A rumor detection method and system based on comprehensive information of propagation process

Abstract

A rumor detection method based on comprehensive information of the propagation process includes the following steps: S1: obtaining a public rumor data set; S2: constructing an event propagation graph and a joint graph; S3: text preprocessing and feature extraction; S4: user feature extraction; S5: model training to obtain the optimal performance model; S6: model prediction to obtain model prediction performance and indicators. The present invention also includes a rumor detection system based on comprehensive information of the propagation process, including a data acquisition module, a feature extraction module, a model training module and a model prediction module connected in sequence. The present invention uses the comprehensive information in the propagation process to perform rumor identification and detection, and the comprehensive information includes: text content, user characteristics and propagation structure, and develops a new rumor identification model Text-SAGPool to accurately and quickly identify rumors on social media. Better capture of rumor information on social media and practicality and efficiency in unknown events are achieved.

Description

Technical Field

[0001] The present invention relates to the field of data mining and computer technology, and is a rumor detection method and system based on comprehensive information of a propagation process. Background Art

[0002] While the development of the Internet has promoted the rapid development of economy, culture and life, it has also been accompanied by the emergence of many problems that affect social security and stability. Rumors, as a widespread social problem on social media, are having a significant impact on social stability and people's personal safety. For example, during the 2016 US presidential election, there were a large number of rumors about the two presidential candidates on the mainstream social media Twitter and Facebook in the United States. These massive rumors directly or indirectly affected the results of the election. In addition, during the 2019 COVID-19 pandemic, rumors on social media led to a large number of misunderstandings about epidemic prevention policies and measures, causing serious social unrest. Many people even refused to wear masks due to the "bewitching" of rumors on social media, causing the epidemic to rage. There are many such rumor accidents on social media. Therefore, it is of great significance to design a rumor detection method to identify whether the information spread on social media is a rumor.

[0003] The present invention focuses on the problem of rumor detection on social media. CN202010757089.1 discloses a Weibo rumor detection method combined with an attention mechanism, which mainly uses the text content of Weibo rumor information and combines the deep learning method Text CNN-Attention model to train text representation; CN201810576095.X discloses a Weibo rumor detection method based on a gradient boosting tree, which first constructs training samples according to features, then uses the training samples to train the rumor detection model, and then performs multiple training on the training sample set to obtain multiple regression tree models, each regression tree gives a prediction value, and the prediction values ​​of multiple regression trees are combined to obtain the final Weibo rumor detection model; CN202110903231.3 uses a bidirectional encoder representation technology based on sentiment polarity to perform text analysis on source post information. Data enhancement; then construct the propagation relationship, and at the same time, add the enhanced text vector features to the root node of the propagation relationship, combine with the propagation relationship to form a tree propagation structure, and then start from the root node, the forward tree propagation structure from front to back formed by forwarding and response relationships and the reverse tree propagation structure from back to front formed by gathering the scattered nodes in the community to the upper root node are input into FRBi-GCN respectively; finally, perform node enhancement operations on all nodes including the root node and forwarding node in the propagation tree, and use the softmax classifier to identify whether the event is a rumor. It can be seen that most of the current work on rumor information detection is based on a certain type of information such as text content or propagation structure to identify rumors. The model is trained by a certain type of information in the information propagation process and then the trained model is used to identify rumor information. It is worth affirming that the use of text content or propagation structure information can identify rumors on social media. However, rumor recognition methods based on text content often require certain linguistic expert knowledge, and rumor recognition methods based on propagation structure require rumor propagation to have a certain scale, so its early rumor detection performance is slightly inferior. Therefore, it is necessary to propose a more accurate and faster rumor identification method that can accurately identify unknown events on social media and determine whether the event is rumor information. This is of great significance to the network environment in terms of applicability and accuracy. Summary of the invention

[0004] In order to overcome the defects of traditional rumor detection systems that use a single type of information to identify rumors, resulting in low recognition accuracy and inability to accurately identify rumors at an early stage, the present invention develops a new rumor detection method and system based on comprehensive information of the propagation process.

[0005] The present invention uses the comprehensive information involved in the process of information dissemination on social media to identify rumors, and realizes comprehensive modeling of the information dissemination process by extracting the text content, propagation structure and user information of the information respectively, and uses a graph branch module and a text branch module to capture the propagation structure characteristics, user characteristics and text representation respectively. The SAGPool model is used in the graph branch to comprehensively utilize the propagation structure characteristics and user characteristics, and the text branch uses a convolutional neural network to capture the text representation. Finally, the connection of the outputs of the two branches is realized through a fully connected layer, and then input into the softmax layer to give a predicted label, which greatly improves the detection and identification of rumors on social media and the applicability and efficiency in unknown information.

