A time-aware heterogeneous graph neural rumor detection model

Abstract

The application discloses a time sequence perceived heterogeneous graph neural rumor detection model. In recent years, the development of online social media greatly accelerates the breeding and spread of rumors, and the harmfulness of rumors makes the automatic detection technology of rumors widely concerned by researchers. The application simultaneously considers the global structural relationship between events and the time sequence relationship of internal message transmission of the events, models the two kinds of relationships together by taking a heterogeneous graph as a carrier, and proposes a new time sequence perceived heterogeneous graph neural rumor detection model. The model captures the time sequence relationship between the internal forwarding (or comment) posts of the events by using a time sequence perceived self-attention mechanism, fuses the forwarding (or comment) posts with time sequence information with source posts, and obtains local time sequence representations of the events. Then, the model captures the global structural relationship between the events by using an element-level attention mechanism, learns global structural representations of the events, and finally fuses the two kinds of representations for detecting rumors.

Description

TECHNICAL FIELD

[0001] The application provides a time-aware heterogeneous graph neural rumor detection model, and belongs to the technical field of rumor detection. BACKGROUND

[0002] In recent years, the development of online social media has greatly accelerated the spread of rumors, and the harmfulness of rumors has attracted widespread attention from researchers in the field of automatic rumor detection. Early rumor detection methods mostly use feature engineering to extract effective features from text content, user configuration information and propagation structure. Such methods are time-consuming and require a lot of human resources, and the artificially constructed features are subjective and lack high-level feature representation. With the development of deep learning, deep neural networks have achieved good results in many natural language processing tasks such as sentiment analysis, machine translation and text classification. Inspired by this, researchers have begun to use deep learning models to model text content and propagation structure, and have proposed many effective rumor detection methods. Recently, graph-based methods use graph neural networks to model the structural features of message propagation, transforming the rumor detection task into a graph classification task and achieving good results. However, these methods only consider the local propagation structure of posts within an event, ignoring the global structural relationships of events on social media. Yuan et al. (Yuan Chun-yuan, Ma Qian-wen, Zhou Wei, et al. Jointly embedding the local and global relations of heterogeneous graph for rumor detection [C] / / 2019 IEEE international conference on data mining (ICDM). IEEE, 2019: 796-805.) believe that each event is not an independent individual, and events may be connected due to the participation of the same user. Ignoring the connection between events will inevitably limit the detection performance of the model. Therefore, they studied the relationship between events from the perspective of heterogeneous networks and proposed a joint global and local relationship heterogeneous graph to capture the local semantic relationship and global structural information of message propagation. Although this model has achieved good results, it ignores the time sequence information in the message propagation process within the event. SUMMARY

[0003] The purpose of the present application is to overcome the above problems, a time-aware heterogeneous graph rumor detection model is proposed, which comprises a heterogeneous graph construction module, a local time sequence information coding module, a global structure information coding module and a rumor classification module.

[0004] S01: Construction of a heterogeneous graph, which includes two parts: based on the interaction relationship between the internal forwarding (comment) post and the source post of the event and the interaction relationship between the user and the event to construct a heterogeneous graph; each type of node in the heterogeneous graph is initialized to represent;

[0005] S02: Local time sequence feature extraction of the event, which includes two parts: using a time-aware self-attention mechanism to mine the local time sequence information inside the event to obtain a response post representation with time sequence information Then the response post with time sequence information is fused into the representation of the source post to obtain the local time sequence representation of the event

[0006] S03: Global structure feature extraction of the event, which includes two parts: calculating the attention vector γ i of the specific user u j in different aspects j ; the attention vector γ j is aggregated with all users participating in the event c i in an element multiplication manner to capture the global structure relationship between events and events;

[0007] S04: After obtaining the local time sequence representation and the global structure representation of each event node, the two features are spliced as the final representation of the event node, and the prediction result of the event is calculated through the full connection layer and the softmax function, that is, the probability value of each label Finally, a loss function is defined to update the model parameters to obtain the optimal value.

