A robust rumor detection method on social media

Abstract

The present invention relates to the field of electronic digital data processing technology, and in particular to a robust rumor detection method on social media, comprising the following steps: constructing events in social media data as static propagation graphs, and constructing event and label sample pairs to form a rumor detection data set; constructing a rumor dynamic propagation graph and a dynamic augmented graph containing multi-dimensional fusion information; obtaining a rumor dynamic propagation graph representation and a dynamic augmented graph representation; unsupervised comparative learning to capture the representation information of the rumor dynamic propagation graph and the representation information of the dynamic augmented graph; and performing rumor classification through a deep neural classification model to complete robust rumor detection on social media. The method provided by the present invention can make good use of the temporal information of the dynamic propagation graph and the dynamic augmented graph, and can effectively use comparative learning to obtain graph representation, thereby better completing the rumor detection task on social media.

Description

Technical Field

[0001] The present invention relates to the technical field of electronic digital data processing, and in particular to a robust rumor detection method on social media. Background Art

[0002] Social media has become the main platform for people to obtain news, express opinions and participate in social interactions. However, social media is also a double-edged sword. It has the risk of rumors spreading rapidly and posing a serious threat to social stability, public trust and economic order. Therefore, it is urgent to effectively detect rumors on social media.

[0003] Traditional detection methods rely on hand-crafted features and machine learning classifiers to identify rumors, but these methods are limited by the subjectivity and constraints of hand-crafted features. The development of deep learning has promoted rumor detection from text content, users, and propagation structures. Methods based on propagation structures have achieved excellent detection performance, among which graph neural networks have been proposed to model propagation graphs for rumor detection. However, existing methods only consider the static propagation structure of the final state of the rumor, while ignoring the temporal dynamics of the propagation graph. At the same time, in the actual propagation process of social media messages, there are some missing and unreliable interactions in the propagation structure, and the perturbation of the propagation structure will affect the performance of propagation-based methods. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a robust rumor detection method on social media, which can dynamically model the spread of rumors in social media, obtain two information representations: a dynamic propagation graph of rumors and a dynamic augmented graph, and improve the accuracy and efficiency of rumor detection by fusing multi-dimensional information.

[0005] The present invention is achieved through the following technical solutions:

[0006] A robust rumor detection method on social media comprises the following steps:

[0007] S1: Obtain social media data, construct events in the social media data into static propagation graphs, and construct event and label sample pairs to form a rumor detection dataset;

[0008] S2: Based on the static propagation graph in the rumor detection dataset, a dynamic rumor propagation graph and a dynamic augmented graph containing multi-dimensional fusion information are constructed;

[0009] S3: The rumor dynamic propagation graph and dynamic augmented graph containing multi-dimensional fusion information are used as the input of the graph encoder and encoded, and finally the rumor dynamic propagation graph representation and dynamic augmented graph representation are obtained;

[0010] S4: Unsupervised comparative learning is performed based on the rumor dynamic propagation graph representation and the dynamic augmented graph representation to capture the representation information of the rumor dynamic propagation graph and the dynamic augmented graph;

[0011] S5: Integrate the representation information of the rumor dynamic propagation graph and the dynamic augmented graph, classify rumors through a deep neural classification model, and complete robust rumor detection on social media.

[0012] Furthermore, when events in social media data are constructed as a static propagation graph in step S1, the nodes of the static propagation graph are composed of the individual posts that constitute each event, and the edges of the static propagation graph are composed of the forwarding or comment relationships between the posts and their response posts.

[0013] Furthermore, when events in social media data are constructed as a static propagation graph in step S1, each constructed static propagation graph is a connected graph. If the static propagation graph is not connected, the maximum connected graph is taken as the static propagation graph.

[0014] Optimized, when constructing events in social media data as static propagation graphs, events with less than 3 comments in social media data are eliminated.

[0015] Optimized, in step S2, a time snapshot construction method is adopted to construct a rumor dynamic propagation graph containing multi-dimensional fusion information.

