Dynamic graph-based rumor detection method, system and device, product and medium

Abstract

The invention relates to the technical field of artificial intelligence, and provides a rumor detection method, system and device based on a dynamic graph, a product and a medium, and the method comprises the steps: obtaining rumor data, obtaining a rumor text sequence from the rumor data, and constructing a rumor propagation graph; dividing the rumor text sequence to establish a snapshot propagation graph of rumor snapshots; calculating a time decay parameter of a rumor node in the snapshot propagation graph, and calculating a rumor period parameter of the rumor snapshot through the time decay parameter; the rumor period parameters are fused into a GIN encoder, an edge weight GIN encoder is obtained, multi-layer node embedding updating is conducted on rumor snapshots in the snapshot propagation graph through the edge weight GIN encoder, and a snapshot dynamic propagation graph is obtained; inputting the snapshot dynamic propagation graph into a bidirectional long short-term memory network to obtain a coding graph representation of the rumor snapshot; and inputting the coding graph representation into the deep neural network and completing rumor detection. The rumor detection effect can be effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and particularly to a rumor detection method, system, device, product and medium based on a dynamic graph. Background Art

[0002] Social media has become a major channel for people to obtain news, express opinions and conduct social interactions. However, it also brings potential risks, especially the proliferation and rapid spread of rumors, which pose a serious threat to social stability, public trust and economic order. Therefore, it is particularly important to detect rumors on social media in a timely and effective manner.

[0003] Traditional rumor detection methods usually rely on manually designed features and machine learning classifiers, and these methods are limited by the subjectivity and limitations of feature design. With the progress of deep learning technology, researchers have begun to detect rumors from multiple perspectives such as text content, user behavior and propagation structure. In particular, the detection method based on the propagation structure, and the graph neural network is used to model the information propagation process. Among them, the dynamic graph method considering time information has achieved remarkable detection results. However, existing rumor detection models using dynamic graphs usually only focus on coarse-grained time information and ignore the fine-grained time dynamics within and between individual snapshots.

[0004] The rumor detection method based on the fine-grained dynamic graph neural network is an innovative research issue with important research significance and application prospects. Summary of the Invention

[0005] The present invention aims to at least solve one of the technical problems existing in the related art. For this purpose, the present invention provides a rumor detection method, system, device, product and medium based on a dynamic graph, so as to achieve rapid and accurate detection of online rumors.

[0006] The present invention provides a rumor detection method based on a dynamic graph, including: S1: Obtain rumor data, obtain a rumor text sequence and a rumor source node from the rumor data, and construct a rumor propagation graph through the rumor source node; S2: Divide the rumor text sequence to obtain a segmented rumor sequence, and establish a snapshot propagation graph of the rumor snapshot through the segmented rumor sequence and the rumor propagation graph; S3: Calculate the time decay parameter of the rumor node in the snapshot propagation graph, and calculate the rumor cycle parameter of the rumor node through the time decay parameter; S4: Obtain a GIN encoder, fuse the rumor cycle parameter into the GIN encoder to obtain an edge weight GIN encoder, and perform multi-layer node embedding update on the rumor snapshot through the edge weight GIN encoder to obtain a snapshot dynamic propagation graph; S5: Establish a bidirectional long short-term memory network, input the snapshot dynamic propagation graph into the bidirectional long short-term memory network, and obtain the encoded graph representation of the rumor snapshot; S6: Obtain a deep neural network, input the encoded graph representation into the deep neural network, and complete rumor detection through the deep neural network.

[0007] According to the rumor detection method based on a dynamic graph provided by the present invention, step S1 further includes: S11: Determine the target software, obtain rumor data from the target software, detect the rumor data to obtain potential rumor events, extract and sort rumor texts related to the potential rumor events from the rumor data to obtain a rumor text sequence; S12: Use the rumor text with the earliest time in the rumor text sequence as the rumor source node, and construct the rumor propagation graph according to the generation time and propagation order of the rumor text.

[0008] According to the rumor detection method based on a dynamic graph provided by the present invention, step S2 further includes: S21: Extract the timestamps of the rumor text sequence, divide the rumor text sequence according to the timestamps to obtain a segmented rumor sequence including rumor text blocks; S22: Divide and texturize the rumor propagation graph through the rumor text blocks in the segmented rumor sequence to obtain the rumor snapshot, and obtain the snapshot propagation graph according to the rumor snapshot.