[0006] The technical solutions adopted by the present invention to achieve the above-mentioned invention object are as follows:

[0007] A rumor detection method based on comprehensive information of a propagation process, characterized in that it comprises the following steps:

[0008] S1: Data acquisition: Obtain public rumor datasets from social media platforms. Each rumor dataset contains comprehensive information on the spread of several rumor events and non-rumor events. Each rumor and non-rumor event contains a source post and several reply posts. The comprehensive information includes spread structure information, text information, and user information. In addition, each event has a classification label corresponding to whether the event is a rumor.

[0009] S2: Construction of event propagation graph and joint graph: Analyze each rumor event or non-rumor event in the dataset, extract its propagation structure information, and model it into an event propagation graph. The root node in the event propagation graph represents the source post, and the other nodes represent the users involved in the event propagation. The edges between the nodes represent the social behavior between the users. All event propagation graphs in the dataset are connected together to form a joint graph of the entire dataset.

[0010] S3: Text preprocessing: Both the source posts and the reply posts contain text content of different lengths. The text content of each rumor event and non-rumor event in the dataset is integrated into a segment. The segment is cleaned using natural language processing tools, and the word2vec tool is used to obtain the text embedding vector of the segment.

[0011] S4: User feature extraction: For each user participating in event propagation, 8 user features are used to characterize the user;

[0012] S5: Model training: Input the joint graph, text embedding vector and user features obtained in S2, S3 and S4 into the designed rumor detection model: Text-SAGPool, use the CNN module in the model to extract text representation, and input the user features as the node features of each node in the joint graph into the SAGPool module in the model to obtain the graph representation of each rumor event or non-rumor event, combine the above-obtained text representation and graph representation together, and finally pass through the MLP layer in the model to obtain the prediction result of the model, calculate the loss of the prediction result and the event label, and use the loss to adjust the model parameters until the training model is obtained after the iteration is completed;

[0013] S6: Model prediction: Process new propagation events into the format of input data and input them into the trained model to predict unknown events.

[0014] Preferably, in step S2:

[0015] S2.1: The rumor events and non-rumor events are all real propagation events in Internet social media. The publisher of the source post is used as the root node of the event propagation graph. The users who participated in the propagation of the source post are generated as other nodes in the propagation graph according to their social behavior actions and time sequence. The social behavior actions between users are used as the edges between nodes to obtain the event propagation graph.

[0016] S2.2: The event propagation graph constructed by S2.1 above is further combined into a joint graph. The joint graph is composed of the event propagation graphs of each event in the data set. There are no edges between the event propagation graphs. The propagation structure of each event in the data set can be comprehensively represented by the joint graph. The joint graph can be regarded as a collection of event propagation graphs of all rumor events and non-rumor events in the data set.

[0017] S2.3: In this method, the adjacency matrix is ​​obtained from the joint graph obtained above as the input of the subsequent model.

[0018] Preferably, in step S3:

[0019] S3.1: Each rumor event and non-rumor event in the dataset contains a different number of posts, and each post contains text information of different lengths. For each rumor event or non-rumor event, all sentences in the event are connected to obtain a segment, which contains all the text information of the entire event;

[0020] S3.2: Perform word segmentation on the above-obtained paragraph, and after word segmentation, a list of all the words in the paragraph can be obtained;

[0021] S3.3: Remove stop words from the word list after word segmentation, use the existing stop word library to filter, remove all words in the word list that are in the stop word library, and obtain a word list without stop words;

[0022] S3.4: Count the word frequencies in the word list excluding stop words, and sort the words by their frequency of occurrence. Extract the top 200 words from the sorted word list and input them into the trained word2vec model to obtain the text embedding vectors of these words.

[0023] Preferably, in step S4:

[0024] S4.1: Extract user information from the dataset. The dataset contains user information for each post and its reply posts. The following eight user features are extracted from the dataset to describe the user: whether the account is verified, whether the account has description text content, whether the account geographic location is verified, the number of account friends, the number of fans, the number of accounts followed by both parties, the length of the account nickname, and the length of the account description text;

[0025] S4.2: After extracting the eight user features, the non-Boolean features in the features need to be normalized by min-max as shown below:

[0026]

[0027] In the formula, x' represents the normalized value of a single data set, min is the minimum value in the data column, and max represents the maximum value in the data column.