[0016] Furthermore, the method for constructing a rumor dynamic propagation graph containing multi-dimensional fusion information in step S2 is as follows:

[0017] S211: Divide each event into multiple time snapshots equally according to the time span to obtain a dynamic propagation graph of rumors;

[0018] S212: Take the source post of the event as the root node of the time snapshot, form an edge for each response relationship, establish the corresponding propagation graph, model the propagation structure of the rumor dynamic propagation graph, and thus construct a rumor dynamic propagation graph containing multi-dimensional fusion information.

[0019] Furthermore, the method for constructing a dynamic augmented graph containing multi-dimensional fusion information in step S2 is as follows:

[0020] S221: Augmenting the time snapshot of each dynamic propagation graph using a node mask to obtain a corresponding dynamic augmented graph;

[0021] S222: As time passes, masking operations are performed on nodes of the time snapshot mask of the previous dynamic augmented graph in the time snapshot of the next dynamic augmented graph to construct a dynamic augmented graph containing multi-dimensional fusion information.

[0022] Further, in step S3, the rumor dynamic propagation graph representation and the dynamic augmented graph representation are obtained according to the following method:

[0023] S31: For a given rumor propagation dynamic graph, the hidden feature vector of the graph encoder is updated and encoded using the aggregation function (1):

[0024] (1);

[0025] in: The representation graph encoder is Layer Node The hidden feature vector of Indicates that the multi-layer perceptron aggregates and updates node features. represents the learning parameters, The representation graph encoder is The learning parameters of the layer, The representation graph encoder is Layer Node The hidden feature vector of The representation graph encoder is Layer Node The hidden feature vector of Representation graph encoder node Neighbor nodes of

[0026] S32: The image encoder obtains the first The propagation graph of the time snapshot is represented by the formula (3). The augmented graph representation of time snapshots is:

[0027] (2);

[0028] (3);

[0029] in: Indicates The node representation of the propagation graph of a time snapshot is: represents the adjacency matrix, Indicates The propagation graph feature matrix of time snapshots, Indicates The propagation graph representation of a time snapshot is It means obtaining the node representation of the graph encoding for the content in the brackets. It means taking the mean of the contents in brackets to form a graph. Indicates The node representation of the augmented graph of time snapshots, Indicates The augmented graph feature matrix of time snapshots, Indicates A graphical representation of the augmented graph of a time snapshot;

[0030] S33: According to The propagation graph representation of each time snapshot obtains the dynamic propagation graph representation sequence of the same event. The graph representation of the augmented graph of each time snapshot obtains a sequence of dynamic augmented graph representations of the same event;

[0031] S34: A bidirectional long short-term memory model is used to model the forward and backward dynamic propagation graph representation sequences and dynamic augmentation graph representation sequences, capturing long-term dependencies and identifying potential patterns of the dynamic propagation graph and dynamic augmentation graph;

[0032] S35: Connect the forward state and the backward state in the latent mode of the dynamic propagation graph and the dynamic augmentation graph to obtain the graph representation of the bidirectional long short-term memory model, and then obtain the batch rumor dynamic propagation graph representation and the batch dynamic augmentation graph representation.

[0033] Furthermore, in step S4, the following method is used to perform unsupervised comparative learning based on the rumor dynamic propagation graph representation and the dynamic augmented graph representation to capture the representation information of the rumor dynamic propagation graph and the representation information of the dynamic augmented graph:

[0034] S41: Using unsupervised contrastive learning, the similarity matrix in the batch rumor dynamic propagation graph representation and the dynamic augmented graph representation is calculated according to formula (4) to capture the similarity between the two vector representations:

[0035] (4);

[0036] in: Represents the similarity matrix in the batch rumor dynamic propagation graph representation and the dynamic augmented graph representation, Represents the dynamic propagation graph of batch rumor processing, represents the batch dynamically augmented graph representation, represents an exponential function with a natural constant as base, represents the cosine similarity function, represents the temperature coefficient;

[0037] S42: Using the similarity between the two vector representations, the unsupervised contrast loss is calculated according to formula (5);

[0038] (5);

[0039] in: represents the unsupervised contrast loss, represents the identity matrix, Represents the cross entropy loss of calculating the similarity matrix;

[0040] S43: The unsupervised contrast loss and the cross entropy loss are added to train the model to capture the representation information of the rumor dynamic propagation graph and the dynamic augmented graph during the training process.

