A rumor detection method and system based on causal feature discovery

Abstract

The application discloses a rumor detection method and system based on causal feature discovery, and the method comprises the following steps: S1, data extraction and format conversion are performed on the data obtained from a social network, source posts, reply posts and a propagation path of an event are obtained, and a rumor propagation graph is constructed; S2, text feature extraction is performed on the text of the source posts and the reply posts after the processing in S1, and the text is converted into a vector form; S3, the rumor propagation graph after the processing in S1 is pre-trained, and a pre-trained rumor detection model is obtained; S4, a mask generator is used to divide the rumor propagation graph after the processing in S1 into causal subgraphs and non-causal subgraphs; S5, the pre-trained rumor detection model obtained in S3 is used to enhance the operation of the causal subgraphs and the non-causal subgraphs divided in S4, and false correlation is eliminated; S6, the causal subgraphs obtained in S4 and the rumor propagation graph after the processing in S1 are embedded respectively, and the graph features of the two are extracted; and S7, the causal subgraph features and the rumor propagation graph features obtained in S6 are connected to obtain a representation vector, the representation vector is input into a classifier, a target prediction is calculated, and rumor bias is removed.

Description

Technical Field

[0001] This invention relates to the field of rumor detection technology, specifically to a rumor detection method and system based on causal feature discovery. Background Technology

[0002] In today's society, the rapid development of social media provides people with a platform for quickly obtaining and sharing information. However, the other side of this phenomenon is the widespread dissemination of rumors, which often spread online at an alarming rate, causing enormous harm to individuals and society. To effectively reduce the negative impact of rumors, automated rumor detection technologies have emerged and are gradually becoming a focus of attention in academia and industry. However, a common problem in existing research is that researchers often rely too heavily on the statistical correlation between rumor propagation graphs and labels, neglecting the non-causal features that may be hidden within the rumor propagation graph. This hasty approach may lead to spurious correlations in existing rumor detection methods, resulting in poor performance when dealing with out-of-distribution datasets. Specifically, a rumor propagation graph is a complex network structure that contains information from multiple dimensions, including users, information, and propagation paths. Some of this information is causally related, such as a user's behavior directly influencing the spread of a rumor; while some is non-causal, meaning they may only have a statistical correlation without an actual causal relationship. If researchers do not adequately identify and control these non-causal features, they may mistakenly use them as the basis for rumor detection, leading to inaccurate detection results.

[0003] Furthermore, the spread of rumors is a dynamic process influenced by various factors, including socio-psychological factors, cultural background, and technological means. These factors may manifest differently on different datasets, requiring rumor detection methods to have strong generalization capabilities. However, if the detection method relies too heavily on statistical correlation without a deep understanding of the underlying mechanisms of rumor spread, its performance will often be significantly reduced when faced with datasets whose distribution differs from the training data.

[0004] Therefore, in order to improve the accuracy and robustness of rumor detection and address the problem that existing methods rely too heavily on statistical correlation and thus perform poorly on new datasets, this invention designs a rumor detection method based on causal feature discovery. This method can effectively improve the accuracy and robustness of rumor detection. Furthermore, in order to apply the research results of this invention to real-world scenarios, we have further developed a rumor detection system, aiming to transform this invention into a practical tool to serve society more broadly. Summary of the Invention

[0005] This invention overcomes the problem that existing rumor detection algorithms hastily utilize the statistical correlation between rumor propagation graphs and labels, while neglecting the non-causal features that may be hidden in the rumor propagation graphs. It proposes a rumor detection method and system based on the discovery of causal features.

[0006] The technical concept of this invention is as follows: Due to the importance of structural features in rumor detection and the rapid development of graph-based methods, more and more researchers are adopting rumor propagation-based methods for rumor detection. However, most machine learning methods focus on optimizing the accuracy and efficiency between input data and labels, often neglecting the non-causal features hidden in the rumor propagation graph. This may lead to poor model performance and lack of robustness when facing out-of-distribution data, and current rumor detection technologies are rarely applied in real-world scenarios. Therefore, we propose a rumor detection method and system based on causal feature discovery, which not only effectively improves the model's performance and generalization ability but also demonstrates strong practicality and adaptability in real-world applications.

