Key user identification method, device and system in derivative topic network

Abstract

The invention belongs to the field of social network information processing, and particularly relates to a key user identification method, device and system in a derivative topic network. The method comprises the following steps: acquiring data information in a topic network; extracting a native topic, a derivative topic user relationship network, a native topic, a derivative topic propagation network, user basic attributes and intimacy among users from the data information in the topic network; extracting user feature vector representation through a community label propagation algorithm according to the basic attributes of the users and the intimacy between the users; extracting topic vector representations of the native topics and the derivative topics through a joint distribution adaptive algorithm according to the native topic propagation network and the derivative topic propagation network; and inputting the user feature vector representation and the topic vector representation into a user influence dynamic time-space model based on an attention mechanism, and outputting the user as an opinion leader user or a long-tail user. According to the method, the key users in the derivative topic network can be accurately identified.

Description

Technical Field

[0001] The present invention belongs to the field of Internet application technology, relates to data analysis in social networks, and in particular to a method, device and system for identifying key users in a derivative topic network. Background Art

[0002] The rise and popularity of social media has enabled a large number of users to freely express their opinions, share information, and participate in extensive discussions and interactions. This interactive behavior has built a huge user network in social networks, in which the influence of users plays a vital role in information dissemination. However, the booming development of social networks has also brought a series of problems. In particular, in some special events, a series of new topics have been derived, which have aroused heated discussions among netizens. The information spreads quickly and widely, which has brought challenges to the information management of social networks. At the same time, it may also trigger widespread online public opinion, and its scope of influence and speed of dissemination are very huge.

[0003] The influence analysis of derivative topics has a significant impact in many fields. On social network platforms, the influence analysis of derivative topics can reveal the degree of user participation in topics, the frequency of interaction, and the path of information dissemination, thereby helping the platform to better understand user needs and preferences, help identify potential target customer groups, understand their needs and preferences, optimize recommendation algorithms, and thus formulate more accurate recommendation strategies and improve user experience.

[0004] At present, researchers mainly divide the influence analysis of derived topic nodes into three aspects. The first is based on the research of social network structure, which mainly focuses on the relationship and connection structure between nodes in social networks, as well as the community structure and information propagation path in social networks, and identifies the key nodes in social networks through the above indicators and methods. The second is based on the research of user behavior, which mainly analyzes a series of communication behaviors and activities such as likes, comments and forwarding of topics by users in social networks. By analyzing these behavioral data, the user influence score is calculated to identify users who have made important contributions to information dissemination and social influence in social networks. The third is based on the research of topic information, which mainly analyzes text sentiment, distinguishes user emotional tendencies, and then analyzes user influence. Fundamentally speaking, the above technologies analyze node influence from different angles through the existing core content of social networks.

[0005] As the amount of social network data continues to expand and deep learning technology matures, nonlinear models such as neural networks and deep learning have gradually become the mainstream choice for user influence analysis models. However, these existing technologies cannot accurately describe the spread scale of original topics and derivative topics, as well as the differences in users' cognition of original topics and derivative topics, making it difficult to find key users in the early stages of derivative topic spread. Summary of the invention

[0006] In view of the problems existing in the prior art, the present invention proposes a method, device and system for identifying key users in a derivative topic network, which realizes the analysis of the influence of early users on derivative topics based on the scale of dissemination and cognitive differences. The present invention utilizes the relationship between the original topic and the derivative topic, and combines the user's own characteristics, the semantic information of the derivative topic theme and the self-learning ability of the deep neural network to discover key users in the early stage of the spread of the derivative topic.

