Social network violent aggressive user detection method and system based on moral convergence

Abstract

The invention provides a social network violent aggressive user detection method and system based on moral convergence, and the method comprises the steps: firstly obtaining a moral foundation in a user social network text, constructing a moral framework, then constructing a user representation combined with the moral framework, and combining with the user embedding learning of TaoDean identity fusion, thereby achieving the detection of the violent aggressive user of the social network. And finally, realizing violent aggressive user identification through a multi-layer perceptron by utilizing learned user embedding. According to the method, the moral convergence phenomenon of the social network user's ideology level can be modeled, and violent aggressive users in the social network can be accurately detected, so that the problem of inaccurate detection due to neglect of the ideology level influence factors in the existing detection method is effectively solved.

Description

Technical Field

[0001] The present invention relates to the technical field of detecting violent radicalized users on social networks, and in particular to a method and system for detecting violent radicalized users on social networks based on moral convergence. Background Art

[0002] The development of social media has promoted extensive exchanges of ideas and opinions. People can understand each other's views and ideas through posting, forwarding, replying and other means. Therefore, social media is often used by so-called "extremist" groups. They often operate anonymously, promote radical views, use political and social issues to divide society, and change other people's thoughts, attitudes or behaviors. Social media users are at risk of radicalization, and serious radicalization can lead to real-life violent behavior. In order to curb the occurrence of violent phenomena, the identification of violent radicalization on social networks has become an important research focus. Existing literature has achieved some encouraging results by analyzing user characteristics such as emotions and user activity to identify violent radicalization, but these studies ignore the impact of ideology and values ​​on violent radicalization. Summary of the invention

[0003] In order to overcome the deficiency of existing methods for identifying violent radicalized users that ignore user ideology, the present invention provides a method and system for detecting violent radicalized users in social networks based on moral convergence, which can model the moral convergence phenomenon at the ideological level of social network users and accurately detect violent radicalized users in social networks, thereby effectively solving the problem of inaccurate detection caused by existing detection methods ignoring influencing factors at the ideological level.

[0004] The technical solution of the present invention is:

[0005] The method for detecting violent radicalized users on social networks based on moral convergence comprises the following steps:

[0006] Step 1: Obtain the moral foundation in users’ social network texts and construct a moral framework:

[0007] Extract entities from user social network texts, establish relationship definition rules and constraints for entities, predict the moral basis of user social network texts through inference rules and constraints, obtain word vectors of important unigrams of each moral basis, and concatenate word vectors to obtain the moral framework;

[0008] Step 2: Build a user representation that incorporates an ethical framework:

[0009] Obtain unigrams of user social network texts, calculate the frequency and mutual information score of unigrams in user social network texts, and comprehensively sort them by frequency and mutual information score to obtain unigrams at the top of the ranking; encode the unigrams at the top of the ranking by BERT and use PCA to reduce the dimension, obtain the first principal component after the unigram dimension reduction, and concatenate them to obtain the basic profile P of each user. i ; Use the attention mechanism to obtain the moral framework that the user pays attention to, and finally obtain the user embedding combined with the moral framework as the user representation;

[0010] Step 3: Learning user embeddings combined with moralized identity fusion:

[0011] Use the user representation combined with the moral framework obtained in step 2 as node information, construct a heterogeneous graph G with users as nodes and user relationships as edges;

[0012] Calculate the similarity of user moral frameworks and set a similarity threshold to determine whether there is an edge between user nodes to obtain the identity fusion relationship graph IFG;

[0013] Combine the original heterogeneous graph G with IFG to obtain the refined graph G';

[0014] Based on the refined graph G', a relational graph neural network is used to learn user embedding;

[0015] Step 4: Using the user embedding learned in step 3, identify violently radicalized users through a multi-layer perceptron.

[0016] Furthermore, in step 1, entities in the user's social network text are extracted, and relationship definition rules and constraints of the entities are established. The process of predicting the moral basis of the user's social network text by inference rules and constraints is as follows:

[0017] Define the moral foundation m in the user's social network text t and the entity radical form k that has a dependent relationship with the moral foundation; establish the relationship definition rules and constraints of the entity as follows:

[0018]

[0019] r1:

[0020] r2:

[0021] r3:

[0022] Rule r1 models the dependency between text t and moral foundation m, rule r2 models the dependency between text t and radical morphology k, and constraint r3 is the constraint between time, radicality and text;

[0023] According to the relationship definition rules and constraints, a factor graph is constructed using the user social network text t and the entity radical form k that has a dependency relationship with the moral foundation, and MAP (Maximum A Posteriori) reasoning is performed to obtain the optimal score assignment, and based on the MAP reasoning result, the moral foundation m of the user social network text is obtained. i .