[0007] The technical solution adopted by the present invention to achieve the above-mentioned objectives is as follows:

[0008] A rumor detection method based on causal feature discovery includes the following steps:

[0009] S1. Extract and convert the data obtained from social networks to obtain the source posts, reply posts and propagation paths of the event, and construct a rumor propagation graph;

[0010] S2. Extract text features from the event source post and reply text processed in S1 and convert them into vector form;

[0011] S3. Pre-train the rumor propagation graph processed in S1 to obtain a pre-trained rumor detection model;

[0012] S4. Divide the rumor propagation graph processed in S1 into causal subgraphs and non-causal subgraphs using a mask generator;

[0013] S5. Train the mask generator in S4 with the pre-trained rumor detection model obtained in S3 to obtain more accurate causal and non-causal subgraphs.

[0014] S6. Embed the causal subgraph obtained in S4 and the rumor propagation graph processed in S1 respectively, and extract the graph features of the two.

[0015] S7. Connect the causal subgraph features and rumor propagation graph features obtained in S6 to obtain a representation vector. Input the representation vector into the classifier to calculate the target prediction and achieve rumor de-biasing.

[0016] Preferably, step S1 specifically includes:

[0017] Representing event-based rumor datasets as ,in Indicates the number of events. Each event... Modeled as a directed graph ,in Represents the i-th event post or reply text, edge This indicates a reply relationship. For a set of nodes, where For the event The number of posts in For the event In the post, j is the post Index sorted by the order in which events are published.

[0018] Preferably, step S2 specifically includes:

[0019] Use a pre-trained Word2Vec model to process posts Converting text into word vectors The word vectors are then converted into text feature vectors for the posts using the following formula:

[0020] (1)

[0021] Preferably, step S3 specifically includes:

[0022] To train a pre-trained rumor detection model, first, given a directed graph... Then, a graph neural network is used to extract network features, and a multilayer perceptron is used to classify the embeddings. The cross-entropy loss function is employed, with the following formula:

[0023] (2)

[0024] (3)

[0025] (4)

[0026] Where y represents a genuine rumor label. Representing a graph neural network, Indicates the readout layer.

[0027] Preferably, step S4 specifically includes:

[0028] A mask generator for partitioning causal and non-causal subgraphs is designed, specifically for a given directed graph. The attention scores of nodes and edges are calculated using node-attention multilayer perceptrons and edge-attention multilayer perceptrons, and then the corresponding soft masks are constructed using the attention scores of nodes and edges.

[0029] Preferably, step S5 specifically includes:

[0030] A pre-trained rumor detection model is used to partition and predict causal and non-causal subgraphs. The formulas are as follows:

[0031] (5)

[0032] (6)

[0033] (7)

[0034] in, (*) represents the pre-trained rumor detection model from step 3. For the model's prediction of the rumor label Y, The non-causal subgraph intervention loss is CE(*), which is the cross-entropy loss, and y is the true rumor label.

[0035] (8)

[0036] (9)

[0037] Indicates false label KL(*) represents the KL divergence. This represents a Uniform-Label. By optimizing the causal subgraph intervention loss, the mask generator can accurately identify non-causal subgraphs from the rumor propagation graph.

[0038] Ultimately, the mask generator loses... This can be expressed as the sum of all losses:

[0039] (10)

[0040] in, and The hyperparameters for the intervention loss of causal and non-causal subgraphs are respectively. To amplify the losses.

[0041] Preferably, step S6 specifically includes:

[0042] Extracting causal subgraphs from the rumor propagation graph using a mask generator Then, two independent graph neural networks were used to analyze the rumor propagation graph. and causal graph Embedding is performed to extract the graph features from both:

[0043] (11)

[0044] (12)

[0045] in, (*) represents the readout function, implemented using an average pooling layer. and These are the embedding vectors of the rumor propagation graph and the causal subgraph, respectively.

[0046] Preferably, step S7 specifically includes:

[0047] The embedding vectors of the rumor propagation graph and causal subgraph obtained in step S6 and A concatenation operation is performed to obtain a representation vector. This representation vector is then fed into a classifier to calculate the target's predicted label, using the following formula:

[0048] (13)

[0049] (14)

[0050] in, For classifiers, For predicting labels, The loss function is cross-entropy, which is used here to minimize the loss. This enables the debiased rumor detector to correctly identify rumors.

