Rumor detection method based on sliding window

Abstract

The invention relates to the technical field of artificial intelligence, and provides a rumor detection method based on a sliding window, and the method comprises the steps: obtaining a rumor source node, and constructing a rumor propagation graph through the rumor source node; calculating a rumor propagation reference rate through the initial rumor window, performing window dynamic adjustment, and updating the rumor propagation reference rate to obtain a rumor window set; calculating node depth and matching window codes, obtaining an initial memory library, calculating node depth information through the matching window codes and the depth embedding vectors, obtaining rumor event representation increments through the node depth information and a memory attention enhancement mechanism, and obtaining rumor event representation increments; the rumor event representation increment is merged into an initial memory bank, the initial memory bank is iterated, and the rumor event representation is obtained; and obtaining a deep neural network, inputting the rumor event representation into the deep neural network, and completing rumor detection through the deep neural network. The rumor detection effect can be effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a rumor detection method based on a sliding window. Background Art

[0002] Social media has become the core platform for people to obtain information, express opinions, and maintain social connections. Nevertheless, the potential risks behind it cannot be ignored. In particular, the disorderly spread and rapid spread of rumors have posed severe challenges to social stability, public trust, and economic order. Therefore, how to efficiently and accurately identify and contain the spread of rumors on social media has become a key issue that needs to be solved urgently.

[0003] Traditional rumor detection mainly relies on manually designed features and machine learning methods to identify rumors. With the development of deep learning, the emergence of various neural networks has improved the performance of rumor detection. Researchers have explored the static propagation graph of rumor events and achieved excellent detection performance. Further exploration of the temporal dynamics of events has been carried out, and dynamic graphs have been proposed to simulate the spread of events on social media. These methods are usually constructed by graph neural networks, emphasizing the transformation and aggregation of node features, but they cannot capture the detailed temporal features of the spread, such as speed, depth, and breadth. In this way, rumors cannot be identified from the perspective of temporal features, which may lead to the failure to identify rumors in a timely and accurate manner, and the failure to obtain the comprehensive spread characteristics of rumors. Summary of the Invention

[0004] The present invention aims to solve at least one of the technical problems existing in the related art. For this purpose, the present invention provides a rumor detection method based on a sliding window to achieve efficient detection of rumors on social media.

[0005] The present invention provides a rumor detection method based on a sliding window, including: S1: Obtain rumor content data, obtain a rumor content sequence and a rumor source node from the rumor content data, and construct a rumor propagation graph through the rumor content sequence and the rumor source node; S2: Determine an initial rumor window, calculate a rumor propagation benchmark rate through the initial rumor window, perform window dynamic adjustment according to the rumor propagation benchmark rate, and update the rumor propagation benchmark rate to obtain a set of rumor windows; S3: Calculate the node depth through the rumor propagation graph and the set of rumor windows, and obtain a depth embedding vector according to the node depth; S4: Calculate a matching window encoding, obtain an initial memory bank, calculate node depth information through the matching window encoding and the depth embedding vector, obtain a rumor event representation increment through the node depth information and a memory-enhanced attention mechanism, and merge the rumor event representation increment into the initial memory bank and iterate the initial memory bank to obtain a rumor event representation; S5: Obtain a deep neural network, input the rumor event representation into the deep neural network, and complete rumor detection through the deep neural network.

[0006] For a rumor detection method based on a sliding window provided by the present invention, step S1 specifically includes: S11: Determine the target software, obtain rumor content data from the target software, detect the rumor content data to obtain latent rumor events, extract and sort rumor content texts related to the latent rumor events from the rumor content data to obtain a rumor content sequence; S12: Use the rumor content text with the earliest time in the rumor content sequence as the rumor source node, and construct the rumor propagation graph according to the generation time and propagation order of the rumor content text.