[0041] Furthermore, the following method is used to fuse the representation information of the rumor dynamic propagation graph and the representation information of the dynamic augmented graph, and the rumor classification is performed through a deep neural classification model:

[0042] S51: Connect the representation information of the captured rumor dynamic propagation graph and the representation information of the dynamic augmented graph to form integrated information, and then input the integrated information into the fully connected layer and activation function layer of the deep neural classification model for training, and calculate the predicted probability distribution according to formula (6):

[0043] (6);

[0044] in: represents the predicted probability distribution, represents the activation function, represents the weight, Indicates integrated information. Indicates bias;

[0045] S52: Calculate the classification cross entropy loss function during training according to formula (7):

[0046] (7);

[0047] in: represents the classification cross entropy loss, represents the number of rumor events, Indicates one of the rumor events in the rumor event. Indicates the type of rumor. Indicates the total number of rumor categories, Indicates the label value of the rumor event, represents the probability distribution of rumor event prediction;

[0048] S53: Calculate the loss of the deep neural classification model according to formula (8), and train the deep neural classification model to classify rumors and complete robust rumor detection on social media:

[0049] (8);

[0050] in: represents the loss of the deep neural classification model, Represents the hyperparameter used to balance the loss function.

[0051] Beneficial effects of the invention:

[0052] The present invention provides a robust rumor detection method on social media, which integrates the dynamic propagation method and dynamic augmentation method of generating multi-dimensional information, and designs unsupervised comparative learning to obtain high-quality representations of the dynamic propagation graph and dynamic augmentation graph of rumors of the same event, thereby more effectively performing rumor detection and improving the robustness and effectiveness of the rumor detection task. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is a schematic diagram of the process of the present invention.

[0054] Figure 2 It is a diagram of experimental results of different dynamic graph construction methods of the present invention.

[0055] Figure 3 It is a graph of experimental results of different augmentation methods of the present invention. DETAILED DESCRIPTION

[0056] A robust rumor detection method on social media includes the following steps, and its flow chart is as follows: Figure 1 As shown:

[0057] S1: Obtain social media data, construct events in the social media data into static propagation graphs, and construct event and label sample pairs to form a rumor detection dataset;

[0058] Specifically, we can collect the datasets RumorEval and TWITTER as source data, extract the text content, propagation structure and authenticity labels of the posts to construct the rumor detection dataset. RumorEval and TWITTER are two currently available rumor detection datasets, and the samples in the datasets are all from the popular social media platform Twitter.

[0059] The specific statistical information of the data set is shown in Table 1:

[0060] Table 1

[0061]

[0062] Specifically, when events in social media data are constructed as a static diffusion graph, the nodes of the static diffusion graph are composed of the individual posts that constitute each event, and the edges of the static diffusion graph are composed of the forwarding or comment relationships between the posts and their response posts.

[0063] When events in social media data are constructed as static propagation graphs, each constructed static propagation graph is a connected graph. If the static propagation graph is not connected, the maximum connected graph is taken as the static propagation graph.

[0064] Specifically, when constructing events in social media data into a static propagation graph, events with less than 3 comments in the social media data may be preferably eliminated.

[0065] S2: Based on the static propagation graph in the rumor detection dataset, a dynamic rumor propagation graph and a dynamic augmented graph containing multi-dimensional fusion information are constructed;

[0066] In order to capture the dynamic patterns of propagation graphs, there are three main existing dynamic graph construction methods: time snapshot, sequential snapshot, and fixed time snapshot.

[0067] Time snapshots divide the entire propagation process of an event into multiple stages at the same time intervals, providing an intuitive simulation of the time evolution of rumor propagation.

[0068] Sequential snapshots structure each stage based on the number of posts to reflect the user’s engagement intensity.

[0069] Fixed-time snapshots divide the propagation into phases with uniform time intervals according to fine-grained time units to emphasize detailed temporal granularity while potentially ignoring post-dynamics.

