A rumor detection method based on videoization of topic space

Abstract

The present invention relates to a rumor detection method based on videoization of topic space, including: obtaining original topic data with tag information through an API interface provided by a social platform; creating a rumor topic detection model, which includes: a data preprocessing module, a data feature extraction module, a 3DCNN convolutional neural network, a compare module, a Topic2RGB module, and a Topic2Video module; training the rumor topic detection model with the original topic data as training samples; obtaining target topic data and inputting the target topic data into the trained rumor topic detection model to output the detection result of the target topic. The present invention analyzes the topic after videoizing the trend development characteristics of the topic, which can enable the social platform to more intuitively identify rumor topics, gain more trust and thus obtain more benefits, and at the same time can provide guarantee for social stability.

Description

Technical Field

[0001] The present invention belongs to the field of network public opinion monitoring, and specifically relates to a rumor detection method based on topic space videoization. Background Art

[0002] Rumors can affect the economic stability of a region at a very low cost, and can also gain political benefits by manipulating media opinion. Therefore, the spread of online rumors is faster than that of traditional rumors, and the impact is greater. In order to curb or reduce the negative impact of rumors, we need to effectively detect rumor topics among hundreds of millions of online topics.

[0003] Existing rumor detection methods are content-based rumor detection methods, which detect rumors by observing the text features and the changing trends of user features on the time axis during the rumor propagation process. Other methods are network structure-based rumor detection methods, which detect the propagation source by establishing a rumor source node estimator through the social network topology graph. They do not consider the impact of combining text sentiment features and topic propagation features to make topic video on rumor detection, and cannot monitor rumors more intuitively and efficiently. Therefore, conducting research on topic video can effectively combine the multi-dimensional features of topics, open up new ideas for rumor detection research, and is of great significance for the control and guidance of public opinion analysis. Summary of the invention

[0004] In order to solve the problems existing in the background technology, the present invention starts from the confrontational emotional tendency of topic comments, arranges the pixels in the entire time period into a video through representation learning and time slicing methods, detects rumor videos using the designed compare module, and more intuitively identifies rumor topics, providing a rumor detection method based on topic space videoization, including:

[0005] S1: Obtaining original topic data with label information through the API interface provided by the social platform; wherein the label information includes: rumor topics and non-rumor topics; the original topic data includes: the text of the original topic, the user's comments, the time when the original topic was published, the user's account level, the number of the user's fans, the friend relationship between users, the comment relationship between users and the time when the user commented;

[0006] S2: Creating a rumor topic detection model, the rumor topic detection model includes: a data preprocessing module, a data processing module, a 3DCNN convolutional neural network, a compare module, a Topic2RGB module and a Topic2Video module;

[0007] S3: Use the original topic data as training samples to train the rumor topic detection model;

[0008] S4: Obtain the target topic data, input the target topic data into the trained rumor topic detection model, and output the detection result of the target topic.

[0009] The present invention has at least the following beneficial effects

[0010] According to the present invention, the user data under the rumor topic is used to videoize the rumor topic, and after the trend development characteristics of the topic are videoized, analysis is carried out, which can enable the social platform to more intuitively identify the rumor topic, and classify and detect the rumor video through the CNN convolutional neural network and the compare algorithm, and can accurately classify the rumor topic and timely prevent the spread of the rumor, win more trust for the social platform and thus obtain more benefits, and at the same time can also provide guarantee for social stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 The flowchart of the method of the present invention;

[0012] Figure 2 The schematic diagram of the Topic2RGB module of the present invention for calculating user emotions;

[0013] Figure 3 The schematic diagram of the Topic2Video module of the present invention for videoizing the topic space;

[0014] Figure 4 The schematic diagram of the two-dimensional space frame imaging of the present invention. SPECIFIC IMPLEMENTATION METHOD

[0015] The following uses specific specific examples to illustrate the implementation manners of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific implementation manners, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the diagrams provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner. Without conflict, the following embodiments and the features in the embodiments can be combined with each other.

[0016] Please refer to Figure 1 , the present invention provides a rumor detection method based on topic space videoization, including:

[0017] S1: Obtain the original topic data with tag information through the API interface provided by the social platform; among them, the tag information includes: rumor topics and non-rumor topics; the original topic data includes: the text of the original topic, user comments, the time when the original topic was released, the user's account level, the number of the user's fans, the friendship relationship between users, the comment relationship between users, and the time when the user made a comment;

[0018] The way to obtain the original topic data can be obtained from the API provided by the official website of the public social platform.