[0007] In a first aspect of the present invention, the present invention provides a method for identifying key users in a derived topic network, the method comprising:

[0008] Acquire data information in the topic network; the data information includes basic user information, basic topic information and user relationship network;

[0009] Extracting original topic and derived topic user relationship network, original topic and derived topic propagation network, user basic attributes and intimacy between users from the data information in the topic network;

[0010] According to the basic attributes of users and the intimacy between users, the user feature vector representation is extracted through the community label propagation algorithm;

[0011] According to the original topic and derived topic propagation network, the topic vector representation of the original topic and derived topic is extracted through the joint distribution adaptive algorithm;

[0012] The user feature vector representation and the topic vector representation are input into a user influence dynamic spatiotemporal model based on an attention mechanism, and the output is an opinion leader user or a long-tail user.

[0013] In a second aspect of the present invention, the present invention further provides a key user identification device in a derived topic network, comprising:

[0014] A data acquisition module is used to acquire data information in the topic network; the data information includes basic user information, basic topic information and user relationship network;

[0015] A data extraction module is used to extract the original topic, derived topic user relationship network, the original topic, derived topic propagation network, user basic attributes and user intimacy from the data information in the topic network;

[0016] User feature extraction module, used to extract user feature vector representation based on user basic attributes and user intimacy through community label propagation algorithm;

[0017] The topic feature extraction module is used to extract the topic vector representation of the original topic and the derived topic through the joint distribution adaptive algorithm based on the original topic and the derived topic propagation network;

[0018] The key user identification module is used to input the user feature vector representation and the topic vector representation into the user influence dynamic spatiotemporal model based on the attention mechanism, and output the opinion leader user or the long-tail user.

[0019] In a third aspect of the present invention, the present invention further provides a key user identification system in a derived topic network, comprising:

[0020] The transceiver module is used to obtain data information in the topic network; the data information includes basic user information, basic topic information and user relationship network;

[0021] A processing module extracts the original topic and derived topic user relationship network, the original topic and derived topic propagation network, the basic attributes of users and the intimacy between users from the data information in the topic network; extracts the user feature vector representation through the community label propagation algorithm based on the basic attributes of users and the intimacy between users; extracts the topic vector representation of the original topic and the derived topic through the joint distribution adaptive algorithm based on the original topic and derived topic propagation network; inputs the user feature vector representation and the topic vector representation into the user influence dynamic spatiotemporal model based on the attention mechanism, and outputs the opinion leader user or the long-tail user;

[0022] A display module is used to display the opinion leader users or long-tail users output by the processing module.

[0023] Beneficial effects of the present invention:

[0024] The method, device and system for identifying key users in the above-mentioned derived topic network extract the original topic, derived topic user relationship network, the original topic, derived topic propagation network, user basic attributes and user intimacy by acquiring data information in the topic network; extract the user feature vector representation through the community label propagation algorithm based on the user basic attributes and user intimacy; extract the topic vector representation of the original topic and derived topic through the joint distribution adaptive algorithm based on the original topic and derived topic propagation network; input the user feature vector representation and the topic vector representation into the user influence dynamic spatiotemporal model based on the attention mechanism, and output it as an opinion leader user or a long-tail user, thereby improving the recognition accuracy of key users of the derived topic network in the social network platform. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 This is a schematic diagram of a key user identification scenario in an embodiment of the present invention;

[0026] Figure 2 is a flow chart of a key user identification method according to an embodiment of the present invention;

[0027] Figure 3 is a schematic diagram of implicit influence of users based on intimacy between users according to an embodiment of the present invention;

[0028] Figure 4 Schematic diagram of a topic vector representation method based on emotion according to an embodiment of the present invention;

[0029] Figure 5 Schematic diagram of a dynamic spatiotemporal model of user influence according to an embodiment of the present invention;

[0030] Figure 6 is a schematic diagram of the structure of a key user identification device according to an embodiment of the present invention;

[0031] Figure 7 It is a schematic diagram of the structure of a key user identification system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0032] In order to better illustrate the technical solution of the present invention and make the advantages more concise and clear, the problem to be solved by the present invention is first specifically explained below, and then the specific implementation methods of the present invention are further described in detail with reference to the accompanying drawings of the specification.

