A method and system for analyzing user opinions in a social network
By constructing a candidate topic dimension set for social networks and using an improved DBSCAN clustering algorithm, the problems of flexibility and granularity in user opinion analysis in social networks are solved, achieving more accurate user opinion identification and clustering, and supporting strategy or product improvement.
Patent Information
- Application Number
- CN202511851390.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-10
- Publication Date
- 2026-03-03
- Estimated Expiration
- 2045-12-10
AI Technical Summary
Existing technologies lack flexibility and granularity in analyzing user opinions on social networks, failing to accurately identify differences in user levels, resulting in the drowning out of important user voices and poor clustering effects.
By acquiring social network data, a candidate topic dimension set is constructed. A large language model is used for screening and scoring. Combined with an improved DBSCAN clustering algorithm, clustering is performed based on user representation vectors to distinguish key users from ordinary users and form multiple user clusters.
It improves the accuracy and efficiency of user opinion analysis, enabling a better understanding of the views and attitudes of individuals and groups on specific issues, identifying the voices of different communities and attracting users, and supporting the improvement of related strategies or products.
Smart Images
Figure CN121280164B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of online community opinion analysis technology, and in particular to a method and system for analyzing user opinions in social networks. Background Technology
[0002] To understand the views and attitudes of individuals and groups on specific macro issues, commonly used technical methods include:
[0003] (1) The polarization level is assessed on the individual pre-defined dimensions using the continuous view model, but the selection of the pre-defined dimensions is based on the theory of specific macro issues, which cannot guarantee matching specific datasets and lacks flexibility;
[0004] (2) The analysis is based on the attributes filled in by social users, but this method ignores the lack and inaccuracy of user tags in real social networks;
[0005] (3) A large number of social users who express their opinions are divided into groups using methods such as community discovery or clustering to grasp the degree of polarization of users' opinions from the perspective of the group; however, these methods do not distinguish the different levels of users. In reality, there are huge differences in the voice of different individuals. Treating everyone equally in a general way may cause the voices of important users to be buried under the massive amount of content. When clustering a large number of users, the clustering effect is not precise enough, and the effect of user clustering is difficult to guarantee, which in turn affects the effect of opinion analysis. Summary of the Invention
[0006] To address the technical problems existing in the background art, this invention proposes a method and system for analyzing user opinions in social networks.
[0007] In a first aspect, the present invention proposes a method for analyzing user opinions in social networks, comprising:
[0008] Obtain social data from social networks on a given macro topic; where social data includes post content, the users who posted the posts, and user interaction data;
[0009] Based on a given macro-level issue, a set of candidate topic dimensions is constructed. For each candidate topic dimension in the set, a pre-defined large language model is used to filter the data in the social dataset, and the polarization degree of each filtered data is scored to obtain the polarization score of each data on that candidate topic dimension.
[0010] Based on the polarization score of each data point on the candidate topic dimension, calculate the dispersion of each candidate topic dimension; sort the candidate topic dimensions according to the dispersion of each candidate topic dimension, and select the top N candidate topic dimensions to construct a set of identifiable topic dimensions.
[0011] For each topic dimension in the set of identifiable topic dimensions, the post content of each user is summarized, and the semantics obtained from the summarization are vectorized to obtain the vector representation of each user in each topic dimension.
[0012] The vector representations of each user across all topic dimensions are concatenated to obtain the representation vector for each user.
[0013] The improved DBSCAN clustering algorithm is used to cluster the representation vectors of all users to obtain the clustering results.
[0014] Preferably, the content of each user's posts is summarized, specifically including:
[0015] For each topic dimension in the set of identifiable topic dimensions, a large language model is used to summarize and categorize each user's posts to obtain the semantics of each user in that topic dimension.
[0016] Preferably, the improved DBSCAN clustering algorithm is used to cluster the representation vectors of all users to obtain the clustering results, specifically including:
[0017] Set the minimum number of neighbors, MinPts;
[0018] Calculate the Euclidean distance between each user's representation vector and the representation vectors of all other users;
[0019] Based on the Euclidean distance between each user's representation vector and the representation vectors of all other users, find the MinPts nearest neighbors for each user;
[0020] Determine the spherical neighborhood radius of each user based on each user's MinPts nearest neighbors;
[0021] Sort all users' sphere neighborhood radii in ascending order;
[0022] Use the sphere radius of users located at a preset percentile value as the neighborhood radius threshold;
[0023] The corrected sphere neighborhood radius for each user is determined based on the sphere neighborhood radius and neighborhood radius threshold for each user.