[0070] Specifically, a time snapshot construction method can be used to construct a rumor dynamic propagation graph containing multi-dimensional fusion information.

[0071] The traditional time snapshot method focuses on capturing the propagation path of rumors that evolve over time in the network, and records the propagation status at different times through a series of time snapshots. However, this method often ignores the complexity and dynamics of the spatial structure dimension in the rumor propagation process. In order to overcome this limitation, this method innovatively reconstructs and optimizes the limitations of time snapshots, and introduces the spatial structure dimension of the augmented graph into the framework of the dynamic propagation graph. The multi-dimensional fusion of temporal information and spatial information not only retains the basic structure of nodes and edges in the original propagation graph, but also expands the data dimension of the graph by introducing additional node and edge relationships, while better simulating the uncertainty of the propagation process and improving the robustness of rumor detection experimental results.

[0072] The specific method of constructing a rumor dynamic propagation graph containing multi-dimensional fusion information is as follows:

[0073] S211: Divide each event into multiple time snapshots equally according to the time span to obtain a dynamic propagation graph of rumors;

[0074] S212: Take the source post of the event as the root node of the time snapshot, form an edge for each response relationship, establish the corresponding propagation graph, model the propagation structure of the rumor dynamic propagation graph, and thus construct a rumor dynamic propagation graph containing multi-dimensional fusion information.

[0075] By conducting comparative experiments on time snapshots, sequential snapshots, and fixed time snapshots, the evaluation indicators are the accuracy of the overall category and the F1 value of the rumor category. It can be found that on the RumorEval and TWITTER datasets, the accuracy of the dynamic graph construction method of time snapshots and the F1 value of the rumor category are higher than the other two methods.

[0076] The F1 in the table refers to the harmonic mean of precision and recall.

[0077] Therefore, the present invention selects time snapshot as the construction method of the dynamic propagation graph. The specific comparison results are shown in Table 2:

[0078] Table 2

[0079]

[0080] The specific method of constructing a dynamic augmented graph containing multi-dimensional fusion information is as follows:

[0081] S221: Augmenting the time snapshot of each dynamic propagation graph using a node mask to obtain a corresponding dynamic augmented graph;

[0082] The time snapshot feature matrix of the dynamic augmented graph is expressed as

[0083]

[0084] in: Indicates The augmented graph feature matrix of time snapshots, represents the mask matrix, , which contains randomly generated Row zero vector, represents the number of nodes in the rumor propagation graph, represents the embedding vector dimension of each node, represents the mask rate, represents element-wise multiplication, Indicates The propagation graph feature matrix of time snapshots, , Represents the field of real numbers.

[0085] S222: As time passes, masking operations are performed on nodes of the time snapshot mask of the previous dynamic augmented graph in the time snapshot of the next dynamic augmented graph to construct a dynamic augmented graph containing multi-dimensional fusion information.

[0086] Propagation-based rumor detection models often suffer from missing and unreliable interactions in the propagation structure. Graph augmentation methods can alleviate this problem and improve the robustness of the model by modifying node or edge attributes during training.

[0087] There are three common graph augmentation methods: node masking, edge dropping, and feature masking. Node masking randomly masks node features during training, allowing the model to learn robust representations without certain node information. Edge dropping randomly removes some edge connections during training, allowing the model to robustly predict nodes when some connection information is lost. Feature masking randomly masks node dimension features during training, allowing the model to maintain generalization even when feature information is incomplete.

[0088] By conducting comparative experiments on node masking, edge discarding, and feature masking, the evaluation indicators are the accuracy of the overall category and the F1 value of the rumor category. It can be found that on the RumorEval and TWITTER datasets, the accuracy of the node mask augmentation method and the F1 value of the rumor category are higher than the other two methods. Therefore, the present invention selects node masking as the construction method of the dynamic augmentation graph. The specific comparison results are shown in Table 3:

[0089] Table 3

[0090]

[0091] S3: The rumor dynamic propagation graph and dynamic augmented graph containing multi-dimensional fusion information are used as the input of the graph encoder and encoded, and finally the rumor dynamic propagation graph representation and dynamic augmented graph representation are obtained;