[0007] For a rumor detection method based on a sliding window provided by the present invention, step S2 specifically includes: S21: Determine the initial number of rumors, determine the initial rumor window according to the initial number of rumors, obtain the rumor timestamps of the initial rumor window, and calculate the rumor propagation benchmark rate according to the rumor timestamps; S22: Obtain the number of rumor windows and determine the window to be adjusted, divide by the number of rumor windows to obtain the initial number of nodes, dynamically adjust the window to be adjusted according to the initial number of nodes and the rumor propagation benchmark rate to obtain an adjusted window, and update the rumor propagation benchmark rate; S23: Repeat step S22 with the updated rumor propagation benchmark rate, so as to obtain the rumor window set through the adjusted window.

[0008] For a rumor detection method based on a sliding window provided by the present invention, in step S3, obtain a window adjacency matrix through the rumor window set and the rumor propagation graph, calculate the node depth through the window adjacency matrix, and obtain the depth embedding vector according to the node depth.

[0009] For a rumor detection method based on a sliding window provided by the present invention, step S4 specifically includes: S41: Calculate the matching window encoding through a multi-layer perceptron function to obtain an initial memory bank, extract the first depth information and the second depth information from the depth embedding vector, and calculate the node depth information through the first depth information, the second depth information, and the matching window encoding; S42: Obtain a trainable parameter matrix, transform the trainable parameter matrix through the node depth information to obtain a query vector, a key vector, and a value vector; S43: Obtain the attention parameters of the query vector, the key vector, and the value vector, input the attention parameters into the multi-head attention function to obtain the multi-head attention representation, and obtain the rumor event representation increment through the multi-head attention representation and the node depth information; S44: Incorporate the rumor event representation increment into the initial memory bank and update the window adjacency matrix. Iterate the initial memory bank through the updated window adjacency matrix to obtain the rumor event representation.

[0010] According to a rumor detection method based on a sliding window provided by the present invention, in step S44, when iterating the initial memory bank, select a new adjusted window from the rumor window set and repeat steps S3 to S43, incorporate the obtained rumor event representation increment into the initial memory bank until all the adjusted windows are used to calculate the node depth and iterate the initial memory bank.

[0011] According to a rumor detection method based on a sliding window provided by the present invention, step S5 specifically includes: S51: Obtain an initial deep neural network, input the rumor event representation into the fully connected layer and the activation function layer of the initial deep neural network to obtain an initial neural network output; S52: Calculate the cross-entropy of the initial neural network output, train the initial deep neural network through the cross-entropy to obtain a target deep neural network, input the rumor event representation into the target deep neural network to obtain a rumor representation judgment result, and complete rumor detection through the rumor representation judgment result.

[0012] According to a rumor detection method based on a sliding window provided by the present invention, in step S1, the rumor propagation graph includes a rumor node set, an adjacency matrix, and a feature matrix.

[0013] According to a rumor detection method based on a sliding window provided by the present invention, in step S43, obtain the vector dimensions of the query vector, the key vector, and the value vector, and obtain the attention parameters through the vector dimensions and a softmax transformation.

[0014] According to a rumor detection method based on a sliding window provided by the present invention, in step S41, obtain the first node depth information through the first depth information and the initial memory bank, obtain the second node depth information through the second depth information and the matching window encoding, and the node depth information includes the first node depth information and the second node depth information.

[0015] One or more of the above technical solutions in the embodiments of the present invention have at least one of the following technical effects: A rumor detection method based on a sliding window provided by the present invention effectively adjusts the window dynamically for rumor information through the benchmark rate of rumor propagation, thereby reflecting the dynamics of rumor information dissemination. Through a memory-enhanced attention mechanism, it effectively enhances the learning of the structural information features and information features of rumor propagation, and combines them into the initial memory bank for rumor detection, effectively improving the accuracy and effectiveness of the rumor detection task.

[0016] Additional aspects and advantages of the present invention will be given in part in the following description, will become apparent in part from the following description, or will be understood through the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order 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.

[0018] Figure 1 It is a schematic flowchart of a rumor detection method based on a sliding window provided by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0019] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions in the present invention will be clearly and completely described below. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of 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.

[0020] 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.

