Method, device and equipment for false information detection based on propagation forest

Abstract

The application discloses a false information detection method, device and equipment based on a propagation forest, and the method comprises the following steps: acquiring propagation data of each source blog in a blog propagation in a social network for a source blog set to be detected, wherein the propagation data comprises propagation content and propagation relationship; extracting text features of the propagation content; calculating a propagation prototype and constructing a propagation tree after dividing the source blog set into K class clusters according to the text features; constructing a propagation forest based on the propagation prototype and the propagation tree; aggregating node neighborhood representations in the propagation forest to obtain embedding representations of the nodes; calculating enhanced embedding representations of source blog nodes and subsequent propagation blog nodes according to corresponding text features and embedding representations respectively; and classifying the enhanced embedding representations of the source blog nodes and the subsequent propagation blog nodes by applying a detection model to obtain false detection results of each source blog in the source blog set. The application can improve the performance of a false information detection task.

Description

Technical Field

[0001] This invention relates to the field of data mining technology, specifically to a method, apparatus, and equipment for detecting false information based on propagation forests. Background Technology

[0002] In recent years, thanks to the convenience of social media, a large amount of misinformation has proliferated. Therefore, how to detect misinformation in a timely and accurate manner is an urgent research topic in the field of social network analysis, with significant research and practical application implications.

[0003] Most disinformation campaigns revolve around a specific trending topic or public opinion event; therefore, there are potential semantic connections between different pieces of misinformation. Furthermore, many pieces of misinformation are actually orchestrated by large malicious organizations, making their dissemination patterns similar. Therefore, skillfully utilizing the similar semantic connections and dissemination patterns of multiple misinformation dissemination lines can play a positive role in achieving more accurate misinformation detection.

[0004] Existing detection methods focus only on deep modeling of the propagation of individual pieces of information, ignoring the shared propagation patterns behind the propagation of different pieces of information; or they only focus on fine-grained (such as word-level) semantic associations. Since words have different meanings in different contexts, such approaches will introduce noisy semantic information and even limit the performance of existing fake information detection methods. Summary of the Invention

[0005] This invention addresses the key technical problem of how to utilize the similar semantic associations and propagation patterns of multiple false information propagation paths. It provides a false information detection method, apparatus, and device based on propagation forests to improve the performance of false information detection tasks.

[0006] The specific technical solution of the present invention is as follows:

[0007] According to a first aspect of the present disclosure, the present invention provides a method for detecting false information based on a propagation forest, the method comprising the following steps:

[0008] For the set of source blog posts to be detected, obtain the dissemination data of each source blog post on social networks; wherein, the dissemination data includes: dissemination content and dissemination relationship;

[0009] Extract the text features of the disseminated content;

[0010] After dividing the source blog post set into K clusters according to text features, the center vector of each cluster is used as the propagation prototype, and a propagation tree for each cluster is constructed based on the text features and propagation relationship corresponding to each cluster.

[0011] Based on the propagation prototype and the propagation tree, a propagation forest is constructed; wherein, the nodes of the propagation forest include: source blog post nodes, propagation prototype nodes, and subsequent propagation blog post nodes in the propagation tree, and the edges of the propagation forest include: connections between the source blog post nodes and the propagation prototype nodes defined by the pseudo-labels of the clusters, connections between the source blog post nodes and subsequent propagation blog post nodes, and corresponding connections between the subsequent propagation blog post nodes; the initial representations of the source blog post nodes and the subsequent propagation blog post nodes are text features, and the initial representation of the propagation prototype nodes is the center vector of the cluster;

[0012] Aggregate the neighborhood representations of nodes in the propagation forest to obtain the embedded representation of the nodes;

[0013] For the source blog post node and the subsequent propagated blog post node, an enhanced embedding representation is calculated based on the corresponding text features and embedding representations, respectively;

[0014] The enhanced embedding representations of the source blog post nodes and the subsequent propagated blog post nodes are classified using a detection model to obtain the false detection results of each source blog post in the source blog post set.

[0015] Furthermore, the subsequent dissemination blog post nodes include at least one of forwarding nodes and comment nodes.

[0016] Furthermore, the step of dividing the source blog post set into K clusters based on text features, and using the center vector of each cluster as the propagation prototype, includes:

[0017] K vectors are randomly initialized; each vector corresponds to a propagation prototype.

[0018] For each text feature in the set of text features, the text feature is assigned to the nearest propagation prototype in the semantic space. To generate several clusters Where t represents the number of iteration rounds;

[0019] According to each cluster Calculate the cluster based on the text features in the text. The center vector is used to obtain the propagation prototype.