[0009] According to the rumor detection method based on a dynamic graph provided by the present invention, step S4 further includes: S41: Obtain a GIN encoder, fuse the rumor period parameter into the GIN encoder to obtain an edge-weighted GIN encoder, and the expression of the edge-weighted GIN encoder is: where, is the hidden feature vector of the m-th layer of the v-th rumor node, represents the encoder parameter of the hidden feature vector of the m-th layer that can be learned, is the hidden feature vector of the (m - 1)-th layer of the v-th rumor node, is the rumor period parameter of the v-th rumor node, is the hidden feature vector of the (m - 1)-th layer of the u-th neighboring node, is the set of neighboring nodes of the v-th rumor node, ReLU() represents activation using an activation function, and MLP() represents using a multi-layer perceptron function; S42: Update node embeddings in multiple hidden layers of the rumor snapshot through multiple edge weight GIN encoders to obtain the snapshot hidden layer output; S43: Repeat the steps between calculating the time decay parameter of the rumor node until obtaining the snapshot hidden layer output, to obtain multiple snapshot hidden layer outputs, and perform mean pooling on all the snapshot hidden layer outputs to obtain the snapshot dynamic propagation graph.

[0010] According to the rumor detection method based on a dynamic graph provided by the present invention, step S5 further includes: S51: Obtain a forward long short-term memory network and a backward long short-term memory network, and establish the bidirectional long short-term memory network through the forward long short-term memory network and the backward long short-term memory network; S52: Input the snapshot dynamic propagation graph into the forward long short-term memory network to obtain the snapshot forward encoded graph representation, and input the snapshot dynamic propagation graph into the backward long short-term memory network to obtain the snapshot backward encoded graph representation; S53: Concatenate the snapshot forward encoded graph representation and the snapshot backward encoded graph representation to obtain the encoded graph representation of the rumor snapshot.

[0011] According to the rumor detection method based on a dynamic graph provided by the present invention, step S6 further includes: S61: Obtain a deep neural network, and input the encoded graph representation into the fully connected layer and the activation function layer of the deep neural network to obtain the neural network output; S62: Adjust the weights and biases of the neural network output to obtain the rumor snapshot judgment result, and complete rumor detection through the rumor snapshot judgment result.

[0012] The present invention also provides a rumor detection system based on a dynamic graph, including: Rumor propagation graph module: used to obtain rumor data, obtain the rumor text sequence and the rumor source node from the rumor data, and construct a rumor propagation graph through the rumor source node; Snapshot propagation graph module: used to divide the rumor text sequence to obtain a segmented rumor sequence, and establish a snapshot propagation graph of the rumor snapshot through the segmented rumor sequence and the rumor propagation graph; Rumor cycle parameter module: used to calculate the time decay parameter of the rumor node in the snapshot propagation graph, and calculate the rumor cycle parameter of the rumor node through the time decay parameter; Snapshot Dynamic Propagation Graph Module: Used to obtain a GIN encoder, fuse the rumor cycle parameters into the GIN encoder to obtain an edge-weighted GIN encoder, and perform multi-layer node embedding updates on the rumor snapshot through the edge-weighted GIN encoder to obtain a snapshot dynamic propagation graph; Encoded Graph Representation Module: Used to establish a bidirectional long short-term memory network, input the snapshot dynamic propagation graph into the bidirectional long short-term memory network to obtain an encoded graph representation of the rumor snapshot; Rumor Detection Module: Used to obtain a deep neural network, input the encoded graph representation into the deep neural network, and complete rumor detection through the deep neural network.

[0013] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the steps of the rumor detection method based on a dynamic graph as described in any one of the above are implemented.

[0014] The present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the rumor detection method based on a dynamic graph as described in any one of the above are implemented.

[0015] The present invention also provides a computer program product. The computer program product includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions. When the program instructions are executed by a computer, the computer can execute the steps of the rumor detection method based on a dynamic graph as described in any one of the above.

[0016] One or more of the above technical solutions in the embodiments of the present invention have at least one of the following technical effects: The rumor detection method, system, device, product, and medium based on a dynamic graph provided by the present invention use an edge-weighted GIN encoder to perceive time information and use multi-layer node embedding updates to explore the time information within and between snapshots, thereby more effectively performing rumor detection. The present invention effectively constructs an edge-weighted GIN encoder to obtain fine-grained time features, establishes a multi-layer node embedding update mechanism, and enhances the model's learning of temporal dynamic representations through a bidirectional long short-term memory network, greatly improving the effectiveness of the rumor detection task.

[0017] The additional aspects and advantages of the present invention will be partially given in the following description, partially become apparent from the following description, or be understood through the practice of the present invention. Description of the Drawings

[0018] To more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0019] Figure 1 It is a schematic flowchart of the rumor detection method based on a dynamic graph provided by the present invention.

[0020] Figure 2 It is a schematic structural diagram of the rumor detection system based on a dynamic graph provided by the present invention.

[0021] Figure 3 It is a schematic structural diagram of the rumor detection device based on a dynamic graph provided by the present invention.

