A Network Rumor Detection Method Based on Spatiotemporal Feature Learning of Dual Dynamic Graph Convolutional Networks

Abstract

This invention proposes a network rumor detection method based on spatiotemporal feature learning using a dual dynamic graph convolutional network. The method includes: transforming events into a predetermined number of propagation structures based on time; constructing user interaction graphs and information propagation graphs using these propagation structures; learning graph features from the user interaction graphs and information propagation graphs using a convolutional network; concatenating the graph features from the previous layer with the feature vectors of the root nodes of the next layer to obtain enhanced graph features; obtaining the output features of a temporal fusion unit; further obtaining the final features of the event based on the output features of the temporal fusion unit; and finally, obtaining the classification result using a multilayer perceptron. This invention not only comprehensively captures the spatiotemporal features of information propagation and user interaction but also effectively integrates content features, propagation features, and user features, improving the accuracy of rumor detection, especially showing significant advantages in the early detection stage.

Description

Technical Field

[0001] This invention relates to the fields of computer science and artificial intelligence, and in particular to a method for detecting online rumors based on spatiotemporal feature learning of dual dynamic graph convolutional networks. Background Technology

[0002] The rapid development of social networks has brought about a transformation in the way information is disseminated; however, this has also led to the problem of the spread of online rumors. Therefore, researching and developing online rumor detection technologies, especially automated rumor identification based on social network data, has become an important research topic.

[0003] Most existing rumor detection methods focus on text content analysis, using techniques such as sentiment analysis and keyword extraction to identify rumors from message content. However, with the increasingly sophisticated information disguise capabilities of rumor spreaders, relying solely on content features is no longer effective in identifying rumors. In recent years, rumor detection methods based on propagation characteristics have gradually become a research hotspot. Graph Neural Networks (GNNs), as an emerging graph-structured data processing method, can effectively learn the propagation structure characteristics in networks and has achieved good results in rumor detection on social networks.

[0004] Existing GNN-based rumor detection methods primarily focus on the spatial structural features of graphs, capturing information propagation paths by learning the relationships between nodes and edges. However, most of these methods neglect the temporal structural features of information propagation, failing to fully consider the temporal evolution and dynamic changes of information during propagation. Furthermore, although some studies have attempted to combine spatiotemporal features to utilize the propagation temporal structure for rumor detection, these methods often ignore the spatiotemporal characteristics of user interactions, failing to fully explore the impact of user interaction patterns on rumor propagation. Summary of the Invention

[0005] In view of the above situation, the main objective of this invention is to propose a network rumor detection method based on spatiotemporal feature learning of dual dynamic graph convolutional networks, so as to solve the above-mentioned technical problems.

[0006] This invention proposes a method for detecting online rumors based on spatiotemporal feature learning using dual dynamic graph convolutional networks. The method includes the following steps:

[0007] Step 1: Divide the event into a preset number of time-series states based on time, and then combine the relationship between forwarding and commenting to transform each time-series state into a propagation structure;

[0008] Step 2: Obtain users' social characteristics, and combine these characteristics with the propagation structure to construct a user interaction graph;

[0009] Obtain the original posts and response information, and combine the obtained original posts, response information, and propagation structure to construct an information propagation graph;

[0010] Step 3: Use two different convolutional networks to learn the user interaction graph and the information propagation graph respectively, and obtain the features of the user interaction graph and the information propagation graph respectively;

[0011] Step 4: For the user interaction graph features, concatenate the features of the previous layer with the root node feature vector of the next layer to obtain the enhanced user interaction graph features.

[0012] For information propagation graph features, the features of the previous layer are concatenated with the feature vector of the root node of the next layer to obtain the enhanced information propagation graph features.

[0013] Step 5: Construct a temporal fusion unit based on the cross-attention mechanism. Input the enhanced user interaction graph features and the enhanced user interaction graph features as input sequences into the temporal fusion unit to obtain the output features of the temporal fusion unit.

[0014] Step 6: Perform global average pooling on the output features of the temporal fusion unit to obtain the output features after global average pooling;

[0015] Semantic features are obtained based on the initial features in the information propagation graph and combined with the pre-trained and pre-drilled model;

[0016] The output features after global average pooling are concatenated with the semantic features to obtain the final features of the event.

