A method and system for user role analysis in online communities

Abstract

This invention discloses a method and system for analyzing user roles in online communities, relating to the field of social networks. The method includes: acquiring community network interaction data on preset social network platforms regarding preset category viewpoints; wherein the community network interaction data includes posts and the users who posted or forwarded the posts; determining the best topic on the preset category viewpoint; calculating the semantic polarization degree of each user on the best topic on the preset category viewpoint based on the community network interaction data; and classifying each user into online community user roles based on the semantic polarization degree of each user on the best topic, obtaining the classification result. This invention enables a more detailed understanding and role classification of communities in social networks that exhibit a binary structure on preset category viewpoints, and analyzes whether users are living in information cocoons of preset category viewpoints.

Description

Technical Field

[0001] This invention relates to the field of social network technology, and in particular to a method and system for analyzing user roles in online communities. Background Technology

[0002] On social media, users tend to form online communities based on shared beliefs. These homogeneous communities, composed of individuals with similar ideologies, limit individuals' exposure to diverse viewpoints, creating an "echo chamber" effect. Typically, methods such as community segmentation and semantic analysis can be used to identify individual and group communities from the perspectives of network topology and multimodal content understanding, assigning community tags to individuals to analyze which community a user belongs to and what beliefs they hold.

[0003] However, existing technologies typically categorize individuals simply into users of a particular community, without considering the more complex roles within social networks. For example, there are users who fully embrace different ideologies, and users who absorb multiple viewpoints but only support one particular ideology. Secondly, while they usually examine which ideology a user's opinion leans towards, they don't consider the strength of that leaning, potentially leading to misjudgments. That is, some users may simply express opinions close to a particular ideology, rather than being strong supporters of that ideology, especially in communities where the pre-defined categories of viewpoints present a binary structure. Summary of the Invention

[0004] To address the technical problems existing in the background art, this invention proposes a method and system for analyzing user roles in online communities.

[0005] Firstly, the present invention proposes a method for analyzing user roles in online communities, comprising:

[0006] Collect community network interaction data on preset social networking platforms regarding preset categories of viewpoints; the community network interaction data includes posts and the users who posted or forwarded the posts;

[0007] Identify the best topic within the predefined category of viewpoints;

[0008] Based on social network interaction data, calculate the semantic polarization of each user's best topic in the preset category of viewpoints;

[0009] Based on the degree of semantic polarization of each user on the best topic, the users are divided into network community user roles, resulting in the following division results: users of the first preset community, users of the second preset community, and cross-community users. Among them, users of the first preset community are further divided into users within the first preset community and users at the boundary of the first preset community, and users of the second preset community are further divided into users within the second preset community and users at the boundary of the second preset community. The first preset community and the second preset community present a binary structure in terms of preset category viewpoints.

[0010] Preferably, determining the best topic within a preset category of viewpoints specifically includes:

[0011] Obtain a set of candidate topics related to viewpoints within a preset category;

[0012] First, determine the official social media accounts or opinion leader accounts for the first and second preset communities respectively;

[0013] Each preset community's official social media account or opinion leader's account will be used to create a separate set of posts for each preset community on each topic within the candidate topic set.

[0014] The post set for each topic in each preset community is encoded using a preset vector encoding model to obtain the vector set for each topic in each preset community.

[0015] Select the vector pairs with the largest vector distance under each topic from the vector sets of the two preset communities; find the post pairs corresponding to the vector pairs with the largest vector distance under each topic in reverse search based on the vector pairs of the two preset communities.

[0016] The third pre-defined language model is used to evaluate the degree of deviation of posts from the two pre-defined communities under various topics, and the evaluation scores of the posts of the two pre-defined communities under various topics are obtained.

[0017] Based on the evaluation scores of posts in each topic across the two preset communities, the topic with the highest evaluation score is selected as the best topic for the two preset communities in terms of preset category viewpoints.

[0018] Preferably, based on social network interaction data, the semantic polarization degree of each user's optimal topic in a preset category of viewpoints is calculated, specifically including:

[0019] Based on social network interaction data, a user set is constructed, and the posts of each user in the social network interaction data are combined into a post set for each user.

[0020] For each user, the first preset language model is used to determine whether each post in their post collection is relevant to the best topic; if not, the post is removed from the post collection.

[0021] If so, then keep it;

[0022] After iterating through each post in the user's post collection, determine whether the number of posts in the current post collection on the best topic is greater than a preset threshold; if not, end the process.

[0023] If so, the semantic polarization of each post of the user on the best topic is calculated using the preset second language model to obtain the semantic polarization of each post of the user on the best topic; the average of the semantic polarization of each post of each user on the best topic is calculated and used as the semantic polarization of each user on the best topic.

[0024] Preferably, based on the degree of semantic polarization of each user on the optimal topic, the user roles in the online community are divided to obtain the division results, specifically including:

[0025] Based on the degree of semantic polarization of each user on the best topic, adaptive community segmentation is performed on each user to obtain the adaptive community segmentation results for each user; among them, the adaptive community segmentation results are divided into three types: first preset community users, second preset community users, and cross-community users.

[0026] After completing the adaptive community segmentation for all users, adaptive internal role segmentation is performed on users in the first preset community and users in the second preset community to obtain internal role segmentation results. The internal role segmentation results are divided into four types: users inside the first preset community, users at the boundary of the first preset community, users inside the second preset community, and users at the boundary of the second preset community.

