A method for early rumor detection based on multi-element feature fusion

Abstract

The application belongs to the technical field of rumor detection of social networks, and particularly relates to an early rumor detection method based on multi-feature fusion, which comprises the following steps: obtaining a social network data set to be detected, and extracting user emotional features and user credit features of each user node in a rumor propagation network; using a multi-regression model containing interaction items to model the nonlinear interaction of the user emotional features and the user credit features, and calculating the initial feature influence score of each user node; inputting the initial feature influence score into a two-layer graph attention network for node feature aggregation, and outputting an embedding representation matrix containing propagation topology correlation; setting a time sequence sliding window on a time axis, calculating the cognitive game influence degree of each user node, and constructing a dynamic gain matrix; performing feature fusion and classification discrimination on the embedding representation matrix and the dynamic gain matrix, and outputting early rumor detection results. The application can effectively improve the feature expression capability and detection accuracy in the early rumor scenario.

Description

Technical Field

[0001] This invention belongs to the field of social network rumor detection technology, specifically involving an early rumor detection method based on multi-feature fusion. Background Technology

[0002] With the rapid development of social networking platforms, the speed of online information dissemination has significantly increased, allowing users to quickly spread information through forwarding, commenting, and interacting. In this process, rumors are also more easily amplified through social platforms, causing widespread dissemination and negatively impacting public opinion and perception within a short period. Therefore, how to achieve rapid and accurate identification of rumors in their early stages has become a crucial issue in the governance of social network information.

[0003] Existing rumor detection methods typically model from the perspectives of text content, user attributes, and propagation structure. However, in the early stages of rumor dissemination, the number of users involved is small, the propagation chain has not yet fully developed, and the propagation network structure is relatively sparse, resulting in limited usable data and insufficient feature representation. Furthermore, there are often complex nonlinear interactions between user sentiment features, user credibility features, and propagation structure features. If only a single feature or a simple concatenation method is used for modeling, it is difficult to fully explore the correlations between multiple features, thus affecting the accuracy of the detection results.

[0004] Researchers have proposed various innovative methods in existing rumor detection technologies. For example, one study proposed a detection model that integrates opinion divergence and sentiment game theory. This method extracts features based on topics and users, and combines polynomial equations and three-party cognitive game theory to construct the influence of sentiment interaction. Finally, it uses a Long Short-Term Memory (LSTM) network to complete rumor detection. Another technique uses the BERT model to extract semantic and sentiment features from the main text and comments separately, calculates the sentiment difference vector between the two, and uses a fully connected network to identify and classify the fused features. In addition, there are methods that generate heterogeneous graphs of social networks by preprocessing user and social response texts to achieve early detection and warning of rumors; or input user attributes into a model based on sparse representation and tensor completion, and combine evolutionary game theory and neural networks to predict the spread trend of user participation in rumor or debunking topics. However, the above-mentioned existing technologies still have certain limitations in practical applications. First, user sentiment features and credibility features are usually just directly concatenated as independent inputs, failing to explicitly model the nonlinear interaction relationship between these two types of features. Secondly, existing graph attention networks are mostly limited to topological aggregation of original features, lacking a mechanism for co-designing feature interaction modeling results with propagation topological aggregation. Finally, in terms of temporal dynamic modeling, existing methods generally lack cognitive game influence calculation and dynamic gain calibration mechanisms based on temporal sliding windows, making it difficult to accurately capture the dynamic changes during rumor propagation. At different times, user participation behavior is often influenced by multiple factors such as their historical behavior, social neighbor behavior, and the overall popularity of the topic. Existing methods still fall short in characterizing these dynamic changes in user behavior and suppressing abnormal fluctuations, making it difficult to effectively utilize dynamic gain features to improve early rumor detection.

