A rumor detection method based on topic visualization

Abstract

The present invention belongs to the field of social network analysis, and specifically relates to a rumor detection method based on topic visualization, including: performing data cleaning processing on the acquired data; visualizing the comment topology network in two dimensions, mapping each comment node to each pixel point in the image; establishing an emotional interaction function through internal and external factors of the user, building an emotional interaction model in combination with evolutionary game theory and quantifying the comment emotion; embedding the quantified emotion into the corresponding pixel points in the topic image, convolving the topic two-dimensional topic image through a convolutional neural network, using a function to convert the output of the network into probability values ​​of two categories, and the category corresponding to the maximum probability value is used as the prediction result. The present invention visualizes the comment topology network to mine the secrecy in the rumor topic comments, improves the accuracy of rumor detection, and provides a new direction for the study of the mechanism of social network rumors.

Description

Technical Field

[0001] The present invention belongs to the field of social network analysis, and in particular relates to a rumor detection method based on topic visualization. Background Art

[0002] In recent years, with the development of smart mobile terminals, the Internet has changed the way people interact and communicate with its low cost, simple access, and fast information dissemination. Online social networking (OSN) platforms have become an indispensable part of people's daily lives. For many people, social media and online portals are more popular than traditional newspapers in terms of news search and reading. Although people can obtain a variety of information on the Internet, this information is not always beneficial to us, and it can also bring negative byproducts-rumors. Rumors refer to unverified statements initiated from different sources that spread from one point to another in the network on a regular basis. Compared with the influence of traditional media on the spread of rumors, online social platforms, with their characteristics of participation, openness, communication, dialogue, and community, have enabled the spread of rumors to achieve the characteristics of cross-regional, cross-group, and short-term.

[0003] In addition, while social networks provide people with convenient information exchange and interaction, they also reduce the cost of spreading false information, and gradually develop into a hotbed for the rampant growth of online rumors. In previous studies on rumor detection, most of them focused on the content characteristics, propagation characteristics and user relationship networks of rumors. The impact of comment topological relationships on rumor topic detection was not considered. Therefore, research on the comment topological structure in the rumor topic space can better explore the confidentiality of comment relationships under rumor topics, and also grasp the impact of user subjective consciousness and personal attributes on rumor topics. It is of great significance to the purification of the network environment, social stability and national security.

[0004] In recent years, with the spread of the epidemic, rumors about the virus have emerged in an endless stream. Manual rumor refutation alone can no longer keep up with the speed of rumor spread, so the automatic rumor detection technology has attracted a large number of researchers. The research on rumor detection is currently divided into retrospective rumor detection, early rumor detection, and real-time rumor detection. With the rapid development of neural networks, rumor detection methods are no longer limited to analyzing rumors based on content and network structure. Gradually, they have realized the importance of multimedia information in the rumor detection process and began to use image labels and external knowledge to assist rumor detection. The advantages of deep learning in processing massive data are applied to dealing with complex problems in social networks.

[0005] With the continuous expansion of social network data and the gradual maturity of deep learning technology, nonlinear models such as neural networks and deep learning have gradually become the mainstream choice for analyzing complex networks. Ma et al. (Ma J, Gao W, Wong K F. Rumor detection on twitter with tree-structured recursive neural networks [C]. Association for Computational Linguistics, 2018.) proposed a model based on topic propagation network. The propagation tree model was established through two RvNN models based on bottom-up and top-down tree structures, and the high-level representation of tree nodes in different branches was enhanced through recursion. Finally, the content semantics was linked to propagation to detect rumors. This paper shows that topic rumors can be effectively detected using topic propagation structure. Inspired by this paper and image forgery recognition technology, this paper proposes a rumor detection method based on topic visualization, which combines image structure and comment topological relationship. It can not only detect topic rumors more accurately, but also analyze the relationship between topic comment topological relationship and image pixel position.

