CN103020116A - Method for automatically screening influential users on social media networks - Google Patents

Abstract

Translated fromChinese

本发明是一种在社交媒体网络上自动筛选有影响力用户的方法，包括步骤如下：步骤S1：利用超图模型为兴趣社交媒体网络中的用户、兴趣对象及其相互作用关系建模；步骤S2：采用超图约束的正则化主题概率模型，利用兴趣对象的内容信息和内容信息之间的相似性关系作为约束，自动学习得到隐含的兴趣主题；步骤S3：对每个用户和兴趣对象进行主题影响力排序，采用相似性传播模型及在超图上的用户和兴趣对象及相互之间的超边传播主题影响力，直到稳态，然后排序可得到特定主题下的有影响力的用户。本发明能够真实而准确地反映用户影响力在社交媒体网络中的分布。

The present invention is a method for automatically screening influential users on a social media network, comprising the following steps: Step S1: using a hypergraph model to model users, interest objects and their interaction relationships in the social media network of interest; S2: Using the regularized topic probability model constrained by the hypergraph, using the content information of the object of interest and the similarity relationship between the content information as constraints, automatically learn to obtain the hidden topic of interest; Step S3: For each user and object of interest Sorting the topic influence, using the similarity propagation model and users and interest objects on the hypergraph and the hyperedge propagation topic influence between each other, until the steady state, and then sorting can get influential users under a specific topic . The invention can truly and accurately reflect the distribution of user influence in social media networks.

Description

Translated fromChinese

在社交媒体网络上自动筛选有影响力用户的方法Method for Automatic Screening of Influential Users on Social Media Networks

技术领域technical field

本发明属于数字信息处理技术领域，具体涉及一种社交媒体网络的数据筛选技术，特别是基于多媒体内容与链接分析的主题敏感有影响力用户的筛选方法。The invention belongs to the technical field of digital information processing, and in particular relates to a social media network data screening technology, in particular to a screening method for theme-sensitive and influential users based on multimedia content and link analysis.

背景技术Background technique

社交媒体网络的出现和繁荣发展，改变了人们获取和消费信息的方式。各种社交媒体网络为人们提供了一个可以便捷创造和分享兴趣内容的平台。比如，新浪、腾讯微博的短讯图片分享，twitter的短讯，Flickr的图片分享等。然而，一个显著的存在问题是，人们在便捷获取信息的同时，也面临信息过载的问题。人们获取信息时，会倾向于获取自己感兴趣的内容和把有影响力的用户作为信息源。从社交媒体网络中筛选出在某一领域或主题下有影响力的用户或是兴趣对象，成为当前学术界和工业界关注的热点。通过筛选出主题敏感的用户，一种“兴趣达人”或“意见领袖”，从而商家可以进行影响力营销推广，用户可以更好地有目标性地获取所感兴趣所需要的知识信息。The advent and prosperity of social media networks has changed the way people obtain and consume information. Various social media networks provide a platform for people to easily create and share interesting content. For example, Sina and Tencent Weibo share text messages, twitter text messages, and Flickr picture sharing, etc. However, an obvious problem is that while people obtain information conveniently, they also face the problem of information overload. When people obtain information, they tend to obtain the content they are interested in and use influential users as information sources. Screening out influential users or objects of interest in a certain field or topic from social media networks has become a hot spot in academia and industry. By screening out users with sensitive topics, a kind of "interested people" or "opinion leaders", merchants can carry out influence marketing promotion, and users can better obtain the knowledge and information they are interested in in a targeted manner.