[0024] Furthermore, in step 1, before constructing the factor graph, for rules r1 and r2, BERT is used to encode the user social network text t, and for rule r2, a feedforward neural network is used to encode the radical form k, and the encoding of the radical form k is connected with the BERT (Bidirectional Encoder Representations from Transformers) encoding of the user social network text t to obtain the final rule-based representation.

[0025] Furthermore, in step 1, the word vectors of the important unigrams of each moral foundation are obtained, and the word vectors are concatenated to obtain the moral framework process as follows:

[0026] The frequency of the unigrams of each moral foundation and the mutual information scores between the unigrams and the moral dictionary vocabulary are obtained, and the unigrams in the top several positions are obtained by combining the frequency and mutual information scores. The unigrams are then reduced in dimension using PCA (Principal Component Analysis), and the first principal component is used as the final representation of the unigrams e(m i ) ; concatenate the final representations of the unigrams in the first several positions of the sort to obtain the moral basis m i The moral framework of i .

[0027] Furthermore, in step 2, the attention mechanism is used to obtain the moral framework that the user pays attention to, and the process of obtaining the user embedding X combined with the moral framework is:

[0028] According to the formula

[0029] α ij =softmax j (a(M j ,P i ))

[0030] Calculate the user's basic profile P i Corresponding to a certain moral foundation Moral framework M j The attention coefficient α ij , a(·) is the similarity function;

[0031] Use the attention coefficient α ijWeight different moral frameworks to obtain a comprehensive moral framework corresponding to different moral foundations of users

[0032] M u =∑α ij M j

[0033] Finally, the user embedding X combined with the moral framework is obtained as:

[0034] X=concat(M u ,P i ).

[0035] Furthermore, in step 3, the edge between the node pairs ui and uj in the identity fusion graph IFG is obtained by the following formula:

[0036]

[0037] Among them, ∈∈[0,1] is the threshold used to control the density of generated IFG; S(M ui ,M uj ) is a parameterized function used to determine the user’s degree of moral integration.

[0038] Furthermore, in step 3, based on the refined graph G', the specific process of using the relational graph neural network to learn user embedding is:

[0039] The relationship graph neural network includes multiple R-GCNs layers; the input is the adjacency matrix A∈R under each user relationship |V|×|V| , and the feature matrix X of the user node set V |V|×|V| , the output is user embedding; the propagation rule of each layer is expressed as:

[0040]

[0041] in, is the representation of user embedding i at layer l+1, and is the set learnable weight, represents the neighbor set of user embedding i under user relation r∈R, and R is a set of different user relations.

[0042] Furthermore, in step 3, when learning user embedding, the joint loss during training is

[0043]

[0044] in To detect the loss, is the identity fusion loss, λ gsr is the hyperparameter of the identity fusion process, is the self-supervision loss, λ ssl is the hyperparameter of the self-supervision process.

[0045] Furthermore, the detection loss The cross entropy loss between a given user label and a predicted result is: ν c is the set of labeled user nodes, Y i For a given user tag, To predict the results;

[0046] Identity Fusion Loss A' is the adjacency matrix of G';

[0047] Self-supervised loss Y iq For a node pair (v i ,v q ), Prediction labels obtained by combining self-supervised graph neural networks.

[0048] The social network violent radical user detection system based on moral convergence includes: a moral modeling module using relationship learning, a user representation construction module combined with a moral framework, a user embedding learning module combined with moral identity fusion, and a violent radical user identification module;

[0049] The moral modeling module using relational learning receives user social network text, extracts entities from the user social network text, defines rules and constraints according to the relationships between the entities, predicts moral foundation labels by inferring the rules and constraints, thereby obtaining word vectors of important unigrams of each moral foundation, and concatenates the word vectors to obtain a moral framework;

[0050] The user representation construction module combined with the moral framework generates a user representation based on a unigram in the user's social network text, and fuses the user representation with the moral framework to obtain a user representation combined with the moral framework as an input for subsequent stages;

[0051] The user embedding learning module combined with moral identity fusion obtains multiple user relationships, and constructs a heterogeneous graph with users as nodes and multiple user relationships as edges, and uses the user representation combined with the moral framework as node information to learn user embedding using a relationship graph neural network;

[0052] The violent radicalized user identification module is implemented by a multi-layer perceptron, the input of which is user embedding, and the output of which is a classification label of whether the user is a violent radicalized user.