[0027] Based on the adaptive community segmentation results and the internal role segmentation results, the segmentation results are obtained.

[0028] Preferably, based on the semantic polarization degree of each user on the optimal topic, adaptive community segmentation is performed on each user to obtain the adaptive community segmentation results for each user, specifically including:

[0029] Adaptively determine the preset community segmentation threshold range; wherein, the preset community segmentation threshold range is: ,and ;

[0030] Determine whether the semantic polarization of each user on the optimal topic falls within the preset community segmentation threshold range. Inside;

[0031] If yes, then the user is determined to be a cross-community user; if not, then it is determined whether the user's semantic polarization on the best topic is below the preset community segmentation threshold. If yes, then the user is determined to be a user of the first preset community; if no, then the user is determined to be a user of the second preset community.

[0032] Preferably, adaptively determining the preset community segmentation threshold range specifically includes:

[0033] Each user's posts on the best topic are pieced together to obtain a representation of each user's posts;

[0034] Each user's post representation is vectorized to obtain the representation vector for each user;

[0035] The KMeans clustering algorithm was used to cluster the representation vectors of all users, resulting in three clusters, which were denoted as the first preset community cluster, the second preset community cluster, and the cross-community cluster, respectively.

[0036] The statistical mapping represents the semantic polarization of all users in the first preset community cluster on the best topic, and the values ​​of the first preset percentiles from low to high are assigned as thresholds. The second preset percentile value, from high to low, is assigned to the threshold. .

[0037] Preferably, adaptive internal role classification is performed on the users of the first preset community to obtain the internal role classification results, specifically including:

[0038] For all users in the first preset community, calculate the Euclidean distance from the representation vector of each user in the first preset community to the geometric center, and take the value of the third preset percentile from the nearest to the farthest as the boundary threshold.

[0039] Based on the Euclidean distance from the representation vector of each user in the first preset community to the geometric center, users whose Euclidean distance to the geometric center exceeds the boundary threshold within the first preset community are extracted and used as the candidate set of boundary users of the first preset community.

[0040] Calculate the proportion of each user's K nearest neighbors in the candidate set of boundary users of the first preset community;

[0041] When the ratio is less than the preset ratio, the user is determined to be a user on the first preset community boundary.

[0042] Users in the first preset community, excluding those at the boundary of the first preset community, are marked as users within the first preset community.

[0043] Preferably, the users of the second preset community are subjected to adaptive internal role classification to obtain the internal role classification results, specifically including:

[0044] For all users in the second preset community, calculate the Euclidean distance from the representation vector of each user in the second preset community to the geometric center, and take the value of the third preset percentile from the nearest to the farthest as the boundary threshold.

[0045] Based on the Euclidean distance from the representation vector of each user in the second preset community to the geometric center, users whose Euclidean distance to the geometric center exceeds the boundary threshold within the second preset community are extracted and used as the boundary user candidate set of the second preset community.

[0046] Calculate the proportion of each user's K nearest neighbors in the candidate set of boundary users of the second preset community;

[0047] When this ratio is less than the preset ratio, the user is determined to be a user of the second preset community boundary.

[0048] Users in the second preset community, excluding those at the community's boundary, are marked as users within the second preset community.

[0049] Secondly, this invention also proposes a user role analysis system in online communities, comprising:

[0050] The acquisition module is used to acquire community network interaction data on preset social network platforms regarding preset categories of viewpoints; wherein, the community network interaction data includes posts and the users who posted or forwarded the posts;

[0051] The processing module is used to determine the best topic in the preset category viewpoints and calculate the semantic polarization degree of each user's best topic in the preset category viewpoints based on community network interaction data. According to the semantic polarization degree of each user in the best topic, the module classifies each user into network community user roles and obtains the classification results. The classification results are divided into three types: first preset community users, second preset community users, and cross-community users. The first preset community users are further divided into two types: users within the first preset community and users at the boundary of the first preset community. The second preset community users are further divided into two types: users within the second preset community and users at the boundary of the second preset community. The first preset community and the second preset community present a binary structure in the preset category viewpoints.

[0052] Preferably, determining the best topic within a preset category of viewpoints specifically includes:

[0053] Obtain a set T of candidate topics related to viewpoints within a preset category;

[0054] First, determine the official social media accounts or opinion leader accounts for the first and second preset communities respectively;

[0055] Each preset community's official social media account or opinion leader's account will be used to create a separate set of posts for each preset community on each topic within the candidate topic set.

[0056] The post set for each topic in each preset community is encoded using a preset vector encoding model to obtain the vector set for each topic in each preset community.

[0057] Select the vector pairs with the largest vector distance under each topic from the vector sets of the two preset communities; find the post pairs corresponding to the vector pairs with the largest vector distance under each topic in reverse search based on the vector pairs of the two preset communities.

[0058] The third pre-defined language model is used to evaluate the degree of deviation of posts from the two pre-defined communities under various topics, and the evaluation scores of the posts of the two pre-defined communities under various topics are obtained.

[0059] Based on the evaluation scores of posts in each topic across the two preset communities, the topic with the highest evaluation score is selected as the best topic for the two preset communities in terms of preset category viewpoints.