目前针对有影响力用户的筛选，现有的方法有：一种是专家发现方法，即给定一个主题，鉴别出有相关的技能或经验的人。现有的工作主要集中在文本数据上，没有涉及多媒体数据，即各种用户感兴趣的信息载体，比如音频、图片、视频等。另一种是社交媒体网络的影响力分析，即分析社交媒体网络并对社交媒体网络中的影响力进行建模，了解社交媒体网络的动态发展情况。现有主要工作是在社交网络中鉴别影响力的存在或者是在同质网络中量化影响力。At present, there are existing methods for screening influential users: one is the expert discovery method, which is to identify people with relevant skills or experience given a topic. Existing work mainly focuses on text data, and does not involve multimedia data, that is, various information carriers of interest to users, such as audio, pictures, videos, etc. The other is the influence analysis of social media networks, that is, analyzing social media networks and modeling the influence in social media networks to understand the dynamic development of social media networks. Existing major work is to identify the existence of influence in social networks or to quantify influence in homogeneous networks.

然而，上述方法不能完全真实准确地反映用户影响力在社交网络中的分布，用户影响力在社交网络中是一个连续性的可量化的变量，并且用户的影响力是主题敏感的，即在不同的主题上，用户的影响力分布是不同的。传统的方法，一方面大多方法局限于文本数据处理度量用户影响力，而实际上社交网络中包含丰富的多媒体数据，这些信息对用户影响力建模具有重要的作用。另一方面传统方法是对用户一般化的影响力建模，没有考虑主题敏感的影响力建模。However, the above methods cannot completely reflect the distribution of user influence in social networks completely, truly and accurately. User influence is a continuous and quantifiable variable in social networks, and user influence is subject-sensitive, that is, in different On the topic of , the influence distribution of users is different. Traditional methods, on the one hand, most methods are limited to text data processing to measure user influence, but in fact social networks contain rich multimedia data, and this information plays an important role in modeling user influence. On the other hand, the traditional method is to model the general influence of users, without considering the topic-sensitive influence modeling.

发明内容Contents of the invention

(一)要解决的技术问题(1) Technical problems to be solved

本发明所要解决的技术问题是如何自动地从社交媒体网络中关于特定的主题筛选出影响力的用户，并克服当前方法只在文本数据上为用户影响力建模的局限和仅限于用户全局影响力的度量。The technical problem to be solved by the present invention is how to automatically screen out influential users on specific topics from social media networks, and overcome the limitations of current methods that only model user influence on text data and are limited to the global influence of users. measure of force.

(二)技术方案(2) Technical solutions

为解决上述技术问题，本发明提出一种在社交媒体网络上自动筛选有影响力用户的方法，该方法包括步骤如下：步骤S1：利用超图模型为兴趣社交媒体网络中的用户、兴趣对象及其相互作用关系建模；步骤S2：采用超图约束的正则化主题概率模型，利用兴趣对象的内容信息和内容信息之间的相似性关系作为约束，自动学习得到隐含的兴趣主题；步骤S3：对每个用户和兴趣对象进行主题影响力排序，采用相似性传播模型及在超图上的用户和兴趣对象及相互之间的超边传播主题影响力，直到稳态，然后排序可得到特定主题下的有影响力的用户。In order to solve the above-mentioned technical problems, the present invention proposes a method for automatically screening influential users on a social media network. The method includes the following steps: Step S1: Use the hypergraph model to identify users, interest objects and users in the social media network of interest. Its interaction relationship modeling; Step S2: Using the regularized topic probability model constrained by the hypergraph, using the content information of the object of interest and the similarity relationship between the content information as constraints, automatically learn to obtain the hidden topic of interest; Step S3 : Sorting the topic influence of each user and object of interest, using the similarity propagation model and the user and object of interest on the hypergraph and the hyperedge propagation topic influence between them, until the steady state, and then sorting can get a specific Influential users under the topic.

(三)有益效果(3) Beneficial effects

本发明利用社交媒体网络中包含的各种媒体内容自动地发现潜在的主题，并分析相应的主题下的有影响力用户，能够在多模态异质网络中利用多媒体数据和各种社交链接关系挖掘出主题敏感的用户。并且，本发明能够真实而准确地反映用户影响力在社交媒体网络中的分布，筛选出社交媒体网络中主题敏感的有影响力用户。The present invention utilizes various media contents included in social media networks to automatically discover potential themes, and analyzes influential users under corresponding themes, and can utilize multimedia data and various social link relationships in multimodal heterogeneous networks Dig out users who are sensitive to the subject. Moreover, the present invention can truly and accurately reflect the distribution of user influence in the social media network, and screen out influential users with sensitive topics in the social media network.