[0053] Beneficial Effects

[0054] In view of the problem that the traditional method of identifying violent radical users in social networks ignores the user ideology, the present invention extracts entities in the user social network text, establishes the relationship definition rules and constraints of the entities, predicts the moral basis of the user social network text through inference rules and constraints, obtains the word vectors of the important unigrams of each moral basis, and splices the word vectors to obtain the moral framework, and generates the user embedding combined with the moral framework as the user representation; in addition, the user relationship is used to construct a heterogeneous graph, and the identity fusion relationship graph of the user's moralization is constructed based on the similarity threshold function. The two are added to obtain a refined graph. These operations give full play to the role of the user's ideological level in violent radicalization identification. In addition, the user embedding is learned by using a relationship graph neural network based on self-supervised learning to provide a guarantee for the accuracy of the model. Compared with the traditional method of identifying violent radical users in social networks, the method of the present invention models and obtains the user embedding under the user moral framework at the ideological level and captures the phenomenon of moral convergence, fully utilizes the moral fusion characteristics at the ideological level, comprehensively considers user characteristics, and enables user ideology to play a role in the identification of violent radical users, providing new ideas for the accurate identification of violent radical users on social networks and providing intelligent assistance for network security construction.

[0055] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which:

[0057] Figure 1 This is a principle block diagram of the method of the present invention. DETAILED DESCRIPTION

[0058] Embodiments of the present invention are described in detail below. The embodiments are exemplary and intended to be used to explain the present invention, but should not be construed as limiting the present invention.

[0059] In response to the demand for detecting violently radicalized users in social networks, the present invention proposes a method for detecting violently radicalized users in social networks based on moral convergence, which can model the moral convergence phenomenon at the ideological level of social network users and accurately detect violently radicalized users in social networks, thereby effectively solving the problem of inaccurate detection in existing detection methods due to ignoring the influencing factors at the ideological level.

[0060] like Figure 1 As shown in FIG. 1 , the method for detecting violent radicalized users on social networks based on moral convergence includes the following steps:

[0061] Step 1: Obtain the moral foundation in users’ social network texts and construct a moral framework:

[0062] Extract entities from user social network texts, use combined neural symbols to represent the dependency between moral foundations and related entities, establish relationship definition rules and constraints for entities, and predict the moral foundations of user social network texts by inferring rules and constraints.

[0063] Define the moral foundation m in the user's social network text t and the entity radical form k that has a dependency relationship with the moral foundation as the method input; in order to define the dependency relationship between entities, the relationship definition rules and constraints of the entities are established as follows:

[0064]

[0065] And use the usage rule set to model the dependencies between relationships:

[0066] r1:

[0067] r2:

[0068] r3:

[0069] Rule r1 models the dependency between text t and moral foundation m, rule r2 models the dependency between text t and radical form k, and constraint r3 is the constraint between time, radicality and text.

[0070] According to the relationship definition rules and constraints, a factor graph is constructed using the user social network text t and the entity radical form k that has a dependency relationship with the moral foundation, and MAP reasoning is performed to obtain the optimal score assignment, and based on the MAP reasoning result, the moral foundation m of the user social network text is obtained. i To formalize this process, the rules are written as linear inequalities, so that MAP inference is defined as a linear program as follows:

[0071]

[0072] where each rule r has input features x r and the predictor variable y r , defines the potential energy ψ r (x r ,y r ), with weight ≡ r Added to the linear program.

[0073] like Figure 1As shown in the figure, before constructing the factor graph, for rules r1 and r2, BERT is used to encode the user social network text t. For rule r2, a feedforward neural network is used to encode the radical form k in Onehot form, and the encoding of the radical form k is connected with the BERT encoding of the user social network text t to obtain the final rule-based representation.

[0074] After obtaining the moral foundation of the user's social network text, the word vectors of the important unigrams of each moral foundation are obtained, and the word vectors are concatenated to obtain the moral framework, which is as follows:

[0075] Identify the moral foundation of each social network text of the user, obtain the word frequency of each moral foundation unigram and the mutual information score between the unigram and the moral dictionary vocabulary, calculate the product of the mutual information score and the frequency of each unigram for comprehensive sorting, and obtain the unigrams whose product is in the top 200 of the sorting; and use PCA to reduce the dimension of the unigram, and use the first principal component as the final representation of the unigram e(m i ) ; concatenate the final representations of the top 200 unigrams to obtain the moral foundation m i The moral framework of i .