[0006] In summary, the existing technical problems are:

[0007] 1. Complexity and diversity of user sentiment tendencies: In the topic comment space, the sentiment of user comments is related to the user's attributes. The sentiment tendency of user comments is not only affected by authoritative users, but also by the user's recent status.

[0008] 2. Image forgery recognition is to identify whether it is a forged image by comparing the smoothness, consistency, continuity, regularity, periodicity and image background noise of the forged area with the original image at the pixel point. How to establish a connection between the characteristics of the topic image and the real image characteristics is the key issue studied in this paper. Summary of the invention

[0009] In order to solve the above technical problems, the present invention proposes a rumor detection method based on topic visualization, comprising:

[0010] S1: Obtain all comments and content and user information data containing rumors and non-rumor topics through social network public APIs or public data websites;

[0011] S2: Perform data cleaning on the acquired data to obtain structured data and delete abnormal or invalid data;

[0012] S3: The processed topic content and comments are converted into vector representations through the Doc2vec algorithm. The adversarial strength of the comments is quantified by the Topic2Image algorithm. The user comments are embedded into a low-dimensional dense space through the adversarial strength to obtain a two-dimensional representation of the comment topology network. The two-dimensional representation of the comment topology network is converted into a pixel array by cutting and diffusing operations while keeping the relative positions of the comment nodes unchanged to obtain the comment privacy image.

[0013] S4: Quantify the emotional interaction influence according to the emotional interaction influence model based on evolutionary game theory, modify the user's emotions through the emotional interaction influence, and mine the user's true emotions;

[0014] S5: Fill the mined user's real emotion into the pixel array of the comment privacy image to obtain a two-dimensional topic image containing comment privacy and comment emotion;

[0015] S6: Convolutional neural network (CNN) is used to convolve the topic two-dimensional topic image, and the output of CNN is converted into probability values ​​of two categories using Softmax function, and the category corresponding to the maximum probability value of the two categories is taken as the prediction result.

[0016] Beneficial effects of the invention: The invention proposes for the first time the idea of ​​visualizing the comment topology network, and has achieved good results in the field of rumor detection. At the same time, by visualizing topic comments, the secrecy in rumor topic comments is effectively excavated (the greater the antagonism between comments, the greater the distance between them in space), which provides a new direction for the study of the mechanism of social network rumors. The effectiveness of the topic visualization rumor detection method also contributes to purifying the network environment and maintaining social stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is a flow chart of the present invention;

[0018] Figure 2 This is an example diagram of topic image detection of the present invention;

[0019] Figure 3 This is a schematic diagram of mining the privacy between comments using the Topic2Image algorithm of the present invention;

[0020] Figure 4 A schematic diagram of the mechanism for quantifying the emotional interaction effect in the present invention;

[0021] Figure 5 Schematic diagram of the CNN neural network of the present invention. DETAILED DESCRIPTION

[0022] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0023] A rumor detection method based on topic visualization, such as Figure 1 As shown, including:

[0024] S1: Obtain all comments and content and user information data containing rumors and non-rumor topics through social network public APIs or public data websites;

[0025] S2: Perform data cleaning on the acquired data to obtain structured data and delete abnormal or invalid data;

[0026] S3: The processed topic content and comments are converted into vector representations through the Doc2vec algorithm, and the adversarial strength of the comments is quantified. The user comments are embedded into a low-dimensional dense space through the adversarial strength to obtain a two-dimensional representation of the comment topology network. The two-dimensional representation of the comment topology network is converted into a pixel array by cutting and diffusing operations while keeping the relative positions of the comment nodes unchanged, and the comment privacy image is obtained.

[0027] S4: Quantify the emotional interaction influence according to the emotional interaction influence model based on evolutionary game theory, modify the user's emotions through the emotional interaction influence, and mine the user's true emotions;

[0028] S5: Fill the mined user's real emotion into the pixel array of the comment privacy image to obtain a two-dimensional topic image containing comment privacy and comment emotion;

[0029] S6: Convolutional neural network (CNN) is used to convolve the two-dimensional topic image, and the Softmax function is used to convert the output of CNN into probability values ​​of two categories. The category corresponding to the maximum probability value of the two categories is used as the prediction result, that is, to determine whether the topic image is a rumor, such as Figure 2 shown.