[0051] The present invention also provides a rumor detection system based on causal feature discovery, comprising: an application layer, a platform layer, and a support layer;

[0052] The application layer is used in practical applications. It is applied to actual application scenarios in the form of browser plugins. When browsing posts, the plugin will automatically extract the post information of the current page to detect rumors and return the results to be displayed on the page.

[0053] The platform layer is used to crawl information from web pages, detect rumors from the crawled information, and finally store the results. It is mainly divided into three modules: Weibo crawler, rumor detection, and storage. The Weibo crawler module mainly obtains information from Weibo posts, the rumor detection module mainly inputs the obtained Weibo information into a trained model to detect rumors, and the storage module mainly saves the detection results.

[0054] The support layer is used to build a runtime environment that supports the entire system, including a Python environment for executing programs.

[0055] The beneficial effects of this invention are as follows:

[0056] (1) This invention employs an innovative causal reasoning framework to ensure the accuracy and reliability of detection results by identifying and blocking these backdoor paths. In this way, we can not only more accurately identify and filter rumors, but also improve the performance and efficiency of the entire system;

[0057] (2) This invention innovatively improves the mask generator by employing a soft masking method combined with a pre-trained model for training, in order to more accurately capture causal subgraphs closely related to the rumor detection results. To further enhance the performance of the mask generator, a causal enhancement loss is introduced as a penalty term. This strategy effectively guides the mask generator to focus on identifying and constructing purer causal subgraphs. In this way, not only is the accuracy of the mask generator improved, but its robustness in complex data environments is also enhanced, thus achieving superior performance in the rumor detection task.

[0058] (3) The present invention designs a rumor debias detector, which effectively improves the rumor debiasing ability of the model. Attached Figure Description

[0059] Figure 1 This is a structural diagram of the present invention;

[0060] Figure 2(a) shows the causal graph of the mask generator training for the causal part; Figure 2(b) shows the causal graph of the mask generator training for the non-causal part.

[0061] Figure 3 To compare the proposed method with the baseline model on the Weibo dataset;

[0062] Figure 4 To compare the proposed method with the baseline model on the PHEME dataset;

[0063] Figure 5 This is a schematic diagram of the system structure according to an embodiment of the present invention;

[0064] Figure 6 This is a schematic diagram of the application layer according to an embodiment of the present invention;

[0065] Figure 7 A flowchart of the Weibo crawler process;

[0066] Figure 8 This is a flowchart for rumor detection. Detailed Implementation

[0067] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0068] Reference Figure 1 This paper describes a rumor detection method based on causal feature discovery, which includes the following steps:

[0069] S1. Extract and convert the data obtained from social networks to obtain the source posts, reply posts and propagation paths of the event, and construct a rumor propagation graph;

[0070] S2. Extract text features from the event source post and reply text processed in S1 using Word2Vec and convert them into vector form;

[0071] S3. The rumor propagation graph processed in S1 is pre-trained using GNN to obtain a pre-trained rumor detection model.

[0072] S4. Use a mask generator to divide the rumor propagation graph processed in S1 into causal subgraphs and non-causal subgraphs.

[0073] S5. Train the mask generator in S4 with the pre-trained rumor detection model obtained in S3 to obtain more accurate causal and non-causal subgraphs.

[0074] S6. Embed the causal subgraph obtained in S4 and the rumor propagation graph processed in S1 using GNN to extract the graph features of both.

[0075] S7. Connect the causal subgraph features and rumor propagation graph features obtained in S6 to obtain a representation vector. Input the representation vector into the classifier to calculate the target prediction and achieve rumor de-biasing.

[0076] Step S1 specifically includes:

[0077] This invention uses the Weibo and PHEME datasets to repackage the original data according to the format of the event to which the post belongs, post ID, post text, posting time, and reply ID. The post text is preprocessed text information. Then, a rumor propagation graph is constructed based on the relationship between the source post and the replies.