[0024] Redefine direct density reachability; where direct density reachability is defined as: user u is within the sphere neighborhood of user v, and user v is within the sphere neighborhood of user u;
[0025] The DBSCAN clustering algorithm is used to cluster the representation vectors of all users based on the corrected sphere neighborhood radius and the redefined direct density reachability, and the clustering results are obtained.
[0026] Preferably, ;
[0027] In the formula, Let u be the radius of the sphere's neighborhood. The neighborhood radius threshold, Let be the corrected sphere neighborhood radius for user u.
[0028] Preferably, the DBSCAN clustering algorithm is used to cluster the representation vectors of all users based on the corrected spherical neighborhood radius and the redefined direct density reachability, specifically including:
[0029] Step a: Set MinPts;
[0030] Step b: Randomly select a user from all users and determine whether the number of users within the sphere's neighborhood radius of that user is greater than or equal to MinPts; if yes, mark the user as a core user and reconstruct a new cluster for that core user; if no, mark the user as a noisy user.
[0031] Step c: Recursively determine each unmarked user within the user's sphere neighborhood radius until all users within the user's sphere neighborhood radius are marked; wherein, in the recursive determination, if the number of users within a user's sphere neighborhood radius is greater than or equal to MinPts, then mark the user as a core point and add all users within its sphere neighborhood radius to the cluster; if the number of users within a user's sphere neighborhood radius is less than MinPts, then mark the user as a boundary user;
[0032] Step d: Randomly select the next user from the remaining users of all users, and repeat steps b and c to cluster into new clusters until all users have been selected.
[0033] Preferably, before randomly selecting a user from all users, the method further includes:
[0034] All users are divided into key users and ordinary users;
[0035] Specifically, one user is randomly selected from all users, including:
[0036] Randomly select one key user from the key users;
[0037] This includes randomly selecting the next user from the remaining users of all users, specifically including:
[0038] The next key user is randomly selected from the remaining key users of the key user group.
[0039] Preferably, social data also includes interaction data of the users who posted the posts;
[0040] After clustering the representation vectors of all users using the DBSCAN clustering algorithm based on the corrected spherical neighborhood radius and the redefined direct density reachability, the following steps are also included:
[0041] For users who are not categorized into any cluster, identify the key users who have the most positive interactions with that user in the interaction data.
[0042] Determine whether the number of positive interactions between the user and the key user is within the preset ranking of the number of positive interactions with the key user;
[0043] If so, the user is considered to have the same view as the key user, and the user is added to the cluster where the key user is located;
[0044] If not, then remove the user completely.
[0045] Secondly, this invention also proposes a user opinion analysis system in a social network, comprising:
[0046] The acquisition module is used to acquire social data from social networks on a given macro topic; the social data includes post content and the users who posted it.
[0047] The processing module is used to construct a set of candidate topic dimensions based on a given macro-level issue. For each candidate topic dimension in the set, a pre-defined large language model is used to filter the data in the social dataset, and the polarization degree of each filtered data point is scored to obtain a polarization score for each data point on that candidate topic dimension. Based on the polarization score of each data point on that candidate topic dimension, the dispersion of each candidate topic dimension is calculated. The candidate topic dimensions are sorted according to their dispersion, and the top N candidate topic dimensions are selected to construct a set of identifiable topic dimensions. For each topic dimension in the set of identifiable topic dimensions, the post content of each key user is summarized, and the summarized semantics are vectorized to obtain the vector representation of each key user on each topic dimension. The vector representations of each key user across all topic dimensions are concatenated to obtain the representation vector of each key user. An improved DBSCAN clustering algorithm is used to cluster the representation vectors of all users to obtain the clustering results.