[0076] Step 2: Build a user representation that incorporates an ethical framework:

[0077] like Figure 1 As shown in the figure, the unigrams of the user's social network text are obtained, the word frequency and mutual information score of the unigrams in the user's social network text are calculated, the product of the mutual information score and the frequency of each unigram is calculated for comprehensive sorting, and the unigrams in the top 200 of the product sorting are obtained; the unigrams in the top 200 are encoded by BERT and reduced in dimension using PCA, the first principal component after the unigram dimension reduction is obtained and concatenated to obtain the basic profile P of each user. i In this embodiment, the top 10 unigrams of a certain user are:

[0078]

[0079] The attention mechanism is used to obtain the moral framework that a user i pays attention to, and finally the user embedding combined with the moral framework is obtained as the user representation; according to the formula

[0080] α ij =softmax j (a(M j ,P i ))

[0081] Calculate user basic profile P i Corresponding to a certain moral foundation Moral framework M j The attention coefficient α ij, a(·) is the similarity function; the attention coefficient α is used ij Weight different moral frameworks to obtain a comprehensive moral framework corresponding to different moral foundations of users

[0082] M u =∑α ij M j

[0083] Finally, the user embedding X combined with the moral framework is obtained as:

[0084] X=concat(M u ,P i ).

[0085] Step 3: Learning user embeddings combined with moralized identity fusion:

[0086] like Figure 1 As shown, the user representation combined with the moral framework obtained in step 2 is used as node information, and a heterogeneous graph G is constructed with users as nodes and user relationships as edges. There are three types of relationship edges in the heterogeneous graph G, including user-follow-user, user-reply-user, and user-mention-user. An independent adjacency matrix is ​​constructed for each relationship to distinguish them.

[0087] Calculate the similarity of the user's moral framework, set the similarity threshold, determine whether there is an edge between user nodes, and obtain the identity fusion relationship graph IFG. The edge between the node pairs ui and uj in the identity fusion relationship graph IFG is obtained by the following formula:

[0088]

[0089] Among them, ∈∈[0,1] is the threshold used to control the generated IFG density,

[0090] After experiments, the test results of threshold ∈ are:

[0091]

[0092] Therefore, in this embodiment, ∈=0.90; S(M ui ,M uj ) is a parameterized function used to determine the degree of moral integration of users, which is achieved by calculating the cosine similarity of the moral frameworks that two users follow:

[0093] S(M ui ,M uj )=Γ(M ui ,M uj ),

[0094] Where Γ represents the similarity function, and cosine similarity is used here.

[0095] The original heterogeneous graph G is combined with IFG to obtain a refined graph G', and embedding learning is continued on this refined graph. In order to avoid overfitting of downstream tasks, the identity fusion loss is obtained through the L1 norm A' is the adjacency matrix of G'.

[0096] Based on the refined graph G', a relational graph neural network is used to learn user embedding;

[0097] The relationship graph neural network includes multiple R-GCNs layers; the input is the adjacency matrix A∈R under each user relationship |V|×|V| , and the feature matrix X of the user node set V |V|×|V| , the output is user embedding; the propagation rule of each layer is expressed as:

[0098]

[0099] in, is the representation of user embedding i at layer l+1, and is the set learnable weight, represents the neighbor set of user embedding i under user relation r∈R, and R is a set of different user relations.

[0100] For each user node v i ∈V, h i Output to a fully connected layer and softmax function to predict whether the user is a radicalized user, the corresponding detection loss The cross entropy loss between a given user label and a predicted result is: ν c is the set of labeled user nodes, Y i For a given user tag, For the prediction results.

[0101] In this embodiment, user embedding learning is enhanced by a similarity-based self-supervision module. Specifically, for a node v i , define two sets of nodes V hi and V li , respectively containing M and v i There are nodes with high similarity and M nodes with v i Based on the similarity between node pairs, V hi The nodes in and the nodes v i The node pairs formed are defined as positive samples, and V li The nodes in and the nodes v iThe node pairs formed are defined as negative samples. In each round of training, each node pair (v i , v q ) are connected by embedding, denoted as X iq , and assign a label Y to each node pair embedding iq When v q ∈V hi When Y iq =1, when v q ∈V li When Y iq = 0. Then, embed the node pair into X iq Feed into a fully connected neural network to get predicted labels The self-supervised loss is

[0102]

[0103] Then when learning user embedding, the joint loss during training is

[0104]

[0105] in To detect the loss, is the identity fusion loss, λ gsr is the hyperparameter of the identity fusion process, is the self-supervision loss, λ ssl is the hyperparameter of the self-supervision process. After experiments, λ gsr The test results are:

[0106]

[0107] λ ssl The test results are:

[0108]

[0109] Therefore, in this embodiment, λ gsr =0.01, in this embodiment, λ ssl =1.

