A rumor detection method and system based on adaptive causal subgraph learning

Abstract

The application relates to the field of rumor detection, and provides a rumor detection method and system based on adaptive causal subgraph learning, which comprises the following steps: establishing a rumor propagation event; obtaining attention coefficients between nodes, performing feature coding through the attention coefficients between nodes, and obtaining rumor node representations; extracting the attention coefficients and a feature matrix, obtaining a node environment mask and an edge environment mask, enhancing the feature matrix through the node environment mask, obtaining an enhanced feature matrix, weighting the attention coefficients through the edge environment mask, and obtaining target attention coefficients; obtaining causal subgraph representations and environment subgraph representations, obtaining combined features, calculating divergence loss of the combined features; constructing an identification model, obtaining an identification result by using the identification model, calculating an identification label loss of the identification result through the divergence loss, training the identification model according to the identification label loss, and completing rumor detection through the trained identification model.

Description

Technical Field

[0001] This invention relates to the field of rumor detection technology, and in particular to a rumor detection method and system based on adaptive causal subgraph learning. Background Technology

[0002] Social media has become a primary medium for communication and information dissemination in daily life. With its widespread use, a vast amount of user-generated content is constantly being generated and shared online. However, this openness also accelerates the spread of rumors. Therefore, effective rumor detection on social media is crucial for mitigating its negative impact. Graph neural networks (GNNs) have demonstrated excellent performance in modeling rumor propagation, contributing to a deeper understanding of the underlying propagation patterns. Researchers explored rumor detection using GNNs for globally propagating information. Building upon this, a masking strategy was added to the global propagation graph, exploring various subgraph learning strategies for rumor detection. While these methods provide insights into different subgraph learning techniques, two significant limitations exist. First, subgraphs obtained through fixed-scale or heuristic sampling may limit the expressive power of graph semantics. Second, causal subgraphs in existing methods may not maintain robustness and stability under environmental or structural perturbations. Summary of the Invention

[0003] This invention aims to at least solve one of the technical problems existing in related technologies. To this end, this invention provides a rumor detection method and system based on adaptive causal subgraph learning, achieving highly accurate detection of rumor content on social media.

[0004] This invention provides a rumor detection method based on adaptive causal subgraph learning, comprising:

[0005] S1: Obtain rumor samples and establish rumor propagation events based on the rumor samples;

[0006] S2: Extract features from the rumor propagation event to obtain the attention coefficients between nodes, and encode the features using the attention coefficients to obtain the rumor node representation;

[0007] S3: Extract the attention coefficients and feature matrix, obtain the node environment mask and edge environment mask through the rumor node representation, enhance the feature matrix through the node environment mask to obtain the enhanced feature matrix, and use the edge environment mask to weight the attention coefficients to obtain the target attention coefficients;

[0008] S4: Obtain the causal subgraph representation and the environment subgraph representation by enhancing the feature matrix and the target attention coefficient. Obtain the combined features based on the causal subgraph representation and the environment subgraph representation, and calculate the divergence loss of the combined features.

[0009] S5: Build an identification model, use the identification model to obtain the identification results, calculate the identification label loss of the identification results through divergence loss, train the identification model based on the identification label loss, and complete the detection of rumors through the trained identification model.

[0010] According to the rumor detection method based on adaptive causal subgraph learning provided by the present invention, in step S1, a rumor event is determined from the rumor sample, the propagation nodes and propagation paths of the rumor event are extracted, a rumor propagation graph is established through the propagation nodes and propagation paths, and the rumor propagation event is established using the rumor propagation graph.

[0011] According to the rumor detection method based on adaptive causal subgraph learning provided by the present invention, in step S2, node feature representations are extracted from the rumor propagation event, the node feature representations are activated and normalized to obtain the attention coefficients between nodes.

[0012] According to the rumor detection method based on adaptive causal subgraph learning provided by the present invention, in step S2, an encoder is determined, the node feature representation is weighted according to the attention coefficient between nodes and encoded using the encoder to obtain the rumor node representation.