[0020] Based on the distribution of the aforementioned text features and various propagation prototypes Calculate the loss function φ;

[0021] If the loss function φ does not converge, let t = t + 1, and return to the step of assigning each text feature in the text feature set to the nearest propagation prototype in the semantic space. To generate several clusters

[0022] If the loss function φ converges, the propagation prototype is obtained.

[0023] Furthermore, the loss function in, Representing a cluster The number of text features described in the text, For the cluster In the text feature, β represents the smoothing term.

[0024] Further, the aggregation of node neighborhood representations in the propagation forest to obtain the embedding representation of the node includes:

[0025] By using graph convolutional networks to aggregate and propagate neighborhood information of nodes in a forest, node representations are obtained.

[0026] The nodes are represented using graph convolutional layers. The data is passed to the propagation tree to obtain the embedded representation of the node.

[0027] Further, the step of calculating enhanced embedding representations for the source blog post node and the subsequent propagated blog post nodes based on text features and embedding representations, respectively, includes:

[0028] Calculate the enhanced embedding representation of the source blog node Among them, W e b represents the first learnable parameter. e Let x represent the second learnable parameter. i This indicates the source blog post node x i The corresponding text features, This indicates the source blog post node x i The corresponding embedded representation;

[0029] Calculate the enhanced embedding representation of the subsequent propagated blog post nodes. Among them, c ij This indicates the subsequent propagation blog post node c. ij The corresponding text features, the This indicates the subsequent propagation blog post node c. ij The corresponding embedded representation.

[0030] According to a second aspect of the present disclosure, the present invention provides a fake information detection device based on a propagation forest, the device comprising:

[0031] The data acquisition module is used to acquire the dissemination data of each source blog post in the social network for the set of source blog posts to be detected; wherein, the dissemination data includes: dissemination content and dissemination relationship;

[0032] The data processing module is used to extract the text features of the disseminated content; after dividing the source blog post set into K clusters according to the text features, the center vector of each cluster is used as the propagation prototype, and the propagation tree of each cluster is constructed based on the text features and the propagation relationship corresponding to each cluster.

[0033] The propagation forest construction module is used to divide the source blog posts in the propagation data into K clusters according to their text feature sets, use the center vector of each cluster as the propagation prototype, and construct a propagation forest based on the propagation prototype and the propagation tree. The nodes of the propagation forest include: source blog post nodes, propagation prototype nodes, and subsequent propagation blog post nodes under the propagation tree. The edges of the propagation forest include: connections between the source blog post nodes and the propagation prototype nodes defined by the pseudo-labels of the clusters, connections between the source blog post nodes and subsequent propagation blog posts, and corresponding connections between the subsequent propagation blog posts. The initial representations of the source blog post nodes and the subsequent propagation blog posts are text features, and the initial representation of the propagation prototype nodes is the center vector of the cluster.

[0034] The vector representation enhancement module is used to aggregate the node neighborhood representations in the propagation forest to obtain the embedding representation of the node; for the source blog post node and the subsequent propagation blog post node, the enhanced embedding representation is calculated according to the corresponding text features and embedding representations respectively;

[0035] The information detection module is used to classify the enhanced embedded representations of the source blog post nodes and the subsequent propagated blog post nodes in combination with the original false information model, and obtain the false detection results of each source blog post in the source blog post set.

[0036] According to a third aspect of the present disclosure, the present invention provides an electronic device including a memory and a processor, wherein the memory stores a computer program, the computer program being loaded and executed by the processor to implement any of the above-described methods for detecting false information based on propagation forests.

[0037] According to a fourth aspect of the present disclosure, the present invention provides a computer-readable storage medium having stored thereon computer program instructions, characterized in that, when executed by a processor, the program instructions implement any of the above-described methods for detecting false information based on propagation forests.

[0038] Compared with the prior art, the present invention has at least the following advantages: 1) The framework provided by the present invention can explore coarse-grained semantic associations between multiple propagation lines, and can effectively avoid noise problems in fine-grained semantics;

[0039] 2) The framework provided by this invention can model the potential coarse-grained semantic and topological relationships between multiple propagation trees, thereby mining more accurate and richer high-order semantic and structural features for false information detection;

[0040] 3) The framework provided by this invention is universal and can be used to improve any propagation tree-based method for detecting false information.

[0041] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0042] Figure 1 The flowchart illustrates the implementation of the propagation forest-based fake information detection framework provided by this invention.

[0043] Figure 2 The diagram shows the structure of the false information detection system based on propagation forest provided by this invention. Detailed Implementation

[0044] The present invention will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that the described embodiments are only intended to facilitate the understanding of the present invention and do not constitute any limitation thereof.