[0060] Preferably, based on social network interaction data, the semantic polarization degree of each user's optimal topic in a preset category of viewpoints is calculated, specifically including:

[0061] Based on social network interaction data, a user set is constructed, and the posts of each user in the social network interaction data are combined into a post set for each user.

[0062] For each user, the first preset language model is used to determine whether each post in their post collection is relevant to the best topic; if not, the post is removed from the post collection.

[0063] If so, then keep it;

[0064] After iterating through each post in the user's post collection, determine whether the number of posts in the current post collection on the best topic is greater than a preset threshold; if not, end the process.

[0065] If so, the semantic polarization of each post of the user on the best topic is calculated using the preset second language model to obtain the semantic polarization of each post of the user on the best topic.

[0066] Calculate the average semantic polarization of each user's posts on the best topic, and use it as the semantic polarization of each user on the best topic.

[0067] Preferably, based on the degree of semantic polarization of each user on the optimal topic, the user roles in the online community are divided to obtain the division results, specifically including:

[0068] Based on the degree of semantic polarization of each user on the best topic, adaptive community segmentation is performed on each user to obtain the adaptive community segmentation results for each user; among them, the adaptive community segmentation results are divided into three types: first preset community users, second preset community users, and cross-community users.

[0069] After completing the adaptive community segmentation for all users, adaptive internal role segmentation is performed on users in the first preset community and users in the second preset community to obtain internal role segmentation results. The internal role segmentation results are divided into four types: users inside the first preset community, users at the boundary of the first preset community, users inside the second preset community, and users at the boundary of the second preset community.

[0070] Based on the adaptive community segmentation results and the internal role segmentation results, the segmentation results are obtained.

[0071] Preferably, based on the semantic polarization degree of each user on the optimal topic, adaptive community segmentation is performed on each user to obtain the adaptive community segmentation results for each user, specifically including:

[0072] Adaptively determine the preset community segmentation threshold range; wherein, the preset community segmentation threshold range is: ,and ;

[0073] Determine whether the semantic polarization of each user on the optimal topic falls within the preset community segmentation threshold range. Inside;

[0074] If yes, then the user is determined to be a cross-community user; if not, then it is determined whether the user's semantic polarization on the best topic is below the preset community segmentation threshold range. If yes, then the user is determined to be a user of the first preset community; if no, then the user is determined to be a user of the second preset community.

[0075] Preferably, adaptively determining the preset community segmentation threshold range specifically includes:

[0076] Each user's posts on the best topic are pieced together to obtain a representation of each user's posts;

[0077] Each user's post representation is vectorized to obtain the representation vector for each user;

[0078] The KMeans clustering algorithm was used to cluster the representation vectors of all users, resulting in three clusters, which were denoted as the first preset community cluster, the second preset community cluster, and the cross-community cluster, respectively.

[0079] The statistical mapping represents the semantic polarization of all users in the first preset community cluster on the best topic, and the values ​​of the first preset percentiles from low to high are assigned as thresholds. The second preset percentile value, from high to low, is assigned to the threshold. .

[0080] Preferably, adaptive internal role classification is performed on the users of the first preset community to obtain the internal role classification results, specifically including:

[0081] For all users in the first preset community, calculate the Euclidean distance from the representation vector of each user in the first preset community to the geometric center, and take the value of the third preset percentile from the nearest to the farthest as the boundary threshold.

[0082] Based on the Euclidean distance from the representation vector of each user in the first preset community to the geometric center, users whose Euclidean distance to the geometric center exceeds the boundary threshold in the first preset community are extracted and used as the boundary user candidate set of the first preset community.

[0083] Calculate the proportion of each user's K nearest neighbors in the candidate set of boundary users of the first preset community;

[0084] When the ratio is less than the preset ratio, the user is determined to be a user on the first preset community boundary.

[0085] Users in the first preset community, excluding those at the boundary of the first preset community, are marked as users within the first preset community.

[0086] Preferably, the users of the second preset community are subjected to adaptive internal role classification to obtain the internal role classification results, specifically including:

[0087] For all users in the second preset community, calculate the Euclidean distance from the representation vector of each user in the second preset community to the geometric center, and take the value of the third preset percentile from the nearest to the farthest as the boundary threshold.

[0088] Based on the Euclidean distance from the representation vector of each user in the second preset community to the geometric center, users whose Euclidean distance to the geometric center exceeds the boundary threshold within the second preset community are extracted and used as the candidate set of boundary users of the second preset community.

[0089] Calculate the proportion of each user's K nearest neighbors in the candidate set of boundary users of the second preset community;

[0090] When this ratio is less than the preset ratio, the user is determined to be a user of the second preset community boundary.

[0091] Users in the second preset community, excluding those at the community's boundary, are marked as users within the second preset community.

