An Interpretable False Information Detection Method and Device Based on Credible Evidence Reasoning and Spatiotemporal Feature Aggregation

Abstract

The present invention discloses an interpretable false information detection method and device for credible evidence reasoning and spatio-temporal feature aggregation. The method includes: from two perspectives of the user's own authority and information semantic interaction, using the TrustRank random walk idea to calculate the credibility index of comments during the information dissemination process, taking the comments as evidence, and discovering high-quality evidence; combining the credibility index of the evidence, through the long short-term memory network and the graph attention network, aggregating the temporal neighborhood and spatial neighborhood information of the evidence, and selecting and combining the spatio-temporal structural features of the information to enrich the node semantic representation and enhance the reliable representation of the evidence; through the multi-head attention mechanism, modeling the semantic interaction between the evidence and the source information, capturing the false part of the source information, and at the same time modeling the implicit stance of the evidence capsule in a transparent manner through the dynamic routing mechanism. The present invention not only enhances the interpretability of the false information detection result, but also improves the discrimination performance.

Description

Technical Field

[0001] The present invention belongs to the technical field of information content security, and particularly to the detection of false information based on content semantics for social media. An interpretable false information detection method and device based on credible evidence reasoning and spatio-temporal feature aggregation are proposed. Background Art

[0002] Social media has become an important place for users to share and exchange information due to its openness and anonymity; at the same time, due to the low threshold and fast speed of providing and spreading online news, it has also provided a breeding ground for the rapid spread of false news. False information on social media not only threatens the security of the cyber space, but also plays an important role especially in major events. False information will seriously interfere with the public's cognition, causing the masses to make wrong decisions, thus having a serious negative impact on the order, economy, and society of the real world.

[0003] At present, false information detection methods can be divided into two categories: content-based and social context-based. False information detection models based on content semantics mainly model through the specific language style of information, including early manual modeling methods such as extracting linguistic features, topic and sentiment features, and methods using deep neural networks to mine deep implicit semantic features of text in recent years. Such methods have achieved certain results in identifying false news articles, but the information published on social media has characteristics such as shortness and non-standard expression compared with news articles, with less effective information and more noise data, and it is difficult for existing models to extract key features for detection.

[0004] During the process of information dissemination on social media, it will trigger discussions and doubts among users, and the comment content of users often provides key information beneficial to model prediction. Therefore, related research uses methods such as multi-task learning frameworks and co-attention network models to model the semantic interaction between the source information and comments, learn the views of users on relevant information content, and help the model infer which information is false. Such methods further improve the accuracy of model detection, but ignore the local characteristics of semantic interaction during the information dissemination process, that is, in the information dissemination process, replies usually respond to their direct ancestor nodes rather than the source post. Therefore, some research has begun to focus on the research of information dissemination patterns, modeling the information dissemination process as a dissemination tree or graph structure, and using recurrent neural networks, graph neural networks, etc. to learn the high-level spatial structure semantic features of information during the dissemination process.

[0005] Although these methods have achieved good improvements in the performance of detecting misinformation in social media, there are still some limitations. First, due to the low-threshold nature of social media, there may be low-quality comment information during the information dissemination process. When the model treats all disseminated nodes equally, it will introduce noise, making the node semantic representation based on structure learning unreliable. Therefore, it is necessary to consider the credibility of comment users or content to reduce the interference to such misinformation detection models. In addition, existing methods focus on using deep learning models, integrating more external information, and automatically mining implicit features to improve the model's performance in misinformation discrimination. However, as the complexity of the deep model increases, the internal decision-making process of the model becomes increasingly difficult to explain and verify. However, relevant psychological research shows that the spreadability of misinformation is closely related to the importance and ambiguity of events. Therefore, simply labeling information as false is usually not enough, and the model also needs to automatically give judgment bases to enhance its interpretability. Summary of the Invention

[0006] In view of the current situation that most misinformation detection methods mainly use deep learning models, combine social context, aggregate neighborhood features from time or space dimensions to enrich the semantic representation of information, and have achieved certain results. However, when aggregating neighborhood information, these models ignore the reality that the quality of social media information is uneven, and treat all node information equally. Therefore, it may bring noise to the model, making the node semantic representation unreliable. At the same time, these models focus on using deep learning to automatically mine implicit features to improve the detection performance of the model, ignoring the problem of the interpretability of the internal decision-making process and discrimination results of the model. The present invention proposes an interpretable misinformation detection method and device based on credible evidence reasoning and spatio-temporal feature aggregation.

[0007] To achieve the above object, the present invention adopts the following technical solutions:

[0008] On the one hand, the present invention proposes an interpretable misinformation detection method based on credible evidence reasoning and spatio-temporal feature aggregation, including:

[0009] From the two perspectives of the user's own authority and information semantic interaction, use the TrustRank random walk idea to calculate the credibility index of comments during the information dissemination process, and take the comments as evidence to discover high-quality evidence;

[0010] Combined with the credibility index of evidence, through the long short-term memory network and the graph attention network, aggregate the temporal neighborhood and spatial neighborhood information of evidence, and select and combine the spatio-temporal structural features of information to enrich the node semantic representation and enhance the reliable representation of evidence;

[0011] Model the semantic interaction between evidence and source information through the multi-head attention mechanism, capture the false part of the source information, and at the same time model the implicit stance of the evidence capsule in a transparent manner through the dynamic routing mechanism.

[0012] Furthermore, from the two perspectives of the user's own authority and information semantic interaction, use the TrustRank random walk idea to calculate the credibility index of posts during the information dissemination process, and find that high-quality evidence includes:

[0013] Based on the source tweet information, the text content of comments / forwards, the interaction relationship between information, and the basic information of the users participating in the discussion in the given sample, calculate the user's own authority of the information publisher and the degree of recognition between information. Use the source tweet information and the text content of comments / forwards in the given sample as nodes, use the user's own authority of the information publisher as the node weight, and use the degree of recognition between information as the edge weight to construct an information dissemination network with node weights and edge weights;

[0014] Use the TrustRank random walk idea to calculate the credibility index of comments during the information dissemination process:

[0015] r (t+1) =(αS+(1-α)J)r (t) (5)

[0016] Where r (t) and r (t+1) respectively represent the probability vectors of all nodes being visited before and after update, that is, the credibility of comments. α represents the probability that the user walks along the edge, then 1-α represents the probability that the user randomly jumps according to the node weight. S is the edge transition matrix, S is composed of the probabilities of all nodes walking along the edge, J is the jump matrix, and J is composed of the probabilities of all nodes jumping according to the node weights;

[0017] After multiple iterations, equation (5) finally converges to a stable vector, and then high-quality evidence is obtained.

[0018] Furthermore, it also includes: through the PCA principal component analysis method, weighted fusion of the basic information of the users participating in the discussion is performed to obtain the authority of the information publisher; the basic information of the users participating in the discussion includes whether real-name authentication is completed, the number of fans, whether there is a homepage introduction, the number of friends, whether there is location information, and the number of likes / number of published information.

[0019] Furthermore, calculate the degree of recognition between information according to the following formula:

[0020] w ij =sign(content i ,content j )*simlar(content i, content j ) (2)

[0021] where w ij represents the degree of recognition of the information published by blog post i for blog post j, and sign(content i , content j ) represents the emotional tendency between blog posts, and simlar(content i , content j ) represents the semantic similarity between blog posts.

