A multi-task rumor detection model and method based on a dynamic network

Abstract

The application provides a multi-task rumor detection model and method based on a dynamic network, which can capture dynamic information and improve detection effect. The application combines dynamic spatial features and time sequence text features captured over time by taking a time step as a basic sharing unit, and uses an attention mechanism to update output features, thereby forming a multi-task rumor detection model and method based on a dynamic network. First, a rumor event propagation graph at different times is constructed, and a dynamic graph convolution network is used to model the propagation graph at different times to capture dynamic structural features of the rumor event. Then, a time sequence sharing layer is used as a shared network to interact features of the event dynamic structural features and the time sequence text features at each same time step. Finally, an attention mechanism is used to focus on information having an important influence between different tasks and different times, as an input of a classifier, to perform prediction.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and in particular to a multi-task rumor detection model and method based on dynamic networks. Background Technology

[0002] In recent years, the Internet and social media platforms have gradually become the main way for people to obtain news, but at the same time, they have also brought great convenience to the spread of rumors. The social panic and serious consequences caused by rumors have attracted the attention of existing researchers.

[0003] Traditional research focuses more on text content and social context, largely ignoring various factors and the implicit relationship between social media and the spread of rumors. However, with the development of artificial intelligence, deep learning has been widely applied in rumor detection. Researchers are beginning to model rumor propagation patterns as propagation trees or graphs, using structural similarity to distinguish between rumors and non-rumors. For example, Vaibhav V et al. (Do sentence interactions matter? leveraging sentence level representations for fake news classification[J]. arXiv preprint arXiv:1910.12203,2019) modeled news samples as graphs, transforming sentence classification into graph classification to ultimately identify rumors. However, this method of constructing propagation patterns is static; it only reflects the propagation pattern of rumors in their later stages or after they have ended. In the real world, the evolution of rumors is dynamic and changes over time, and this method cannot capture the dynamic characteristics of rumors as they evolve.

[0004] Dynamic graph neural networks (RNNs) can effectively address the aforementioned problems. Over time, several detection methods based on dynamic principles have been proposed. For example, Ma et al. (Ma J, Gao W, Mitra P, et al. Detecting rumors from microblogs with recurrent neural networks[J]. 2016) constructed a set of time-continuous static graph sequences by building static graphs of the same propagation event at different time stages. The Chinese patent "A Multi-Task Rumor Detection Method Based on Bidirectional Propagation Graphs" (application number: 202110454550.0) treats text content as sequence data and uses RNNs to capture the changing contextual information between different texts in the sequence over time. Although these models have achieved good detection results, their performance is limited by the number of time steps. When the number of time steps is too small, good dynamic effects cannot be achieved. Summary of the Invention

[0005] The purpose of this invention is to overcome the above problems and propose a multi-task rumor detection model and method based on dynamic networks that can capture dynamic information and improve detection performance. This invention organically combines the captured dynamic spatial features and temporal text features that change over time by using time steps as the basic shared unit, and uses an attention mechanism to update the output features, thereby forming a multi-task rumor detection model and method based on dynamic networks.

[0006] First, a propagation graph of rumor events at different times is constructed. A dynamic graph convolutional network is used to model the propagation graph at different times to capture the dynamic structural features of rumor events. Then, a temporal sharing layer is used as a shared network to perform feature interaction between the dynamic structural features of the event and the temporal text features at each same time step. Finally, an attention mechanism is used to focus on information that has an important impact on different tasks and different times, and use it as input to the classifier for prediction.

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a multi-task rumor detection model and method based on dynamic networks, specifically including the following steps:

[0008] S1. Construct a dynamic graph of rumor propagation: Dynamic-GCN. Design a fine-grained dynamic graph neural network based on time steps. It uses a dynamic graph neural network to construct the rumor event propagation graph at different times. It can enhance the efficiency of multi-task sharing by capturing the dynamic structural features of rumor events with smoother transitions.

[0009] like Figure 1The Dynamic-GCN module is located at the bottom center. The spread of rumors evolves dynamically over time. For the same event, each time step represents a rumor propagation network. As time progresses, the number of these networks dynamically increases, eventually forming a dynamic propagation network. Based on the rule that the propagation network dynamically changes over time, a corresponding propagation graph is constructed at each time step. Nodes represent different posts, and edges represent the forwarding and replying relationships between these posts. Specifically, we define the fine-grained dynamic graph based on time steps as: a fine-grained dynamic pattern where only one node propagates at each time step.

[0010] S2. To capture temporal text features, a text content extraction network based on LSMT, Content-LSTM, was designed. It is an improvement on the LSTM network as a baseline model to capture temporal text features that change over time and to use it for shared interaction in multiple tasks to improve model performance.

[0011] First, a time-series-based structural feature H is obtained through fine-grained dynamic graph convolution computation based on time steps. i ={H i0 H i1 ,...,H ir The difference between adjacent time steps lies in the generation of a certain node (post), and their arrangement follows a strict temporal order. To ensure that this dynamic nature of the dynamic graph is related to the development of Content-LSTM in time steps, thus meeting the requirements of time-step-based integration in TES, this invention uses the node diffusion order as the arrangement order of the corresponding rumor event text content. Text content with the same order is modeled to form a set of time-step-based text feature sequences, which are used for the calculation of the shared layer of subsequent time-series events.