[0008] The step S01 specifically comprises:

[0009] S11: First, the event and the related user are abstracted as two different types of nodes in the network, and the edge relationship between the user node and the event node is established according to the participation of the user to the event (the user has the behavior of forwarding or commenting on the post in the event). And, in each event, there is a source post and a series of response posts. According to the time delay of the response post after the source post is published, the response post is constructed as a time sequence, so that each source post corresponds to a response sequence. Finally, an event-user heterogeneous graph with time sequence information is constructed.

[0010] S12: The initialization representation of each node in the heterogeneous graph is performed, and the specific steps include:

[0011] S12-1: The event node is initialized to represent. The essence of the event is the text content of the source post and the response post, which is initialized in the form of word vector and encoded by CNN. Specifically, the number of words in each post is fixed to L, and when the number of words is less than L, it is filled with 0; when the number of words exceeds L, it is truncated. Then the Word2Vec algorithm is trained on the corpus in a specific field to obtain the vector representation of each word. For words that do not appear in the pre-trained word vector library, the invention uses a uniform distribution to initialize, and the word vector is fine-tuned during training. Let the initial vector of each word be j represents the jth word in the post, and each post with L words can be represented as:

[0012] x 1:L =[x1;x2;…x L ],

[0013] where “;” is a concatenation operation,

[0014] Further, the CNN is used to encode the sentence sequence. Given a sentence sequence x 1:L composed of word vectors, a one-dimensional convolution operation is performed on each possible window by the convolution layer of the CNN:

[0015] e i =σ(W*x i:i+h-1 ),

[0016] to obtain the feature map where is a convolution kernel with size h. Then the maximum pooling operation selects the maximum value of each feature map, and the initialization vector representation of each post is obtained by concatenation operation. For the i-th event, the source post representation is and each response post is represented as The matrix composed of response posts in this event is denoted as

[0017] S12-2: The user node is initialized to represent. The attribute information of the user (including gender, age, number of fans, number of attention, etc.) is encoded to obtain the initialization vector representation of the user node. For the user information that cannot be obtained, it is initialized by normal distribution.

[0018] The step S02 specifically includes:

[0019] S21: Adopting the self-attention mechanism of time sequence perception to mine the local time sequence information in the event, capture the differences in the response posts generated by rumor events and non-rumor events at different time stages and the potential time sequence dependent relationship between the response posts.

[0020] S21-1: In order to encode the time delay information of each response post, a position embedding is generated for each response post using the position encoding (PE) formula in the Transformer model:

[0021]

[0022]

[0023] Wherein, pos represents the position of the response post in the sequence, d represents the dimension of PE, 2k represents the even dimension, and 2k+1 represents the odd dimension (that is, 2k≤d, 2k+1≤d).

[0024] S21-2: Associate the embedding of each response post with its corresponding position embedding to capture the time sequence information between the response posts:

[0025]

[0026] S21-3: Utilize the multi-head attention mechanism to focus on important response posts. The self-attention mechanism can explicitly give greater weight to information that has a greater impact on itself and weight the information to itself, greatly enriching the representation of the node, and the multi-head can consider the influence of external information as much as possible:

[0027]

[0028]

[0029] S22: Fuse the response posts with time sequence information into the representation of the source post to obtain the local time sequence representation of the event There are many specific fusion strategies, such as average pooling, maximum pooling, splicing, etc. The invention regards a series of response posts as first-order neighbor nodes of the corresponding source post, and adopts the aggregation function in the graph attention network for fusion, and the specific calculation is as follows:

[0030] α ii =softmax(LeakyReLU(a T [m i ;m i ]),

[0031]

[0032]

[0033] where, is a sigmoid activation function, and ii , respectively, represent the attention scores between node i and itself and node i and node j, respectively, and ij N(m i ) is the neighbor nodes of node i, i.e., all response posts corresponding to the current source post, is the weight parameter of the node feature transformation of the layer.