[0048] Preferably, the improved DBSCAN clustering algorithm is used to cluster the representation vectors of all users to obtain the clustering results, specifically including:
[0049] Set the minimum number of neighbors, MinPts;
[0050] Calculate the Euclidean distance between each user's representation vector and the representation vectors of all other users;
[0051] Based on the Euclidean distance between each user's representation vector and the representation vectors of all other users, find the MinPts nearest neighbors for each user;
[0052] Determine the spherical neighborhood radius of each user based on each user's MinPts nearest neighbors;
[0053] Sort all users' sphere neighborhood radii in ascending order;
[0054] Use the sphere radius of users located at a preset percentile value as the neighborhood radius threshold;
[0055] The corrected sphere neighborhood radius for each user is determined based on the sphere neighborhood radius and neighborhood radius threshold for each user.
[0056] Redefine direct density reachability; where direct density reachability is defined as: user u is within the sphere neighborhood of user v, and user v is within the sphere neighborhood of user u;
[0057] The DBSCAN clustering algorithm is used to cluster the representation vectors of all users based on the corrected sphere neighborhood radius and the redefined direct density reachability, and the clustering results are obtained.
[0058] The proposed method and system for analyzing user opinions in social networks in this invention constructs multiple user clusters for a given macro-topic by summarizing and clustering. This helps to understand the opinions and attitudes of individuals and groups on each topic dimension of a given macro-topic on social networks, understand the voices of different communities on the network and the users they attract, and thus improve the relevant strategies or products for the given macro-topic. Attached Figure Description
[0059] Figure 1 This is a flowchart illustrating a user opinion analysis method in a social network according to an embodiment of the present invention. Detailed Implementation
[0060] 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.
[0061] Firstly, referring to Figure 1 The present invention proposes a method for analyzing user opinions in social networks, comprising:
[0062] Obtain social data from social networks regarding a given macro-level issue; whereby social data includes post content and the users who posted the posts;
[0063] Based on social data, establish a set of identifiable topic dimensions;
[0064] In each topic dimension of the identifiable topic dimension set, the post content of each key user is summarized, and the summarized semantics are vectorized to obtain the vector representation of each key user in each topic dimension.
[0065] The vector representations of each key user across all topic dimensions are concatenated to obtain the representation vector of each key user.
[0066] The improved DBSCAN clustering algorithm is used to cluster the representation vectors of all users to obtain the clustering results.
[0067] This invention establishes a set of identifiable topic dimensions based on social data. For each topic dimension, the post content of key users is summarized to obtain vector representations of each key user across all topic dimensions. This dynamically and adaptively selects several conflicting topic dimensions under a macro-level issue, retaining only opinions on identifiable topic dimensions while eliminating user statements on irrelevant topics, thus eliminating noise and improving the accuracy and efficiency of subsequent clustering. Furthermore, the vector representations of each key user across all topic dimensions are concatenated to obtain a representation vector for each key user, enabling a... Instead of multiple posts per user, each user is represented by a single vector, significantly reducing the amount of data. Finally, an improved DBSCAN clustering algorithm is used to cluster the representation vectors of all users, resulting in clustering results that can group users for a given macro-topic, yielding multiple user clusters. This improves the accuracy of user tags and helps understand the views and attitudes of individuals and groups on each topic dimension of a given macro-topic on social networks. It also helps to understand the voices of different communities on the internet and the users they attract, grasping the polarization of market demand at the group level, thereby improving relevant strategies or products for the given macro-topic.
[0068] In the process of acquiring social data on a given macro topic from social networks, for the social platforms to be investigated, the relevant post content and user data of the posters can be obtained by means of API acquisition, keyword search, etc. to obtain the corresponding post content and corresponding authors.
[0069] Of course, in this embodiment, open data from various websites or apps is obtained through legal means, in accordance with the Robots Exclusion Protocol of each website or app, and is used only for opinion analysis.
[0070] In this embodiment, based on social data, a set of identifiable topic dimensions that meet preset discrimination conditions is established, specifically including:
[0071] Based on a given macro-level issue type, construct a set of candidate topic dimensions;
[0072] For each candidate topic dimension in the candidate topic dimension set, a pre-defined large language model is used to filter the data in the social dataset to remove data that is irrelevant to that candidate topic dimension.
[0073] The polarization degree of each filtered data point is scored to obtain a polarization score for each data point on the candidate topic dimension;
[0074] The degree of dispersion of each candidate topic dimension is calculated based on the polarization score of each data point on that candidate topic dimension.
