A community portrait construction method based on multi-level aggregation

Abstract

The application discloses a kind of community portrait construction methods based on multilevel aggregation, and the community portrait construction method at least includes the following steps: S1: based on social relationship data, user information is represented in vector, and form graph model;S2: based on group identification method, different family is divided;S3: based on opinion leader identification algorithm, opinion leader is identified in different family;S4: based on the opinion leader identified in S3, constructs portrait representation system;S5: based on the portrait representation system constructed, using static community portrait generation method or dynamic community portrait generation method completes community portrait construction.The community portrait construction method of the application solves the problem that traditional portrait system cannot be systematically represented.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence, and in particular relates to a method for constructing community profiles based on multi-level aggregation. Background Technology

[0002] With the rapid development of internet technology, the scale of information dissemination is accelerating, and the number of online social groups is also increasing. How to summarize the characteristics of social groups is a key focus in the fields of public opinion supervision and group recommendation.

[0003] Currently, methods for constructing community profiles can be mainly divided into two categories. The first is feature fusion, which combines personalized user profiles within a community to form a community profile. The second is statistical analysis, which uses mathematical statistical analysis techniques to mine common interest characteristics of users within a community to construct a community profile. Meanwhile, people participate in discussions of current hot news, follow bloggers and topics of interest, and build their own friend circles on social networks, forming social groups of varying sizes and natures. Therefore, how to construct appropriate profile models for social groups of different natures has become another major challenge in the field of profiling. If traditional profile representation methods are used, based on statistical information to analyze users, it is difficult to characterize the features of a community. Therefore, researching methods for representing and constructing social group profiles is highly valuable.

[0004] Currently, profiling technology is mainly used in specialized fields such as e-commerce platforms, libraries, and finance. Profiling models also rely on domain experts for construction. In specific fields, expert experience is used to build a profiling system that incorporates the professional knowledge of that field. Only such a system can accurately depict a user profile.

[0005] However, in the social networking field, user chat content is incredibly diverse, and users have vastly different interests and hobbies. Traditional profiling methods cannot accurately represent a community's profile system. Therefore, how to generate social group profiles has become an urgent problem to be solved. Summary of the Invention

[0006] The purpose of this invention is to overcome the problems of existing technologies by disclosing a community profile construction method based on multi-level aggregation. This invention solves the problem that traditional profile systems cannot systematically represent the community profile.

[0007] The objective of this invention is achieved through the following technical solution:

[0008] A method for constructing community profiles based on multi-level aggregation, the method comprising at least the following steps:

[0009] S1: Based on social relationship data, user information is vectorized to form a graph model;

[0010] S2: Based on the group identification method, different groups are divided;

[0011] S3: Based on the opinion leader identification algorithm, opinion leaders are identified in different groups;

[0012] S4: Based on the opinion leaders identified in S3, construct a profile representation system;

[0013] S5: Based on the constructed profile representation system, the community profile is constructed using either a static or dynamic community profile generation method.

[0014] According to a preferred embodiment, in step S1, the user information includes: basic user information, interaction information between users, blog post content, user following, comments, and likes data.

[0015] According to a preferred embodiment, in step S1, a graph convolutional neural network is used to characterize community features, and user information is vectorized to obtain a graph model, which includes several user features.

[0016] According to a preferred embodiment, step S2 is to cluster several user features obtained in the graph model based on the density peak clustering algorithm, thereby completing the feature extraction of users and dividing them into different groups.

[0017] According to a preferred embodiment, the profile representation system in step S4 represents the identified opinion leaders according to three levels.

[0018] According to a preferred implementation, the first level of the portrait representation system includes eight primary categories: sports, finance, entertainment, society, science and technology, military, health, and international.

[0019] According to a preferred embodiment, the static community profile generation method in step S5 includes the following steps:

[0020] S511: Update user information using influence calculation methods;

[0021] S512: Map user information to the profile representation system constructed in step S4 to form a new user profile;

[0022] S513: Using a fusion strategy, merge the new user profiles obtained in step S512 into a static community profile.