[0034] The step S03 specifically comprises:

[0035] S31: Based on the global structure relationship between events established by common users, we consider how to learn the global structure features of event nodes Inspired by the element-level attention mechanism widely used in recommendation system tasks, the present application proposes a user embedding element-oriented attention mechanism, which assumes that each dimension of the user embedding reflects different aspects of information about the user, and these different attributes of the user have different effects on the propagation of messages. The specific process is as follows:

[0036] For a specific user i involved in event c , the attention vector j of the user u j in different aspects is calculated:

[0037] γ j = tanh(W c ·u j +b),

[0038] where, is a feature transformation matrix, is an attention vector in different aspects. The larger the value is, the greater the influence of the kth aspect of the user embedding u j on the propagation of messages.

[0039] S32: The attention vectors j of all users involved in event c i are aggregated in an element-wise product manner to capture the global structure relationship between events:

[0040]

[0041] The step S04 specifically comprises:

[0042] S41: Splice the local temporal representation of the event with the global structure representation as the final representation of the event node, and calculate the prediction result of the event through the full connection layer and the softmax function, that is, the probability value of the event for each label:

[0043]

[0044] Wherein, Fc(·) is a full connection layer, and the dimension of the output is consistent with the category of classification.

[0045] S42: Finally, the loss function of the model is defined as the cross entropy between the prediction result and the real label:

[0046]

[0047] Wherein, r is the number of categories of classification, theta is the parameter of the whole model, y i ∈{0,1,2,3}(Twitter), y i ∈{0,1}(Weibo) is the real label value.

[0048] Compared with the prior art, the beneficial effects achieved by the implemented time-aware heterogeneous graph neural rumor detection model of the application include:

[0049] The application fully considers the local temporal relationship between the internal forwarding (or comment) post and the source post of the event and the global structure relationship between events, and jointly models the local temporal information and global structure information to complete the rumor detection task.

[0050] The application is based on the interaction relationship between the forwarding (or comment) post and the source post, models the local temporal relationship between the response post through position coding, uses the multi-head attention mechanism to focus on important response posts, and then fuses the source post and the response post to obtain the local temporal representation of each event node.

[0051] The application is based on the interaction relationship between the user and the event, and uses the element-level attention mechanism to learn the global structure representation of each event node to capture complex and diverse propagation structure features.

[0052] The application has carried out a series of experiments on three real-world data sets. The experimental results show that the model proposed in the application has achieved better effects than existing models in rumor classification and rumor early detection tasks. BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 1 It is the overall framework diagram of the time-aware heterogeneous graph neural rumor detection model - SHGN model. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solutions and advantages of the present application more clear, specific embodiments are combined below, and the present application is further described in detail with reference to the drawings, but not used to limit the scope of the present application.

[0055] Figure 1 is a sequence-aware heterogeneous graph neural rumor detection model-SHGN (Sequence-aware Heterogeneous Graph Neural Rumor Detection) model overall framework diagram. As shown in Figure 1 , it includes four modules: a heterogeneous graph construction module, a local temporal information encoding module, a global structure information encoding module, and a rumor classification module. Specifically, it includes

[0056] S01: Construction of a heterogeneous graph, which includes two parts: construction of a heterogeneous graph based on the interaction relationship between the internal forwarding (comments) of the event and the source post, and the interaction relationship between the user and the event; and initialization of the representation of each type of node in the heterogeneous graph.

[0057] Specifically, the construction of the S01 heterogeneous graph includes the following steps:

[0058] S11: First, events and related users are abstracted as two different types of nodes in the network, and the edge relationship between user nodes and event nodes is established according to the participation of the user in the event (the user has the behavior of forwarding or commenting on the post in the event). And, in each event, there is a source post and a series of response posts. According to the time delay of the response post after the source post is published, the response post is constructed as a time sequence, so that each source post corresponds to a response sequence. Finally, an event-user heterogeneous graph with time sequence information is constructed.