[0022] Furthermore, the credibility index of the combined evidence aggregates the temporal neighborhood and spatial neighborhood information of the evidence through a long short-term memory network and a graph attention network, and selects and combines the spatio-temporal structural features of the information to enrich the node semantic representation and enhance the reliable representation of the evidence, including:

[0023] Integrating the credibility of information during the process of using bidirectional LSTM to model the semantic representation of evidence information in two directions:

[0024]

[0025]

[0026]

[0027] where represent the hidden states of the forward LSTM and the backward LSTM respectively; l represents the number of hidden units of the LSTM; [;] represents the vector concatenation operation; represents the initial semantic representation of evidence post i, d represents the dimension of the vector, when i = 0, c i represents the initial semantic representation of the source information; r i ∈[0, 1] represents the credibility of the node;

[0028] Using the power exponent of r i to replace represents the temporal structure semantic representation of evidence i;

[0029] Enhancing the reliable representation of nodes by introducing the credibility of nodes during the GAT neighborhood aggregation process:

[0030]

[0031] where α ij represents the contribution degree of neighbor node j to the semantic enhancement of node i, represents a learnable shared parameter matrix that acts on each node in the network, and q represents the dimension of the node semantic representation after a linear transformation; is a learnable parameter weight vector;

[0032] Adopt a multi-head attention mechanism to capture diverse representations of spatial structure relationships:

[0033]

[0034] Among them, || represents the concatenation operation, and W k are both parameters of the k-th head, σ(·) represents the ELU activation function, and N i represents the set of nodes directly connected to node i, represents the spatial structure semantic representation of evidence i;

[0035] Convert the temporal structure semantic representation and spatial structure semantic representation of evidence i to the same semantic space; determine the importance of the temporal structure semantic representation and spatial structure semantic representation through a fully connected layer with a sigmod activation function, and perform weighted fusion on the two semantic representations according to the importance to obtain the weighted fusion information representation.

[0036] Furthermore, the modeling of the semantic interaction between evidence and source information through the multi-head attention mechanism to capture the spurious part of the source information, and at the same time modeling the implicit stance of the evidence capsule in a transparent manner through the dynamic routing mechanism includes:

[0037] Adopt a multi-head attention mechanism to model the semantic interaction between evidence and source information:

[0038]

[0039]

[0040] Among them, X represents the set of weighted fusion information representations; R C represents the semantic representation of the source information at the word granularity; represents the number of heads in the attention mechanism, represents the set of underlying evidence capsules, and p represents the dimension of the evidence capsule semantic representation after one attention;

[0041] Obtain category capsules through the dynamic routing mechanism:

[0042]

[0043] Among them, P j|i represents the probability that evidence e i thinks the source information belongs to label j; is a learnable parameter matrix, where d v represents the dimension of the class capsule, and d e represents the dimension of the evidence capsule, and d e = np; the squash() function is used to compress the norm of the capsule to between 0 and 1;

[0044] After obtaining the class capsule through the dynamic routing mechanism, select the class capsule with the largest norm as the label of the authenticity of the source information.

[0045] On the other hand, the present invention proposes an interpretable false information detection device for credible evidence reasoning and spatio-temporal feature aggregation, including:

[0046] A TrustRank-based evidence credibility reasoning module, which is used to calculate the credibility index of comments in the information dissemination process from two perspectives of the user's own authority and information semantic interaction, using the TrustRank random walk idea, taking the comments as evidence, and discovering high-quality evidence;

[0047] A spatio-temporal structure perception module, which is used to combine the credibility index of evidence, and through the long short-term memory network and the graph attention network, aggregate the time neighborhood and space neighborhood information of the evidence, and select and combine the spatio-temporal structure features of the information to enrich the node semantic representation and enhance the reliable representation of the evidence;

[0048] An evidence aggregation module based on a capsule network, which is used to model the semantic interaction between evidence and source information through a multi-head attention mechanism, capture the false part of the source information, and at the same time model the implicit stance of the evidence capsule in a transparent manner through a dynamic routing mechanism.

[0049] Furthermore, the TrustRank-based evidence credibility reasoning module is specifically used for:

[0050] Based on the source tweet information, comment / retweet text content, interaction relationship between information, and basic information of the users participating in the discussion in a given sample, calculate the user's own authority of the information publisher and the degree of recognition between information. Using the source tweet information, comment / retweet text content in the given sample as nodes, taking the user's own authority of the information publisher as the node weight, and taking the degree of recognition between information as the edge weight, construct an information dissemination network with node weights and edge weights;

[0051] Use the TrustRank random walk idea to calculate the credibility index of comments in the information dissemination process:

[0052] r (t+1) =(αS+(1 - α)J)r (t) (5)

[0053] where r (t) and r(t+1) They respectively represent the probability vectors of all nodes being visited before and after the update, that is, the credibility of the comments. α represents the probability that the user walks along the edge, then 1 - α represents the probability that the user randomly jumps according to the node weight. S is the edge transition matrix, and S is composed of the probabilities of all nodes walking along the edge. J is the jump matrix, and J is composed of the probabilities of all nodes jumping according to the node weights;

[0054] After multiple iterations, Equation (5) finally converges to a stable vector, and then high-quality evidence is obtained.

[0055] Furthermore, it also includes: by using the PCA principal component analysis method, the basic information of the users participating in the discussion is weighted and fused to obtain the authority of the information publishing user; the basic information of the users participating in the discussion includes whether real-name authentication is completed, the number of fans, whether there is a homepage introduction, the number of friends, whether there is location information, the number of likes / the number of published information.

[0056] Furthermore, the recognition degree between information is calculated according to the following formula:

[0057] w ij = sign(content i , content j ) * simlar(content i , content j ) (2)

[0058] where w ij represents the recognition degree of the information published by blog post i to blog post j. sign(content i , content j ) represents the emotional tendency between blog posts, and simlar(content i , content j ) represents the semantic similarity between blog posts.

[0059] Furthermore, the spatio-temporal structure perception module is specifically used for:

[0060] Integrating the credibility of information in the process of using bidirectional LSTM to model the semantic representation of evidence information in two directions:

[0061]

[0062]

[0063]

[0064] Among them, respectively represent the hidden states of the forward LSTM and the backward LSTM; l represents the number of hidden units of the LSTM; [;] represents the concatenation operation of vectors; represents the initial semantic representation of the evidence post i, d represents the dimension of the vector, when i = 0, c i represents the initial semantic representation of the source information; r i ∈[0,1] represents the credibility of the node;

[0065] use the power exponent of r i to replace represents the temporal structure semantic representation of the evidence i;

[0066] By introducing the credibility of the node in the GAT neighborhood aggregation process, enhance the reliable representation of the node:

[0067]

[0068] where, α ij represents the contribution degree of the neighbor node j to the semantic enhancement of the node i, represents the learnable shared parameter matrix, which acts on each node in the network, q represents the dimension of the node semantic representation after a linear transformation; is the learnable parameter weight vector;

[0069] Adopt the multi-head attention mechanism to capture the diverse representations of the spatial structure relationship:

[0070]

[0071] where, || represents the concatenation operation, and W k are both the parameters of the k-th head, σ(·) represents the ELU activation function, N i represents the set of nodes directly connected to the node i, represents the spatial structure semantic representation of the evidence i;

[0072] Convert the temporal structure semantic representation and the spatial structure semantic representation of the evidence i to the same semantic space; determine the importance of the temporal structure semantic representation and the spatial structure semantic representation through the fully connected layer with the sigmod activation function, and perform weighted fusion on the two semantic representations through the importance to obtain the weighted fusion information representation.

