A method for identifying susceptible users of fake information based on dynamic graph attention networks

Abstract

This invention discloses a method for identifying susceptible users of false information based on dynamic graph attention networks, comprising the following steps: Step 1: Mapping the propagation relationship of social networks into a directed graph; Step 2: Using a deep learning model to identify social bots and deleting social bot nodes; Step 3: Calculating user credibility features; Step 4: Calculating the user's global trust score; Step 5: Calculating the user's local trust score; Step 6: Using a dynamic graph attention mechanism, calculating the trust score between users based on their trust scores; Step 7: Normalizing the trust score to a value between 0 and 1; Step 8: Using a graph convolutional neural network to aggregate user credibility and obtain an adjacency matrix carrying neighbor trust scores; Step 9: Using a feedforward neural network to learn node representations carrying neighbor information; Step 10: Using a cross-entropy loss function to classify nodes and output a ternary classification of the user when facing false information: susceptible user / harmless user / debunking user.

Description

Technical Field

[0001] This invention relates to the field of misinformation control, and in particular to a method for identifying users susceptible to misinformation based on dynamic graph attention networks. Background Technology

[0002] The frequent outbreaks of misinformation on social media over the past decade have brought about a series of serious consequences, detrimental to long-term social stability. Currently, most methods for combating misinformation focus on content detection; however, the scope and impact of misinformation depend on the users who promote it. Every user is both a consumer and a disseminator of internet information. As nodes in social networks, internet users constantly receive and spread information daily. Therefore, the main actors in the spread of misinformation are users, and behind these users are people in the real world. Thus, to control the spread of misinformation on social networks, it is crucial to proactively identify users who may spread misinformation. Some users unintentionally become nodes in the spread of misinformation because they spread news without verifying its authenticity. This study refers to natural users who indiscriminately spread unverified information on social networks as misinformation-sensitive users. A deeper understanding of misinformation-sensitive users helps to proactively predict whether nodes in the dissemination path may further spread false / debunked information, providing solutions for the identification and control of misinformation on social media platforms.

[0003] Most current research focuses on detecting misinformation. Due to the privacy concerns and difficulty in collecting user data on social media, research on users spreading misinformation is still in its early stages. Domestic and international researchers largely focus on malicious users, social bots, and paid commenters on social networks. However, these studies suffer from two problems: first, they presuppose that users spreading misinformation are malicious or social bots controlled by software algorithms, ignoring users on social networks who readily believe false information; second, current research rarely addresses the psychological motivations of users spreading information and the trust relationships between users. In psychology, trust refers to a psychological state based on positive expectations of another's intentions and behaviors, a willingness to expose one's weaknesses without fear of being exploited, and this trust is subjective—two users with similar neighbors do not always trust the same neighbor. In sociology, an individual's motivation for spreading information can be divided into two categories: one is the tendency to spread information that aligns with their own cognition, i.e., intrinsic motivation; the other is the spread of information from people they trust, i.e., extrinsic motivation. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, this invention provides a method for identifying susceptible users of false information based on dynamic graph attention networks.

[0005] To achieve the aforementioned objectives of the invention, the technical solution adopted to solve its technical problems is as follows:

[0006] A method for identifying susceptible users of fake information based on dynamic graph attention networks includes the following steps:

[0007] Step 1: Map the news dissemination relationship graph in the social networking platform into a directed graph;

[0008] Step 2: Use a deep learning model to identify social bots in the social network and delete social bot nodes from the directed graph;

[0009] Step 3: Extract features from four dimensions based on the user's historical tweets and basic user information. After feature filtering and verification, integrate them into user credibility features, namely basic user features, emotional features, psychological features, and personality features.

[0010] Step 4: Use a social media trust calculation algorithm to calculate the trust level between users as the user's global trust level;

[0011] Step 5: Calculate the number of reposts a user has as their local trust level;

[0012] Step 6: Employ a dynamic graph attention mechanism to calculate the importance score between users, i.e., the trust score, based on the level of trust between them;

[0013] Step 7: Normalize the trust score to a value between 0 and 1;

[0014] Step 8: Based on the trust level, use a graph convolutional neural network to proportionally aggregate the trust levels of users extracted in Step 3 to obtain an adjacency matrix carrying the trust levels of neighbors.

[0015] Step 9: Use a feedforward neural network to learn node representations carrying neighbor information;

[0016] Step 10: Use the cross-entropy loss function to classify nodes and output a ternary classification of users when faced with false information: susceptible user / harmless user / debunking user.

[0017] Furthermore, step 1 includes the following:

[0018] Based on the publicly available dataset MinFN containing ten news events N1-N 10 Each news event includes a true news item and a fake news item, along with their propagation relationships. These relationships are mapped to a directed graph on the social network. Nodes in the directed graph represent users who spread the news, and edges represent the propagation relationships between users. The direction of the edges indicates the direction of news transmission.

[0019] Furthermore, step 2 includes the following:

[0020] To eliminate the influence of social bots on experimental results, a tweet-level model is proposed, which is an LSTM model that incorporates auxiliary information. User account information is used as auxiliary input. The tweet text is then segmented into sentences and words, converted into a set of word vectors by a pre-trained BERT model, and then fed into the LSTM model. The output is then concatenated with the auxiliary information input and fed into a two-layer fully connected neural network with ReLU activation function to generate the final output, achieving binary classification between social bots and natural users. After identifying social bot users, social bot nodes are removed from the information propagation network.