[0033] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0034] The terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or modules is not necessarily limited to those steps or modules clearly listed, but may include other steps or modules that are not clearly listed or inherent to these processes, methods, products or devices. The division of modules that appears in this article is only a logical division. There may be other division methods when implemented in actual applications, such as multiple modules can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between each other shown or discussed can be through some interfaces, and the indirect coupling or communication connection between modules can be electrical or other similar forms, which are not limited in this article. Furthermore, the modules or sub-modules described as separate components may or may not be physically separated, may or may not be physical modules, or may be distributed in multiple circuit modules, and some or all of the modules may be selected according to actual needs to achieve the purpose of the embodiments of the present invention.

[0035] Considering the current status of user influence research, there are still several challenges in using information related to native topics to identify key users in the derived topic network:

[0036] 1. It is difficult to discover the hidden influence of key users in the early stage of the outbreak of derivative topics. At this stage, it is difficult to discover influential key users because the topic is still new, the user participation and influence are small, and the participating users have not yet formed clear opinions and attitudes.

[0037] 2. Cognitive differences caused by user domain migration. User domain migration leads to cognitive differences in knowledge between the original topic domain and the derived topic domain, which in turn affects the user's attitude and behavior towards the topic and also has a certain impact on the user's influence.

[0038] 3. The influence of users of derivative topics is in a dynamic state. Key users of original topics may lack relevant expertise in the domain of derivative topics, thus transforming from key users to long-tail users; at the same time, new and more influential key users may emerge.

[0039] Based on the above technical issues, Figure 1 FIG. 1 is an application environment diagram of a key user identification method in an embodiment. Figure 1, the key user identification method is applied to the terminal 101 and the server 102. The terminal 101 and the server 102 are connected through a network. The terminal 101 sends data information in the topic network to the server 102; and receives the result that the user is an opinion leader user or a long-tail user from the server 102; wherein the server 102 obtains the data information in the topic network from the terminal 101; the data information includes user basic information, topic basic information and user relationship network; extracts the original topic, derived topic user relationship network, original topic, derived topic propagation network, user basic attributes and user intimacy from the data information in the topic network; according to the user basic attributes and user intimacy, extracts the user feature vector representation through the community label propagation algorithm; according to the original topic and derived topic propagation network, extracts the topic vector representation of the original topic and the derived topic through the joint distribution adaptive algorithm; the terminal 110 inputs the user feature vector representation and the topic vector representation into the user influence dynamic spatiotemporal model based on the attention mechanism, and outputs the opinion leader user or the long-tail user. It is understandable that the user identification method in the derivative topic network provided in the embodiment of the present application can also be applied to a system including a terminal and a server, and implemented through the interaction between the terminal and the server. Among them, the terminal 101 can include but is not limited to various personal computers, laptops, tablet computers, etc. The server 102 can be implemented by an independent server or a server cluster composed of multiple servers.

[0040] like Figure 2 As shown, in some embodiments, the present application embodiment provides a method for identifying users in a derived topic network, the method comprising:

[0041] 201. Acquire data information in a topic network; the data information includes basic user information, basic topic information, and a user relationship network;

[0042] In an embodiment of the present invention, the method of obtaining data is to directly query the real-time social network data in the database provided by the enterprise. The database stores active data, social data, etc. generated by all users in various social network platforms. The user may be a user who has logged in to the social network platform. What needs to be obtained here is the user's basic information (such as age, region, number of fans, number of friends, etc.), the interaction of native topics and derived topics (such as likes, favorites and forwarding) and the historical behavior data of users who have interacted with them. The participation in native topics and derived topics requires the basic information of the interacting users and the basic information of the topic. The historical behavior of the interactors of native topics and derived topics includes the information related to the native topics and derived topics that the user has interacted with in history. And the data is preprocessed.

