A method for early detection and warning of rumors based on social networks

Abstract

This invention discloses an early detection and warning method for rumors based on social networks. The method first preprocesses the user and social response text along the user's propagation path to generate a social network heterogeneous graph. This graph is then loaded into the social network for early rumor detection, thereby identifying rumors. This early rumor detection method can better capture user characteristics, textual information features, and propagation characteristics in the early stages of posting on social platforms, enabling the early detection of rumors within a small propagation range.

Description

Technical Field

[0001] This invention relates to the field of rumor detection, and in particular to the field of early rumor detection on social media platforms. Background Technology

[0002] With its low cost, rapid dissemination, and wide reach, social media has become a primary platform for communication and information acquisition in people's daily lives, and is gradually becoming the main carrier of information. While social platforms have brought great convenience to people's lives, they have also provided channels for the spread of rumors. The rapid and widespread dissemination of rumors on online platforms has had a significant impact on society and the economy.

[0003] Currently, there are two main types of rumor detection technologies: text-content-based methods and information dissemination path-based methods. Text-content-based methods detect rumors by comparing them with known facts or analyzing textual features and writing style. While this method can detect rumors in their early stages of spread and has a good accuracy rate, it cannot adapt to changes in the text style of false information. Furthermore, existing methods are designed for long texts, while social media platforms typically use short texts, leading to data sparsity and hindering rumor detection. Information dissemination path-based methods detect rumors by constructing trees or graphs to analyze the dissemination path. This method is more robust than content-based methods, but it cannot detect rumors at an early stage. Also, many users along the dissemination path hide or delete their social media records due to privacy concerns, increasing the difficulty of detection.

[0004] Therefore, to solve the above problems, a rumor early detection and warning technology is proposed for online social platforms, which uses information such as post source text, user information, and social response text as raw input. Summary of the Invention

[0005] The method of this invention is used to detect false information on social media platforms and to detect rumors on social media platforms. It can improve the accuracy and efficiency of detection and can block the spread of false information and rumors in their early stages.

[0006] To achieve the above objectives, the technical solution adopted in this invention is to integrate text information, dissemination user paths, and dissemination users to design an early rumor detection model based on social networks, namely the ER-CM model. After training this ER-CM model on false information and rumors on social platforms, the features of false information and rumors are extracted. The early rumor detection model (ER-CM model) based on social networks of this invention mainly includes:

[0007] A bidirectional encoder based on Transformers (BERT) is used to represent the text information content, resulting in a feature representation of the text.

[0008] A Graph Attention Network (GAT) is used to aggregate user information and social response text information in heterogeneous graphs of social networks to obtain user representations and text representations.

[0009] Residual Gated Graph ConvNets (RGGCN) are used to aggregate heterogeneous information such as users, text, and knowledge entities in heterogeneous graphs of social networks to obtain user representations and text representations.

[0010] Global Mean Pooling (GMP) is used to aggregate information from the user set and text set in the heterogeneous graph of the social network to obtain global user and text labels.

[0011] The present invention provides an early detection and warning platform for rumors on a social platform, which includes a social network heterogeneous graph generation module, a node information representation module, and a rumor detection and classification module;

[0012] The social network heterogeneous graph generation module has two main aspects: first, it generates a user social network heterogeneous graph UG from the user set USER based on the user propagation path; second, it generates a graph based on the text set TEXT. USER The first aspect is to obtain the text social network heterogeneous graph TG from the Wikipedia entity set WE and the ConceptNet entity set CE; the second aspect is to fuse UG and TG to generate the social network heterogeneous graph GG.

[0013] The node information representation module uses the GAT model and the RGGCN model to generate user representations. and text representation

[0014] The rumor detection and classification module is based, firstly, on user characteristics. and text representation The first step is to obtain the heterogeneous graph representation of social networks (MGG); the second step is to use a fully connected layer (FC) based on MGG to determine whether it is a rumor.

[0015] Compared with existing methods, the method of this invention has the following advantages: by linking text with entities from Wikipedia and ConceptNet knowledge graphs, entity features are fused with text features, enriching the feature representation of the text. In the process of constructing a heterogeneous graph of social networks, the method fully utilizes the early-stage user, text, and propagation information features of rumors on social networks. Compared with existing methods, the performance of early rumor detection is improved when the number of propagating users is small. Attached Figure Description

[0016] Figure 1 It is a heterogeneous graph of user social networks.

[0017] Figure 2 It is a heterogeneous graph of text-based social networks.

[0018] Figure 3 This is a diagram of the early rumor detection structure based on social networks constructed by this invention.

[0019] Figures 4-9 This is a comparison chart of early detection and evaluation indicators for rumors using different detection methods. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0021] In this invention, the information text, user, and propagation path from the same online social platform are used. The user social network heterogeneous graph is denoted as UG, the text social network heterogeneous graph as TG, and the social network heterogeneous graph as GG. The early rumor detection model based on social networks is denoted as the ER-CM model.

[0022] Part 1: Early Detection and Warning Platforms for Rumors on Social Media Platforms

[0023] See Figure 3 As shown, the present invention targets an early detection and warning platform for rumors on social platforms, which includes a social network heterogeneous graph generation module, a node information representation module, and a rumor detection and classification module.