[0021] In the description of the embodiments of the present invention, it should be noted that unless otherwise clearly specified 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.

[0022] In the description of this specification, the description referring to terms such as "one embodiment", "some embodiments", "example", "specific example", or "some examples" means 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 expressions 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.

[0023] The following is combined with Figure 1 to describe the implementation scheme of the present invention: Figure 1 It is a schematic flowchart of a rumor detection method based on a sliding window provided by the present invention, including first obtaining rumor content data and constructing a rumor propagation graph; then performing window dynamic adjustment to obtain a rumor window set; subsequently calculating the node depth to obtain a depth embedding vector; then obtaining the increment of the rumor event representation, performing initial memory bank iteration to obtain the rumor event representation; and finally completing rumor detection through a deep neural network.

[0024] The present invention provides a rumor detection method based on a sliding window, including: S1: Obtain rumor content data, obtain a rumor content sequence and a rumor source node from the rumor content data, and construct a rumor propagation graph through the rumor content sequence and 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 content data from the target software, detect the rumor content data to obtain latent rumor events, extract and sort rumor content texts related to the latent rumor events from the rumor content data to obtain a rumor content sequence; S12: Use the rumor content text with the earliest time in the rumor content sequence as the rumor source node, and construct the rumor propagation graph according to the generation time and propagation order of the rumor content text.

[0025] In step S1, the rumor propagation graph includes a rumor node set, an adjacency matrix, and a feature matrix.

[0026] For the above steps, the specific implementation methods in this embodiment are as follows: First, determine at least one Internet social media software or Internet forum, etc. as the target software. In this embodiment, Twitter is selected as the target software. Then, obtain the content such as posts, interactive speeches, and topics in the target software to obtain rumor content data including rumors. Subsequently, identify and monitor the rumor content data to find latent rumor events. When a certain event suspected of being a rumor appears repeatedly in the rumor content data, it is determined as a latent rumor event, and the rumor content text related to the latent rumor event is extracted from the rumor content data, that is, the posts, interactive speeches, and topics including the latent rumor event, and sorted in chronological order from the front to the back to obtain the rumor content sequence.

[0027] Then, take the rumor content text with the earliest time in the rumor content sequence as the rumor source node, and use the rumor source node as the starting point to establish a rumor propagation graph G according to the generation time and propagation order of the rumor content text, and , where R is the set of rumor nodes. Here, each rumor content text is regarded as a rumor node; A is the adjacency matrix, which can be used to represent the mutual relationship between rumor nodes. Here, the mutual relationship can be the mutual comments, forwards, etc. between rumor nodes; X is the feature matrix, and the feature matrix includes the text features of the content in the rumor content text corresponding to the rumor node.

[0028] S2: Determine the initial rumor window, calculate the rumor propagation benchmark rate through the initial rumor window, perform window dynamic adjustment according to the rumor propagation benchmark rate, and update the rumor propagation benchmark rate to obtain a set of rumor windows; Furthermore, the purpose of this stage is to calculate the rumor propagation benchmark rate, perform window dynamic adjustment according to the rumor propagation benchmark rate, and update the rumor propagation benchmark rate to obtain a set of rumor windows. Specifically, step S2 specifically includes: S21: Determine the initial number of rumors, determine the initial rumor window according to the initial number of rumors, obtain the rumor timestamps of the initial rumor window, and calculate the rumor propagation benchmark rate according to the rumor timestamps; S22: Obtain the number of rumor windows and determine the window to be adjusted. Divide through the number of rumor windows to obtain the initial number of nodes, perform window dynamic adjustment on the window to be adjusted according to the initial number of nodes and the rumor propagation benchmark rate to obtain the adjusted window, and update the rumor propagation benchmark rate; S23: Repeat step S22 through the updated rumor propagation benchmark rate, so as to obtain the set of rumor windows through the adjusted window.