[0017] Step 7: Input the final features of the event into the multilayer perceptron, and then process them through the activation function to obtain the classification result.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0019] 1. This invention constructs a dual dynamic graph convolutional network, which can comprehensively capture the spatiotemporal features of information dissemination and user interaction, and effectively integrate text, dissemination and user features, significantly improving the detection accuracy of online rumors;

[0020] 2. This invention introduces a cross-attention mechanism for temporal fusion, which enhances the model's ability to perceive temporal changes and is particularly suitable for the early detection of rumors;

[0021] 3. This invention improves the robustness of the model by using a dual dynamic graph convolutional network, enabling it to cope with complex social network data and enhance the generalization ability and stability of rumor detection. This effectively solves the problems of insufficient multimodal feature fusion and difficulty in capturing temporal changes in existing methods.

[0022] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by means of embodiments of the invention. Attached Figure Description

[0023] Figure 1 This is a flowchart of the network rumor detection method based on spatiotemporal feature learning of dual dynamic graph convolutional networks proposed in this invention;

[0024] Figure 2 This is a model architecture diagram of the network rumor detection method based on spatiotemporal feature learning of dual dynamic graph convolutional networks proposed in this invention;

[0025] Figure 3 This is a flowchart of the cross-attention calculation process for the network rumor detection method based on spatiotemporal feature learning of dual dynamic graph convolutional networks proposed in this invention. Detailed Implementation

[0026] Embodiments of the present invention are described in detail below. Examples of these embodiments are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0027] These and other aspects of the embodiments of the present invention will become clear from the following description and accompanying drawings. In these descriptions and drawings, some specific embodiments of the present invention are specifically disclosed to illustrate some ways of implementing the principles of the embodiments of the present invention; however, it should be understood that the scope of the embodiments of the present invention is not limited thereto.

[0028] Please see Figure 1 , Figure 2 This invention proposes a method for detecting online rumors based on spatiotemporal feature learning of dual dynamic graph convolutional networks. The method includes the following steps:

[0029] Step 1: Divide the event into a preset number of time-series states based on time, and then combine the relationship between forwarding and commenting to transform each time-series state into a propagation structure;

[0030] Specifically, in this step, the time of the last response information of the event and the time of the original post are recorded, and the time span is statistically calculated based on the two times. The time span is then divided into a certain number of time intervals. Subsequently, the time is divided into multiple time-series states according to the time intervals, and each time-series state is transformed into a propagation structure according to the relationship between forwarding and commenting.

[0031] Step 2: Obtain users' social characteristics, and combine these characteristics with the propagation structure to construct a user interaction graph;

[0032] Obtain the original posts and response information, and combine the obtained original posts, response information, and propagation structure to construct an information propagation graph;

[0033] Specifically, in this step, both the user interaction graph and the information propagation graph are composed of a set of nodes, a set of edges, and an adjacency matrix.

[0034] Step 3: Use two different convolutional networks to learn the user interaction graph and the information propagation graph respectively, and obtain the features of the user interaction graph and the information propagation graph respectively;

[0035] In step 3, two different convolutional networks are used to learn the user interaction graph and the information propagation graph respectively, obtaining the features of the user interaction graph and the information propagation graph. The corresponding relationship in the process is as follows:

[0036] ;

[0037] in, Indicates the first Layered user interaction graph features Indicates the first Features of layered information propagation graphs This indicates that the process has been performed using an activation function. Represents a normalized symmetric adjacency matrix. Indicates the first Layered user interaction graph features Indicates the first Features of layered information propagation graphs , , and All indicate the first Trainable parameters in a multilayer convolutional network Indicates an event, Indicates timing state;

[0038] The definition of a normalized symmetric adjacency matrix is:

[0039] ;

[0040] in, Represents the adjacency matrix. Represents the identity matrix. Degree matrix, This represents the matrix obtained by adding the adjacency matrix and the identity matrix.

[0041] Step 4: For the user interaction graph features, concatenate the features of the previous layer with the root node feature vector of the next layer to obtain the enhanced user interaction graph features.

[0042] For information propagation graph features, the features of the previous layer are concatenated with the feature vector of the root node of the next layer to obtain the enhanced information propagation graph features.