[0013] According to the rumor detection method based on adaptive causal subgraph learning provided by the present invention, in step S3, a multilayer perceptron network is constructed, a node causal mask is obtained through the multilayer perceptron network and the rumor node representation, an edge causal mask is obtained through the node causal mask and the multilayer perceptron network, and a node environment mask and an edge environment mask are obtained based on the node causal mask and the edge causal mask.

[0014] According to the rumor detection method based on adaptive causal subgraph learning provided by the present invention, in step S3, attention coefficients and a feature matrix including a causal feature matrix and an environmental feature matrix are extracted from the rumor node representation. The node environment mask, the causal feature matrix and the environmental feature matrix are used for enhancement to obtain the enhanced feature matrix including the enhanced causal feature matrix and the enhanced environmental feature matrix. The edge environment mask and the attention coefficients are used for weighting and normalization to obtain the target attention coefficients.

[0015] According to the rumor detection method based on adaptive causal subgraph learning provided by the present invention, in step S4, the target attention coefficient and the causal feature matrix are encoded to obtain causal node representations, and the causal node representations are subjected to mean pooling to obtain the causal subgraph representations.

[0016] The target attention coefficient and the environment feature matrix are encoded to obtain the environment node representation. The environment node representation is then subjected to mean pooling to obtain the environment subgraph representation.

[0017] According to the rumor detection method based on adaptive causal subgraph learning provided by the present invention, in step S4, random noise is determined, the combined feature is constructed through the random noise, the causal subgraph representation and the environmental subgraph representation, and the divergence loss of the combined feature and the uniform class distribution is calculated.

[0018] According to the rumor detection method based on adaptive causal subgraph learning provided by the present invention, in step S5, the cross-entropy loss of the identification result is calculated, a weighting coefficient is determined, the divergence loss is weighted using the weighting coefficient, and then added to the cross-entropy loss to obtain the identification label loss.

[0019] This invention also provides a rumor detection system based on adaptive causal subgraph learning, comprising:

[0020] The rumor propagation event module includes obtaining rumor samples and establishing rumor propagation events based on these samples.

[0021] Rumor Node Representation Module: Used to extract features from rumor propagation events, obtain attention coefficients between nodes, and encode features using these attention coefficients to obtain rumor node representations;

[0022] Target attention coefficient module: used to extract attention coefficients and feature matrix, obtain node environment mask and edge environment mask through rumor node representation, enhance the feature matrix through node environment mask to obtain enhanced feature matrix, and use edge environment mask to weight attention coefficients to obtain target attention coefficients;

[0023] The divergence loss module is used to obtain causal subgraph representations and environmental subgraph representations through the enhanced feature matrix and target attention coefficients, obtain combined features based on the causal subgraph representations and environmental subgraph representations, and calculate the divergence loss of the combined features.

[0024] Rumor detection module: Used to build an identification model, obtain identification results using the identification model, calculate the identification label loss of the identification results through divergence loss, train the identification model based on the identification label loss, and complete the detection of rumors using the trained identification model.

[0025] The above-described one or more technical solutions in the embodiments of the present invention have at least one of the following technical effects:

[0026] This invention provides a rumor detection method and system based on adaptive causal subgraph learning. It utilizes node and edge context masks to extract causal subgraphs for rumor detection. The node and edge context masks adaptively learn the importance of posts within the propagation graph and their interrelationships, guiding the model to identify the causal subgraph most relevant to the rumor's veracity. The node context mask helps identify key nodes, while the edge context mask guides the learning of key edges, resulting in a robust causal subgraph. This allows the model to learn subgraphs that remain stable under noisy or mixed conditions, improving the effectiveness of rumor detection.

[0027] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0029] Figure 1 This is a flowchart illustrating a rumor detection method based on adaptive causal subgraph learning provided by the present invention.

