmax layer, and the classifier in the rumor detection model performs binary classification processing on the weighted average vector and outputs a corresponding rumor detection result, including:

[0163] Inputting the weighted average vector into the fully connected layer, and outputting the original prediction values ​​of the rumor category and the non-rumor category;

[0164] The original prediction values ​​of the rumor category and the non-rumor category are input into the softmax layer, and the original prediction values ​​are converted into corresponding probabilities through the softmax function, and the category with the highest probability is selected as the rumor detection result.

[0165] As an example, non-volatile storage media can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) as external cache memory. By way of illustration and not limitation, RAM can be obtained in many forms such as synchronous RAM (SRAM), dynamic RAM, (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), Synchlink DRAM (SLDRAM), and direct Rambus RAM (DRRAM). The disclosed memory components or memories of the operating environment described herein are intended to include one or more of these and / or any other suitable types of memory.

[0166] In summary, the present invention discloses a rumor detection method, device, system and medium, the method includes: collecting multimodal information to be detected, constructing a corresponding social network graph according to the multimodal data; performing entity relationship fusion on the social network graph through a graph convolutional neural network, and updating the node features of the social network graph; inputting the updated social network graph into a trained rumor detection model, performing weighted averaging on the node vectors and edge vectors of each node in the updated social network graph through the Momentum parameter in the model, and obtaining a corresponding weighted average vector; performing binary classification processing on the weighted average vector through the classifier in the rumor detection model, and outputting the corresponding rumor detection result. The detection of multimodal information is supported by constructing a social network graph and graph convolution processing, and the Momentum parameter is introduced to perform weighted processing on node features, thereby increasing the accuracy and stability of rumor detection.

[0167] Of course, those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing related hardware (such as a processor, a controller, etc.) through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium, and the computer program can include the processes of the above-mentioned method embodiments when executed. The storage medium can be a memory, a disk, a floppy disk, a flash memory, an optical storage device, etc.

[0168] It should be understood that the application of the present invention is not limited to the above examples. For ordinary technicians in this field, improvements or changes can be made based on the above description. All these improvements and changes should fall within the scope of protection of the claims attached to the present invention.

Claims

1. A rumor detection method, characterized in that: include: Collecting multimodal information to be detected, and constructing a corresponding social network graph based on the multimodal data; Performing entity relationship fusion on the social network graph through a graph convolutional neural network to update node features of the social network graph; The updated social network graph is input into the trained rumor detection model, and the node vector and edge vector of each node in the updated social network graph are weighted averaged through the Momentum parameter in the model to obtain the corresponding weighted average vector; The weighted average vector is subjected to binary classification processing by the classifier in the rumor detection model, and a corresponding rumor detection result is output.

2. The rumor detection method according to claim 1, characterized in that: The collecting of the multimodal information to be detected and constructing a corresponding social network graph according to the multimodal data includes: Collecting multimodal information to be detected, wherein the multimodal information includes user information, propagation mode information and propagation content information; According to the user information, the propagation mode information and the propagation content information, a user feature vector, a propagation mode vector and an information content vector are respectively extracted; The user feature vector is used as a node vector, and the corresponding edge vector is calculated according to the propagation mode vector and the information content vector, and the corresponding social network graph is constructed through a graph neural network.

3. The rumor detection method according to claim 2, characterized in that: The calculating of the corresponding edge vector according to the propagation mode vector and the information content vector specifically includes: A weighted sum is performed on the propagation mode vector and the information content vector to obtain a corresponding edge vector.

4. The rumor detection method according to claim 1, characterized in that: The Momentum parameters are trained by the following steps: Calculate the user rumor rate and initialize the Momentum parameter, where the Momentum parameter is used to adjust the gradient update direction and size of the model; When updating the gradient during model training, the learning rate is updated according to the current Momentum parameter and the user's rumor-making rate, and the Momentum parameter is adjusted by the updated learning rate until the model training is completed.

5. The rumor detection method according to claim 4, characterized in that: The calculating of the user rumor spreading rate and initializing the Momentum parameter specifically includes: The user rumor rate is calculated by the formula Urr = (1 + 2^2) * (Nrx + Ntx) / (2^2 * Nrx + Ntx), where Urr is the user rumor rate, Nrx is the number of received rumors, and Ntx is the number of sent rumors; Initialize the Momentum parameter to the user's rumor rate.

6. The rumor detection method according to claim 5, characterized in that: The updating of the learning rate according to the current Momentum parameter and the user's rumor-mongering rate specifically includes: Through the formula learning_rate'=learning_rate*Momentum weight coefficient Update the learning rate, where learning_rate' is the updated learning rate, learning_rate is the learning rate before the update, Momentum is the current Momentum parameter, and Urr is the user's rumor rate.

7. The rumor detection method according to claim 1, characterized in that: The classifier includes a fully connected layer and a softmax layer. The classifier in the rumor detection model performs binary classification on the weighted average vector and outputs a corresponding rumor detection result, including: Inputting the weighted average vector into the fully connected layer, and outputting the original prediction values ​​of the rumor category and the non-rumor category; The original prediction values ​​of the rumor category and the non-rumor category are input into the softmax layer, and the original prediction values ​​are converted into corresponding probabilities through the softmax function, and the category with the highest probability is selected as the rumor detection result.

8. A rumor detection device, characterized in that: include: A graph construction module, used to collect multimodal information to be detected and construct a corresponding social network graph based on the multimodal data; A graph convolution module, used to perform entity relationship fusion on the social network graph through a graph convolutional neural network, and update node features of the social network graph; A weighted processing module is used to input the updated social network graph into the trained rumor detection model, and perform weighted averaging of the node vectors and edge vectors of each node in the updated social network graph through the Momentum parameter in the model to obtain a corresponding weighted average vector; The detection and classification module is used to perform binary classification processing on the weighted average vector through the classifier in the rumor detection model, and output the corresponding rumor detection result.

9. A rumor detection system, characterized in that: The system includes at least one processor; and, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the rumor detection method described in any one of claims 1-7.

10. A non-volatile computer-readable storage medium, characterized in that: The non-volatile computer-readable storage medium stores computer-executable instructions, which, when executed by one or more processors, enable the one or more processors to execute the rumor detection method described in any one of claims 1-7.

Citation Information

Patent Citations

  • Social media multi-modal rumor detection method based on propagation heterogeneous graph modeling

    CN112035669A

  • Network information detection method and device based on information propagation tree, equipment and medium

    CN114722254A

  • Method and device for performing video retrieval by using text

    CN116383439A

  • Early rumor detection method and device, equipment and medium

    CN116861863A