[0043] 202. Extracting the original topic, derived topic user relationship network, the original topic, derived topic propagation network, user basic attributes and intimacy between users from the data information in the topic network;

[0044] In social networks, the influence of users is affected by many factors, including social relationships, interests, hobbies, historical behaviors, etc. Based on this, extract user-related attributes:

[0045] In some embodiments, a single message propagation network is composed of a set of all users participating in topic propagation and a set of propagation paths of all users during topic propagation; it can be expressed as:

[0046]

[0047] Formula (1) is a single message propagation network. Indicates participation in T i The set of all users who spread the topic, Indicates that all users in T i The set of propagation paths in the process of topic propagation, together they constitute T i Topic Communication Network

[0048] In some embodiments, the original and derived topic propagation networks N OT 、N DT It is composed of the user sets participating in the original and derived topic propagation and the total number of messages in the corresponding topic propagation network respectively; it can be expressed as:

[0049]

[0050] Formula (2) represents the original topic propagation network. Formula (3) represents the derived topic propagation network. is the i-th user in the native topic propagation network, is a single message propagation network, |S| is the total number of messages in the original topic propagation network, and |M| is the total number of messages in the derived topic propagation network.

[0051] In some embodiments, the original and derived topic user relationship network N OR 、N DR It is composed of the user set and user relationship set participating in the original and derived topic propagation, respectively, and can be expressed as:

[0052]

[0053] Formula (4) represents the native topic user relationship network. Common user relationship networks include likes, comments, reposts, and mentions. The network type is represented by R. Represents the i-th user who participated in the original topic, It means that the i-th user and the j-th user in the original topic network have R types of relationships. Formula (5) is expressed as the derived topic user relationship network.

[0054] In some embodiments, the basic attributes of a user may be composed of the number of fans, the number of friends, the certification field such as the entertainment field, the certification community such as the celebrity super community, and the emotional characteristics of the user. The basic attributes of the user Att(u i ) is expressed as:

[0055] Att(u i )=[fansNum(u i ),friNum(u i ),field(u i ),cerCom(u i ),emo(u i )] (6)

[0056] Formula (6) represents the basic attributes of users. fansNum(u i ) represents user u i The number of fans, friNum(u i ) represents user u i The number of friends, field(u i ) represents user u i The authentication field may be null. i ) represents user u i The authentication community, there may be multiple values, emo(u i ) represents user u i The emotional features can be the result of sentiment analysis based on the user's social behavior, speech content, etc.

[0057] In some embodiments, the user intimacy is determined by the similarity score, difference score and weighted score of the same attribute between users and the same interaction mode between users. i ,u j ) is expressed as:

[0058]

[0059] Among them, I(u i ,u j ) represents user u i and user u j The degree of intimacy between m represents the weight of attribute m, S(u i ,u j ,m) represents user ui and user u j The similarity score on attribute m, P(u i ,m)-P(u j ,m) represents user u i and user u j In the difference for attribute m, w n represents the weight of the nth interaction mode, B(u i ,u j ,n) represents user u i and user u j The score on the nth interaction method.

[0060] 203. According to the basic attributes of users and the intimacy between users, the user feature vector representation is extracted through the community label propagation algorithm;

[0061] In the embodiment of the present invention, this embodiment utilizes the strong correlation between the original topic and the derived topic, starting from the original topic, by designing user intimacy and introducing a community label propagation algorithm, to obtain the community characteristics of the node.

[0062] In some embodiments, the method for extracting the user feature vector representation includes processing the intimacy between users and the frequency of label occurrence by using a propagation algorithm based on a speaker monitor, and calculating the acceptance probability of each label; based on the acceptance probability of each label, calculating the probability that the user belongs to the label; based on the probability that the user belongs to the label, obtaining a node label set, the number of connected communities and the size of the community, and generating a user feature vector representation through an embedding model.

