A method for discovering interest communities in social networks

Abstract

The present invention provides a method for discovering interest communities in social networks, which relates to the technical field of community discovery. The present invention establishes a new deep learning model based on the traditional Text CNN model, which can support the input of multiple social network texts, and combines the similarity between social network texts to propose a multi-text interest modeling method based on Text CNN combined with similarity, and proposes a user interest feature modeling method combining network structure and interactive behavior. The LM network algorithm is used to construct a user influence model, and then the results based on SMB-TextCNN are adjusted according to the model results, attention relationships and @ behavior information. Finally, based on the results of SIBUIM, a method for discovering interest communities based on k-means overlap is proposed. This method takes into account the structurality of social networks and the content of nodes, and can divide Sina Weibo users into overlapping interest communities.

Description

Technical Field

[0001] The present invention relates to the technical field of community discovery, and in particular to a method for discovering interest communities in social networks. Background Art

[0002] With the development of smart mobile devices, people are spending an increasing proportion of their time online. They also spend a significant amount of time browsing and acquiring information on social networks. Social networks offer users the ability to post hypertext messages, comment on others' blogs, and chat with others. Users can experience the joy of social interaction on social networks, making them of great research value.

[0003] Online social networks have become an indispensable bridge connecting the virtual world and the physical world. Analyzing user groups with similar interests around a particular topic through Weibo users' text and online interactions has significant implications for interest discovery, public opinion monitoring, and psychological research on social networks. For example, in online public opinion monitoring, mining interest communities within hot Weibo events can help understand online public opinion and the composition of different interest communities. In the field of mental health analysis, categorizing Weibo users into interest communities facilitates targeted psychological counseling and professional intervention for high-stress individuals. In recommendation systems, this can aid in advertising placement.

[0004] Community discovery involves dividing network nodes into clusters, using clustering to create clusters with high cohesion and low coupling. Interest-based community discovery involves finding communities by grouping users on social networks based on their interests.

[0005] Current research on interest community discovery faces three challenges. First, interest community discovery research often relies on structural analysis, content analysis, or a combination of both within social networks. For content-based interest community discovery, widely used methods include topic extraction methods like LDA. However, the accuracy of these methods is affected by prior knowledge and is unstable. While the widely recognized TextCNN algorithm, which is based on text content classification, can solve single-text classification problems, it is less effective at classifying multiple blog posts. Second, while interest community discovery methods that combine structure and content are more specific than single-source approaches, they neglect the social nature of social networks and fail to consider user interactions such as comments, reposts, likes, and @s. Third, current methods for discovering overlapping interest communities rely on label propagation algorithms, operating solely on network structure. Furthermore, label acquisition relies primarily on user label information, which can be inaccurate and biased. Discovering non-overlapping interest communities does not account for the diverse interests of users within social networks and therefore lacks practical application value. Therefore, this research has significant implications and applications in areas such as online public opinion, public mental health, and personalized recommendations.

[0006] In summary, traditional interest community discovery ignores important factors such as multi-text content generated by users, interactive behaviors, and user label accuracy, and the existing research results based on interest community discovery are not ideal. Summary of the Invention

[0007] In response to the problems existing in the prior art, the present invention provides a method for discovering interest communities for social networks. It establishes a new deep learning model based on the traditional Text CNN model, which can support multiple social network text inputs and combines the similarity between social network texts. It proposes a multi-text interest modeling method based on Text CNN combined with similarity (Similarity considerations and Multi-Blog input interest model based TextCNN, SMB-TextCNN), and proposes a user interest feature modeling method based on network structure and interaction behavior (with structure and interaction behavior user interest model, SIBUIM). The user influence model is constructed using the LM network algorithm, and the results based on SMB-TextCNN are adjusted according to the model results, attention relationships and @ behavior information. Finally, based on the results of SIBUIM, an overlap interest community discovery method based on k-means (overlap interest community discovery method based on k-means, K-OICD) is proposed. This method takes into account the structurality of social networks and the content of nodes, and can divide Sina Weibo users into overlapping interest communities.

[0008] A method for discovering interest communities on social networks, with the following specific steps:

[0009] Step 1: Change the single text input of the Text CNN convolutional neural network model to multiple text inputs, and use the TextCNN convolutional neural network to generate the user's interest feature matrix. The obtained interest feature matrix is adjusted using the text similarity between multiple social network texts of each user;

[0010] Step 1.1: Read multiple social network texts from the same user ID and segment the texts using the precise patterns in the Chinese word segmentation module Jieba implemented in Python.

