A specific user mining method based on group propagation

Abstract

The application discloses a specific user mining method based on group communication, and first proposes the concept of group average communication contribution and calculates the group average communication contribution by analyzing the definition of groups in communication courses and the characteristics of group communication; then the specific Twitter user mining is completed by collecting Twitter data, constructing a follow network, extracting network attribute features, constructing a group clustering algorithm and a group label reassignment algorithm. The method measures the average interaction closeness between individuals and individuals in groups by proposing the group average communication contribution, proposes a Twitter user communication group mining method based on group features, and solves the problem that the clustering results of the existing group clustering algorithm cannot well reflect the group features.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of cyberspace big data information service, and particularly relates to a specific user mining method based on group propagation. BACKGROUND

[0002] With the development of the information age, network socialization has become a part of many people's daily life, and social network platforms have become an important tool for people to conduct network socialization.

[0003] At present, the global larger social network platforms include Weibo, Twitter and FaceBook, etc. Among them, Twitter has a high user volume and popularity in the world, and its daily active user volume has reached about 200 million. The huge information flow in Twitter provides data support for user attribute modeling and group feature analysis, and the data collection method provided by Twitter also provides technical support for actual research.

[0004] The user group mining problem in Twitter can be converted into the clustering problem of user groups. According to the system, the clustering algorithm can be further divided into partition-based clustering, density-based clustering, hierarchical clustering and graph-based clustering. Among them, the partition-based clustering takes the spatial distance between nodes in the class as the clustering target, and the spatial distance between nodes in different classes is far enough; the density-based clustering divides the nodes into groups by judging whether the node density in a certain region around the current node meets the minimum threshold; the hierarchical clustering method generates a hierarchical "tree" by calculating the similarity between nodes of different categories. This "tree" can directly reflect the relationship between the node category distribution and the number of clusters; the graph-based clustering calculates the similarity between nodes as the edge weight through node attribute, determines whether there is an edge between nodes by comparing the similarity and the edge threshold, and then cuts the edge through the edge weight to obtain several connected branches as the result of group division.

[0005] The clustering process of the above existing algorithm does not model the characteristics of the group features mentioned in the propagation course and lacks consideration of the influence of social relations on the interaction behavior between users, so the clustering result cannot better reflect the group features. SUMMARY

[0006] To solve the above technical problems, the application provides a specific user mining method based on group propagation, which first determines the premise that information in a social network can be propagated through the characteristics of group propagation, then analyzes the characteristics that a group and individuals in the group should have through the definition of the group in the propagation course, measures the average close degree of interaction of individuals and individuals in the group through the basic assumption and "group average propagation contribution" defined based on the basic assumption, and finally constructs a Twitter user propagation group mining method based on the characteristics of the group based on the index, so as to solve the problem that the clustering results of the existing group clustering algorithm cannot well reflect the characteristics of the group.

[0007] The technical scheme adopted by the application is as follows: a specific user mining method based on group propagation, and the specific steps are as follows:

[0008] S1, calculating group average propagation contribution;

[0009] S2, collecting Twitter data;

[0010] S3, constructing a follow network and extracting network attribute features;

[0011] S4, constructing a group clustering algorithm and a group label reassignment algorithm, and completing specific Twitter user mining.

[0012] Further, the step S1 is specifically as follows:

[0013] A group is a collection of a plurality of people with a specific common goal and a sense of belonging, and an interactive relationship exists between the people.

[0014] The interactive relationship refers to the follow relationship between users, and the common goal and the sense of belonging are measured. It is provided that the strength of the individual to the common goal and the sense of belonging of the group is reflected by the close degree of interaction between the individual and the group. The concept of group average propagation contribution is proposed to measure the average close degree of interaction between the individual and each node in the group.

[0015] The specific calculation method of the group average propagation contribution is as follows:

[0016]

[0017]

[0018]

[0019]

[0020] wherein, represents the node average propagation contribution within the group, the influence of a node on every node within the group and the ability of a node to be influenced by every node within the group; the propagation contribution of a node within the group; the propagation contribution of a node within the group; the number of nodes contained within the group; the influence of a node on other nodes within the group, the influence of a node on other nodes within the group; the influence of a node on other nodes within the group; the weight of the attention edge from a node to a node ; the weight of the attention edge from a node to a node ; the set of source nodes corresponding to all attention edges whose destination nodes are ; the set of destination nodes corresponding to all attention edges pointed by .