[0005] Therefore, there is an urgent need to propose an early rumor detection method that can comprehensively utilize multiple features, propagation structure features, and dynamic behavioral features to further improve the accuracy, stability, and reliability of early rumor detection. Summary of the Invention

[0006] The technical problems to be solved by this invention include: First, how to effectively extract and jointly model user emotional characteristics and user credibility characteristics in the early stage of rumor dissemination to improve the ability to express user node features; Second, how to use a multivariate regression model with interaction terms and a graph attention network to collaboratively model the nonlinear interaction relationship and propagation topology relationship between user nodes to obtain a more discriminative node embedding representation; Third, how to combine user historical behavior, social neighbor behavior and overall topic popularity to construct dynamic gain features and fuse them with the propagation structure embedding representation for discrimination, thereby improving the accuracy and stability of early rumor detection.

[0007] To address the aforementioned technical problems, the present invention employs the following technical solution: an early rumor detection method based on multi-feature fusion. In the same detection process, the following six steps are executed sequentially: S1, acquire the social network dataset to be detected, construct a rumor propagation network, and extract user sentiment features and user reputation features of each user node in the rumor propagation network; S2, use a multivariate regression model with interaction terms to perform nonlinear interaction modeling on the user sentiment features and user reputation features extracted in step S1, and calculate the initial feature influence score of each user node; S3, use the initial feature influence score obtained in step S2 as the initial attribute of each user node in the rumor propagation network, input it into a two-layer graph attention network for node feature aggregation, and output the feature influence score. S4. Set a time-series sliding window on the time axis and calculate the cognitive game influence of each user node based on the user's historical behavior participation, the influence strength of the current adversarial behavior of social neighbor nodes, and the overall popularity of the topic at any given time. S5. Calculate the adversarial cognitive gain based on the cognitive game influence obtained in step S4, and compress the adversarial cognitive gain using a calibration mechanism based on a dynamic threshold to construct a dynamic gain matrix. S6. Perform feature fusion and classification on the embedding representation matrix output in step S3 and the dynamic gain matrix constructed in step S5 to output the early rumor detection results of the social network dataset.

[0008] The above six steps are sequentially linked in the same detection process: the user sentiment features and user reputation features extracted in step S1 support the multivariate regression modeling with interaction terms in step S2; the initial feature influence score output in step S2 serves as the input basis for the graph attention network node aggregation in step S3; the embedding representation matrix output in step S3, together with the cognitive game influence degree and dynamic gain matrix constructed in steps S4 and S5, support the feature fusion and classification in step S6, forming a complete technical closed loop of "feature interaction modeling - topology aggregation - temporal dynamic gain - fusion classification".

[0009] Compared with existing technologies, the beneficial effects of this invention are as follows: by extracting user sentiment features and user reputation features, and using a multivariate regression model with interaction terms to characterize the nonlinear interaction relationship between multivariate features, the ability to model user node features can be improved; by introducing a two-layer graph attention network to weighted aggregate the node relationships in the propagation structure, the ability to express propagation topology association information can be enhanced; by combining cognitive game influence, adversarial cognitive gain and dynamic threshold calibration mechanism to construct a dynamic gain matrix, and fusing it with the embedded representation matrix for discrimination, the dynamic behavioral change characteristics in the rumor propagation process can be more fully characterized, thereby effectively improving the accuracy, stability and reliability of early rumor detection. Attached Figure Description

[0010] Figure 1This is an overall flowchart of an early rumor detection method based on multi-feature fusion according to the present invention;

[0011] Figure 2 This is a flowchart of the propagation structure feature aggregation based on graph attention network in this invention;

[0012] Figure 3 This is a flowchart of the dynamic gain calibration and dynamic gain matrix construction process in this invention. Detailed Implementation

[0013] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0014] The symbols are defined as follows:

[0015] User nodes in a rumor-spreading network;

[0016] User node There are neighboring nodes with propagation relationships;

[0017] : Indicates the time index;

[0018] : indicates a rumor-spreading network, in which Represents a set of user nodes. Represents the set of propagation edges;

[0019] User node The emotional characteristics;

[0020] User node The creditworthiness characteristics;

[0021] User node The input feature vector;

[0022] User node Influence score on the initial characteristics of rumor spread;