[0073] Furthermore, the evidence aggregation module based on the capsule network is specifically used for:

[0074] Adopt the multi-head attention mechanism to model the semantic interaction between the evidence and the source information:

[0075]

[0076]

[0077] Where X represents the set of information representations after weighted fusion; R C represents the semantic representation of the source information word granularity; represents the number of heads in the attention mechanism, represents the set of underlying evidence capsules, and p represents the dimension of the semantic representation of the evidence capsules after one attention;

[0078] The category capsules are obtained through the dynamic routing mechanism:

[0079]

[0080] Where P j|i represents the probability that the evidence e i thinks the source information belongs to the label j; is a learnable parameter matrix, where d v represents the dimension of the category capsules, and d e represents the dimension of the evidence capsules, and d e = np; The squash() function is used to compress the norm of the capsule to between 0 and 1;

[0081] After obtaining the category capsules through the dynamic routing mechanism, the category capsule with the largest norm is selected as the label of the authenticity of the source information.

[0082] Compared with the prior art, the beneficial effects of the present invention are:

[0083] First, based on the TrustRank random walk idea, the present invention calculates the credibility index of relevant posts from two perspectives: the user's own authority and the network propagation structure, and discovers high-quality evidence; secondly, by designing a spatio-temporal structure perception step, combined with the credibility index, it aggregates the spatio-temporal neighborhood features of information and enhances the reliable representation of information; finally, a two-layer capsule network is designed to model the implicit stance of evidence in a transparent manner, while capturing the controversial points (false parts) of the source information, and enhancing the interpretability of the model decision-making process and results. The experimental results on two public datasets (PHEME, CED) show that the present invention can not only enhance the interpretability of the false information detection results, but also improve the model discrimination performance (F1 value) by 3.7% and 2.7% compared with the current advanced algorithms. Brief Description of the Drawings

[0084] Figure 1 It is the overall framework diagram of an interpretable false information detection method model (TRSA) for credible evidence reasoning and spatio-temporal feature aggregation according to an embodiment of the present invention;

[0085] Figure 2Schematic diagram of the information propagation network architecture with node weights and edge weights constructed for the embodiments of the present invention;

[0086] Figure 3 Performance comparison of the model under different damping coefficients;

[0087] Figure 4 Viewable diagram of the discrimination process of a certain sample (labeled False) of the model on the test set;

[0088] Figure 5 Attention degree of word-level semantics during the propagation of true and false information;

[0089] Figure 6 Comparison of the attention degree distribution of the word-level semantic granularity of the source information by the evidence;

[0090] Figure 7 Early detection performance of the model. Detailed implementation manners

[0091] The following further explains the present invention in conjunction with the accompanying drawings and specific embodiments:

[0092] As Figure 1 shown, the overall framework of an interpretable false information detection method for credible evidence reasoning and spatio-temporal feature aggregation mainly includes three parts:

[0093] (1) TrustRank-based credible reasoning: From the two perspectives of the user's own authority and information semantic interaction, use the TrustRank random walk idea to calculate the credibility index of comments during the information propagation process, and use the comments as evidence to discover high-quality evidence;

[0094] (2) Spatio-temporal structure perception: Combine the credibility index of the evidence, and through the long short-term memory network and the graph attention network, aggregate the temporal neighborhood and spatial neighborhood information of the evidence, and select and combine the spatio-temporal structure features of the information to enrich the node semantic representation and enhance the reliable representation of the evidence;

[0095] (3) Evidence aggregation based on the capsule network: Through the multi-head attention mechanism, model the semantic interaction between the evidence and the source information, capture the false part of the source information, and at the same time model the implicit stance of the evidence capsule in a transparent manner through the dynamic routing mechanism.

[0096] Specifically, the method includes:

[0097] 1. TrustRank-based credible evidence reasoning

[0098] In the process of information dissemination, users will form their own opinions on the received information and express their views and positions through forwarding or commenting. Therefore, users' replies to information can provide some evidence (clues) for the authenticity of the information. Due to the openness of social media, there may be low-quality reply information during the information dissemination process. The present invention attempts to select highly credible comments as the basis for judging the authenticity of statements. We believe that the credibility of a comment depends on the degree of recognition of its content by other participating comment users during the information interaction process. At the same time, the higher the user's own authority, the more reliable the comment content they publish. That is, users are more likely to receive information published by users with high authority. Therefore, the idea of random walk similar to PageRank is used to calculate the credibility index of comment evidence during the information dissemination process. The probability that the random walker finally reaches each node can be regarded as the credibility of the evidence.

[0099] (1) Construction of an information dissemination network with node weights and edge weights

[0100] First, based on the source tweet information, comment / retweet text content, interaction relationship between information, and basic information of participating users in the given sample, using the source tweet information and comment / retweet text content in the given sample as nodes, taking the authority of the information publishing user itself as the node weight, and taking the degree of recognition between information as the edge weight, an information dissemination network with node weights and edge weights is constructed. Specifically as Figure 2 shown. Among them, the depth of the node color represents the size of the node weight, the darker the color, the greater the node weight; the depth of the edge color represents the size of the edge weight, the darker the color, the greater the edge weight.

[0101] Use w i to represent the authority of the information publishing user, which is mainly expressed through the user's own metadata, including whether real-name authentication is completed (V), the number of followers (FL), whether there is a homepage profile (D), the number of friends (FR), whether there is location information (GEO), the number of likes (F, English dataset) or the number of published information (M, Chinese dataset). The present invention uses the PCA principal component analysis method to weight and fuse these attributes, specifically as shown in Equation (1):

[0102] w i = λ1V i + λ2FL i + λ3FR i + λ4D i + λ5GEO i + λ6F(M) (1)

[0103] Use w ijIt represents the degree of recognition of the information released by blog post j by blog post i. The present invention calculates the recognition degree of users for the upper-layer information in the network from two perspectives of information (content) semantics and sentiment, specifically as shown in Equation (2):

[0104] w ij =sign(content i ,content j )*simlar(content i ,content j ) (2)

[0105] Among them, sign(content i ,content j ) represents the sentiment tendency between blog posts, and simlar(content i ,content j ) represents the semantic similarity between blog posts. If the sentiment polarities between blog posts are the same, then sign(content i ,content j )=1, otherwise it is equal to 0. That is, although there is semantic interaction in the process of real information dissemination, since user u i has not affected u j 's view (i.e., there is no emotional resonance), therefore, the value of w ij is 0. The sentiment tendency calculation method of the present invention mainly uses the interface provided by the Baidu AI platform, and the similarity calculation method mainly adopts soft consistency measurement, taking the average word embedding between information as the semantic similarity.

[0106] (2) Calculation of the evidence credibility index based on random walk

[0107] According to the previous assumptions, the random walk mode of the random walker in the network can be summarized into two types: randomly jumping according to the weight of the network node (i.e., considering the authority of the user who publishes the blog post) and randomly walking along the edges in the network (i.e., considering the information interaction situation in the dissemination process).

[0108] The probability that the random walker jumps from node i to node j according to the weight is defined as j ij :

[0109]

[0110] The probability of jumping according to the node weight for all nodes can be represented by the jump matrix J, where each column element is the same, that is, the value of j ij is the same as the value of J j .