[0059] S12: Initialization of the representation of each node in the heterogeneous graph, including the following steps:

[0060] S12-1: Initialization of the representation of the event node. The internal essence of the event is the text content of the source post and the response post, which is initialized in the form of word vector and encoded using CNN. Specifically, the number of words in each post is fixed to L, and when the number of words is less than L, it is filled with 0; when the number of words exceeds L, it is truncated. Then, the Word2Vec algorithm is trained on a specific domain corpus to obtain the vector representation of each word. For words that do not appear in the pre-trained word vector library, the present application uses a uniform distribution for initialization, and the word vector is kept adjustable during training. Let the initial vector of each word be j represents the jth word in the post, and each post with L words can be represented as:

[0061] x 1:L =[x1;x2;…xL ],

[0062] where “;” is the concatenation operation,

[0063] Further, the CNN is used to encode the sequence of sentences. Given a sequence of sentences x composed of word vectors 1:L , the one-dimensional convolution operation is performed on each possible window by the convolution layer of the CNN:

[0064] e i = σ(W*x i:i+h-1 ),

[0065] The feature map is obtained where is a convolution kernel with a size of h. Then the maximum pooling operation selects the maximum value of each feature map, and the concatenation operation is used to obtain the initial vector representation of each post. For the i-th event, the source post representation is Each response post is represented as The matrix composed of the response posts in the event is denoted as

[0066] S12-2: Initialization of user node representation. The attribute information of the user (including gender, age, number of fans, number of attention, etc.) is encoded to obtain the initial vector representation of the user node. For the user information that cannot be obtained, the normal distribution is used for initialization.

[0067] S02: Local temporal feature extraction of the event, which includes two parts: a time-aware self-attention mechanism is used to mine the local temporal information within the event, and the response post representation with temporal information is obtained Then the response post with temporal information is fused into the representation of the source post to obtain the local temporal representation of the event

[0068] Specifically, the S02 local temporal information encoding includes the following steps:

[0069] S21: A time-aware self-attention mechanism is used to mine the local temporal information within the event, and the differences between rumor events and non-rumor events in different time stages and the potential temporal dependency between response posts are captured.

[0070] S21-1: In order to encode the time delay information of each response post, a position embedding is generated for each response post using the Positional Encoding (PE) formula in the Transformer model:

[0071]

[0072]

[0073] where pos represents the position of the response post in the sequence, d represents the dimension of the PE, 2k represents the even dimension, and 2k+1 represents the odd dimension (i.e. 2k≤d, 2k+1≤d);

[0074] S21-2: associate the embedding of each response post with its corresponding position embedding to capture the timing information between response posts:

[0075]

[0076] S21-3: focus on important response posts using a multi-head attention mechanism. The self-attention mechanism can explicitly give more weight to information that has a greater impact on itself and weight the information to itself, greatly enriching the representation of the node, and the multi-head can consider the influence of external information as much as possible:

[0077]

[0078]

[0079] S22: fuse the response posts with timing information into the representation of the source post to obtain the local timing representation of the event There are many specific fusion strategies, such as average pooling, maximum pooling, and concatenation operations. The present application regards a series of response posts as first-order neighbor nodes of the corresponding source post and uses the aggregation function in the graph attention network for fusion, and the specific calculation is as follows:

[0080] α ii =softmax(LeakyReLU(a T [m i ;m i ])),

[0081]

[0082]

[0083] where, is a sigmoid activation function, and α ii , α ij represent the attention scores between node i and itself and node i and node j, respectively, N(m i ) is the neighbor node of node i, i.e. all response posts corresponding to the current source post, is the weight parameter of the node feature transformation of this layer.

[0084] S03: Extraction of global structural features of events. This module consists of two parts: calculating the participating events c i A specific user u in j Attention vectors γ in different aspects j ; Use the attention vector γ j and participation in events c i All users in are aggregated in an element-wise manner to capture the global structural relationship between events.