[0030] Figure 2 This is a schematic diagram of the structure of a rumor detection system based on adaptive causal subgraph learning provided by the present invention.

[0031] Figure label:

[0032] 100. Rumor propagation event module; 200. Rumor node representation module; 300. Target attention coefficient module; 400. Divergence loss module; 500. Rumor detection module. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention. The following embodiments are used to illustrate this invention but cannot be used to limit the scope of this invention.

[0034] In the description of the embodiments of the present invention, it should be noted that the terms "first", "second" and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0035] In the description of the embodiments of the present invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "connected" and "linked" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms in the embodiments of the present invention based on the specific circumstances.

[0036] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0037] The following is combined Figures 1 to 2 Specific embodiments of the present invention are described below. Figure 1 This is a flowchart illustrating a rumor detection method based on adaptive causal subgraph learning provided by the present invention, including:

[0038] S1: Obtain rumor samples and establish rumor propagation events based on the rumor samples;

[0039] Furthermore, the objective of this stage is to establish a rumor propagation event. Specifically, in step S1, a rumor event is determined from the rumor sample, the propagation nodes and propagation paths of the rumor event are extracted, a rumor propagation graph is established through the propagation nodes and propagation paths, and the rumor propagation graph is used to establish the rumor propagation event.

[0040] The specific implementation method for the above steps in this embodiment is as follows:

[0041] First, it's necessary to acquire posts from social media, extract their text content and propagation structure, and label them with authenticity tags to obtain rumor samples. From these samples, a specific event mentioned is identified as a potential rumor event, and its propagation nodes and paths are extracted. Propagation nodes are the various posts mentioning the rumor event, and propagation paths are the routes through which the rumor content spreads from an initial post to other posts via reposts, citations, etc.; these are also known as the edge relationships between posts. A rumor propagation graph G is then constructed using the propagation nodes and paths. Here, V represents the set of nodes in the propagation of the rumor event, and E represents the edge relationships of the rumor event. By combining the rumor propagation graph with the rumor event itself and whether the rumor event is true or not, i.e., the authenticity label, we can obtain the rumor propagation event.

[0042] S2: Extract features from the rumor propagation event to obtain the attention coefficients between nodes, and encode the features using the attention coefficients to obtain the rumor node representation;

[0043] Furthermore, the objective of this stage is to obtain the attention coefficients between nodes, and then perform feature encoding to obtain the rumor node representation. Specifically, in step S2, node feature representations are extracted from the rumor propagation event, and the node feature representations are activated and normalized to obtain the attention coefficients between nodes.

[0044] In step S2, an encoder is determined, the node feature representation is weighted according to the inter-node attention coefficient, and encoded using the encoder to obtain the rumor node representation.

[0045] The specific implementation method for the above steps in this embodiment is as follows:

[0046] First, it is necessary to extract node feature representations from the rumor propagation event, that is, the information features possessed by each node in the rumor event. Then, the node feature representations are activated and normalized to obtain the inter-node attention coefficients between the i-th and j-th propagation nodes. :

[0047]

[0048]

[0049] in, Let be the node feature representation of the i-th propagation node. Let j be the node feature representation of the j-th propagation node. Let T represent the learnable parameter vector, and T denote the transpose. This is the first trainable weight matrix. Let LeakyReLU() be the second trainable weight matrix. Let be the unnormalized inter-node attention coefficient between the i-th propagation node and the j-th propagation node. Let be the unnormalized inter-node attention coefficient between the i-th propagation node and the k-th neighboring node. Let exp() be the set of neighboring nodes of the i-th propagation node, and let exp() denote the exponential function. Let be the inter-node attention coefficient between the i-th propagation node and the j-th propagation node.