[0029] For the above steps, the specific implementation methods in this embodiment are as follows: First, determine the number of rumor nodes included in the rumor propagation graph and use it as the initial number of rumors. Then, set the initial number of rumor windows according to the initial number of rumors and experience, and divide the number of rumor nodes by the initial number of rumor windows to obtain the number of rumor nodes included in each rumor window, so that the rumor nodes are evenly distributed in several rumor windows. Since the rumor nodes in the rumor propagation graph are arranged in chronological order, the rumor window where the rumor source node is located is the initial rumor window. Subsequently, obtain the rumor timestamps of the initial rumor window, that is, the sending times of each rumor node in the initial rumor window. Then, calculate the time difference between the j-th rumor node in the initial rumor window and the subsequent rumor node to obtain the timestamp interval of the j-th rumor node , thereby calculating the rumor propagation rate of the i-th rumor window as the rumor propagation benchmark rate: where J is the number of rumor nodes in the rumor window.

[0030] Subsequently, obtain the number of rumor windows. Here, the number of rumor windows is the initial number of rumor windows minus 1. The window to be adjusted can be the initial rumor window here. And equally divide the rumor nodes outside the window to be adjusted by the number of rumor windows to obtain the initial rumor window set including rumor propagation windows and the initial node number b. Here, the initial node number is the number of rumor nodes included in each rumor window in the initial rumor window set. Then, dynamically adjust the window of the window to be adjusted according to the initial node number and the rumor propagation benchmark rate. The adjustment method is as follows: First, calculate the (i + 1)-th rumor propagation rate of the rumor propagation window after the window to be adjusted , and then adjust the number of rumor nodes in the window to be adjusted: where is the number of rumor nodes in the window to be adjusted after window dynamic adjustment, and α is a dynamic adjustment hyperparameter determined according to experience. Through window dynamic adjustment, the size of the window to be adjusted can be adjusted according to the rumor propagation rate of the rumor propagation window after the window to be adjusted, that is, add a part of the rumor nodes in the rumor propagation window after the window to be adjusted to the window to be adjusted, or add a part of the rumor nodes in the window to be adjusted to the rumor propagation window after the window to be adjusted. Take the adjusted window to be adjusted as the adjusted window, and take the rumor propagation rate of the rumor propagation window after the window to be adjusted as the new rumor propagation benchmark rate to complete the update of the rumor propagation benchmark rate.

[0031] Finally, step S22 is repeatedly executed with the updated benchmark rate of rumor propagation, that is, each time it is repeatedly executed, 1 is subtracted from the number of rumor propagation windows included in the initial rumor window set obtained in the previous execution to obtain a new number of rumor windows. And each time it is repeatedly executed, the rumor nodes other than the adjusted window are equally divided according to the number of rumor windows, and the rumor propagation window after the adjusted window is used as the window to be adjusted, so that step S22 can be repeatedly executed until all the rumor propagation windows are adjusted. Among them, the size of the last rumor propagation window does not need to be adjusted, and all the adjusted windows after adjustment are used as the rumor window set.

[0032] S3: Calculate the node depth through the rumor propagation graph and the rumor window set, and obtain the depth embedding vector according to the node depth; Furthermore, the purpose of this stage is to calculate the node depth so as to obtain the depth embedding vector. Specifically, in step S3, a window adjacency matrix is obtained through the rumor window set and the rumor propagation graph, the node depth is calculated through the window adjacency matrix, and the depth embedding vector is obtained according to the node depth.

[0033] For the above steps, the specific implementation in this embodiment is as follows: First, take the first adjusted window in the rumor window set as the reference window Then, select an adjusted window from the rumor window set as the matching window Obtain the mutual relationship between the rumor nodes of the reference window from the rumor propagation graph to get the reference adjacency matrix And obtain the mutual relationship between the rumor nodes of the matching window to get the matching adjacency matrix Take the reference adjacency matrix and the matching adjacency matrix as the window adjacency matrix, and calculate the node depth depth through the window adjacency matrix: Among them, ComputeDepth() represents a function for calculating the depth of each node in the propagation graph. Finally, obtain the depth embedding layer and obtain the depth embedding vector D according to the node depth: Among them, represents a trainable depth embedding layer.