[0024] The social network heterogeneous graph generation module has two main aspects: first, it generates a user social network heterogeneous graph UG from the user set USER based on the user propagation path; second, it generates a graph based on the text set TEXT. USER The first aspect is to obtain the text social network heterogeneous graph TG from the Wiki entity set WE and the ConceptNet entity set CE; the second aspect is to fuse UG and TG to generate the social network heterogeneous graph GG.

[0025] The node information representation module uses the GAT model and the RGGCN model to generate user representations. and text representation

[0026] The rumor detection and classification module is based, firstly, on user characteristics. and text representation The first step is to obtain the heterogeneous graph representation of social networks (MGG); the second step is to use a fully connected layer based on MGG to determine whether it is a rumor.

[0027] Part Two: Constructing the ER-CM Model

[0028] This invention targets the feature extraction of early rumors and early misinformation on social network platforms, and constructs an early rumor detection model (ER-CM model) based on social networks based on these features. The construction of the ER-CM model includes the following steps.

[0029] Modeling step one: Set the propagation user length ζ;

[0030] In this invention, the propagation user length is denoted as ζ; the ζ is set to an array.

[0031] For example: the propagation user length ζ is set to 5, 10, 15, 20, 25, 30, 40, 50, that is, ζ = [5, 10, 15, 20, 2, 5, 30, 40, 50].

[0032] Modeling step two: Obtaining the user propagation path;

[0033] In this invention, the user propagation path is denoted as propa. Propa refers to the transmission path between users on the same online social platform obtained based on ζ.

[0034] The multiple users existing on the user propagation path propa are represented in set form, denoted as user set USER, and The subscript i represents the user's identifier, and the subscript B represents the total number of users along the user's propagation path.

[0035] user1 represents the first user.

[0036] user2 represents the second user.

[0037] user i This represents the i-th user.

[0038] user j This represents the j-th user.

[0039] user B This indicates the last user.

[0040] For ease of explanation, the user mentioned above iAlso known as any user, the user j Also referred to as another arbitrary user, and user user i user j It is not the same user on the user propagation path.

[0041] In this invention, any user i Including the number of words in the self-introduction Number of characters in a nickname Number of characters in nickname Number of followers Number of fans Number of texts created Number of likes Save this post Difference between account creation time and post creation time Is the account verified? Allow location services User gender User i These twelve original features are a set of user profile information. Used, i.e.

[0042]

[0043] Similarly, we can deduce that the j-th user... j The user profile information set, denoted as and

[0044]

[0045] Similarly, the user profile information set for the first user, user1, is denoted as... and

[0046]

[0047] Similarly, the user profile information set for the second user, user2, is denoted as... and

[0048]

[0049] Similarly, we can deduce that the last user is... B The user profile information set, denoted as and

[0050]

[0051] In this invention, the user profile information set carried by the user set USER on the user propagation path propa is denoted as the user-carried information set MPR, and

[0052]

[0053] Modeling step three: Generate the initial feature set of the text;

[0054] In this invention, social response text information published by each user in the user set USER is obtained from the propagation user path propa, and the initial text features are generated after being processed by the BERT pre-trained model.

[0055] Modeling step 301: Obtaining social response text information;

[0056] Because any user on a social networking platform i It enables actions such as forwarding, commenting, and replying to comments on a specific post, generating social response text messages during these actions, denoted as... Therefore, all social response text information published by users on the propa user path are represented as a set, denoted as TEXT. USER ,and

[0057]

[0058] This represents the text message posted by user1.

[0059] This represents the text message posted by user2.

[0060] Indicates user i Published text information.

[0061] Indicates user j Published text information.

[0062] Indicates user B Published text information.

[0063] In this invention, for forwarding TEXT USER For users who respond with text messages on social media, this invention considers their text messages to be the source post text messages forwarded by that user. Therefore, the TEXT... USER The set length is assigned the value ζ.

[0064] Modeling step 302, application of BERT pre-trained model;

[0065] Using BERT pre-trained models on the social response text set TEXT USER The process is performed to generate an initial feature set of the text, denoted as FT, and The number of initial text feature sets (FTs) is the same as the number of social response text sets (TEXTs).

[0066] Representing text information The initial features of the text obtained after BERT.

[0067] Representing text information The initial features of the text obtained after BERT.

[0068] Representing text information The initial features of the text obtained after BERT.

[0069] Representing text information The initial features of the text obtained after BERT.

[0070] Representing text information The initial features of the text obtained after BERT.

[0071] In this invention, the BERT pre-trained model is referenced from Devlin J., Chang MW, Lee K., et al. BERT: Pre-Training of Deep Bidirectional Transformers for Language Understanding[J]. 2018.

[0072] Modeling step four: Preprocess the social response text set;

[0073] Modeling step 401 involves linking information from Wikipedia entries using the TAGME tool;

[0074] In this invention, the TAGME toolkit is used to process any text information. Linking to Wikipedia as an entity yields the linked Wikipedia entry – the descriptive text, denoted as .

[0075] Any text information Multiple Wikipedia descriptive texts can be represented as a set, denoted as Wikipedia descriptive text set. and The subscript W is The total number of entities linked to Wikipedia.

[0076] express The first Wikipedia entry linked to Wikipedia – the descriptive text.

[0077] express The second Wikipedia entry linked to Wikipedia – the descriptive text.

[0078] express The last Wikipedia entry linked to Wikipedia - the description text.