[0092] Specifically, the rumor dynamic propagation graph representation and dynamic augmented graph representation can be obtained according to the following method:

[0093] S31: For a given rumor propagation dynamic graph, the hidden feature vector of the graph encoder is updated and encoded using the aggregation function (1):

[0094] (1);

[0095] in: The representation graph encoder is Layer Node The hidden feature vector of Indicates that the multi-layer perceptron aggregates and updates node features. represents the learning parameters, The representation graph encoder is The learning parameters of the layer, The representation graph encoder is Layer Node The hidden feature vector of The representation graph encoder is Layer Node The hidden feature vector of Representation graph encoder node Neighbor nodes of

[0096] S32: The image encoder obtains the first The propagation graph of the time snapshot is represented by the formula (3). The augmented graph representation of time snapshots is:

[0097] (2);

[0098] (3);

[0099] in: Indicates The node representation of the propagation graph of a time snapshot is: represents the adjacency matrix, Indicates The propagation graph feature matrix of time snapshots, Indicates The propagation graph representation of a time snapshot is It means obtaining the node representation of the graph encoding for the content in the brackets. It means taking the mean of the contents in brackets to form a graph. Indicates The node representation of the augmented graph of time snapshots, Indicates The augmented graph feature matrix of time snapshots, Indicates A graphical representation of the augmented graph of a time snapshot;

[0100] S33: According to The propagation graph representation of each time snapshot obtains the dynamic propagation graph representation sequence of the same event. The graph representation of the augmented graph of each time snapshot obtains a sequence of dynamic augmented graph representations of the same event;

[0101] S34: A bidirectional long short-term memory model is used to model the forward and backward dynamic propagation graph representation sequences and dynamic augmentation graph representation sequences, capturing long-term dependencies and identifying potential patterns of the dynamic propagation graph and dynamic augmentation graph;

[0102] The bidirectional long short-term memory model is used to model the forward and backward dynamic propagation graph representation sequences and dynamic augmentation graph representation sequences. The process of capturing long-term dependencies and identifying the potential patterns of dynamic propagation graphs and dynamic augmentation graphs is described as follows:

[0103]

[0104] in: represents the forward hidden state of the dynamic propagation graph, A graph representation representing a dynamic propagation graph, represents the forward long short-term memory network, represents the backward hidden state of the dynamic propagation graph, represents the backward long short-term memory network, represents the forward hidden state of the dynamic augmented graph, Graph representation representing the dynamically augmented graph, Represents the backward hidden state of the dynamic augmented graph;

[0105] S35: Connect the forward state and the backward state in the latent mode of the dynamic propagation graph and the dynamic augmentation graph to obtain the graph representation of the bidirectional long short-term memory model, and then obtain the batch rumor dynamic propagation graph representation and the batch dynamic augmentation graph representation.

[0106] S4: Unsupervised comparative learning is performed based on the rumor dynamic propagation graph representation and the dynamic augmented graph representation to capture the representation information of the rumor dynamic propagation graph and the dynamic augmented graph;

[0107] Specifically, the following method can be used to perform unsupervised comparative learning based on the rumor dynamic propagation graph representation and the dynamic augmented graph representation to capture the representation information of the rumor dynamic propagation graph and the dynamic augmented graph representation information:

[0108] S41: Using unsupervised contrastive learning, the similarity matrix in the batch rumor dynamic propagation graph representation and the dynamic augmented graph representation is calculated according to formula (4) to capture the similarity between the two vector representations:

[0109] (4);

[0110] in: Represents the similarity matrix in the batch rumor dynamic propagation graph representation and the dynamic augmented graph representation, Represents the dynamic propagation graph of batch rumor processing, represents the batch dynamically augmented graph representation, represents an exponential function with a natural constant as base, represents the cosine similarity function, represents the temperature coefficient;

[0111] This method uses unsupervised contrastive learning. Given a propagation graph representation and a corresponding augmentation graph representation, there exists Possible representations of .in, For batch processing The propagation diagram shows that For batch processing The augmented graph shows that Represents the number of rumor events. of is a correct match between the propagation graph representation and augmented graph representation of the same event, while the rest is a wrong match, that is, in the unsupervised contrastive learning designed by the present invention, The propagation graph representation and augmented graph representation samples of the same event are positive samples, and the rest samples are negative samples. In the designed unsupervised comparison strategy, the goal is to maximize similarity of positive samples, while minimizing In order to calculate the similarity of negative samples. The similarity of vector pairs is constructed by To capture the similarity matrix in the batch rumor propagation dynamic graph representation and the dynamic augmented graph representation.