[0034] S4: Calculate the matching window encoding, obtain the initial memory bank, calculate the node depth information through the matching window encoding and the depth embedding vector, obtain the rumor event representation increment through the node depth information and the memory-enhanced attention mechanism, merge the rumor event representation increment into the initial memory bank and iterate the initial memory bank to obtain the rumor event representation; Further, the purpose of this stage is to calculate the node depth information, obtain the rumor event representation increment through the memory-enhanced attention mechanism, and iterate the initial memory bank. Specifically, step S4 specifically includes: S41: Calculate the matching window encoding through a multi-layer perceptron function and obtain the initial memory bank. Extract the first depth information and the second depth information from the depth embedding vector, and calculate the node depth information through the first depth information, the second depth information, and the matching window encoding; S42: Obtain the trainable parameter matrix, and transform the trainable parameter matrix through the node depth information to obtain the query vector, the key vector, and the value vector; S43: Obtain the attention parameters of the query vector, the key vector, and the value vector, input the attention parameters into the multi-head attention function to obtain the multi-head attention representation, and obtain the rumor event representation increment through the multi-head attention representation and the node depth information; S44: Incorporate the rumor event representation increment into the initial memory bank and update the window adjacency matrix, and iterate the initial memory bank through the updated window adjacency matrix to obtain the rumor event representation.

[0035] In step S44, when iterating the initial memory bank, select a new adjusted window from the rumor window set and repeat steps S3 to S43, incorporate the obtained rumor event representation increment into the initial memory bank until all the adjusted windows are used to calculate the node depth and iterate the initial memory bank.

[0036] In step S41, obtain the first node depth information through the first depth information and the initial memory bank, obtain the second node depth information through the second depth information and the matching window encoding, and the node depth information includes the first node depth information and the second node depth information.

[0037] In step S43, obtain the vector dimensions of the query vector, the key vector, and the value vector, and obtain the attention parameters through the vector dimensions and a softmax transformation.

[0038] For the above steps, the specific implementation in this embodiment is as follows: First, encode the feature matrix of the rumor nodes in the reference window through a multi-layer perceptron function to obtain the initial memory bank , and encode the feature matrix of the rumor nodes in the matching window through a multi-layer perceptron function to obtain the matching window encoding S: Among them, MLP() represents the multi-layer perceptron function. Then, the first depth information of the reference window is extracted from the deep embedding vector and the second depth information of the matching window . The node depth information including the first node depth information and the second node depth information is calculated through encoding and computing using the first depth information, the second depth information, and the matching window: Among them, is the first node depth information, is the second node depth information.

[0039] Subsequently, a trainable parameter matrix including the trainable query parameter matrix , the key parameter matrix and the value parameter matrix is obtained. The trainable parameter matrix is transformed through the node depth information to obtain the query vector Q, the key vector K, and the value vector V: Next, the vector dimensions of the query vector, the key vector, and the value vector are obtained, and the attention parameter β(Q, K, V) of the query vector, the key vector, and the value vector is calculated: Among them, is the transpose of K, softmax() is the softmax transformation of the content in the brackets, is the vector dimension of the query vector, the key vector, and the value vector. The attention parameter is input into the multi-head attention function to obtain the multi-head attention representation including multiple multi-head attention parameters: Among them, is the h-th multi-head attention parameter, H is the number of multi-head attention parameters, and A() is the multi-head attention function. The multi-head attention representations are concatenated to obtain the target multi-head attention representation : Among them, is the concatenation operation of the content in the brackets. The target multi-head attention representation is concatenated with the second node depth information to obtain the rumor event representation increment : Among them, Concat() represents concatenating the content within the parentheses. Subsequently, the incremental representation of the rumor event is merged into the initial memory bank, and a new non-repeating adjusted window is selected as the matching window. The merged initial memory bank is used as the initial memory bank for the next iteration, and steps S3 to S43 are repeated. Each repetition yields a new incremental representation of the rumor event, which is merged into the initial memory bank, thereby iterating the initial memory bank until all adjusted windows have been used to calculate the node depth, participating in this process and iterating the initial memory bank. The initial memory bank at this time is used as the target memory bank. , and finally, mean pooling operation is performed on the target memory bank to obtain the rumor event representation Z: Among them, MEAN() represents performing mean pooling operation on the content within the parentheses.