[0111] Edge - weight - based random walk means that the random walker moves along the edges in the network and selects a neighboring node to visit directly with a certain probability. If the probabilities of all nodes for edge - based random walks can be represented by the edge transition matrix \(S\), the probability of edge - based transfer between users is expressed as follows:

[0112]

[0113] Suppose the probability of a user randomly jumping according to node weights is \(1 - \alpha\), then \(\alpha\) represents the probability of a user's edge - based walk. Let \(r\) (t) and \(r\) (t+1) respectively represent the probability vectors of all nodes being visited (i.e., the credibility of nodes) before and after update. Then

[0114] \(r\) (t+1) =(αS+(1 - α)J)r (t) (5)

[0115] Related research shows that: when the transition matrix satisfies the irreducible and aperiodic properties, this type of random walk algorithm will converge to a unique vector. Therefore, after multiple iterations, equation (5) will eventually converge to a stable vector.

[0116] 2. Evidence Representation Based on Spatiotemporal Structure Awareness

[0117] To obtain more reliable semantic representations of source information and evidence information, it is necessary to incorporate the calculated node credibility into the evidence representation step to avoid the problem of noise introduced to the model due to uneven quality of reply information. At the same time, to alleviate the semantic sparsity of short texts on social media and characterize semantic interactions in the information dissemination process, the present invention aggregates the time - neighborhood and space - neighborhood information of nodes from two perspectives: the time series and the space structure of information dissemination, to enrich the semantic representation of nodes. It specifically includes three sub - steps: time - series representation, space - structure representation, and spatiotemporal feature fusion.

[0118] (1) Time - series Representation

[0119] First, obtain the initial semantic representations \(C=\{c_0,c_1,c_2,\cdots,c\) m \} of the source information and its related evidence posts through a pre - trained model, where \(d\) represents the dimension of the vector. When \(i = 0\), it represents the source information, and the rest represent evidence elements, and the evidence elements are sorted according to their posting times. To capture the temporal dependence relationship of information in the dissemination process and consider the reliability of time - neighborhood information at the same time, the present invention incorporates the credibility of information in the process of using bidirectional LSTM to model the semantic representations of evidence information in two directions. Specifically, as shown in equations (6) - (7):

[0120]

[0121]

[0122]

[0123] Among them, respectively represent the hidden states of the forward LSTM and the backward LSTM; l represents the number of hidden units of the LSTM; [;] represents the concatenation operation of vectors. r i ∈[0,1] represents the credibility of the node. Here, in order to avoid the r in the calculation i being too small, resulting in the vanishing gradient, use the power exponent of r i to replace represents the temporal structure semantic representation of evidence i (when i = 0, it represents the source information).

[0124] (2) Spatial structure representation

[0125] In recent years, many efficient graph neural networks have been used to learn the high-level structural features of the propagation network. For example, GAT can utilize structural information and the attention mechanism to learn the importance of different neighbor nodes to the target node, and aggregate neighbor information through weighted summation to enhance the feature representation of the target node. When aggregating neighbor nodes, although GAT can automatically capture the contribution degree of different nodes to the target node, it ignores the credibility of the neighborhood information. Therefore, such an aggregation operation may introduce noise. The present invention enhances the reliable representation of nodes by introducing the credibility of nodes during the GAT neighborhood aggregation process. Specifically, as shown in Equation (9):

[0126]

[0127] Among them, α ij represents the contribution degree of neighbor node j to the semantic enhancement of node i, represents the learnable shared parameter matrix, which acts on each node in the network; q represents the dimension of the node semantic representation after a linear transformation; is the learnable parameter weight vector.

[0128] Adopt the multi-head attention mechanism to capture diverse representations of spatial structure relationships, specifically as shown in Equation (10):

[0129]

[0130] Among them, || represents the concatenation operation, and W k are both the parameters of the k-th head, σ(·) represents the ELU activation function, N i represents the set of nodes directly connected to the node, represents the spatial structure semantic representation of evidence i (when i = 0, it represents the source information).

[0131] (3) Spatiotemporal Feature Fusion

[0132] To capture the semantic features of information from multiple perspectives, the spatiotemporal feature fusion is used to select and combine the temporal semantic representation and the spatial structure semantic representation of evidence, so as to obtain the spatiotemporal structure semantic representation of information. Considering that the temporal sequence semantic representation and the spatial structure semantic representation of evidence are not in the same semantic space, therefore, it is first necessary to further transform them into the same semantic space. The specific process is shown in Equation (11):

[0133]

[0134] Secondly, to determine the importance of the temporal structure semantic representation and the spatial structure semantic representation, it is calculated through a fully connected layer with a sigmod activation function, as shown in Equation (12). Finally, the two semantic representations are weighted and fused by the importance.

[0135] z = σ(W2[h' i ; m' i ) (12)

[0136] x i = z⊙h' i +(1 - z)⊙m' i (13)

[0137] are learnable parameters; σ(·) represents a fully connected layer with a sigmod activation function; is the information representation obtained after fusion, and h represents the dimension of the output representation after fusion.

[0138] 3 Evidence Aggregation Based on Capsule Network

[0139] To model the implicit stance of evidence on the authenticity of source information, the present invention introduces a capsule network into our model. The capsule network uses neuron vectors (capsules) to replace individual nodes in traditional neural networks, and adopts dynamic routing to realize the mapping from the current capsule to the high-level capsule, which can efficiently model the relationship between the local and the whole. Among them, the capsule can represent both the probability of target discrimination and the features of the current target. Compared with neural networks, its process is interpretable. Therefore, introducing a capsule network increases the transparency of the model inference process. The specific process is as Figure 1 shown in the evidence semantic aggregation part.

[0140] (1) Semantic Interaction between Evidence and Source Information Based on Multi-Head Attention

[0141] Although spatio-temporal feature fusion can efficiently aggregate the spatio-temporal neighborhood information of nodes to obtain reliable evidence representations, it cannot model the fine-grained semantic interaction between evidence and source information, nor can it model the view of the evidence on the authenticity of the source information. The attention mechanism can obtain the attention weights of each position in the sentence during the encoding process through a series of transformation operations of the matrix. In order to capture the focus points (false parts) of the comments on the source information, the present invention adopts a multi-head attention mechanism to model the semantic interaction between evidence and source information.

[0142] First, the pre-trained model is used to perform word-level semantic representation R of the source information C = BERT(s), where S = {w1, w2,..., w f}}, f represents the length of the text of the claim, Secondly, the obtained set of evidence representations (information representations) is regarded as Q, and R C is regarded as K and V. By performing matrix multiplication on the transpose of X and R C , the attention weights of each comment (evidence) for each word in the source information can be obtained; each comment is mapped to the semantic space of the source information in the form of a weighted sum. Specifically, as shown in Equation (14):

[0143]

[0144] To enhance the expressive power of the model and prevent the model from overly concentrating its attention on a specific position, first, Q, K, and V are linearly transformed through different transformations to map them to different spaces; secondly, attention calculations are performed in different spaces in parallel to obtain the encoded representation information of each comment (evidence) in different sub-spaces; finally, multiple attention results are concatenated, and non-linear features are applied through the ReLu activation function to obtain the final information representation. The specific calculation process is as follows:

[0145]

[0146] Among them, represents the number of heads in the attention mechanism, represents the set of underlying evidence capsules; p represents the dimension of the semantic representation of the evidence capsules after one attention.