[0092] The proposed user role analysis method and system in the online community can provide a more detailed understanding and role classification of communities that present a binary structure based on preset category viewpoints in social networks. It can analyze whether users live in information cocoons based on preset category viewpoints and can serve as basic tags for subsequent key user mining in social networks. Attached Figure Description

[0093] Figure 1 This is a flowchart illustrating a user role analysis method in an online community proposed in this invention. Detailed Implementation

[0094] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0095] Firstly, referring to Figure 1 The present invention proposes a method for analyzing user roles in online communities, comprising:

[0096] Collect community network interaction data on preset social networking platforms regarding preset categories of viewpoints; the community network interaction data includes posts and users who posted or forwarded the posts;

[0097] Identify the best topic within the predefined category of viewpoints;

[0098] Based on social network interaction data, calculate the semantic polarization of each user's best topic in the preset category of viewpoints;

[0099] Based on the degree of semantic polarization of each user on the best topic, the users are divided into network community user roles, resulting in the following division results: users of the first preset community, users of the second preset community, and cross-community users. Among them, users of the first preset community are further divided into users within the first preset community and users at the boundary of the first preset community, and users of the second preset community are further divided into users within the second preset community and users at the boundary of the second preset community. The first preset community and the second preset community present a binary structure in terms of preset category viewpoints.

[0100] This invention enables a more detailed understanding and role classification of communities in social networks that exhibit a binary structure based on preset category viewpoints, analyzes whether users live in information cocoons based on preset category viewpoints, and can serve as a basic tag for subsequent key user mining in social networks.

[0101] In the process of obtaining community network interaction data related to a preset community from a preset social network platform, the API is used to obtain the community network interaction data related to the preset community from the preset social network platform, and the corresponding content and authors are obtained by searching by keywords, etc., and the content of important users such as opinion leaders of the preset community is obtained separately.

[0102] Among these, determining the best topics within the preset category of viewpoints specifically includes:

[0103] Obtain a set T of candidate topics related to viewpoints within a preset category;

[0104] First, determine the official social media accounts or opinion leader accounts for the first and second preset communities respectively;

[0105] Each preset community's official social media account or opinion leader's account will be used to create a separate set of posts for each preset community on each topic within the candidate topic set.

[0106] The post set for each topic in each preset community is encoded using a preset vector encoding model to obtain the vector set for each topic in each preset community.

[0107] Select the vector pairs with the largest vector distance under each topic from the vector sets of the two preset communities; find the post pairs corresponding to the vector pairs with the largest vector distance under each topic in reverse search based on the vector pairs of the two preset communities.

[0108] The third pre-defined language model is used to evaluate the degree of deviation of posts from the two pre-defined communities under various topics, and the evaluation scores of the posts of the two pre-defined communities under various topics are obtained.

[0109] Based on the evaluation scores of posts in each topic across the two preset communities, the topic with the highest evaluation score is selected as the best topic for the two preset communities in terms of preset category viewpoints.

[0110] In this embodiment, the degree of semantic polarization of a user on the optimal topic is: ;

[0111] in, In the formula, This indicates the degree of semantic polarization of the i-th post content of the current user on the best topic. This represents the content of the i-th post by the current user on the best topic. This represents the best topic. This indicates the posts in the post pair belonging to the first pre-defined community. This indicates the post in the post pair that represents the second preset community. This represents the second largest language model.

[0112] Specifically, through The second language model is guided by methods such as assigning a score of 0 to posts from the first pre-defined community and a score of 1 to posts from the second pre-defined community. The i-th content of the current user is then assigned a score based on semantic similarity. Rate it.

[0113] The first, second, and third language models in this embodiment are all existing technologies, as long as they can achieve the corresponding functions in this embodiment.

[0114] It should be understood that the set of candidate topics for preset category viewpoints in this embodiment can be obtained from official accounts in the first preset community and the second preset community.

[0115] Considering that the first and second preset communities in the binary structure sometimes converge and sometimes oppose on topics related to preset category viewpoints, in order to accurately calculate the semantic polarization degree of the content published by each user on preset category viewpoints, this embodiment calculates the semantic polarization degree of the optimal topic for each user on preset category viewpoints based on community network interaction data, specifically including:

[0116] Based on social network interaction data, a user set is constructed, and the posts of each user in the social network interaction data are combined into a post set for each user.

[0117] For each user, the first preset language model is used to determine whether each post in their post collection is relevant to the best topic; if not, the post is removed from the post collection.

[0118] If so, then keep it;

[0119] After iterating through each post in the user's post collection, determine whether the number of posts in the current post collection on the best topic is greater than a preset threshold; if not, end the process.

[0120] If so, the semantic polarization of each post of the user on the best topic is calculated using the preset second language model to obtain the semantic polarization of each post of the user on the best topic.

[0121] Calculate the average semantic polarization of each user's posts on the best topic, and use it as the semantic polarization of each user on the best topic.

[0122] In this embodiment, the first large language model is a lightweight large language model, and the second large language model is a large language model with a larger number of parameters than the first large language model, in order to adapt to different computing power requirements.

[0123] The preset vector encoding model in this embodiment can be selected from existing technologies.

[0124] In this embodiment, users within the first or second preset community typically exhibit strong support for the position of their respective first or second preset community, and their information sources are limited to their own community.

[0125] Users at the boundaries of the first or second preset community typically exhibit strong support for the position of their respective first or second preset community. However, information sources span across communities, which can usually break the echo chamber effect and should be given special consideration when identifying key users.

[0126] While users from different communities may lean towards the boundaries of the first or second pre-defined community, their stance is not strong enough to warrant a clear categorization into a particular pre-defined community.

[0127] In this embodiment, based on the degree of semantic polarization of each user on the optimal topic, the user roles in the online community are divided, and the division results are obtained, specifically including:

[0128] Based on the degree of semantic polarization of each user on the best topic, adaptive community segmentation is performed on each user to obtain the adaptive community segmentation results for each user; among them, the adaptive community segmentation results are divided into three types: first preset community users, second preset community users, and cross-community users.