[0085] Specifically, S03 global structure information encoding includes the following steps:

[0086] S31: Based on the global structural relationship between events established by common users, we consider how to learn the global structural features of event nodes Inspired by the element-level attention mechanism widely used in recommendation system tasks, this paper proposes an attention mechanism guided by user embedding elements. It assumes that each dimension of the user embedding reflects different aspects of the user, and these different user attributes will have different effects on the spread of messages. The specific process is as follows:

[0087] For participation in event c i A specific user in Calculate user u j Attention vectors γ in different aspects j :

[0088] γ j =tanh(W c ·u j +b),

[0089] in, is the feature transformation matrix, are attention vectors of different aspects. The larger the value, the more embedded the user is. j The k-th aspect has a greater impact on message propagation.

[0090] S32: Using attention vector γ j and participation in events c i All users in are aggregated element-wise to capture the global structural relationship between events:

[0091]

[0092] S04: After obtaining the local temporal representation and global structural representation of each event node, the two features are concatenated as the final representation of the event node. The prediction result of the event is calculated through the fully connected layer and the softmax function, that is, the probability value of the event for each label. Finally, define a loss function, and update the model parameters to obtain the optimal value.

[0093] Specifically, S04 rumor classification includes the following steps:

[0094] S41: splice the local time sequence representation and the global structure representation as the final representation of the event node, and calculate the prediction result of the event through the full connection layer and the softmax function, that is, the probability value of the event for each label:

[0095]

[0096] Wherein, Fc(·) is a full connection layer, and the dimension of the output is consistent with the category of classification.

[0097] S42: Finally, the loss function of the model is defined as the cross entropy between the prediction result and the real label:

[0098]

[0099] Wherein, r is the number of categories of classification, θ is the parameter of the whole model, y i ∈{0,1,2,3}(Twitter), y i ∈{0,1}(Weibo) is the real label value.

[0100] The above only describes the preferred embodiments of the present application, and it should be pointed out that for ordinary skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, and these improvements and refinements should be considered as the protection scope of the present application.

Claims

1. A time-aware heterogeneous graph neural rumor detection model, the model comprising four modules: a heterogeneous graph construction module, a local temporal information encoding module, a global structure information encoding module, and a rumor classification module, wherein, The rumor detection data set is used to construct an event-user heterogeneous graph, and an embedding technology is used to initialize the representation of each node in the graph; the local time sequence information coding module learns the response post representation with time sequence information based on the time sequence relationship between the event internal forwarding or comment posts, and then fuses the content information of the source post itself to obtain the local time sequence representation of each event node The global structure information coding module learns the global structure representation of each event node based on the interaction relationship between the user and the event by using an element-level attention mechanism The rumor classification module fuses the local time sequence representation and the global structure representation of the event to predict the probability of the current event being a rumor; and the model specifically includes the following implementation steps: S01: Heterogeneous graph construction: construct a heterogeneous graph based on the time sequence relationship between the internal forwarding or comment posts of the event and the source post and the interaction relationship between the user and the event; initialize the representation of each type of node in the heterogeneous graph; S02: local timing feature extraction of the event: adopt the self-attention mechanism of timing perception to mine the local timing information inside the event, and obtain the response sticker representation with timing information Then, the response sticker with timing information is fused into the representation of the source sticker to obtain the local timing representation of the event Then, the response sticker with timing information is fused into the representation of the source sticker to obtain the local timing representation of the event S03: Global structure feature extraction of events: compute the attention vector γ i for a specific user u j in different aspects j ; aggregate the attention vectors γ j for all users participating in an event c i in an element-wise product manner to capture the global structure relationship between events and events; S04: After obtaining the local temporal representation and the global structure representation of each event node, the two features are spliced as the final representation of the event node. The prediction result of the event is calculated through the full connection layer and the softmax function, i.e. the probability value of each label Finally, a loss function is defined to update the model parameters to obtain the optimal value.