[0050] The encoder is then determined; in this embodiment, a dual-layer GATv2 encoder is selected as the feature encoder. Subsequently, the node feature representations are weighted in multiple layers based on the inter-node attention coefficients to obtain the i-th propagation node's... Layer-weighted node feature representation :

[0051]

[0052] in, Represents a nonlinear function. For the i-th propagation node and its k-th neighboring node, the distance between the i-th propagation node and the k-th neighboring node is... Attention coefficients between nodes in a layer For the first The trainable weight matrix of the layer. For the first The weighted node feature representation of the k-th neighbor node in the first layer is used, while the first layer uses the node feature representation of the k-th neighbor node. Multiple weighting is performed using the above method, and the weighted node feature representation of the last layer is output as the final node feature representation. This is then input into the encoder for encoding to obtain the rumor node representation.

[0053] S3: Extract the attention coefficients and feature matrix, obtain the node environment mask and edge environment mask through the rumor node representation, enhance the feature matrix through the node environment mask to obtain the enhanced feature matrix, and use the edge environment mask to weight the attention coefficients to obtain the target attention coefficients;

[0054] Furthermore, the objective of this stage is to obtain the node environment mask and the edge environment mask, and to use these masks for enhancement and weighting to obtain the target attention coefficient. Specifically, in step S3, a multilayer perceptron network is constructed. A node causal mask is obtained through the multilayer perceptron network and the rumor node representation. An edge causal mask is obtained through the node causal mask and the multilayer perceptron network. Finally, the node environment mask and the edge environment mask are obtained based on the node causal mask and the edge causal mask.

[0055] In step S3, attention coefficients and a feature matrix including a causal feature matrix and an environmental feature matrix are extracted from the rumor node representation. The node environment mask, the causal feature matrix, and the environmental feature matrix are used for enhancement to obtain the enhanced feature matrix including the enhanced causal feature matrix and the enhanced environmental feature matrix. The edge environment mask and the attention coefficients are used for weighting and normalization to obtain the target attention coefficients.

[0056] The specific implementation method for the above steps in this embodiment is as follows:

[0057] First, construct a multilayer perceptron network. Next, the uniform random noise in the rumor propagation process is identified. Based on the uniform random noise, the noise interference of the multilayer perceptron network can be obtained. By using the sigmoid activation function to activate the noise interference and the multilayer perceptron, the Gumbel-Sigmoid function can be obtained. The rumor node is represented using a multilayer perceptron network and the i-th propagation node. Obtain the node causality mask of the i-th propagation node. :

[0058]

[0059] The node causality masks of each propagation node can be combined to form a node causality mask matrix. Subsequently, the edge causality mask between the i-th propagation node and the j-th propagation node is obtained through node causality masking and a multilayer perceptron network. :

[0060]

[0061] in, Let j be the node causality mask for the j-th propagation node. Let j be the rumor node representing the j-th propagation node. This is a multilayer perceptron network that senses two elements. Similarly, the edge causal masks of each propagation node can form an edge causal mask matrix. The node environment mask can then be obtained. and edge environment mask :

[0062]

[0063]

[0064] Subsequently, the causal attention coefficient of rumor propagation between the i-th propagation node and the j-th propagation node can be extracted from the rumor node representation. and environmental attention coefficient The attention coefficient can also be used to extract causal feature matrix X and environmental feature matrix. The characteristic matrix of .

[0065] By enhancing the causal feature matrix, we can obtain the enhanced causal feature matrix. :

[0066]

[0067] in, This is element-wise multiplication.

[0068] Similarly, by using node environment masks to enhance the environment feature matrix, an enhanced environment feature matrix can be obtained. :

[0069]

[0070] This yields the enhanced feature matrix, which includes both the causal feature matrix and the environmental feature matrix. Subsequently, the causal attention coefficients are weighted and normalized to obtain the target causal attention coefficients between the i-th and j-th propagation nodes. :

[0071]

[0072]

[0073] in, The weighted causal attention coefficients between the i-th propagation node and the j-th propagation node are... The weighted causal attention coefficients between the i-th propagation node and its k-th neighboring propagation node are... This is a pre-defined temperature coefficient. Similarly, the edge environment mask factor between the i-th propagation node and the j-th propagation node is extracted from the edge environment mask. This allows for the weighting and normalization of the environmental attention coefficients, yielding the target environmental attention coefficients between the i-th and j-th propagation nodes. :

[0074]

[0075]

[0076] in, The weighted environmental attention coefficients between the i-th and j-th propagation nodes are denoted as , and the target attention coefficients include the target environmental attention coefficients and the target causal attention coefficients.