[0021] Further, the step S2 is specifically as follows:

[0022] collecting user data, attention relationships between users, historical tweets published by users and tweet information data of specific topics based on a data collection interface provided by Twitter.

[0023] Further, the step S3 is specifically as follows:

[0024] S31, construction of an attention network;

[0025] S311, adding high-influence users in the data collected in step S2 to the network node set, and then adding to the network node set all followers of the high-influence users who have forwarded tweets with a specified propagation scale l above; wherein the high-influence users are users whose tweet forwarding scales are % or above,

[0026] the numerical values of , are set artificially according to actual requirements; ,

[0027] ​S312, collect the follower list of all nodes in the existing network node set, then delete the part of the list that has been in the network node set and perform multiple non-replacement random sampling, and add the objects obtained after sampling to the network node set, and the number of samplings is equal to the number of nodes in the network node set before sampling;

[0028] S313, construct the connection according to the follow relationship between users to obtain a complete information propagation follow network;

[0029] S32, extraction of network attribute characteristics;

[0030] Constructing a user keyword vector, first, replacing the symbols in the historical tweets of all users in the studied follow network with spaces, the user historical tweets refer to the latest Tweets published by the user, these symbols include: '\!"#$&()*+,-. / :;<=>?@[\\]^_{|}·~‘’', then using the word_tokenize method in the nltk.tokenize module of Python to perform word segmentation and extraction on the content of the historical tweets published by the users in the follow network, and then removing illegal words in the segmented words through the synsets method in the nltk.wordnet module and using the stopwords set in the nltk.corpus to judge whether the segmented words are stop words to filter the current segmented result, finally obtaining the corresponding valid keyword set;

[0031] Among them, The value is set artificially according to actual needs.

[0032] Statistical effective keyword set, each keyword in the set is mentioned by how many users in the historical tweets of the follow network, and all keywords in the effective keyword set that are mentioned more than the total number of users in the network % are retained to obtain the corresponding bag of words;

[0033] Among them, The value is set artificially according to actual needs.

[0034] According to the number of corresponding keywords in the bag of words , construct a user keyword vector and a tweet keyword vector with dimensions respectively;

[0035] Among them, the value of each position in the user keyword vector respectively represents the frequency of the corresponding keyword appearing in the historical tweets published by the user; the value of each position in the tweet keyword vector respectively represents the frequency of the corresponding keyword appearing in the propagation tweet.

[0036] After obtaining the user keyword vector, the follow edge the two user nodes corresponding to and the Euler distance between the user keyword vectors corresponding to, to obtain the attention edge of the weight.

[0037] Further, the step S4 is specifically as follows:

[0038] S41, construction of the group clustering algorithm;

[0039] The group clustering algorithm is constructed, and the algorithm input includes: the entire attention network ; the multiple of the average propagation contribution ; the minimum size of the group ; the number of times of reassigning the group label of the node ; the maximum number of attempts of group division .

[0040] Wherein, G represents the attention network constructed in step S3, represents the multiple of the average propagation contribution within the group of the candidate node that can be added to the current group before the actual group clustering is performed, compared with the average propagation contribution outside the group.

[0041] Finally, the group label corresponding to each node in is output, which is specifically as follows:

[0042] (1) initialize the initial group label of all nodes ;

[0043] (2) initialize the unclassified node set , which contains all nodes in the network;

[0044] (3) initialize the classified node set , which is initially an empty set;

[0045] (4) initialize the current group number ;

[0046] (5) initialize the number of attempts of the current group division 0;

[0047] (6) determine whether the number of nodes that have not been classified in the current group is zero, if yes, call the group label reassignment algorithm, otherwise sort all nodes that have not been classified in the current group in descending order according to the number of unclassified nodes in the attention relationship thereof;

[0048] (7) traverse each node sorted to take the node as the first node in the current group, and further take the nodes having the attention relationship with the node as the initial candidate node set​ ;

[0049] (8) Traverse the nodes in , if the current node is a classified node, continue to process the next node, otherwise, calculate the intra-group average propagation contribution and the extra-group average propagation contribution , only when is greater than or equal to times of , it is added to the current group, if satisfies the condition of joining the current group, it is further removed from , added to , and is added to , and , finally, the unclassified nodes in the neighborhood of are added to ;

[0050] (9) When the nodes in are traversed, it is further judged whether the result of the current group division is reasonable, if the number of nodes in the current divided group is greater than or equal to , the result of the current group division is reasonable; let = 1, . Otherwise, the result of the current group division is unreasonable, let add one, and is restored to the state before the current group division; then, it is judged whether the number of attempts of group division is greater than the maximum number of attempts of group division , if not, return to step (7), otherwise, the nodes in are added to the group with the largest propagation contribution;

[0051] (10) Call the group label reassignment algorithm to correct the group label of each classified node;

[0052] (11) Output the group number of each node in the attention network.