[0019] Perform data cleaning on the obtained original topic data, including: unifying the time format of the original topic data, and deleting null values or outliers in the original topic data.

[0020] Store the cleaned original topic data in the local database, and uniformly standardize the storage and naming of the data through the table structure. Moreover, storing the original topic data in the database can also improve the retrieval and reuse efficiency of the data and facilitate the mapping of inter-table relationships. The rumor topic is a topic that does not conform to the facts. The topic data includes topics such as military, music, and entertainment, which can be directly obtained on the social platform. Under each topic, users can express their own comments and views. Similarly, users can also comment on the comments of other users.

[0021] S2: Create a rumor topic detection model, which includes: a data preprocessing module, a data feature extraction module, a 3DCNN convolutional neural network, a compare module, a Topic2RGB module, and a Topic2Video module;

[0022] S3: Use the original topic data as training samples to train the rumor topic detection model;

[0023] S31: Input the original topic data into the data processing module to calculate the support rate of user comments, the credibility of users, the friendship relationship graph, the comment relationship graph, the user relationship graph, the probability of users being affected, the popularity of the original topic, and the relevance between the user's comment and the original topic;

[0024] S311: Calculate the support rate of user comments according to the number of positive and negative words in the user's comments;

[0025]

[0026] Among them, Tri(c i ) represents the support rate of user comments, PosWord(c i ) represents the number of positive words in all comments of the i-th user, negWord(c i) represents the number of negative words in the comments of the \(i\)-th user.

[0027] S312: Calculate the credibility of the user according to the user's account level and the number of the user's fans;

[0028] Cre(u i ) = baseCre + authCra(u i ) + α × fansNum(u i )

[0029] Among them, Cre(u i ) represents the credibility of the user, baseCre represents the basic credibility parameter, authCra(u i ) represents the account level of the \(i\)-th user, α represents the attenuation factor parameter, a random number between α ∈ (0, 1); fansNum(u i ) represents the number of fans of the \(i\)-th user, where baseCre is randomly set to a number greater than 0 by those skilled in the art.

[0030] S313: Generate a friend relationship graph with the user as a node according to the friend relationship between users; generate a comment relationship graph with the user as a node according to the comment relationship between users; and merge the friend relationship graph and the comment relationship graph to generate a user relationship graph;

[0031]

[0032] G n (u i ) = G f (u i ) ∩ G c (u i )

[0033] Among them, G f (u i ) represents the \(i\)-th user node in the friend relationship graph, G c (u i ) represents the \(i\)-th user node in the comment relationship graph, Gra(u i ) represents the \(i\)-th user node in the user relationship graph.

[0034] S314: Calculate the probability that the user is affected according to the friend relationship between users and the comments of the user;

[0035]

[0036] Among them, Fri(u i ) represents the probability that the user is affected, TriNum(u i ) represents the user u iThe number of users who have made positive comments among the friend users of, FriNum(u i ) represents user u i The number of users who have made comments among the friend users of, where the positive comment is a comment in which the number of positive words in the user's comment is more than the number of negative words.

[0037] S315: Calculate the popularity of the original topic according to the time when the original topic is published and the time when the user makes a comment;

[0038]

[0039] Among them, Pop(t) represents the popularity of the original topic at time t, n represents the release time of the original topic, retNum(t) represents the number of user comments on the original topic at time t, and w represents the adjustment factor parameter.

[0040] S316: Represent the user's comment and the text of the original topic by doc2vec embedding vectors to obtain the user comment vector and the original topic text vector, and calculate the relevance between the user's comment and the original topic according to the user comment vector and the original topic text vector using the cosine similarity formula:

[0041]

[0042] Among them, Rel(c i ) represents the relevance between the user's comment and the original topic, topVec represents the original topic text vector, and comVec(c i ) represents the user comment vector.

[0043] S32: The data preprocessing module eliminates the irrelevant users in the user relationship graph according to the relevance between the user's comment and the original topic to obtain the final user relationship graph;

[0044] Among the user's comments, there are some comments that are irrelevant to the original topic. These irrelevant comments will affect the accuracy of the rumor topic detection model. Therefore, it is necessary to eliminate the irrelevant user comments from the user's topic space network. At the same time, the lower-level comments of the irrelevant comments are also useless. When all the user's comments are irrelevant comments, it is necessary to eliminate the irrelevant users.