[0147] (2) Evidence aggregation based on the dynamic routing mechanism

[0148] In the false information detection task, there are three types of high-level category capsules, namely false category capsules, true category label capsules, and uncertain category capsules. In the capsule network, the top-level capsules are aggregated from the underlying capsules. For example, for the false category capsules, they should be aggregated from the evidence supporting that the source information is false. Therefore, for the category capsules, it can be described by Equation (16):

[0149]

[0150] Among them, P j|i represents the probability that the evidence e i thinks the source information belongs to label j, and is calculated by an iterative algorithm. is a learnable parameter matrix, where d v represents the dimension of the class capsule, and d e represents the dimension of the evidence capsule, and d e = np. In order to make the norm of the class capsule represent the probability that the information belongs to this class, and at the same time increase the non-linear features, therefore, a squash operation is required to compress the norm of the capsule to between 0 and 1, as shown below:

[0151]

[0152] (3) Classification

[0153] After obtaining the class capsule through the dynamic routing mechanism, select the class capsule with the largest probability (norm) as the label of the authenticity of the source information.

[0154] p j = ||v j ||, j ∈ (0, 1, 2) (18)

[0155] Finally, in order to measure the difference between the prediction result and the true value, the present invention uses cross-entropy as the loss function of the model:

[0156]

[0157] Among them, θ is the parameter of the entire model, and y i ∈ {0, 1, 2} (Twitter), y i ∈ {0, 1} (Weibo) represents belonging to the true label value.

[0158] In order to verify the effectiveness of the method model of the present invention, an experiment is tried on a real dataset, and the following four questions are answered:

[0159] Question 1: Compared with the existing false information detection methods, can the model TRSA proposed by the present invention obtain better performance.

[0160] Question 2: Does each module of the model contribute to the performance of false information detection.

[0161] Question 3: Does the evidence credible reasoning and evidence semantic aggregation module make the result of false information detection easier to understand.

[0162] Question 4: What is the performance of the model in the early detection of false information?

[0163] 1. Model performance

[0164] Experimental dataset description: We validated our model on two real datasets; PHEME (an English dataset, the data mainly comes from foreign Twitter) and CED (a Chinese dataset, the data mainly comes from the domestic Sina platform). Among them, the PHEME dataset contains three types of labels, namely true, false, and uncertain; CED only contains two labels, true and false. The detailed statistical information of the datasets is shown in Table 1 as follows:

[0165] Table 1 Dataset statistics

[0166]

[0167]

[0168] Comparison methods: To verify the effectiveness of the method proposed in the present invention, performance comparison is carried out with the currently most advanced representative methods, which can be roughly divided into methods based on feature engineering, methods based on text sequence modeling, and methods based on information propagation network structure.

[0169] 1) Methods based on feature engineering

[0170] DTC: From four perspectives of text content, user characteristics, forwarding behavior, and propagation mode, design multi-dimensional statistical features including (number of followers, number of following, whether the identity is verified, sentiment score of the source tweet, average sentiment score of forwards, etc.), and use decision trees to make a determination on the credibility of information.

[0171] SVM-TS: Use manually crafted features to conduct overall statistics on posts and determine the truth or falsehood of information based on the support vector machine model.

[0172] 2) Methods based on text sequence modeling

[0173] HSA_BLSTM: Based on the hierarchical features of false events (a false event is composed of source information and multiple forwarded or commented posts, and each post is composed of words), use bidirectional LSTM and self-attention mechanism to learn the word-level, post set, and event-level representations of the text.

[0174] DTCA: Use the comment information of users as the evidence source for judging the authenticity of statements. First, use the decision tree method to construct a three-level credible evidence judgment index to screen high-credible evidence; secondly, by designing a co-attention network, enhance the semantic interaction between evidence and source information and capture the suspicious points of information.

[0175] EmotionEnhance: By measuring the difference in emotions between information publishers and audiences, a two-way emotional feature set is constructed and added to the fake news detector as a supplement and enhancement.

[0176] 1) Methods based on the information dissemination network structure

[0177] GLAN: First, the attention mechanism is used to enrich the semantics of the source tweet by aggregating relevant retweet messages; second, a heterogeneous network is constructed based on the source tweet, retweets, and users, and the graph attention neural network is used to capture rich structural information, thereby improving the performance of fake information detection.

[0178] BiGCN: Use two-layer graph convolutional neural network to explore the causal features of information propagation from top to bottom and the structural features from bottom to top, and at the same time integrate the source post information into each layer of GCN to enhance the influence of the source information.

[0179] DDGCN: Use GCN to model the features of the information dissemination structure and entity knowledge structure within each time period, and perform incremental fusion of these dynamic features by designing a time fusion unit.

[0180] To answer Question 1, first explore the performance changes of the model under different damping coefficient conditions to determine the optimal damping parameter. Specifically as Figure 3 shown.

[0181] The value range of the damping coefficient is [0, 1). When the damping coefficient is equal to 0, it means that the random walker only jumps according to the authority of the user during the walk, without considering the actual dissemination network (that is, the credibility of the finally obtained evidence is only determined by the user's own authority); when the damping coefficient is infinitely close to 1, it means that the random walker almost ignores the authority of the user and jumps along the actual dissemination network. The reason why 1 cannot be taken is that the information dissemination network is not a strongly connected graph, resulting in the possible failure of TrustRank. It can be seen from the figure that on the Pheme dataset, when the damping coefficient is 0.8, the model performance reaches the optimal; on the CED dataset, the best value of the damping coefficient is 0.7.

[0182] Secondly, compare the method model of the present invention with 8 representative models, and use four evaluation indicators: accuracy, recall rate, Macro F1 value, and precision rate to measure the performance of the models on the real dataset. The specific results are shown in Table 2.

[0183] Table 2 Comparison results of different models on Weibo and Pheme datasets

[0184]

[0185] 1) Deep learning methods are significantly better than machine learning methods based on feature engineering. The most fundamental reason is that deep learning models can automatically learn implicit high-level semantic representations, while traditional machine learning methods rely on feature engineering and can only capture obvious false information at the presentation layer, bringing certain limitations to the model.

[0186] 2) Models that increase the semantic interaction between statement content and comments (DTCA, EmotionEnhanced) perform better than models that directly concatenate content semantics and comment semantics (HAS_BLSTM), and at the same time make the model have a certain degree of interpretability. DTCA automatically captures the controversial points of the source information through the co-attention mechanism, and EmotionEnhanced BiGRU constructs dual emotional features by combining psychological principles to improve the detection performance of false information.

[0187] 3) Models based on the information dissemination structure are superior to models based on text semantic modeling. For example, the precision rates of GLAN, BiGCN, and DDGCN are on average 0.5 - 3.2 percentage points higher than those of DTCA on two datasets. This shows that mining the implicit structural features of information dissemination is very helpful for improving the detection performance of false information. However, the accuracy rate is 1.5 percentage points lower than that of DTCA. This is because DTCA uses decision trees to filter out some low-confidence noisy comments.

[0188] 4) The indicators of the model proposed in the present invention on two real datasets are superior to most text sequence models and information dissemination structure models. Compared with DTCA, the model proposed in the present invention enriches the comment semantic information of users from both temporal and spatial perspectives, and its performance is 5.7%, 3.05%, 6.95%, and 5.3% higher in the four evaluation indicators of the two real datasets. Compared with the information dissemination network structure model, it is on average 3%, 4%, 2.65%, and 3.2% higher in the four indicators. This is because these models treat all comments equally, which will introduce noise. Our model reduces this risk by calculating the credibility index of comments.