[0079] Modeling step 402, application of BERT pre-trained model;

[0080] Using BERT pre-trained models After processing, the Wikipedia-entity feature set is obtained, denoted as... and Quantity and The quantities are consistent.

[0081] This refers to Wikipedia - description text. The first Wikipedia entity feature obtained after training with BERT.

[0082] This refers to Wikipedia - description text. The second Wikipedia entity feature obtained after training with BERT.

[0083] This refers to Wikipedia - description text. The last Wikipedia entity feature obtained after training with BERT.

[0084] Similarly, text information Wikipedia - a collection of descriptive texts, denoted as and Its corresponding Wikipedia entity feature set is denoted as and

[0085] Similarly, text information Wikipedia - a collection of descriptive texts, denoted as and Its corresponding Wikipedia entity feature set is denoted as and

[0086] Similarly, text information Wikipedia - a collection of descriptive texts, denoted as and Its corresponding Wikipedia entity feature set is denoted as and

[0087] Modeling step 403: Feature merging based on Wikipedia entry information;

[0088] In this invention, TEXT USER The Wikipedia-description text sets corresponding to all texts are merged to obtain the Wiki Entity Set MWE.

[0089]

[0090] In this invention, TEXT USER The Wikipedia entity feature set corresponding to all texts is merged to obtain the Wikipedia entity feature set MFW.

[0091]

[0092] In this invention, the TAGME toolkit is referenced from Ferragina P., Scaiella U. TAGME: On-the-Fly Annotation of Short Text Fragments (by Wikipedia Entities) [A]. Proceedings of the 19th ACM international conference on Information and knowledge management-CIKM'10 [C]. Toronto, ON, Canada: ACM Press, 2010: 1625.

[0093] Modeling step 404 involves linking based on the ConceptNet knowledge graph in the COCOEX tool;

[0094] In this invention, the TAGME toolkit is used to process any text information. Link entities with the ConceptNet knowledge graph to obtain the linked graph-description text, denoted as...

[0095] Any text information Multiple corresponding graph-description texts are represented as a set, denoted as _ ... and The subscript C is The total number of entities linked to the ConceptNet knowledge graph.

[0096] express The first graph linked to the ConceptNet knowledge graph – the description text.

[0097] express The second graph linked to the ConceptNet knowledge graph – the description text.

[0098] express The last graph linked to the ConceptNet knowledge graph – the description text.

[0099] Modeling step 405: Application of the ConceptNet-Numberbatch toolkit;

[0100] Using the ConceptNet-Numberbatch toolkit After processing, a map-entity feature set is obtained, denoted as... and Quantity and The quantities are consistent.

[0101] Representation of the graph - description text The first graph – entity features – after post-processing using the ConceptNet-Numberbatch toolkit.

[0102] Representation of the graph - description text The second graph – entity features – is post-processed using the ConceptNet-Numberbatch toolkit.

[0103] Representation of the graph - description text The last graph feature processed by the ConceptNet-Numberbatch toolkit – entity features.

[0104] Similarly, text information The map-description text set, denoted as and Its corresponding graph-entity feature set is denoted as and

[0105] Similarly, text information The map-description text set, denoted as and Its corresponding graph-entity feature set is denoted as and

[0106] Similarly, text information The map-description text set, denoted as and Its corresponding graph-entity feature set is denoted as and

[0107] Modeling step 406: Feature merging based on ConceptNet knowledge graph;

[0108] In this invention, TEXT USER The graph-description text sets corresponding to all texts in the ConceptNet knowledge graph representation are merged to obtain the graph entity set MCE.

[0109]

[0110] In this invention, TEXT USER The graph-entity feature sets of all texts corresponding to the ConceptNet knowledge graph representation are merged to obtain the graph entity feature set MFC.

[0111]

[0112] In this invention, ConceptNet and ConceptNet-Numberbatch refer to Speer R., Chin J., Havasi C. ConceptNet 5.5: An Open Multilingual Graph of General Knowledge[J].arXiv, 2018.

[0113] In the present invention, the COCOEX toolkit refers to Becker M., Korfhage K., Frank A. COCO-EX: ATool for Linking Concepts from Texts to ConceptNet [A]. Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: System Demonstrations [C]. Online: Association for Computational Linguistics, 2021: 119–126.

[0114] Modeling step five: Construct a heterogeneous graph of the social network;

[0115] In this invention, based on the user set USER and the text set TEXT USER The text initial feature set FT, wiki entity feature set MFW, graph entity feature set MFC, and user-carried information set MPR are used to construct the heterogeneous graph of the social network.

[0116] Modeling step 501: Construct a heterogeneous graph of user social networks based on users and the information they carry;

[0117] In this invention, the heterogeneous graph of the user social network is denoted as UG.

[0118] This will propagate to any user on the user path propa. i As a node in UG The The node features are node With nodes The connection edges between them are constructed according to three relationships: forwarding time sequence, comment reply, and cosine similarity.

[0119] In this invention, the forwarding time sequence refers to sorting all users according to the time they participated in forwarding or commenting, generating a forwarding user time sequence linked list, and connecting users with adjacent time sequences using edges, i.e., users j It is user i The first user to participate in the propagation path, then user i and user j There is an edge connecting them, with a weight of 1.

[0120] In this invention, the comment reply relationship refers to establishing an edge between users based on the comment reply relationship, i.e., user. i For user j If a user replies to or shares a post or comment, then the user... i and user j There exists an edge with a weight of 1.