[0028] Preferably, in step S5:

[0029] S5.1: First, construct the Text-SAGPool model. There are two variants of Text-SAGPool, namely Text-SAGPoolg and Text-SAGPoolh models. Both variant models are composed of two branches, namely text branch and graph branch. The text branches are the same, and are composed of three CNN modules with kernel sizes of 3, 4, and 5 and three maximum pooling layers connected in sequence. For the Text-SAGPoolg model, three graph convolutional layers connected in sequence are used to perform three graph convolution operations, and a graph pooling layer is used to perform a graph pooling operation. The output of the last graph pooling layer is used as the final graph representation output. For the Text-SAGPoolh model, its graph branch contains three layers of graph convolutional layers and three layers of graph pooling layers, and after each graph pooling operation is completed, it will be input to the next graph convolution layer for graph convolution operation. Finally, the output of the three graph pooling layers is used as the final graph representation output. Finally, the outputs of the text branch and the graph branch are connected and input to the MLP layer to obtain the final prediction label;

[0030] S5.2: The input of the Text-SAGPool model is the preprocessed text embedding vector User characteristics and dissemination of structural information Where N represents the sum of rumor events and non-rumor events in the dataset, and M represents the number of all users in the dataset. At the same time, rumor events and non-rumor events in the dataset need to be divided into training and test sets in a ratio of 8:2.

[0031] S5.3: Use the training set to train the Text-SAGPoolg and Text-SAGPoolh models respectively. During the training process, the predicted labels output by the Text-SAGPoolg and Text-SAGPoolh models are compared with the true labels of the training set, and the Loss is calculated according to the following formula:

[0032]

[0033] According to the Loss obtained above, the parameters of the model are continuously optimized until the iteration is completed;

[0034] Preferably, in step S6:

[0035] S6.1: After completing the data preprocessing and model training of the model according to the above steps, a trained model can be obtained, and the classification results of the test events can be obtained using the model.

[0036] A system for implementing a rumor detection method based on comprehensive information of a propagation process of the present invention comprises a data acquisition module, a feature extraction module, a model training module and a model prediction module connected in sequence;

[0037] The data acquisition module acquires a public data set corresponding to the training model from a public website;

[0038] The data preprocessing module preprocesses the text information in the data set, obtains a text embedding matrix, obtains 8 user features from the user profile in the data set to represent the user, generates an event propagation graph for each event in the data set, and constructs a joint graph based on the event propagation graph;

[0039] The model training module inputs the preprocessed data into the Text-SAGPool model, and the text embedding vector obtains the text representation of the corresponding event through the text branch, and the graph structure and user features obtain the graph representation of the corresponding event through the graph branch. The text representation and the graph representation are connected together to obtain the event representation, and the final classification result is obtained through the MLP layer. The model is trained with the purpose of reducing the error value between the classification result and the actual result through the cross entropy loss function. After multiple iterations, the model training is saved until it reaches stability;

[0040] The model prediction module uses the known event data set as the training sample of the model and the untrained data set as the test sample to evaluate the classification and detection performance of the model for news events. This module can also be directly used in practical applications.

[0041] The present invention utilizes the existing public rumor dataset and uses the comprehensive information of the propagation process of the propagation events on social media to realize the accurate and rapid identification of rumor events on social media. By creatively using the CNN module and the SAGPool module, the rumor recognition performance is greatly improved, and the detection of rumor information on social media and the practicality and efficiency in unknown events are better realized.

[0042] The beneficial effects of the present invention are:

[0043] (1) This paper proposes for the first time a rumor detection model Text-SAGPool based on a comprehensive and sophisticated propagation process. This model can identify rumors by using the text content, propagation structure information, and user information when the information is propagated on social media. It can accurately and quickly identify rumors on social media.

[0044] (2) The present invention is compared with many advanced methods on three existing public datasets of different sizes. Our system shows good and stable performance on all three datasets. In addition, in early detection experiments, our method is able to accurately identify rumors at an early stage.

[0045] (3) Through a series of ablation experiments, the present invention proves that the text content, propagation structure information and user information involved in the propagation process all have a significant effect on improving the rumor recognition performance.

[0046] (4) Finally, we also performed rumor identification on unknown events on social media to highlight the stable performance of our model in real application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a structural flow chart of the present invention.