[0045] Eliminating the irrelevant users in the user relationship graph according to the relevance between the user's comment and the original topic to obtain the final user relationship graph includes:

[0046] Traverse all the comments of each user, and eliminate the user comments with a relevance lower than the set threshold between all the comments of the user and the original topic and the comments of the remaining users under this comment; if all the comments of the user are eliminated, the user is eliminated as an irrelevant user from the user relationship graph.

[0047] Please refer to Figure 2 , S33: The Topic2RGB module pixelates the user based on the support rate of the user's comment, the user's credibility, and the probability of the user being affected to obtain the pixel value of the user in RGB;

[0048] S331: Calculate the support rate of the user for the topic according to the final user relationship graph and the support rate of the user's comment;

[0049]

[0050] Among them, γ represents the inversion parameter, Tri(c i ) represents the support rate of the comment of user u i comment, Tri(c j ) represents the support rate of the comment of user u i 's comment on the comment of user u j , Sup(u i,t ) represents the support rate of the user for the topic at time t, where, in the present invention

[0051] S332: Calculate the true emotional tendency of the user using game theory based on the support rate of the user for the topic and the probability of the user being affected, and convert the credibility of the user and the true emotional tendency of the user into pixel values of RGB to obtain the pixel value of the user in RGB.

[0052] Preferably, the true emotional tendency of the user includes:

[0053] Pro sup (u i,t ) = Sup(u i,t ) + η × Fri(u i )

[0054] Pro opp (u i,t ) = Opp(u i,t ) + η × (1 - Fri(u i ))

[0055] Opp(u i,t ) = 1 - Sup(u i,t )

[0056]

[0057] Among them, Opp(u i,t ) represents the opposition rate of the user for the topic at time t, Fri(u i ) represents the probability of the user being affected, Sup(u i,t ) represents the support rate of the user for the topic at time t, Teri(ui,t ) represents the true emotional tendency of the user, η represents the convergence factor, Pro sup (u i,t ) represents the profit function of the user supporting the topic in game theory, Pro opp (u i,t ) represents the profit function of the user opposing the topic in game theory, Mut sup (u i,t ) represents the emotional influence of the user's attitude of supporting the topic after game quantization, Mut opp (u i,t ) represents the emotional influence of the user's attitude of opposing the topic after game quantization, Tri(u i,t ) = 1 indicates that user u i supports the topic at time t, Tri(u i,t ) = 0 indicates that user u i opposes the topic at time t, and the value of η is randomly set by those skilled in the art.

[0058] Preferably, the pixel values of the user in RGB include:

[0059]

[0060] Rgb(u i,t ) = λ × Cre(u i )

[0061] wherein, Pix(u i,t ) represents the color of the pixel point corresponding to user u i at time t, Rgb(u i,t ) represents the brightness value of the pixel point corresponding to user u i at time t, Cre(u i ) represents the credibility of user u i , λ represents the pixel brightness base value, and λ takes the value of 25 in the present invention.

[0062] Please refer to Figure 3 , S34: The Topic2Video module videoizes the final user relationship graph according to the time when the user publishes comments and the pixel values of the user in RGB to obtain the original topic video;

[0063] S341: Use the node embedding algorithm to map the user nodes in the final user relationship graph to a two-dimensional space to obtain the first user space distribution;

[0064]

[0065] c0 = u

[0066] π μν = αpq (μ, v)·ω μν

[0067]

[0068] Among them, ω represents the regularization factor, E represents the set of edges in the user relationship network, and ω μν represents the weight of the edge (μ, v), and π μν is the transition probability from user node v to user node μ in the non-regularized user relationship network. c0 represents the source topic, and u represents the initiator of the source topic. α pq (t, ν) represents the probability of selecting the next node from user node ν, p represents the probability that user node v selects a node near the previous node next time, q represents the probability that user node v does not select the previous node next time, and d μv represents the distance between user node ν and user node μ.

[0069] The above algorithm belongs to the second-order graph random walk algorithm, aiming to transform the user relationship graph into a sequence of user nodes, and then use the T-SNE algorithm for non-linear dimensionality reduction to transform the sequence of user nodes into a two-dimensional space.