[0121] In this invention, the cosine similarity relationship is used The cosine similarity between users is used to measure user similarity in order to uncover potential relationships between users. First, a complete graph of user relationships is constructed, with the edge weights being the cosine similarity of user features. Then, a sparse graph is generated from the edges to obtain the potential connections between users.

[0122] In this invention, user i With user j cosine similarity for:

[0123]

[0124]

[0125]

[0126]

[0127] For user i A collection of user profile information.

[0128] For user j A collection of user profile information.

[0129] For user i The number of words in your self-introduction.

[0130] For user i The number of characters in a username.

[0131] For user i User gender.

[0132] For user j The number of words in your self-introduction.

[0133] For user j The number of characters in a username.

[0134] For user j User gender.

[0135] In this invention, triples are used to describe the relationships between edges between users, denoted as ME. U ,and It is user i With user j The edge relationships between them.

[0136] For example: Indicates user i With user j Adding propagation links in the time sequence, user i For user jComments were generated, and the user feature similarity between the two was 0.8. The heterogeneous user social network graph is then represented as UG = (USER, ME...). U MPR), such as Figure 1 As shown.

[0137] Modeling step 502 involves constructing a heterogeneous graph of the text social network based on text information, initial text features, two entity sets and their feature sets;

[0138] In this invention, the text social network heterogeneous graph is denoted as TG.

[0139] Text set TEXT USER Any text in the Wikipedia entity set WE and ConceptNet entity set CE are taken as text nodes, and their node features are FT. Any entity in the Wikipedia entity set WE and ConceptNet entity set CE are taken as entity nodes, and their node features are FC and FW, respectively.

[0140] Edges, denoted as ME, are established between text nodes based on their comment / reply relationships. T That is, user i For user j If a post or comment receives a reply or is forwarded, then and There are edges between them, with a weight of 1. Similar to the comment reply relationship in the heterogeneous graph of the user social network UG, bidirectional edges are used.

[0141] To avoid confusion regarding entity source information, there are no edge connections between entity nodes. The edge between an entity node and a text node is denoted as EE. T Connected by a one-way edge according to their relationship, i.e., text The Wikipedia entity we exists in China. j Or ConceptNet entity ce j Then we j or ce j and There exists an edge with the direction of entity pointing to text and a weight of 1. The one-way edge is to prevent text information from flowing back into entity information and affecting the correctness of entity information.

[0142] In this invention, the heterogeneous graph of a text social network is represented as TG = {TEXT∪WE∪CE,ME} T ∪EE T ,FT∪FW∪FC}, such as Figure 2 As shown.

[0143] Modeling step 503: Construct a social network heterogeneous graph based on the user social network heterogeneous graph and the text social network heterogeneous graph;

[0144] The user social network heterogeneous graph UG and the text social network heterogeneous graph TG are merged to generate a social network heterogeneous graph GG.

[0145] In this invention, there is a one-way connection between text information and the user. When text information is injected into user information features, i.e., the user... i Published text but and user i There exists an edge with direction . Point to user i The weight is 1.

[0146] In this invention, only users who forward the message and the source user share the same viewpoint, meaning there exists an edge where the source text points to the forwarding user.

[0147] Modeling step six: Calculate graph representation based on heterogeneous graphs of social networks;

[0148] Modeling step 601, User set feature representation

[0149] In this invention, the user set feature representation is denoted as... and The subscript l is the identifier of the network layer, and l = [0, 1, ..., L], where L is the total number of network layers.

[0150] It is the representation of user1 at level l.

[0151] It is the representation of user2 at level l.

[0152] It is user B The representation at layer l.

[0153] Initially, the user set features, i.e., the 0th layer representation, consist of the user's original profile information, denoted as... and

[0154]

[0155] For user i A collection of user profile information.

[0156] For user j A collection of user profile information.

[0157] For user B A collection of user profile information.

[0158] Modeling step 602, text set feature representation

[0159] In this invention, the text set feature representation is denoted as... and

[0160]

[0161] It is text The representation at layer l.

[0162] It is text The representation at layer l.

[0163] It is text The representation at layer l.

[0164] Initially, the text set features, i.e., the 0th layer representation, are the original concise information of the text, denoted as... and

[0165]

[0166] Modeling step 603, Wikipedia entity representation;

[0167] In this invention, the Wikipedia entity representation is denoted as... and The subscript K represents the total number of entities in the merged entity set. K represents the total number of entities in the text set TEXT. USER The entity set is obtained by merging the Wikipedia entities and their features corresponding to all texts in the dataset.

[0168] It is a representation of the Wikipedia entity we1.

[0169] It is a representation of the Wikipedia entity we2.

[0170] It is a Wikipedia entity we K The representation of.

[0171] Initially, Wikipedia entities were represented as . and

[0172] Modeling step 604, based on ConceptNet entity representation;

[0173] In this invention, ConceptNet entity representation is denoted as... and The subscript O represents the total number of entities in the merged entity set. The O represents the total number of entities in the text set TEXT. USERThe entity set is obtained by merging the ConceptNet entities and their features corresponding to all texts in the dataset.

[0174] It is a representation of entity ce1.

[0175] It is a representation of entity ce2.

[0176] It is a physical entity CE O The representation of.