[0048] Figure 2 It is a system module connection diagram of the present invention.

[0049] Figure 3 It is the overall structure of the network constituted by the model. DETAILED DESCRIPTION

[0050] In order to overcome the defects of low recognition accuracy and inability to accurately identify rumors in the early stage of traditional rumor detection systems using a single type of information to identify rumors, the present invention develops a new rumor detection method and system based on comprehensive information of the propagation process. The system uses comprehensive information involved in the process of information propagation on social media to identify rumors, and extracts the text content, propagation structure and user information of the information respectively to achieve comprehensive modeling of the information propagation process, and uses a graph branch module and a text branch module to capture propagation structure features, user features and text representation respectively. The SAGPool model is used in the graph branch to comprehensively utilize propagation structure features and user features, and the text branch uses a convolutional neural network to capture text representation. Finally, the connection of the outputs of the two branches is realized through a fully connected layer, and then input to the softmax layer to give a predicted label, which greatly improves the detection and recognition of rumors on social media and the applicability and efficiency in unknown information.

[0051] like Figure 1-Figure 3 As shown, the technical solution adopted by the present invention to achieve the above-mentioned invention object is as follows:

[0052] A rumor detection method based on comprehensive information of a propagation process, characterized in that it comprises the following steps:

[0053] S1: Data acquisition: Obtain public rumor datasets from social media platforms. Each rumor dataset contains comprehensive information on the spread of several rumor events and non-rumor events. Each rumor and non-rumor event contains a source post and several reply posts. The comprehensive information includes spread structure information, text information, and user information. In addition, each event has a classification label corresponding to whether the event is a rumor.

[0054] S2: Construction of event propagation graph and joint graph: Analyze each rumor event or non-rumor event in the dataset, extract its propagation structure information, and model it into an event propagation graph. The root node in the event propagation graph represents the source post, and the other nodes represent the users involved in the event propagation. The edges between the nodes represent the social behavior between the users. All event propagation graphs in the dataset are connected together to form a joint graph of the entire dataset.

[0055] S3: Text preprocessing: Both the source posts and the reply posts contain text content of different lengths. The text content of each rumor event and non-rumor event in the dataset is integrated into a segment. The segment is cleaned using natural language processing tools, and the word2vec tool is used to obtain the text embedding vector of the segment.

[0056] S4: User feature extraction: For each user participating in event propagation, 8 user features are used to characterize the user;

[0057] S5: Model training: Input the joint graph, text embedding vector and user features obtained in S2, S3 and S4 into the designed rumor detection model: Text-SAGPool, use the CNN module in the model to extract text representation, and input the user features as the node features of each node in the joint graph into the SAGPool module in the model to obtain the graph representation of each rumor event or non-rumor event, combine the above-obtained text representation and graph representation together, and finally pass through the MLP layer in the model to obtain the prediction result of the model, calculate the loss of the prediction result and the event label, and use the loss to adjust the model parameters until the training model is obtained after the iteration is completed;

[0058] S6: Model prediction: Process new propagation events into the format of input data and input them into the trained model to predict unknown events.

[0059] In step S2:

[0060] S2.1: The dataset contains rumor events and non-rumor events, each of which is a real propagation event in the Internet social media. The publisher of the source post is taken as the root node of the event propagation graph, and the users who participate in the propagation of the source post are generated as other nodes in the propagation graph according to their social behavior actions and time sequence. The social behavior actions between users are used as the edges between nodes to obtain the event propagation graph. The social behavior actions between users involved include comments, replies, and forwarding, which are uniformly regarded as reply posts to the source posts;

[0061] S2.2: The event propagation graph constructed by S2.1 above is further combined into a joint graph. The joint graph is composed of the event propagation graphs of each event in the data set. There are no edges between the event propagation graphs. The propagation structure of each event in the data set can be comprehensively represented by the joint graph. The joint graph can be regarded as a collection of event propagation graphs of all rumor events and non-rumor events in the data set.

[0062] S2.3: In this method, the adjacency matrix is ​​obtained from the joint graph obtained above as the input of the subsequent model.