[0070] S342: Cut the time into N identical time periods according to the time when the user posts comments, and form the second user space distribution by combining all the users who post comments in each time period in the first user space distribution to obtain the second user space distributions of N different time periods; and transform the second user space distribution into the original topic frame image through the cutting and diffusion algorithm;

[0071] Assume that the time interval from the publication of the source topic to the last comment is [n, m], and then divide this time interval into N time periods. All the comment users of the source topic belong to the first user space, and the users participating in the topic in each time period belong to the second user space. Obtain the cutting distance of the second user space according to the coordinate positions of the user nodes in the second user space.

[0072]

[0073] Among them, dis(t) represents the cutting distance of the second user space corresponding to the t-th time period, and s max represents the maximum value in the positive X-axis direction of the user space, y max represents the maximum value in the positive Y-axis direction of the user space, nodeNum represents the number of user nodes in the user space, ξ represents the number of pixel grids assigned to each user node, and ξ is 1.3 in the present invention.

[0074] Subsequently, the second user space is divided into pixel grids according to the calculated cutting distance. To address the issue of multiple user nodes in some grids, the redundant user nodes are moved to the surrounding empty node positions using the expansion algorithm, as Figure 4 shown.

[0075] S343: Colorize the original topic frame image according to the pixel values of the user in RGB to obtain a colorized original topic frame image;

[0076] S344: Standardize the color topic frame images through the expansion algorithm and combine the standardized color topic frame images in chronological order to obtain the original topic video;

[0077] The standardization of the color topic frame images includes:

[0078]

[0079] where w max and l max respectively refer to the maximum values of the width and length among all color topic frame images, while w and l respectively represent the width and height of the current color topic frame image, d x and d y respectively represent the moving distances of the current frame image in the positive x-axis and y-axis directions. Through the above algorithm, the topic space is segmented into a set of standard frame images, and then the set of standard frame images is arranged into a topic video in chronological order.

[0080] S35: 3DCNN classifies the original topic video to obtain the initial classification prediction result of the original topic video;

[0081]

[0082] where P represents the initial classification prediction result of the original topic video, softmax represents the classifier. In the present invention, the output of the classifier softmax has only two categories, 0 and 1, indicating that the topic belongs to a rumor topic and a non-rumor topic, represents the topic video, and cnn represents the convolutional layer.

[0083] S36: The compare module calculates the final prediction result of the original topic video using the compare algorithm based on the classification prediction result of the original topic video and the number of different color pixel points in the last frame of the color topic frame image;

[0084]

[0085] where, represents the classification prediction result of 3DCNN, and image sup(m) represents the number of green pixels in the last frame image of the topic video, image opp (m) represents the number of red pixels in the last frame image of the topic video, Pop(t) represents the topic popularity in the time period t, Pop(t) - Pop(t - 1) represents the topic popularity increment in the time period t, σ represents the judgment basis for whether the topic popularity reaches the peak in the time period t, and σ of the present invention is -10, P n represents whether the topic is a rumor.

[0086] When P n = 1, it indicates that the topic is a non-rumor topic. When P n = 0, it indicates that the topic is a rumor topic. By using the compare algorithm designed based on the number of pixels of different colors in the last frame of the color topic frame image and the prediction result of the CNN convolutional model to perform a secondary prediction on the topic, the rumor topic can be predicted more accurately.

[0087] S37: According to the final prediction result of the original topic video and the label information of the original topic video, use the cross-entropy loss function and update the parameters of the rumor topic detection model through the backpropagation mechanism to complete the training of the rumor topic detection model.

[0088] The above preferred embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Those skilled in the art should understand that various changes can be made in form and details without departing from the scope defined by the claims of the present invention.