[0129] After completing the adaptive community segmentation for all users, adaptive internal role segmentation is performed on users in the first preset community and users in the second preset community to obtain internal role segmentation results. The internal role segmentation results are divided into four types: users inside the first preset community, users at the boundary of the first preset community, users inside the second preset community, and users at the boundary of the second preset community.

[0130] Based on the adaptive community segmentation results and the internal role segmentation results, the segmentation results are obtained.

[0131] This embodiment can adaptively classify users in the online community user role classification, thereby improving the accuracy of the classification.

[0132] In a further embodiment, based on the semantic polarization degree of each user on the optimal topic, adaptive community segmentation is performed on each user to obtain the adaptive community segmentation results for each user, specifically including:

[0133] Adaptively determine the preset community segmentation threshold range; wherein, the preset community segmentation threshold range is: ,and ;

[0134] Determine whether the semantic polarization of each user on the optimal topic falls within the preset community segmentation threshold range. Inside;

[0135] If yes, then the user is determined to be a cross-community user; if not, then it is determined whether the user's semantic polarization on the best topic is below the preset community segmentation threshold. If yes, then the user is determined to be a user of the first preset community; if no, then the user is determined to be a user of the second preset community.

[0136] When a user's semantic polarization level on the best topic is located at Within, the user was determined to be... That is, users across communities;

[0137] When a user's semantic polarization level on the best topic is located at Outside the range and below The user was determined to be That is, the first preset community users;

[0138] When a user's semantic polarization level on the best topic is located at Outside the range and higher than The user was determined to be That is, the second preset community users.

[0139] In this embodiment, the adaptive community segmentation results for each user are represented as follows: ;

[0140] In the formula, Indicates user The results of the role classification For the first preset community users, For the second preset community users, For cross-community users, Indicates user On the best topic The degree of semantic polarization.

[0141] In a further embodiment, adaptively determining the preset community segmentation threshold range specifically includes:

[0142] Each user's posts on the best topic are pieced together to obtain a representation of each user's posts;

[0143] Each user's post representation is vectorized to obtain the representation vector for each user;

[0144] The KMeans clustering algorithm was used to cluster the representation vectors of all users, resulting in three clusters, which were denoted as the first preset community cluster, the second preset community cluster, and the cross-community cluster, respectively.

[0145] The statistical mapping represents the semantic polarization of all users in the first preset community cluster on the best topic, and the values ​​of the first preset percentiles from low to high are assigned as thresholds. The second preset percentile value, from high to low, is assigned to the threshold. .

[0146] After semantic vectorization, the numerous users of the first preset community (Community A), cross-community, and the second preset community (Community B) are typically arranged sequentially and occupy a relatively contiguous space. This embodiment maps the clustering results to the polarization degree of key issues, rather than directly defining communities by clusters. The reasons are twofold: first, interpretability—"attitude towards the best issue" has a more direct meaning than unsupervised clustering; and second, scalability—new users can be directly determined based on polarization scores without the need for re-clustering.

[0147] After completing the pre-defined community segmentation for all users, it is necessary to further segment the users within each community. During segmentation, the primary consideration is the source of their information. If a user simultaneously meets the following criteria: 1. A certain percentage of their information comes from users outside their own community; 2. Their semantics fall between those of users within their own community and those from other communities, then they are considered a community boundary user and will receive more attention in further key user mining and other applications. The rest are considered internal users of their own community.

[0148] In a further embodiment, adaptive internal role classification is performed on the first preset community users to obtain the internal role classification result, specifically including:

[0149] For all users in the first preset community, calculate the Euclidean distance from the representation vector of each user in the first preset community to the geometric center, and take the value of the third preset percentile from the nearest to the farthest as the boundary threshold.

[0150] Based on the Euclidean distance from the representation vector of each user in the first preset community to the geometric center, users whose Euclidean distance to the geometric center exceeds the boundary threshold within the first preset community are extracted and used as the candidate set of boundary users of the first preset community.

[0151] Calculate the proportion of each user's K nearest neighbors in the candidate set of boundary users of the first preset community;

[0152] When the ratio is less than the preset ratio, the user is determined to be a user on the first preset community boundary.

[0153] Users in the first preset community, excluding those at the boundary of the first preset community, are marked as users within the first preset community.

[0154] In a further embodiment, adaptive internal role classification is performed on the users of the second preset community to obtain the internal role classification result, specifically including:

[0155] For all users in the second preset community, calculate the Euclidean distance from the representation vector of each user in the second preset community to the geometric center, and take the value of the third preset percentile from the nearest to the farthest as the boundary threshold.

[0156] Based on the Euclidean distance from the representation vector of each user in the second preset community to the geometric center, users whose Euclidean distance to the geometric center exceeds the boundary threshold within the second preset community are extracted and used as the boundary user candidate set of the second preset community.

[0157] Calculate the proportion of each user's K nearest neighbors in the candidate set of boundary users of the second preset community;

[0158] When this ratio is less than the preset ratio, the user is determined to be a user of the second preset community boundary.

[0159] Users in the second preset community, excluding those at the community's boundary, are marked as users within the second preset community.