[0063] In step S3:

[0064] S3.1: Each rumor event and non-rumor event in the dataset contains a different number of posts, and each post contains text information of different lengths. For each rumor event or non-rumor event, all sentences in the event are connected to obtain a segment, which contains all the text information of the entire event;

[0065] S3.2: Use the natural language processing tool jieba to perform word segmentation on the above-obtained paragraph. After word segmentation according to the professional mode, a list of all the words in the paragraph can be obtained;

[0066] S3.3: Remove stop words from the word list after word segmentation, use the stop word library of Harbin Institute of Technology to filter, remove all words in the word list that are in the stop word library, and obtain a word list without stop words;

[0067] S3.4: Perform word frequency statistics on the words in the word list without stop words as follows:

[0068]

[0069] After calculating the frequency of each word in the word list excluding stop words, sort them from high to low according to the frequency, take the top 200 words for word embedding operation, and use the pre-trained Word2Vec model developed by TencentAI lab to obtain the word embedding vectors of these 200 words.

[0070] In step S4:

[0071] S4.1: Extract user information from the dataset. The dataset contains user information for each post and its reply posts. The eight user features shown in Table 1 are extracted from the dataset to describe the user.

[0072] Table 1 User characteristics

[0073] serial number name describe type 1 Verified Is the account verified? Boolean 2 User_description Whether the account has a user description Boolean 3 User_geo_enabled Is the account's geographic location verified? Boolean 4 Friends_count The number of friends following the account Numeric 5 Followers_count Number of followers of the account Numeric 6 Bi_followers_count Number of two-way followers of the account Numeric 7 Username_length Account nickname length Numeric 8 User_description_length The length of the user description of the account Numeric

[0074] S4.2: After extracting the eight user features, the non-Boolean features in the features need to be normalized by min-max as shown below:

[0075]

[0076] In the formula, x' represents the normalized value of a single data set, min is the minimum value in the data column, and max represents the maximum value in the data column.

[0077] In step S5:

[0078] S5.1: First construct the Text-SAGPool model, such as Figure 3 As shown, there are two variants of Text-SAGPool, namely Text-SAGPoolg and Text-SAGPoolh models. Both variant models are composed of two branches, namely text branch and graph branch. The text branches are the same, and are composed of three CNN modules with kernel sizes of 3, 4, and 5 and three maximum pooling layers connected in sequence. For the Text-SAGPoolg model, three graph convolutional layers connected in sequence are used to perform three graph convolution operations, and a graph pooling layer is used to perform a graph pooling operation, and the output of the last graph pooling layer is used as the final graph representation output. For the Text-SAGPoolh model, its graph branch contains three layers of graph convolutional layers and three layers of graph pooling layers, and after each graph pooling operation is completed, it will be input to the next graph convolution layer for graph convolution operation. Finally, the output of the three graph pooling layers is used as the final graph representation output. Finally, the outputs of the text branch and the graph branch are connected and input to the MLP layer to obtain the final prediction label;

[0079] S5.2: The input of the Text-SAGPool model is the preprocessed text embedding vector User characteristics and dissemination of structural information Where N represents the sum of rumor events and non-rumor events in the dataset, and M represents the number of all users in the dataset. At the same time, rumor events and non-rumor events in the dataset need to be divided into training and test sets in a ratio of 8:2.

[0080] S5.3: Use the training set to train the Text-SAGPoolg and Text-SAGPoolh models respectively. During the training process, the predicted labels output by the Text-SAGPoolg and Text-SAGPoolh models are compared with the true labels of the training set, and the Loss is calculated according to the following formula:

[0081]

[0082] According to the Loss obtained above, the parameters of the model are continuously optimized until the iteration is completed;

[0083] In step S6:

[0084] S6.1: After completing the data preprocessing and model training of the model according to the above steps, a trained model can be obtained, and the classification results of the test events can be obtained using the model. Specifically, the specific parameters of the model training are shown in Table 2;

[0085] Table 2 Model training parameters

[0086] serial number Variable Name describe Numeric 1 NUM_CLASSES Determine the number of categories 2 2 EPOCHS Iteration rounds 150 3 HIDDEN_DIM Hidden layer size 32 4 LEARNING_RATE Learning Rate 0.01 5 WEIGHT_DECAY Weight decay rate 0.0001 6 FILTER_NUM Number of convolution kernels 32

[0087] A system for implementing a rumor detection method based on comprehensive information of a propagation process of the present invention comprises: a data acquisition module, a feature extraction module, a model training module and a model prediction module;

[0088] The data acquisition module acquires a public data set corresponding to the training model from a public website;

[0089] The data preprocessing module preprocesses the text information in the data set, obtains the text embedding matrix, obtains 8 user features from the user profile in the data set to represent the user, generates an event propagation graph for each event in the data set, and constructs a joint graph based on the event propagation graph, specifically including:

[0090] S2.1: The rumor events and non-rumor events are all real propagation events in Internet social media. The publisher of the source post is used as the root node of the event propagation graph. The users who participated in the propagation of the source post are generated as other nodes in the propagation graph according to their social behavior actions and time sequence. The social behavior actions between users are used as the edges between nodes to obtain the event propagation graph.