[0023] According to a preferred embodiment, the calculation process of the influence calculation method in step S511 includes:

[0024] (1) Steps to calculate user intimacy: The following formula is used to calculate user intimacy:

[0025]

[0026] Among them, Relation u,v This represents the direct, intimate relationship between users u and v, D. u,v This represents the similarity distance between users u and v;

[0027] and,

[0028] A u and A v Let u and v represent the rating sets for users u and v respectively, and set their values ​​to be in the range of [0,1].

[0029] (2) The user similarity calculation is performed using the following formula:

[0030]

[0031] sim u,v r represents the similarity between members u and v within a group. u,i and r v,i This represents the ratings given by users u and v to topic i. and This represents the average ratings of users u and v for all topics;

[0032] (3) Determine whether someone is an opinion leader and construct a formula for calculating the influence of members.

[0033] The influence of the leader members is:

[0034] The influence of non-leader members is:

[0035] λ is the influence factor of the leader's influence.

[0036] (4) The steps to update user interest preferences include: inputting the initial interest vector of each user in the community according to the member influence calculation formula, and then updating the interests of each member.

[0037] According to a preferred embodiment, in step S513, the fusion strategy is as follows:

[0038] If a member in the community has a lower than set threshold for liking a topic i, the average strategy is used to calculate the average of all members' liking for the topic as the community's liking score.

[0039] When a member in the community expresses a liking for topic i that exceeds a set threshold, the "happiest" strategy is adopted, and the lowest score given by a community member for that topic is used as the community score.

[0040] That is, through a fusion strategy, the degree of user liking for a topic within the community is transformed into the community's degree of liking, denoted as R. G ,in,

[0041]

[0042] InfluceR u,i It is the new score of user u against i after the community update. It is the set threshold.

[0043] According to a preferred embodiment, the dynamic community profile generation method in step S5 includes the following steps:

[0044] S521: Vectorize the dynamic community and input it into a graph neural network;

[0045] S522: Obtain the scoring matrix using graph attention mechanism;

[0046] S533: Map the first K vectors in the rating matrix to the image representation system constructed in step S4 using a similarity-based image mapping algorithm to form a dynamic image.

[0047] According to a preferred embodiment, the graph neural network in step S521 specifically includes an input layer, a network layer, and an output layer; wherein, the first layer of the model is a word embedding layer, which is the input layer, in which the graph structure data is embedded into words, so that natural language is converted into vectors that the model can understand.

[0048] The aforementioned main solution of the present invention and its various further alternative solutions can be freely combined to form multiple solutions, all of which are solutions that can be adopted and are claimed by the present invention. Those skilled in the art, after understanding the solution of the present invention, will realize that there are many combinations based on existing technology and common knowledge, all of which are technical solutions to be protected by the present invention, and will not be exhaustively listed here.

[0049] The beneficial effects of this invention are as follows: This invention proposes a multi-level network community profiling system, dividing community profiles into three levels, from coarse to fine and from shallow to deep, solving the problem of traditional profiling systems being unable to systematically represent data. Simultaneously, a similarity-based profile mapping algorithm is designed to map user data to multi-level profiles. Addressing issues such as large fluctuations in community size and low utilization of user relationships, this invention proposes static and dynamic community profiling methods. It utilizes an influence propagation algorithm to simulate the community's evolution process, employs a profile fusion strategy to fuse user profiles into group profiles, utilizes a graph attention mechanism to mine group characteristics, and then maps the extracted key features to multi-level profiles. Attached Figure Description

[0050] Figure 1 This is a flowchart illustrating the community profile construction method of the present invention;

[0051] Figure 2 This is a schematic diagram of the profile representation system in the community profile construction method of this invention;

[0052] Figure 3 This is a schematic diagram of the influence calculation process in the community profile construction method of this invention. Detailed Implementation

[0053] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0054] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0055] Example 1:

[0056] refer to Figures 1 to 3 As shown in the figure, a community profile construction method based on multi-level aggregation is illustrated. The community profile construction method includes at least the following steps.