[0023] Feature weights in a multiple regression model containing interaction terms;

[0024] : The coefficient of the interaction term between sentiment features and credibility features;

[0025] User node Node embedding representation after aggregation via graph attention network;

[0026] User node with neighboring nodes Attention coefficient between them;

[0027] : An embedding representation matrix consisting of the embedding representations of all user nodes;

[0028] : Window length of the time-series sliding window;

[0029] User node At any moment The degree of influence of cognitive game theory;

[0030] User node At any moment Historical behavioral participation;

[0031] User node Social neighbor nodes at any time The intensity of the impact of the current adversarial behavior;

[0032] :time The overall popularity of the topic;

[0033] , , The weight coefficients of each influencing factor in the cognitive game influence;

[0034] User node Counter-cognitive gains;

[0035] :user In the sliding window Inside Normalized cognitive probability at any given moment;

[0036] : The mean of the adversarial cognitive gain within the current sliding window;

[0037] : The fluctuation range of the counter-cognitive gain within the current sliding window;

[0038] : A dynamic gain matrix formed by splicing the adversarial cognitive gains after calibration of each user node;

[0039] The fused representation is obtained by fusing the embedding representation matrix and the dynamic gain matrix;

[0040] : The predicted probability that the topic to be detected is a rumor or not;

[0041] : Represents the cross-entropy loss function.

[0042] An early rumor detection method based on multi-feature fusion, such as Figure 1 As shown, the specific steps include:

[0043] S1: Obtain the social network dataset to be detected, and extract user sentiment features and user credibility features for each user node in the rumor propagation network, including:

[0044] S11: Obtain the original post text, comment text, forwarded text, user attribute information, and propagation relationship information between users corresponding to the topic to be detected. The social network dataset includes the original post text, comment text, forwarded text, user attribute information, and propagation relationship information between users for the topic to be detected. Data can be obtained from public data interfaces of social network platforms or imported from existing public rumor datasets.

[0045] S12. Preprocess the raw data. The preprocessing includes deleting duplicate records, removing invalid user data, completing missing fields, and standardizing the text format to improve the accuracy of subsequent feature extraction, propagation structure modeling, and rumor detection.

[0046] S13: Extract user sentiment characteristics from each user node in the rumor propagation network. With user reputation characteristics Among them, user sentiment features are derived from the viewpoints expressed in user posts and are used to represent the emotional polarity of users in topics related to rumors; user credibility features are constructed based on user authentication status, number of followers, and number of followers and are used to represent the credibility and social influence of users on social platforms.

[0047] S2: Using a multiple regression model with interaction terms, nonlinear interaction modeling is performed on user sentiment features and user reputation features to calculate the initial feature influence score for each user node, including:

[0048] S21: Based on the user sentiment characteristics of user nodes With user reputation characteristics Constructing node input feature vectors The formula is defined as follows:

[0049]

[0050] S22: Utilize a multiple regression model with interaction terms to perform regression calculations on the node input feature vectors, obtaining the initial feature influence score of each user node on rumor propagation. The formula is defined as follows:

[0051]

[0052] S23: In the aforementioned multiple regression model containing interaction terms, This is a bias term used to represent the basic influence level of user nodes; , The first term represents the feature weights, used to indicate the differences in the degree of influence of different features on the spread of rumors; the second term represents the coefficient of the interaction term between sentiment and credibility features, used to measure the joint influence between the two types of features. This is achieved by introducing the interaction term. This model explicitly models the nonlinear relationship between two types of features within the framework of multiple regression. The least squares criterion is used to fit the parameters of the multiple regression model containing interaction terms, and the formula is defined as follows:

[0053]

[0054] S3: Input the initial feature influence scores into a two-layer graph attention network for node feature aggregation, and output an embedding representation matrix containing propagation topological associations, including:

[0055] S31: After constructing the initial representation of the high-dimensional rumor, the initial feature influence score will be learned through a multivariate regression model with interaction terms. As the initial attribute of a node in a rumor propagation network; a graph attention network is used to model the non-uniform information aggregation relationship between users, with the following formula defined:

[0056]

[0057] S32: In the In layer GAT, node pairs Attention coefficient Calculated using the following formula:

[0058]

[0059] S33: Based on the attention coefficient The features of incoming neighbor nodes are weighted and aggregated, and the node feature representation is updated. The formula for the update rule is defined as follows:

[0060]

[0061] S34: Among them, This represents all incoming neighbor nodes of node u; different weights are assigned to different neighbor nodes through an attention mechanism, thereby distinguishing important information from noise information during feature aggregation.