[0030] Data can be obtained from public data websites or through mature social network public APIs. What needs to be obtained here is all the comments on rumor and non-rumor topics during their life cycle and the basic information and historical behavior information of the comment publisher. The basic information of the comment publisher needs to obtain the number of fans and followers of the user; the user's historical behavior information includes the user's historical reputation and the post information published by the user in the past month. And pre-process the data.

[0031] The raw data usually obtained is unstructured and cannot be used directly for data analysis. Simple data cleaning can make most of the unstructured data structured, delete abnormal or invalid data, and reduce the complexity of subsequent calculations.

[0032] like Figure 3 As shown in the figure, Doc2vec is used to convert topic content and comments into vector representations, and the adversarial strength of comments is quantified. User comments are embedded into a low-dimensional dense space through the adversarial strength to obtain a two-dimensional representation of the comment topology network. Finally, cutting and diffusion operations are performed while keeping the relative positions of comment nodes unchanged to convert the two-dimensional distribution of comment nodes into a pixel array.

[0033] Comment nodes are continuously distributed in two-dimensional space. We convert the two-dimensional distribution of comment nodes into a pixel array through the following two-step method. The first step is to cut the two-dimensional space with a certain distance vertically and horizontally to form a grid space containing all nodes. Each grid contains at most one comment node. The second step is to keep only one node in the grid with multiple nodes through diffusion operation.

[0034] The strength of the adversarial relationship is reflected in the positional relationship between two comments. The farther apart they are, the greater the strength of the adversarial relationship. The angle and distance between the comments give the positional relationship between the two comments. The strength of the adversarial relationship between the comments is quantified as:

[0035]

[0036] Among them, Force(u i ,u j ) distance Represents user u i ,u j The straight-line distance of the comments in space,

[0037] Force(u i ,u j ) angle Represents user u i ,u j The relative angle of the comments in space, and Represents user u i ,u j The vector representation of the kth comment, where m represents the dimension of the vector.

[0038] Cutting and diffusion operations are performed while keeping the relative positions of the comment nodes unchanged to convert the two-dimensional representation of the comment topology network into a pixel array, including:

[0039] Cutting operation: according to the distribution of comment nodes in two-dimensional space, all nodes are moved parallel to the first quadrant, and the cutting distance d is determined. Cut from the origin to the positive direction of the x and y axes at a distance of d until all comment nodes are included. After cutting, a grid space containing all comment nodes is formed, and each grid will contain zero or more comment nodes.

[0040] According to the distribution of comment nodes in two-dimensional space, all nodes are moved parallel to the first quadrant and the cutting distance d is determined, which is expressed as:

[0041] Move parallel to the first quadrant:

[0042]

[0043] Determine the cutting distance:

[0044]

[0045] Among them, x new ,y new is the coordinates of all points after moving to the first quadrant, x old ,y old is the coordinate of the comment node after embedding into the two-dimensional space, x min ,y min ,x max ,y max It is the maximum and minimum value of the coordinates of the comment node after being embedded in the two-dimensional space. Node_num is the number of nodes under a topic.

[0046] Diffusion operation: start from the central grid according to the diffusion rules, and process each grid in a clockwise spiral order. When processing a grid containing multiple nodes, retain the node closest to the center, and diffuse the redundant nodes in a clockwise direction to the grids without nodes around them. After the diffusion is completed, each grid contains at most one comment node. Each grid is regarded as a pixel point, and each pixel point represents at most one user comment.

[0047] The diffusion operation is expressed as:

[0048]

[0049] Among them, N i,j It means that the grid that needs to diffuse nodes is in the i-th row and j-th column. Empty(·) means searching for empty grids in a clockwise direction according to the three previous rules.