[0022] Reference numerals: 100, rumor propagation graph module; 200, snapshot propagation graph module; 300, rumor cycle parameter module; 400, snapshot dynamic propagation graph module; 500, encoded graph representation module; 600, rumor detection module; 810, processor; 820, communication interface; 830, memory; 840, communication bus. Detailed implementation manners

[0023] To make the objectives, technical solutions, and advantages of the present invention clearer, the following will clearly and completely describe the technical solutions in the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art without creative efforts based on the embodiments in the present invention belong to the scope protected by the present invention. The following embodiments are used to illustrate the present invention but cannot be used to limit the scope of the present invention.

[0024] In the description of the embodiments of the present invention, it should be noted that the terms "first", "second", and "third" are only used for descriptive purposes and cannot be construed as indicating or implying relative importance.

[0025] In the description of the embodiments of the present invention, it should be noted that unless otherwise clearly defined and limited, the terms "connected" and "coupled" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium. For those of ordinary skill in the art, the specific meanings of the above terms in the embodiments of the present invention can be understood according to specific circumstances.

[0026] In the description of this specification, the descriptions referring to terms such as "one embodiment", "some embodiments", "example", "specific example", or "some examples" mean that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the embodiments of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples. In addition, without contradiction, those skilled in the art can combine and combine the different embodiments or examples described in this specification and the features of different embodiments or examples.

[0027] The following is combined with Figures 1 to 3 Describe the implementation scheme of the present invention: Figure 1 It is a schematic flowchart of a rumor detection method based on a dynamic graph. First, rumor data is obtained and a rumor propagation graph is constructed. Then, a snapshot propagation graph of the rumor snapshot is established. Subsequently, rumor cycle parameters are calculated and a GIN encoder with edge weights is constructed through the rumor cycle parameters. Then, multi-layer node embedding updates are performed on the rumor snapshot, a bidirectional long short-term memory network is established, and an encoded graph representation is obtained. Finally, the encoded graph representation is input into a deep neural network to complete rumor detection.

[0028] The present invention provides a rumor detection method based on a dynamic graph, including: S1: Obtain rumor data, obtain a rumor text sequence and a rumor source node from the rumor data, and construct a rumor propagation graph through the rumor source node; Furthermore, the purpose of this stage is to obtain rumor data from social media, determine the rumor source node, and then establish a rumor propagation graph. Among them, step S1 further includes: S11: Determine the target software, obtain rumor data from the target software, detect the rumor data to obtain potential rumor events, extract and sort the rumor texts related to the potential rumor events from the rumor data to obtain a rumor text sequence; S12: Use the rumor text with the earliest time in the rumor text sequence as the rumor source node, and construct the rumor propagation graph according to the generation time and propagation order of the rumor text.

[0029] For the above steps, the specific implementation methods in this embodiment are as follows: First, determine at least one social media software as the target software. In this embodiment, the selected target software is Twitter. Then, extract the posts and speeches in the target software to obtain rumor data including rumors. Subsequently, detect and monitor the rumor data to find potential rumor events. When a certain event suspected of being a rumor appears multiple times in the rumor data, it is regarded as a potential rumor event, and the rumor text related to the potential rumor event is extracted from the rumor data, that is, the posts or speeches including the potential rumor event, and sorted in chronological order from front to back to obtain a rumor text sequence.

[0030] Next, take the rumor text with the earliest time in the rumor text sequence as the rumor source node, and starting from the rumor source node, construct a rumor propagation graph G according to the generation time and propagation order of the rumor text, and , where V represents the set of rumor nodes. Here, each rumor text is regarded as a rumor node; A represents the adjacency matrix, which is used to represent the mutual relationship between rumor nodes. The mutual relationship here includes forwarding, commenting, etc. between rumor nodes; X is the feature matrix, and the feature matrix includes the text features of the content in the rumor text corresponding to the rumor node.

[0031] S2: Divide the rumor text sequence to obtain a segmented rumor sequence, and establish a snapshot propagation graph of rumor snapshots through the segmented rumor sequence and the rumor propagation graph; Furthermore, the purpose of this stage is to divide the rumor text sequence according to the time stamp to obtain a segmented rumor sequence including rumor text blocks, and finally obtain a snapshot propagation graph including rumor snapshots. Among them, step S2 further includes: S21: Extract the time stamps of the rumor text sequence, and divide the rumor text sequence according to the time stamps to obtain a segmented rumor sequence including rumor text blocks; S22: Divide and texturize the rumor propagation graph through the rumor text blocks in the segmented rumor sequence to obtain the rumor snapshots, and obtain the snapshot propagation graph according to the rumor snapshots.

[0032] For the above steps, the specific implementation methods in this embodiment are as follows: First, extract the time stamps of the rumor text sequence. The time stamp is the time corresponding to each rumor text in the rumor text sequence, and divide the rumor text sequence at the same time interval according to the time stamps. In this embodiment, it is necessary to divide the rumor text sequence into S blocks with the same time interval, then the time interval is calculated as: , where is the time stamp of the last rumor text in the rumor text sequence, is the timestamp of the first rumor text in the rumor text sequence. After partitioning, a chunked rumor sequence including rumor text chunks can be obtained, and each rumor text chunk contains rumor texts within the same time interval.