[0177] Initially, the ConceptNet entity representation is denoted as... and

[0178] In this invention, Wikipedia entity representation and ConceptNet entity representation It remains unchanged in the model, that is

[0179] Modeling step 605 is based on the aggregation of the GAT method and the RGGCN method;

[0180] In this invention, the GAT method is first used to aggregate and generate user and text feature representations; then the RGGCN method is used to inject text representations into user features and entity features into text features, and the calculation formula is as follows.

[0181]

[0182] Inject feature representations of user features into text representations.

[0183] Injecting feature representations of entity features into text features.

[0184] In this invention, after L-layer transformation, the final output user representation is obtained. and final text representation

[0185] It is the representation of user1 at level L.

[0186] It is the representation of user2 at level L.

[0187] It is user B The representation at layer L.

[0188] It is text The representation at layer L.

[0189] It is text The representation at layer L.

[0190] It is text The representation at layer L.

[0191] In this invention, the GAT method is referenced from Velikovi P., Cucurul G., Casanova A., et al. GraphAttention Networks[J].2017.

[0192] In this invention, the RGGCN method is referenced from Bresson X., Laurent T. Residual Gated GraphConvNets[J].arXiv,2018.

[0193] Modeling step 606: Graph representation of heterogeneous graphs in social networks;

[0194] In this invention, the generated social network heterogeneous graph is GG, and the GMP method is used to represent users. and text representation Pooling is performed, and then a threshold is used to perform global representation MGG on the heterogeneous graph of the social network. The calculation formula is as follows:

[0195]

[0196]

[0197]

[0198]

[0199] This represents the global features of the heterogeneous graph UG representing the user's social network.

[0200] It is a global feature of the heterogeneous graph TG of text social networks.

[0201] g represents the threshold value.

[0202] exp represents the exponent.

[0203] θ U This indicates the weight of user parameters.

[0204] θ T This represents the weight of the text information parameters.

[0205] b represents the bias value.

[0206] MGG represents the global features of a heterogeneous graph in a social network.

[0207] Modeling step seven: Fitting and detecting rumors;

[0208] Using MGG as the input vector for the fitted prediction, the information type is finally obtained, i.e., the rumor detection result y, where y∈{0,1}, 0 indicates that it is not a rumor, and 1 indicates that it is a rumor. The calculation formula is as follows:

[0209] y=θ·MGG+b (11)

[0210] θ represents the parameter weight.

[0211] b represents the bias value.

[0212] Part Three: Methods for Early Rumor Detection Using the ER-CM Model

[0213] The first step in the detection process is to initialize and set the current propagation user length.

[0214] The ER-CM model constructed according to this invention is installed on a network social platform. The network social platform is started, and the ER-CM model is initialized. The current propagation user length is set, denoted as ζ. 当前 .

[0215] Step two of the detection process: Obtain the user propagation path;

[0216] According to ζ 当前 Obtain the transmission paths between users on the same online social platform to get the data belonging to ζ. 当前 user propagation path 当前 Then record the propa 当前 All users existing on the platform 当前 And a user profile information set carried by each user.

[0217] Step 3 involves detection and generating an initial feature set for the text.

[0218] During the detection process, those belonging to the propa 当前 The social response text information posted by users is categorized, and initial text features are generated after processing by the BERT pre-trained model.

[0219] Step four of the detection process involves linking entities in Wikipedia entries and the ConceptNet knowledge graph.

[0220] The TAGME toolkit is used to link entities in Wikipedia. Then, the Wikipedia feature associations are output after training in the BERT pre-trained model. Finally, the Wikipedia entities and their features are merged to obtain the Wikipedia entity set MWE and the Wikipedia entity feature set MFW.

[0221] The TAGME toolkit is used to link entities in the ConceptNet knowledge graph, and then the graph feature association is completed in the ConceptNet-Numberbatch toolkit. Finally, ConceptNet entities and their features are merged to obtain the graph entity set MCE and the graph entity feature set MFC.

[0222] Step 5 of the detection process: Construct a heterogeneous graph of the current user's social network;

[0223] Initially, the current user's social network heterogeneous graph UG 当前 =(USER 当前 ,UE U ,MPR) is an empty set.

[0224] The propagation user path obtained in step two will be detected. 当前 The user set USER 当前 Any user as UG 当前 The nodes in the data have the following characteristics: Edges between nodes are constructed based on three relationships: forwarding order, comment reply, and cosine similarity. Triples are used to describe the relationships between users, denoted as ME. U .

[0225] Any text in the text set obtained in step three is used as a text node.

[0226] Any entity in the Wikipedia entity set and ConceptNet entity set obtained in step four is taken as an entity node, and its node features are as follows.

[0227] Edges are established between text nodes based on their comment / reply relationships; no edges connect entity nodes. The edge between an entity node and a text node is denoted as EE. T By connecting the elements according to their affiliation with unidirectional edges, we obtain the heterogeneous graph TG of the text social network. 当前 .

[0228] UG will be configured according to the correspondence between users and text. 当前 With TG 当前 Merging to generate heterogeneous graphs of social networks (GG) 当前 .

[0229] Detection step six: graphical representation;

[0230] Generate user set representations using GAT and RGGCN models. and text set representation User representation using GMP method and text representation Pooling, then for GG 当前 Using thresholding techniques to generate graphical representations of MGG 当前 .

[0231] Step seven of the detection process: fitting and detecting rumors.

[0232] MGG 当前 The input information type y∈{0,1} is used to determine whether it is a rumor or not.