[0012] S3. Construct a time-series event sharing layer and design a time-series event-based sharing method (TES) as the sharing layer. The time step is used as the basic sharing unit to realize the sharing and interaction of dynamic structural features and time-series text features between the first two modules. This can reduce noise interference and effectively improve the prediction effect.

[0013] Based on the aforementioned rules governing the dynamic changes of the propagation network over time, and the temporal order of node generation, the dynamic structural feature H was obtained. i and time series features H content They are related in the sequence of events of propagation; in other words, for a propagation event at the same time stage, its temporal and spatial characteristic sequences are captured respectively, and their temporal and dynamic characteristics are consistent in the development process of the post. In this invention, as... Figure 1In the middle section, a temporal event sharing layer is proposed. It combines Dynamic-GCN and Content-LSTM with time steps as the basic computational unit, enabling interactive sharing and joint training of features from both at the same time stage. Based on the hardware parameter sharing approach in multi-task learning mentioned in the paper "All-in-one: Multi-task learning for rumour verification[J].arXiv preprintarXiv:1806.03713,2018.", temporal event sharing is realized, thereby improving the sharing frequency and efficiency of Dynamic-GCN and Content-LSTM.

[0014] S4. Attention Mechanism Feature Extraction and Classification: Through the preceding calculations and the calculations using TES, the feature representations after feature sharing are obtained, namely the feature representations of Dynamic-GCN, Content-LSTM, and TES layers: H ir H content H share To understand which features have a more significant impact on prediction across different time series and tasks, this invention uses a multi-head attention mechanism to update them, obtaining an updated final event feature representation H for prediction. content-t H i-t Finally, the data are input into a classifier for calculation to predict rumors.

[0015] Preferably, the problem is defined before step S1, and the rumor detection task is defined as a binary classification problem, where ε = {ε1,...,ε}. n} is a set of examples of rumor detection events, ε i Let n be the i-th event and n be the number of events; represent the set of posts and comments as... s i It's the source tweet, c ij This is the j-th comment text; m i It is event ε i The number of posts in the text; if the source tweet is considered as the 0th tweet, then... It can be represented as To propagate an event ε i The order in which each post spreads is represented chronologically. t represents time, and ε represents the final event of a rumor's spread. i It can be represented as

[0016] Each post in the propagation event ε iDivided into γ time steps along the time dimension, where γ is the number of posts m for each propagation event. i Determine, γ∈{1,2,...,γ}, for a propagation event ε i At different time steps γ, there are different propagation states; at the first time step ε... i0 ={(c i0 ,t i0 )}, in the second time step ε i1 ={(c i0 ,t i0 ),(c i1 ,t i1 Therefore, event ε i The final expression at all time steps is ε i ={ε i0 ,ε i1 ,...,ε iγ};

[0017] For the same propagation event ε i Construct different propagation structure diagrams at different time steps. Where ε i Represents the set of tweets in the i-th event. V represents the set of tweets in the i-th event, γ represents the time step of different tweets in the i-th event, and V ir E ir These represent the node set and edge set at time γ in the i-th event, respectively. This represents the propagation diagram at time γ in the i-th event.

[0018] Preferably, in step S1, for the propagation event ε i At time t0, only one source post exists at node0. At time t1, two responding tweets interact with the source post, namely node 1 and node 2. At time t2, only one tweet responds to the source post at node 3. At time t3, node 4 and node 5 simultaneously respond to node 1. Based on this relationship, a coarse-grained dynamic graph state is initially obtained. However, at some times, multiple nodes propagate simultaneously, and now a finer-grained partitioning is needed.

[0019] In subsequent calculations, to enable the dynamic graph network to perform time-step-based feature interactions in the time-series event sharing module, this invention sorts multiple nodes propagating at the same time, ultimately forming a single time step where only the state of one node is propagated, such as... Figure 3As shown. For node1 and node2, which propagate simultaneously at time t1, although their creation times are the same, they are sorted according to the order of their node numbers. Finally, based on the above node propagation rules and their response and forwarding node numbers, static graphs are constructed at different time steps, creating a fine-grained dynamic graph sequence based on time steps for subsequent dynamic graph convolution operations.

[0020] Preferably, the coarse-grained dynamic graph states are further subdivided to construct a time-step-based fine-grained dynamic graph sequence.

[0021] For propagation event ε i ={ε i0 ,ε i1 ,...,ε iγ Construct propagation graph snapshots at different times. It contains a propagation graph The feature matrix of all nodes, m ir Let F be the number of nodes in the propagation graph at that moment, and let F represent the eigenvectors of the nodes; its adjacency matrix is ​​defined as... As shown in equations (1) and (2), the dynamic graph convolutional layer is defined as:

[0022]

[0023] in, For degree matrix, It is an adjacency matrix. For the propagation diagram of the upper layer The feature matrix of all nodes, W l Let σ be the weight matrix, and σ be the nonlinear activation function.

[0024] Propagation graph for all time steps Perform the operations in equation (1) to obtain the spatial features of all time steps:

[0025] H i ={H i0 H i1 ,...,H ir} (2)

[0026] At this point, the propagation structure characteristics of the same propagation event at each moment have been obtained, forming a dynamic feature sequence H. i The dynamic structural feature sequence is taken as one of the inputs in the subsequent time-series event sharing layer calculation; at the same time, the dynamic structural feature H at the last moment is used as one of the inputs in the calculation. ir As the final output of the dynamic graph convolutional network module, it serves as the input for subsequent classification network computation.