[0021] Furthermore, step 3 includes the following:

[0022] Based on the user's basic information, usable basic user characteristics are extracted, including authentication status and registration duration. The registration duration is calculated in days, with a specific date as the time node.

[0023] Emotional features were extracted using the NRC-EIL emotional intensity dictionary and a degree adverb dictionary. Based on users' historical tweets, their emotional tendencies were categorized into eight types: funny, happy, sad, afraid, unhappy, indifferent, encouraging, and angry. The calculation process included calculating the emotional tendencies of sentences and tweets, and finally classifying the user's overall emotion. Specifically, the user's emotional score was calculated through the following three steps:

[0024] For each user's tweet, after data preprocessing, sentence segmentation, and stop word removal, negative words were selected. Sentiment scores for each sentence were calculated based on the NRC-EIL dictionary and a degree adverb dictionary, using the following formula:

[0025] Senti_s core = (-1) 否定词个数 × Degree value × Sentiment word score

[0026] After obtaining the sentiment score for each sentence, the sentiment words of the same category in a tweet are summed, and the scores of these sentiment categories are compared. The sentiment category with the largest absolute value is taken as the sentiment tendency of the tweet, as shown in the following formula:

[0027] Senti t =max(|sentiment category, sum(sentiment word score)|)

[0028] After obtaining the sentiment tendency of each tweet, the category with the most sentiment tendency among the user's 100 historical tweets is selected as the user's sentiment tendency after statistical analysis, using the following formula:

[0029]

[0030] In the formula, n represents the number of tweets;

[0031] Psychological motivational characteristics include four factors: anxiety, uncertainty, attention-seeking, and engagement.

[0032] Anxiety factors were measured using the anxiety category in the LIWC dictionary, which includes tension and fear. The level of anxiety was measured by counting the number of words in the anxiety category in the LIWC dictionary used in user tweets.

[0033] Uncertainty factors are also calculated using categories related to uncertainty in the LIWC dictionary. The degree of uncertainty of a user is represented by the number and percentage of tweets containing words related to uncertainty.

[0034] The attention-seeking factor is determined by calculating the average number of tweets posted by a user over a 30-day period, including retweets, within a month of the most recent tweet posted by that user. This helps infer whether users with a psychological motivation to seek attention will post frequently to maintain account activity. The formula is as follows:

[0035]

[0036] Engagement factor measures a user's influence by calculating the ratio of followers to subscribers, using the following formula:

[0037]

[0038] Among them, TFF represents the fan follower ratio, num follower num represents the number of followers of a specific account. following This indicates the number of followers a particular account has;

[0039] The personality trait is predicted by using an unsupervised testing tool to analyze 25 of the user's tweets to identify the Big Five personality traits. The output is the user's score between 0 and 1 for each of the five personality types. Finally, the personality type with the highest score among the five types is taken as the user's representative personality.

[0040] Furthermore, step 4 includes the following:

[0041] The global trust calculation for users employs the previously proposed TSM algorithm to measure the trust between users, which serves as the edge weight between user nodes. The algorithm takes a graph G(V,E) as input and assigns each node in the network a pair of complementary trust scores, called trust level and credibility. The output is a tuple (ti, tw), where trust level ti quantifies a node's tendency to trust its neighbors, and trust level tw quantifies a neighbor's willingness to trust that node.

[0042]

[0043]

[0044] Where u, v, x∈V, ti(v) and tw(u) are the credibility scores of v and u respectively, w(v,x) is the edge weight from v to x, out(v) is the set of target nodes of the outgoing edge of v, in(u) is the set of source nodes of the incoming edge of u, and s is the participation score of the social network, which takes a value between 0 and 1, representing the potential risk that participants take when creating links in the network.

[0045] Furthermore, step 5 includes the following:

[0046] Local trust among users is calculated through retweet behavior. The degree to which a user trusts others is measured by the number of retweets in a user's tweets, denoted as tr1. The degree to which other users trust a user is measured by the average number of retweets of a user's tweets, denoted as tr2.

[0047] Furthermore, step 6 includes the following:

[0048] Based on the trust value Tr∈(ti,tw,tr1,tr2) obtained from the trust calculation, an importance score is assigned to each neighbor j (j∈Ni) of node i: First, the trust values ​​of nodes i and j, including local and global trust scores, are concatenated. Then, a parameterized projection matrix W is assigned for linear transformation to obtain a scalar. Finally, a shared attention mechanism a is used to perform self-attention. The calculation of this mechanism is based on the importance coefficient of the trust. The non-normalized trust score between nodes i and j is represented as follows:

[0049] e ij =a(W[Tr i ||Tr j ])

[0050] Among them, e ij Tr represents the trust score, which measures the importance of node j to i, and Tr represents the trust value obtained through trust calculation.

[0051] Furthermore, step 7 includes the following:

[0052] When aggregating neighbor information, it is necessary to normalize the attention coefficients of all neighbors of each node. The normalized attention weight α ij That is the true edge weight:

[0053] Normalize it using the softmax function, as shown in the formula below:

[0054]

[0055] Among them, eij It normalizes all neighbors of node i.