[0233] Example 1

[0234] The experimental dataset used in this embodiment is the Weibo dataset. Compared with the RvNN, GRU, PPC_RNN+CNN, BiGCN, and KZWANG methods, the performance of the method in early rumor detection in this invention was evaluated. The performance of the model in early rumor detection was tested on the Weibo dataset, with a set number of users spreading the rumor ζ = [5, 10, 15, 20, 2, 5, 30, 40, 50]. An additional experiment with ζ = 100 was included to highlight the effectiveness of this invention in early rumor detection. The parameters of the social network-based early rumor detection model proposed in this invention are updated using the momentum-driven SGD algorithm, with a momentum set to 0.9, an initial learning rate of 0.01, a batch size of 64, 128 epochs, a dropout layer ratio of 0.5, and a training / test set ratio of 0.3. A step-rate adjustment strategy (StepLR) was used to adjust the learning rate, with a step size of 32 epochs and an adjustment factor of 0.1, all implemented using PyTorch.

[0235] Figures 4-9 The experimental results are shown, with the horizontal axis representing the number of users who disseminated the information, and the vertical axis representing accuracy, precision, recall, and F1 score, respectively. As can be seen from the figure, the accuracy, precision, recall, and F1 score of this invention are significantly better than the baseline models in the early stages of rumor dissemination, except for a few points. The accuracy was improved by 4.33%, 3.16%, 3.08%, 4.91%, 4.13%, 3.86%, 3.86%, and 4.16%, respectively, demonstrating the effectiveness of this invention in the early detection and warning of rumors.

[0236] In Example 1, the Weibo dataset and GRU method are referenced from Ma J., Gao W., Mitra P., et al. Detecting Rumors from Microblogs with Recurrent Neural Networks [A]. Proceedings of the Twenty-Fifth International Joint Conference on Artificial Intelligence [C]. New York, New York, USA: AAAI Press, 2016: 3818–3824.

[0237] In Example 1, the RvNN method refers to Ma J., Gao W., Wong KFRumor Detection on Twitter with Tree-Structured Recursive Neural Networks [A]. Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) [C]. Melbourne, Australia: Association for Computational Linguistics, 2018:1980–1989.

[0238] In Example 1, the PPC_RNN+CNN method refers to Liu Y., Wu YFEearly Detection of FakeNews on Social Media Through Propagation Path Classification with Recurrentand Convolutional Networks[J]. Proceedings of the AAAI Conference on Artificial Intelligence, 2018, 32(1).

[0239] In Example 1, the BiGCN method is referenced from Bian T., Xiao X., Xu T., et al. Rumor Detection on Social Media with Bi-Directional Graph Convolutional Networks[J]. Proceedings of the AAAI Conference on Artificial Intelligence, 2020, 34(01): 549–556.

[0240] In Example 1, the KZWANG method is referenced from Ke Z., Li Z., Zhou C., et al. Rumor Detection on Social Media via Fused Semantic Information and a Propagation Heterogeneous Graph[J]. Symmetry, 2020, 12(11): 1806.

Claims

1. A rumour early detection and warning platform for a social platform, characterized in that: It includes a social network heterogeneous graph generation module, a node information representation module, and a rumor detection and classification module; The social network heterogeneous graph generation module, firstly, involves generating user sets... The heterogeneous graph of user social networks is obtained based on the user propagation path. ; The second aspect is based on the text set Wiki Entity Set and ConceptNet entity set Obtain the heterogeneous graph of the text social network The third aspect is to and Fusion generates heterogeneous graphs of social networks ; The node information representation module uses the GAT model and the RGGCN model to generate user representations. and text representation ; The rumor detection and classification module is based, firstly, on user characteristics. and text representation Obtain heterogeneous graph representations of social networks The second aspect is based on Use a fully connected layer (FC) to determine if it is a rumor; Design an ER-CM model for early rumor detection based on social networks, which has the following construction steps: Modeling Step 1: Set the Propagation User Length ; Set the propagation user length It is an array; Modeling step two: Obtaining the user propagation path; The user propagation path is denoted as... The It refers to the basis Obtain the transmission path between users on the same online social platform; User propagation path Multiple users existing on the platform are represented in set form, denoted as user set. ; User propagation path The set of user profile information for all users is denoted as the user-carried information set. ; Modeling step three: Generate the initial feature set of the text; Modeling step 301: Obtaining social response text information; Propagation User Path All social response text messages posted by users on the platform are denoted as follows: The The set length is assigned a value. ; Modeling step 302, application of BERT pre-trained model; Using BERT pre-trained models on social response text sets The process is performed to generate an initial feature set for the text, denoted as... ; Modeling step four: Preprocess the social response text set; Modeling step 401 involves linking information from Wikipedia entries using the TAGME tool; Modeling step 402, application of BERT pre-trained model; Modeling step 403: Feature merging based on Wikipedia entry information; Will The Wikipedia entity set is obtained by merging the Wikipedia descriptive text sets corresponding to all texts. ; Will The Wikipedia entity feature set corresponding to all texts is merged to obtain the Wikipedia entity feature set. ; Modeling step 404 involves linking based on the ConceptNet knowledge graph in the COCOEX tool; Modeling step 405: Application of the ConceptNet-Numberbatch toolkit; Modeling step 406: Feature merging based on ConceptNet knowledge graph; Will The graph-description text sets corresponding to all texts in the ConceptNet knowledge graph representation are merged to obtain the graph entity set. ; Will The graph-entity feature sets of all texts corresponding to the ConceptNet knowledge graph representation are merged to obtain the graph entity feature set. ; Modeling step five: Construct a heterogeneous graph of the social network; Based on user set Text set Initial feature set of text Wikipedia entity feature set Graph entity feature set and user-carried information sets To complete the construction of the heterogeneous graph of the social network; Modeling step 501: Construct a heterogeneous graph of user social networks based on users and the information they carry; A heterogeneous graph of user social networks, denoted as ; Modeling step 502 involves constructing a heterogeneous graph of the text social network based on text information, initial text features, two entity sets and their feature sets; Text-based social network heterogeneous graph, denoted as ; Text set Any text in the text is used as a text node, and the text node is characterized by: ; Wiki Entity Set and ConceptNet entity set Any entity is taken as an entity node, and its node characteristics are as follows: , ; Modeling step 503: Construct a social network heterogeneous graph based on the user social network heterogeneous graph and the text social network heterogeneous graph; Modeling step six: Calculate graph representation based on heterogeneous graphs of social networks; Modeling step 601, User set feature representation ; Modeling step 602, text set feature representation ; Modeling step 603, Wikipedia entity representation; Modeling step 604, based on ConceptNet entity representation; Modeling step 605 is based on the aggregation of the GAT method and the RGGCN method; First use The method aggregates and generates user and text feature representations; then it uses... The method injects text representations into user features and entity features into text features; Modeling step 606: Graph representation of heterogeneous graphs in social networks; In the generated social network heterogeneous graph Use The method represents the user and text representation Pooling is then used, followed by a thresholding method to perform a global representation of the heterogeneous graph of the social network. ; Modeling step seven: Fitting and detecting rumors; Will The rumor detection result is obtained by using the input vector as the fitted prediction.