[0112] S42: Using the similarity between the two vector representations, the unsupervised contrast loss is calculated according to formula (5);

[0113] (5);

[0114] in: represents the unsupervised contrast loss, represents the identity matrix, , Represents the cross entropy loss of calculating the similarity matrix;

[0115] S43: The unsupervised contrast loss and the cross entropy loss are added to train the model to capture the representation information of the rumor dynamic propagation graph and the dynamic augmented graph during the training process.

[0116] Contrastive learning is divided into supervised contrastive learning and unsupervised contrastive learning. Supervised contrastive learning refers to contrastive learning performed when the data already has clear labels. In unsupervised contrastive learning, the model does not require any label information, but instead builds an effective data representation by learning the characteristics of the data itself. This learning method can effectively use a large amount of unlabeled data to improve the performance of the model. Through unsupervised contrastive learning, the model can learn the similarities and differences between data, thereby building a more effective and robust data representation in the feature space. This representation not only helps improve the performance of downstream tasks, but also reveals the potential structure and relationship between data.

[0117] S5: Integrate the representation information of the rumor dynamic propagation graph and the dynamic augmented graph, classify rumors through a deep neural classification model, and complete robust rumor detection on social media.

[0118] Specifically, the following method can be used to integrate the representation information of the rumor dynamic propagation graph and the representation information of the dynamic augmented graph, and classify rumors through a deep neural classification model:

[0119] S51: Connect the representation information of the captured rumor dynamic propagation graph and the representation information of the dynamic augmented graph to form integrated information, and then input the integrated information into the fully connected layer and activation function layer of the deep neural classification model for training, and calculate the predicted probability distribution according to formula (6):

[0120] (6);

[0121] in: represents the predicted probability distribution, represents the activation function, represents the weight, Indicates integrated information. Indicates bias;

[0122] S52: Calculate the classification cross entropy loss function during training according to formula (7):

[0123] (7);

[0124] in: represents the classification cross entropy loss, represents the number of rumor events, Indicates one of the rumor events in the rumor event. Indicates the type of rumor. Indicates the total number of rumor categories, Indicates the label value of the rumor event, represents the probability distribution of rumor event prediction;

[0125] S53: Calculate the loss of the deep neural classification model according to formula (8), and train the deep neural classification model to classify rumors and complete robust rumor detection on social media:

[0126] (8);

[0127] in: represents the loss of the deep neural classification model, Represents the hyperparameter used to balance the loss function.

[0128] By comparing the rumor detection method provided by the present invention with the rumor detection method commonly used in the prior art, it is found that the rumor detection method provided by the present invention is superior to the other baseline models in all indicators. The specific experimental results are shown in Table 4, where DCL represents the rumor detection method model provided by the present invention, and the rest are several rumor detection method models in the prior art:

[0129] Table 4

[0130]

[0131] In the table, F represents rumor and T represents non-rumor, which confirms the effectiveness of the temporal dynamics and contrastive learning of the propagation graph of the present invention. Models that only use propagation only capture spatial information, which makes them susceptible to noise and heavily dependent on a large amount of labeled data. Only using static propagation graphs to capture information cannot learn a robust representation of dynamic propagation graphs. In contrast, the rumor detection method provided by the present invention emphasizes the temporal dynamics of the propagation graph and enhances the ability to capture spatial and temporal information through contrastive learning of robust representations. In addition, contrastive learning can obtain high-quality feature representations and achieve excellent performance.