[0057] Step S1: Based on social relationship data, vectorize user information to form a graph model.

[0058] Preferably, in step S1, the user information includes: basic user information, interaction information between users, blog post content, user following, comments, and likes data.

[0059] Preferably, in step S1, a graph convolutional neural network is used to characterize community features, and user information is vectorized to obtain a graph model, which includes several user features.

[0060] Specifically, a Graph Convolutional Network (GCN) model is used to characterize community features and vectorize user features. Assuming the input data has C-dimensional features, after feature extraction by the GCN network, its feature dimension is transformed to F-dimensional, and then clustering is performed using the F-dimensional features.

[0061] Graph Convolutional Networks (GCNs) are graph models based on convolutional neural networks. They can extract useful features from complex graph data and utilize these features to perform functions such as graph classification, graph embedding, and edge prediction. Assuming a dataset consists of N nodes, each with M-dimensional features, then these feature matrices X are N*M-dimensional matrices, and their neighbor matrices A are N*N-dimensional matrices. These matrices constitute the model's input. In GCNs, the propagation process is as follows...

[0062] As shown in Equation 1.

[0063]

[0064] A represents the information matrix composed of other matrix nodes connected to node A. This represents the information matrix plus its own information. D represents the matrix composed of the degree information of each node in the graph network. This represents the information matrix plus its own degree. H represents the node attribute matrix, which is a matrix tensor composed of all node attribute columns. The matrix calculation process, such as

[0065] As shown in Equation 2.

[0066]

[0067] Graph convolution is actually an operation on feature attributes and adjacency matrices. Specifically, the adjacency matrix of each convolutional layer is multiplied by the feature attributes and weights of that layer, and then an activation function is applied. The result is used as a new feature attribute to continue the loop operation.

[0068] Step S2: Based on the group identification method, different groups are divided.

[0069] Preferably, step S2 is based on the density peak clustering algorithm to cluster several user features obtained in the graph model, thereby completing the feature extraction of users and dividing them into different groups.

[0070] Traditional community discovery methods suffer from high time complexity when dealing with complex social network data. Density Peak Clustering (DPC), an unsupervised clustering algorithm, significantly reduces computational time complexity. Therefore, this paper introduces the DPC algorithm, combining it with the feature representation results of GCN to provide more accurate clustering results. For each data point, its local density and minimum distance (also known as delta distance) to higher density points are calculated. A decision graph is drawn based on the local density and delta distance, selecting data points with higher local density and larger delta distances as cluster centers. For each non-center point, it is assigned to the cluster containing its nearest higher density point.

[0071] By employing graph convolutional neural networks (Graph Convolutional Neural Networks) and Directed Polymerization (DPC) clustering, user features can be effectively extracted and segmented from complex social data. The superior performance of Graph Convolutional Neural Networks in graph feature extraction makes them highly suitable for community identification. Graph Convolutional Neural Networks can represent nodes as vectors and combine them into communities of different sizes using DPC clustering. This method is more efficient and faster than traditional clustering algorithms, and also offers significant advantages in terms of time complexity.

[0072] Step S3: Based on the opinion leader identification algorithm, identify opinion leaders in different groups.

[0073] Specifically, this embodiment identifies leaders in a social network graph using the following method. A betweenness centrality index is added to the K-kernel decomposition algorithm, calculated as shown in Equation 3. ij Representative node v i With node v j The number of shortest paths that exist between them, n ij (v k ) represents n ij Includes node v k The number of shortest paths. Betweenness centrality measures the importance of a node's position in a social network graph.

[0074]

[0075] When calculating the k-core value of a node, the influencing factor of betweenness centrality is added, and the calculation of its core value is redefined as shown in Equation 4. Here, λ is a constant with a value of 10, which ensures that the final k-value is greater than or equal to 0.