[0189] 2. Ablation experiment

[0190] To answer Question 2 and verify the effectiveness of each model proposed in the present invention, a series of ablation experiments were conducted, mainly including 3 modules:

[0191] (1) w / o TrustRank: Remove the part of calculating the credibility index;

[0192] (2)w / o Structure: It includes three parts: - w / o Sequential removes the temporal characteristics of information propagation; - w / o Spatial removes the spatial characteristics of information propagation; - w / o Spatial&Sequential removes both the spatio-temporal characteristics of information propagation.

[0193] (3)w / o Evidence aggregation: It means removing the evidence aggregation part based on the capsule network in the model.

[0194] Table 3 Comparison results of different variants on Weibo and Pheme datasets

[0195]

[0196]

[0197] As can be seen from Table 2, on both datasets, the performance of all ablation variants is worse than that of the complete TRSA. In particular, when removing the spatio-temporal characteristics of information propagation, the F1 score on the PHEME dataset decreased by 5.5%, and on the CED dataset it decreased by 6.7%. This shows the importance of temporal and spatial propagation structure information for improving model performance. In addition, the results show that compared with retaining the time series characteristics on both datasets, the performance degradation of the model retaining the spatial structure characteristics is smaller. Therefore, it can be further explained that the spatial structure characteristics are more important than the temporal structure characteristics. When removing the trust-aware evidence reasoning module, the F1 scores of PHEME and CED decreased by 3.6% and 2.9% respectively. This shows that the impact of low-quality comments on the model's performance can be mitigated through the evidence credibility index. The replacement of the evidence aggregation module led to a 4.8% decrease in the F1 score on PHEME and a 3.8% decrease on CED. It proves the necessity of aggregating evidence semantics to improve performance.

[0198] 3. Interpretability of the model

[0199] Based on the evidence credibility reasoning of TrustRank and the evidence aggregation based on the capsule network, the process of model discrimination becomes more transparent and the results are more interpretable. Therefore, to answer Question 3, the present invention reveals the internal reasons for the predicted information to be false by visualizing the evidence credibility index, the distribution of semantic interaction attention weights, and the distribution of evidence capsule stances. Figure 4 Describes the results of a specific sample in the test set.

[0200] First, by accumulating the attention values obtained from the semantic interaction between highly credible evidence and source information, the attention level of each word during the information dissemination process can be obtained, which is represented by the size and color of the word. The larger the word font and the darker the color, the higher the attention the word receives during the information dissemination process, and the more likely it is to arouse controversy among people. Starting from Figure 4 It can be observed that "Emergency", "distress", and "#4U9525" have been widely discussed by users during the information dissemination process, which further indicates that the model proposed in the present invention can automatically capture the controversial points of information. At the same time, in order to further illustrate the differences in the attention distribution of users to true and false information content during the information dissemination process, we randomly selected 3 false messages and 3 true messages from the test set, and visualized the weight distribution of each word in the source information through a heat map for comparative analysis. As Figure 5 shown, among them, the horizontal direction from left to right represents the word sequence, and in the vertical direction, the first three represent false messages (0 - 2), and the last three represent true messages (3 - 5). The results show that during the dissemination process of false information, some semantics will arouse extensive attention from users; while during the dissemination process of true information, the attention received by each part of the semantics is relatively uniform.

[0201] Secondly, gephi is used to draw the information dissemination network, where the size of the node is determined by the credibility index of the node. The higher the credibility index of the node, the larger the node. As Figure 4 shown, the black nodes represent the source information, and the remaining nodes represent relevant forwarded or commented posts. It is found that after the TrustRank calculation, relevant posts such as "I suspect that no pilot would say 'Emergency', but would say 'Mayday'", "No, then you would say 'PANPAN'. Trust me, I'm a pilot! Also: 'Mayday' is when life is in danger...", "By the way: The loss of cabin pressure in a passenger plane is a typical case of 'Mayday' #4u9525", etc. are given high credibility weights as evidence to prove that the source information is false. Among them, "PANPAN" and "Mayday" that appear in these comments belong to internationally common radio crisis call signals, thus proving the falsity of "Emergency" in the source information. This shows that the TrustRank module can discover highly reliable evidence to explain the results of the model's judgment.

[0202] In addition, in order to more objectively measure the support of these evidences for the results, the probability of the aggregation of underlying evidence capsules to high-level category capsules is visualized to reveal the implicit stance distribution of the evidences. Starting from Figure 6 it can be seen that generally speaking, the selected high-quality evidences all oppose the semantic content of the source information; at the same time, combined with the attention heat map of each evidence ( Figure 5),It can be further found that Evidence 1, 2, 3, 6, and 7 mainly discuss "Emergency" and "distress" in the source information; while 4 and 5 focus on the emergency situation of Flight ##4U9525 at that time.

[0203] 4. Early Detection Research of the Model

[0204] One of the important goals of false information detection is to detect false information as early as possible so as to intervene in a timely manner and minimize its influence range. To answer Question 4 and verify whether the verification model has excellent performance in the early detection of false information, the present invention designs an early detection experiment on two datasets, Pheme and CED. The specific method is as follows: Sort all comments or forwarded posts according to their release time; by changing the number of received comments (0%, 20%, 40%, 60%, 80%, 100%), the change in the detection performance of the model is evaluated. Figure 7 Shows the results of the early detection of the model on the two datasets. It can be observed that for the model proposed by the present invention, when only the first 40% of the comments or disseminated posts are obtained, the precision rates of the model can reach 85.2% and 91.2%. These results are better than those of the other comparison models, indicating that our model performs well in early detection. At the same time, it can be found that for the three models, GLAN, BiGCN, and DDGCN, the accuracy increases relatively slowly over time, while the model proposed by the present invention and the DTCA model have obvious performance improvements. This is because as the number of relevant discussion posts increases, the information dissemination structure becomes complex and the types of remarks are diverse. Since both this model and the DTCA model have modules for filtering noisy posts, it shows that the model has good robustness.

[0205] Based on the above embodiments, the present invention also proposes an interpretable false information detection device for credible evidence reasoning and spatio-temporal feature aggregation, including:

[0206] A credible evidence reasoning module based on TrustRank, which is used to calculate the credibility index of comments in the information dissemination process from two perspectives of the user's own authority and information semantic interaction by using the TrustRank random walk idea, and regard the comments as evidence to discover high-quality evidence;

[0207] A spatio-temporal structure perception module, which is used to combine the credibility index of evidence, aggregate the time neighborhood and space neighborhood information of the evidence through a long short-term memory network and a graph attention network, and select and combine the spatio-temporal structure features of the information to enrich the node semantic representation and enhance the reliable representation of the evidence;

[0208] The evidence aggregation module based on capsule network is used to model the semantic interaction between evidence and source information through the multi-head attention mechanism, capture the false part of the source information, and at the same time model the implicit stance of evidence capsules in a transparent manner through the dynamic routing mechanism.