[0160] The candidate sets of boundary users for the first and second preset communities are represented as follows: ;

[0161] In the formula, Users in the candidate set of boundary users of the first preset community. Users in the boundary user candidate set of the second preset community The percentile corresponding to the boundary threshold. The percentile of the Euclidean distance from the representation vector of each user to the geometric center.

[0162] The partitioning result in this embodiment is expressed as follows: ;

[0163] In the formula, This indicates the user segmentation results. This refers to users within the first preset community. This indicates the users at the first preset community boundary. This refers to users within the second preset community. This indicates the second preset community boundary users. The percentile of the Euclidean distance from the user's representation vector to the geometric center. This represents the percentile corresponding to the boundary threshold. The proportion of each user's K nearest neighbors in the candidate set of boundary users of the first or second preset community.

[0164] Among them, the proportion of each user's K nearest neighbors in the candidate set of boundary users of the first preset community is that of the user in this community. .

[0165] Thus, this embodiment completes the classification of all user roles. Specifically, based on determining the optimal topic and calculating the polarization score, firstly, through vectorization and clustering, the semantic contours of the first preset community, cross-community, and second preset community are identified, and the range of the three types of people in terms of the optimal topic polarization score is further obtained; secondly, within the community, users who simultaneously meet the criteria of having many information sources outside their own community and having a lot of semantic overlap with those outside their own community are designated as community boundary users.

[0166] In one specific embodiment, the first, second and third preset percentiles are all 95%, and the preset ratio is 50%.

[0167] Secondly, this invention also proposes a user role analysis system in online communities, comprising:

[0168] The acquisition module is used to acquire community network interaction data on preset social network platforms regarding preset categories of viewpoints; wherein, the community network interaction data includes posts and the users who posted or forwarded the posts;

[0169] The processing module is used to determine the best topic in the preset category of viewpoints; based on the social network interaction data, it calculates the semantic polarization degree of each user's best topic in the preset category of viewpoints; according to the semantic polarization degree of each user's best topic, it divides each user into social network user roles, and obtains the division results; among them, the division results are divided into three types: first preset community users, second preset community users, and cross-community users; the first preset community users are further divided into two types: users within the first preset community and users at the boundary of the first preset community, and the second preset community users are further divided into two types: users within the second preset community and users at the boundary of the second preset community. The first preset community and the second preset community present a binary structure in the preset category of viewpoints.

[0170] Preferably, determining the best topic within a preset category of viewpoints specifically includes:

[0171] Obtain a set of candidate topics related to viewpoints within a preset category;

[0172] First, determine the official social media accounts or opinion leader accounts for the first and second preset communities respectively;

[0173] Each preset community's official social media account or opinion leader's account will be used to create a separate set of posts for each preset community on each topic within the candidate topic set.

[0174] The post set for each topic in each preset community is encoded using a preset vector encoding model to obtain the vector set for each topic in each preset community.

[0175] Select the vector pairs with the largest vector distance under each topic from the vector sets of the two preset communities; find the post pairs corresponding to the vector pairs with the largest vector distance under each topic in reverse search based on the vector pairs of the two preset communities.

[0176] The third pre-defined language model is used to evaluate the degree of deviation of posts from the two pre-defined communities under various topics, and the evaluation scores of the posts of the two pre-defined communities under various topics are obtained.

[0177] Based on the evaluation scores of posts in each topic across the two preset communities, the topic with the highest evaluation score is selected as the best topic for the two preset communities in terms of preset category viewpoints.

[0178] Preferably, based on social network interaction data, the semantic polarization degree of each user's optimal topic in a preset category of viewpoints is calculated, specifically including:

[0179] Based on social network interaction data, a user set is constructed, and the posts of each user in the social network interaction data are combined into a post set for each user.

[0180] For each user, the first preset language model is used to determine whether each post in their post collection is relevant to the best topic; if not, the post is removed from the post collection.

[0181] If so, then keep it;

[0182] After iterating through each post in the user's post collection, determine whether the number of posts in the current post collection on the best topic is greater than a preset threshold; if not, end the process.

[0183] If so, the semantic polarization of each post of the user on the best topic is calculated using the preset second language model.

[0184] Preferably, based on the degree of semantic polarization of each user on the optimal topic, the user roles in the online community are divided to obtain the division results, specifically including:

[0185] Based on the degree of semantic polarization of each user on the best topic, adaptive community segmentation is performed on each user to obtain the adaptive community segmentation results for each user; among them, the adaptive community segmentation results are divided into three types: first preset community users, second preset community users, and cross-community users.

[0186] After completing the adaptive community segmentation for all users, adaptive internal role segmentation is performed on users in the first preset community and users in the second preset community to obtain internal role segmentation results. The internal role segmentation results are divided into four types: users inside the first preset community, users at the boundary of the first preset community, users inside the second preset community, and users at the boundary of the second preset community.

[0187] Based on the adaptive community segmentation results and the internal role segmentation results, the segmentation results are obtained.

[0188] In a further embodiment, based on the semantic polarization degree of each user on the optimal topic, adaptive community segmentation is performed on each user to obtain the adaptive community segmentation results for each user, specifically including:

[0189] Adaptively determine the preset community segmentation threshold range; wherein, the preset community segmentation threshold range is: ,and ;

[0190] Determine whether the semantic polarization of each user on the optimal topic falls within the preset community segmentation threshold range. Inside;

[0191] If yes, then the user is determined to be a cross-community user; if not, then it is determined whether the user's semantic polarization on the best topic is below the preset community segmentation threshold. If yes, then the user is determined to be a user of the first preset community; if no, then the user is determined to be a user of the second preset community.