[0076]

[0077] The importance of a node is represented by the sum of the two values. The main process of the algorithm is shown in Table 1.

[0078] The basic idea is to calculate the k-value of a node, which is the product of node degree and node betweenness centrality. Starting with a node with k=1, first delete the nodes with k=1 and their connecting edges, forming subgraph G1. Repeat this step, deleting the nodes with k=2 and their connecting edges, forming subgraph G2, and so on, until all nodes in the graph have been traversed. Finally, the node with the largest k-value is the key node in the graph, i.e., the opinion leader. This algorithm requires traversing all vertices in the graph, with a time complexity of O(n) in the best case and O(n^2) in the worst case. 2 The algorithm has an average time complexity of O(nlogn) and a space complexity of S(n+m). It improves upon the traditional k-core decomposition algorithm by introducing the concept of betweenness centrality, making the k-value more representative of node importance and to some extent diminishing the importance of nodes with high degrees but located on the edges of the group graph.

[0079] Table 1 Improved K-kernel Decomposition Algorithm

[0080]

[0081] Step S4: Based on the opinion leaders identified in S3, construct a profile representation system.

[0082] In various fields, experts have created domain profiles, which are typically used to comprehensively and multidimensionally represent information within that domain. For example, a profile for the field of faults includes detailed information such as the faulty equipment, the nature of the fault, and the equipment manufacturer. However, unlike other fields, social networks are filled with various categories of topics, such as sports, finance, or astronomy, and each of these topics contains even more diverse sub-topics. Therefore, it is impossible to represent a user's topics or interests with a single professional domain profile.

[0083] Preferably, the profile representation system in step S4 represents the identified opinion leaders according to three levels.

[0084] The first level divides community interests into eight primary categories, including finance, sports, and entertainment. The second level, drawing on the classification systems of content platforms such as Baidu and Zhihu, constructs a secondary system containing 100 categories. The third level is built by crawling relevant articles using Python, achieving comprehensive coverage of content information.

[0085] For example, the blog post "The new Chinese men's basketball team begins training in Shanghai to prepare for the sixth window of the FIBA ​​Basketball World Cup qualifiers" can be mapped to articles on related topics within the first-level category of sports, the second-level category of basketball, and the third-level category. Similarly, the blog post "Pfizer discloses its 2022 financial report, with total revenue reaching $100.3 billion, a year-on-year increase of 23%" can be mapped to the first-level category of finance and health, and the second-level categories of stocks and pharmaceuticals.

[0086] A multi-level social group profiling system is proposed. A three-level social profiling system is designed. The first level uses eight categories to represent the profile; the second level refines the profile using 128 categories to represent information; and the third level introduces external articles for calculating profile mapping relationships. Experiments show that the construction of multi-level profiles not only forms a systematic representation of social information but also lays the foundation for subsequent social group profiling.

[0087] Step S5: Based on the constructed profile representation system, complete the community profile construction using either a static community profile generation method or a dynamic community profile generation method.

[0088] This embodiment proposes two types of social group profile generation models: static community profile generation and dynamic community profile generation. In static community profile generation, based on stable social relationships among users, an influence propagation algorithm is used to simulate the community's evolution. After the community reaches stability, a profile fusion strategy is used to merge user profiles into community profiles. In dynamic community profile generation, a graph attention mechanism is used to mine important features within the group, and then the extracted key features are mapped to multi-level profiles.

[0089] Preferably, the static community profile generation method in step S5 includes the following steps: S511: Update user information using an influence calculation method; S512: Map user information to the profile representation system constructed in step S4 using a similarity-based profile mapping algorithm to form a new user profile; S513: Use a fusion strategy to fuse the new user profile obtained in step S512 into a static community profile.