[0027] The dynamic structural feature H was obtained through time-step-based fine-grained dynamic graph convolution calculation. i ={H i0 H i1 H ir It is a time-series-based structural feature embedding where the difference between adjacent time steps lies in the generation of a certain node (post), and their arrangement follows a strict temporal order. To ensure that the dynamism of the dynamic graph is related to the development of the Content-LSTM over time steps, thus meeting the requirement of combining them based on time steps in the time-series event sharing layer, this invention uses the node diffusion order during the construction of the dynamic graph as the arrangement order of the corresponding rumor event text content. Text content with the same order is modeled to form a set of time-step-based text feature sequences for subsequent computation.

[0028] Preferably, step S2 uses an LSTM network as a baseline model to improve it into a new model, Content-LSTM, for propagating event ε. i All included tweets are represented as Used as input to Content-LSTM, as in equations (3)-(8):

[0029] i t =σ(W ii X i_content_t +b it +W hi h (t-1) +b hi (3)

[0030] f t =σ(W if X i_content_t +b if +W hf h (t-1) +n hf (4)

[0031] g t =tanh(W ig X i_content_t +b ig +W hg h (t-1) +b hg (5)

[0032] o t =σ(W io X i_content_t +b io +W ho h (t-1) +b ho (6)

[0033] c t =f t *c (t-1) +i t *g t (7)

[0034] h content-t =o t *tanh(c t (8)

[0035] Where X i_content_t h is the feature vector input at a certain time. (t-1) W is the hidden state from the previous moment. * b * Here are the corresponding weight matrix and bias.

[0036] At this point, the text feature sequences of the same propagation event at each moment form a time feature sequence H. content As shown in equation (9), it is denoted as:

[0037] H content ={h content-0 h content-1 , ..., h content-t} (9)

[0038] Time series features H content With the dynamic structural sequence feature H in formula (2) i The order of propagation events is one-to-one, and this time series feature is used as one of the inputs for subsequent time series event sharing layer calculations; at the same time, the sequence feature h at the last moment is also used. content-t The final output of Content-LSTM is used as input in subsequent classification network computations.

[0039] Based on the preceding calculations, the dynamic structural feature H was obtained. i and time series features H content Based on the consistency of their temporal and dynamic characteristics in the development process of posts, this invention designs a temporal event sharing layer. Using time steps as the basic unit of calculation, it combines dynamic graph features and rumor text content features within the same time phase for interactive sharing and joint training. Hardware parameter sharing is used to achieve temporal event sharing, thereby improving the sharing frequency and efficiency. The specific calculation steps are as follows:

[0040] Preferably, step S3 includes the following steps:

[0041] S31. At time t0, take the dynamic graph and construct the dynamic structural features during the construction process. The corresponding Node_0 node feature that is currently propagating. and time series features H content- Here, the propagation structure and content features of the corresponding tweet with ID 0 are taken simultaneously.

[0042] S32. The propagation structure and content features of the tweets are spliced ​​together and mapped through a linear layer. This is used as the input to the cell block at time t0 in the time-series event sharing layer. After gating calculation, the output of the first time step is obtained, as shown in equations (10)-(16):

[0043] X share-0 =Linear(Concat(H i0 H content-0 (10)

[0044] i t =σ(W ii X share-0 +b it +W hi h (t-1) +b hi (11)

[0045] f t =σ(W if X share +b if +W hf h (t-1) +n hf (12)

[0046] g t =tanh(W ig X share-0 +b ig +W hg h (t-1) +b hg (13)

[0047] o t =σ(W io X share +b io +W ho h (t-1) +b ho (14)

[0048] c t =f t *c (t-1) +i t *g t (15)

[0049] H share =o t *tanh(ct (16)

[0050] In the above formula, X sha，e-0 It is the input of the time-series event sharing layer at time t0, W * b * It is the weight matrix and the bias, h (t-1) It is the hidden layer vector from the previous time step, randomly initialized and generated at time t0, H share-0 It is the output;

[0051] At this point, a similar operation is being performed in the first cell block of the Content-LSTM, except that the input is the text content representation at time t0, and the output is the time series feature h at time t0. content-0 ;

[0052] S33, splicing h content- and H share-0 And through linear mapping, the hidden state of Content-LSTM at time t0 is updated as the hidden state of the Content-LSTM cell block at the next time step, as shown in equation (17):

[0053] h content-0 =Linear(Concat(h) content-0 H share-0 (17)

[0054] The calculation process of Content-LSTM at the next moment is based on equations (3)-(8) in the process of extracting text content features of rumor events;

[0055] S34. Repeat S31 to S33 until all time steps are completed, and finally obtain the output of the time sequence event sharing layer at each time step, as shown in equation (18):

[0056] H share ={H share-0 H share-1 H share-t} (18).

[0057] It is important to note that during the computation process, the time-series event sharing layer also continuously interacts with information from the dynamic graph construction process and the rumor event text content extraction process in chronological order, thereby achieving time-series event sharing.