[0077] S4: Obtain the causal subgraph representation and the environment subgraph representation by enhancing the feature matrix and the target attention coefficient. Obtain the combined features based on the causal subgraph representation and the environment subgraph representation, and calculate the divergence loss of the combined features.

[0078] Furthermore, the objective of this stage is to obtain the combined features and calculate the divergence loss of the combined features. Specifically, in step S4, the target attention coefficients and the causal feature matrix are encoded to obtain causal node representations, and the causal node representations are subjected to mean pooling to obtain the causal subgraph representation;

[0079] The target attention coefficient and the environment feature matrix are encoded to obtain the environment node representation. The environment node representation is then subjected to mean pooling to obtain the environment subgraph representation.

[0080] In step S4, random noise is determined, and the combined feature is constructed using the random noise, causal subgraph representation, and environmental subgraph representation. The divergence loss of the combined feature and the uniform class distribution is calculated.

[0081] The specific implementation method for the above steps in this embodiment is as follows:

[0082] By inputting the set A of all target causal attention coefficients and the causal feature matrix into the encoder, the causal node representation can be obtained. :

[0083]

[0084] Here, GAT() represents the encoder. Then, mean pooling is performed on the causal node representations to obtain the causal subgraph representation. :

[0085]

[0086] Here, MEAN() represents mean pooling. Similarly, it represents the set of all target and environment attention coefficients within the target attention coefficient set. The environmental feature matrix is ​​input into the encoder for encoding, which yields the environmental node representation. :

[0087]

[0088] Next, mean pooling is performed on the environmental node representation to obtain the environmental subgraph representation. :

[0089]

[0090] Random noise in the rumor propagation process was subsequently identified. This allows for the construction of combined features. :

[0091]

[0092] Finally, based on the propagation type of the rumor data, the uniform category distribution u of each type of rumor is determined, and the divergence loss between the combined features and the uniform category distribution is calculated. :

[0093]

[0094] Here, Softmax() represents a normalized exponential function transformation of the content within the parentheses. This indicates the divergence loss between the two values ​​within the parentheses. This represents the weight coefficients of the trainable combined features.

[0095] S5: Build an identification model, use the identification model to obtain the identification results, calculate the identification label loss of the identification results through divergence loss, train the identification model based on the identification label loss, and complete the detection of rumors through the trained identification model.

[0096] Furthermore, the purpose of this stage is to train the recognition model and use the trained model to detect rumors. Specifically, in step S5, the cross-entropy loss of the recognition result is calculated, a weighting coefficient is determined, the divergence loss is weighted using the weighting coefficient, and then added to the cross-entropy loss to obtain the recognition label loss.

[0097] The specific implementation method for the above steps in this embodiment is as follows:

[0098] Here, the first step is to construct a recognition model capable of identifying whether the content being spread is a rumor. This involves inputting a causal subgraph representation into the recognition model to obtain the probability distribution of whether a rumor spread event is indeed a rumor. The cross-entropy loss between the probability distribution and the actual truth labels is calculated. A pre-determined tradeoff coefficient is used to weight the divergence loss. The weighted divergence loss is then added to the cross-entropy loss to obtain the label recognition loss. The recognition model is trained with the goal of minimizing the label recognition loss, resulting in a trained model. This trained model is then used to identify and label rumors.

[0099] The effectiveness of a rumor detection method based on adaptive causal subgraph learning was also verified here. Table 1 compares the results of our method with other methods for identifying rumors. Datasets 1, 2, and 3 are publicly available social media datasets; Bi-GCN, EBGCN, GACL, RDEA, TrustRD, DCE-RD, and PSGT are existing rumor detection methods; CARD is the method provided in this invention. Here, T indicates that the rumor event is not a rumor, and F indicates that the rumor event is a rumor.