[0110] Step 4: Using the user embedding learned in step 3, identify violently radicalized users through a multi-layer perceptron.

[0111] Based on the above method, this embodiment also proposes a social network violent radicalization user detection system based on moral convergence, including: a moral modeling module using relationship learning, a user representation construction module combined with a moral framework, a user embedding learning module combined with moral identity fusion, and a violent radicalization user identification module;

[0112] The moral modeling module using relational learning receives user social network text, extracts entities from the user social network text, defines rules and constraints according to the relationships between the entities, predicts moral foundation labels by inferring the rules and constraints, thereby obtaining word vectors of important unigrams of each moral foundation, and concatenates the word vectors to obtain a moral framework;

[0113] The user representation construction module combined with the moral framework generates a user representation based on a unigram in the user's social network text, and fuses the user representation with the moral framework to obtain a user representation combined with the moral framework as an input for subsequent stages;

[0114] The user embedding learning module combined with moral identity fusion obtains multiple user relationships, and constructs a heterogeneous graph with users as nodes and multiple user relationships as edges, and uses the user representation combined with the moral framework as node information to learn user embedding using a relationship graph neural network;

[0115] The violent radicalized user identification module is implemented by a multi-layer perceptron, the input of which is user embedding, and the output of which is a classification label of whether the user is a violent radicalized user.

[0116] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention without departing from the principles and intent of the present invention.

Claims

1. A method for detecting violent radicalized users on social networks based on moral convergence, characterized by: The following steps are involved: Step 1: Obtain the moral foundation in users’ social network texts and construct a moral framework: Extract entities from user social network texts, establish relationship definition rules and constraints for entities, predict the moral basis of user social network texts through inference rules and constraints, obtain word vectors of important unigrams of each moral basis, and concatenate word vectors to obtain the moral framework; Step 2: Build a user representation that incorporates an ethical framework: Obtain unigrams of user social network texts, calculate the frequency and mutual information score of unigrams in user social network texts, and comprehensively sort them by frequency and mutual information score to obtain unigrams at the top of the ranking; encode the unigrams at the top of the ranking by BERT and use PCA to reduce the dimension, obtain the first principal component after the unigram dimension reduction, and concatenate them to obtain the basic profile P of each user. i ; Use the attention mechanism to obtain the moral framework that the user pays attention to, and finally obtain the user embedding combined with the moral framework as the user representation; Step 3: Learning user embeddings combined with moralized identity fusion: Use the user representation combined with the moral framework obtained in step 2 as node information, construct a heterogeneous graph G with users as nodes and user relationships as edges; Calculate the similarity of user moral frameworks and set a similarity threshold to determine whether there is an edge between user nodes to obtain the identity fusion relationship graph IFG; Combine the original heterogeneous graph G with IFG to obtain the refined graph G'; Based on the refined graph G', a relational graph neural network is used to learn user embedding; Step 4: Using the user embedding learned in step 3, identify violently radicalized users through a multi-layer perceptron.

2. According to claim 1, a method for detecting violent radicalized users in social networks based on moral convergence is characterized by: In step 1, entities in the user's social network text are extracted, and relationship definition rules and constraints of the entities are established. The process of predicting the moral basis of the user's social network text by inference rules and constraints is as follows: Define the moral foundation m in the user's social network text t and the entity radical form k that has a dependent relationship with the moral foundation; establish the relationship definition rules and constraints of the entity as follows: r1: r2: r3: Rule r1 models the dependency between text t and moral foundation m, rule r2 models the dependency between text t and radical morphology k, and constraint r3 is the constraint between time, radicality and text; According to the relationship definition rules and constraints, a factor graph is constructed using the user social network text t and the entity radical form k that has a dependency relationship with the moral foundation, and MAP (Maximum A Posteriori Maximum A Posterior Probability) reasoning is performed to obtain the optimal score assignment, and based on the MAP reasoning result, the moral foundation m of the user social network text is obtained. i .