[0043] In step 4, for the user interaction graph features, the features from the previous layer are concatenated with the root node feature vector of the next layer to obtain the enhanced user interaction graph features. The corresponding relationship in this process is as follows:

[0044] ;

[0045] in, Indicates the first Enhanced user interaction graph features This indicates a splicing operation. Indicates the first The feature vector of the root node of the layer user interaction graph;

[0046] For information propagation graph features, the features of the previous layer are concatenated with the feature vector of the root node of the next layer to obtain the enhanced information propagation graph features. The corresponding relationship in this process is as follows:

[0047] ;

[0048] in, Indicates the first Features of the information propagation graph after layer enhancement Indicates the first The feature vector of the root node in the layer information propagation graph;

[0049] Specifically, this step obtains enhanced user interaction graph features and enhanced information propagation graph features to fully highlight the role of the user information and semantic information of the original post in rumor detection.

[0050] Step 5: Construct a temporal fusion unit based on the cross-attention mechanism. Input the enhanced user interaction graph features and the enhanced user interaction graph features as input sequences into the temporal fusion unit to obtain the output features of the temporal fusion unit.

[0051] Please see Figure 3 In step 5, a temporal fusion unit is constructed based on the cross-attention mechanism. The definition of the temporal fusion unit is:

[0052] ;

[0053] in, This indicates that cross-attention processing has been performed. Indicates that it comes from the input sequence The query, Indicates that it comes from the input sequence The transpose of the keyword. Indicates that it comes from the input sequence The value, The feature dimensions of keywords and These represent two different input sequences;

[0054] The enhanced user interaction graph features and the enhanced user interaction graph features are used as input sequences to the time-series fusion unit to obtain the output features of the time-series fusion unit. The corresponding relationship in the process is as follows:

[0055] ;

[0056] in, Indicates the first Output characteristics of the temporal fusion unit at each time step. Indicates the first Output characteristics of the temporal fusion unit at each time step. Indicates the first Enhanced user interaction graph features at different times. Indicates the first The enhanced information propagation graph features at any given moment.

[0057] Specifically, this step uses a temporal fusion unit to fuse two dynamic spatiotemporal features, and inputs the fusion result into the graph convolutional network and temporal fusion unit at the next time step to achieve temporal interaction between the two dynamic graphs, thereby further enhancing the model's ability to perceive temporal changes.

[0058] Furthermore, in Figure 3 middle, and Both represent the sequence length. , , and Both represent feature dimensions. , and All of these represent learning parameters.

[0059] Step 6: Perform global average pooling on the output features of the temporal fusion unit to obtain the output features after global average pooling;

[0060] Semantic features are obtained based on the initial features in the information propagation graph and combined with the pre-trained and pre-drilled model;

[0061] The output features after global average pooling are concatenated with the semantic features to obtain the final features of the event.

[0062] In step 6, the output features of the temporal fusion unit are subjected to global average pooling to obtain the output features after global average pooling. The corresponding relationship in this process is as follows:

[0063] ;

[0064] in, This represents the output features after global average pooling. This indicates an average pooling operation;

[0065] The output features after global average pooling are concatenated with the semantic features to obtain the final features of the event. The corresponding relationship in this process is as follows:

[0066] ;

[0067] in, Representing the ultimate characteristic of time, Indicates semantic features.

[0068] Step 7: Input the final features of the event into the multilayer perceptron, and then process them through the activation function to obtain the classification result;

[0069] In step 7, the final features of the event are input into the multilayer perceptron, and then processed by the activation function to obtain the classification result. The corresponding relationship in this process is as follows:

[0070] ;

[0071] in, Indicates the classification result. This indicates that the process is handled by a multilayer perceptron. This indicates that the process has been performed using an activation function. The parameters representing the linear layer, This represents the bias term for the linear layer.