[0075] The candidate topic dimensions are sorted according to their degree of dispersion to obtain a sequence of candidate topic dimensions.
[0076] Select the top N candidate topic dimensions from the candidate topic dimension sequence to construct a set of identifiable topic dimensions.
[0077] This embodiment is configured in such a way that the established set of identifiable topic dimensions can be flexibly adapted to the current social dataset, enabling more accurate identification and understanding of individual characteristics in a multidimensional topic space.
[0078] Of course, to ensure the rationality and discriminativeness of the scoring, before using a pre-set large language model to score the polarization degree of each data point in the social dataset and obtaining the polarization score of each data point on the candidate topic dimension, the following steps are also included:
[0079] The preset large language model is fine-tuned so that it can correctly reject scores, that is, it does not score content that is irrelevant to the current dimension.
[0080] This embodiment can also set sub-dimensions for each candidate topic dimension, and use a preset large language model to perform polarization scoring on each data in the social dataset on each sub-dimension of each candidate topic dimension, and then take the average of each sub-dimension as the polarization score of the data on the candidate topic dimension.
[0081] In practice, existing technologies commonly used in the industry can be used to fine-tune the preset large language model, which will not be elaborated further.
[0082] In this embodiment, the post content of each key user is summarized, specifically including:
[0083] For each topic dimension in the set of identifiable topic dimensions, a large language model is used to summarize and categorize each user's posts to obtain the semantics of each user in that topic dimension.
[0084] In this embodiment, the improved DBSCAN clustering algorithm is used to cluster the representation vectors of all users to obtain the clustering results, specifically including:
[0085] Set the minimum number of neighbors, MinPts;
[0086] Calculate the Euclidean distance between each user's representation vector and the representation vectors of all other users;
[0087] Based on the Euclidean distance between each user's representation vector and the representation vectors of all other users, find the MinPts nearest neighbors for each user;
[0088] Determine the spherical neighborhood radius of each user based on each user's MinPts nearest neighbors;
[0089] Sort all users' sphere neighborhood radii in ascending order;
[0090] The sphere neighborhood radius of users located at a preset percentile value is used as the sphere neighborhood radius to correct the value of the sphere neighborhood radius Eps;
[0091] The corrected sphere neighborhood radius for each user is determined based on the sphere neighborhood radius and neighborhood radius threshold for each user.
[0092] Redefine direct density reachability; where direct density reachability is defined as: user u is within the sphere neighborhood of user v, and user v is within the sphere neighborhood of user u;
[0093] The DBSCAN clustering algorithm is used to cluster the representation vectors of all users based on the sphere neighborhood radius and the redefined direct density reachability, and the clustering results are obtained.
[0094] This embodiment can adaptively correct the sphere neighborhood radius according to the quantile and redefine the direct density reachability to make the direct density reachability have a "bidirectional" meaning, thereby preventing chain elongation to a certain extent, causing the cluster edges to tend to be arc-shaped, and avoiding long strips or crescent shapes.
[0095] Because KMeans clustering algorithms are insufficient in removing outliers, and a large number of outliers naturally exist in the current scenario—that is, ordinary users whose views are not close to those of any key users—this embodiment chooses the DBSCAN algorithm, which can naturally exclude outliers. Furthermore, a spherical adaptive clustering radius and bidirectional reachability requirements are designed to guide the formation of approximately spherical, high-density regions centered on the semantics of key users, avoiding the problems of peculiar clustering shapes and lack of prominent central nodes in traditional DBSCAN clustering.
[0096] In a further embodiment, the DBSCAN clustering algorithm is used to cluster the representation vectors of all users based on the corrected spherical neighborhood radius and the redefined direct density reachability, specifically including:
[0097] Step a: Set MinPts;
[0098] Step b: Randomly select a user from all users and determine whether the number of users within the sphere's neighborhood radius of that user is greater than or equal to MinPts; if yes, mark the user as a core user and reconstruct a new cluster for that core user; if no, mark the user as a noisy user.