[0078] Step S2 specifically includes:

[0079] Use a pre-trained Word2Vec model to process posts Converting text into word vectors The word vectors are then converted into text feature vectors for the posts using the following formula:

[0080] (1)

[0081] Step S3 specifically includes:

[0082] To train a pre-trained rumor detection model, first, given a directed graph... The model utilizes a GNN for network feature extraction, and then a multilayer perceptron (MLP) is used to classify the embeddings. To optimize the pre-trained rumor detection model, a cross-entropy loss function is employed, with the following formula:

[0083] (2)

[0084] (3)

[0085] (4)

[0086] Where y represents a genuine rumor label. Representing a graph neural network, Indicates the readout layer.

[0087] Preferably, step S4 specifically includes:

[0088] A mask generator for partitioning causal and non-causal subgraphs is designed, specifically for a given directed graph. Using node attention multilayer perceptron (*) and side-attention multilayer perceptron (*) Calculate the attention score of the node and the attention score of the edge, and then construct the corresponding soft mask using the attention scores of the node and the edge, as shown in the following formula:

[0089] (5)

[0090] (6)

[0091] (7)

[0092] (8)

[0093] (9)

[0094] (10)

[0095] in, and These are the text features of node i and node j, respectively, and || denotes the vector concatenation operation. and These represent the probabilities that a node belongs to a causal subgraph and a non-causal subgraph, respectively. and Then, these represent the probabilities that an edge belongs to a causal subgraph and a non-causal subgraph, respectively. and Let these be the soft masks corresponding to the nodes and edges, respectively. Finally, the causal and non-causal subgraphs can be represented as follows:

[0096] (11)

[0097] (12)

[0098] in, This represents the Hadmard product.

[0099] Preferably, step S5 specifically includes:

[0100] To train the mask generator to accurately divide the rumor propagation graph into causal and non-causal subgraphs, while ensuring that the rumor detection result Y is only affected by the causal subgraphs to eliminate false correlations, the training data needs to be intervened to block S. Y, as shown in Figure 2(a). Traditional rumor detection methods have a backdoor path S. This leads to a false correlation between the non-causal subgraph S and the rumor detection result Y. To eliminate the false correlation, the non-causal subgraph S can be intervened with do(S= To block the aforementioned path, the process described above can be represented by the following formula:

[0101] (13)

[0102] (14)

[0103] (15)

[0104] in, (*) represents the pre-trained rumor detection model from step 3. To predict the rumor label Y using the model, and to identify causal subgraphs, One-Hot labels are used as the optimization objective of the mask generator. is the non-causal subgraph intervention loss, CE(*) is the cross-entropy loss, and y is the real One-Hot rumor label. By optimizing the non-causal subgraph intervention loss, the mask generator can accurately identify the causal subgraph from the rumor propagation graph.

[0105] To more accurately unravel the causal aspect, auxiliary pseudo-prediction labels are introduced. It is only affected by S. As shown in Figure 2(b), there may be a front door path. This makes C and There are spurious correlations. Intervention in the causal subgraph (do(C=)) This is used to block the aforementioned path. This process can be represented as follows:

[0106] (16)

[0107] (17)

[0108] Indicates false label KL(*) represents the KL divergence. This represents a Uniform-Label. By optimizing the causal subgraph intervention loss, the mask generator can accurately identify non-causal subgraphs from the rumor propagation graph.

[0109] To capture purer causal subgraphs that are more closely related to the outcome, a causal enhancement loss is introduced as a penalty term. The causal enhancement loss consists of node causal enhancement loss and edge causal enhancement loss:

[0110]

[0111]

[0112]

[0113] in, and These are the node causal enhancement loss and the edge causal enhancement loss, respectively. To enhance the causal effect, ultimately, the mask generator loss... This can be expressed as the sum of all losses:

[0114] (twenty one)

[0115] in, and , respectively, are the hyperparameters of the intervention loss for causal and non-causal subgraphs.

[0116] Preferably, step S6 specifically includes:

[0117] Extracting causal subgraphs from the rumor propagation graph using a mask generator Then, two independent GNNs were used to analyze the rumor propagation graph. and causal graph Embedding is performed to extract the graph features from both:

[0118] (twenty two)

[0119] (twenty three)

[0120] in, (*) represents the readout function, implemented using an average pooling layer. and These are the embedding vectors of the rumor propagation graph and the causal subgraph, respectively.