[0209] Furthermore, the evidence credibility reasoning module based on TrustRank is specifically used for:

[0210] Based on the source tweet information, the text content of comments / retweets, the interaction relationship between information, and the basic information of users participating in the discussion in a given sample, calculate the authority of the information publishing user himself and the degree of recognition between information. Taking the source tweet information and the text content of comments / retweets in the given sample as nodes, taking the authority of the information publishing user himself as the node weight, and taking the degree of recognition between information as the edge weight, construct an information propagation network with node weights and edge weights;

[0211] Use the TrustRank random walk idea to calculate the credibility index of comments during the information propagation process:

[0212] r (t+1) =(αS+(1 - α)J)r (t) (5)

[0213] where r (t) and r (t+1) respectively represent the probability vectors of all nodes being visited before and after update, that is, the credibility of comments. α represents the probability that the user walks along the edge, then 1 - α represents the probability that the user randomly jumps according to the node weight. S is the edge transition matrix, S is composed of the probabilities of all nodes walking along the edge, J is the jump matrix, and J is composed of the probabilities of all nodes jumping according to the node weights;

[0214] After multiple iterations, equation (5) finally converges to a stable vector, and then high-quality evidence is obtained.

[0215] Furthermore, it also includes: through the PCA principal component analysis method, the basic information of users participating in the discussion is weighted and fused to obtain the authority of the information publishing user; the basic information of users participating in the discussion includes whether real-name authentication is completed, the number of fans, whether there is a homepage introduction, the number of friends, whether there is location information, and the number of likes / number of published information.

[0216] Furthermore, calculate the degree of recognition between information according to the following formula:

[0217] w ij =sign(content i ,content j )*simlar(content i ,content j)(2)

[0218] where w ij represents the degree of recognition of the information posted by blog post i for blog post j, and sign(content i , content j ) represents the emotional tendency between blog posts, and simlar(content i , content j ) represents the semantic similarity between blog posts.

[0219] Furthermore, the spatio-temporal structure perception module is specifically used for:

[0220] Integrating the credibility of information during the process of using bidirectional LSTM to model the semantic representation of evidence information in two directions:

[0221]

[0222]

[0223]

[0224] Among them, respectively represent the hidden states of the forward LSTM and the backward LSTM; l represents the number of hidden units of the LSTM; [;] represents the vector concatenation operation; represents the initial semantic representation of evidence post i, d represents the dimension of the vector, and when i = 0, c i represents the initial semantic representation of the source information; r i ∈[0, 1] represents the credibility of the node;

[0225] Use the power exponent of r i to replace to represent the temporal structure semantic representation of evidence i;

[0226] By introducing the credibility of the node during the GAT neighborhood aggregation process, enhance the reliable representation of the node:

[0227]

[0228] Among them, α ij represents the contribution degree of neighbor node j to the semantic enhancement of node i, represents the learnable shared parameter matrix, which acts on each node in the network, and q represents the dimension of the node semantic representation after a linear transformation; is the learnable parameter weight vector;

[0229] Adopt the multi-head attention mechanism to capture diverse representations of spatial structure relationships:

[0230]

[0231] Among them, || represents the splicing operation, and W k are both parameters of the k-th head, σ(·) represents the ELU activation function, N i represents the set of nodes directly connected to node i, represents the spatial structure semantic representation of evidence i;

[0232] Convert the temporal structure semantic representation and spatial structure semantic representation of evidence i to the same semantic space; determine the importance of the temporal structure semantic representation and spatial structure semantic representation through a fully connected layer with a sigmod activation function, and perform weighted fusion on the two semantic representations through the importance to obtain the weighted fusion information representation.

[0233] Furthermore, the evidence aggregation module based on the capsule network is specifically used for:

[0234] Adopt the multi-head attention mechanism to model the semantic interaction between evidence and source information:

[0235]

[0236]

[0237] where X represents the set of weighted fusion information representations; R C represents the semantic representation of the source information at the word granularity; represents the number of heads in the attention mechanism, represents the set of bottom-level evidence capsules, p represents the dimension of the semantic representation of the evidence capsule after one attention;

[0238] Obtain the category capsule through the dynamic routing mechanism:

[0239]

[0240] where P j|i represents the probability that evidence e i thinks the source information belongs to label j; is a learnable parameter matrix, where d v represents the dimension of the category capsule, d e represents the dimension of the evidence capsule, d e = np; the squash() function is used to compress the norm of the capsule to between 0 and 1;

[0241] After obtaining the category capsule through the dynamic routing mechanism, select the category capsule with the largest norm as the label of the authenticity of the source information.

[0242] In summary, based on the TrustRank random walk idea, the present invention calculates the credibility index of relevant posts from two perspectives: the user's own authority and the network propagation structure, and discovers high-quality evidence. Secondly, by designing a spatio-temporal structure perception step and combining the credibility index, the spatio-temporal neighborhood features of information are aggregated to enhance the reliable representation of information. Finally, a two-layer capsule network is designed to transparently model the implicit stance of evidence while capturing the controversial points (false parts) of the source information, enhancing the interpretability of the model decision-making process and results. Experimental results on two public datasets (PHEME, CED) show that the present invention can not only enhance the interpretability of false information detection results, but also improve the model discrimination performance (F1 value) by 3.7% and 2.7% compared with the current advanced algorithms.

[0243] The above are only the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. An interpretable misinformation detection method based on credible evidence reasoning and spatio-temporal feature aggregation, characterized in that, Including: From the two perspectives of the user's own authority and information semantic interaction, using the TrustRank random walk idea to calculate the credibility index of comments in the information dissemination process, taking comments as evidence, and discovering high-quality evidence; Combining the credibility index of evidence, through the long short-term memory network and the graph attention network, aggregating the temporal neighborhood and spatial neighborhood information of evidence, and selecting and combining the spatio-temporal structural features of information to enrich the node semantic representation and enhance the reliable representation of evidence; Through the multi-head attention mechanism, modeling the semantic interaction between evidence and source information, capturing the false part of the source information, and at the same time modeling the implicit stance of evidence capsules in a transparent manner through the dynamic routing mechanism; The above-mentioned calculating the credibility index of posts in the information dissemination process from the two perspectives of the user's own authority and information semantic interaction, and discovering high-quality evidence includes: Based on the source tweet information, the text content of comments / forwards, the interaction relationship between information, and the basic information of users participating in the discussion in the given sample, calculating the user's own authority of information release and the degree of recognition between information. Taking the source tweet information and the text content of comments / forwards in the given sample as nodes, taking the user's own authority of information release as the node weight, and taking the degree of recognition between information as the edge weight, constructing an information dissemination network with node weights and edge weights; Using the TrustRank random walk idea to calculate the credibility index of comments in the information dissemination process: r (t+1) =(αS+(1 - α)J)r (t) (5) where r (t) and r (t+1) respectively represent the probability vectors of all nodes being visited before and after the update, that is, the credibility of the comments. α represents the probability that the user walks along the edge, then 1 - α represents the probability that the user randomly jumps according to the node weight. S is the edge transition matrix, which is composed of the probabilities of all nodes walking along the edge. J is the jump matrix, which is composed of the probabilities of all nodes jumping according to the node weights; After multiple iterations, equation (5) finally converges to a stable vector, and then high-quality evidence is obtained; The above-mentioned combining the credibility index of evidence, through the long short-term memory network and the graph attention network, aggregating the temporal neighborhood and spatial neighborhood information of evidence, and selecting and combining the spatio-temporal structural features of information to enrich the node semantic representation and enhance the reliable representation of evidence includes: Integrating the credibility of information into the process of modeling the semantic representation of evidence information in two directions using bidirectional LSTM; Among them, respectively represent the hidden states of the forward LSTM and the backward LSTM; l represents the number of hidden units of the LSTM; [;] represents the concatenation operation of vectors; represents the initial semantic representation of the evidence post i, d represents the dimension of the vector. When i = 0, c i represents the initial semantic representation of the source information; r i ∈[0,1] represents the credibility of the node; Replace with the power exponent of r i ; denote the temporal structure semantic representation of evidence i; Enhancing the reliable representation of nodes by introducing the credibility of nodes in the GAT neighborhood aggregation process; Among them, α ij represents the contribution degree of neighbor node j to the semantic enhancement of node i, represents a learnable shared parameter matrix, which acts on each node in the network. q represents the dimension of the node semantic representation after a linear transformation; is a learnable parameter weight vector; Adopting the multi-head attention mechanism to capture diverse representations of spatial structure relationships; Among them, || represents the concatenation operation, and W k are both parameters of the k-th head, σ(·) represents the ELU activation function, N i represents the set of nodes directly connected to node i, represents the spatial structure semantic representation of evidence i; Converting the temporal structure semantic representation and spatial structure semantic representation of evidence i to the same semantic space; determining the importance of the temporal structure semantic representation and spatial structure semantic representation through a fully connected layer with a sigmod activation function, and weighted fusing the two semantic representations through the importance to obtain the weighted fused information representation; The above-mentioned through the multi-head attention mechanism, modeling the semantic interaction between evidence and source information, capturing the false part of the source information, and at the same time modeling the implicit stance of evidence capsules in a transparent manner through the dynamic routing mechanism includes: Adopting the multi-head attention mechanism to model the semantic interaction between evidence and source information; Among them, X represents the set of information representations after weighted fusion; R C represents the semantic representation of the source information word granularity; n represents the number of heads in the attention mechanism, represents the set of underlying evidence capsules, and p represents the dimension of the semantic representation of the evidence capsules after one attention; Obtaining category capsules through the dynamic routing mechanism; where P j|i represents the probability that the evidence e i is considered to belong to the label j for the source information; is a learnable parameter matrix, where d v represents the dimension of the class capsule, and d e represents the dimension of the evidence capsule, and d e = np; the squash() function is used to compress the norm of the capsule to between 0 and 1; After obtaining the category capsules through the dynamic routing mechanism, selecting the category capsule with the largest norm as the label of the authenticity of the source information.