[0056] Furthermore, step 8 includes the following:

[0057] After obtaining the edge weights between the center node and its neighboring nodes, a graph convolutional neural network (GCN) with two hidden layers is used to aggregate the node credibility features Cr. The model input is the adjacency matrix A, which uses the attention coefficients as edge weights, and the output is the node feature matrix Z, which is an N×F matrix, where N is the number of nodes and F is the feature vector output by each node, as shown in the following formula:

[0058] H (l+1) =f(H (l) A)

[0059] Where H represents the hidden layer, and A represents the attention coefficient matrix, i.e. H(0) = Cr; H (L) =Z; L represents the number of layers, and f(·) represents the function for computing the hidden states of the graph convolution model;

[0060] The user credibility features are input into a graph convolutional neural network with two hidden layers, as shown in the following formula:

[0061]

[0062] in, I is the identity matrix, which ensures that its own features are included during the aggregation of the credibility features of neighbors; yes The diagonal matrix of node degree, This is equivalent to dividing all elements in each row by the out-degree of the node in that row.

[0063] Furthermore, step 9 includes the following:

[0064] Using the credibility features of nodes in the neighborhood of node i and the network structure, node representations are learned from the graph using an attention-based adjacency matrix paired with a feedforward neural network, as shown in the following formula:

[0065]

[0066] Where X represents the credibility feature. W(0) and W(1) are the weight matrices from the input layer to the hidden layer and from the hidden layer to the output layer, respectively.

[0067] Furthermore, step 10 includes the following:

[0068] Gradient descent is used to learn the weight matrix, followed by cross-entropy loss for node classification, as shown in the following formula:

[0069]

[0070] Among them, Y l The label represents the node, and f represents the credibility feature.

[0071] By employing the above technical solutions, this invention has the following advantages and positive effects compared with the prior art:

[0072] This invention presents a method for identifying users susceptible to misinformation based on dynamic graph attention networks. It treats users as nodes in information dissemination and mines user data from publicly available datasets to identify features related to user misinformation dissemination behavior, including basic features, emotional features, psychological features, and personality features. A social media trust algorithm is used to measure the level of trust between users. To address the dynamic nature of trust relationships between users on social media platforms, a framework for identifying users susceptible to misinformation based on dynamic graph attention networks is proposed. Finally, a series of comparative experiments verify the effectiveness of the extracted features and the superiority of the framework in identifying users susceptible to misinformation, providing a practical solution for controlling misinformation on social media platforms. Attached Figure Description

[0073] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0074] Figure 1 This is a flowchart of a method for identifying susceptible users of false information based on a dynamic graph attention network according to the present invention;

[0075] Figure 2 This is a flowchart of the user emotion classification process of the present invention;

[0076] Figure 3 This is a schematic diagram of the psychological motivation feature extraction method of the present invention;

[0077] Figure 4 This is a framework diagram of the false information-sensitive user identification model of the present invention;

[0078] Figure 5 This is a schematic diagram of the network topology of the MinFN dataset for false information in this invention. Detailed Implementation

[0079] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0080] like Figure 1 As shown, this embodiment discloses a method for identifying susceptible users of fake information based on dynamic graph attention networks, including the following steps:

[0081] Step 1: Map the news dissemination relationship graph in the social networking platform into a directed graph;

[0082] Furthermore, step 1 includes the following:

[0083] Based on the publicly available dataset MinFN containing ten news events N1-N 10 Each news event includes a true news item and a fake news item, along with their propagation relationships. These relationships are mapped to a directed graph on the social network. Nodes in the directed graph represent users who spread the news, and edges represent the propagation relationships between users. The direction of the edges indicates the direction of news transmission.

[0084] Step 2: Use a deep learning model to identify social bots in the social network and delete social bot nodes from the directed graph;

[0085] Furthermore, step 2 includes the following:

[0086] To eliminate the influence of social bots on experimental results, a tweet-level model is proposed, which is an LSTM model that incorporates auxiliary information. User account information is used as auxiliary input. The tweet text is then segmented into sentences and words, converted into a set of word vectors by a pre-trained BERT model, and then fed into the LSTM model. The output is then concatenated with the auxiliary information input and fed into a two-layer fully connected neural network with ReLU activation function to generate the final output, achieving binary classification between social bots and natural users. After identifying social bot users, social bot nodes are removed from the information propagation network.

[0087] Step 3: Extract features from four dimensions based on the user's historical tweets and basic user information. After feature filtering and verification, integrate them into user credibility features, namely basic user features, emotional features, psychological features, and personality features.

[0088] Furthermore, step 3 includes the following:

[0089] Based on the user's basic information, usable basic user characteristics are extracted, including authentication status and registration duration. The registration duration is calculated in days, with a specific date (e.g., January 1, 2022) as the time node.