[0045] Figure 1 This invention provides a flowchart of the implementation of a fake information detection framework based on propagation forests, specifically including the following steps:

[0046] Step A: Obtain the dissemination data of blog posts on social networks and extract features from the text content.

[0047] Step A1: Obtain propagation data from the social network (including propagation content and propagation relationships during the propagation process), and define the propagation tree of sample i as G based on the propagation data. i = <V i E i >, among which It is a set of nodes, where each element represents a blog post in the propagation process. Specifically, x i For the source blog post, For subsequent blog posts; It is a set of edges, where This represents the connection between the s-th node and the t-th node in the set of node i, representing a propagation relationship (such as forwarding, commenting, etc.) in the propagation process. Let the adjacency matrix of the propagation tree corresponding to sample i be initially defined as:

[0048]

[0049] Step A2: For any sample i, extract the source blog post x i and subsequent blog posts Text features, definition For the source blog post x i Textual features, To spread blog posts ij Text features, where d0 is the dimension of the text features.

[0050] Step B: Generate a propagation prototype based on the source blog post content of the propagation data in Step A, and construct a propagation forest to explore the potential relationships between similar propagation trees;

[0051] Step B1: Propagate the prototype generation module, given the text features {x1,…,x} of the source blog posts in the dataset. N}, where N is the number of samples. This module divides a given dataset into K clusters by minimizing the objective function φ, defining the center of each cluster as {m1,…,m K The center vector is the propagation prototype, denoted as {m1,…,m}. K The loss function is defined as follows:

[0052]

[0053] Where m represents the propagation prototype, β is the smoothing term, Z is the number of sample points contained in each cluster, and x z These are the text features within cluster z. The specific process is described below:

[0054] 1) Randomly initialize K vectors to represent the propagation prototype, i.e., m1,…,m K ;

[0055] 2) Based on the loss function φ, repeat the following two processes within a finite number of iterations until φ converges:

[0056] a. Text features x for each sample i i It is assigned to the propagation prototype that is closest in the semantic space, i.e., argmin. k ||x i -m k ||2

[0057] b. For each cluster, recalculate the cluster center to update the corresponding propagation prototype, i.e.

[0058] Step B2: Based on the generated propagation prototype and propagation tree, construct the propagation forest G. forest =(V forest E forest ), where V forest Denotes the set of nodes, E forest Represents the set of edges.

[0059] Specifically, V forest =V S UV R UV C , where V S ={x1,…,x N} represents the set of nodes formed by the source blog posts, where N is the number of samples. The initial features of this type of node are the text features of the corresponding content; V R ={m1,…,m K} represents the set of nodes composed of propagation prototypes, where K is the number of propagation prototypes. The initialization characteristics of this type of node adopt the propagation prototype m from step B1 above; V C This represents the set of nodes formed by subsequent propagation blog posts in each propagation tree. The initialization features of this type of node are initialized using the text features of the corresponding content.

[0060] There are three types of edges: First, based on the pseudo-label of the cluster, the connection between the source blog post node and the propagation prototype node is defined, and the edge weight is based on the probability that the source blog post is assigned to that cluster. Second, based on the propagation relationship, the connection between the source blog post node and subsequent propagation blog post nodes is defined, and the edge weight is set to 1. Third, based on the propagation relationship, the connection between subsequent propagation blog post nodes is defined, and the edge weight is set to 1.

[0061] Step C: Based on the propagation forest constructed in Step B, aggregate the node neighborhood representations in the propagation forest and mine the semantic and structural associations between similar propagation trees; specifically including:

[0062] First, graph convolutional layers are used to aggregate neighborhood information of nodes in the propagation forest to capture higher-order semantics and propagation structure features in similarity propagation. The message passing formula for layer 1 in the graph is calculated as follows:

[0063]

[0064] in, This represents the high-order features of each node output after the first graph convolutional layer. Representative of the forest G forestInitialization characteristics of nodes. ij It is the weight of node i and node j, a ii It is the weight of the node itself, defined as a. ii =1. W (1) and It is a trainable parameter matrix, and σ represents the activation function. Let i represent the set of neighboring nodes of node i.

[0065] Secondly, the graph convolutional layer is used again to pass the aggregated node information from the upper layer to similar propagation trees. The graph message passing calculation method of the second layer is as follows:

[0066]

[0067] W represents the high-order features of each node output after the second graph convolutional layer. (2) and It is a trainable parameter matrix.