2. The interpretable false information detection method based on credible evidence reasoning and spatio-temporal feature aggregation according to claim 1, wherein, Also including: Through the PCA principal component analysis method, the basic information of users participating in the discussion is weighted and fused to obtain the authority of the information release user; The basic information of the participating users includes whether real-name authentication is completed, the number of fans, whether there is a homepage profile, the number of friends, whether there is location information, and the number of likes / number of posts.

3. The interpretable false information detection method based on credible evidence reasoning and spatio-temporal feature aggregation according to claim 1, characterized in that Calculate the degree of recognition between information according to the following formula: w ij = sign(content i , content j ) * simlar(content i , content j ) (2) where w ij represents the degree of recognition of the information published by blog post i for blog post j, sign(content i , content j ) represents the emotional tendency between blog posts, and simlar(content i , content j ) represents the semantic similarity between blog posts.

4. An interpretable misinformation detection device for credible evidence reasoning and spatio-temporal feature aggregation, characterized in that, Including: A TrustRank-based evidence credibility reasoning module, which is used to calculate the credibility index of comments in the information dissemination process from two perspectives of the user's own authority and information semantic interaction, using the TrustRank random walk idea, taking the comments as evidence, and discovering high-quality evidence; A spatio-temporal structure perception module, which is used to combine the credibility index of evidence, and through the long short-term memory network and the graph attention network, aggregate the time neighborhood and space neighborhood information of the evidence, and select and combine the spatio-temporal structure features of the information to enrich the node semantic representation and enhance the reliable representation of the evidence; A capsule network-based evidence aggregation module, which is used to model the semantic interaction between evidence and source information through a multi-head attention mechanism, capture the false part of the source information, and at the same time model the implicit stance of the evidence capsule in a transparent manner through a dynamic routing mechanism; The TrustRank-based evidence credibility reasoning module is specifically used for: Based on the source tweet information, comment / retweet text content, interaction relationship between information, and basic information of the participating users in the given sample, calculate the user's own authority of the information publisher and the degree of recognition between information. Use the source tweet information and comment / retweet text content in the given sample as nodes, use the user's own authority of the information publisher as the node weight, and use the degree of recognition between information as the edge weight to construct an information dissemination network with node weights and edge weights; Use the TrustRank random walk idea to calculate the credibility index of comments in the information dissemination process: r (t+1) =(αS + (1 - α)J)r (t) (5) where r (t) and r (t+1) respectively represent the probability vectors of all nodes being visited before and after the update, that is, the credibility of the comments. α represents the probability that the user walks along the edge, then 1 - α represents the probability that the user randomly jumps according to the node weight. S is the edge transition matrix, and S is composed of the probabilities of all nodes walking along the edge. J is the jump matrix, and J is composed of the probabilities of all nodes jumping according to the node weights; After multiple iterations, formula (5) finally converges to a stable vector, and then high-quality evidence is obtained; The spatio-temporal structure perception module is specifically used for: Integrate the credibility of information in the process of modeling the semantic representation of evidence information in two directions using bidirectional LSTM: Among them, respectively represent the hidden states of the forward LSTM and the backward LSTM; l represents the number of hidden units of the LSTM; [;] represents the concatenation operation of vectors; represents the initial semantic representation of the evidence post i, d represents the dimension of the vector. When i = 0, c i represents the initial semantic representation of the source information; r i ∈[0,1] represents the credibility of the node; Replace with the power exponent of r i ; denote the temporal structure semantic representation of evidence i; Enhance the reliable representation of nodes by introducing the credibility of nodes in the GAT neighborhood aggregation process: Among them, α ij represents the contribution degree of neighbor node j to the semantic enhancement of node i, represents a learnable shared parameter matrix, which acts on each node in the network. q represents the dimension of the node semantic representation after a linear transformation; is a learnable parameter weight vector; Adopt a multi-head attention mechanism to capture diverse representations of spatial structure relationships: Among them, || represents the concatenation operation, and W k are both parameters of the k-th head, σ(·) represents the ELU activation function, N i represents the set of nodes directly connected to node i, represents the spatial structure semantic representation of evidence i; Convert the temporal structure semantic representation and spatial structure semantic representation of evidence i to the same semantic space; determine the importance of the temporal structure semantic representation and spatial structure semantic representation through a fully connected layer with a sigmod activation function, and weight and fuse the two semantic representations through the importance to obtain a weighted fusion information representation; The capsule network-based evidence aggregation module is specifically used for: Adopt a multi-head attention mechanism to model the semantic interaction between evidence and source information: E = MultiHeadAttention(X, R C , R C ) = ReLU([Head1||Head2||Head3...Head n ) where X represents the set of information representations after weighted fusion; R C represents the semantic representation of the source information word granularity; n represents the number of heads in the attention mechanism, represents the set of underlying evidence capsules, and p represents the dimension of the semantic representation of the evidence capsules after one attention; Obtain category capsules through a dynamic routing mechanism: where P j|i represents the probability that the evidence e i is considered to belong to the label j for the source information; is a learnable parameter matrix, where d v represents the dimension of the class capsule, and d e represents the dimension of the evidence capsule, and d e = np; the squash() function is used to compress the norm of the capsule to between 0 and 1; After obtaining the category capsules through the dynamic routing mechanism, select the category capsule with the largest norm as the label of the authenticity of the source information.

Images