[0062] S35: A two-layer graph attention network is used to model the rumor propagation structure, and the final output is the embedding representation matrix of the nodes in the graph structure. The formula is defined as follows:

[0063]

[0064] Through the above steps, joint modeling of multi-feature fusion and propagation structure based on MLR-GAT was achieved.

[0065] S4: Set a time-series sliding window on the timeline, and calculate the cognitive game influence of each user node based on the user's historical behavior engagement, the strength of the current adversarial behavior of social neighbor nodes, and the overall popularity of the topic at any given moment, including:

[0066] S41: In the spread of rumors, user behavior is dynamic and complex, driven by a variety of heterogeneous factors, and exhibits obvious temporal evolution characteristics; user judgment is influenced by social neighbors, their own early behavior, and the popularity of the topic, thus depicting the temporal dynamic evolution of user behavior from a temporal perspective.

[0067] S42: Interference from social neighbor behavior, early user dissemination behavior, and current topic popularity collectively constitute the dynamic environment of user cognitive evolution within the dissemination network; based on the above interaction mechanism, user nodes are defined. At any moment The Cognitive Game Influence (CGI) is defined by the following formula:

[0068]

[0069] S43: Among them, Indicates user Historical engagement in the topic This indicates the strength of the current adversarial behavior of its social neighbor nodes. express The overall popularity of the topic at any given moment , , The coefficients for adjusting the weights of various factors are used to model their game-theoretic influence preferences.

[0070] S5: Calculate the adversarial cognitive gain based on the cognitive game influence, and compress the adversarial cognitive gain using a dynamic threshold-based calibration mechanism to construct a dynamic gain matrix, including:

[0071] S51: Since the cognitive game influence (CGI) is a time-varying sequence, to quantify its fluctuation trend and characterize the stability of user behavior under information interference, adversarial cognitive gain (ACG) is introduced to measure the degree of cognitive uncertainty in user behavior. The formula is defined as follows:

[0072]

[0073] S52: Among them, This indicates the length of the sliding window. Indicates the user's position within the sliding window. The normalized cognitive probability is defined by the following formula:

[0074]

[0075] S53: After sliding along the time axis, a complete cognitive gain sequence is obtained, and the adversarial cognitive gains of all user nodes are combined into a gain matrix to represent the cognitive fluctuation characteristics of nodes in the time dimension.

[0076] S54: To eliminate anomalous noise in the propagation network caused by extreme disturbances or isolated nodes, a dynamic threshold calibration mechanism based on a time-series sliding window is introduced to compress the adversarial cognitive gain. The calibration process formula is defined as follows:

[0077]

[0078] S55: Among them, This represents the mean of the adversarial cognitive gain within the current sliding window. This represents the fluctuation range of the adversarial cognitive gain within the current sliding window; the calibrated adversarial cognitive gains of all user nodes under the sliding window are concatenated to form a dynamic gain matrix. As shown in the formula.

[0079]

[0080] S6: Perform feature fusion and classification on the embedding representation matrix and dynamic gain matrix to output early rumor detection results for the social network dataset, including:

[0081] S61: Embedding in user structure modeling and behavioral dynamic characteristics Subsequently, a dual-channel convolutional neural network fusion model was constructed to combine the two types of information mentioned above for rumor detection, resulting in a fused representation. As shown in the formula:

[0082]

[0083] S62: Merge representation The input is processed by a fully connected layer for higher-order feature transformation, and finally, a Softmax layer outputs a predicted probability value for whether the current topic is a rumor or not. The formula is defined as follows:

[0084]

[0085] S63. To make the final binary classification decision, a discrimination threshold is set. When the probability prediction value is greater than or equal to the discrimination threshold, the current topic to be detected is determined to be a rumor topic; when the probability prediction value is less than the discrimination threshold, the current topic to be detected is determined to be a non-rumor topic.