附图说明Description of drawings

图1是本发明在社交媒体网络上自动筛选有影响力用户的方法的流程图；Fig. 1 is the flowchart of the method for automatically screening influential users on a social media network in the present invention;

图2是根据本发明的基于视觉内容构建的同质超边示意图；Fig. 2 is a schematic diagram of a homogeneous hyperedge constructed based on visual content according to the present invention;

图3是根据本发明的基于文本内容构建的同质超边示意图；Fig. 3 is a schematic diagram of a homogeneous hyperedge constructed based on text content according to the present invention;

图4是根据本发明的异质超边示意图；Fig. 4 is a schematic diagram of a heterogeneous hyperedge according to the present invention;

图5是本发明的超图中影响力消息传播示意图；Fig. 5 is a schematic diagram of influence message dissemination in the hypergraph of the present invention;

图6a和图6b是根据本发明的一个实施例的方法所得到的代表性用户和图片。Fig. 6a and Fig. 6b are representative users and pictures obtained by the method according to an embodiment of the present invention.

具体实施方式Detailed ways

为使本发明的目的、技术方案和优点更加清楚明白，以下结合具体实施例，并参照附图，对本发明作进一步的详细说明。In order to make the object, technical solution and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with specific embodiments and with reference to the accompanying drawings.

本发明的目标是筛选出社交媒体网络中主题敏感的有影响力用户。本发明中的社交媒体网络指的是指为用户提供的一个可以创造和分享媒体信息的平台，例如图片分享网站Flickr。本发明中所称的用户指的是社交媒体网络中的主体对象即人，所称的兴趣对象指的是由用户创造和分享的特定对象，如图片、视频、音乐。所谓主题指的是兴趣对象在语义层次上的聚合表达，具体表现形式为语义相近的词的概率分布。有影响力用户是指在社交网络中能够对其他用户的网络行为比如转发、评论等产生直接或间接影响的用户，所谓的影响力定义为当用户的情绪、意见或行为受到其他人的作用的一种表现形式。The goal of the present invention is to screen out influential users who are subject-sensitive in social media networks. The social media network in the present invention refers to a platform for users to create and share media information, such as the picture sharing website Flickr. The user referred to in the present invention refers to the main object in the social media network, that is, the person, and the referred object of interest refers to a specific object created and shared by the user, such as pictures, videos, and music. The so-called theme refers to the aggregated expression of the object of interest at the semantic level, and the specific form is the probability distribution of words with similar semantics. Influential users refer to users who can directly or indirectly influence other users' network behaviors such as forwarding and commenting in social networks. The so-called influence is defined as when the user's emotions, opinions or behaviors are influenced by other people a form of expression.

本发明的社交媒体网络指的是以兴趣对象为中心的，为用户提供的一个创造和分享兴趣对象的平台，兴趣对象可以是短讯、图片、视频、音乐等，在社交媒体网络中存在丰富的多媒体数据，多模态和异质的，比如在Flickr中，存在文本、图片、视频，并且在用户和图片之间存在评论、转发、喜爱等链接关系。The social media network of the present invention refers to a platform for users to create and share interest objects centered on interest objects. Interest objects can be short messages, pictures, videos, music, etc., and there are abundant Multi-modal and heterogeneous multimedia data, for example, in Flickr, there are text, pictures, and videos, and there are links such as comments, reposts, and favorites between users and pictures.