[0099] Step c: Recursively determine each unmarked user within the user's sphere neighborhood radius until all users within the user's sphere neighborhood radius are marked; wherein, in the recursive determination, if the number of users within a user's sphere neighborhood radius is greater than or equal to MinPts, then mark the user as a core point and add all users within its sphere neighborhood radius to the cluster; if the number of users within a user's sphere neighborhood radius is less than MinPts, then mark the user as a boundary user;
[0100] Step d: Randomly select the next user from the remaining users of all users, and repeat steps b and c to cluster into new clusters until all users have been selected.
[0101] In a further embodiment, before randomly selecting a user from all users, the method further includes:
[0102] All users are divided into key users and ordinary users;
[0103] Specifically, one user is randomly selected from all users, including:
[0104] Randomly select one key user from the key users;
[0105] This includes randomly selecting the next user from the remaining users of all users, specifically including:
[0106] The next key user is randomly selected from the remaining key users of the key user group.
[0107] This embodiment differentiates between key users and ordinary users. When identifying which users hold opposing viewpoints on a macro-level issue, it aggregates ordinary users around key users. This embodiment summarizes online opinions on a specific macro-level issue as follows: centering on a prominent influencer or official account, it identifies which ordinary users share the same viewpoint. This summary is not only clear and concise but also facilitates the implementation of subsequent tasks.
[0108] This embodiment ensures that each cluster has a key user at or near the center, and several ordinary users who share the same viewpoint as him. This avoids the problems of peculiar cluster shapes and lack of prominent central nodes in traditional DBSCAN clusters, making it more suitable for business scenarios.
[0109] In one specific embodiment, the social data also includes the number of followers and activity level of the posting user; wherein, based on the social data, each user is assigned a role to obtain a role assignment result, specifically including: based on the social data, each user is assigned a role according to the number of followers and activity level to obtain a role assignment result.
[0110] In another specific embodiment, users with influential or official badges in a given macro-topic area are designated as key users, while the remaining users are designated as ordinary users.
[0111] Of course, this embodiment can also use existing key user mining methods to mine key users from social data.
[0112] In one specific embodiment, the improved DBSCAN clustering algorithm, using Euclidean distance as a metric, is as follows in the semantic vector space:
[0113] The first step is to target each user Find its MinPts nearest neighbors Calculate the sphere neighborhood radius (Eps) for each user, instead of using a single global Eps value; where the sphere neighborhood radius is... Note that MinPts still uses the traditional DBSCAN algorithm to obtain the same value globally.
[0114] The second step is to include all users. sphere neighborhood radius Sort the spheres from smallest to largest, and take the sphere's neighborhood radius value corresponding to the 90th percentile. This value is denoted as the neighborhood radius threshold. Then set:
[0115] In the formula, Let u be the radius of the sphere's neighborhood. The neighborhood radius threshold, Let be the corrected sphere neighborhood radius for user u;
[0116] according to It can be seen that when a user's ball neighborhood is too sparse, that is... The largest 10% of users will not meet the Eps and MinPts thresholds required for improving the DBSCAN clustering algorithm and cannot be used as core points in the cluster growth process. It should be noted that key users such as big V accounts and official accounts are usually in semantically dense regions and will not be in the sparsest 10%, so they are not expected to be removed. If they fall into the sparsest 10%, they will be removed logically without affecting subsequent steps.
[0117] The third step is to define direct density reachability in the improved DBSCAN clustering algorithm as: user u is in the spherical neighborhood of the representation vector of user v, and user v is in the spherical neighborhood of user u; this setting can prevent chain elongation to a certain extent, and make the cluster edges tend to be arc-shaped, avoiding long strips or crescent shapes.
[0118] The fourth step is to set a unique value MinPts and the sphere neighborhood radius for each user. The clustering results are obtained by setting each key user as the initial core point set and executing the standard DBSCAN clustering process.
[0119] Considering that a large number of ordinary users on social networks exhibit diverse behavioral characteristics, with some providing richer content and others engaging in more interaction, this embodiment also includes interaction data from the users who posted the content.
[0120] After clustering the representation vectors of all users using the DBSCAN clustering algorithm based on the corrected spherical neighborhood radius and the redefined direct density reachability, the following steps are also included:
[0121] For users who are not categorized into any cluster, identify the key users who have the most positive interactions with that user in the interaction data.