[0058] The preceding calculations yielded feature representations after feature sharing, namely the feature representations of Dynamic-GCN, Content-LSTM, and TES layers: H ir H content H shareTo understand which features have a more significant impact on prediction across different time series and tasks, this invention employs a multi-head attention mechanism for feature extraction. LSTM networks for rumor detection typically use the latent feature representation of the last time series as the predicted feature vector because the sequence data is processed through a gating mechanism, and the output of the last gating operation contains the hidden states of all preceding gating operations. Therefore, this invention uses the features of the last time series for computation to obtain the convolutional features H of the dynamic graph convolutional network at the final time step. it This will be used for subsequent calculations. The specific implementation details are as follows:

[0059] Preferably, step S4 specifically includes the following steps:

[0060] S41, H sha respectively with H content-t and H i-t The inputs X1 and X2 of the attention layer are concatenated and mapped through a linear layer, as shown in equations (19-20):

[0061] x1 = Linear(Concat(H) content-t H sha (19)

[0062] X2 = Linear(Concat(H) i-t H sha (20)

[0063] S42, X1, and X2 are linearly transformed into three projections with the same dimensions as themselves, namely query Q, key K, and value V, which serve as the inputs to the attention layer, as shown in equations (21)-(23).

[0064] MultiHead(Q,K,V)=Concat(head1,...,head h W O (twenty one)

[0065]

[0066]

[0067] In the above formula, MultiHead(Q, K, V) is the output of the attention layer, and Q, K, and V represent respectively. W O It is the weight matrix, W i It is the initialization of different multi-head matrices, d k It is the dimension of K; when K is large, the output is scaled.

[0068] S43. Connect the output of the attention layer with H respectively.share and H i-t They are concatenated and linearly mapped to update them, as in equations (24)-(25).

[0069] H content-t =Linear(Concat(H sha H content-t )) (twenty four)

[0070] H i-t =Linear(Concat(H share-t H i-t (25)

[0071] S44. Obtain the updated final event feature representation H used for prediction. content-t H i-t For each of them, a corresponding fully connected layer and activation layer are designed, as shown in equations (26)-(27).

[0072] Y C =softmax(w C H content-t +b C (26)

[0073] Y DG =softamx(w DG H i-t +b DG (27)

[0074] w and b are their respective weights and bias parameters. Finally, the cross-entropy loss function is used as the classification loss, as shown in equations (28)-(29).

[0075]

[0076]

[0077] Y i It is the true label of the i-th event;

[0078] During training, the two loss functions are assigned corresponding coefficients, added together to obtain the total loss function, and then backpropagated, as shown in equation (30).

[0079] Loss = W1 * L C +W2 * L DG (30)

[0080] Preferably, model training is also included.

[0081] This invention provides a multi-task rumor detection model and method based on dynamic networks, which has the following beneficial effects:

[0082] 1. This invention proposes a network called DGCN-TES, which models dynamic spatial structure, temporal structure, and text content information under a unified architecture. DGCN-TES includes a temporally based fine-grained dynamic graph convolutional layer and a text content-based LSTM layer for processing spatial and content information that changes over time.

[0083] 2. This invention designs a temporal event sharing module to share and interact with features from different time sequences in two tasks. Utilizing an attention mechanism, it focuses on features beneficial to prediction across different tasks and time sequences, thereby improving network performance.

[0084] 3. This invention received positive feedback on two publicly available datasets. Experimental results show that the proposed model outperforms other methods and performed well in early detection studies. Furthermore, to explore the impact of different temporal granularities on the performance of dynamic networks, additional experiments were conducted. The results revealed that a smoother temporal granularity is more suitable for dynamic networks. Attached Figure Description

[0085] The present invention will be further described below with reference to the accompanying drawings and embodiments:

[0086] Figure 1 This is a diagram of the overall structure of the model of the present invention;

[0087] Figure 2 It is a coarse-grained dynamic graph based on time steps;

[0088] Figure 3 It is a fine-grained dynamic graph based on time steps;

[0089] Figure 4 This is a structural diagram of the timing event sharing layer;

[0090] Figure 5 This is a diagram of the model training process. Detailed Implementation

[0091] like Figure 1 As shown, a multi-task rumor detection model and method based on dynamic networks specifically includes the following steps:

[0092] S1. Construct a dynamic graph of rumor propagation: Dynamic-GCN. Design a fine-grained dynamic graph neural network based on time steps. It uses a dynamic graph neural network to construct the rumor event propagation graph at different times. It can enhance the efficiency of multi-task sharing by capturing the dynamic structural features of rumor events with smoother transitions.

[0093] like Figure 1The Dynamic-GCN module is located at the bottom center. The spread of rumors evolves dynamically over time. For the same event, each time step represents a rumor propagation network. As time progresses, the number of these networks dynamically increases, eventually forming a dynamic propagation network. Based on the rule that the propagation network dynamically changes over time, a corresponding propagation graph is constructed at each time step. Nodes represent different posts, and edges represent the forwarding and replying relationships between these posts. Specifically, we define the fine-grained dynamic graph based on time steps as: a fine-grained dynamic pattern where only one node propagates at each time step.

[0094] S2. To capture temporal text features, a text content extraction network based on LSMT, Content-LSTM, was designed. It is an improvement on the LSTM network as a baseline model to capture temporal text features that change over time and to use it for multi-task sharing and interaction to improve model performance.