概括地说，本发明利用社交媒体网络中包含的各种媒体内容自动地发现潜在的主题，并分析相应的主题下的有影响力用户。本发明能够在多模态异质网络中利用多媒体数据和各种社交链接关系挖掘出主题敏感的用户。下面具体说明本发明的实施方式。In a nutshell, the present invention utilizes various media contents included in social media networks to automatically discover potential topics, and analyzes influential users under the corresponding topics. The invention can mine theme-sensitive users by using multimedia data and various social link relationships in a multi-modal heterogeneous network. Embodiments of the present invention will be specifically described below.

图1所示为本发明的有影响力用户筛选方法的流程图。如图1所示，本发明包括三个步骤：S1、超图构建(hypergraph construction)；S2、兴趣主题分布学习(Topic of interest distribution learning)；S3、主题敏感影响力排序(Topic sensitive influence ranking)。下面分别说明各个步骤。FIG. 1 is a flow chart of the screening method for influential users of the present invention. As shown in Figure 1, the present invention includes three steps: S1, hypergraph construction; S2, Topic of interest distribution learning; S3, Topic sensitive influence ranking . Each step is described below.

S1、超图构建S1, hypergraph construction

所谓超图指的是能够表示多阶关系的图。在超图中，包含节点和超边G＝(V，E，w)，其中节点表示不同类型的对象，而超边可以连接多于两个节点表示相互之间高阶关系。超图能够对包含高阶关系的对象网络进行建模。The so-called hypergraph refers to a graph that can represent multi-order relationships. In the hypergraph, there are nodes and hyperedges G=(V, E, w), where nodes represent different types of objects, and hyperedges can connect more than two nodes to represent high-order relationships between each other. Hypergraphs are capable of modeling object networks containing high-order relationships.

步骤S1是运用超图模型来为社交媒体网络中的用户、兴趣对象及其相互作用关系进行建模的步骤。在社交媒体网络中，用户和兴趣对像是最基本的元素，其间存在多种链接关系，比如用户可以评论、转发、喜爱和评论一个兴趣对象。Step S1 is a step of using a hypergraph model to model users, interest objects and their interaction relationships in a social media network. In a social media network, users and objects of interest are the most basic elements, and there are various links between them. For example, users can comment, forward, like and comment on an object of interest.

在本发明中，用超图节点表示社交媒体网络中的用户(user)和兴趣对象(object of interest，OI)；超边分为两种类型：同质(homogeneous)超边和异质(heterogeneous)超边。In the present invention, hypergraph nodes are used to represent users (user) and interest objects (object of interest, OI) in social media networks; hyperedges are divided into two types: homogeneous (homogeneous) hyperedges and heterogeneous (heterogeneous) hyperedges ) hyperedge.

同质超边用于表示兴趣对象之间的内容相似性，包括视觉内容相似性和文本内容相似性，异质超边用于表示用户和兴趣对象之间的高阶社交链接关系，如用户和兴趣对象之间存在的喜欢和评论关系。Homogeneous hyperedges are used to represent content similarity between objects of interest, including visual content similarity and textual content similarity, and heterogeneous hyperedges are used to represent high-order social link relationships between users and objects of interest, such as users and The like and comment relationships that exist between objects of interest.

图2为构建基于视觉内容相似性的同质超边的示意图，如图2所示，本发明采用K近邻的方法，即对于每一个兴趣对象，找到其K个最近邻的兴趣对象，然后用一条同质超边连接这些节点，并且权重设为1。Fig. 2 is a schematic diagram of constructing a homogeneous hyperedge based on visual content similarity. As shown in Fig. 2, the present invention adopts the method of K nearest neighbors, that is, for each object of interest, find its K nearest neighbors of interest, and then use A homogeneous hyperedge connects these nodes with a weight set to 1.

对于文本内容相似性，本发明构建基于文本标签的同质超边，图3为构建基于文本相似性的超边的示意图，如图3所示，首先从所有兴趣对象的文本元数据抽取一个词典，然后对于每一个词，为所有包含该词的兴趣对象建立一条超边，并且权重设为1。For the similarity of text content, the present invention builds a homogeneous hyperedge based on text tags. Figure 3 is a schematic diagram of building a hyperedge based on text similarity. As shown in Figure 3, a dictionary is first extracted from the text metadata of all objects of interest , and then for each word, establish a hyperedge for all interest objects containing the word, and set the weight to 1.