[0068] Step D: Design an embedding enhancement module based on the root node. Utilize the node representations learned in Step C to enhance the node representations of the source and subsequent propagation posts extracted by the original fake information detection model. The enhanced node representations of the source and subsequent propagation posts are obtained according to the following calculation method:

[0069]

[0070] in, To spread the source blog node x in the forest i The learned node representation, x i Let x′ be the original blog post node representation before enhancement for sample i. i This represents the enhanced source blog post node; To propagate subsequent blog post nodes in the forest ij The learned node representation, c ij Let c′ be the pre-boost node representation of the j-th propagation node in the propagation tree of sample i. ij Its enhanced propagation node representation; W e and b e These are learnable parameters.

[0071] Step E: Based on the enhanced feature representation obtained in Step D, a classifier is used to detect false information. The enhanced source blog post node representation x′ based on sample i is then used. i The subsequent propagation of blog post nodes is represented by c′. ij Using existing fake information detection models, predict the category label of sample i.

[0072] In summary, the present invention has the following technical effects:

[0073] I. The framework provided by this invention can simultaneously model the potential coarse-grained semantic and topological relationships between multiple propagation trees, thereby mining more accurate and richer high-order semantic and structural features for fake information detection. On public datasets, this framework improves the detection accuracy by about 2% to 5% compared with existing fake information detection methods.

[0074] Second, the framework provided by this invention is practical and universal. Validated on public datasets, this framework can effectively improve the detection accuracy of existing propagation tree-based fake information detection methods.

[0075] like Figure 2 As shown, the present invention also provides a false information detection device based on propagation forest, the device comprising: a data acquisition and preprocessing module, a propagation forest construction module, a propagation embedding enhancement module based on propagation forest, and a false information detection module.

[0076] The data acquisition module 100 is used to acquire the dissemination data of each source blog post in the social network for the set of source blog posts to be detected; wherein, the dissemination data includes: dissemination content and dissemination relationship;

[0077] The data processing module 200 is used to extract the text features of the disseminated content; after dividing the source blog post set into K clusters according to the text features, the center vector of each cluster is used as the propagation prototype, and a propagation tree of the cluster is constructed based on the text features and the propagation relationship corresponding to each cluster.

[0078] The propagation forest construction module 300 is used to divide the source blog posts in the propagation data into K clusters according to their text feature sets, use the center vector of each cluster as the propagation prototype, and construct a propagation forest based on the propagation prototype and the propagation tree. The nodes of the propagation forest include: source blog post nodes, propagation prototype nodes, and subsequent propagation blog post nodes under the propagation tree. The edges of the propagation forest include: connections between the source blog post nodes and the propagation prototype nodes defined by the pseudo-labels of the clusters, connections between the source blog post nodes and subsequent propagation blog posts, and corresponding connections between the subsequent propagation blog posts. The initial representations of the source blog post nodes and the subsequent propagation blog posts are text features, and the initial representation of the propagation prototype nodes is the center vector of the cluster.

[0079] The vector representation enhancement module 400 is used to aggregate the node neighborhood representations in the propagation forest to obtain the embedding representation of the node; for the source blog post node and the subsequent propagation blog post node, the enhanced embedding representation is calculated according to the corresponding text features and embedding representation respectively;

[0080] The information detection module 500 is used to classify the enhanced embedded representations of the source blog post nodes and the subsequent propagated blog post nodes in combination with the original false information model, and obtain the false detection results of each source blog post in the source blog post set.

[0081] For details regarding the specific execution process and beneficial effects of the device module, please refer to the description of the above method embodiments; further details will not be provided here.

[0082] In an exemplary embodiment, a computer device is also provided, the computer device including a memory and a processor, the memory storing a computer program, the computer program being loaded and executed by the processor to implement the above-described method for detecting false information based on propagation forest.

[0083] In an exemplary embodiment, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed by a processor, implements the above-described method for detecting false information based on a propagating forest.

[0084] In an exemplary embodiment, a computer program product is also provided, which, when run on a computer device, causes the computer device to perform the aforementioned method for detecting misinformation based on a propagation forest.