[0063] like Figure 3 As shown, assuming that the derived topic user relationship network contains 14 users, this embodiment is based on the user relationship network, and the community label propagation algorithm is used to divide the network into communities to obtain the community and community characteristics of each user. It can be seen that user 2, user 5 and user 11 form a user community; user 1, user 4, user 6, user 7 and user 12 form a user community; user 3, user 8, user 9, user 10, user 12, user 13 and user 14 form a user community. These community characteristics are used as the basic attributes of the user to generate a preliminary user feature vector representation, and then these features are further updated and optimized through a two-layer graph convolutional network to obtain a more accurate user feature vector representation. Specifically, this embodiment uses the intimacy between users as the basis for selecting labels for neighbor nodes, and calculates the acceptance probability of each label by the intimacy between users and the frequency of label occurrence. Assuming u i As the initial node, u j As a speaker, I(u i ,u j) represents user u i and user u j The intimacy between them. Then the probability of each label receiving is:

[0064]

[0065] Where l is a tag in the tag list, counts(l) is the frequency of the tag, and sum(l) is the total number of tags.

[0066] In the propagation algorithm based on speaker listener, each user node will obtain labels from its neighboring nodes. Based on the probability of receiving each label, the label with the highest probability can be selected for acceptance and propagated to the neighboring nodes. After multiple rounds of label propagation and updating, each user node will have a final label set; obtain the user label set, the number of connected communities and the size of the community; according to the user label set, the number of connected communities and the size of the community, generate a user feature vector representation through the embedding model. Each label set represents one or more communities to which the node belongs. The number of communities connected to the node is the number of different labels to which the node ultimately belongs. Assume that the node u i The label set is L(u i ), then the number of communities connected to the node is the number of labels in the label set. i The number of connected communities is:

[0067] Num cerCom (u i )=|L(u i )| (9)

[0068] The size of the community to which a node belongs refers to the number of all other nodes in the community to which the node belongs. The label set L(u i ) can determine the community to which a node belongs. If node u i If a node belongs to label l, then the community corresponding to this label is the community composed of all nodes labeled l. The community size is the number of nodes in this community. Let N l is a set of nodes with label l, then:

[0069] Num size (u i ,l)=|N l | (10)

[0070] Among them, N l is the set of all nodes with label l, |N l | represents the size of the community, that is, the number of nodes in the community corresponding to label l.

[0071] The number of connections between nodes refers to the number of nodes u i The number of edges between nodes and other nodes, that is, the degree of the node, is denoted by deg(u i ). Node u i The degree of deg(u i ) represents node u i The number of connections between it and its neighboring nodes. That is:

[0072] Num degree (u i )=deg(u i )=|N(u i )| (11)

[0073] Among them, N(u i ) represents node u i The neighbor node set, |N(u i )| represents the node u i The number of connected neighbor nodes.

[0074] The embodiment of the present invention uses modularity as a community division evaluation index. The larger the modularity value, the better the division quality. Modularity is a quantitative index that is not affected by subjective judgment and can objectively reflect the quality of community division. Modularity is used to optimize the basic attributes of users composed of the number of connected communities and the size of the community, thereby further optimizing the user feature vector representation; making the final user feature vector representation more accurate.

[0075] Furthermore, the embodiment of the present invention uses the number of communities to which a node is connected, the size of the community, and the number of connections between nodes as the user's own attributes. According to formulas (9)-(11), the single user feature vector M(u i )={Num cerCom (u i ),Num size (u i ,l),Num degree (u i )}, the community user feature vector is represented by X = N × M (u i ), N represents the number of nodes in a single community network.

[0076] In some embodiments, this embodiment extracts the user feature vector representation through an embedding model composed of a two-layer graph convolutional network, and the graph convolutional network updates the node embedding vector by propagating node feature information layer by layer. When processing graph data, the graph convolutional network can capture the local neighbor information of the node, thereby generating useful node embeddings. By processing the initial user feature vector representation X, the adjacency matrix A, and the degree matrix D through the embedding model, the final user feature vector representation can be extracted more accurately.