[0053] S42, construction of the group label reassignment algorithm;

[0054] The group label reassignment algorithm is constructed, and the algorithm inputs include: the attention network ; , the group label of each node in the attention network , the number of times of group label reassignment of the node , the weight of each attention edge , and a group label set .

[0055] Final output Each node Corrected group labels The details are as follows:

[0056] 1) Judgment Is it greater than zero? If so, then let Otherwise output Each node Corrected group labels ;

[0057] 2) Traversing the graph Each node in Calculate its intra-group propagation contribution to each group in the existing group, add the node to the group with the largest intra-group propagation contribution, and return to step 1).

[0058] The beneficial effects of this invention are as follows: First, the method analyzes the definition of a group and the characteristics of group communication in communication studies textbooks, proposes and calculates the concept of average group communication contribution, then collects Twitter data, constructs an attention network and extracts network attribute features, and finally constructs a group clustering algorithm and a group label reassignment algorithm to complete the mining of specific Twitter users. This invention measures the average degree of interaction between an individual and other individuals in a group by proposing the average group communication contribution, and proposes a Twitter user communication group mining method based on group characteristics, solving the problem that the clustering results of existing group clustering algorithms cannot adequately reflect group characteristics. Attached Figure Description

[0059] Figure 1 This is a flowchart of a specific user mining method based on group propagation according to the present invention. Detailed Implementation

[0060] The method of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0061] like Figure 1 The flowchart of a specific user mining method based on group propagation according to the present invention is shown below, and the specific steps are as follows:

[0062] A method for identifying specific users based on group propagation, with the following specific steps:

[0063] S1, Calculate the average propagation contribution of the group;

[0064] S2, Twitter data collection;

[0065] S3, focus on the construction of network and extraction of network attribute features;

[0066] S4, construct group clustering algorithm and group label reassignment algorithm, complete specific Twitter user mining.

[0067] In the embodiment, the step S1 is specifically as follows:

[0068] The relevant definition of group is that a plurality of people with specific common goals and sense of belonging, and interactive relationship is a group. The essence of group communication is to connect and realize the common goal and cooperative will. Therefore, the user group under the condition of group communication should have interactive relationship, and at the same time have specific common goal and sense of belonging. For the interactive relationship, since the method research scene of the present application is the user communication group mining in Twitter, the information will be transmitted between users through the attention relationship, and the transmission process of information in the group is also the process of connecting and realizing the common goal and cooperative will of users. Therefore, the interactive relationship considered in the embodiment refers to the attention relationship between users, and the measurement of common goal and sense of belonging.

[0069] In the embodiment, the strength of individual to group common goal and sense of belonging is embodied by the closeness of individual and group interaction. According to the basic assumption, the concept of group average transmission contribution is proposed to measure the average closeness of individual and each node in the group.

[0070] The specific calculation method of group average transmission contribution is as follows:

[0071]

[0072]

[0073]

[0074]

[0075] Among them, represents the node The average transmission contribution in the group Reflects the influence of node On each node in the group and the ability to be affected by each node in the group; represents the transmission contribution of node In the group ; represents the number of nodes contained in the group ; represents the influence of node On the group influence of other nodes, representing nodes influenced by other nodes in the group influenced by other nodes in the group; representing the weight of the attention edge from node to node . representing the weight of the attention edge from node to node . representing the set of source nodes corresponding to all attention edges whose destination nodes are . representing the set of destination nodes corresponding to all attention edges pointed by .

[0076] In this embodiment, the step S2 is specifically as follows:

[0077] In order to model the determination of the user attention relationship in the studied social network and the average propagation contribution to the group of users, the user data, the attention relationship between users, the historical tweets published by the users and the tweet information data of a specific topic in a specific region are collected based on the data collection interface provided by Twitter.