Claims

1. A rumor detection method based on the videoization of topic space, characterized in that, Including: S1: Obtain the original topic data with tag information through the API interface provided by the social platform; wherein, the tag information includes: rumor topics and non-rumor topics; the original topic data includes: the text of the original topic, user comments, the time when the original topic was released, the user's account level, the number of the user's fans, the friendship relationship between users, the comment relationship between users, and the time when the user posted a comment; S2: Create a rumor topic detection model, and the rumor topic detection model includes: a data preprocessing module, a data feature extraction module, a 3DCNN convolutional neural network, a compare module, a Topic2RGB module, and a Topic2Video module; S3: Use the original topic data as a training sample to train the rumor topic detection model; The step of using the original topic data as a training sample to train the rumor topic detection model includes: S31: Input the original topic data into the data processing module to calculate the support rate of user comments, the credibility of users, the friendship relationship graph, the comment relationship graph, the user relationship graph, the probability of users being affected, the popularity of the original topic, and the relevance between the user's comment and the original topic; S32: The data preprocessing module eliminates irrelevant users in the user relationship graph according to the relevance between the user's comment and the original topic to obtain the final user relationship graph; S33: The Topic2RGB module pixelates users according to the support rate of user comments, the credibility of users, and the probability of users being affected to obtain the pixel values of users in RGB; S34: The Topic2Video module videoizes the final user relationship graph according to the time when the user posted a comment and the pixel values of users in RGB to obtain the original topic video; S35: The 3DCNN classifies the original topic video to obtain the initial classification prediction result of the original topic video; S36: The compare module calculates the final prediction result of the original topic video by using the compare algorithm according to the classification prediction result of the original topic video and the number of different color pixel points in the last frame of the color topic frame image; S37: Update the parameters of the rumor topic detection model by using the cross-entropy loss function according to the final prediction result of the original topic video and the tag information of the original topic video and through the backpropagation mechanism to complete the training of the rumor topic detection model; S4: Obtain the target topic data, input the target topic data into the trained rumor topic detection model, and output the detection result of the target topic.

2. The rumor detection method based on the videoization of topic space according to claim 1, characterized in that, The specific steps of S31 include: S311: Calculate the support rate of user comments according to the number of positive and negative words in the user's comment; S312: Calculate the credibility of users according to the user's account level and the number of the user's fans; S313: Generate a friendship relationship graph with users as nodes according to the friendship relationship between users; generate a comment relationship graph with users as nodes according to the comment relationship between users; and merge the friendship relationship graph and the comment relationship graph to generate a user relationship graph; S314: Calculate the probability of users being affected according to the friendship relationship between users and the user's comment; S315: Calculate the popularity of the original topic based on the time when the original topic was published and the time when the user posted a comment. S316: Represent the user's comment and the text of the original topic as vectors through doc2vec embedding to obtain the user comment vector and the original topic text vector, and calculate the relevance between the user's comment and the original topic according to the user comment vector and the original topic text vector using the cosine similarity formula.

3. The rumor detection method based on the videoization of topic space according to claim 1, characterized in that, The step of removing irrelevant users in the user relationship graph according to the relevance between the user's comment and the original topic to obtain the final user relationship graph includes: Traverse all comments of each user, and remove the user comments with a relevance lower than the set threshold to the original topic and the comments of the remaining users under this comment; if all comments of a user are removed, then the user is removed as an irrelevant user from the user relationship graph.

4. The rumor detection method based on the videoization of topic space according to claim 1, characterized in that, The step of pixelating the user according to the support rate of the user comment, the credibility of the user, and the probability of the user being affected to obtain the pixel values of the user in RGB includes: S331: Calculate the support rate of the user for the topic according to the final user relationship graph and the support rate of the user comment. S332: Calculate the true emotional tendency of the user using game theory according to the support rate of the user for the topic and the probability of the user being affected, and convert the credibility of the user and the true emotional tendency of the user into pixel values in RGB to obtain the pixel values of the user in RGB.

5. The rumor detection method based on the videoization of topic space according to claim 1, characterized in that, The step of videoizing the final user relationship graph according to the time when the user posted a comment and the pixel values of the user in RGB to obtain the original topic video includes: S341: Use the node embedding algorithm to map the user nodes in the final user relationship graph to a two-dimensional space to obtain the first user space distribution. S342: Cut the time into the same N time periods according to the time when the user posted a comment, and form the second user space distribution by all users who posted comments in each time period in the first user space distribution to obtain the second user space distributions of N different time periods; and convert the second user space distribution into the original topic frame image through the cutting and diffusion algorithm. S343: Colorize the original topic frame image according to the pixel values of the user in RGB to obtain the colored original topic frame image. S344: Standardize the colored topic frame image through the expansion algorithm and combine the standardized colored topic frame images in chronological order to obtain the original topic video.

6. The rumor detection method based on the videoization of topic space according to claim 1, characterized in that,The final prediction result of the original topic video includes: Among them, represents the classification prediction result of the 3DCNN, and image sup (m) represents the number of green pixels in the last frame image of the topic video, and image opp (m) represents the number of red pixels in the last frame image of the topic video. Pop(t) represents the topic popularity in the t time period, Pop(t) - Pop(t - 1) represents the topic popularity increment in the t time period, and σ represents the judgment basis for whether the topic popularity reaches the peak in the t time period. In the present invention, σ is -10, and P n represents whether the topic is a rumor.

Images