[0077] 204. According to the original topic and the derived topic propagation network, the topic vector representations of the original topic and the derived topic are extracted by a joint distribution adaptive algorithm;

[0078] In social networks, the influence of users on the dissemination of original topics and derived topics changes due to domain migration. The cognitive differences caused by this migration lead to inconsistency in the distribution of content data. In order to improve the generalization ability of the recognition model on derived topics, the embodiment of the present invention adopts a joint distribution adaptation method to optimize the performance of the model in the two topic content spaces. The topic vector representation of the original topic and the derived topic is extracted in the following way:

[0079] Extract message content and user content through the original topic propagation network and the derived topic propagation network;

[0080] Extract keywords from message content and user content to obtain feature vectors of topic content and user sentiment;

[0081] The feature vector of topic content and the feature vector of user emotion are combined with the number of users participating in the original topic and derived topics, and the topic vector representation of the original topic and derived topics is obtained through a joint distribution adaptive algorithm.

[0082] like Figure 4 As shown, the embodiment of the present invention can extract the characteristics of topics and users from the propagation network of native topics and derived topics through user feature vector representation and topic semantic information matrix, thereby obtaining user sentiment difference characteristics; combined with the joint distribution adaptive algorithm, by optimizing the performance of the model between the source domain and the target domain, a more accurate topic vector representation is obtained.

[0083] Specifically, the embodiment of the present invention divides the data into two parts: topic content and user content. OT 、N DT Message content and user collections are extracted from the dataset and embedded into the topic semantic information matrix to explore the similarities and differences of user emotions. Next, the topic content and user content are preprocessed, and keywords are extracted through the term frequency-inverse document frequency (TF-IDF) and text ranking (TextRank) algorithms to obtain the characteristics of topics and user emotions. Finally, the TE2Vec method is used to generate the feature vectors of native topics and derived topics:

[0084]

[0085] Among them, T O and T DRepresented as the feature vector of topic content, and Represented as the feature vector of user emotion, C O and C D Represented as the number of users participating in the original topic and the derived topic respectively.

[0086] It should be noted that emo(u i ) reflects the user’s overall emotional tendency, and It is the specific manifestation of the emotional characteristics corresponding to these overall emotional tendencies in original topics and derived topics. and It is an emotion vector generated based on the user’s emotion expression in different topics. i ) is a relatively basic emotional feature representation. and It is a high-dimensional emotional feature vector generated based on the application and changes of the emotional feature in a specific topic scenario.

[0087] source domain in, Target domain in, Assume V O =V D , Y O =Y D , Using pseudo labels, combined with marginal MMD metric and class conditional probability distribution MMD metric, we find an exchange variable W. In the self-space corresponding to W, The differences in the derived topic user sentiment domain are significantly reduced, and the content features W T E.

[0088] In this embodiment, the topic vector representation of the original topic is regarded as the source domain, and the topic vector representation of the derived topic is regarded as the target domain; by minimizing the square of the marginal distribution difference between the source domain and the target domain, and the square of the difference between each type of sample in the source domain and the target domain, the optimal topic vector representation of the original topic and the derived topic is obtained. The overall optimization goal is:

[0089]

[0090] Among them, H is the center matrix, U is all the users who participated in the original topic and the derived topic, tr(W T UM0U T W) represents the square of the marginal distribution difference between the source domain and the target domain, tr(W T UM V U TW) represents the square of the difference between the Vth class samples in the source domain and the target domain. It is calculated as follows:

[0091]

[0092] The source domain and the target domain contain multiple categories, and each category corresponds to a certain number of samples. The Vth category samples refer to samples belonging to the Vth category, which can be divided according to different features, such as sentiment categories, topic categories, etc.

[0093] 205. Input the user feature vector representation and the topic vector representation into a user influence dynamic spatiotemporal model based on an attention mechanism, and output an opinion leader user or a long-tail user.

[0094] In the embodiment of the present invention, it is considered that the influence of users may grow or weaken over time because their social relationships, activities and content may change. Therefore, in order to analyze the influence of users more accurately, this embodiment considers the dynamic changes of user influence in time and space.