[0132] And through ablation experiments, it can be seen that the present invention performs better because of (1) the importance of spatial structural information of dynamic propagation graph; (2) the effectiveness of dynamic augmentation graph; (3) contrastive learning can effectively capture the intrinsic relationship in data features and obtain high-quality feature representation; (4) the temporal information of dynamic graph is very important for improving performance. The specific ablation experiment results are shown in Table 5:

[0133] Table 5

[0134]

[0135] The robustness of this method is explored by using early rumor detection and incremental mask rate experiments. The experimental results of different dynamic graph construction methods are shown in the figure. Figure 2 As shown in Figure 2, in the early rumor detection experiment, this method consistently achieves higher accuracy than other models at each deadline in the RumorEval and TWITTER datasets, and has the best robustness in early rumor detection. Figure 3 As shown in the figure, in the incremental mask rate experiment, the accuracy of this method decreases more slowly in both RumorEval and TWITTER datasets, and it always outperforms the comparison model. In addition, the accuracy change of this method is significantly smaller than that of other models, and it can maintain a high accuracy even at a higher masking rate. The experimental results show that the model proposed by this method is robust and can effectively alleviate the impact of missing data.

[0136] In summary, the robust rumor detection method on social media provided by the present invention can make good use of the temporal information of the dynamic propagation graph and the dynamic augmented graph, and can effectively use contrastive learning to obtain graph representation, thereby better completing the rumor detection task on social media.

[0137] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A robust rumor detection method on social media, characterized by: The steps include: S1: Obtain social media data, construct events in the social media data into static propagation graphs, and construct event and label sample pairs to form a rumor detection dataset; S2: Based on the static propagation graph in the rumor detection dataset, a dynamic rumor propagation graph and a dynamic augmented graph containing multi-dimensional fusion information are constructed; Among them, the method of constructing a rumor dynamic propagation graph containing multi-dimensional fusion information is as follows: S211: Divide each event equally into multiple time snapshots according to the time span to obtain a dynamic propagation graph of rumors; S212: The source post of the event is used as the root node of the time snapshot, each response relationship forms an edge, a corresponding propagation graph is established, and the propagation structure of the rumor dynamic propagation graph is modeled, thereby constructing a rumor dynamic propagation graph containing multi-dimensional fusion information; The method of constructing a dynamic augmented graph containing multi-dimensional fusion information is as follows: S221: Augment each time snapshot using a node mask to obtain a corresponding dynamic augmented graph; S222: as time passes, masking operations are performed on nodes of the time snapshot mask of the previous dynamic augmented graph in the time snapshot of the next dynamic augmented graph to construct a dynamic augmented graph containing multi-dimensional fusion information; S3: The rumor dynamic propagation graph and dynamic augmented graph containing multi-dimensional fusion information are used as the input of the graph encoder and encoded, and finally the rumor dynamic propagation graph representation and dynamic augmented graph representation are obtained according to the following method; S31: For a given rumor propagation dynamic graph, the hidden feature vector of the graph encoder is updated and encoded using the aggregation function (1): （1）； in: The representation graph encoder is Layer Node The hidden feature vector of Indicates that the multi-layer perceptron aggregates and updates node features. represents the learning parameters, The representation graph encoder is The learning parameters of the layer, The representation graph encoder is Layer Node The hidden feature vector of The representation graph encoder is Layer Node The hidden feature vector of Representation graph encoder node Neighbor nodes of S32: The image encoder obtains the first The propagation graph of the time snapshot is represented by the formula (3). The augmented graph representation of time snapshots is: （2）； (3)； in: Indicates The node representation of the propagation graph of a time snapshot is: represents the adjacency matrix, Indicates The propagation graph feature matrix of time snapshots, Indicates The propagation graph representation of a time snapshot is It means obtaining the node representation of the graph encoding for the content in the brackets. It means taking the mean of the contents in brackets to form a graph. Indicates The node representation of the augmented graph of time snapshots, Indicates The augmented graph feature matrix of time snapshots, Indicates A graphical representation of the augmented graph of a time snapshot; S33: According to The propagation graph representation of each time snapshot obtains the dynamic propagation graph representation sequence of the same event. The graph representation of the augmented graph of each time snapshot obtains a sequence of dynamic augmented graph representations of the same event; S34: A bidirectional long short-term memory model is used to model the forward and backward dynamic propagation graph representation sequences and dynamic augmentation graph representation sequences, capturing long-term dependencies and identifying potential patterns of the dynamic propagation graph and dynamic augmentation graph; S35: Connect the forward state and the backward state in the potential mode of the dynamic propagation graph and the dynamic augmentation graph to obtain the graph representation of the bidirectional long short-term memory model, and then obtain the batch rumor dynamic propagation graph representation and the batch dynamic augmentation graph representation; S4: Unsupervised comparative learning is performed based on the rumor dynamic propagation graph representation and the dynamic augmented graph representation to capture the representation information of the rumor dynamic propagation graph and the dynamic augmented graph; S5: By integrating the representation information of the rumor dynamic propagation graph and the dynamic augmented graph, we classify rumors through a deep neural classification model and complete robust rumor detection on social media.