[0095] First, a time-series-based structural feature H is obtained through fine-grained dynamic graph convolution computation based on time steps. i ={H i0 H i1 ,...,H ir The difference between adjacent time steps lies in the generation of a certain node (post), and their arrangement follows a strict temporal order. To ensure that this dynamic nature of the dynamic graph is related to the development of Content-LSTM in time steps, thus meeting the requirements of time-step-based integration in TES, this invention uses the node diffusion order as the arrangement order of the corresponding rumor event text content. Text content with the same order is modeled to form a set of time-step-based text feature sequences, which are used for the calculation of the shared layer of subsequent time-series events.

[0096] S3. Construct a time-series event sharing layer and design a time-series event-based sharing method (TES) as the sharing layer. The time step is used as the basic sharing unit to realize the sharing and interaction of dynamic structural features and time-series text features between the first two modules. This can reduce noise interference and effectively improve the prediction effect.

[0097] Based on the aforementioned rules governing the dynamic changes of the propagation network over time, and the temporal order of node generation, the dynamic structural feature H was obtained. i and time series features H content They are related in the sequence of events of propagation; in other words, for a propagation event at the same time stage, its temporal and spatial characteristic sequences are captured respectively, and their temporal and dynamic characteristics are consistent in the development process of the post. In this invention, as... Figure 1In the middle section, a temporal event sharing layer is proposed. It combines Dynamic-GCN and Content-LSTM with time steps as the basic computational unit, enabling interactive sharing and joint training of features from both at the same time stage. Based on the hardware parameter sharing approach in multi-task learning mentioned in the paper "All-in-one: Multi-task learning for rumour verification[J].arXiv preprintarXiv:1806.03713,2018.", temporal event sharing is realized, thereby improving the sharing frequency and efficiency of Dynamic-GCN and Content-LSTM.

[0098] S4. Attention Mechanism Feature Extraction and Classification: Through the preceding calculations and the calculations using TES, the feature representations after feature sharing are obtained, namely the feature representations of Dynamic-GCN, Content-LSTM, and TES layers: H ir H content H shar To understand which features have a more significant impact on prediction across different time series and tasks, this invention uses a multi-head attention mechanism to update them, obtaining an updated final event feature representation H for prediction. content-t H i-t Finally, the data are input into a classifier for calculation to predict rumors.

[0099] Preferably, the problem is defined before step S1, and the rumor detection task is defined as a binary classification problem, where ε = {ε1,...,ε}. n} is a set of examples of rumor detection events, ε i Let n be the i-th event and n be the number of events; represent the set of posts and comments as... s i It's the source tweet, c ij This is the j-th comment text; m i It is event ε i The number of posts in the text; if the source tweet is considered as the 0th tweet, then... It can be represented as To propagate an event ε i The order in which each post spreads is represented chronologically. t represents time, and ε represents the final event of a rumor's spread. i It can be represented as

[0100] Each post in the propagation event ε iDivided into γ time steps along the time dimension, where γ is the number of posts m for each propagation event. i Determine, γ∈{1,2,...,γ}, for a propagation event ε i At different time steps γ, there are different propagation states; at the first time step ε... i0 ={(c i0 ,t i0 )}, in the second time step ε i1 ={(c i0 ,t i0 ),(c i1 ,t i1 Therefore, event ε i The final expression at all time steps is ε i ={ε i0 ,ε i1 ,...,ε iγ};

[0101] For the same propagation event ε i Construct different propagation structure diagrams at different time steps. Where ε i Represents the set of tweets in the i-th event. V represents the set of tweets in the i-th event, γ represents the time step of different tweets in the i-th event, and V ir E ir These represent the node set and edge set at time γ in the i-th event, respectively. This represents the propagation diagram at time γ in the i-th event.

[0102] Preferred, such as Figure 2 As shown, in step S1, for the propagation event ε i At time t0, only one source post exists at node 0. At time t1, two responding tweets interact with the source post, namely node 1 and node 2. At time t2, only one tweet responds to the source post at node 3. At time t3, node 4 and node 5 simultaneously respond to node 1. Based on this relationship, we can initially derive the following: Figure 2 The state is a coarse-grained dynamic graph. However, at certain moments, multiple nodes propagate simultaneously, so a finer-grained partitioning is now needed.

[0103] In subsequent calculations, to enable the dynamic graph network to perform time-step-based feature interactions in the time-series event sharing module, this invention sorts multiple nodes propagating at the same time, ultimately forming a single time step where only the state of one node is propagated, such as... Figure 3As shown. For node1 and node2, which propagate simultaneously at time t1, although their creation times are the same, they are sorted according to the order of their node numbers. Finally, based on the above node propagation rules and their response and forwarding node numbers, static graphs are constructed at different time steps, creating a fine-grained dynamic graph sequence based on time steps for subsequent dynamic graph convolution operations.

[0104] Preferably, the coarse-grained dynamic graph states are further subdivided to construct a time-step-based fine-grained dynamic graph sequence.

[0105] For propagation event ε i ={ε i0 ,ε i1 ,...,ε iγ Construct propagation graph snapshots at different times. It contains a propagation graph The feature matrix of all nodes, m ir Let F be the number of nodes in the propagation graph at that moment, and let F represent the eigenvectors of the nodes; its adjacency matrix is ​​defined as... As shown in equations (1) and (2), the dynamic graph convolutional layer is defined as:

[0106]

[0107] in, For degree matrix, It is an adjacency matrix. For the propagation diagram of the upper layer The feature matrix of all nodes, W l Let σ be the weight matrix, and σ be the nonlinear activation function.