[0050] The diffusion rules are as follows: first, redundant points can only diffuse to higher-level grids in the same layer or lower layer; second, points diffused from adjacent grids will not diffuse back to the original grids; third, they are not forced to diffuse to the grids of the same layer, that is, when there are no suitable grids to diffuse to around, they should be forced to diffuse to higher layers according to the relative positions of redundant points and reserved points.

[0051] After the diffusion is completed, each grid contains at most one comment node. Each grid is regarded as a pixel point, and each pixel point represents at most one user comment, thus obtaining the comment privacy image Ima(topic).

[0052] Emotional influence function Emot(u i ) is composed of two aspects: user factors and external factors. First, the user factors E are constructed from the user's recent status and activity. in (u i ), construct external factors E from two aspects: topic popularity and user influence out (u i ):

[0053] E in (u i )=Act(u i )×Status(u i )

[0054] E out (u i )=UserInf(u i )×PeriodSpike(topic)

[0055] Among them, E in (u i ) represents the user's own factors constructed from the user's recent status and activity level, E out (u i ) represents the external factors constructed from the two aspects of topic popularity and user influence, Act(u i ) represents user u i The emotional activity of Status(u i ) represents user u i Recent status of UserInf(u i ) represents user u i The influence of the topic, PeriodSpike(topic) indicates the periodic peak value of the topic.

[0056] User emotional activity Act(u i): Considering that some users accept external information and express their emotions, the comments of users with high emotional activity will be relatively more important, so the user emotional activity is defined as:

[0057]

[0058] Among them, Act(u i ) represents user u i The emotional activity of Num ret , Num ori Respectively represent the number of historical reposts and original Weibo posts of the user.

[0059] User recent status Status(u i ): The emotions expressed by users in topics and comments are affected by their recent status. If their status is negative, their emotional expression will also be negative. Therefore, the user's recent status is defined as:

[0060]

[0061] Among them, N p ,N n They represent the number of positive and negative words in the microblogs posted by the user in the past month, P represents the probability that the user's recent status is positive, Status(u i ) represents user u i The recent status.

[0062] Periodic peak value of topic PeriodSpike(topic): Rumors and non-rumors have an obvious feature in the time series. Rumors often have multiple periodic peaks, while non-rumors usually have only one significant peak. Therefore, the peak value of topic heat is defined as:

[0063] PeriodSpike(topic)=Count(h)

[0064] Where PeriodSpike represents the peak value of the topic’s popularity, H = {h 1 ,h 2 ,h 3 ,…,h n} is a list of topic popularity sampled by time, h n represents the topic popularity at time n, and Count(·) is the statistical peak function.

[0065] User Inf(u i ): The influence of information released by different opinion leaders is different. Some opinion leaders may guide the judgment of their fans. Therefore, the influence of users is defined based on their historical reputation points, number of fans and number of followers:

[0066]

[0067] Among them, UserInf(u i ) represents user u i The influence of W i ∈[0, 1] is the weight of the user’s historical credit score, Represents user u i Number of fans and followers.

[0068] Then, the emotional influence function is constructed by integrating the internal and external factors of the user and using the multivariate linear regression algorithm as follows:

[0069]

[0070] Among them, Emot positive (u i ) represents the positive emotion influence function, Emot negative (u i ) represents the negative sentiment influence function, ρ 0 , 1 , 2 Represents the first, second, and third partial regression coefficients obtained by training using the multivariate linear regression algorithm. Respectively represent the positive and negative sentiment information for user u i external influencing factors.