[0192] When a user's semantic polarization level on the best topic is located at Within, the user was determined to be... That is, users across communities;

[0193] When a user's semantic polarization level on the best topic is located at Outside the range and below The user was determined to be That is, the first preset community users;

[0194] When a user's semantic polarization level on the best topic is located at Outside the range and higher than The user was determined to be That is, the second preset community users.

[0195] In a further embodiment, adaptively determining the preset community segmentation threshold range specifically includes:

[0196] Each user's posts on the best topic are pieced together to obtain a representation of each user's posts;

[0197] Each user's post representation is vectorized to obtain the representation vector for each user;

[0198] The KMeans clustering algorithm was used to cluster the representation vectors of all users, resulting in three clusters, which were denoted as the first preset community cluster, the second preset community cluster, and the cross-community cluster, respectively.

[0199] The statistical mapping represents the semantic polarization of all users in the first preset community cluster on the best topic, and the values ​​of the first preset percentiles from low to high are assigned as thresholds. The second preset percentile value, from high to low, is assigned to the threshold. .

[0200] Specifically, both the first and second preset percentiles are 95%.

[0201] In a further embodiment, adaptive internal role classification is performed on the first preset community users to obtain the internal role classification result, specifically including:

[0202] For all users in the first preset community, calculate the Euclidean distance from the representation vector of each user in the first preset community to the geometric center, and take the value of the third preset percentile from the nearest to the farthest as the boundary threshold.

[0203] Based on the Euclidean distance from the representation vector of each user in the first preset community to the geometric center, users whose Euclidean distance to the geometric center exceeds the boundary threshold within the first preset community are extracted and used as the candidate set of boundary users of the first preset community.

[0204] Calculate the proportion of each user's K nearest neighbors in the candidate set of boundary users of the first preset community;

[0205] When the ratio is less than the preset ratio, the user is determined to be a user on the first preset community boundary.

[0206] Users in the first preset community, excluding those at the boundary of the first preset community, are marked as users within the first preset community.

[0207] Specifically, the third preset percentile is 95%, and the preset proportion is 50%.

[0208] In a further embodiment, adaptive internal role classification is performed on the users of the second preset community to obtain the internal role classification result, specifically including:

[0209] For all users in the second preset community, calculate the Euclidean distance from the representation vector of each user in the second preset community to the geometric center, and take the value of the third preset percentile from the nearest to the farthest as the boundary threshold.

[0210] Based on the Euclidean distance from the representation vector of each user in the second preset community to the geometric center, users whose Euclidean distance to the geometric center exceeds the boundary threshold within the second preset community are extracted and used as the boundary user candidate set of the second preset community.

[0211] Calculate the proportion of each user's K nearest neighbors in the candidate set of boundary users of the second preset community;

[0212] When this ratio is less than the preset ratio, the user is determined to be a user of the second preset community boundary.

[0213] Users in the second preset community, excluding those at the community's boundary, are marked as users within the second preset community.

[0214] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for analyzing user roles in a network community, characterized by, The method comprises the following steps: obtaining community network interaction data on a preset category of views published in a preset social network platform, wherein the community network interaction data comprises posts and post users or forwarding users; determining the best topic on the preset category of views; constructing a user set according to the community network interaction data, and forming a post set of each user in the community network interaction data; for each user, using a preset first large language model to determine whether each post in the post set is related to the best topic; if not, deleting the post from the post set; if yes, retaining the post; after traversing each post in the post set of the user, determining whether the number of posts on the best topic in the post set is greater than a preset number threshold; if not, ending; if yes, using a preset second large language model to calculate the semantic polarization degree of each post of the user on the best topic, and obtaining the semantic polarization degree of each post of the user on the best topic; calculating the average value of the semantic polarization degree of each post of each user on the best topic, and taking the average value as the semantic polarization degree of each user on the best topic; dividing the users into network community user roles according to the semantic polarization degree of each user on the best topic, and obtaining a division result; wherein the division result is divided into three types: a first preset community user, a second preset community user, and a cross-community user; wherein the first preset community user is divided into a first preset community internal user and a first preset community boundary user, the second preset community user is divided into a second preset community internal user and a second preset community boundary user, and the first preset community and the second preset community present a binary structure on the preset category of views.

2. The method of claim 1, wherein, The method for determining the best topic on the preset category of views comprises the following steps: obtaining a candidate topic set on the preset category of views; determining the official social media accounts or opinion leader accounts of the first preset community and the second preset community respectively; forming a post set of each topic of each preset community by publishing posts of the official social media accounts or opinion leader accounts of each preset community under each topic in the candidate topic set; using a preset vector encoding model to encode the post set of each topic of each preset community, and obtaining a vector set of each topic of each preset community; selecting a vector pair with the largest vector distance under each topic from the vector sets of each topic of the two preset communities; and finding a post pair corresponding to the vector pair according to the vector pair with the largest vector distance under each topic of the two preset communities; using a preset third large language model to evaluate the stand-off degree of the post pair of the two preset communities under each topic, and obtaining an evaluation score of the post pair of the two preset communities under each topic; selecting the topic with the first evaluation score as the best topic of the two preset communities on the preset category of views according to the evaluation scores of the post pair of the two preset communities under each topic.