[0090] Furthermore, the calculation process of the influence calculation method in step S511 includes:

[0091] (1) Steps for calculating user intimacy: Equation 5 is used to complete the user intimacy calculation:

[0092]

[0093] Among them, Relation u,v This represents the direct, intimate relationship between users u and v, D. u,vThis represents the similarity distance between users u and v;

[0094] and,

[0095] A u and A v Let u and v represent the rating sets for users u and v respectively, and set their values ​​to be in the range of [0,1].

[0096] (2) Steps for calculating user similarity: Equation 7 is used to complete the user similarity calculation:

[0097]

[0098] sim u,v r represents the similarity between members u and v within a group. u,i and r v,i This represents the ratings given by users u and v to topic i. and This represents the average ratings of users u and v for all topics;

[0099] (3) Determine whether someone is an opinion leader and construct a formula for calculating the influence of members. The influence of a leader member is:

[0100]

[0101] The influence of non-leader members is:

[0102]

[0103] λ is the influence factor of the leader's influence.

[0104] (4) The steps to update user interest preferences include: inputting the initial interest vector of each user in the community according to the member influence calculation formula, and then updating the interests of each member.

[0105] Furthermore, in step S513, the fusion strategy is as follows:

[0106] If a member in the community has a lower than set threshold for liking a topic i, the average strategy is used to calculate the average of all members' liking for the topic as the community's liking score.

[0107] When a member in the community expresses a liking for topic i that exceeds a set threshold, the "happiest" strategy is adopted, and the lowest score given by a community member for that topic is used as the community score.

[0108] That is, through a fusion strategy, the degree of user liking for a topic within the community is transformed into the community's degree of liking, denoted as R. G ,in,

[0109]

[0110] InfluceR u,i It is the new score of user u against i after the community update. It is the set threshold.

[0111] After obtaining the results of user evolution influenced by other members of the community, it is necessary to further obtain the community's interests and hobbies, i.e., the community profile. The evolved user profiles of each user are then fused to obtain the desired community profile. This embodiment discloses a new profile fusion algorithm that can effectively collect the interests of all members in the community, thereby determining their maximum interest value in a particular area. This algorithm can be implemented using various methods such as the mean strategy, the painless average strategy, the fair strategy, and the happiest strategy. Different strategies are suitable for different scenarios, and the most suitable strategy should be selected based on the size of the community.

[0112] This embodiment combines the mean strategy and the happiest strategy to propose a profile fusion strategy suitable for social network graphs. The mean strategy sums and then averages the interest vectors of community members to obtain the community's interest vector. This method is suitable for communities with high user similarity and moderate size. The happiest strategy selects the user with the highest score in the community to represent the entire community. This algorithm performs well with communities that have high user similarity and are large in size.

[0113] Preferably, the dynamic community profile generation method in step S5 includes the following steps: S521: vectorize the dynamic community and input it into a graph neural network; S522: use a graph attention mechanism to obtain a rating matrix; S533: map the first K vectors in the rating matrix to the profile representation system constructed in step S4 using a similarity-based profile mapping algorithm to form a dynamic profile.

[0114] The method for generating dynamic community profiles in step S5 specifically includes:

[0115] Based on existing research on graph model-related algorithms, this paper employs a graph attention algorithm to construct a community profiling model. Graph neural networks can deeply represent user relationships within a community and also possess self-learning capabilities. Furthermore, the introduction of the attention mechanism allows the graph neural network to assign different weights to different nodes, thus making the model focus more on important node information while ignoring different features. Therefore, this paper proposes a dynamic community profiling model based on the graph attention mechanism, which mainly consists of three parts: an input layer, a network layer, and an output layer.

[0116] The first layer of the model is the word embedding layer, which is also the input layer. This layer requires word embedding of the graph-structured data to convert natural language into vectors that the model can understand. This model uses the pre-trained language model BERT to encode sentences, as shown in Equations 11 and 12. Where x... n H represents the nth statement sequence. n This represents the state vector of the nth layer.