[0091] S2.2: The event propagation graph constructed by S2.1 above is further combined into a joint graph. The joint graph is composed of the event propagation graphs of each event in the data set. There are no edges between the event propagation graphs. The propagation structure of each event in the data set can be comprehensively represented by the joint graph. The joint graph can be regarded as a collection of event propagation graphs of all rumor events and non-rumor events in the data set.

[0092] S2.3: In this method, the adjacency matrix is ​​obtained from the joint graph obtained above as the input of the subsequent model.

[0093] S3.1: Each rumor event and non-rumor event in the dataset contains a different number of posts, and each post contains text information of different lengths. For each rumor event or non-rumor event, all sentences in the event are connected to obtain a segment, which contains all the text information of the entire event;

[0094] S3.2: Perform word segmentation on the above-obtained paragraph, and after word segmentation, a list of all the words in the paragraph can be obtained;

[0095] S3.3: Remove stop words from the word list after word segmentation, use the existing stop word library to filter, remove all words in the word list that are in the stop word library, and obtain a word list without stop words;

[0096] S3.4: Count the word frequencies in the word list excluding stop words, and sort the words by their frequency of occurrence. Extract the top 200 words from the sorted word list and input them into the trained word2vec model to obtain the text embedding vectors of these words.

[0097] S4.1: Extract user information from the dataset. The dataset contains user information for each post and its reply posts. The following eight user features are extracted from the dataset to describe the user: whether the account is verified, whether the account has description text content, whether the account geographic location is verified, the number of account friends, the number of fans, the number of accounts followed by both parties, the length of the account nickname, and the length of the account description text;

[0098] S4.2: After extracting the eight user features, the non-Boolean features in the features need to be normalized by min-max as shown below:

[0099]

[0100] In the formula, x' represents the normalized value of a single data set, min is the minimum value in the data column, and max represents the maximum value in the data column.

[0101] The model training module inputs the preprocessed data into the Text-SAGPool model, and the text embedding vector obtains the text representation of the corresponding event through the text branch, and the graph structure and user features obtain the graph representation of the corresponding event through the graph branch. The text representation and the graph representation are connected together to obtain the event representation, and the final classification result is obtained through the MLP layer. The model is trained with the purpose of reducing the error value between the classification result and the actual result through the cross entropy loss function. After multiple iterations, the model training is stable and the model is saved, which specifically includes:

[0102] S5.1: First, construct the Text-SAGPool model. There are two variants of Text-SAGPool, namely Text-SAGPoolg and Text-SAGPoolh models. Both variant models are composed of two branches, namely text branch and graph branch. The text branches are the same, and are composed of three CNN modules with kernel sizes of 3, 4, and 5 and three maximum pooling layers connected in sequence. For the Text-SAGPoolg model, three graph convolutional layers connected in sequence are used to perform three graph convolution operations, and a graph pooling layer is used to perform a graph pooling operation. The output of the last graph pooling layer is used as the final graph representation output. For the Text-SAGPoolh model, its graph branch contains three layers of graph convolutional layers and three layers of graph pooling layers, and after each graph pooling operation is completed, it will be input to the next graph convolution layer for graph convolution operation. Finally, the output of the three graph pooling layers is used as the final graph representation output. Finally, the outputs of the text branch and the graph branch are connected and input to the MLP layer to obtain the final prediction label;

[0103] S5.2: The input of the Text-SAGPool model is the preprocessed text embedding vector User characteristics and dissemination of structural information Where N represents the sum of rumor events and non-rumor events in the dataset, and M represents the number of all users in the dataset. At the same time, rumor events and non-rumor events in the dataset need to be divided into training and test sets in a ratio of 8:2.

[0104] S5.3: Use the training set to train the Text-SAGPoolg and Text-SAGPoolh models respectively. During the training process, the predicted labels output by the Text-SAGPoolg and Text-SAGPoolh models are compared with the true labels of the training set, and the Loss is calculated according to the following formula:

[0105]

[0106] According to the Loss obtained above, the parameters of the model are continuously optimized until the iteration is completed;

[0107] The model prediction module uses rumor events and non-rumor events divided into test samples in the data set as test samples to evaluate the model's detection performance for unknown events. This module can also be directly used in practical applications, specifically including:

[0108] S6.1: After completing the data preprocessing and model training of the model according to the above steps, a trained model can be obtained, and the classification results of the test events can be obtained using the model.