对于异质超边，本发明主要考虑两种：For heterogeneous hyperedges, the present invention mainly considers two types:

一种异质超边是“拥有者-多个兴趣对象-单一用户”(owner-OIs-user)的超边，其连接的是拥有者(用户A)和另一个用户B以及他们之间的交互的多个兴趣对象，用户B对用户A的多个兴趣对象表现出兴趣行为，比如评论或喜欢；该超边的权重为1。A heterogeneous hyperedge is a hyperedge of "owner-OIs-user" (owner-OIs-user), which connects the owner (user A) with another user B and the relationship between them Multiple interest objects of interaction, user B shows interest behaviors, such as comments or likes, on multiple interest objects of user A; the weight of this hyperedge is 1.

另一种异质超边是拥有者-单一兴趣对象-多个用户(owner-OI-users)的超边，其连接的是拥有者(用户A)和一个兴趣对象以及对该兴趣对象产生兴趣行为的多个用户。该超边的权重为1。Another heterogeneous hyperedge is the hyperedge of owner-single object of interest-multiple users (owner-OI-users), which connects the owner (user A) with an object of interest and is interested in the object of interest behavior of multiple users. The weight of this hyperedge is 1.

图4为上述两种异质超边的示意图，箭头表示用户和兴趣对象间的某种链接关系。Fig. 4 is a schematic diagram of the above two kinds of heterogeneous hyperedges, and the arrows indicate a certain link relationship between users and interest objects.

S2、兴趣主题分布学习S2. Interest topic distribution learning

在社交媒体网络中，每一个兴趣对象既包含有内容信息，也包含有上下文元数据信息，内容信息包括文本、音频、视频等信息，上下文元数据信息包括标签、时间、位置等信息。该步骤S2采用超图正则化主题概率模型，利用兴趣对象的内容信息和内容信息之间的相似性作为约束，自动学习得到隐含的兴趣主题。In a social media network, each interest object contains both content information and contextual metadata information. Content information includes text, audio, video and other information, and contextual metadata information includes label, time, location and other information. In step S2, the hypergraph regularized topic probability model is used, and the content information of the object of interest and the similarity between the content information are used as constraints to automatically learn the hidden topic of interest.

假设一个集合包含有N个兴趣对象O＝{o₁，o₂，…，o_N}，享有K个主题Z＝{z₁，…，z_K}，每一个兴趣对象表示成一个基于词袋的特征向量W＝{w₁，w₂，…，w_M}。把每一个兴趣对象看作一个文档，附带的文本中的单词作为词，共同享有的主题作为主题，我们采用概率潜在语义索引(PLSI)来对每一个兴趣对象的产生和共生词率来进行建模，产生过程如下：Suppose a collection contains N interest objects O={o₁ , o₂ ,…,o_N }, and has K topics Z={z₁ ,…,z_K }, each interest object is expressed as a bag-of-words-based The eigenvector of W={w₁ , w₂ , . . . , w_M }. Considering each object of interest as a document, the words in the accompanying text as words, and the shared topics as topics, we use probabilistic latent semantic indexing (PLSI) to build the generation and co-occurrence word rate of each object of interest. model, the generation process is as follows:

以概率P(o_i)选择一个兴趣对象o_i；Select an object of interest o_i with probability P(o_i );

以概率P(z_k|o_i)选择一个潜在的兴趣主题z_k；Select a potential topic of interest z_k with probability P(z_k |o_i );

以概率P(wj_|z_k)产生一个单词w_j。Generate a word w_j with probability P(wj_| z_k ).