2. The timing-aware heterogeneous graph neural rumor detection model of claim 1, wherein, The step S01 specifically includes: S11: First, abstract the event and related users as two different types of nodes in the network, and establish the edge relationship between the user node and the event node according to the user's participation in the event, that is, the user's forwarding or commenting behavior on the post in the event; and in each event, there is a source post and a series of response posts; according to the time delay of the response post after the source post is published, the response post is constructed as a time sequence, so that each source post corresponds to a response sequence; finally, an event-user heterogeneous graph with time sequence information is constructed; S12: Initialize the representation of each node in the heterogeneous graph, and the specific steps include: S12-1: initialize the event node representation: initialize in the way of word vector and encode it by CNN; specifically, fix the number of words in each post as L, fill with 0 when the number of words is less than L, and truncate when the number of words exceeds L; then train on the corpus by Word2Vec algorithm to get the vector representation of each word, initialize the words that do not appear in the pre-trained word vector library with uniform distribution, and keep the word vector fine-tuned during training; record the initial vector of each word as j represents the jth word in the post, and each post with L words is represented as: x 1:L = [x1; x2;... ; x L ], wherein " ; " is a concatenation operation, Further, the sentence sequence is encoded using a CNN: given a sentence sequence x 1:L one-dimensional convolution operation is performed on each possible window by the convolutional layer of the CNN: e i = σ(W * x i:i+h-1 ), get the feature map where, is a convolution kernel with size h; then use the max-pooling operation select the maximum value of each feature map, and then get the initial vector representation of each post by the splicing operation; for the i-th event, the source post representation is each response post is represented as the matrix composed of the response posts in the event is denoted as S12-2: Initialize the representation of the user node: encode the attribute information of the user, including gender, age, number of fans, and number of attention, to obtain the initial vector representation of the user node, and initialize the user information that cannot be obtained through normal distribution.

3. The timing-aware heterogeneous graph neural rumor detection model of claim 2, wherein, Step S02 specifically includes: S21: Use the self-attention mechanism with time sequence perception to mine the local time sequence information in the event, capture the differences between the response posts generated by rumor events and non-rumor events at different time stages, and the potential time sequence dependency relationship between the response posts; S21-1: In order to encode the time delay information of each response post, use the position encoding formula in the Transformer model to generate a position embedding for each response post: Wherein, pos represents the position of the response post in the sequence, d represents the dimension of PE, 2k represents the even dimension, and 2k+1 represents the odd dimension, that is, 2k≤d, 2k+1≤d; S21-2: Associate the embedding of each response post with its corresponding position embedding to capture the time sequence information between the response posts: S21-3: Use the multi-head attention mechanism to focus on important response posts: S22: Fusing the response patches with temporal information into the representation of the source patch, obtaining the local temporal characterization of the event A series of response patches are considered as first-order neighbor nodes of the corresponding source patch, and an aggregation function in the graph attention network is used for fusion, and the specific calculation is as follows: a ii = softmax(LeakyReLU(aT[m i ; m i ]), in, is the sigmoid activation function, α ii , α ij Represents the attention scores between node i and itself and between node i and node j, N(m i ) are the neighbor nodes of node i, that is, all the response posts corresponding to the current source post, is the weight parameter for feature transformation of nodes in this layer.

4. The timing-aware heterogeneous graph neural rumor detection model of claim 1, wherein, The step S03 specifically includes: S31: The global structural relationship between events based on common users, the specific process is as follows: For a particular user u participating in an event c i Compute attention vector γ for user u j in different aspects j :​ gamma j = tanh(W c · u j +b), wherein, is a feature transformation matrix, is an attention vector for a different aspect; is larger, the influence of the kth aspect of the user embedding u j on the message propagation is larger; S32: Use attention vector γ j All users participating in events c i are aggregated in an element-wise product fashion to capture the global structural relationship between events and events:

5. The timing-aware heterogeneous graph neural rumor detection model of claim 1, wherein, The step S04 specifically includes: S41: Concatenate the local time sequence representation of the event and the global structure representation as the final representation of the event node, and calculate the prediction result of the event through the full connection layer and the softmax function, that is, the probability value of each label of the event: Wherein, Fc(·) is a full connection layer, and the output dimension is consistent with the number of categories; S42: Finally, define the loss function of the model as the cross entropy between the prediction result and the true label: where r is the number of categories, 0 is the parameter of the entire model, y i ∈{0,1,2,3}Twitter, y i ∈{0,1}Weibo is the real label value.

Images