[0109] S6.2: Compare the predicted labels with the true labels, calculate the accuracy, precision, recall, and evaluation methods.

[0110] The data acquisition module, feature extraction module, model training module and model prediction module are connected in sequence.

[0111] As mentioned above, the rumor datasets used in the present invention are all from mainstream social media (Weibo and Twitter) at home and abroad. The datasets in the datasets are input into the system for iterative training to obtain a model with excellent accuracy, precision, recall and other indicators for the experimental set, validation set and test set. As shown in Table 3:

[0112] Table 3 Rumor detection results of two variant models of Text-SAGPool on different datasets

[0113]

[0114] After using our system to detect rumors on three different real rumor datasets, it can be seen that the method of the present invention has almost excellent performance on the three datasets. It can achieve accurate recognition and detection of news events on mainstream social media platforms at home and abroad, which further proves the superior performance and practicality of our invention.

[0115] The contents described in the embodiments of this specification are merely an enumeration of the implementation forms of the inventive concept. The protection scope of the present invention should not be regarded as limited to the specific forms described in the embodiments. The protection scope of the present invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. A rumor detection method based on comprehensive information of the propagation process, characterized by: The following steps are involved: S1: Data acquisition: Obtain public rumor datasets from social media platforms. Each rumor dataset contains comprehensive information on the spread of several rumor events and non-rumor events. Each rumor and non-rumor event contains a source post and several reply posts. The comprehensive information includes spread structure information, text information, and user information. In addition, each event has a classification label corresponding to whether the event is a rumor. S2: Construction of event propagation graph and joint graph: Analyze each rumor event or non-rumor event in the dataset, extract its propagation structure information, and model it into an event propagation graph. The root node in the event propagation graph represents the source post, and the other nodes represent the users involved in the event propagation. The edges between the nodes represent the social behavior between the users. All event propagation graphs in the dataset are connected together to form a joint graph of the entire dataset. S3: Text preprocessing: Both the source posts and the reply posts contain text content of different lengths. The text content of each rumor event and non-rumor event in the dataset is integrated into a segment. The segment is cleaned using natural language processing tools, and the word2vec tool is used to obtain the text embedding vector of the segment. S4: User feature extraction: For each user participating in event propagation, 8 user features are used to characterize the user; S5: Model training: Input the joint graph, text embedding vector and user features obtained by S2, S3 and S4 into the designed rumor detection model: Text-SAGPool, use the CNN module in the model to extract text representation, and input the user features as the node features of each node in the joint graph into the SAGPool module in the model to obtain the graph representation of each rumor event or non-rumor event, combine the above-obtained text representation and graph representation together, and finally pass the MLP layer in the model to obtain the prediction result of the model, calculate the loss of the prediction result and the event label, and use the loss to adjust the model parameters until the trained model is obtained after the iteration is completed; specifically including: S5.1: First, construct the Text-SAGPool model. There are two variants of Text-SAGPool, namely Text-SAGPoolg and Text-SAGPoolh models. Both variant models are composed of two branches, namely text branch and graph branch. The text branches are the same, and are composed of three CNN modules with kernel sizes of 3, 4, and 5 and three maximum pooling layers connected in sequence. For the Text-SAGPoolg model, three graph convolutional layers connected in sequence are used to perform three graph convolution operations, and a graph pooling layer is used to perform a graph pooling operation. The output of the last graph pooling layer is used as the final graph representation output. For the Text-SAGPoolh model, its graph branch contains three layers of graph convolutional layers and three layers of graph pooling layers, and after each graph pooling operation is completed, it will be input to the next graph convolution layer for graph convolution operation. Finally, the output of the three graph pooling layers is used as the final graph representation output. Finally, the outputs of the text branch and the graph branch are connected and input into the MLP layer to obtain the final prediction label. S5.2: The input of the Text-SAGPool model is the preprocessed text embedding vector User characteristics and dissemination of structural information Where N represents the sum of rumor events and non-rumor events in the dataset, and M represents the number of all users in the dataset. At the same time, rumor events and non-rumor events in the dataset need to be divided into training and test sets in a ratio of 8:

2. S5.3: Use the training set to train the Text-SAGPoolg and Text-SAGPoolh models respectively. During the training process, the predicted labels output by the Text-SAGPoolg and Text-SAGPoolh models are compared with the true labels of the training set, and the Loss is calculated according to the following formula: According to the Loss obtained above, the parameters of the model are continuously optimized until the iteration is completed; S5.4: Complete training and save the optimal model; S6: Model prediction: Process new propagation events into the format of input data and input them into the trained model to predict unknown events.