一对兴趣对象和单词的观察概率如下计算：The observation probability for a pair of an object of interest and a word is computed as follows:

$\mathrm{<span><span>P</span>P</span>}<span><span>(</span>(</span>{\mathrm{<span><span>o</span>o</span>}}_{\mathrm{<span><span>i</span>i</span>}}<span><span>,</span>,</span>{\mathrm{<span><span>w</span>w</span>}}_{\mathrm{<span><span>j</span>j</span>}}<span><span>)</span>)</span><span><span>=</span>=</span>\mathrm{<span><span>P</span>P</span>}<span><span>(</span>(</span>{\mathrm{<span><span>o</span>o</span>}}_{\mathrm{<span><span>i</span>i</span>}}<span><span>)</span>)</span>\mathrm{<span><span>P</span>P</span>}<span><span>(</span>(</span>{\mathrm{<span><span>w</span>w</span>}}_{\mathrm{<span><span>j</span>j</span>}}<span><span>|</span>|</span>{\mathrm{<span><span>o</span>o</span>}}_{\mathrm{<span><span>i</span>i</span>}}<span><span>)</span>)</span><span><span>=</span>=</span>\mathrm{<span><span>P</span>P</span>}<span><span>(</span>(</span>{\mathrm{<span><span>o</span>o</span>}}_{\mathrm{<span><span>i</span>i</span>}}<span><span>)</span>)</span>\underset{\mathrm{<span><span>k</span>k</span>}<span><span>=</span>=</span>\mathrm{<span><span>1</span>1</span>}}{\overset{\mathrm{<span><span>K</span>K</span>}}{\mathrm{<span><span>\&Sigma;</span>\&Sigma;</span>}}}\mathrm{<span><span>P</span>P</span>}<span><span>(</span>(</span>{\mathrm{<span><span>w</span>w</span>}}_{\mathrm{<span><span>j</span>j</span>}}<span><span>|</span>|</span>{\mathrm{<span><span>z</span>z</span>}}_{\mathrm{<span><span>k</span>k</span>}}<span><span>)</span>)</span>\mathrm{<span><span>P</span>P</span>}<span><span>(</span>(</span>{\mathrm{<span><span>z</span>z</span>}}_{\mathrm{<span><span>k</span>k</span>}}<span><span>|</span>|</span>{\mathrm{<span><span>o</span>o</span>}}_{\mathrm{<span><span>i</span>i</span>}}<span><span>)</span>)</span><span><span>-</span>-</span><span><span>-</span>-</span><span><span>-</span>-</span><span><span>(</span>(</span>\mathrm{<span><span>1</span>1</span>}<span><span>)</span>)</span>$

包含参数有{P(w_j|z_k)，P(z_k|o_i)}，我们通过优化似然可以得到，Including parameters are {P(w_j |z_k ), P(z_k |o_i )}, we can get by optimizing the likelihood,