[0122] Determine whether the number of positive interactions between the user and the key user is within the preset ranking of the number of positive interactions with the key user;
[0123] If so, the user is considered to have the same view as the key user, and the user is added to the cluster where the key user is located;
[0124] If not, then remove the user completely.
[0125] With this configuration, this embodiment can identify users who are not in any key user's cluster but have positive interactions (such as liking) with a key user, and whose frequency of positive interactions ranks high among the key user's interacting users. Ordinary users believe that this ordinary user shares the same view as the key user who interacts the most. Threshold The value can be selected as needed, for example, between 0.1 and 1. Furthermore, ordinary users who are not in any key user clusters in semantic clustering and have little interaction with key users are removed from the dataset.
[0126] Thus, this embodiment has completed all two major steps: mining several identifiable topic dimensions for specific macro issues and simplifying and focusing all user content onto these topic dimensions; in the semantic vector space, through the improved DBSCAN clustering, clustering a large number of ordinary users who share the same views as each key user into the corresponding cluster.
[0127] Positive interactions include liking and sharing.
[0128] Secondly, the present invention also proposes a user opinion analysis system in a social network, comprising: an acquisition module for acquiring social data on a given macro topic in the social network; wherein the social data includes post content and the user who posted the post;
[0129] The processing module is used to establish a set of identifiable topic dimensions based on social data; for each topic dimension in the set of identifiable topic dimensions, the post content of each key user is summarized, and the summarized semantics are vectorized to obtain the vector representation of each key user in each topic dimension; the vector representations of each key user in all topic dimensions are concatenated to obtain the representation vector of each key user; the improved DBSCAN clustering algorithm is used to cluster the representation vectors of all users to obtain the clustering results.
[0130] In this embodiment, based on social data, a set of identifiable topic dimensions that meet preset discrimination conditions is established, specifically including:
[0131] Based on a given macro-level issue type, construct a set of candidate topic dimensions;
[0132] For each candidate topic dimension in the candidate topic dimension set, a pre-defined large language model is used to filter the data in the social dataset to remove data that is irrelevant to that candidate topic dimension.
[0133] The polarization degree of each filtered data point is scored to obtain a polarization score for each data point on the candidate topic dimension;
[0134] The degree of dispersion of each candidate topic dimension is calculated based on the polarization score of each data point on that candidate topic dimension.
[0135] The candidate topic dimensions are sorted according to their degree of dispersion to obtain a sequence of candidate topic dimensions.
[0136] Select the top N candidate topic dimensions from the candidate topic dimension sequence to construct a set of identifiable topic dimensions.
[0137] In this embodiment, the improved DBSCAN clustering algorithm is used to cluster the representation vectors of all users to obtain the clustering results, specifically including:
[0138] Set the minimum number of neighbors, MinPts;
[0139] Calculate the Euclidean distance between each user's representation vector and the representation vectors of all other users;
[0140] Based on the Euclidean distance between each user's representation vector and the representation vectors of all other users, find the MinPts nearest neighbors for each user;
[0141] Determine the spherical neighborhood radius of each user based on each user's MinPts nearest neighbors;
[0142] Sort all users' sphere neighborhood radii in ascending order;
[0143] Use the sphere radius of users located at a preset percentile value as the neighborhood radius threshold;
[0144] The corrected sphere neighborhood radius for each user is determined based on the sphere neighborhood radius and neighborhood radius threshold for each user.
[0145] Redefine direct density reachability; where direct density reachability is defined as: user u is within the sphere neighborhood of user v, and user v is within the sphere neighborhood of user u;
[0146] The DBSCAN clustering algorithm is used to cluster the representation vectors of all users based on the corrected sphere neighborhood radius and the redefined direct density reachability, and the clustering results are obtained.
[0147] In a further embodiment, the DBSCAN clustering algorithm is used to cluster the representation vectors of all users based on the corrected spherical neighborhood radius and the redefined direct density reachability, specifically including:
[0148] Step a: Set MinPts;
[0149] Step b: Randomly select a user from all users and determine whether the number of users within the sphere's neighborhood radius of that user is greater than or equal to MinPts; if yes, mark the user as a core user and reconstruct a new cluster for that core user; if no, mark the user as a noisy user.