[0072] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0073] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0074] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for detecting online rumors based on spatiotemporal feature learning of dual dynamic graph convolutional networks, characterized in that, The method includes the following steps: Step 1: Divide the event into a preset number of time-series states based on time, and then combine the relationship between forwarding and commenting to transform each time-series state into a propagation structure; Step 2: Obtain users' social characteristics, and combine these characteristics with the propagation structure to construct a user interaction graph; Obtain the original posts and response information, and combine the obtained original posts, response information, and propagation structure to construct an information propagation graph; Step 3: Use two different convolutional networks to learn the user interaction graph and the information propagation graph respectively, and obtain the features of the user interaction graph and the information propagation graph respectively; Step 4: For the user interaction graph features, concatenate the features of the previous layer with the root node feature vector of the next layer to obtain the enhanced user interaction graph features. For information propagation graph features, the features of the previous layer are concatenated with the feature vector of the root node of the next layer to obtain the enhanced information propagation graph features. Step 5: Construct a temporal fusion unit based on the cross-attention mechanism. Input the enhanced user interaction graph features and the enhanced user interaction graph features as input sequences into the temporal fusion unit to obtain the output features of the temporal fusion unit. Step 6: Perform global average pooling on the output features of the temporal fusion unit to obtain the output features after global average pooling; Semantic features are obtained based on the initial features in the information propagation graph and combined with the pre-trained and pre-drilled model; The output features after global average pooling are concatenated with the semantic features to obtain the final features of the event. Step 7: Input the final features of the event into the multilayer perceptron, and then process them through the activation function to obtain the classification result; In step 3, two different convolutional networks are used to learn the user interaction graph and the information propagation graph respectively, obtaining the user interaction graph features and the information propagation graph features respectively. The corresponding relationship in the process is as follows: ； in, Indicates the first Layered user interaction graph features Indicates the first Features of layered information propagation graphs This indicates that the process has been performed using an activation function. Represents a normalized symmetric adjacency matrix. Indicates the first Layered user interaction graph features Indicates the first Features of layered information propagation graphs , , and All indicate the first Trainable parameters in a multilayer convolutional network Indicates an event, Indicates timing state; In the process of learning the user interaction graph and the information propagation graph using two different convolutional networks respectively, and obtaining the features of the user interaction graph and the information propagation graph respectively, the definition of the normalized symmetric adjacency matrix is: ； in, Represents the adjacency matrix. Represents the identity matrix. Degree matrix, This represents the matrix obtained by adding the adjacency matrix and the identity matrix; In step 4, for the user interaction graph features, the features of the previous layer are concatenated with the root node feature vector of the next layer to obtain the enhanced user interaction graph features. The corresponding relationship in this process is as follows: ； in, Indicates the first Enhanced user interaction graph features This indicates a splicing operation. Indicates the first The feature vector of the root node of the layer user interaction graph; In step 4, for the information propagation graph features, the features of the previous layer are concatenated with the root node feature vector of the next layer to obtain the enhanced information propagation graph features. The corresponding relationship in this process is as follows: ； in, Indicates the first Features of the information propagation graph after layer enhancement Indicates the first The feature vector of the root node in the layer information propagation graph; In step 5, a temporal fusion unit is constructed based on the cross-attention mechanism. The definition of the temporal fusion unit is: ; in, This indicates that cross-attention processing has been performed. Indicates that it comes from the input sequence The query, Indicates that it comes from the input sequence The transpose of the keyword. Indicates that it comes from the input sequence The value, The feature dimensions of keywords and These represent two different input sequences; In step 5, the enhanced user interaction graph features and the enhanced user interaction graph features are used as input sequences to the time-series fusion unit to obtain the output features of the time-series fusion unit. The corresponding relationship in the process is as follows: ； in, Indicates the first Output characteristics of the temporal fusion unit at each time step. Indicates the first Output characteristics of the temporal fusion unit at each time step. Indicates the first Enhanced user interaction graph features at different times. Indicates the first Enhanced information propagation graph features at any given moment; In step 6, the output features of the temporal fusion unit are subjected to global average pooling to obtain the output features after global average pooling. The corresponding relationship in this process is as follows: ； in, This represents the output features after global average pooling. This indicates an average pooling operation; In step 6, the output features after global average pooling are concatenated with the semantic features to obtain the final features of the event. The corresponding relationship in this process is as follows: ； in, Representing the ultimate characteristic of time, Represents semantic features; In step 7, the final features of the event are input into a multilayer perceptron, and then processed by an activation function to obtain the classification result. The corresponding relationship in this process is as follows: ； in, Indicates the classification result. This indicates that the process is handled by a multilayer perceptron. This indicates that the process has been performed using an activation function. The parameters representing the linear layer, This represents the bias term for the linear layer.

Images