[0033] Then, partition and textify the rumor propagation graph according to the rumor text chunks, that is, convert the content corresponding to the rumor text chunks in the rumor propagation graph into natural language to obtain a rumor snapshot. The rumor snapshot includes the content, propagation path, etc. of the rumor text corresponding to each rumor text chunk. Then, obtain the snapshot propagation graph of the s-th rumor snapshot : Among them, is the set of rumor nodes of the s-th rumor snapshot, is the adjacency matrix of the s-th rumor snapshot, is the feature matrix of the s-th rumor snapshot.

[0034] S3: Calculate the time decay parameter of the rumor nodes in the snapshot propagation graph, and calculate the rumor cycle parameter of the rumor nodes through the time decay parameter; Furthermore, the purpose of this stage is to introduce the time decay parameter and calculate the rumor cycle parameter of the rumor snapshot through the time decay parameter to present the characteristics of rumor propagation in terms of time and cycle. Specifically, since in the process of rumor propagation, the longer the time interval between its posts and response posts, the lower the rumor popularity can be considered. Therefore, for the rumor nodes in the snapshot propagation graph, the time decay parameter : Among them, α is the time decay factor set according to experience, t is the timestamp of the v-th rumor node in the snapshot propagation graph, is the timestamp of the parent node of the v-th rumor node in the snapshot propagation graph. Here, the meaning of the parent node is the node of the rumor source of this rumor node. For example, when this rumor node forwards, quotes, or mentions the content in a previous rumor node, the previous rumor node can be considered as the parent node of this rumor node.

[0035] Then, based on the time decay parameter, use the cosine function to obtain the rumor cycle parameter of the rumor nodes in the rumor snapshot : Among them, is the learnable first rumor cycle coefficient, is the learnable second rumor cycle coefficient.

[0036] S4: Obtain a GIN encoder, fuse the rumor cycle parameter into the GIN encoder to obtain an edge-weighted GIN encoder, and perform multi-layer node embedding updates on the rumor snapshot through the edge-weighted GIN encoder to obtain a snapshot dynamic propagation graph; Further, the purpose of this stage is to combine the rumor cycle coefficient with a GIN encoder (Graph Isomorphism Network) to obtain an edge-weighted GIN encoder, and use the GIN encoder to perform multi-layer node embedding updates on the rumor snapshot to obtain a snapshot propagation graph. Among them, step S4 further includes: S41: Obtain a GIN encoder, fuse the rumor cycle parameter into the GIN encoder to obtain an edge-weighted GIN encoder. The expression of the edge-weighted GIN encoder is: Among them, is the hidden feature vector of the m-th layer of the v-th rumor node, represents the encoder parameter of the learnable hidden feature vector of the m-th layer, is the hidden feature vector of the (m - 1)-th layer of the v-th rumor node, is the rumor cycle parameter of the v-th rumor node, is the hidden feature vector of the (m - 1)-th layer of the u-th neighboring node, is the set of neighboring nodes of the v-th rumor node, ReLU() represents activation using an activation function, and MLP() represents using a multi-layer perceptron function; S42: Perform node embedding updates in multiple hidden layers of the rumor snapshot through multiple edge-weighted GIN encoders to obtain a snapshot hidden layer output; S43: Repeat the steps between calculating the time decay parameter of the rumor node until obtaining the snapshot hidden layer output to obtain multiple snapshot hidden layer outputs, and perform mean pooling on all the snapshot hidden layer outputs to obtain the snapshot dynamic propagation graph.

[0037] For the above steps, the specific implementation in this embodiment is as follows: First, obtain a GIN encoder. The GIN encoder can effectively construct a graph representation, but it does not contain time information and cannot reflect the influence of time in the rumor propagation process. Therefore, by fusing the rumor cycle parameter into the GIN encoder, an edge-weighted GIN encoder can be obtained: Among them, is the hidden feature vector of the m-th layer of the v-th rumor node, represents the encoder parameter of the learnable hidden feature vector of the m-th layer, is the (m - 1)-th layer hidden feature vector of the v-th rumor node, is the rumor period parameter of the v-th rumor node, is the (m - 1)-th layer hidden feature vector of the u-th neighboring node, is the set of neighboring nodes of the v-th rumor node. ReLU() represents activation using an activation function, and MLP() represents using a multi-layer perceptron function. Here, neighboring nodes are rumor nodes adjacent to the v-th rumor node in the rumor snapshot, and both neighboring nodes and rumor nodes are located in the rumor snapshot. In the edge-weighted GIN encoder, there are multiple layers, and each layer can output the hidden feature vector of that layer. During the calculation of the first-layer hidden feature vector, and are both the text features of the rumor node and neighboring nodes directly obtained from the feature matrix of the snapshot propagation graph of the rumor snapshot.