2. The construction of the ER-CM model in the early rumor detection and warning platform for social platforms according to claim 1 includes the following steps: Modeling Step 1: Set the Propagation User Length ; Set the propagation user length It is an array; Modeling step two: Obtaining the user propagation path; The user propagation path is denoted as... The It refers to the basis Obtain the transmission path between users on the same online social platform; User propagation path Multiple users existing on the platform are represented in set form, denoted as user set. ; Any user The set of user profile information possessed is denoted as ,Right now ; User propagation path The set of user profile information for all users is denoted as the user-carried information set. ; Modeling step three: Generate the initial feature set of the text; Modeling step 301: Obtaining social response text information; Propagation User Path All social response text messages posted by users on the platform are denoted as follows: The The set length is assigned a value. ; Modeling step 302, application of BERT pre-trained model; Using BERT pre-trained models on social response text sets The process is performed to generate an initial feature set for the text, denoted as... ; Modeling step four: Preprocess the social response text set; Modeling step 401 involves linking information from Wikipedia entries using the TAGME tool; Using the TAGME toolkit to process any text information Linking to Wikipedia results in the linked Wikipedia entry – descriptive text, denoted as . ; Any text information Multiple Wikipedia descriptive texts can be represented as a set, denoted as Wikipedia descriptive text set. ; Modeling step 402, application of BERT pre-trained model; Using BERT pre-trained models After processing, the Wikipedia-entity feature set is obtained, denoted as... ; Modeling step 403: Feature merging based on Wikipedia entry information; Will The Wikipedia entity set is obtained by merging the Wikipedia descriptive text sets corresponding to all texts. ; Will The Wikipedia entity feature set corresponding to all texts is merged to obtain the Wikipedia entity feature set. ; Modeling step 404 involves linking based on the ConceptNet knowledge graph in the COCOEX tool; Using the TAGME toolkit to process any text information Link entities with the ConceptNet knowledge graph to obtain the linked graph-description text, denoted as... ; Any text information Multiple corresponding graph-description texts are represented as a set, denoted as _ ... ; Modeling step 405: Application of the ConceptNet-Numberbatch toolkit; Using the ConceptNet-Numberbatch toolkit After processing, a map-entity feature set is obtained, denoted as... ; Modeling step 406: Feature merging based on ConceptNet knowledge graph; Will The graph-description text sets corresponding to all texts in the ConceptNet knowledge graph representation are merged to obtain the graph entity set. ; Will The graph-entity feature sets of all texts corresponding to the ConceptNet knowledge graph representation are merged to obtain the graph entity feature set. ; Modeling step five: Construct a heterogeneous graph of the social network; Based on user set Text set Initial feature set of text Wikipedia entity feature set Graph entity feature set and user-carried information sets To complete the construction of the heterogeneous graph of the social network; Modeling step 501: Construct a heterogeneous graph of user social networks based on users and the information they carry; A heterogeneous graph of user social networks, denoted as ; Propagation user path Any user on As One of the nodes The The node features are ;node With nodes The connection edges between them are constructed according to three relationships: forwarding time sequence, comment reply, and cosine similarity. Forwarding time sequence relationship refers to sorting all users according to the time they participated in forwarding or commenting, generating a forwarding user time sequence linked list, and connecting users with adjacent time sequences with an edge weight of 1; Comment reply relationship refers to establishing edges between users based on the comment reply relationship, with each edge having a weight of 1; Cosine similarity between users for: ； ； ； ； For users A collection of user profile information; For users A collection of user profile information; For users The number of words in a self-introduction; For users The number of characters in the username; For users User gender; For users The number of words in a self-introduction; For users The number of characters in the username; For users User gender; Use triples to describe the edge relationships between users. ; User With users The edge relationship between them; Modeling step 502 involves constructing a heterogeneous graph of the text social network based on text information, initial text features, two entity sets and their feature sets; Text-based social network heterogeneous graph, denoted as ; Text set Any text in the text is used as a text node, and the text node is characterized by: ; Wiki Entity Set and ConceptNet entity set Any entity is taken as an entity node, and its node characteristics are as follows: , ; Modeling step 503: Construct a social network heterogeneous graph based on the user social network heterogeneous graph and the text social network heterogeneous graph; heterogeneous graphs of user social networks Heterogeneous graphs of text-based social networks Merging to generate heterogeneous graphs of social networks ; Modeling step six: Calculate graph representation based on heterogeneous graphs of social networks; Modeling step 601, User set feature representation ; User set feature representation, denoted as subscript This is the identifier for the network layer. Initially, the user set features, i.e., the 0th layer representation, consist of the user's original profile information, denoted as... ,and ; For users A collection of user profile information; For users A collection of user profile information; For users A collection of user profile information; Modeling step 602, text set feature representation ; Text set feature representation, denoted as ; Initially, the text set features, i.e., the 0th layer representation, are the original concise information of the text, denoted as... ,and ; Modeling step 603, Wikipedia entity representation; Wikipedia entity representation, denoted as ; Initially, Wikipedia entities were represented as . ,and ; Modeling step 604, based on ConceptNet entity representation; ConceptNet entity representation, denoted as ; Initially, the ConceptNet entity representation is denoted as... ,and ; Modeling step 605 is based on the aggregation of the GAT method and the RGGCN method; First use The method aggregates and generates user and text feature representations; then it uses... The method injects text representations into user features and entity features into text features, with the calculation formula as follows; ； ； Injecting feature representations of user characteristics into text representations; Injecting feature representations of entity features into text features; Modeling step 606: Graph representation of heterogeneous graphs in social networks; In the generated social network heterogeneous graph Use The method represents the user and text representation Pooling is then used, followed by a thresholding method to perform a global representation of the heterogeneous graph of the social network. The calculation formula is as follows: ； ； ； ； Represents a heterogeneous graph of user social networks Global features; It is a text-based social network heterogeneous graph Global features; Indicates the threshold value; Indicates an exponent; Indicates the weight of user parameters; Indicates the weight of text information parameters; Indicates the bias value; Represents the global features of a heterogeneous graph in a social network; Modeling step seven: Fitting and detecting rumors; Will The input vector used for fitting and prediction ultimately yields the information type, i.e., the rumor detection result. ,and , This indicates that it is not a rumor. To indicate that it is a rumor, the calculation formula is as follows: ； Indicates parameter weights; This represents the bias value.