[0011] Step 1.2: Use the CBOW (Continuous Bag-of-Words Model) training model in the word2vec model to vectorize the word segmentation results;

[0012] Step 1.3: Input the word vector obtained in step 1.2 into the Text CNN convolutional neural network model. After the convolution layer, pooling layer, and fully connected layer, the user's network social text interest feature matrix T = {I1, I2, ...I N} T ;

[0013] Step 1.4: Use cosine similarity to calculate the text similarity between each online social text of the user, and use the obtained online social text similarity as a weight to adjust the online social text interest feature matrix;

[0014] Step 2: Use the LM neural network algorithm to build a user influence model, classifying user influence into four levels: high, relatively high, medium, and low. Then, adjust the results based on the SMB-TextCNN based on the user influence model results, user follow-up relationships, and @ behavior information;

[0015] Step 2.1: Extract the number of followers, posts, likes, comments, and reposts of users on the social networking platform as features of the influence model, and merge the user data with the same ID;

[0016] Step 2.2: Use the user data obtained in step 2.1 to train and verify the LM neural network classifier, and classify user influence into four levels: high, relatively high, medium, and low;

[0017] Step 2.3: Use the LM neural network classifier trained in step 2.2 to classify user u i The follower list of the follower set UF i Classify the influence and select users with high influence to form the set HF i ;

[0018] Step 2.4: Calculate user u separately i With the collection HF i The similarity between the social network texts of all users in the , and the interest feature matrix obtained in step 1.4 is adjusted using the obtained text similarity;

[0019] Step 2.5: Calculate user u i The interest matrix of the elements in the @user set P is constructed by the @list of @, and its interest matrix and u are calculated. i The similarity of the interest matrix is calculated, and the obtained result is used to adjust the interest feature matrix obtained in step 2.4 to obtain the final interest feature matrix F i ;

[0020] Step 3: Based on the idea of k-means clustering, use the interest feature matrix F iDiscover and segment overlapping communities among users of social network platforms;

[0021] Step 3.1: Calculate the sum of each element in each interest category in each user's interest feature matrix. The sum of each element in the i-th row of the matrix for interest category k is denoted as Sum ik , Sum of all users in interest category k k The value of k ={Sum 1k ,Sum 2k ,…Sum zk}.

[0022] Step 3.2: The set Y of interest categories is L = {Y1, Y2, ... Y m}, perform k-means algorithm on each element in set L, that is, the set of communities is recorded as set C, and Y k The results of internal community division are sorted from high to low according to the position of the centroid of each community, and are recorded as {C k1 , C k2 , C ka}.

[0023] Step 3.3: Assume that there are y small communities under each interest community. These y small communities represent the distribution of interest communities under each interest category. Therefore, for each element in L, the first y communities are taken as the community division result of this method in each category.

[0024] The formula for calculating the social network text similarity adjustment user interest feature matrix in step 1.4 is as follows:

[0025] V={ω1I1,ω2I2,...ω N I N} T

[0026] V is the adjusted interest matrix, ωi represents the weight of Ii, and the value of ωi is described as follows. Cosine similarity is used as the similarity metric, and the calculation method is as follows:

[0027]

[0028] Among them I i , I j Represents user u i The interest distribution vectors of any two social network texts are calculated. After calculating the similarity between any two social network texts, the resulting matrix is recorded as B. The adjustment formula for matrix B is as follows:

[0029] A=f(B-δE)

[0030]

[0031] E represents the identity matrix, δ is the threshold, if B ij -δ>0,A ij A value of 1 indicates that the two social network texts have a great similarity, otherwise it is 0. i ) indicates A i The number of 1s in , (A) represents the total number of elements in the matrix A, and the value of ωi is as follows:

[0032]

[0033] In step 2.4, user u is used i High-influence user collection HF i The formula of the interest feature matrix obtained by adjusting the interest feature matrix in step 1.4 is as follows:

[0034]

[0035] Among them |HF i | represents the set HF i The number of elements in , j represents the subscript of the jth user in HF, λ i is the adjustment parameter, g(V i ,V j ) is a measure of V j V i The contribution standard is calculated as follows:

[0036]

[0037] Matrix H is user u i and the high-influence user collection HF i The similarity between the interest feature matrices of , Sum(H) is the sum of the probabilities of all users in a certain interest category, and is calculated as follows:

[0038]

[0039]

[0040] The adjustment formula of the user interest matrix of the @ behavior in step 2.5 is as follows:

[0041]

[0042] Where user u i The feature matrix of interest is V' i ={I i1 ',I i2 ',...I iN '} T , user ui The user set composed of @ lists is P. Any element u in P j The interest matrix is V j '={I j1 ',I j2 ',...I jN '} T , P ik , P jk For user u i The probability that a social network text of its @ user belongs to a certain interest category.