[0038] Then, embedding updates are performed through multiple edge-weighted GIN encoders in multiple hidden layers of the rumor snapshot. In this embodiment, the number of hidden layers is 2 layers: Among them, EAGIN() represents using the edge-weighted GIN encoder for all rumor nodes in the rumor snapshot, represents the output of the initial first-layer snapshot hidden layer of the s-th rumor snapshot, represents the initial input for embedding update of the s-th rumor snapshot. Here, the initial input is , represents the output of the first-layer snapshot hidden layer of the s-th rumor snapshot, β represents the learnable first hidden layer coefficient, () represents the first multi-layer perceptron function, () represents the second multi-layer perceptron function, represents the output of the transformed first-layer snapshot hidden layer of the (s - 1)-th rumor snapshot, γ represents the learnable second hidden layer coefficient, represents the output of the initial second-layer snapshot hidden layer of the s-th rumor snapshot, represents the output of the second-layer snapshot hidden layer of the s-th rumor snapshot, represents the output of the transformed second-layer snapshot hidden layer of the (s - 1)-th rumor snapshot. The parameters of the first multi-layer perceptron function and the second multi-layer perceptron function are different. The solutions for the output of the transformed first-layer snapshot hidden layer and the output of the transformed second-layer snapshot hidden layer are as follows: Among them, ET Layer() indicates processing using a vector transformation layer, represents the output of the initial first-layer snapshot hidden layer of the (s - 1)-th rumor snapshot, represents the output of the initial second-layer snapshot hidden layer of the (s - 1)-th rumor snapshot. When the rumor snapshot is the first rumor snapshot, the values of the transformed first-layer snapshot hidden layer output and the transformed second-layer snapshot hidden layer output are 0. In this embodiment, the output of the second-layer snapshot hidden layer is used as the snapshot hidden layer output.

[0039] Finally, for all rumor snapshots, repeat the steps from calculating the time decay parameter of the rumor node to obtaining the snapshot hidden layer output, to obtain multiple snapshot hidden layer outputs. Perform mean pooling on the snapshot hidden layer output of the s-th rumor snapshot to obtain the sub-snapshot propagation graph of the s-th rumor snapshot : Among them, MEAN() indicates performing mean pooling on the content within the parentheses. Perform mean pooling on all snapshot hidden layer outputs to obtain the snapshot dynamic propagation graph , and .

[0040] S5: Establish a bidirectional long short-term memory network, input the snapshot dynamic propagation graph into the bidirectional long short-term memory network to obtain the encoded graph representation of the rumor snapshot; Furthermore, the purpose of this stage is to process the dynamic propagation graph through a bidirectional long short-term memory network to enhance the learning of time features by the dynamic propagation graph. Among them, step S5 further includes: S51: Obtain a forward long short-term memory network and a backward long short-term memory network, and establish the bidirectional long short-term memory network through the forward long short-term memory network and the backward long short-term memory network; S52: Input the snapshot dynamic propagation graph into the forward long short-term memory network to obtain a snapshot forward encoded graph representation, and input the snapshot dynamic propagation graph into the backward long short-term memory network to obtain a snapshot backward encoded graph representation; S53: Concatenate the snapshot forward encoded graph representation and the snapshot backward encoded graph representation to obtain the encoded graph representation of the rumor snapshot.

[0041] For the above steps, the specific implementation in this embodiment is as follows: First, obtain a forward long short-term memory network and a backward long short-term memory network, and then concatenate the forward long short-term memory network and the backward long short-term memory network to obtain a bidirectional long short-term memory network.

[0042] Next, input the snapshot dynamic propagation graph into the forward long short-term memory network to capture the forward hidden state of the snapshot dynamic propagation graph over time, and obtain the snapshot forward encoded graph representation : Among them, () indicates that the content within the parentheses is processed by the forward long short-term memory network

[0043] At the same time, input the snapshot dynamic propagation graph into the backward long short-term memory network to capture the backward hidden state of the snapshot dynamic propagation graph over time, and obtain the snapshot backward encoded graph representation : Among them, indicates that the content within the parentheses is processed by the backward long short-term memory network

[0044] Finally, concatenate the snapshot forward encoded graph representation and the snapshot backward encoded graph representation to obtain the encoded graph representation of the rumor snapshot .