[0040] S5: Obtain a deep neural network, input the rumor event representation into the deep neural network, and complete rumor detection through the deep neural network.

[0041] Furthermore, the purpose of this stage is to complete rumor detection through a deep neural network. Specifically, step S5 includes: S51: Obtain an initial deep neural network, input the rumor event representation into the fully connected layer and activation function layer of the initial deep neural network to obtain an initial neural network output; S52: Calculate the cross-entropy of the initial neural network output, train the initial deep neural network through the cross-entropy to obtain a target deep neural network, input the rumor event representation into the target deep neural network to obtain a rumor representation judgment result, and complete rumor detection through the rumor representation judgment result.

[0042] For the above steps, the specific implementation in this embodiment is as follows: First, obtain an initial deep neural network, then input the rumor event representation into the fully connected layer and activation function layer of the initial neural network to obtain an initial neural network output. Then calculate the cross-entropy of the initial neural network output. The initial deep neural network can be trained through the cross-entropy to improve its performance, thereby obtaining a target deep neural network. Input the rumor event representation into the target deep neural network to obtain a rumor representation judgment result, and thus determine whether the latent rumor event is a rumor, completing rumor detection.

[0043] The present invention also verifies the effectiveness of a rumor detection method based on a sliding window. First, several rumor detection methods such as EBGCN, Bi-GCN, GACL, RDEA, TrustRD, DynGCN, and PSGT are selected for comparison, and the effectiveness of each method is measured by indicators such as accuracy, precision, recall, and F1 value. SWAM is the method proposed by the present invention, and Table 1 is a comparison table of the results of the rumor judgment experiments of the present invention and other methods.

[0044] Table 1 Comparison table of rumor judgment results between the present invention and other methods

[0045] Among them, T indicates that the latent rumor event is non-rumor, F indicates that the latent rumor event is rumor, DRWeibo is to use Weibo as the target software to obtain rumor data and conduct detection, and 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.

[0046] In addition, the present invention also conducts an ablation experiment on a rumor detection method based on a sliding window. Among them, SWAM is the complete method proposed by the present invention, and the effectiveness of each method is measured by indicators such as accuracy, precision, recall, and F1 value. After removing S3, S2, and S4, the model needs to be adjusted adaptively to ensure the normal operation of the model. Table 2 is the result table of the ablation experiment of the present invention.

[0047] Table 2 Result table of the ablation experiment of the present invention.

[0048] It can be seen that removing the above steps will significantly affect the performance of the rumor detection method provided by the present invention.

[0049] 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 such an understanding, the above technical solution, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. The 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.

[0050] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than 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; and 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 various embodiments of the present invention.

Claims

1. A rumor detection method based on sliding window, characterized in that: include: S1: Obtain rumor content data, obtain rumor content sequence and rumor source node from the rumor content data, and construct a rumor propagation graph through the rumor content sequence and rumor source node; S2: determining an initial rumor window, calculating a rumor propagation benchmark rate through the initial rumor window, dynamically adjusting the window according to the rumor propagation benchmark rate and updating the rumor propagation benchmark rate to obtain a rumor window set; S3: Calculate the node depth through the rumor propagation graph and the rumor window set, and obtain the deep embedding vector according to the node depth; S4: Calculate the matching window encoding, obtain the initial memory library, calculate the node depth information through the matching window encoding and the deep embedding vector, obtain the rumor event representation increment through the node depth information and the memory enhancement attention mechanism, merge the rumor event representation increment into the initial memory library and iterate the initial memory library to obtain the rumor event representation; S5: Obtain a deep neural network, input the rumor event representation into the deep neural network, and complete rumor detection through the deep neural network.