[0078] In this embodiment, the step S3 is specifically as follows:

[0079] S311, 5 high-influence users in the data collected in step S2 are added to the network node set as source users, and then all the followers of the source users who have forwarded the tweets with a propagation scale of 5 specified propagation scales above 6000 are added to the network node set;

[0080] Among them, the high-influence user is a user whose tweet forwarding scale is above 2000 by more than 50%.

[0081] S312, the follower list of all nodes in the existing network node set is collected, then the part of which has been in the network node set is deleted and multiple non-replacement random sampling is performed, and the objects obtained after sampling are added to the network node set, and the number of sampling is equal to the number of nodes in the network node set before sampling;

[0082] S313, the connection edge is constructed according to the attention relationship between users, and the complete information propagation attention network is obtained;

[0083] S32, extraction of network attribute characteristics;

[0084] The user keyword vector is constructed as follows. Firstly, symbols in the historical tweets of all users in the studied attention network are replaced with spaces, the historical tweets of a user referring to the last 1000 tweets published by the user, the symbols including: '\!"#$&()*+,-. / :;<=>?@[\\]^_{|}·~‘’', and then the contents in the historical tweets published by the users in the attention network are tokenized and extracted by using the word_tokenize method in the nltk.tokenize module of Python, and then the illegal words in the tokenization are removed by using the synsets method in the nltk.wordnet module, and whether the tokenization is a stop word is judged by using the stopwords set in the nltk.corpus to filter the current tokenization result, and finally the corresponding valid keyword set is obtained;

[0085] The number of times each keyword in the valid keyword set is mentioned in the historical tweets of the users in the attention network is counted, and all keywords in the valid keyword set whose number of times of being mentioned is greater than 1% of the total number of users in the network are reserved to obtain the corresponding bag-of-words.

[0086] According to the number of keywords in the bag-of-words , a user keyword vector and a tweet keyword vector are respectively constructed, the dimensions of the user keyword vector and the tweet keyword vector being .

[0087] In the user keyword vector, the value of each position represents the frequency of occurrence of the corresponding keyword in the historical tweets published by the user, and in the tweet keyword vector, the value of each position represents the frequency of occurrence of the corresponding keyword in the propagation tweet.

[0088] After obtaining the user keyword vector, the Euler distance between the user keyword vectors corresponding to two user nodes and and is calculated to obtain the weight of the attention edge .

[0089] In the embodiment, the step S4 is specifically as follows:

[0090] S41, construction of a group clustering algorithm;

[0091] The group clustering algorithm is constructed, and the algorithm input includes: the entire attention network ; the multiple of the average propagation contribution ; the minimum size of the group ; the number of times of reassigning the group label of the node ; the maximum number of attempts of group division .

[0092] wherein G represents the constructed attention network of step S3, represents the ratio of the average intra-group propagation contribution of the candidate nodes that can be added to the current group before performing the actual group clustering to the average extra-group propagation contribution.

[0093] Finally output the group label corresponding to each node in the network. The specific steps are as follows:

[0094] (1) Initialize the initial group label of all nodes ;

[0095] (2) Initialize the unclassified node set , which contains all nodes in the network.

[0096] (3) Initialize the classified node set , which is initially an empty set.

[0097] (4) Initialize the current group number ;

[0098] (5) Initialize the number of attempts for the current group division 0.

[0099] (6) Determine whether the number of nodes that have not been classified is zero. If so, call the group label reassignment algorithm. Otherwise, sort all nodes that have not been classified in descending order according to the number of unclassified nodes in their attention relationships.

[0100] (7) Traverse each node in the sorted order, take the node as the first node in the current group, and further take the nodes related to the node in the attention relationship as the initial candidate node set .

[0101] (8) Traverse the nodes in , if the current node is a classified node, continue processing the next node, otherwise calculate the average intra-group propagation contribution and the average extra-group propagation contribution of the node. Only when is greater than or equal to times , the node can be added to the current group. If satisfies the condition of being added to the current group, further remove from , add to , let , and finally add the unclassified nodes in the neighborhood of to ;

[0102] (9) When the traversal is complete After identifying the nodes, further determine whether the result of this group partitioning is reasonable. If the number of nodes in the current partitioned group is greater than or equal to... If the group division result is reasonable, then the result is reasonable. +=1, Otherwise, the group division results would be unreasonable, leading to... Add one, and Restore the group to its state before the current group partitioning; then determine the number of partitioning attempts. Is it greater than the maximum number of attempts for group partitioning? If it is not greater than, return to step (7); otherwise, The nodes in the sequence are added to the group that contributes the most to the propagation.