[0043] Beneficial effects of the present invention: The present invention proposes a method for discovering interest communities on social networks. This method can not only discover interest communities with high modularity, but also has an acceptable running time. This method can not only monitor the direction of public opinion in social networks, but also help improve various recommendation algorithms, and can also provide psychological research institutions with emotional change analysis results of people on sensitive topics. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 This is a diagram showing the word segmentation results in step 1.1 in an embodiment of the present invention;

[0045] Figure 2 This is a graph showing the word vector results in step 1.2 of an embodiment of the present invention;

[0046] Figure 3 This is a framework diagram of the Text CNN convolutional neural network model in an embodiment of the present invention;

[0047] Figure 4 This is a comparison experiment chart of the user classification accuracy of the SMB-Text CNN model under different input methods in an embodiment of the present invention;

[0048] Figure 5 The experimental results of comparing the accuracy of different vector dimensions in the embodiment of the present invention are shown;

[0049] Figure 6 This is a comparison chart of the classification accuracy of the four models used in the embodiments of the present invention: SMB-Text CNN, Text CNN, LDA, and TF-IDF;

[0050] Figure 7 This is a flowchart of the influence model algorithm established using the LM God General network algorithm in step 2 of an embodiment of the present invention;

[0051] Figure 8 This is a comparison chart of the influence results and accuracy of the LM neural network applied in the embodiment of the present invention to the decision tree classifier, the naive Bayes classifier, and the random walk PageRank algorithm;

[0052] Figure 9 This is a comparison chart of the accuracy of the SIBUIM algorithm proposed in step 2 and the SMB-Text CNN algorithm proposed in step 1, as well as the Text CNN algorithm in the embodiment of the present invention;

[0053] Figure 10 This is a comparison chart of the intra-class tightness CP value and inter-class tightness SP value using the SMB-Text CNN algorithm and SIBUIM algorithm applied by K-OICD in an embodiment of the present invention, as well as directly using the Text CNN algorithm and a separate COPRA algorithm. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solutions and advantages of the present invention clearer, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments. The specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0055] A method for discovering interest communities on social networks, the specific method is as follows:

[0056] In this example, a total of 12 million raw Weibo posts were acquired, and 1 million Weibo user data was stored in a database, referred to as Dataset A. Additionally, a dataset of 35,000 tagged posts was collected online. These tagged posts were categorized into 21 categories: IT, finance, media, animation, real estate, advertising, health, education, travel, food, women, automobiles, art, lifestyle, fashion, sports, publishing, campus, games, entertainment, and parenting. This dataset is referred to as Dataset B. In this example, Dataset C was constructed by selecting users from Dataset A with fewer Weibo posts, fewer followers, and fewer reposts and comments. Dataset C was also constructed by collecting online Sina Big V user data and Weibo Hall of Fame data.

[0057] Step 1: Change the single text input of the Text CNN convolutional neural network model to multiple text inputs, and use the TextCNN convolutional neural network to generate the user's interest feature matrix. Use the text similarity between multiple social network texts of each user to adjust the obtained interest feature matrix. The specific steps are as follows:

[0058] Step 1.1: Read multiple social network texts of the same user ID and segment the multiple social network texts of the user using the precise pattern of the Chinese word segmentation module Jieba implemented in Python. The segmentation results are as follows: Figure 1 As shown;

[0059] Step 1.2: Use the CBOW (Continuous Bag-of-Words Model) training model in the word2vec model to vectorize the word segmentation results. The vectorization results are as follows: Figure 2 As shown;

[0060] Step 1.3: Input the word vector obtained in step 1.2 into the Text CNN convolutional neural network model, as shown in Figure 3 As shown in the figure, after the convolution layer, pooling layer, and fully connected layer operations, the user's network social text interest feature matrix T = {I1, I2, ...I N} T ;

[0061] Step 1.4: Use cosine similarity to calculate the text similarity between each online social text of the user, and use the obtained online social text similarity as a weight to adjust the online social text interest feature matrix;

[0062] The formula for adjusting the user interest feature matrix based on blog post similarity is as follows:

[0063] V={ω1I1,ω2I2,...ω N I N} T (1)

[0064] V is the adjusted interest matrix, ω i Representative I i The weight, ω i The values are explained as follows. This method uses cosine similarity as the similarity metric, which is calculated as follows:

[0065]

[0066] Among them I i , I j Represents user u i The interest distribution vectors of any two blog posts are calculated. After calculating the similarity between any two blog posts, the resulting matrix is recorded as B. The adjustment formula for matrix B is as follows:

[0067] A=f(B-δE) (3)

[0068]

[0069] E represents the identity matrix, δ is the threshold, if B ij -δ>0,A ij A value of 1 indicates that the two blog posts have a great similarity, otherwise it is 0. i ) represents the number of 1s in Ai, (A) represents the total number of elements in the matrix A, ω iThe values are as follows:

[0070]

[0071] In this example, the length of each blog post will vary, and the input layer length of the SMB-TextCNN algorithm model is a fixed value, so this example needs to unify the length of the input blog posts. This example tests the user classification accuracy of the SMB-TextCNN model under three methods: maximum, minimum, and average input length. Figure 4 shown.

[0072] This example compares the effect of word2vec output word vector dimension size on the accuracy of SMB-TextCNN model. This experiment selects 100, 150, 200, 250, 300, 350, and 400 dimensions as candidate dimensions to observe the trend of accuracy as the dimension changes. Figure 5 We can see that if the word vector dimension is large, overfitting will occur, while if the dimension is too small, the accuracy will be affected. When the vector dimension is below 300, the accuracy increases as the dimension increases. When the vector dimension exceeds 300, the accuracy begins to decrease. Here, 300 dimensions is selected as the optimal dimension.

[0073] To demonstrate the validity of the SMB-TextCNN model proposed in this example, a comparative experiment was designed, comparing the SMB-TextCNN algorithm model with common text classification algorithms such as TextCNN, LDA, and TF-IDF. However, this example proposes that the final user classification can be multi-classified, so the traditional single-class classification standard is no longer applicable. Therefore, the correctness of the classification results for this experiment are shown in Table 1.

[0074] Since TextCNN can only process a single blog post, the example combines all blog posts under the same user ID as input and conducts experimental comparisons. Finally, the experimental results are as follows Figure 6 shown.

[0075] Step 2: Use the LM neural network algorithm to build a user influence model, classifying user influence into four levels: high, relatively high, medium, and low. Then, adjust the results based on the SMB-TextCNN based on the user influence model results, user follow-up relationships, and @ behavior information;

[0076] Step 2.1: Extract the number of followers, number of posts, number of likes, number of comments, and number of reposts of users on the social networking platform as features of the influence model. Merge the user data with the same ID. The merge results are shown in Table 2.

[0077] Table 1 Description of classification correctness

[0078]

[0079] Table 2 Final data form

[0080] User ID Number of fans Number of Weibo posts Likes Number of comments Number of forwards Classification results 1802464801 1380 383 2790 732 223 high 2153467364 240 948 679 283 92 Low …… …… …… …… …… …… ……

[0081] Step 2.2: Use the user data obtained in step 2.1 to train and verify the LM neural network classifier, and divide the user influence into four levels: high, relatively high, medium, and low. The influence model algorithm established by the LM neural network algorithm is as follows: Figure 7 As shown;

[0082] This example uses Dataset C to verify the accuracy of the proposed LM neural network and decision tree classifier, naive Bayes classifier, and random walk PageRank algorithm. The data set is divided as shown in Table 3. Among the four indicators, this example classifies users in Dataset C as high if two values exceed the above average. Similarly, if a user has two values that exceed the average of the big V user data, it is considered high. The accuracy comparison chart is shown in the figure below. Figure 8 As shown;

[0083] Table 3 Dataset division

[0084] content Total quantity / piece Mean / piece Number of comments 118431871 165870.97 (Max=8342455) Number of forwarding 152082044 213000.06 (Max=12528208) Number of fans 6341220030 888126.09 (Max=14707929) Number of Weibo posts 1701001 2382.37 (Max=31546)

[0085] Step 2.3: Use the LM neural network classifier trained in step 2.2 to classify user u i The follower list of the follower set UF i Classify the influence and select users with high influence to form the set HF i ;

[0086] Step 2.4: Calculate user u separately i With the collection HF i The similarity between the social network texts of all users in is used to adjust the interest feature matrix obtained in step 1.4 using the obtained text similarity.