[0117] H0 = x1 + x2 + ... + x n (11)

[0118] H n =Transformer(H n-1 (12)

[0119] The BERT model obtains a vector representation of each node, which expresses the relationships between nodes in the dimension of vector space. These node vectors then serve as input to a graph attention neural network to uncover important relationships between nodes.

[0120] Graph neural networks (Graph Neural Networks) can deeply represent user relationships within a community and possess self-learning capabilities. Furthermore, the introduction of attention mechanisms allows Graph Neural Networks to assign different weights to different nodes, enabling the model to focus more on important node information while ignoring different features. The adjacency matrix A of the graph model G and the vector representation Z = Z1, Z2, ..., Zn obtained through the embedding layer are given. n As input to the network layer. Where Z i ∈R d Let the dimension of a node be d. The adjacency matrix A indicates whether there are edges between nodes. The adjacency matrix A = A + I, with the identity matrix I added to incorporate the node's own features. The adjacency matrix then needs to be normalized to prevent the multiplication of the adjacency matrix with the feature matrix from altering the original feature distribution. As shown in Equations 13 and 14... yes The degree matrix.

[0121]

[0122]

[0123] To enhance the eigenvalues ​​of the node vectors, they can be transformed using matrix W, as shown in Equation 15. W represents the vector of node i in the l-th layer of the network. l This represents the parameter matrix to be learned in the l-th layer.

[0124]

[0125] The initial attention calculation formula for a node is shown in Equation 16. Vector i and vector l are concatenated (|| represents vector concatenation), then the result is multiplied by the dot product of the parameter matrix W to be learned, and finally multiplied by the attention weight α, where α represents the importance of vector l to vector i. Finally, the LeakyReLU activation function is applied.

[0126]

[0127] After obtaining the initial attention scores for all neighboring nodes of the target node, it is necessary to calculate the attention weights of the nodes and process them using softmax, as shown in Equation 17.

[0128]

[0129] Then, the attention weights of all neighboring nodes of node i are multiplied and summed by weight to obtain the vector representation of the target node, as shown in Equation 18.

[0130]

[0131] To better capture the features of node vectors, a multi-head attention mechanism is used to build the model, which involves setting up K independent attention mechanisms in each layer.

[0132] The dynamic community profiling model proposed in this embodiment is a supervised training model. The final output of the model is a weight matrix A of the relationships between nodes. Based on the characteristics of the attention mechanism—focusing on important edges while ignoring edges with less influence—the resulting weight matrix A represents which edges in the community are significant and which edges have a smaller impact on the community. The formula is shown in Equation 19, a ij This represents the weight between node i and node j.

[0133]

[0134] After obtaining the weight matrix A, we need to extract the TopK edges to represent the characteristics of the entire community. In practical terms, the TopK edges represent the most popular or meaningful statements discussed by users within the community. These statements can effectively characterize the topics that the dynamic community focuses on. Regarding the selection of the value of K, considering that the number of edges between nodes in a graph structure increases exponentially with the size of the nodes, a piecewise approach is used to select the size of K, as shown in Equation 20, where |V| represents the number of edges.

[0135]

[0136] After obtaining the TopK edges of the ratings in the weight matrix A, a similarity-based profile mapping algorithm is used to map the content represented by these edges to a multi-level profile system, thereby obtaining a profile of the dynamic community.

[0137] This invention proposes a multi-level network community profiling system, dividing community profiles into three levels, from coarse to fine and from shallow to deep, solving the problem of traditional profiling systems' inability to systematically represent data. Simultaneously, a similarity-based profile mapping algorithm is designed to map user data to multi-level profiles. Addressing issues such as large fluctuations in community size and low utilization of user relationships, this invention proposes static and dynamic community profiling methods. It utilizes an influence propagation algorithm to simulate the community's evolution, employs a profile fusion strategy to merge user profiles into group profiles, leverages graph attention mechanisms to mine group features, and then maps the extracted key characteristics onto multi-level profiles.