2. According to claim 1, a rumor detection method based on sliding window is characterized in that: Step S1 specifically includes: S11: determining a target software, obtaining rumor content data from the target software, detecting the rumor content data to obtain a latent rumor event, extracting rumor content texts related to the latent rumor event from the rumor content data and sorting them to obtain a rumor content sequence; S12: Taking the earliest rumor content text in the rumor content sequence as the rumor source node, and constructing the rumor propagation graph according to the generation time and propagation order of the rumor content text.

3. The rumor detection method based on sliding window according to claim 1, characterized in that: Step S2 specifically includes: S21: Determine the initial rumor number, determine the initial rumor window according to the initial rumor number, obtain the rumor timestamp of the initial rumor window, and calculate the rumor propagation benchmark rate according to the rumor timestamp; S22: Obtain the number of rumor windows and determine the window to be adjusted, divide the rumor windows by the number of rumor windows to obtain the number of initial nodes, dynamically adjust the window to be adjusted according to the number of initial nodes and the rumor propagation benchmark rate to obtain an adjusted window, and update the rumor propagation benchmark rate; S23: Repeat step S22 using the updated rumor propagation benchmark rate, thereby obtaining the rumor window set through the adjusted window.

4. The rumor detection method based on sliding window according to claim 1 is characterized in that: In step S3, a window adjacency matrix is ​​obtained through the rumor window set and the rumor propagation graph, the node depth is calculated through the window adjacency matrix, and the depth embedding vector is obtained according to the node depth.

5. The rumor detection method based on sliding window according to claim 1 is characterized in that: Step S4 specifically includes: S41: Calculate the matching window code through a multi-layer perceptron function and obtain an initial memory library, extract the first depth information and the second depth information from the depth embedding vector, and calculate the node depth information through the first depth information, the second depth information and the matching window code; S42: Obtain a trainable parameter matrix, and transform the trainable parameter matrix according to the node depth information to obtain a query vector, a key vector, and a value vector; S43: Obtain attention parameters of the query vector, the key vector, and the value vector, input the attention parameters into a multi-head attention function to obtain a multi-head attention representation, and obtain an increment of the rumor event representation through the multi-head attention representation and the node depth information; S44: Incrementally merge the rumor event representation into the initial memory bank and update the window adjacency matrix, iterate the initial memory bank through the updated window adjacency matrix to obtain the rumor event representation.

6. The rumor detection method based on sliding window according to claim 5 is characterized in that: In step S44, when iterating the initial memory bank, a new adjusted window is selected from the rumor window set and steps S3 to S43 are repeated to merge the obtained rumor event representation increments into the initial memory bank until all the adjusted windows are used to calculate the node depth and iterate the initial memory bank.

7. The rumor detection method based on sliding window according to claim 1 is characterized in that: Step S5 specifically includes: S51: Obtain an initial deep neural network, input the rumor event representation into the fully connected layer and activation function layer of the initial deep neural network, and obtain an initial neural network output; S52: Calculate the cross entropy of the initial neural network output, train the initial deep neural network through the cross entropy to obtain a target deep neural network, input the rumor event representation into the target deep neural network to obtain a rumor representation judgment result, and complete rumor detection through the rumor representation judgment result.

8. The rumor detection method based on sliding window according to claim 1, characterized in that: In step S1, the rumor propagation graph includes a rumor node set, an adjacency matrix and a feature matrix.

9. The sliding window-based rumor detection method according to claim 5, characterized in that: In step S43, the vector dimensions of the query vector, the key vector and the value vector are obtained, and the attention parameter is obtained by performing a softmax transformation on the vector dimensions.

10. The sliding window-based rumor detection method according to claim 5, characterized in that: In step S41, first node depth information is obtained through the first depth information and the initial memory library, and second node depth information is obtained through the second depth information and the matching window encoding. The node depth information includes the first node depth information and the second node depth information.

Images