[0108] Propagation graph for all time steps Perform the operations in equation (1) to obtain the spatial features of all time steps:

[0109] H i ={H i0 H i1 ,...,H ir} (2)

[0110] At this point, the propagation structure characteristics of the same propagation event at each moment have been obtained, forming a dynamic feature sequence H. i The dynamic structural feature sequence is taken as one of the inputs in the subsequent time-series event sharing layer calculation; at the same time, the dynamic structural feature H at the last moment is used as one of the inputs in the calculation. ir As the final output of the dynamic graph convolutional network module, it serves as the input for subsequent classification network computation.

[0111] The dynamic structural feature H was obtained through time-step-based fine-grained dynamic graph convolution calculation. i ={H i0 H i1 ,...,H ir It is a time-series-based structural feature embedding where the difference between adjacent time steps lies in the generation of a certain node (post), and their arrangement follows a strict temporal order. To ensure that the dynamism of the dynamic graph is related to the development of the Content-LSTM over time steps, thus meeting the requirement of combining them based on time steps in the time-series event sharing layer, this invention uses the node diffusion order during the construction of the dynamic graph as the arrangement order of the corresponding rumor event text content. Text content with the same order is modeled to form a set of time-step-based text feature sequences for subsequent computation.

[0112] Preferably, step S2 uses an LSTM network as a baseline model to improve it into a new model, Content-LSTM, for propagating event ε. i All included tweets are represented as Used as input to Content-LSTM, as in equations (3)-(8):

[0113] i t =σ(W ii X i_content_t +b it +W hi h (t-1) +b hi (3)

[0114] f t =σ(W if X i_content_t +b if +W hf h (t-1) +n hf (4)

[0115] g t =tanh(W ig X i_content_t +b ig +W hg h (t-1) +b hg (5)

[0116] o t =σ(W io X i_content_t +b io +W ho h (t-1) +b ho (6)

[0117] c t =f t *c (t-1) +i t *g t (7)

[0118] h content- =o t *tanh(c t (8)

[0119] Where X i_content_t h is the feature vector input at a certain time. (t-1) W is the hidden state from the previous moment. * b * Here are the corresponding weight matrix and bias.

[0120] At this point, the text feature sequences of the same propagation event at each moment form a time feature sequence H. content As shown in equation (9), it is denoted as:

[0121] H content ={h content-0 h content-1 , ..., h content-t} (9)

[0122] Time series features H content With the dynamic structure sequence feature H in formula (2) i The order in which the propagation events occur is one-to-one, and this time series feature is taken as one of the inputs for the subsequent time series event sharing layer calculation; at the same time, the sequence feature h of the last moment is also used. content-t The final output of Content-LSTM is used as input in subsequent classification network computations.

[0123] Based on the preceding calculations, the dynamic structural feature H was obtained. i and time series features H content Based on the consistency of their temporal and dynamic characteristics in the development process of posts, this invention designs a temporal event sharing layer. Using time steps as the basic unit of calculation, it combines dynamic graph features and rumor text content features within the same time phase for interactive sharing and joint training. Hardware parameter sharing is used to achieve temporal event sharing, thereby improving the sharing frequency and efficiency. The specific calculation steps are as follows:

[0124] Preferably, step S3 includes the following steps:

[0125] S31, such as Figure 4 As shown, at time t0, dynamic structural features during the construction process are constructed by taking dynamic graphs. The corresponding Node_0 node feature that is currently propagating. and time series features H content-0 Here, the propagation structure characteristics and content characteristics of the corresponding tweet with number 0 are taken at the same time.

[0126] S32. The propagation structure and content features of the tweets are spliced ​​together and mapped through a linear layer. This is used as the input to the cell block at time t0 in the time-series event sharing layer. After gating calculation, the output of the first time step is obtained, as shown in equations (10)-(16):

[0127] X sha =Linear(Concat(H i0 H conte (10)

[0128] i t =σ(W ii X share +b it +W hi h (t-1) +b hi (11)

[0129] f t =σ(W if X share +b if +W hf h (t-1) +n hf (12)

[0130] g t =tanh(W ig X share- +b ig +W hg h (t-1) +b hg (13)

[0131] o t =σ(W io X share-0 +b io +W ho h (t-1) +b ho (14)

[0132] c t =f t *c (t-1) +i t *g t (15)

[0133] H share-0 =o t *tanh(ct (16)

[0134] In the above formula, X share It is the input of the time-series event sharing layer at time t0, W * b * It is the weight matrix and the bias, h (t-1) It is the hidden layer vector from the previous time step, randomly initialized and generated at time t0, H share-0 It is the output;

[0135] At this point, a similar operation is being performed in the first cell block of the Content-LSTM, except that the input is the text content representation at time t0, and the output is the time series feature h at time t0. content-0 ;

[0136] S33, splicing h content-0 and H share-0 And through linear mapping, the hidden state of Content-LSTM at time t0 is updated as the hidden state of the Content-LSTM cell block at the next time step, as shown in equation (17):

[0137] h content-0 =Linear(Concat(h) content-0 H share-0 (17)

[0138] The calculation process of Content-LSTM at the next moment refers to equations (3)-(8) in the process of extracting text content features of rumor events;

[0139] S34. Repeat S31 to S33 until all time steps are completed, and finally obtain the output of the time sequence event sharing layer at each time step, as shown in equation (18):

[0140] H share ={H share-0 H share-1 H share} (18).