[0138] This invention is compared with algorithms commonly used in community leader profiling. Table 2 shows the results for AGP, NC, and χ². 2 The resulting profiles obtained by this method are more concentrated in the sports category, with less attention paid to names of people and places within the topic. Analysis shows that statistical methods have a significant advantage in identifying high-frequency categories, while low-frequency categories are easily overlooked, thus ignoring potentially low-frequency but important categories. Comparing the results of the four algorithms reveals that the profiles obtained by the AGP and NC methods have a higher similarity, χ² 2 The differences between the profile results obtained by the AGP and NC algorithms are significant. Analysis shows that user influence has a substantial impact on the accuracy of community profiling, information that is not reflected in the AGP and NC algorithms. Table 2 shows that in dynamic profiling results, the CGP, LP, and DSGP algorithms all obtained the best category information, while the BNS algorithm obtained a lower score. Analysis reveals that BNS, a community profiling algorithm based on classification, ignores user topics within the community in dynamic communities, focusing instead on individual user profiles, thus resulting in poor performance. Comparing the results of CGP, LP, and DSGP, it can be found that the CGP and LP algorithms both obtained lower scores for the data point "release time," while DSGP obtained higher category scores. Analysis shows that the CGP and LP algorithms rely on node importance for ranking, ignoring implicit relationships between nodes. The DSGP algorithm, based on graph attention mechanisms, can better capture implicit relationships between nodes, thus performing better.

[0139] Table 2 Comparison of Static Community Profile Results

[0140]

[0141]

[0142] Table 3 Comparison of Dynamic Community Profile Results

[0143]

[0144] Table 4. Static profile results of basketball communities

[0145] Model HR@5 AP@5 HR@10 AP@10 AGP 0.0 0.0 0.4 0.267 NC 0.2 0.1 0.6 0.511 <![CDATA[χ 2 ]]> 0.4 0.333 0.7 0.653 SSGP 0.6 0.32 0.8 0.733

[0146] Table 5. Dynamic profiling results of film communities

[0147]

[0148]

[0149] Table 4 presents the comparative experimental results of different community profiling methods in static communities, while Table 5 presents the comparative experimental results of different community profiling methods in dynamic communities. Table 6 shows that in static communities, the SSGP model proposed in this invention achieves better overall results, and in the AP@5 metric, it outperforms SSGP. This is mainly because the SSGP algorithm considers node centrality when calculating user influence, thus assigning higher influence to entities with higher in-degree and out-degree. Meanwhile, the AGP and NC algorithms perform poorly because they focus on categories with higher frequency or popularity, ignoring categories with lower frequency but higher contribution. Furthermore, the AGP algorithm's results are lower than the NC algorithm's because the AGP algorithm only calculates the results by statistically analyzing the highest frequency categories, while the NC algorithm calculates user importance, reducing the interference of frequency on the results to some extent. This also demonstrates the crucial role of user influence in the results of static community profiling. Table 6 shows that in dynamic communities, the DSGP model proposed in this paper achieves better overall results, outperforming the other three algorithms in all four metrics. Furthermore, BNS outperforms the other two algorithms in AP@10, while CPG outperforms them in AP@5. Analysis shows that CGP achieves better Top-5 accuracy at smaller scales, while in Top-10, the BNS algorithm, using influence calculation, outperforms CGP, indicating inconsistent performance across the two metrics. The LP algorithm performs poorly across all four metrics because it assigns higher weights to user attribute and statistical information, which contradicts the characteristics of dynamic groups. Therefore, the proposed DSGP model demonstrates good performance in generating dynamic community profiles.