[0090] Emotional features were extracted using the NRC-EIL emotional intensity dictionary and a degree adverb dictionary. Based on users' historical tweets, their emotional tendencies were categorized into eight types: amused, happy, sad, afraid, annoyed, don't care, inspired, and angry. The calculation process included calculating the emotional tendencies of sentences and tweets, and finally classifying the user's overall emotion. Specifically, the user's emotional score was calculated through the following three steps, as shown in the flowchart below. Figure 2 As shown:

[0091] For each user's tweet, after data preprocessing, sentence segmentation, and stop word removal, negative words were selected. Sentiment scores for each sentence were calculated based on the NRC-EIL dictionary and a degree adverb dictionary, using the following formula:

[0092] Senti_s core = (-1) 否定词个数 × Degree value × Sentiment word score

[0093] After obtaining the sentiment score for each sentence, the sentiment words of the same category in a tweet are summed, and the scores of these sentiment categories are compared. The sentiment category with the largest absolute value is taken as the sentiment tendency of the tweet, as shown in the following formula:

[0094] Senti t =max(|sentiment category, sum(sentiment word score)|)

[0095] After obtaining the sentiment tendency of each tweet, the category with the most sentiment tendency among the user's 100 historical tweets is selected as the user's sentiment tendency after statistical analysis, using the following formula:

[0096]

[0097] In the formula, n represents the number of tweets;

[0098] like Figure 3 As shown, the characteristics of psychological motivation include four factors: anxiety, uncertainty, attention-seeking, and participation.

[0099] Anxiety factors were measured using the anxiety category (anx) of the LIWC (Linguistic Inquiry and Word Count analyzer) dictionary, which includes words such as nervous and afraid. The user's anxiety level was measured by counting the number of words in the anxiety category of the LIWC dictionary used in the user's tweets.

[0100] In addition, research has shown that users who use social media at night or in the early morning are often in a state of anxiety. To determine the user's state, this embodiment also counted the time periods when users posted, divided into four time periods: early morning, morning, afternoon, and night, to explore whether the behavior of spreading false information is related to their activity time. The specific time schedule is shown in Table 1:

[0101] Table 1. Numerical Table of User Posting Time Groups

[0102] Time division Corresponding category Corresponding value 0：00-07：00 early morning 0 07：00-12：00 morning 1 12：00-18：00 afternoon 2 18：00-24：00 night 3

[0103] Uncertainty factors are also calculated using categories related to uncertainty in the LIWC dictionary. The degree of uncertainty of a user is represented by the number and percentage of tweets containing words related to uncertainty.

[0104] The attention-seeking factor is determined by calculating the average number of tweets (including retweets) posted by a user over a month from the most recent tweet date. This helps infer whether users with a psychological motivation to seek attention will post frequently to maintain account activity. The formula is as follows:

[0105]

[0106] Engagement factor measures a user's influence by calculating the ratio of followers to subscribers, using the following formula:

[0107]

[0108] TFF (Twitter follower of following) represents the percentage of followers who are followed, num follower num represents the number of followers of a specific account. following This indicates the number of followers a particular account has;

[0109] The personality traits were predicted using an unsupervised testing tool based on 25 of the user's tweets, representing the Big Five personality traits. The output was a score from 0 to 1 for each of the five personality types. The highest-scoring personality type was then selected as the user's representative personality. In this embodiment, the five personality traits are: Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism.

[0110] Step 4: Use a social media trust calculation algorithm to calculate the trust level between users as the user's global trust level;

[0111] Furthermore, step 4 includes the following:

[0112] The global trust calculation for users employs the previously proposed TSM (Trust Score) algorithm to measure the trust between users, which serves as the edge weight between user nodes. The algorithm takes a graph G(V,E) as input and assigns a pair of complementary trust scores, called Trust and Credibility, to each node in the network. The output is a tuple (ti, tw), where Trust ti quantifies a node's tendency to trust its neighbors, and Trust tw quantifies a neighbor's willingness to trust that node. The TSM algorithm takes the user network, i.e., the directed graph G(V,E), as input and specifies the convergence condition or the maximum allowed number of iterations. In each iteration for each node in the network, Credibility and Credibility are calculated using the following formula:

[0113]

[0114]

[0115] Where u, v, x∈V, ti(v) and tw(u) are the credibility scores of v and u respectively, w(v,x) is the edge weight from v to x, out(v) is the set of target nodes of the outgoing edge of v, in(u) is the set of source nodes of the incoming edge of u, and s is the participation score of the social network, which takes a value between 0 and 1, representing the potential risk that participants take when creating links in the network.

[0116] Step 5: Calculate the number of reposts a user has as their local trust level;

[0117] Furthermore, step 5 includes the following:

[0118] Local trust among users is calculated through retweet behavior. The degree to which a user trusts others is measured by the number of retweets in a user's tweets, denoted as tr1. The degree to which other users trust a user is measured by the average number of retweets of a user's tweets, denoted as tr2.

[0119] Step 6: Employ a dynamic graph attention mechanism to calculate the importance score between users, i.e., the trust score, based on the level of trust between them;

[0120] Furthermore, step 6 includes the following:

[0121] Based on the trust value Tr∈(ti,tw,tr1,tr2) obtained from the trust calculation, an importance score is assigned to each neighbor j (j∈Ni) of node i. First, the trust values ​​of nodes i and j, including local and global trust scores, are concatenated. Then, a parameterized projection matrix W is assigned for linear transformation to obtain a scalar. Finally, a shared attention mechanism a (single-layer feedforward neural network) is used to perform self-attention. The calculation of this mechanism is based on the importance coefficient of the trust. The non-standardized trust score between nodes i and j is represented as follows:

[0122] e ij =a(W[Tr i ||Tr j ])

[0123] Among them, e ij Tr represents the trust score, which measures the importance of node j to i, and Tr represents the trust value obtained through trust calculation.