[0141] It is important to note that during the computation process, the time-series event sharing layer also continuously interacts with information from the dynamic graph construction process and the rumor event text content extraction process in chronological order, thereby achieving time-series event sharing.

[0142] The preceding calculations yielded feature representations after feature sharing, namely the feature representations of Dynamic-GCN, Content-LSTM, and TES layers: H ir H content H shareTo understand which features have a more significant impact on prediction across different time series and tasks, this invention employs a multi-head attention mechanism for feature extraction. LSTM networks for rumor detection typically use the latent feature representation of the last time series as the predicted feature vector because the sequence data is processed through a gating mechanism, and the output of the last gating operation contains the hidden states from all preceding gating operations. Therefore, this invention uses the features of the last time series for computation to obtain the convolutional features H of the dynamic graph convolutional network at the final time step. it This will be used for subsequent calculations. The specific implementation details are as follows:

[0143] Preferably, step S4 specifically includes the following steps:

[0144] S41, H share respectively with H content-t and H i-t The inputs X1 and X2 of the attention layer are concatenated and mapped through a linear layer, as shown in equations (19-20):

[0145] X1 = Linear(Concat(H) content- H shar (19)

[0146] X2 = Linear(Concat(H) i-t H share (20)

[0147] S42, X1, and X2 are linearly transformed into three projections with the same dimensions as themselves, namely query Q, key K, and value V, which serve as the inputs to the attention layer, as shown in equations (21)-(23).

[0148] MultiHead(Q,K,V)=Concat(head1,...,head h W o (twenty one)

[0149]

[0150]

[0151] In the above formula, MultiHead(Q, K, V) is the output of the attention layer, and Q, K, and V represent respectively. W O It is the weight matrix, W i It is the initialization of different multi-head matrices, d k It is the dimension of K; when K is large, the output is scaled.

[0152] S43. Connect the output of the attention layer with H respectively.share-t and H i-t They are concatenated and then linearly mapped to update them, as shown in equations (24)-(25).

[0153] H content-t =Linear(Concat(H share-t H content-t )) (twenty four)

[0154] H i-t =Linear(Concat(H share H i-t (25)

[0155] S44. Obtain the updated final event feature representation H used for prediction. content-t H i-t For each of them, a corresponding fully connected layer and activation layer are designed, as shown in equations (26)-(27).

[0156] Y C =softmax(w C H content-t +b C (26)

[0157] Y DG =softamx(w DG H i-t +b DG (27)

[0158] w and b are their respective weights and bias parameters. Finally, the cross-entropy loss function is used as the classification loss, as shown in equations (28)-(29).

[0159]

[0160]

[0161] Y i It is the true label of the i-th event;

[0162] During training, the two loss functions are assigned corresponding coefficients, added together to obtain the total loss function, and then backpropagated, as shown in equation (30).

[0163] Loss = W1 * L C +W2 * L DG (30)

[0164] Preferably, it also includes model training, the model training process as follows: Figure 5 As shown.

[0165] Verification Example: To investigate the content and dissemination patterns of fake news and to verify the effectiveness of the model, this invention compared the detection results with other models on the Weibo and Phome datasets. The experiment used four evaluation metrics—accuracy, precision, recall, and F1 score—to evaluate the model's performance. The F1 score was used to assess the model's performance. The calculation formulas are shown in equations (31-34), where TP represents the true positive rate, TN represents the true negative rate, FP represents the true negative rate, P represents the false positive rate, and FN represents the false negative rate.

[0166]

[0167]

[0168]

[0169]

[0170] The experimental results of this invention are shown in Table 1:

[0171] Table 1 shows the rumor detection results on the Weibo and Pheme datasets.

[0172]

[0173] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A multi-task rumor detection model and method based on dynamic networks, characterized in that, Specifically, the following steps are included: S1. Construct a dynamic graph of rumor propagation: Dynamic-GCN. Design a fine-grained dynamic graph neural network based on time steps. It uses a dynamic graph neural network to construct a rumor event propagation graph at different times. S2. To capture temporal text features, a text content extraction network based on LSMT, Content-LSTM, was designed. It is an improvement on the LSTM network as a baseline model to capture temporal text features that change over time and is used for shared interaction in multiple tasks. S3. Construct a time-series event sharing layer and design a time-series event-based sharing method TES as the sharing layer. The time step is the basic sharing unit to realize the sharing and interaction of dynamic structural features and time-series text features between the first two modules. S4. Attention Mechanism Feature Extraction and Classification: Through the preceding calculations and the calculations using TES, we obtained the feature representations after feature sharing, namely the feature representations of Dynamic-GCN, Content-LSTM, and TES layers: , , A multi-head attention mechanism is used to update them, resulting in an updated final event feature representation for prediction. , Finally, the data are input into a classifier for calculation to predict rumors. Before step S1, the problem is defined as a binary classification problem. These are examples of rumor detection events. Let n be the i-th event and n be the number of events; represent the set of posts and comments as... , This is the original tweet. This is the j-th comment text; It is an event The number of posts in the database; If the source tweet is considered as the 0th tweet, then Represented as To spread an event The order in which each post spreads is represented chronologically. , Indicates the time frame, the final event in which a rumor spread. Represented as ; Each post in the spread event Divided along the time dimension Each time step The number of posts for each dissemination event Decide, For a communication event At different time steps There are different propagation states, in the first time step In the second time step Therefore, the event The final representation at all time steps is as follows ; For the same transmission event Construct different propagation structure diagrams at different time steps. ;in Represents the set of tweets in the i-th event. Represents the set of tweets in the i-th event. This represents the time step of different tweets in the i-th event. These represent the i-th event, The set of nodes and edges at each moment. Represents the i-th event A time-based propagation diagram; Step S2 uses an LSTM network as a baseline model to improve upon it into a new model, Content-LSTM, for propagating events. All included tweets are represented as , used as the input of Content-LSTM, as in equations (3)-(8): ； ； ； ； ； ； in The feature vector input at a certain moment. This is the hidden state from the previous moment. For the corresponding weight matrix and bias; At this point, the text feature sequences of the same propagation event at each moment form a time feature sequence. As shown in equation (9), it is denoted as: ； Time series features With the dynamic structural sequence characteristics in formula (2) The propagation events occur in a one-to-one sequence, and this time series feature is used as one of the inputs for subsequent time series event sharing layer calculations; simultaneously, the sequence features of the last moment are also used. The final output of Content-LSTM is used as input in subsequent classification network computations.