[0095] The user influence dynamic spatiotemporal model based on attention mechanism regards user influence as a dynamically changing spatiotemporal process, and optimizes the analysis of user behavior patterns and time series data by introducing attention mechanism. It can not only capture the interactive relationships between users, but also determine how these relationships evolve over time and space.

[0096] like Figure 5 As shown in the figure, by combining the user feature vector representation and the topic vector representation, the dynamic spatiotemporal model uses GRU to capture time series features, uses the attention mechanism to improve the importance judgment of different time steps and inputs, and finally uses GCN to capture graph structure information for global modeling. Ultimately, based on these dynamic spatiotemporal features, the model can predict the influence of users in social networks and classify them as opinion leaders or long-tail users.

[0097] Since GRU can capture the characteristics before and after the time series and has significant advantages in predicting the time series, this embodiment captures the characteristics before and after the topic time series by introducing GRU. For each node in the topic network, GRU performs feature mining on the input topic vector representation and user feature vector representation. Under the joint action of the activation function and the weight matrix, the vector representation information is updated and reset. The gated recurrent neural network module based on the time feature effectively obtains the pre-order and periodic characteristics of the vector representation information in the topic network, and has excellent memory function. After being fused by the attention mechanism, the user feature vector representation and the topic vector representation after attention weighting can be obtained; this fusion representation is passed through the softmax layer, and the probability of the user being a long-tail user and an opinion field can be determined. In fact, the final binary classification result is obtained by using the softmax function; the output corresponding to each user node is a one-dimensional vector, and the one-dimensional vector has two values, 0 indicates that the user is a "long-tail user" and 1 indicates that the user is an "opinion leader"; this method can accurately identify the key users in the derived topic network.

[0098] In some preferred embodiments, this embodiment can also optimize the training of the user influence dynamic spatiotemporal model based on the attention mechanism through classification loss to achieve better recognition effect.

[0099] In one embodiment, see Figure 6 , a key user identification device in a derivative topic network, comprising:

[0100] The data acquisition module 301 is used to acquire data information in the topic network; the data information includes basic user information, basic topic information and user relationship network;

[0101] A data extraction module 302 is used to extract the original topic, derived topic user relationship network, the original topic, derived topic propagation network, user basic attributes and user intimacy from the data information in the topic network;

[0102] User feature extraction module 303, used to extract user feature vector representation through community tag propagation algorithm according to user basic attributes and intimacy between users;

[0103] A topic feature extraction module 304 is used to extract topic vector representations of original topics and derived topics through a joint distribution adaptive algorithm based on the original topic and derived topic propagation network;

[0104] The key user identification module 305 is used to input the user feature vector representation and the topic vector representation into the user influence dynamic spatiotemporal model based on the attention mechanism, and output an opinion leader user or a long-tail user.

[0105] In one embodiment, see Figure 7 , a key user identification system in a derived topic network, comprising:

[0106] The transceiver module 401 is used to obtain data information in the topic network; the data information includes basic user information, basic topic information and user relationship network;

[0107] Processing module 402, extracting the original topic, derived topic user relationship network, the original topic, derived topic propagation network, user basic attributes and user intimacy from the data information in the topic network; extracting the user feature vector representation through the community label propagation algorithm based on the user basic attributes and user intimacy; extracting the topic vector representation of the original topic and the derived topic through the joint distribution adaptive algorithm based on the original topic and derived topic propagation network; inputting the user feature vector representation and the topic vector representation into the user influence dynamic spatiotemporal model based on the attention mechanism, and outputting the opinion leader user or the long-tail user;

[0108] The display module 403 is used to display the opinion leader users or long-tail users output by the processing module.

[0109] A person skilled in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing the relevant hardware through a program, and the program can be stored in a computer-readable storage medium, which can include: ROM, RAM, disk or CD, etc.