[0124] Step 7: Normalize the trust score to a value between 0 and 1;

[0125] Furthermore, step 7 includes the following:

[0126] When aggregating neighbor information, it is necessary to normalize the attention coefficients of all neighbors of each node. The normalized attention weight α ij That is the true edge weight:

[0127] Normalize it using the softmax function, as shown in the formula below:

[0128]

[0129] Among them, e ij It normalizes all neighbors of node i.

[0130] The normalization operation leads to the asymmetry of attention weights because the normalization object is not the same for each node during the normalization process, for example, e. ijIt normalizes all neighbors of node i, e ji This normalizes the input for all neighbors of node j, which is well-suited to the situation in social networks. For example, in a social network, a prominent user and an ordinary user follow each other, but the importance of the prominent user to the ordinary user is clearly different from the ordinary user's importance to the prominent user. The complete formula for calculating the attention coefficient in a dynamic graph is shown below:

[0131]

[0132] The attention coefficient α obtained by each node ij Combined into an attention-based adjacency matrix Used for feature aggregation.

[0133] Step 8: Based on the trust level, use a graph convolutional neural network to proportionally aggregate the trust levels of users extracted in Step 3 to obtain an adjacency matrix carrying the trust levels of neighbors.

[0134] Furthermore, step 8 includes the following:

[0135] After obtaining the edge weights between the center node and its neighboring nodes, a graph convolutional neural network (GCN) with two hidden layers is used to aggregate the node credibility features Cr. The model input is the adjacency matrix A, which uses the attention coefficients as edge weights, and the output is the node feature matrix Z, which is an N×F matrix, where N is the number of nodes and F is the feature vector output by each node, as shown in the following formula:

[0136] H (l+1) =a(H (l) A)

[0137] Where H represents the hidden layer, and A represents the attention coefficient matrix, i.e. H(0) = Cr; H (L) =Z; L represents the number of layers, and f(·) represents the function for computing the hidden states of the graph convolution model;

[0138] The user credibility features are input into a graph convolutional neural network with two hidden layers, as shown in the following formula:

[0139]

[0140] in, I is the identity matrix, which ensures that its own features are included during the aggregation of the credibility features of neighbors; yes The diagonal matrix of node degree, This is equivalent to dividing all elements in each row by the out-degree of the node in that row.

[0141] Step 9: Use a feedforward neural network to learn node representations carrying neighbor information;

[0142] Furthermore, step 9 includes the following:

[0143] Using the credibility features of nodes in the neighborhood of node i and the network structure, node representations are learned from the graph using an attention-based adjacency matrix paired with a feedforward neural network, as shown in the following formula:

[0144]

[0145] Where X represents the credibility feature. W(0) and W(1) are the weight matrices from the input layer to the hidden layer and from the hidden layer to the output layer, respectively.

[0146] Step 10: Use the cross-entropy loss function to classify nodes and output a ternary classification of users when faced with false information: susceptible user / harmless user / debunking user.

[0147] Furthermore, step 10 includes the following:

[0148] Gradient descent is used to learn the weight matrix, followed by cross-entropy loss for node classification, as shown in the following formula:

[0149]

[0150] Among them, Y l The label represents the node, and f represents the credibility feature.

[0151] Example:

[0152] This embodiment considers a news event T and its information dissemination network, where V is the set of nodes in the dissemination network and E represents the forwarding relationships. Since the presence of social bots can affect the experimental results, this embodiment uses a social bot identification algorithm to filter out social bots and analyzes real users VK. User nodes that may participate in information dissemination are classified into users prone to spreading false information (false information susceptible users) VF, users prone to spreading debunking information (debunking users) VT, and users who do not spread information (harmless users) VN. First, a dynamic graph attention mechanism is used to assign importance scores (i.e., the weight of each edge) to each user's neighbor nodes based on the trust feature (Tr) between users. Then, the trust features (Cr) of neighbor nodes and the center node are aggregated proportionally according to the weights. The trust feature Cr includes basic features Cr-b, emotional features Cr-s, psychological motivation features Cr-psy, and personality features Cr-per. Finally, users are classified into a ternary category VT / VN / VB. The model architecture diagram is shown below. Figure 4 As shown.

[0153] (1) Dataset preparation

[0154] The model was validated using the MinFN dataset, which contains 10 news events. Each news event includes two news articles and their veracity status on the fact-checking website atlnews.in as labels, totaling 20 files. Each file contains information on the original disseminators (forwarders) of the news article, the propagation relationship network formed by these forwarders, the account information of all users involved in the forwarding network, and 100 historical tweets. The user account information provided by MinFN includes fields such as user ID, nickname, total number of tweets (Statuses Count), number of favorites (Favorate Count), number of followers (Followers Count), number of friends (Friends Count), number of lists (Listed Count), authentication status (Verified), whether protected (Protected), account creation time (Created At), and location (Location), as shown in Table 2. The structure of the user's historical tweet information table is shown in Table 3, and the specific nodes, edges, and number of disseminators are shown in Table 4.