[0045] Among them, CONCAT() indicates concatenating the content within the parentheses

[0046] S6: Obtain a deep neural network, input the encoded graph representation into the deep neural network, and complete rumor detection through the deep neural network

[0047] Furthermore, the purpose of this stage is to judge whether a potential rumor event is a rumor through a deep neural network, obtain a rumor snapshot judgment result, and thus complete rumor detection. Among them, step S6 further includes: S61: Obtain a deep neural network, input the encoded graph representation into the fully connected layer and activation function layer of the deep neural network to obtain a neural network output S62: Adjust the weights and biases of the neural network output to obtain a rumor snapshot judgment result, and complete rumor detection through the rumor snapshot judgment result

[0048] For the above steps, the specific implementation in this embodiment is as follows: First, obtain a deep neural network, and then input the encoded graph representation into the fully connected layer and activation function layer of the deep neural network. The deep neural network can judge the probability that a potential rumor event is a rumor through the encoded graph representation and obtain a neural network output .

[0049] Then, the weights and biases of the neural network output are adjusted, and the softmax function is used for activation to obtain the judgment result of the rumor snapshot. : Among them, is the weight adjustment coefficient determined according to experience, is the bias coefficient determined according to experience, and softmax() means using the softmax function to activate the content in the parentheses. When the judgment result of the rumor snapshot is greater than the rumor threshold set as needed, it can be determined that the potential rumor event is a rumor, thus completing the rumor detection.

[0050] The present invention also verifies the effectiveness of the rumor detection method based on the dynamic graph. Here, first, several rumor detection methods such as BiGCN, Bi-GCN, GACL, RDEA, TrustRD, and DynGCN are selected for comparison, and the effectiveness of each method is measured by indicators such as accuracy, precision, recall rate, and F1 value. FGDGNN is the method proposed by the present invention, and Table 1 is the comparison result table of the rumor judgment experiments of the present invention and other methods.

[0051] Table 1 Comparison table of rumor judgment results between the present invention and other methods Among them, T indicates that the potential rumor event is a non-rumor, F indicates that the potential rumor event is a rumor, RumorEval is a rumor dataset including rumors, TWITTER is to use twitter as the target software to obtain rumor data and conduct detection. It can be seen from Table 1 that the method provided by the present invention has a good effect in detecting rumors.

[0052] Then, an ablation experiment is conducted on the method proposed by the present invention, and the results of the ablation experiment are shown in Table 2: Table 2 Schematic table of the results of the ablation experiment of the method of the present invention

[0053] It can be seen from the results of the ablation experiment that each part of the method proposed by the present invention has made a significant contribution to improving the effect of rumor detection.

[0054] Subsequently, the edge weight GIN encoder used in the method proposed by the present invention is replaced with other encoders for experiments, and the experimental results are shown in Table 3: Table 3 Schematic table of the results of the replacement experiment of the encoder in the present invention

[0055] It can be seen from the experimental results that the edge-weighted GIN encoder proposed in this experiment can achieve the best results in the process of rumor detection.

[0056] The rumor detection device based on a dynamic graph provided by the present invention will be described below. The rumor detection device based on a dynamic graph described below can be mutually corresponding and referred to the rumor detection method based on a dynamic graph described above.

[0057] Figure 2 The structural schematic diagram of the rumor detection system based on a dynamic graph is exemplified, as Figure 2 shown, for executing the rumor detection method based on a dynamic graph as described above, including: Rumor propagation graph module 100: for obtaining rumor data, obtaining a rumor text sequence and a rumor source node from the rumor data, and constructing a rumor propagation graph through the rumor source node; Snapshot propagation graph module 200: for partitioning the rumor text sequence to obtain a segmented rumor sequence, and establishing a snapshot propagation graph of rumor snapshots through the segmented rumor sequence and the rumor propagation graph; Rumor cycle parameter module 300: for calculating the time decay parameter of rumor nodes in the snapshot propagation graph, and calculating the rumor cycle parameter of the rumor nodes through the time decay parameter; Snapshot dynamic propagation graph module 400: for obtaining a GIN encoder, fusing the rumor cycle parameter into the GIN encoder to obtain an edge-weighted GIN encoder, and performing multi-layer node embedding update on the rumor snapshots through the edge-weighted GIN encoder to obtain a snapshot dynamic propagation graph; Encoded graph representation module 500: for establishing a bidirectional long short-term memory network, and inputting the snapshot dynamic propagation graph into the bidirectional long short-term memory network to obtain an encoded graph representation of the rumor snapshots; Rumor detection module 600: for obtaining a deep neural network, inputting the encoded graph representation into the deep neural network, and completing rumor detection through the deep neural network.