3. The method of claim 1, wherein the step of analyzing the user role in the network community is characterized by, According to the semantic polarization degree of each user on the best topic, the network community user roles of each user are divided to obtain a division result, specifically including: According to the semantic polarization degree of each user on the best topic, the adaptive community division of each user is performed to obtain an adaptive community division result of each user; wherein the adaptive community division result is divided into three kinds of first preset community users, second preset community users and cross-community users; After completing the adaptive community division of all users, the adaptive community internal role division of the first preset community users and the second preset community users is performed respectively to obtain an internal role division result; the internal role division result is divided into four kinds of first preset community internal users, first preset community boundary users, second preset community internal users and second preset community boundary users; Based on the adaptive community division result and the internal role division result, a division result is obtained.

4. The method of claim 3, wherein the step of analyzing the user role in the network community is characterized by, According to the semantic polarization degree of each user on the best topic, the adaptive community division of each user is performed to obtain an adaptive community division result of each user, specifically including: Adaptive determination of a preset community division threshold range; wherein the preset community division threshold range is , and ; respectively determine whether semantic polarization degrees of the users on the best topic are in a preset community division threshold range within; If yes, the user is determined as a cross-community user; if no, whether the semantic polarization degree of the user on the optimal topic is lower than a preset community division threshold range ; if yes, the user is determined as a first preset community user; if no, the user is determined as a second preset community user.

5. The method of claim 4, wherein the step of analyzing the user role in the network community is characterized by, The preset community division threshold range is adaptively determined, specifically including: The post expressions of each user on the best topic are spliced respectively to obtain the post expression of each user; The post expression of each user is vectorized respectively to obtain a representation vector of each user; The representation vectors of all users are clustered by using a KMeans clustering algorithm to obtain three clusters, which are respectively denoted as a first preset community cluster, a second preset community cluster and a cross-community cluster; The statistical mapping is the semantic polarization degree of all users in the first preset community cluster on the best topic, and the threshold value is given by taking the value of the first preset percentile from low to high , the threshold value is given by taking the value of the second preset percentile from high to low .

6. The method of claim 5, wherein the step of analyzing the user role in the network community is characterized by, The adaptive community internal role division of the first preset community users is performed to obtain an internal role division result, specifically including: For all users of the first preset community, the Euclidean distance of the representation vector of each user of the first preset community to the geometric center is calculated, and the value of the third preset percentile from near to far is taken as a boundary threshold value; According to the Euclidean distance of the representation vector of each user of the first preset community to the geometric center, users in the first preset community whose Euclidean distance to the geometric center exceeds the boundary threshold value are extracted and taken as a boundary user candidate set of the first preset community; The proportion of the home community in the K-nearest neighbors of each user in the boundary user candidate set of the first preset community is calculated; When the proportion is less than a preset proportion, the user is determined as a first preset community boundary user; And the remaining users of the first preset community except the first preset community boundary users are marked as first preset community internal users.

7. The method of claim 5, wherein the step of analyzing the user role in the network community is characterized by, The adaptive community internal role division of the second preset community users is performed to obtain an internal role division result, specifically including: For all users of the second preset community, the Euclidean distance of the representation vector of each user of the second preset community to the geometric center is calculated, and the value of the third preset percentile from near to far is taken as a boundary threshold value; According to the Euclidean distance of the representation vector of each user of the second preset community to the geometric center, users in the second preset community whose Euclidean distance to the geometric center exceeds the boundary threshold value are extracted and taken as a boundary user candidate set of the second preset community; Calculate the proportion of the current community in the K-nearest neighbors of each user in the second preset community boundary user candidate set; When the proportion is less than a preset proportion, the user is determined to be a second preset community boundary user; And the remaining users in the second preset community except the second preset community boundary user are marked as second preset community internal users.

8. A user role analysis system in an online community, characterized in that, Comprise: An acquisition module is configured to acquire community network interaction data on a preset category of views published in a preset social network platform, wherein the community network interaction data includes posts and post users or forwarding users of the posts; A processing module is configured to determine an optimal topic on the preset category of views, construct a user set according to the community network interaction data, and form a post set of each user by using the posts of each user in the community network interaction data; for each user, a preset first large language model is used to determine whether each post in the post set of the user is related to the optimal topic; if not, the post is deleted from the post set; if yes, the post is retained; after each post in the post set of the user is traversed, it is determined whether the number of posts on the optimal topic in the current post set is greater than a preset number threshold; if not, the process is ended; if yes, a preset second large language model is used to calculate the semantic polarization degree of each post of the user on the optimal topic, and the semantic polarization degree of each post of the user on the optimal topic is obtained; the average value of the semantic polarization degree of each post of each user on the optimal topic is calculated and used as the semantic polarization degree of each user on the optimal topic; according to the semantic polarization degree of each user on the optimal topic, the network community user role of each user is divided, and a division result is obtained; the division result is divided into three types: a first preset community user, a second preset community user and a cross-community user; the first preset community user is divided into a first preset community internal user and a first preset community boundary user, the second preset community user is divided into a second preset community internal user and a second preset community boundary user, and the first preset community and the second preset community present a binary structure on the preset category of views.

Images