[0150] Step c: Recursively determine each unmarked user within the user's sphere neighborhood radius until all users within the user's sphere neighborhood radius are marked; wherein, in the recursive determination, if the number of users within a user's sphere neighborhood radius is greater than or equal to MinPts, then mark the user as a core point and add all users within its sphere neighborhood radius to the cluster; if the number of users within a user's sphere neighborhood radius is less than MinPts, then mark the user as a boundary user;
[0151] Step d: Randomly select the next user from the remaining users of all users, and repeat steps b and c to cluster into new clusters until all users have been selected.
[0152] In a further embodiment, before randomly selecting a user from all users, the method further includes:
[0153] All users are divided into key users and ordinary users;
[0154] Specifically, one user is randomly selected from all users, including:
[0155] Randomly select one key user from the key users;
[0156] This includes randomly selecting the next user from the remaining users of all users, specifically including:
[0157] The next key user is randomly selected from the remaining key users of the key user group.
[0158] In a further embodiment, after clustering the representation vectors of all users based on the corrected spherical neighborhood radius and the redefined direct density reachability using the DBSCAN clustering algorithm, the method further includes:
[0159] For users who are not categorized into any cluster, identify the key users who have the most positive interactions with that user in the interaction data.
[0160] Determine whether the number of positive interactions between the user and the key user is within the preset ranking of the number of positive interactions with the key user;
[0161] If so, the user is considered to have the same view as the key user, and the user is added to the cluster where the key user is located;
[0162] If not, then remove the user completely.
[0163] 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 opinions in social networks, characterized in that, include: Obtain social data from social networks regarding a given macro-level issue; whereby social data includes post content and the users who posted the posts; Based on a given macro-level issue, construct a set of candidate topic dimensions; For each candidate topic dimension in the candidate topic dimension set, the data in the social dataset is filtered using a pre-set large language model, and the polarization degree of each filtered data is scored to obtain the polarization score of each data on that candidate topic dimension. Based on the polarization score of each data point on the candidate topic dimension, calculate the dispersion of each candidate topic dimension; sort the candidate topic dimensions according to the dispersion of each candidate topic dimension, and select the top N candidate topic dimensions to construct a set of identifiable topic dimensions. For each topic dimension in the set of identifiable topic dimensions, the post content of each user is summarized, and the semantics obtained from the summarization are vectorized to obtain the vector representation of each user in each topic dimension. The vector representations of each user across all topic dimensions are concatenated to obtain the representation vector for each user; the improved DBSCAN clustering algorithm is then used to cluster the representation vectors of all users to obtain the clustering results.
2. The method for analyzing user opinions in social networks according to claim 1, characterized in that, The content of each user's posts was summarized, including: For each topic dimension in the set of identifiable topic dimensions, a large language model is used to summarize and categorize each user's posts to obtain the semantics of each user in that topic dimension.
3. The method for analyzing user opinions in social networks according to claim 1, characterized in that, The improved DBSCAN clustering algorithm is used to cluster the representation vectors of all users, and the clustering results are as follows: Set the minimum number of neighbors, MinPts; Calculate the Euclidean distance between each user's representation vector and the representation vectors of all other users; Based on the Euclidean distance between each user's representation vector and the representation vectors of all other users, find the MinPts nearest neighbors for each user; Determine the spherical neighborhood radius of each user based on each user's MinPts nearest neighbors; Sort all users' sphere neighborhood radii in ascending order; Use the sphere radius of users located at a preset percentile value as the neighborhood radius threshold; The corrected sphere neighborhood radius for each user is determined based on the sphere neighborhood radius and neighborhood radius threshold for each user. Redefine direct density reachability; where direct density reachability is defined as: user u is within the sphere neighborhood of user v, and user v is within the sphere neighborhood of user u; The DBSCAN clustering algorithm is used to cluster the representation vectors of all users based on the corrected sphere neighborhood radius and the redefined direct density reachability, and the clustering results are obtained.
4. The method for analyzing user opinions in social networks according to claim 3, characterized in that, ； In the formula, For users The radius of the sphere's neighborhood. The neighborhood radius threshold, For users The corrected sphere neighborhood radius.