[0155] To visualize the information dissemination network, this embodiment uses the dissemination network of false information related to news event N2 as an example. It selects 4,544 nodes and the forwarding relationships between these nodes, and uses Gephi to draw the false information network as follows: Figure 5 As shown, the dissemination network of false information is initiated by major nodes, triggering widespread forwarding, which in turn forms the dissemination network of false information. For each news event, this embodiment will evaluate our model on the false information network (F), the real information network (T), and the network formed by the combination of the two (F+T).

[0156] Table 2 Structure of User Account Information Table

[0157]

[0158]

[0159] Table 3. Structure of the Tweet Information Table

[0160] Field value illustrate Tweet ID The unique identifier of a tweet Tweet type Tweet Type (Original Tweet / Retweet with Comment / Retweet without Comment) Tweet text Tweet text Created at Creation time Favorite count The tweet received [number] likes Retweet count This tweet was retweeted a number of times

[0161] Table 4. Social Network Tables of the MinFN Dataset

[0162]

[0163] Explanation: N1-N 10This represents the number of 10 news events, V represents the number of nodes, E represents the number of edges, and Sp represents the number of original disseminators of the tweet.

[0164] The MinFN dataset is enormous. Taking news event N2 as an example, all the user tweet files involved total 50GB. Selecting all files would burden the computational load of subsequent text processing. Furthermore, the data in the real and fake news networks is unbalanced. Therefore, this embodiment retains the propagation (Sp) nodes and neighbor nodes within 5 hops in the propagation network from the real and fake news networks of the 10 news events in the dataset, while deleting user nodes with empty user profiles. This achieves the required data volume for the experiment and balances the number of users in the real and fake news networks. The final dataset after processing is shown in Table 5. The propagator users in the real and fake news network of news event N4 are extremely unbalanced. This embodiment removes 2547 original forwarders of fake news, retaining the network structure of the remaining nodes to achieve a balance in the number of user nodes in the real and fake news networks.

[0165] Table 5 shows the MinFN dataset after processing.

[0166]

[0167] (2) Parameter settings

[0168] The Trust Score (TSM) algorithm is used to generate topology-based trust metrics by running on each network (including F, T, and F+T networks) to obtain ti and tw for all nodes. Hyperparameters are set using previous recommendations: 100 iterations and a participation score s of 0.391. In each iteration, for each node in the network, trust and credibility are calculated using the equations given in the TSM algorithm equations. Once the metrics have been calculated for each node in the network, the scores are standardized and selected based on user-defined normalization constraints. At the beginning of each iteration, a convergence criterion is checked: the algorithm converges if the difference between the last two iterations is less than the user-defined δ parameter, or if the total number of iterations is greater than the set iteration count k. The detailed process of the algorithm is shown in Algorithm 1.

[0169]

[0170]

[0171] Having obtained the optimal parameter configuration for the model, this embodiment sequentially tunes the learning rate, number of iterations, optimization algorithm, batch size, dropout, number of GAT layers, and number of graph attention heads. It has been verified that the model performs best when the number of GAT layers is set to 2 and the number of graph attention heads is set to 1. The final parameters are shown in Table 6.

[0172] Table 6. Experimental Parameters for Susceptible User Identification

[0173] Parameter name Parameter value Learning rate 0.001 Number of iterations 200 dropout 0.2 batch size 64 Number of sampling neighbors 50 depth 1 GAT layers 2

[0174] (3) Experiment and results comparison

[0175] To verify the effectiveness of the basic features (Cr-b), sentiment features (Cr-s), psychological motivation features (Cr-psy), and personality features (Cr-per) extracted in this embodiment of credibility features (Cr), a comparative experiment was conducted based on a mixed true and false information dissemination network. All disseminators were labeled as spreading true / false information. The classification algorithm used was a traditional machine learning classification algorithm—Support Vector Machine (SVM). These features were added sequentially and compared, outputting a binary classification of users spreading false information / users debunking rumors. Evaluation metrics were used to determine the model's effectiveness, and the results were finally plotted in the experimental results table, as shown in Table 7. SVMA represents based solely on the user's basic features, SVMB represents based on basic features and psychological motivation features, SVMC represents based on basic features, psychological motivation features, and sentiment features, and SVMD represents SVMC with the addition of personality features.

[0176] Table 7 shows the algorithm performance for identifying susceptible users using only credibility features.

[0177]

[0178] Experiment 1 revealed that adding each additional dimension of credibility features improved various metrics for identifying users susceptible to misinformation, demonstrating that each dimension is effective in this area. Specifically, adding psychological motivation features to the second model significantly improved the F1 score, indicating that psychological motivation features are more effective in identifying users susceptible to misinformation. The fourth model, using all four dimensions, achieved the best results in identifying susceptible users, demonstrating that combining these four dimensions effectively improves the model's accuracy. Therefore, all four dimensions can be used to identify users susceptible to misinformation.

[0179] To validate the model, a baseline comparison was conducted using the SCARLET framework based on a static graph attention network across 10 news events in the dataset, representing the fake information network (F), the real information network (T), and a hybrid network (F+T). All four comparative experiments utilized the complete four-dimensional features proposed earlier. The difference lay in the attention coefficient calculation and feature aggregation stages, where static and dynamic attention mechanisms were used to calculate the attention coefficients, respectively, and the node credibility was aggregated using the Graph Convolutional Network (GCN) and the GraphSAGE algorithm. The final experimental results are shown in Table 8.