[0086]

[0087] S64: To improve the model's ability to fit the classification boundary, the cross-entropy loss function is used. The classification model is optimized using the following formula:

[0088]

[0089] in, For real labels, The smaller the loss value, the better the model's recognition performance.

[0090] In a specific embodiment, a rumor detection task on the social networking platform Weibo is taken as an example. The dataset contains a certain number of rumor and non-rumor topics. Each topic includes the original post text, comments and reposts, user attribute information, and propagation relationships. The GAT hidden layer dimension is set to 128, the number of attention heads is 8, and there are two layers of GAT; the temporal sliding window length is... Set to 5; Cognitive game influence weighting coefficient , , , , , ;Discrimination threshold The optimization factor is set to 0.5; the optimizer is Adam, the learning rate is set to 0.001, and the training epochs are 100. The above parameter settings are merely examples of an embodiment and are not intended to limit the scope of the claims. Experimental results show that the present invention outperforms traditional feature concatenation methods, pure graph attention network methods, and baseline methods based on LSTM in terms of accuracy, precision, and F1 score for early rumor detection, verifying the effectiveness of the three-stage collaborative mechanism.

Claims

1. An early rumor detection method based on multi-feature fusion, characterized in that, Perform the following steps sequentially within the same testing process: S1. Obtain the social network dataset to be detected, and extract the user sentiment features and user credibility features of each user node in the rumor propagation network; S2. Using a multivariate regression model with interaction terms, nonlinear interaction modeling is performed on the user sentiment features and user reputation features extracted in step S1 to calculate the initial feature influence score of each user node. S3. The initial feature influence score obtained in step S2 is used as the initial attribute of each user node in the rumor propagation network. It is input into a two-layer graph attention network to perform node feature aggregation and output an embedding representation matrix containing propagation topology associations. S4. Set a time-series sliding window on the timeline, and calculate the cognitive game influence of each user node based on the user's historical behavior participation, the current adversarial behavior influence of social neighbor nodes, and the overall popularity of the topic at any given moment. S5. Calculate the adversarial cognitive gain based on the cognitive game influence obtained in step S4, and compress the adversarial cognitive gain using a calibration mechanism based on dynamic thresholds to construct a dynamic gain matrix. S6. Perform feature fusion and classification on the embedding representation matrix output in step S3 and the dynamic gain matrix constructed in step S5, and output the early rumor detection results of the social network dataset. Specifically, the user sentiment features and user reputation features extracted in step S1 support the multivariate regression modeling with interaction terms in step S2; the initial feature influence score output in step S2 serves as the input basis for the graph attention network node aggregation in step S3; the embedding representation matrix output in step S3, together with the cognitive game influence degree and dynamic gain matrix constructed in steps S4 and S5, support the feature fusion and classification in step S6, forming a complete technical closed loop of "feature interaction modeling - topology aggregation - temporal dynamic gain - fusion classification".

2. The method for early rumor detection based on multi-feature fusion according to claim 1, characterized in that, Step S1: Obtain the social network dataset to be detected, construct the rumor propagation network, and extract the user sentiment features and user credibility features of each user node in the rumor propagation network, including: (a) Obtain the original post text, comment text, forwarded text, user attribute information, and propagation relationship information between users corresponding to the topic to be detected; (b) Construct a rumor propagation network using users as nodes and forwarding or comment propagation relationships as edges; (c) Extract user sentiment features and user reputation features for each user node. The user sentiment features are derived from the viewpoints expressed in the user’s text, and the user reputation features are constructed based on the user’s authentication status, number of followers, and number of followers.