5. The method for analyzing user opinions in social networks according to claim 4, characterized in that, The DBSCAN clustering algorithm is used to cluster the representation vectors of all users based on the corrected spherical neighborhood radius and the redefined direct density reachability. Specifically, it includes: Step a: Set MinPts; Step b: Randomly select a user from all users and determine whether the number of users within the sphere's neighborhood radius of that user is greater than or equal to MinPts; if yes, mark the user as a core user and reconstruct a new cluster for that core user; if no, mark the user as a noisy user. Step c: Recursively determine each unmarked user within the user's sphere neighborhood radius until all users within the user's sphere neighborhood radius are marked; wherein, in the recursive determination, if the number of users within a user's sphere neighborhood radius is greater than or equal to MinPts, then mark the user as a core point and add all users within its sphere neighborhood radius to the cluster; if the number of users within a user's sphere neighborhood radius is less than MinPts, then mark the user as a boundary user; Step d: Randomly select the next user from the remaining users of all users, and repeat steps b and c to cluster into new clusters until all users have been selected.
6. The method for analyzing user opinions in social networks according to claim 5, characterized in that, Before randomly selecting a user from all users, the process also includes: All users are divided into key users and ordinary users; Specifically, one user is randomly selected from all users, including: Randomly select one key user from the key users; This includes randomly selecting the next user from the remaining users of all users, specifically including: The next key user is randomly selected from the remaining key users of the key user group.
7. The method for analyzing user opinions in social networks according to claim 6, characterized in that, Social data also includes interaction data from users who post; After clustering the representation vectors of all users using the DBSCAN clustering algorithm based on the corrected spherical neighborhood radius and the redefined direct density reachability, the following steps are also included: For users who are not categorized into any cluster, identify the key users who have the most positive interactions with that user in the interaction data. Determine whether the number of positive interactions between the user and the key user is within the preset ranking of the number of positive interactions with the key user; If so, the user is considered to have the same view as the key user, and the user is added to the cluster where the key user is located; If not, then remove the user completely.
8. A user opinion analysis system in a social network, characterized in that, include: The acquisition module is used to acquire social data from social networks on a given macro topic; the social data includes post content and the users who posted it. The processing module is used to construct a set of candidate topic dimensions based on a given macro-level issue. For each candidate topic dimension in the set, a pre-defined large language model is used to filter the data in the social dataset, and the polarization degree of each filtered data point is scored to obtain a polarization score for each data point on that candidate topic dimension. Based on the polarization score of each data point on that candidate topic dimension, the dispersion of each candidate topic dimension is calculated. The candidate topic dimensions are sorted according to their dispersion, and the top N candidate topic dimensions are selected to construct a set of identifiable topic dimensions. For each topic dimension in the set of identifiable topic dimensions, the post content of each key user is summarized, and the summarized semantics are vectorized to obtain the vector representation of each key user on each topic dimension. The vector representations of each key user across all topic dimensions are concatenated to obtain the representation vector of each key user. An improved DBSCAN clustering algorithm is used to cluster the representation vectors of all users to obtain the clustering results.
9. The user opinion analysis system in a social network according to claim 8, characterized in that, The improved DBSCAN clustering algorithm is used to cluster the representation vectors of all users, and the clustering results are as follows: Set the minimum number of neighbors, MinPts; Calculate the Euclidean distance between each user's representation vector and the representation vectors of all other users; Based on the Euclidean distance between each user's representation vector and the representation vectors of all other users, find the MinPts nearest neighbors for each user; Determine the spherical neighborhood radius of each user based on each user's MinPts nearest neighbors; Sort all users' sphere neighborhood radii in ascending order; Use the sphere radius of users located at a preset percentile value as the neighborhood radius threshold; The corrected sphere neighborhood radius for each user is determined based on the sphere neighborhood radius and neighborhood radius threshold for each user. Redefine direct density reachability; where direct density reachability is defined as: user u is within the sphere neighborhood of user v, and user v is within the sphere neighborhood of user u; The DBSCAN clustering algorithm is used to cluster the representation vectors of all users based on the corrected sphere neighborhood radius and the redefined direct density reachability, and the clustering results are obtained.
Citation Information
Patent Citations
Social media viewpoint evolution simulation method, device and equipment and storage medium
CN117892510A
Multi-modal social network public opinion hidden danger checking method and system
CN120765004A