${\mathrm{<span><span>L</span>L</span>}}^{<span><span>\&prime;</span>\&prime;</span>}<span><span>=</span>=</span>\underset{\mathrm{<span><span>i</span>i</span>}<span><span>=</span>=</span>\mathrm{<span><span>1</span>1</span>}}{\overset{\mathrm{<span><span>N</span>N</span>}}{\mathrm{<span><span>\&Sigma;</span>\&Sigma;</span>}}}\underset{\mathrm{<span><span>j</span>j</span>}<span><span>=</span>=</span>\mathrm{<span><span>1</span>1</span>}}{\overset{\mathrm{<span><span>M</span>m</span>}}{\mathrm{<span><span>\&Sigma;</span>\&Sigma;</span>}}}\mathrm{<span><span>n</span>no</span>}<span><span>(</span>(</span>{\mathrm{<span><span>o</span>o</span>}}_{\mathrm{<span><span>i</span>i</span>}}<span><span>,</span>,</span>{\mathrm{<span><span>w</span>w</span>}}_{\mathrm{<span><span>j</span>j</span>}}<span><span>)</span>)</span>\mathrm{<span><span>log</span>log</span>}\underset{\mathrm{<span><span>k</span>k</span>}<span><span>=</span>=</span>\mathrm{<span><span>1</span>1</span>}}{\overset{\mathrm{<span><span>K</span>K</span>}}{\mathrm{<span><span>\&Sigma;</span>\&Sigma;</span>}}}\mathrm{<span><span>P</span>P</span>}<span><span>(</span>(</span>{\mathrm{<span><span>w</span>w</span>}}_{\mathrm{<span><span>j</span>j</span>}}<span><span>|</span>|</span>{\mathrm{<span><span>z</span>z</span>}}_{\mathrm{<span><span>k</span>k</span>}}<span><span>)</span>)</span>\mathrm{<span><span>P</span>P</span>}<span><span>(</span>(</span>{\mathrm{<span><span>z</span>z</span>}}_{\mathrm{<span><span>k</span>k</span>}}<span><span>|</span>|</span>{\mathrm{<span><span>o</span>o</span>}}_{\mathrm{<span><span>i</span>i</span>}}<span><span>)</span>)</span><span><span>-</span>-</span><span><span>-</span>-</span><span><span>-</span>-</span><span><span>(</span>(</span>\mathrm{<span><span>2</span>2</span>}<span><span>)</span>)</span>$

为了使学习出的兴趣对象的主题分布保持局部相似性，我们加入兴趣对象的内容，包括文本、视觉特征以及二者之间的内容一致性作为约束项，最终我们得到超图正则化的主题概率模型。通过最大化下面似然目标式得到：In order to maintain the local similarity of the topic distribution of the learned object of interest, we add the content of the object of interest, including text, visual features, and the content consistency between the two as constraints, and finally we get the topic probability of hypergraph regularization Model. It is obtained by maximizing the following likelihood objective:

$\mathrm{<span><span>L</span>L</span>}<span><span>=</span>=</span>{\mathrm{<span><span>L</span>L</span>}}^{<span><span>\&prime;</span>\&prime;</span>}<span><span>-</span>-</span>\mathrm{<span><span>\&lambda;R</span>\&lambda;R</span>}$

$<span><span>=</span>=</span>\underset{\mathrm{<span><span>i</span>i</span>}<span><span>=</span>=</span>\mathrm{<span><span>1</span>1</span>}}{\overset{\mathrm{<span><span>N</span>N</span>}}{\mathrm{<span><span>\&Sigma;</span>\&Sigma;</span>}}}\underset{\mathrm{<span><span>j</span>j</span>}<span><span>=</span>=</span>\mathrm{<span><span>1</span>1</span>}}{\overset{\mathrm{<span><span>M</span>m</span>}}{\mathrm{<span><span>\&Sigma;</span>\&Sigma;</span>}}}\mathrm{<span><span>n</span>no</span>}<span><span>(</span>(</span>{\mathrm{<span><span>o</span>o</span>}}_{\mathrm{<span><span>i</span>i</span>}}<span><span>,</span>,</span>{\mathrm{<span><span>w</span>w</span>}}_{\mathrm{<span><span>j</span>j</span>}}<span><span>)</span>)</span>\mathrm{<span><span>log</span>log</span>}\underset{\mathrm{<span><span>k</span>k</span>}<span><span>=</span>=</span>\mathrm{<span><span>1</span>1</span>}}{\overset{\mathrm{<span><span>K</span>K</span>}}{\mathrm{<span><span>\&Sigma;</span>\&Sigma;</span>}}}\mathrm{<span><span>P</span>P</span>}<span><span>(</span>(</span>{\mathrm{<span><span>w</span>w</span>}}_{\mathrm{<span><span>j</span>j</span>}}<span><span>|</span>|</span>{\mathrm{<span><span>z</span>z</span>}}_{\mathrm{<span><span>k</span>k</span>}}<span><span>)</span>)</span>\mathrm{<span><span>P</span>P</span>}<span><span>(</span>(</span>{\mathrm{<span><span>z</span>z</span>}}_{\mathrm{<span><span>k</span>k</span>}}<span><span>|</span>|</span>{\mathrm{<span><span>o</span>o</span>}}_{\mathrm{<span><span>i</span>i</span>}}<span><span>)</span>)</span><span><span>-</span>-</span>\mathrm{<span><span>\&lambda;</span>\&lambda;</span>}\underset{\mathrm{<span><span>k</span>k</span>}<span><span>=</span>=</span>\mathrm{<span><span>1</span>1</span>}}{\overset{\mathrm{<span><span>K</span>K</span>}}{\mathrm{<span><span>\&Sigma;</span>\&Sigma;</span>}}}{\mathrm{<span><span>f</span>f</span>}}_{\mathrm{<span><span>k</span>k</span>}}^{\mathrm{<span><span>T</span>T</span>}}{\mathrm{<span><span>Lf</span>Lf</span>}}_{\mathrm{<span><span>k</span>k</span>}}<span><span>-</span>-</span><span><span>-</span>-</span><span><span>-</span>-</span><span><span>(</span>(</span>\mathrm{<span><span>3</span>3</span>}<span><span>)</span>)</span>$