[0100] Table 1. Comparison of Rumor Identification Results between This Method and Other Methods

[0101]

[0102] As can be seen, RDEA, TrustRD, and GACL generate subgraphs by randomly deleting edges or nodes, thus introducing random perturbations during training. Meanwhile, DCE-RD employs a top-k node sampling strategy to construct causal subgraphs and their corresponding counterfactual evidence. PSGT focuses on extracting task-relevant attention subgraphs to highlight informative propagation patterns for classification. However, such subgraphs derived from random noise or heuristic top-k sampling inevitably lose crucial semantic and structural information, resulting in incomplete representations of the original propagation graph and limiting the model's inference capabilities. This invention, however, preserves this crucial semantic and structural information. Table 2 also shows the removal of some steps in this method, which significantly reduces the ability to identify rumors, illustrating the importance of each step in this method.

[0103] Table 2. Comparison of rumor identification results between the complete method and the method with some steps removed.

[0104]

[0105] The following describes a rumor detection device based on adaptive causal subgraph learning provided by the present invention. The rumor detection device based on adaptive causal subgraph learning described below and the rumor detection method based on adaptive causal subgraph learning described above can be referred to and correspond to each other.

[0106] Figure 2 An example is a schematic diagram of the structure of a rumor detection system based on adaptive causal subgraph learning, as shown below. Figure 2 As shown, a rumor detection method based on adaptive causal subgraph learning, as described above, includes:

[0107] Rumor propagation event module 100: includes obtaining rumor samples and establishing rumor propagation events based on rumor samples;

[0108] Rumor Node Representation Module 200: Used to extract features from rumor propagation events, obtain attention coefficients between nodes, and encode features using the attention coefficients between nodes to obtain rumor node representations;

[0109] Target attention coefficient module 300: used to extract attention coefficients and feature matrix, obtain node environment mask and edge environment mask through rumor node representation, enhance the feature matrix through node environment mask to obtain enhanced feature matrix, and use edge environment mask to weight the attention coefficients to obtain target attention coefficients;

[0110] Divergence loss module 400: used to obtain causal subgraph representation and environment subgraph representation through the enhanced feature matrix and target attention coefficient, obtain combined features based on causal subgraph representation and environment subgraph representation, and calculate the divergence loss of combined features;

[0111] Rumor Detection Module 500: This module is used to build an identification model, obtain identification results using the model, calculate the identification label loss of the identification results using divergence loss, train the identification model based on the identification label loss, and complete the detection of rumors using the trained identification model.

[0112] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A rumor detection method based on adaptive causal subgraph learning, characterized in that, include: S1: Obtain rumor samples and establish rumor propagation events based on the rumor samples; S2: Extract features from the rumor propagation event to obtain the attention coefficients between nodes, and encode the features using the attention coefficients to obtain the rumor node representation; S3: Extract the attention coefficients and feature matrix, obtain the node environment mask and edge environment mask through the rumor node representation, enhance the feature matrix through the node environment mask to obtain the enhanced feature matrix, and use the edge environment mask to weight the attention coefficients to obtain the target attention coefficients; In step S3, a multilayer perceptron network is constructed, and a node causal mask is obtained through the multilayer perceptron network and the rumor node representation. An edge causal mask is obtained through the node causal mask and the multilayer perceptron network. The node environment mask and the edge environment mask are obtained based on the node causal mask and the edge causal mask. In step S3, attention coefficients and a feature matrix including a causal feature matrix and an environmental feature matrix are extracted from the rumor node representation. The node environment mask, the causal feature matrix, and the environmental feature matrix are used for enhancement to obtain an enhanced feature matrix including an enhanced causal feature matrix and an enhanced environmental feature matrix. The edge environment mask and the attention coefficients are used for weighting and normalization to obtain the target attention coefficients. In this method, a multilayer perceptron network and rumor node representation are used to obtain node causality masks, and the node causality masks of each propagating node are combined to form a node causality mask matrix. By using node causal masks and a multilayer perceptron network, edge causal masks are obtained. The edge causal masks of each propagation node can be combined to form an edge causal mask matrix. Then the node environment mask can be obtained. and edge environment mask : ； S4: Obtain the causal subgraph representation and the environment subgraph representation by enhancing the feature matrix and the target attention coefficient. Obtain the combined feature based on the causal subgraph representation and the environment subgraph representation, and calculate the divergence loss of the combined feature. Among them, mean pooling is performed on the environment node representation to obtain the environment subgraph representation. S5: Build an identification model, use the identification model to obtain the identification results, calculate the identification label loss of the identification results through divergence loss, train the identification model based on the identification label loss, and complete the detection of rumors through the trained identification model.