[0121] Preferably, step S7 specifically includes:

[0122] The embedding vectors of the rumor propagation graph and causal subgraph obtained in step S6 and A concatenation operation is performed to obtain a representation vector. This representation vector is then fed into a classifier to calculate the target's predicted label, using the following formula:

[0123] (twenty four)

[0124] (25)

[0125] in, For classifiers, For predicting labels, The loss function is cross-entropy, which is used here to minimize the loss. This enables the debiased rumor detector to correctly identify rumors.

[0126] To demonstrate that the rumor detection method based on causal feature discovery proposed in this invention outperforms other existing models, we tested its performance on the Weibo and PHEME datasets. The experimental results show that the proposed method outperforms baseline methods. Here, RNN represents modeling tweets in an event as a variable-length temporal sequence and using a recurrent neural network for classification; BU-RVNN uses a bottom-up tree-structured neural network to extract the propagation path features of the text content; BiGCN represents extracting rumor propagation and propagation features through two GCNs; and GCN represents extracting graph propagation features. Ours+LEConv represents the model combined with our framework.

[0127] from Figure 3 As can be seen, the Ours+LEConv model combined with the framework proposed in this invention outperforms all baseline models, with an accuracy of 0.953.

[0128] from Figure 4 As can be seen, the Ours+LEConv model combined with the framework proposed in this invention outperforms all baseline models on the PHEME dataset.

[0129] like Figure 5 As shown, this invention also provides a rumor detection system based on causal feature discovery using the above method, comprising three parts: an application layer, a platform layer, and a support layer.

[0130] The application layer is used in practical application scenarios, applied as a browser plugin, such as on the Weibo platform. When browsing posts, the plugin automatically extracts information such as the current page's post ID, text, reply post ID, and reply text to detect rumors, and returns the results to be displayed on the page, with the effect as follows: Figure 6 As shown, Figure 6 (a) This is the interface for identifying the plugin and submitting rumor detection requests. If the page is correctly identified and the detection results are returned, the rumor detection results will be displayed, such as... Figure 6 As shown in (b);

[0131] The platform layer is used to crawl web pages, detect rumors from the crawled information, and finally store the results. It mainly consists of three modules: a Weibo crawler, a rumor detection module, and a storage and recording module. The Weibo crawler module primarily acquires Weibo post information by retrieving relevant Weibo information based on received requests. The Weibo crawler flowchart is shown below. Figure 7 As shown, the rumor detection module mainly inputs the acquired Weibo information into the trained model for rumor detection. The rumor detection flowchart is as follows. Figure 8 As shown, the storage and recording module mainly saves the detection results, including information such as post ID and text.

[0132] The support layer is used to build a runtime environment that supports the entire system, including a basic Docker runtime container and a Python environment for executing programs, as well as environments such as PyTorch required by the model.

[0133] The embodiments described in this specification are merely examples of implementations of the inventive concept. The scope of protection of this invention should not be considered as limited to the specific forms stated in the embodiments. The scope of protection of this 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 causal feature discovery, characterized in that: Includes the following steps: S1. Extract and convert the data obtained from social networks to obtain the source posts, reply posts and propagation paths of the event, and construct a rumor propagation graph; S2. Extract text features from the event source post and reply text processed in S1 and convert them into vector form; S3. Pre-train the rumor propagation graph processed in S1 to obtain a pre-trained rumor detection model; S4. Divide the rumor propagation graph processed in S1 into causal subgraphs and non-causal subgraphs using a mask generator; S5. Train the pre-trained rumor detection model obtained in S3 on the mask generator in S4 to obtain more accurate causal and non-causal subgraphs; specifically including: To train the mask generator to accurately divide the rumor propagation graph into causal and non-causal subgraphs, while ensuring that the rumor detection result Y is only affected by the causal subgraphs to eliminate false correlations, the training data needs to be intervened. The process is described by the following formula: (7) (8) (9) in, (*) represents the pre-trained rumor detection model from step 3. To predict the rumor label Y using the model, and to identify the causal subgraph, One-Hot labels are used as the optimization objective of the mask generator. The non-causal subgraph intervention loss is CE(*), which is the cross-entropy loss, and y is the real One-Hot rumor label; To more accurately unravel the causal aspect, auxiliary pseudo-prediction labels are introduced. It is only affected by S; the intervention loss is achieved by optimizing the causal subgraph of this process. This enables the mask generator to accurately identify non-causal subgraphs from the rumor propagation graph; To capture purer causal subgraphs that are more closely related to the results, a causal enhancement loss is introduced as a penalty term. This causal enhancement loss consists of node causal enhancement loss and edge causal enhancement loss. Finally, the mask generator loss... Represented as the sum of all losses: (10) in, and The hyperparameters for the intervention loss of causal and non-causal subgraphs are respectively. To amplify the losses; S6. Embed the causal subgraph obtained in S5 and the rumor propagation graph processed in S1 through graph neural networks to extract the graph features of both. S7. Connect the causal subgraph features and rumor propagation graph features obtained in S6 to obtain a representation vector. Input the representation vector into the classifier to achieve rumor debiasing.