[0071] Since the emotions of user comments are affected by topics, opinion leaders, and their own status, this paper proposes an emotional interaction model based on evolutionary game theory, such as Figure 4 As shown in the figure, we quantify the emotional interaction: First, we define two game strategies based on game theory: "positive emotional information" and "negative emotional information". P1 and P2 represent the proportion of positive and negative emotions in topics, parent comments, and users' recent microblogs, respectively, and P1+P2=1. The profit functions of the two strategies are:

[0072] profit positive (u i )=P 1 ×Emot positive (u i )

[0073] profit negative (u i )=P 2 ×Emot negative (u i )

[0074] Among them, profitpositive (u i ) represents the positive emotional information profit function, profit negative (u i ) represents the negative sentiment information benefit function, P1 and P2 represent the proportion of positive and negative sentiment in topics, parent comments and users’ recent microblogs, respectively, and Emot positive (u i ) represents the positive emotion influence function, Emot negative (u i ) represents the negative sentiment influence function.

[0075] Then, the evolutionary game theory is used to measure the emotional interaction:

[0076]

[0077] Among them, Mut positive (u i ),Mut negative (u i ) represent the positive and negative information effects on user u after the game. i Emotional impact, profit positive (u i ),profit negative (u i ) represent the profit functions of the two strategies of positive emotional information and negative emotional information respectively, and e represents the natural base.

[0078] Finally, considering that users are affected by other factors when expressing their emotions, the user's emotions are corrected through emotional interaction to mine the user's true emotions. The correction function is:

[0079]

[0080] Among them, regu(u i ) represents user u i The true emotion, α and β represent the adjustment coefficients, Emotion(u i ) represents user u i Adjust the previous emotional tendency, regu(u i ) represents user u i Adjusted sentiment, Mut positive (u i ),Mut negative (u i ) represent the positive and negative information effects on user u after the game. i Emotional impact.

[0081] like Figure 5As shown in the figure, the two-dimensional topic image is convolved by a convolutional neural network, and the Softmax function is used to convert the output of the CNN into probability values ​​of two categories, which are expressed as:

[0082] p(n,r|topic)=softmax(cnn_model(Ima user&emot (topic)))

[0083] Among them, p(n,r|topic) represents the probability of rumor and non-rumor output by softmax, Ima user&emot (topic) represents the topic image with user emotional features added, topic represents the topic, n represents non-rumor, r represents rumor, and cnn_model() represents the convolution operation of the convolutional neural network on the two-dimensional topic image.

[0084] The category corresponding to the maximum probability value of the two categories is taken as the prediction result, which is expressed as:

[0085]

[0086] Among them, Y=0 means that the topic is not a rumor, Y=1 means that the topic is a rumor, p(r|topic) and p(n|topic) represent the probabilities of the topic being a rumor and a non-rumor, respectively.

[0087] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A rumor detection method based on topic visualization, characterized in that: include: S1: Obtain all comments and content and user information data containing rumors and non-rumor topics through social network public APIs or public data websites; S2: Perform data cleaning on the acquired data to obtain structured data and delete abnormal or invalid data; S3: The processed topic content and comments are converted into vector representations through the Doc2vec algorithm, and the adversarial strength of the comments is quantified. The user comments are embedded into a low-dimensional dense space through the adversarial strength to obtain a two-dimensional representation of the comment topology network. The two-dimensional representation of the comment topology network is converted into a pixel array by cutting and diffusing operations while keeping the relative positions of the comment nodes unchanged, and the comment privacy image is obtained. S4: Quantify the emotional interaction influence according to the emotional interaction influence model based on evolutionary game theory, modify the user's emotions through the emotional interaction influence, and mine the user's true emotions; S5: Fill the mined user's real emotion into the pixel array of the comment privacy image to obtain a two-dimensional topic image containing comment privacy and comment emotion; S6: Convolve the two-dimensional topic image through a convolutional neural network, use the Softmax function to convert the output of the convolutional neural network into probability values ​​of two categories, and take the category corresponding to the maximum probability value of the two categories as the prediction result.