[0103] (10) Call the group label redistribution algorithm to correct the group label of each classified node;

[0104] (11) Output the group ID of each node in the network of interest.

[0105] S42. Construction of the group label reassignment algorithm;

[0106] Since the initial classification of nodes into group labels addresses a local optimum, misclassification of node group labels may occur. To address this issue, this embodiment further proposes a group label reassignment algorithm.

[0107] Construct a group label reassignment algorithm. The algorithm input includes: attention network. ; Each node group tags Number of times the group label is reassigned to a node Each focus edge weight and group tag set .

[0108] Final output Each node Corrected group labels The details are as follows:

[0109] 1) Judgment Is it greater than zero? If so, then let Otherwise output Each node Corrected group labels ;

[0110] 2) Traverse each node in the graph , calculate its intra-group propagation contribution to each group in the existing groups, add the node to the group with the largest group propagation contribution, return to step 1).

[0111] In summary, the method of the present application analyzes the definition of the group in the propagation course and the characteristics of the group propagation, proposes the basic assumption that the strength of the individual's common goal and sense of belonging to the group can be reflected by the close degree of the individual and the group interaction, further proposes the concept of "group average propagation contribution" to measure the average close degree of the individual and each node in the group interaction based on the assumption, and finally proposes the Twitter user propagation group mining method based on the group characteristics based on the "group average propagation contribution", solves the problem that the clustering results of the existing group clustering algorithm cannot better reflect the characteristics of the group.

[0112] Those skilled in the art will appreciate that the embodiments described herein are presented for purposes of illustration and understanding of the principles of the present application and are not intended to be exhaustive or to limit the present application to the precise embodiments described. Many modifications and variations will be apparent to those skilled in the art. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present application should be included in the scope of the claims of the present application.​

Claims

1. A method for identifying specific Twitter users based on group propagation, the specific steps of which are as follows: S1, Calculate the average propagation contribution of the group; S2, Twitter data collection; S3. Focus on network construction and extraction of network attribute features; S4. Construct a group clustering algorithm and a group label redistribution algorithm to complete the mining of specific Twitter users; The specific steps of S1 are as follows: A group is a collection of multiple individuals who share a common goal and a sense of belonging and who have an interactive relationship. Under the conditions of group communication, user groups have an interactive relationship and at the same time have a specific common goal and a sense of belonging. in, The interaction relationship refers to the attention relationship between users, and the measurement of common goals and sense of belonging; the strength of an individual's sense of belonging to the group's common goals is set as reflected by the closeness of the individual's interaction with the group, and the concept of average group propagation contribution is proposed to measure the average closeness of the interaction between an individual and each node in the group. The specific calculation method for the average group propagation contribution is as follows: ； ； ； ； in, Represents a node In the group The average propagation contribution within the node reflects the node's average propagation contribution. The influence on each node within the group and the ability to be influenced by each node within the group; Represents a node In the group Contribution to the spread within; Representing a group The number of nodes contained within; Represents a node For groups The influence generated by other nodes within the system, Represents a node In the group The influence of other nodes within it; Indicates that by node Pointing to node The weight of the edge to focus on; Indicates that by node Pointing to node The weight of the edge to focus on; Indicates the destination node is The set of source nodes corresponding to all the edges of interest; Indicates by The set of destination nodes corresponding to all the edges of interest that are indicated.

2. The method for mining specific Twitter users based on group propagation according to claim 1, characterized in that, The specific details of step S2 are as follows: Based on the data collection interface provided by Twitter, we collect user data in specific regions, the following relationships between users, the historical tweets posted by users, and tweet information on specific topics.