[0180] Table 8. Algorithm performance of static / dynamic attention mechanisms combined with GCN and GraphSAGE

[0181]

[0182] Note: SAGES indicates the use of a static graph attention mechanism with the GraphSAGE model; the SCARLET model uses a static graph attention mechanism with the GCN model; SAGED indicates the use of a dynamic graph attention mechanism with the GraphSAGE model; GCND is a frame, indicating the use of a dynamic graph attention mechanism with the GCN model.

[0183] As shown in Table 8, the framework proposed in this embodiment performs best in the fake information propagation network, followed by the real information propagation network. However, it achieves the best F1 score in all three networks, with recall exceeding precision, indicating that the model performs better in identifying users susceptible to fake information than in identifying users debunking fake information. Furthermore, under the same aggregation method, the dynamic graph attention mechanism outperforms the static graph attention mechanism, and in the feature aggregation step, GCN in the fake (F) network performs better than GraphSAGE, but this is not the case in the real (T) network. This may be because real information propagators have fewer labels than those in the fake information network. In the hybrid real / fake information network, our model framework also performs better than other models, further demonstrating the superiority of the dynamic graph attention mechanism combined with graph convolutional neural networks.

[0184] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for identifying susceptible users of fake information based on dynamic graph attention networks, characterized in that, Includes the following steps: Step 1: Map the news dissemination relationship graph in the social networking platform into a directed graph; Step 2: Use a deep learning model to identify social bots in the social network and delete social bot nodes from the directed graph; Step 3: Extract features from four dimensions based on the user's historical tweets and basic user information. After feature filtering and verification, integrate them into user credibility features, namely basic user features, emotional features, psychological features, and personality features. Step 3 includes the following: Based on the user's basic information, usable basic user characteristics are extracted, including authentication status and registration duration. The registration duration is calculated in days, with a specific date as the time node. Emotional features were extracted using the NRC-EIL emotional intensity dictionary and a degree adverb dictionary. Based on users' historical tweets, their emotional tendencies were categorized into eight types: funny, happy, sad, afraid, unhappy, indifferent, encouraging, and angry. The calculation process included calculating the emotional tendencies of sentences and tweets, and finally classifying the user's overall emotion. Specifically, the user's emotional score was calculated through the following three steps: For each user's tweet, after data preprocessing, sentence segmentation, and stop word removal, negative words were selected. Sentiment scores for each sentence were calculated based on the NRC-EIL dictionary and a degree adverb dictionary, using the following formula: Senti_score = (-1) 否定词个数 × Degree value × Sentiment word score After obtaining the sentiment score for each sentence, the sentiment words of the same category in a tweet are summed, and the scores of these sentiment categories are compared. The sentiment category with the largest absolute value is taken as the sentiment tendency of the tweet, as shown in the following formula: Senti t =max(|sentiment category, sum(sentiment word score)|) After obtaining the sentiment tendency of each tweet, the category with the most sentiment tendency among the user's 100 historical tweets is selected as the user's sentiment tendency after statistical analysis, using the following formula: In the formula, n represents the number of tweets; Psychological motivational characteristics include four factors: anxiety, uncertainty, attention-seeking, and engagement. Anxiety factors were measured using the anxiety category in the LIWC dictionary, which includes tension and fear. The level of anxiety was measured by counting the number of words in the anxiety category in the LIWC dictionary used in user tweets. Uncertainty factors are also calculated using categories related to uncertainty in the LIWC dictionary. The degree of uncertainty of a user is represented by the number and percentage of tweets containing words related to uncertainty. The attention-seeking factor is determined by calculating the average number of tweets posted by a user over a 30-day period, including retweets, within a month of the most recent tweet posted by that user. This helps infer whether users with a psychological motivation to seek attention will post frequently to maintain account activity. The formula is as follows: Engagement factor measures a user's influence by calculating the ratio of followers to subscribers, using the following formula: Among them, TFF represents the fan follower ratio, num follower num represents the number of followers of a specific account. following This indicates the number of followers a particular account has; The personality traits are predicted by using an unsupervised testing tool to analyze 25 of the user's tweets to identify the Big Five personality traits. The output is the user's score between 0 and 1 for each of the five personality types. Finally, the personality type with the highest score among the five personality types is taken as the user's representative personality. Step 4: Use a social media trust calculation algorithm to calculate the trust level between users as the user's global trust level; Step 5: Calculate the number of reposts a user has as their local trust level; Step 6: Employ a dynamic graph attention mechanism to calculate the importance score between users, i.e., the trust score, based on the level of trust between them; Step 7: Normalize the trust score to a value between 0 and 1; Step 8: Based on the trust level, use a graph convolutional neural network to proportionally aggregate the trust levels of users extracted in Step 3 to obtain an adjacency matrix carrying the trust levels of neighbors. Step 9: Use a feedforward neural network to learn node representations carrying neighbor information; Step 10: Use the cross-entropy loss function to classify nodes and output a ternary classification of users when faced with false information: susceptible user / harmless user / debunking user.