3. The method for early rumor detection based on multi-feature fusion according to claim 1, characterized in that, Step S2 uses a multiple regression model with interaction terms to perform nonlinear interaction modeling on the user sentiment features and user reputation features extracted in step S1, and calculates the initial feature influence score for each user node, including: (a) Combine the user sentiment features and user reputation features of each user node into a node input feature vector; (b) The node input feature vectors are regressed using a multivariate regression model with interaction terms to obtain the initial feature influence scores of each user node on the spread of rumors. (c) Introduce an interaction term between user sentiment features and user reputation features into a multiple regression model containing an interaction term. To achieve explicit modeling of the nonlinear relationship between two types of features; (d) The least squares criterion is used to fit the parameters of the multivariate regression model with interaction terms to obtain the corresponding bias terms, feature weights and interaction term coefficients.

4. The early rumor detection method based on multi-feature fusion according to claim 1, characterized in that, Step S3 inputs the initial feature influence score into a two-layer graph attention network for node feature aggregation, and outputs an embedding representation matrix containing propagation topological associations, including: (a) The initial feature influence score is used as the initial attribute of each user node in the rumor propagation network; (b) Input the initial attributes of each user node into the graph attention network; (c) Calculate the attention coefficients between the target node and its neighboring nodes in the graph attention network; (d) Perform weighted aggregation of the features of the neighboring nodes of the target node based on the attention coefficient, and update the feature representation of the target node; (e) A two-layer graph attention network is used to model the rumor propagation structure and output an embedding representation matrix containing the propagation topology.

5. The method for early rumor detection based on multi-feature fusion according to claim 1, characterized in that, Step S4 sets up a time-series sliding window on the timeline. Based on the user's historical behavior engagement, the strength of the current adversarial behavior of social neighbor nodes, and the overall popularity of the topic at any given moment, the cognitive game influence of each user node is calculated, including: (a) Set a timing sliding window on the timeline; (b) Calculate the cognitive game influence of each user node at each moment based on the user's historical behavior participation, the current adversarial behavior of social neighbor nodes, and the overall popularity of the topic at any given time. (c) The cognitive dynamics of each user node during the rumor propagation process are characterized by the influence of cognitive game theory.

6. The early rumor detection method based on multi-feature fusion according to claim 1, characterized in that, Step S5 calculates the adversarial cognitive gain based on the cognitive game influence degree, and compresses the adversarial cognitive gain using a dynamic threshold-based calibration mechanism to construct a dynamic gain matrix, including: (a) Calculate the adversarial cognitive gain based on the cognitive game influence degree and normalized cognitive probability of each user node; (b) Compression of adversarial cognitive gain using a dynamic threshold-based calibration mechanism; (c) The compressed user nodes are spliced ​​together to construct a dynamic gain matrix.

7. The method for early rumor detection based on multi-feature fusion according to claim 1, characterized in that, Step S6 performs feature fusion and classification on the embedding representation matrix and dynamic gain matrix, outputting early rumor detection results for the social network dataset, including: (a) Input the embedding representation matrix and the dynamic gain matrix into a dual-channel convolutional neural network fusion model to perform feature fusion and obtain the fused representation; (b) Input the fused representation into the fully connected layer for high-order feature transformation, and output the probability prediction value of whether the topic to be detected belongs to rumors or not through Softmax; (c) Convert the probability prediction value into a binary label according to the preset discrimination threshold; (d) Output early rumor detection results for the social network dataset.

8. The method for early rumor detection based on multi-feature fusion according to claim 7, characterized in that, When the probability prediction value is greater than or equal to the preset discrimination threshold, the current topic to be detected is determined to be a rumor topic; when the probability prediction value is less than the preset discrimination threshold, the current topic to be detected is determined to be a non-rumor topic.

9. The method for early rumor detection based on multi-feature fusion according to claim 7, characterized in that, The cross-entropy loss function is used to optimize the dual-channel convolutional neural network fusion model to improve the model's ability to fit the classification boundary.

Images