2. The method for robust rumor detection on social media according to claim 1, characterized in that: When events in social media data are constructed as a static propagation graph in step S1, the nodes of the static propagation graph are composed of the posts that constitute each event, and the edges of the static propagation graph are composed of the forwarding or comment relationships between the posts and their response posts.

3. The method for robust rumor detection on social media according to claim 1, characterized in that: When events in social media data are constructed as a static propagation graph in step S1, each constructed static propagation graph is a connected graph. If the static propagation graph is not connected, the maximum connected graph is taken as the static propagation graph.

4. The method for robust rumor detection on social media according to claim 1, characterized in that: When constructing events in social media data into a static propagation graph in step S1, events with less than 3 comments in the social media data are eliminated.

5. The method for robust rumor detection on social media according to claim 1, characterized in that: In step S2, a time snapshot construction method is used to construct a rumor dynamic propagation graph containing multi-dimensional fusion information.

6. The method for robust rumor detection on social media according to claim 1, characterized in that: In step S4, the following method is used to perform unsupervised comparative learning based on the rumor dynamic propagation graph representation and the dynamic augmented graph representation to capture the representation information of the rumor dynamic propagation graph and the representation information of the dynamic augmented graph: S41: Using unsupervised contrastive learning, the similarity matrix in the batch rumor dynamic propagation graph representation and the dynamic augmented graph representation is calculated according to formula (4) to capture the similarity between the two vector representations: （4）； in: Represents the similarity matrix in the batch rumor dynamic propagation graph representation and the dynamic augmented graph representation, Represents the dynamic propagation graph of batch rumor processing, represents the batch dynamically augmented graph representation, represents an exponential function with a natural constant as base, represents the cosine similarity function, represents the temperature coefficient; S42: Using the similarity between the two vector representations, the unsupervised contrast loss is calculated according to formula (5); （5）； in: represents the unsupervised contrast loss, represents the identity matrix, Represents the cross entropy loss of calculating the similarity matrix; S43: The unsupervised contrast loss and the cross entropy loss are added to train the model to capture the representation information of the rumor dynamic propagation graph and the dynamic augmented graph during the training process.

7. The method for robust rumor detection on social media according to claim 6, characterized in that: The following method is used to fuse the representation information of the rumor dynamic propagation graph and the representation information of the dynamic augmented graph, and to classify rumors through a deep neural classification model: S51: Connect the representation information of the captured rumor dynamic propagation graph and the representation information of the dynamic augmented graph to form integrated information, and then input the integrated information into the fully connected layer and activation function layer of the deep neural classification model for training, and calculate the predicted probability distribution according to formula (6): （6）； in: represents the predicted probability distribution, represents the activation function, represents the weight, Indicates integrated information. Indicates bias; S52: Calculate the classification cross entropy loss function during training according to formula (7): （7）； in: represents the classification cross entropy loss, represents the number of rumor events, Indicates one of the rumor events in the rumor event. Indicates the type of rumor. Indicates the total number of rumor categories, Indicates the label value of the rumor event, represents the probability distribution of rumor event prediction; S53: Calculate the loss of the deep neural classification model according to formula (8), and train the deep neural classification model to classify rumors and complete robust rumor detection on social media: （8）； in: represents the loss of the deep neural classification model, Represents the hyperparameter used to balance the loss function.

Images