[0058] On the other hand, Figure 3 The structural schematic diagram of an electronic device is exemplified, as Figure 3 shown, the electronic device may include: a processor 810, a communication interface 820, a memory 830, and a communication bus 840. Among them, the processor 810, the communication interface 820, and the memory 830 complete mutual communication through the communication bus 840. The processor 810 can call the logical instructions in the memory 830 to execute the rumor detection method based on a dynamic graph, and the method includes: S1: Obtain rumor data, acquire a rumor text sequence and a rumor source node from the rumor data, and construct a rumor propagation graph through the rumor source node; S2: Divide the rumor text sequence to obtain a chunked rumor sequence, and establish a snapshot propagation graph of the rumor snapshot through the chunked rumor sequence and the rumor propagation graph; S3: Calculate the time decay parameter of the rumor node in the snapshot propagation graph, and calculate the rumor period parameter of the rumor node through the time decay parameter; S4: Obtain a GIN encoder, fuse the rumor period parameter into the GIN encoder to obtain an edge-weighted GIN encoder, and perform multi-layer node embedding update on the rumor snapshot through the edge-weighted GIN encoder to obtain a snapshot dynamic propagation graph; S5: Establish a bidirectional long short-term memory network, input the snapshot dynamic propagation graph into the bidirectional long short-term memory network to obtain a coded graph representation of the rumor snapshot; S6: Obtain a deep neural network, input the coded graph representation into the deep neural network, and complete rumor detection through the deep neural network.

[0059] In addition, when the logical instructions in the above-mentioned memory 830 are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical discs that can store program codes.

[0060] On the other hand, the present invention also provides a computer program product. The computer program product includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions. When the program instructions are executed by a computer, the computer can execute the dynamic graph-based rumor detection method provided by the above-mentioned various methods. The method includes: S1: Obtain rumor data, acquire a rumor text sequence and a rumor source node from the rumor data, and construct a rumor propagation graph through the rumor source node; S2: Partition the rumor text sequence to obtain a segmented rumor sequence, and establish a snapshot propagation graph of the rumor snapshot through the segmented rumor sequence and the rumor propagation graph; S3: Calculate the time decay parameter of the rumor nodes in the snapshot propagation graph, and calculate the rumor cycle parameter of the rumor nodes through the time decay parameter; S4: Obtain a GIN encoder, fuse the rumor cycle parameter into the GIN encoder to obtain an edge-weighted GIN encoder, and perform multi-layer node embedding updates on the rumor snapshot through the edge-weighted GIN encoder to obtain a snapshot dynamic propagation graph; S5: Establish a bidirectional long short-term memory network, input the snapshot dynamic propagation graph into the bidirectional long short-term memory network to obtain a coded graph representation of the rumor snapshot; S6: Obtain a deep neural network, input the coded graph representation into the deep neural network, and complete rumor detection through the deep neural network.

[0061] On the other hand, the present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it is used to execute the rumor detection method based on a dynamic graph provided by the above-mentioned various methods. The method includes: S1: Obtain rumor data, obtain a rumor text sequence and a rumor source node from the rumor data, and construct a rumor propagation graph through the rumor source node; S2: Partition the rumor text sequence to obtain a segmented rumor sequence, and establish a snapshot propagation graph of the rumor snapshot through the segmented rumor sequence and the rumor propagation graph; S3: Calculate the time decay parameter of the rumor nodes in the snapshot propagation graph, and calculate the rumor cycle parameter of the rumor nodes through the time decay parameter; S4: Obtain a GIN encoder, fuse the rumor cycle parameter into the GIN encoder to obtain an edge-weighted GIN encoder, and perform multi-layer node embedding updates on the rumor snapshot through the edge-weighted GIN encoder to obtain a snapshot dynamic propagation graph; S5: Establish a bidirectional long short-term memory network, input the snapshot dynamic propagation graph into the bidirectional long short-term memory network to obtain a coded graph representation of the rumor snapshot; S6: Obtain a deep neural network, input the coded graph representation into the deep neural network, and complete rumor detection through the deep neural network.

[0062] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. A person of ordinary skill in the art can understand and implement it without creative work.

[0063] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the essence of the above technical solution, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0064] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments or perform equivalent replacements for some of the technical features. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A rumor detection method based on dynamic graph, characterized in that: include: S1: Obtain rumor data, obtain rumor text sequences and rumor source nodes from the rumor data, and construct a rumor propagation graph through the rumor source nodes; S2: Divide the rumor text sequence to obtain a block rumor sequence, and establish a snapshot propagation graph of the rumor snapshot through the block rumor sequence and the rumor propagation graph; S3: Calculate the time decay parameter of the rumor node in the snapshot propagation graph, and calculate the rumor period parameter of the rumor node according to the time decay parameter; S4: Obtain a GIN encoder, fuse the rumor cycle parameter into the GIN encoder, obtain an edge weight GIN encoder, perform multi-layer node embedding update on the rumor snapshot through the edge weight GIN encoder, and obtain a snapshot dynamic propagation graph; S5: establishing a bidirectional long short-term memory network, inputting the snapshot dynamic propagation graph into the bidirectional long short-term memory network, and obtaining a coding graph representation of the rumor snapshot; S6: Obtain a deep neural network, input the encoded graph representation into the deep neural network, and complete rumor detection through the deep neural network.