2. The method for identifying susceptible users of false information based on dynamic graph attention networks according to claim 1, characterized in that, Step 1 includes the following: Based on the public dataset MinFN containing ten news events N1-N 10 Each news event includes a true news item and a fake news item, along with their propagation relationships. These relationships are mapped to a directed graph on the social network. Nodes in the directed graph represent users who spread the news, and edges represent the propagation relationships between users. The direction of the edges indicates the direction of news transmission.

3. The method for identifying susceptible users of false information based on dynamic graph attention networks according to claim 2, characterized in that, Step 2 includes the following: To eliminate the influence of social bots on experimental results, a tweet-level model is proposed, which is an LSTM model that incorporates auxiliary information. User account information is used as auxiliary input. The tweet text is then segmented into sentences and words, converted into a set of word vectors by a pre-trained BERT model, and then fed into the LSTM model. The output is then concatenated with the auxiliary information input and fed into a two-layer fully connected neural network with ReLU activation function to generate the final output, achieving binary classification between social bots and natural users. After identifying social bot users, social bot nodes are removed from the information propagation network.

4. The method for identifying susceptible users of false information based on dynamic graph attention networks according to claim 3, characterized in that, Step 4 includes the following: The global trust calculation for users employs the previously proposed TSM algorithm to measure the trust between users, which serves as the edge weight between user nodes. The algorithm takes a graph G(V,E) as input and assigns each node in the network a pair of complementary trust scores, called trust level and credibility. The output is a tuple (ti, tw), where trust level ti quantifies a node's tendency to trust its neighbors, and trust level tw quantifies a neighbor's willingness to trust that node. Where u, v, x∈V, ti(v) and tw(u) are the credibility scores of v and u respectively, w(v,x) is the edge weight from v to x, out(v) is the set of target nodes of the outgoing edge of v, in(u) is the set of source nodes of the incoming edge of u, and s is the participation score of the social network, which takes a value between 0 and 1, representing the potential risk that participants take when creating links in the network.

5. The method for identifying susceptible users of false information based on dynamic graph attention networks according to claim 4, characterized in that, Step 5 includes the following: Local trust among users is calculated through retweet behavior. The degree to which a user trusts others is measured by the number of retweets in a user's tweets, denoted as tr1. The degree to which other users trust a user is measured by the average number of retweets of a user's tweets, denoted as tr2.

6. The method for identifying susceptible users of false information based on dynamic graph attention networks according to claim 5, characterized in that, Step 6 includes the following: Based on the trust value Tr∈(ti,tw,tr1,tr2) obtained from the trust calculation, an importance score is assigned to each neighbor j (j∈Ni) of node i: First, the trust values ​​of nodes i and j, including local and global trust scores, are concatenated. Then, a parameterized projection matrix W is assigned for linear transformation to obtain a scalar. Finally, a shared attention mechanism a is used to perform self-attention. The calculation of this mechanism is based on the importance coefficient of the trust. The non-normalized trust score between nodes i and j is represented as follows: e ij =a(W[Tr i ||Tr j ]) Among them, e ij Tr represents the trust score, which measures the importance of node j to i, and Tr represents the trust value obtained through trust calculation.

7. The method for identifying susceptible users of false information based on dynamic graph attention networks according to claim 6, characterized in that, Step 7 includes the following: When aggregating neighbor information, it is necessary to normalize the attention coefficients of all neighbors of each node. The normalized attention weight α is then calculated. ij That is the true edge weight: Normalize it using the softmax function, as shown in the formula below: Among them, e ij It normalizes all neighbors of node i.

8. The method for identifying susceptible users of false information based on dynamic graph attention networks according to claim 7, characterized in that, Step 8 includes the following: After obtaining the edge weights between the center node and its neighboring nodes, a graph convolutional neural network (GCN) with two hidden layers is used to aggregate the node credibility features Cr. The model input is the adjacency matrix A, which uses the attention coefficients as edge weights, and the output is the node feature matrix Z, which is an N×F matrix, where N is the number of nodes and F is the feature vector output by each node, as shown in the following formula: H (l+1) =f(H (l) ,A) Where H represents the hidden layer, and A represents the attention coefficient matrix, i.e. H(0) = Cr; H (L) =Z; L represents the number of layers, and f(·) represents the function for computing the hidden states of the graph convolution model; The user credibility features are input into a graph convolutional neural network with two hidden layers, as shown in the following formula: in, I is the identity matrix, which ensures that its own features are included during the aggregation of the credibility features of neighbors; yes The diagonal matrix of node degree, This is equivalent to dividing all elements in each row by the out-degree of the node in that row.

9. The method for identifying susceptible users of false information based on dynamic graph attention networks according to claim 8, characterized in that, Step 9 includes the following: Using the credibility features of nodes in the neighborhood of node i and the network structure, node representations are learned from the graph using an attention-based adjacency matrix paired with a feedforward neural network, as shown in the following formula: Where X represents the credibility feature. W (0) and W (1) These are the weight matrices from the input layer to the hidden layer and from the hidden layer to the output layer, respectively.

10. The method for identifying susceptible users of false information based on a dynamic graph attention network according to claim 9, characterized in that, Step 10 includes the following: Gradient descent is used to learn the weight matrix, followed by cross-entropy loss for node classification, as shown in the following formula: Among them, Y l The label represents the node, and f represents the credibility feature.

Images