2. According to the rumor detection method based on topic visualization according to claim 1, it is characterized in that: Quantify the adversarial strength of comments, expressed as: Among them, Force(u i ,u j ) distance Represents user u i ,u j Comments on the straight-line distance in space, Force(u i ,u j ) angle Represents user u i ,u j The relative angle of the comments in space, and Represents user u i ,u j The vector representation of the kth comment, where m represents the dimension of the vector.

3. The rumor detection method based on topic visualization according to claim 1 is characterized in that: Cutting and diffusion operations are performed while keeping the relative positions of the comment nodes unchanged to convert the two-dimensional representation of the comment topology network into a pixel array, including: Cutting operation: according to the distribution of comment nodes in the two-dimensional space, all nodes are moved parallel to the first quadrant, and the cutting distance d is determined. From the origin to the positive direction of the x and y axes, the distance d is cut until all comment nodes are included. After cutting, a grid space containing all comment nodes is formed, and each grid contains zero or more comment nodes; Diffusion operation: Starting from the central grid, each grid is processed in a clockwise spiral order according to the diffusion rule. When processing a grid containing multiple nodes, the node closest to the center is retained, and redundant nodes are preferentially diffused to the grids without nodes in the surrounding clockwise direction. After the diffusion is completed, each grid contains at most one comment node. Each grid is regarded as a pixel point, and each pixel point represents at most one user comment; The diffusion rules are as follows: first, redundant points can only diffuse to higher-level grids in the same layer or lower layer; second, points diffused from adjacent grids will not diffuse back to the original grids; third, they are not forced to diffuse to the grids of the same layer, that is, when there are no suitable grids to diffuse to around, they should be forced to diffuse to higher layers according to the relative positions of redundant points and reserved points.

4. The rumor detection method based on topic visualization according to claim 3 is characterized in that: According to the distribution of comment nodes in two-dimensional space, all nodes are moved parallel to the first quadrant and the cutting distance d is determined, which is expressed as: Among them, x max ,y max It is the maximum value of the coordinates of the comment node after being embedded in the two-dimensional space, and node_num represents the number of nodes under a topic.

5. The rumor detection method based on topic visualization according to claim 1 is characterized in that: The emotional interaction influence model based on evolutionary game theory is used to quantify the emotional interaction influence, which is expressed as: Among them, Mut positive (u i ),Mut negative (u i ) represent the positive and negative information effects on user u after the game. i Emotional impact, profit positive (u i ),profit negative (u i ) represent the profit functions of the two strategies of positive emotional information and negative emotional information respectively, and e represents the natural base.

6. The rumor detection method based on topic visualization according to claim 1 is characterized in that: Through the emotional interaction, the user's emotions are corrected and the user's true emotions are excavated, which is expressed as: Among them, regu(u i ) represents user u i The true emotion, α and β represent the adjustment coefficients, Emotion(u i ) represents user u i Adjust the previous emotional tendency, regu(u i ) represents user u i Adjusted sentiment, Mut positive (u i ),Mut negative (u i ) represent the positive and negative information effects on user u after the game. i Emotional impact.

7. The rumor detection method based on topic visualization according to claim 1 is characterized in that: The two-dimensional topic image is convolved through a convolutional neural network, and the output of the convolutional neural network is converted into probability values ​​of two categories using the Softmax function, which is expressed as: p(n,r|topic)=softmax(cnn_model(Ima user&emot (topic))) Among them, p(n,r|topic) represents the probability of rumor and non-rumor output by softmax, Ima user&emot (topic) represents the topic image with user emotional features added, topic represents the topic, n represents non-rumor, r represents rumor, and cnn_model() represents the convolution operation of the convolutional neural network on the two-dimensional topic image.

8. The rumor detection method based on topic visualization according to claim 1 is characterized in that: The category corresponding to the maximum probability value of the two categories is taken as the prediction result, which is expressed as: Among them, Y=0 means that the topic is not a rumor, Y=1 means that the topic is a rumor, p(r|topic) and p(n|topic) represent the probabilities of the topic being a rumor and a non-rumor, respectively.

Images