3. According to claim 2, a method for detecting violent radicalized users in social networks based on moral convergence is characterized by: In step 1, before constructing the factor graph, for rules r1 and r2, BERT is used to encode the user social network text t. For rule r2, a feedforward neural network is used to encode the radical form k, and the encoding of the radical form k is connected with the BERT encoding of the user social network text t to obtain the final rule-based representation.

4. According to claim 1, a method for detecting violent radicalized users in social networks based on moral convergence is characterized by: In step 1, the word vectors of the important unigrams of each moral foundation are obtained, and the word vectors are concatenated to obtain the moral framework process as follows: The frequency of the unigrams of each moral foundation and the mutual information scores between the unigrams and the moral dictionary vocabulary are obtained, and the unigrams in the top several positions are obtained by combining the frequency and mutual information scores. PCA is used to reduce the dimension of the unigrams, and the first principal component is used as the final representation of the unigrams e(m i ) ; concatenate the final representations of the unigrams in the first several positions of the sort to obtain the moral basis m i The moral framework of i .

5. According to claim 1, a method for detecting violent radicalized users in social networks based on moral convergence is characterized by: In step 2, the attention mechanism is used to obtain the moral framework that the user pays attention to. The process of obtaining the user embedding X combined with the moral framework is: According to the formula a ij =softmax j (a(M j ,P i )) Calculate the user's basic profile P i Corresponding to a certain moral foundation Moral framework M j The attention coefficient α ij , a(·) is the similarity function; Use the attention coefficient α ij Weight different moral frameworks to obtain a comprehensive moral framework corresponding to different moral foundations of users M u =∑α ij M j Finally, the user embedding X combined with the moral framework is obtained as: X=concat(M u ,P i )。 6. According to claim 1, a method for detecting violent radicalized users in social networks based on moral convergence is characterized by: In step 3, the edge between the node pairs ui and uj in the identity fusion graph IFG is obtained by the following formula: Among them, ∈∈[0,1] is the threshold used to control the density of generated IFG; S(M ui ,M uj ) is a parameterized function used to determine the user’s degree of moral integration.

7. According to claim 1, a method for detecting violent radicalized users in social networks based on moral convergence is characterized by: In step 3, based on the refined graph G', the specific process of using the graph neural network to learn user embedding is as follows: The relationship graph neural network includes multiple R-GCNs layers; the input is the adjacency matrix A∈R under each user relationship |V|×|V| , and the feature matrix X of the user node set V |V|×|V| , the output is user embedding; the propagation rule of each layer is expressed as: in, is the representation of user embedding i at layer l+1, and is the set learnable weight, represents the neighbor set of user embedding i under user relation r∈R, and R is a set of different user relations.

8. According to claim 1, a method for detecting violent radicalized users in social networks based on moral convergence, characterized in that: In step 3, when learning user embedding, the joint loss during training is in To detect the loss, is the identity fusion loss, λ gsr is the hyperparameter of the identity fusion process, is the self-supervision loss, λ ssl is the hyperparameter of the self-supervision process.

9. A method for detecting violent radicalized users on social networks based on moral convergence according to claim 8, characterized in that: Detection loss The cross entropy loss between a given user label and a predicted result is: v c is the set of labeled user nodes, Y i For a given user tag, To predict the results; Identity Fusion Loss A' is the adjacency matrix of G'; Self-supervised loss Y iq For a node pair (v i ,v q ), Prediction labels obtained by combining self-supervised graph neural networks.

10. A social network violent radicalization user detection system based on moral convergence, characterized by: include: A moral modeling module using relational learning, a user representation building module combined with a moral framework, a user embedding learning module combined with moral identity fusion, and a violent radicalized user identification module; The moral modeling module using relational learning receives user social network texts, extracts entities from the user social network texts, defines rules and constraints according to the relationships between the entities, predicts moral foundation labels by inferring the rules and constraints, thereby obtaining word vectors of important unigrams of each moral foundation, and concatenates the word vectors to obtain a moral framework; The user representation construction module combined with the moral framework generates a user representation based on a unigram in the user's social network text, and fuses the user representation with the moral framework to obtain a user representation combined with the moral framework as an input for subsequent stages; The user embedding learning module combined with moral identity fusion obtains multiple user relationships, and constructs a heterogeneous graph with users as nodes and multiple user relationships as edges, and uses the user representation combined with the moral framework as node information to learn user embedding using a relational graph neural network; The violent radicalized user identification module is implemented by a multi-layer perceptron, the input of which is user embedding, and the output of which is a classification label of whether the user is a violent radicalized user.

Images