2. The rumor detection method based on causal feature discovery as described in claim 1, characterized in that: Step S1 specifically includes: Representing event-based rumor datasets as ,in Indicates the number of events; each event Modeled as a directed graph ,in Represents the i-th event post or reply text, edge Indicates a reply relationship; For a set of nodes, where For the event The number of posts in For the event In the post, j is the post Index sorted by the order in which events are published.

3. The rumor detection method based on causal feature discovery as described in claim 1, characterized in that: Step S2 specifically includes: Use a pre-trained Word2Vec model to process posts Converting text into word vectors The word vectors are then converted into text feature vectors for the posts using the following formula: (1)。 4. The rumor detection method based on causal feature discovery as described in claim 1, characterized in that: Step S3 specifically includes: To train a pre-trained rumor detection model, first, given a directed graph... The model utilizes GNN for network feature extraction and a multilayer perceptron (MLP) for embedding classification. To optimize the pre-trained rumor detection model, a cross-entropy loss function is employed, with the following formula: (2) (3) (4) Where y represents a genuine rumor label. This represents a graph neural network. Indicates the readout layer.

5. The rumor detection method based on causal feature discovery as described in claim 1, characterized in that: Step S4 specifically includes: A mask generator for partitioning causal and non-causal subgraphs is designed, specifically for a given directed graph. Using node attention multilayer perceptron (*) and side-attention multilayer perceptron (*) Calculate the attention score of the node and the attention score of the edge, and then construct the corresponding soft mask based on the attention scores of the node and the edge. and Let these be the soft masks corresponding to the nodes and edges, respectively. Finally, the causal and non-causal subgraphs are represented as follows: (5) (6) in, This represents the Hadmard product.

6. The rumor detection method based on causal feature discovery as described in claim 1, characterized in that: Step S6 specifically includes: Extracting causal subgraphs from the rumor propagation graph using a mask generator Then, two independent GNNs were used to analyze the rumor propagation graph. and causal graph Embedding is performed to extract the graph features from both: (11) (12) in, (*) represents the readout function, implemented using an average pooling layer. and These are the embedding vectors of the rumor propagation graph and the causal subgraph, respectively.

7. The rumor detection method based on causal feature discovery as described in claim 1, characterized in that: Step S7 specifically includes: The embedding vectors of the rumor propagation graph and causal subgraph obtained in step S6 and A concatenation operation is performed to obtain a representation vector. This representation vector is then fed into a classifier to calculate the target's predicted label, using the following formula: (13) (14) in, For classifiers, For predicting labels, The loss function is cross-entropy, which is used here to minimize the loss. This reduces the impact of non-causal factors on rumor detection, enabling the biased rumor detector to correctly identify rumors.

8. A rumor detection system based on causal feature discovery, implementing the method as described in claim 1, characterized in that, include: The application layer is used in practical applications. It is applied to actual application scenarios in the form of browser plugins. When browsing posts, the plugin automatically extracts the post ID information of the current page to detect rumors and returns the results to be displayed on the page. The platform layer is used to crawl information from web pages, detect rumors from the crawled information, and finally store and record the results. It is divided into three modules: Weibo crawler, rumor detection, and storage and recording. The Weibo crawler module obtains Weibo post information and retrieves the corresponding Weibo information according to the received requests. The rumor detection module inputs the obtained Weibo information into the trained model to detect rumors. The storage and recording module mainly saves the detection results. The support layer is used to build a runtime environment that supports the entire system, including the basic Docker runtime container and the Python environment for executing programs.

Images