3. The method for mining specific Twitter users based on group propagation according to claim 1, characterized in that, Step S3 is as follows: S31. Focus on network construction; S311, the data collected in step S2 A high-influence user is added as the source user to the network node set, and then all nodes related to the high-influence user are added. Followers who have retweeted tweets with a specified reach of 1 or more are added to the network node set; Among them, the high-influence users are those whose tweets have been retweeted on a large scale. %exist The above users, , The values ​​are set manually according to actual needs; S312. Collect the follower list of all nodes in the existing network node set, then delete the part that is already in the network node set and perform multiple random samplings without replacement, and add the sampled objects to the network node set, and the number of samplings is equal to the number of nodes in the network node set before sampling. S313. Construct edges based on the attention relationships between users to obtain a complete information dissemination and attention network; S32. Extraction of network attribute features; To construct user keyword vectors, firstly, replace all symbols in the historical tweets of all users in the studied follower network with spaces. The user's historical tweets refer to their most recent posts. The algorithm first takes a tweet, then uses the word_tokenize method in the nltk.tokenize module of Python to segment and extract the content of the historical tweets posted by users in the followed network. Then, it uses the synsets method in the nltk.wordnet module to remove illegal words in the segmentation and uses the stopwords set in nltk.corpus to determine whether the segmentation is a stop word to filter the current segmentation result, and finally obtains the corresponding set of effective keywords. in, The values ​​are set manually according to actual needs; The number of times each keyword in the effective keyword set has been mentioned in the historical tweets of users within the network is counted. Keywords mentioned more than the total number of users in the network are retained. % of all keywords, to obtain the corresponding bag of words; in, The values ​​are set manually according to actual needs; Based on the number of keywords in the bag of words Construct dimensions respectively User keyword vectors and tweet keyword vectors; In the user keyword vector, the value at each position represents the frequency of the corresponding keyword in the user's historical tweets; in the tweet keyword vector, the value at each position represents the frequency of the corresponding keyword in the disseminated tweet. After obtaining the user keyword vector, the attention edge is calculated. The corresponding two user nodes and The focus edges are obtained by measuring the Eulerian distance between the corresponding user keyword vectors. The weight.

4. The method for mining specific Twitter users based on group propagation according to claim 1, characterized in that, Step S4 is as follows: S41. Construction of a population clustering algorithm; Construct a population clustering algorithm. The algorithm input includes: the entire attention network. The multiple of average propagation contribution Minimum size of the group Number of times the group label is reassigned to a node Maximum number of attempts for group partitioning ; Where G represents the attention network constructed in step S3, This represents the multiple of the average intra-group propagation contribution of candidate nodes that may be added to the current group before actual group clustering, compared to the average out-of-group propagation contribution. Final output Each node Corresponding group tags The specific steps are as follows: (1) Initialize the initial group label of all nodes. ; (2) Initialize the set of unclassified nodes This includes all nodes in the network; (3) Initialize the set of classified nodes Initially, it is an empty set; (4) Initialize the current group number ; (5) Initialize the number of attempts for the current population partitioning. 0; (6) Determine whether the number of nodes that have not been classified into groups is zero. If so, call the group label redistribution algorithm. Otherwise, sort all nodes that have not been classified into groups in descending order based on the number of unclassified nodes in their attention relationships. (7) After traversing and sorting each node, take that node as the first node in the current group, and further take the nodes that have a relationship with that node as the initial set of candidate nodes. ; (8) Traversal The node in the middle, if the current node If a node has already been classified, proceed to the next node; otherwise, calculate the node's average propagation contribution within the group. Average outside-group transmission contribution Only when Greater than or equal to of Only add it to the current group when it is doubled, if If the conditions for joining the current group are met, then further... Remove from ,Will join in In the middle, order Finally Add unclassified nodes from the neighborhood ; (9) When the traversal is complete After identifying the nodes, further determine whether the result of this group partitioning is reasonable. If the number of nodes in the current partitioned group is greater than or equal to... If the group division result is reasonable, then the result is reasonable. +=1, Otherwise, the group division results would be unreasonable, leading to... Add one, and Restore the group to its state before the current group partitioning; then determine the number of partitioning attempts. Is it greater than the maximum number of attempts for group partitioning? If it is not greater than, return to step (7); otherwise, The nodes in the sequence are added to the group that contributes the most to the propagation. (10) Call the group label redistribution algorithm to correct the group label of each classified node; (11) Output the group ID of each node in the network of interest; S42. Construction of the group label reassignment algorithm; Construct a group label reassignment algorithm. The input of this group label reassignment algorithm includes: attention network. ; Each node group tags Number of times the group label is reassigned to a node Each focus edge weight and group tag set ; Final output Each node Corrected group labels The specific steps are as follows: 1) Judgment Is it greater than zero? If so, then let Otherwise output Each node Corrected group labels ; 2) Traversing the graph Each node in Calculate its intra-group propagation contribution to each group in the existing group, add the node to the group with the largest intra-group propagation contribution, and return to step 1).

Images