3. A method for early detection of rumors on social media platforms using the ER-CM model according to claim 1, characterized in that... It includes the following steps: The first step in the detection process is to initialize and set the current propagation user length. Install the ER-CM model on a social networking platform, start the platform, initialize the ER-CM model, and set the current number of users to propagate the model, denoted as . ; Step two of the detection process: Obtain the user propagation path; in accordance with Obtain the transmission paths between users on the same online social platform to get the belonging User propagation path Then record the above. All users existing on and the user profile information set carried by each user; Step 3 involves detection and generating an initial feature set for the text. During the detection process, those belonging to the above The text information of social responses posted by users is categorized, and initial text features are generated after processing by the BERT pre-trained model; Step four of the detection process involves linking entities in Wikipedia entries and the ConceptNet knowledge graph. The TAGME toolkit is used to perform entity linking on Wikipedia. Then, the BERT pre-trained model is trained to output Wikipedia feature associations. Finally, Wikipedia entities and their features are merged to obtain a Wikipedia entity set. and Wikipedia entity feature set ; The TAGME toolkit is used to perform entity linking in the ConceptNet knowledge graph, then the ConceptNet-Numberbatch toolkit is used to complete graph feature association, and finally ConceptNet entities and their features are merged to obtain the graph entity set. and graph entity feature set ; Step 5 of the detection process: Construct a heterogeneous graph of the current user's social network; Initially, the current user's social network heterogeneous graph It is an empty set; Will User set on Any user as The nodes in the data have the following characteristics: The edges between nodes are constructed based on three relationships: forwarding order, comment reply, and cosine similarity. Triples are used to describe the relationships between users, denoted as . ; Use any text in the text set obtained in step three as a text node; Any entity in the Wikipedia entity set and ConceptNet entity set obtained in step four is taken as an entity node, and its node features are as follows: Edges are established between text nodes according to their comment / reply relationships; no edges connect entity nodes. The edges between entity nodes and text nodes are denoted as... By connecting the elements according to their affiliation with unidirectional edges, a heterogeneous graph of the text social network is obtained. ; Based on the correspondence between users and text and Merging to generate heterogeneous graphs of social networks ; Detection step six: graphical representation; User set representations are generated using the GAT and RGGCN models. and text set representation ;use The method represents the user and text representation Pooling, and then... Generate graph representations using threshold techniques. ; Step 7: Detecting rumors using fitting techniques; Will Input information type In order to distinguish between rumors and non-rumors.

Images