[0150] The similarity of user profiles is also an important indicator for evaluating the performance of community profile models. This paper uses cosine similarity and Kendall's coefficient as indicators to measure the similarity between models. Similarity experiments were conducted by selecting the top-5 category distributions of community profiles generated by different models. As shown in Table 6, the first column represents the two comparison models; the second column represents the cosine similarity results of the two models, denoted by C; and the third column represents the Kendall's rank coefficients of the two models. Analysis of the table shows that in static communities, the similarity with SSGP is the highest, while the similarity with AGP and SSGP is the lowest. This result is similar to the results shown by several models in Table 4, further illustrating that the experimental results with SSGP are superior. In dynamic communities, CGP and DSGP models, as well as BNS and LN models, all achieved the highest similarity, consistent with the trend in Table 5. This indicates that CGP and DSGP models share similar approaches, both utilizing the relationships between users within the community, thus resulting in a certain degree of similarity in their profiles.

[0151] Table 6. Similarity among different community profiling models

[0152]

[0153]

[0154] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for constructing community profiles based on multi-level aggregation, characterized in that, The community profiling construction method includes at least the following steps: S1: Based on social relationship data, user information is vectorized to form a graph model; S2: Based on the group identification method, different groups are divided; S3: Based on the opinion leader identification algorithm, opinion leaders are identified in different groups; S4: Based on the opinion leaders identified in S3, construct a profile representation system; S5: Based on the constructed profile representation system, the community profile is constructed using either a static or dynamic community profile generation method. In step S1, user information includes: basic user information, interaction information between users, blog post content, user following, comments, and likes data; In step S1, a graph convolutional neural network is used to characterize community features and to vectorize user information to obtain a graph model, which includes several user features. Step S2 is based on the density peak clustering algorithm to cluster several user features obtained in the graph model, thereby completing the feature extraction of users and classifying them into different groups; The profiling system in step S4 involves representing the identified opinion leaders according to three levels; The first level of the portrait representation system includes eight primary categories: sports, finance, entertainment, society, technology, military, health, and international. The static community profile generation method in step S5 includes the following steps: S511: Update user information using influence calculation methods; S512: Map user information to the profile representation system constructed in step S4 to form a new user profile; S513: Using a fusion strategy, merge the new user profiles obtained in step S512 into a static community profile; The calculation process for the influence calculation method in step S511 includes: (1) Steps to calculate user intimacy: The following formula is used to calculate user intimacy: in, This indicates the direct and intimate relationship between users u and v. This represents the similarity distance between users u and v; and, and Let u and v represent the rating sets for users u and v respectively, and set their values ​​to be in the range of [0,1]. (2) The user similarity calculation is performed using the following formula: This represents the similarity between members u and v within a group. and This represents the ratings given by users u and v to topic i. and This represents the average ratings of users u and v for all topics; (3) Determine whether someone is an opinion leader and construct a formula for calculating the influence of members. The influence of the leader members is: ; The influence of non-leader members is: The influence factor of leader members' influence weight; (4) The steps to update user interest preferences include: inputting the initial interest vector of each user in the community according to the member influence calculation formula, and then updating the interests of each member; In step S513, the fusion strategy is as follows: When a member of the community discusses a certain topic If the liking level is lower than the set threshold, the average strategy is adopted, and the average liking level of all members on the topic is used as the liking level of the community. When a member of the community discusses a topic When the liking level exceeds the set threshold, the "happiest" strategy is adopted, and the lowest score given by community members to the topic is used as the community score. That is, through a fusion strategy, the degree of user liking for a topic within the community is transformed into the community's degree of liking, denoted as... ,in, Users after the community update right New score, It is the set threshold; The dynamic community profile generation method in step S5 includes the following steps: S521: Vectorize the dynamic community and input it into a graph neural network; The neural network in step S521 specifically includes: an input layer, a network layer, and an output layer; The first layer of the model is the word embedding layer, which is the input layer. In this layer, the graph structure data is embedded with words, so that the natural language is transformed into vectors that the model can understand. S522: Obtain the scoring matrix using graph attention mechanism; S533: Map the first K vectors in the rating matrix to the image representation system constructed in step S4 using a similarity-based image mapping algorithm to form a dynamic image.

Images