[0110] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for identifying key users in a derived topic network, characterized in that: The method comprises: Acquire data information in the topic network; the data information includes basic user information, basic topic information and user relationship network; Extracting original topic and derived topic user relationship network, original topic and derived topic propagation network, user basic attributes and intimacy between users from the data information in the topic network; According to the basic attributes of users and the intimacy between users, the user feature vector representation is extracted through the community label propagation algorithm; According to the original topic and derived topic propagation network, the topic vector representation of the original topic and derived topic is extracted through the joint distribution adaptive algorithm; The user feature vector representation and the topic vector representation are input into a user influence dynamic spatiotemporal model based on an attention mechanism, and the output is an opinion leader user or a long-tail user.

2. The method for identifying key users in a derivative topic network according to claim 1, characterized in that: The intimacy between users is determined by the similarity scores and difference scores between users on the same attribute and the weighted scores between users on the same interaction mode.

3. The method for identifying key users in a derivative topic network according to claim 1, characterized in that: The method for extracting the user feature vector representation includes processing the intimacy between users and the frequency of label occurrence by using a propagation algorithm based on a speaker monitor, and calculating the acceptance probability of each label; according to the acceptance probability of each label, calculating the probability that the user belongs to the label; according to the probability that the user belongs to the label, obtaining the user label set, the number of connected communities and the size of the community; according to the user label set, the number of connected communities and the size of the community, generating the user feature vector representation through an embedding model.

4. The method for identifying key users in a derivative topic network according to claim 3, characterized in that: The acceptance probability of each tag is determined by the ratio of the product of the intimacy between users and the frequency of occurrence of the tag to the total number of tags.

5. The method for identifying key users in a derivative topic network according to claim 1, characterized in that: The method of extracting the topic vector representation of the original topic and the derived topic includes: Extract message content and user content through the original topic propagation network and the derived topic propagation network; Extract keywords from message content and user content to obtain feature vectors of topic content and user sentiment; The feature vector of topic content and the feature vector of user emotion are combined with the number of users participating in the original topic and derived topics, and the topic vector representation of the original topic and derived topics is obtained through a joint distribution adaptive algorithm.

6. The method for identifying key users in a derivative topic network according to claim 5, characterized in that: The method of extracting the topic vector representation of the original topic and the derived topic includes: The topic vector representation of the original topic is regarded as the source domain, and the topic vector representation of the derived topic is regarded as the target domain; By minimizing the square of the marginal distribution difference between the source domain and the target domain, as well as the square of the difference between each type of samples in the source domain and the target domain, the optimal topic vector representation of the original topic and the derived topic is obtained.

7. A key user identification device in a derivative topic network, characterized in that: include: Data acquisition module, used to obtain data information in the topic network; The data information includes basic user information, basic topic information and user relationship network; A data extraction module is used to extract the original topic, derived topic user relationship network, the original topic, derived topic propagation network, user basic attributes and user intimacy from the data information in the topic network; User feature extraction module, used to extract user feature vector representation based on user basic attributes and user intimacy through community label propagation algorithm; The topic feature extraction module is used to extract the topic vector representation of the original topic and the derived topic through the joint distribution adaptive algorithm based on the original topic and the derived topic propagation network; The key user identification module is used to input the user feature vector representation and the topic vector representation into the user influence dynamic spatiotemporal model based on the attention mechanism, and output an opinion leader user or a long-tail user.

8. A key user identification system in a derivative topic network, characterized in that: include: The transceiver module is used to obtain data information in the topic network; The data information includes basic user information, basic topic information and user relationship network; A processing module extracts the original topic and derived topic user relationship network, the original topic and derived topic propagation network, the basic attributes of users and the intimacy between users from the data information in the topic network; extracts the user feature vector representation through the community label propagation algorithm based on the basic attributes of users and the intimacy between users; extracts the topic vector representation of the original topic and the derived topic through the joint distribution adaptive algorithm based on the original topic and derived topic propagation network; inputs the user feature vector representation and the topic vector representation into the user influence dynamic spatiotemporal model based on the attention mechanism, and outputs the opinion leader user or the long-tail user; A display module is used to display the opinion leader users or long-tail users output by the processing module.

Images