[0087] The utilization of user u i High-influence user collection HF i The formula of the interest feature matrix obtained by adjusting the interest feature matrix in step 1.4 is as follows:

[0088]

[0089] Among them |HF i | represents the set HF i The number of elements in , j represents the subscript of the jth user in HF, λ i is the adjustment parameter, g(V i ,Vj ) is a measure of V j V i The contribution standard is calculated as follows:

[0090]

[0091] Matrix H is user u i and the high-influence user collection HF i The similarity between the interest feature matrices of , Sum(H) is the sum of the probabilities of all users in a certain interest category, and is calculated as follows:

[0092]

[0093]

[0094] Step 2.5: Calculate user u i The interest matrix of the elements in the @user set P is constructed by the @list of @, and its interest matrix and u are calculated. i The similarity of the interest matrix is calculated, and the obtained result is used to adjust the interest feature matrix obtained in step 2.4 to obtain the final interest feature matrix F i ;

[0095] The adjustment formula of the @ behavior on the user interest matrix is as follows:

[0096]

[0097] Where user u i The feature matrix of interest is V i '={I i1 ',I i2 ',...I iN '} T , user u i The user set composed of @ lists is P. Any element u in P j The interest matrix is V j '={I j1 ',I j2 ',...I jN '} T , P ik , P jk For user u i The probability that a blog post of a user @ed by the user belongs to a certain interest category;

[0098] In order to verify the rationality of the SIBUIM algorithm proposed in step 2, the SIBUIM accuracy comparison experiment compares SIBUIM with the SMB-TextCNN algorithm and TextCNN algorithm proposed in this paper. Because all three algorithms involve deep learning models, non-deep learning models have been compared in the SMB-TextCNN model performance comparison experiment, so non-deep learning models are not included here for comparison. This experiment will compare the above three models in terms of accuracy. The results of the SIBUIM accuracy comparison experiment are shown in the figure below. Figure 9 As shown;

[0099] Step 3: Based on the idea of k-means clustering, use the interest feature matrix F i The specific steps for discovering and dividing overlapping communities among users on social network platforms are as follows:

[0100] Step 3.1: Calculate the sum of each element in each interest category in each user's interest feature matrix. The sum of each element in the i-th row of the matrix for interest category k is denoted as Sum ik , Sum of all users in interest category k k The value of k ={Sum 1k ,Sum 2k ,…Sum zk};

[0101] Step 3.2: The set Y of interest categories is L = {Y1, Y2, ... Y m}, perform k-means algorithm on each element in set L, that is, the set of communities is recorded as set C, and Y k The results of internal community division are sorted from high to low according to the position of the centroid of each community, and are recorded as {Ck1, Ck2, Cka};

[0102] The number of communities after division is set to a, and the k-means algorithm is used to perform cluster analysis on the elements in L, and the a centroids are the output community centroids;

[0103] Step 3.3: Each interest community has y small communities. These y small communities represent the distribution of interest communities under each interest category. Therefore, for each element in L, the first y communities are taken as the community division result of this method in each category.

[0104] In order to prove the rationality of the algorithm proposed in step 3, this paper compares the CP value and SP value by applying K-OICD to the SMB-TextCNN algorithm and SIBUIM algorithm, as well as directly using the TextCNN algorithm and the separate COPRA algorithm. The experimental results are as follows Figure 10As shown in the figure, it can be seen that the K-OICD algorithm has better performance after the SIBUIM algorithm. It is worth mentioning that the COPRA algorithm did not perform well in this experiment. The reason is that the information in this dataset is not consistent with and inaccurate for COPRA, so the effect of the COPRA algorithm is very poor.

[0105] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments can still be modified, or some or all of the technical features therein can be replaced by equivalents. Therefore, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present invention.