[0085] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for detecting false information based on propagation forest, characterized in that, The method comprises: For a set of source blog posts to be detected, obtain propagation data of each source blog post in the blog posts in the social network; wherein the propagation data comprises propagation content and propagation relationship; extract the text features of the propagation content; After dividing the set of source blog posts into K clusters according to the text features, take the center vector of each cluster as a propagation prototype, and construct a propagation tree for the cluster based on the text features and the propagation relationship corresponding to each cluster; Based on the propagation prototype and the propagation tree, construct a propagation forest; wherein the nodes of the propagation forest include source blog post nodes, propagation prototype nodes, and subsequent propagation blog post nodes in the propagation tree, the edges of the propagation forest include connections between the source blog post nodes and the propagation prototype nodes, connections between the source blog post nodes and subsequent propagation blog post nodes, and corresponding connections between the subsequent propagation blog post nodes, the initial representation of the source blog post nodes and the subsequent propagation blog post nodes is the text feature, and the initial representation of the propagation prototype node is the center vector of the cluster; Aggregate the neighborhood representation of the nodes in the propagation forest to obtain the embedding representation of the nodes; For the source blog post nodes and the subsequent propagation blog post nodes, calculate enhanced embedding representations according to the corresponding text features and embedding representations, respectively; Apply a detection model to the enhanced embedding representations of the source blog post nodes and the subsequent propagation blog post nodes to classify and obtain the fake detection results of each source blog post in the set of source blog posts.

2. The method of claim 1, wherein, The subsequent propagation blog post nodes include at least one of the forwarding nodes and the comment nodes.

3. The method of claim 1, wherein, After dividing the set of source blog posts into K clusters according to the text features, taking the center vector of each cluster as a propagation prototype comprises: randomly initializing K vectors; wherein each vector corresponds to a propagation template for each text feature in the set of text features, assigning the text feature to the nearest propagating prototype in semantic space to generate a number of class clusters where t represents the number of iterations According to the text features in each cluster , a center vector of the cluster is calculated to obtain a propagation prototype based on the distribution of the text features and each propagation prototype computing a loss function φ; In the event that the loss function φ does not converge, let t = t + 1 and return to the assigning of each text feature in the set of text features to the nearest propagated prototype in semantic space to generate a number of class clusters In the case where the loss function φ converges, obtain the propagation prototype.

4. The method of claim 3, wherein, the loss function wherein, denotes a cluster of classes the number of text features in, the text features in the cluster of classes β denotes a smoothing term.

5. The method of claim 1, wherein, The aggregation of the neighborhood representation of the nodes in the propagation forest to obtain the embedding representation of the nodes comprises: The node representation is obtained by aggregating and propagating the neighborhood information of the nodes in the forest using a graph convolution network utilizing a graph convolutional layer to the node representations passing into the propagation tree, resulting in an embedded representation of the node.

6. The method of claim 1, wherein, The calculation of enhanced embedding representations for the source blog post nodes and the subsequent propagation blog post nodes according to the text features and the embedding representations, respectively, comprises: computing an enhanced embedding representation of the source blog node where W e denote first learnable parameters, b e denote second learnable parameters, x i denote the source blog node x i corresponding text features, denote the source blog node x i corresponding embedding representation; computing an enhanced embedding representation of the subsequent propagating blog node where c ij representing the subsequent propagating blog node c ij corresponding to the text features, the representing the subsequent propagating blog node c ij corresponding to the embedding representation.

7. A misinformation detection apparatus based on propagation forest, characterized by, The device comprises: A data acquisition module for obtaining propagation data of each source blog post in the blog posts in the social network for a set of source blog posts to be detected; wherein the propagation data comprises propagation content and propagation relationship; A data processing module for extracting the text features of the propagation content; after dividing the set of source blog posts into K clusters according to the text features, taking the center vector of each cluster as a propagation prototype, and constructing a propagation tree for the cluster based on the text features and the propagation relationship corresponding to each cluster; The propagation forest construction module is configured to divide the source posts in the propagation data into K clusters according to their text feature sets, take the center vector of each cluster as a propagation prototype, and construct a propagation forest based on the propagation prototype and the propagation tree; wherein, the nodes of the propagation forest include source post nodes, propagation prototype nodes, and subsequent propagation post nodes under the propagation tree, and the edges of the propagation forest include connections between the source post nodes and the propagation prototype nodes, connections between the source post nodes and the subsequent propagation post nodes, and corresponding connections between the subsequent propagation post nodes, the initial representation of the source post nodes and the subsequent propagation post nodes is a text feature, and the initial representation of the propagation prototype node is a center vector of the cluster; The vector representation enhancement module is configured to aggregate the neighborhood representation of the nodes in the propagation forest to obtain the embedding representation of the nodes; and for the source post nodes and the subsequent propagation post nodes, respectively calculate enhanced embedding representations according to the corresponding text features and embedding representations. The information detection module is configured to classify the enhanced embedding representations of the source post nodes and the subsequent propagation post nodes in combination with an original false information model to obtain the false detection result of each source post in the source post set.

8. A storage medium having stored therein a computer program, wherein, The computer program is configured to execute the method of any one of claims 1-6 when running. 9.An electronic device comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to execute the computer program to perform the method of any one of claims 1-6.

Images