2. A rumor detection method based on comprehensive information of propagation process as claimed in claim 1, characterized in that: The step S2 specifically includes: S2.1: The rumor events and non-rumor events are all real propagation events in Internet social media. The publisher of the source post is used as the root node of the event propagation graph. The users who participated in the propagation of the source post are generated as other nodes in the propagation graph according to their social behavior actions and time sequence. The social behavior actions between users are used as the edges between nodes to obtain the event propagation graph. S2.2: The event propagation graph constructed by S2.1 above is further combined into a joint graph. The joint graph is composed of the event propagation graphs of each event in the data set. There are no edges between the event propagation graphs. The propagation structure of each event in the data set is comprehensively represented by the joint graph. The joint graph is regarded as a collection of event propagation graphs of all rumor events and non-rumor events in the data set. S2.3: In this method, the adjacency matrix is ​​obtained from the joint graph obtained above as the input of the subsequent model.

3. A rumor detection method based on comprehensive information of propagation process as claimed in claim 1, characterized in that: The step S3 specifically includes: S3.1: Each rumor event and non-rumor event in the dataset contains a different number of posts, and each post contains text information of different lengths. For each rumor event or non-rumor event, all sentences in the event are connected to obtain a segment, which contains all the text information of the entire event; S3.2: Perform word segmentation on the above-obtained paragraph, and obtain a list of all words in the paragraph after word segmentation; S3.3: Remove stop words from the word list after word segmentation, use the existing stop word library to filter, remove all words in the word list that are in the stop word library, and obtain a word list without stop words; S3.4: Count the word frequencies in the word list excluding stop words, and sort the words by their frequency of occurrence. Extract the top 200 words from the sorted word list and input them into the trained word2vec model to obtain the text embedding vectors of these words.

4. A rumor detection method based on comprehensive information of propagation process as claimed in claim 1, characterized in that: The step S4 specifically includes: S4.1: Extract user information from the dataset. The dataset contains user information for each post and its reply posts. The following eight user features are extracted from the dataset to describe the user: whether the account is verified, whether the account has description text content, whether the account geographic location is verified, the number of account friends, the number of fans, the number of accounts followed by both parties, the length of the account nickname, and the length of the account description text; S4.2: After extracting the eight user features, the non-Boolean features in the features need to be normalized by min-max as shown below: In the formula, x ′ Represents the normalized value of a single data set. Min is the minimum value in the data column, and max is the maximum value in the data column.

5. A rumor detection method based on comprehensive information of propagation process as claimed in claim 1, characterized in that: The step S6 specifically includes: S6.1: After completing the data preprocessing and model training of the model according to the above steps, a trained model is obtained, and the classification result of the test event is obtained using the model; S6.2: Compare the predicted labels with the true labels, calculate the accuracy, precision, recall, and evaluation methods.

6. A system for implementing the rumor detection method based on comprehensive information of the propagation process as claimed in claim 1, characterized in that: It includes a data acquisition module, a feature extraction module, a model training module and a model prediction module which are connected in sequence; The data acquisition module acquires a public data set corresponding to the training model from a public website; The data preprocessing module preprocesses the text information in the data set, obtains a text embedding matrix, obtains 8 user features from the user profile in the data set to represent the user, generates an event propagation graph for each event in the data set, and constructs a joint graph based on the event propagation graph; The model training module inputs the preprocessed data into the Text-SAGPool model, and the text embedding vector obtains the text representation of the corresponding event through the text branch, and the graph structure and user features obtain the graph representation of the corresponding event through the graph branch. The text representation and the graph representation are connected together to obtain the event representation, and the final classification result is obtained through the MLP layer. The model is trained with the purpose of reducing the error value between the classification result and the actual result through the cross entropy loss function. After multiple iterations, the model training is saved until it reaches stability; The model prediction module uses the known event data set as the training sample of the model and the untrained data set as the test sample to evaluate the classification and detection performance of the model for news events. This module can also be directly used in practical applications.

Images