Claims

1. A method for discovering interest communities in social networks, characterized by: The following steps are involved: Step 1: Change the single text input of the Text CNN convolutional neural network model to multiple text inputs, and use the Text CNN convolutional neural network to generate a user interest feature matrix. The obtained interest feature matrix is adjusted using the text similarity between multiple social network texts of each user; Step 1.1: Read multiple social network texts from the same user ID and segment the texts using the precise patterns in the Chinese word segmentation module Jieba implemented in Python. Step 1.2: Use the CBOW (Continuous Bag-of-Words Model) training model in the word2vec model to vectorize the word segmentation results; Step 1.3: Input the word vector obtained in step 1.2 into the Text CNN convolutional neural network model. After the convolution layer, pooling layer, and fully connected layer operations, the user's network social text interest feature matrix T = {I1, I2, ...I N } T ; Step 1.4: Use cosine similarity to calculate the text similarity between each online social text of the user, and use the obtained online social text similarity as a weight to adjust the online social text interest feature matrix; Step 2: Use the LM neural network algorithm to build a user influence model, categorizing user influence into four levels: high, relatively high, medium, and low. Then, adjust the results based on the Text CNN based on the user influence model results, user follow-up relationships, and @ behavior information. Step 2.1: Extract the number of followers, posts, likes, comments, and reposts of users on the social networking platform as features of the influence model, and merge the user data with the same ID; Step 2.2: Use the user data obtained in step 2.1 to train and verify the LM neural network classifier, and classify user influence into four levels: high, relatively high, medium, and low; Step 2.3: Use the LM neural network classifier trained in step 2.2 to classify user u i The follower list of the follower set UF i Classify the influence and select users with high influence to form the set HF i ; Step 2.4: Calculate user u separately i With the collection HF i The similarity between the social network texts of all users in the , and the interest feature matrix obtained in step 1.4 is adjusted using the obtained text similarity; Step 2.5: Calculate user u i The interest matrix of the elements in the @user set P is constructed by the @list of @, and its interest matrix and u are calculated. i The similarity of the interest matrix is calculated, and the obtained result is used to adjust the interest feature matrix obtained in step 2.4 to obtain the final interest feature matrix F i ; Step 3: Based on the idea of k-means clustering, use the interest feature matrix F i Discover and segment overlapping communities among users of social network platforms; Step 3.1: Calculate the sum of each element in each interest category in each user's interest feature matrix. The sum of each element in the kth column of the interest category is recorded as Sum ik , Sum of all users in interest category k k The value of k ={Sum 1k ,Sum 2k ,…Sum zk }; Step 3.2: The set Y of interest categories is L = {Y1, Y2, ... Y m }, perform k-means algorithm on each element in set L, that is, the set of communities is recorded as set C, and Y k The results of internal community division are sorted from high to low according to the position of the centroid of each community, and are recorded as {C k1 , C k2 , C ka }; Step 3.3: Assume that there are y small communities under each interest community. These y small communities represent the distribution of interest communities under each interest category. Therefore, for each element in L, the first y communities are taken as the community division result of this method in each category.

2. The method for discovering interest communities in social networks according to claim 1, wherein: The formula for calculating the social network text similarity adjustment user interest feature matrix in step 1.4 is as follows: V={ω1I1,ω2I2,...ω N I N } T V is the adjusted interest matrix, ωi represents the weight of Ii, and the value of ωi is described as follows. Cosine similarity is used as the similarity metric, and the calculation method is as follows: Among them I i , I j Represents user u i The interest distribution vectors of any two social network texts are calculated. After calculating the similarity between any two social network texts, the resulting matrix is recorded as B. The adjustment formula for matrix B is as follows: A=f(B-δE) E represents the identity matrix, δ is the threshold, if B ij -δ>0,A ij A value of 1 indicates that the two social network texts have a great similarity, otherwise it is 0; this paper uses (A i ) indicates A i The number of 1s in , (A) represents the total number of elements in the matrix A, and the value of ωi is as follows:

3. The method for discovering interest communities in social networks according to claim 1, wherein: In step 2.4, user u is used i High-influence user collection HF i The formula of the interest feature matrix obtained by adjusting the interest feature matrix in step 1.4 is as follows: Among them |HF i | represents the set HF i The number of elements in , j represents the subscript of the jth user in HF, λ i is the adjustment parameter, g(V i ,V j ) is a measure of V j V i The contribution standard is calculated as follows: Matrix H is user u i and the high-influence user collection HF i The similarity between the interest feature matrices of user u, Sum(H) is the similarity between user u i and user u in the high-influence user set HFi that he follows j The sum of the interest similarities is calculated as follows:

4. The method for discovering interest communities in social networks according to claim 1, wherein: The adjustment formula of the user interest matrix of the @ behavior in step 2.5 is as follows: Where user u i The feature matrix of interest is V' i ={I i1 ',I i2 ',...I iN '} T , user u i The user set composed of @ lists is P; any element u in P j The interest matrix is V j '={I j1 ',I j2 ',...I jN '} T , P ik , P jk For user u i The probability that a social network text of its @ user belongs to a certain interest category.

Images