2. The rumor detection method based on dynamic graph according to claim 1, characterized in that: Step S1 further comprises: S11: determining a target software, obtaining rumor data from the target software, detecting the rumor data to obtain potential rumor events, extracting rumor texts related to the potential rumor events from the rumor data and sorting them to obtain a rumor text sequence; S12: The rumor text with the earliest time in the rumor text sequence is used as the rumor source node, and the rumor propagation graph is constructed according to the generation time and propagation order of the rumor text.

3. The rumor detection method based on dynamic graph according to claim 1, characterized in that: Step S2 further comprises: S21: extracting the timestamp of the rumor text sequence, and dividing the rumor text sequence according to the timestamp to obtain a block rumor sequence including rumor text blocks; S22: Divide and textualize the rumor propagation graph through the rumor text blocks in the block rumor sequence to obtain the rumor snapshot, and obtain the snapshot propagation graph based on the rumor snapshot.

4. The rumor detection method based on dynamic graph according to claim 1, characterized in that: Step S4 further comprises: S41: Obtain a GIN encoder, merge the rumor cycle parameter into the GIN encoder, and obtain an edge weight GIN encoder. The expression of the edge weight GIN encoder is: in, is the m-th layer hidden feature vector of the v-th rumor node, represents the learnable encoder parameters of the m-th layer hidden feature vector, is the m-1th layer hidden feature vector of the vth rumor node, is the rumor cycle parameter of the vth rumor node, is the m-1th layer hidden feature vector of the uth neighboring node, is the set of neighboring nodes of the vth rumor node, ReLU() indicates activation using an activation function, and MLP() indicates the use of a multi-layer perceptron function; S42: performing node embedding update in multiple hidden layers of the rumor snapshot through multiple edge weight GIN encoders to obtain snapshot hidden layer output; S43: Repeat the steps of calculating the time decay parameter of the rumor node until the snapshot hidden layer output is obtained, obtain multiple snapshot hidden layer outputs, perform mean pooling on all the snapshot hidden layer outputs, and obtain the snapshot dynamic propagation graph.

5. The rumor detection method based on dynamic graph according to claim 1, characterized in that: Step S5 further comprises: S51: Acquire a forward long short-term memory network and a backward long short-term memory network, and establish the bidirectional long short-term memory network through the forward long short-term memory network and the backward long short-term memory network; S52: inputting the snapshot dynamic propagation graph into the forward long short-term memory network to obtain a snapshot forward coding graph representation, and inputting the snapshot dynamic propagation graph into the backward long short-term memory network to obtain a snapshot backward coding graph representation; S53: Concatenate the forward coding graph representation of the snapshot and the backward coding graph representation of the snapshot to obtain the coding graph representation of the rumor snapshot.

6. The rumor detection method based on dynamic graph according to claim 1, characterized in that: Step S6 further comprises: S61: Acquire a deep neural network, input the encoded graph representation into the fully connected layer and activation function layer of the deep neural network, and obtain a neural network output; S62: weight adjustment and biasing are performed on the output of the neural network to obtain a rumor snapshot judgment result, and rumor detection is completed through the rumor snapshot judgment result.

7. A rumor detection system based on dynamic graphs, used to execute the rumor detection method based on dynamic graphs as claimed in any one of claims 1 to 6, characterized in that: include: Rumor propagation graph module: used to obtain rumor data, obtain rumor text sequences and rumor source nodes from the rumor data, and construct a rumor propagation graph through the rumor source nodes; Snapshot propagation graph module: used to divide the rumor text sequence to obtain a block rumor sequence, and establish a snapshot propagation graph of the rumor snapshot through the block rumor sequence and the rumor propagation graph; Rumor cycle parameter module: used to calculate the time decay parameter of the rumor node in the snapshot propagation graph, and calculate the rumor cycle parameter of the rumor node through the time decay parameter; Snapshot dynamic propagation graph module: used to obtain a GIN encoder, fuse the rumor cycle parameter into the GIN encoder, obtain an edge weight GIN encoder, perform multi-layer node embedding update on the rumor snapshot through the edge weight GIN encoder, and obtain a snapshot dynamic propagation graph; Coding graph representation module: used to establish a bidirectional long short-term memory network, input the snapshot dynamic propagation graph into the bidirectional long short-term memory network, and obtain the coding graph representation of the rumor snapshot; Rumor detection module: used to obtain a deep neural network, input the coding graph representation into the deep neural network, and complete rumor detection through the deep neural network.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the rumor detection method based on dynamic graphs as described in any one of claims 1 to 6 are implemented.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the rumor detection method based on dynamic graphs as described in any one of claims 1 to 6 are implemented.

10. A computer program product, the computer program product comprising a computer program stored on a non-transitory computer-readable storage medium, the computer program comprising program instructions, characterized in that: When the program instructions are executed by a computer, the computer can perform the steps of the rumor detection method based on dynamic graphs as described in any one of claims 1 to 6.

Images