2. The rumor detection method based on adaptive causal subgraph learning according to claim 1, characterized in that, In step S1, rumor events are identified from the rumor samples, the propagation nodes and propagation paths of the rumor events are extracted, a rumor propagation graph is established using the propagation nodes and propagation paths, and the rumor propagation event is established using the rumor propagation graph.

3. The rumor detection method based on adaptive causal subgraph learning according to claim 1, characterized in that, In step S2, node feature representations are extracted from the rumor propagation event, and the node feature representations are activated and normalized to obtain the attention coefficients between nodes.

4. The rumor detection method based on adaptive causal subgraph learning according to claim 3, characterized in that, In step S2, an encoder is determined, the node feature representation is weighted according to the inter-node attention coefficient, and encoded using the encoder to obtain the rumor node representation.

5. The rumor detection method based on adaptive causal subgraph learning according to claim 1, characterized in that, In step S4, the target attention coefficients and the causal feature matrix are encoded to obtain causal node representations, and mean pooling is performed on the causal node representations to obtain the causal subgraph representations. The target attention coefficient and the environment feature matrix are encoded to obtain the environment node representation. The environment node representation is then subjected to mean pooling to obtain the environment subgraph representation.

6. The rumor detection method based on adaptive causal subgraph learning according to claim 1, characterized in that, In step S4, random noise is determined, and the combined feature is constructed using the random noise, causal subgraph representation, and environmental subgraph representation. The divergence loss of the combined feature and the uniform class distribution is calculated.

7. The rumor detection method based on adaptive causal subgraph learning according to claim 1, characterized in that, In step S5, the cross-entropy loss of the recognition result is calculated, a weighting coefficient is determined, the divergence loss is weighted using the weighting coefficient, and then added to the cross-entropy loss to obtain the recognition label loss.

8. A rumor detection system based on adaptive causal subgraph learning, used to execute the rumor detection method based on adaptive causal subgraph learning as described in any one of claims 1 to 7, characterized in that, include: The rumor propagation event module includes obtaining rumor samples and establishing rumor propagation events based on these samples. Rumor Node Representation Module: Used to extract features from rumor propagation events, obtain attention coefficients between nodes, and encode features using these attention coefficients to obtain rumor node representations; Target attention coefficient module: used to extract attention coefficients and feature matrix, obtain node environment mask and edge environment mask through rumor node representation, enhance the feature matrix through node environment mask to obtain enhanced feature matrix, and use edge environment mask to weight attention coefficients to obtain target attention coefficients; The divergence loss module is used to obtain causal subgraph representations and environmental subgraph representations through the enhanced feature matrix and target attention coefficients, obtain combined features based on the causal subgraph representations and environmental subgraph representations, and calculate the divergence loss of the combined features. Rumor detection module: Used to build an identification model, obtain identification results using the identification model, calculate the identification label loss of the identification results through divergence loss, train the identification model based on the identification label loss, and complete the detection of rumors using the trained identification model.

Images