2. The multi-task rumor detection model and method based on dynamic networks according to claim 1, characterized in that, In step S1, for the propagation event At time t0, there is only one source post in node 0. At time t1, two responding tweets interact with the source post, namely node 1 and node 2. At time t2, only one tweet responds to the source post in node 3. At time t3, node 4 and node 5 respond to node 1 at the same time. Based on this relationship, we can initially obtain the state of the coarse-grained dynamic graph.

3. The multi-task rumor detection model and method based on dynamic networks according to claim 2, characterized in that, Further fine-grained partitioning of the coarse-grained dynamic graph states constructs a time-step-based fine-grained dynamic graph sequence. For the dissemination event Constructing propagation graph snapshots at different times , It contains a propagation graph Feature matrix of all nodes The number of nodes in the propagation graph at that moment. The eigenvectors of a node are represented; its adjacency matrix is ​​defined as follows: As shown in equations (1) and (2), the dynamic graph convolutional layer is defined as: ； in, For degree matrix, It is an adjacency matrix. For the propagation diagram of the upper layer Feature matrix of all nodes This is the weight matrix. It is a non-linear activation function; Propagation graph for all time steps Perform the operations in equation (1) to obtain the spatial features of all time steps: ； At this point, the propagation structure characteristics of the same propagation event at each moment have been obtained, forming a dynamic feature sequence. The dynamic structural feature sequence is taken as one of the inputs in the subsequent time-series event sharing layer calculation; at the same time, the dynamic structural features of the last time step are used as inputs. As the final output of the dynamic graph convolutional network module, it serves as the input for subsequent classification network computation.

4. The multi-task rumor detection model and method based on dynamic networks according to claim 1, characterized in that, Step S3 includes the following steps: S31. At time t0, take the dynamic structural features during the construction of the dynamic graph. The corresponding Node_0 node feature that is currently propagating. and time series features Here, the propagation structure characteristics and content characteristics of the corresponding tweet with number 0 are taken at the same time. S32. The propagation structure and content features of the tweets are spliced ​​together and mapped through a linear layer. This is used as the input to the cell block at time t0 in the time-series event sharing layer. After gating calculation, the output of the first time step is obtained, as shown in equations (10)-(16): ； ； ； ； ； ； ； In the above formula, It is the input of the time-series event sharing layer at time t0. It consists of the weight matrix and the bias. It is the hidden layer vector from the previous time step, which is randomly initialized and generated at time t0. It is the output; At this point, a similar operation is being performed in the first cell block of the Content-LSTM, except that the input is the text content representation at time t0, and the output is the time series features at time t0. ; S33, splicing and And through linear mapping, the hidden state of Content-LSTM at time t0 is updated as the hidden state of the Content-LSTM cell block at the next time step, as shown in equation (17): ； The calculation process of Content-LSTM at the next moment is referenced to Equations (3)-(8) in the feature extraction process of rumor event text content. S34. Repeat S31 to S33 until all time steps are completed, and finally obtain the output of the time sequence event sharing layer at each time step, as shown in equation (18): ； 。 5. The multi-task rumor detection model and method based on dynamic networks according to claim 4, characterized in that, S4 specifically includes the following steps: S41, will respectively with and The data is concatenated and mapped through a linear layer to obtain the input to the attention layer. As shown in equation (19-20): ； ； S42, After linear transformation, it is transformed into a three-dimensional projection with the same dimensions as itself, namely query, Q; key, K; value, V, which serve as the input to the attention layer, as shown in equations (21)-(23). ； ； ； In the above formula, This is the output of the attention layer, where Q, K, and V represent... , It is a weight matrix. It involves the initialization of different multi-head matrices. It is the dimension of K; when K is large, the output is scaled. S43. Connect the output of the attention layer with... and They are concatenated and linearly mapped to update them, as in equations (24)-(25). ； ； S44. Obtain the updated final event feature representation for prediction. , For each of them, a corresponding fully connected layer and activation layer are designed, as shown in equations (26)-(27). These are their respective weights and bias parameters. Finally, the cross-entropy loss function is used as the classification loss, as shown in equations (28)-(29). It is the true label of the i-th event; During training, the two loss functions are assigned corresponding coefficients, added together to obtain the total loss function, and then backpropagated, as shown in equation (30). 。

Images