其中，L是超图拉普拉斯矩阵，如下表达Among them, L is the hypergraph Laplacian matrix, expressed as follows

$\mathrm{<span><span>R</span>R</span>}<span><span>=</span>=</span>\frac{\mathrm{<span><span>1</span>1</span>}}{\mathrm{<span><span>2</span>2</span>}}\underset{\mathrm{<span><span>k</span>k</span>}<span><span>=</span>=</span>\mathrm{<span><span>1</span>1</span>}}{\overset{\mathrm{<span><span>K</span>K</span>}}{\mathrm{<span><span>\&Sigma;</span>\&Sigma;</span>}}}\underset{\mathrm{<span><span>e</span>e</span>}<span><span>\&Element;</span>\&Element;</span>{\mathrm{<span><span>E</span>E.</span>}}_{\mathrm{<span><span>o</span>o</span>}}}{\mathrm{<span><span>\&Sigma;</span>\&Sigma;</span>}}\underset{{\mathrm{<span><span>o</span>o</span>}}_{\mathrm{<span><span>i</span>i</span>}}<span><span>,</span>,</span>{\mathrm{<span><span>o</span>o</span>}}_{\mathrm{<span><span>j</span>j</span>}}<span><span>\&Element;</span>\&Element;</span>{\mathrm{<span><span>V</span>V</span>}}_{\mathrm{<span><span>o</span>o</span>}}}{\mathrm{<span><span>\&Sigma;</span>\&Sigma;</span>}}\frac{\mathrm{<span><span>w</span>w</span>}<span><span>(</span>(</span>\mathrm{<span><span>e</span>e</span>}<span><span>)</span>)</span>\mathrm{<span><span>h</span>h</span>}<span><span>(</span>(</span>{\mathrm{<span><span>o</span>o</span>}}_{\mathrm{<span><span>i</span>i</span>}}<span><span>,</span>,</span>\mathrm{<span><span>e</span>e</span>}<span><span>)</span>)</span>\mathrm{<span><span>h</span>h</span>}<span><span>(</span>(</span>{\mathrm{<span><span>o</span>o</span>}}_{\mathrm{<span><span>i</span>i</span>}}<span><span>,</span>,</span>\mathrm{<span><span>e</span>e</span>}<span><span>)</span>)</span>}{\mathrm{<span><span>\&delta;</span>\&delta;</span>}<span><span>(</span>(</span>\mathrm{<span><span>e</span>e</span>}<span><span>)</span>)</span>}<span><span>(</span>(</span>\frac{\mathrm{<span><span>P</span>P</span>}<span><span>(</span>(</span>{\mathrm{<span><span>z</span>z</span>}}_{\mathrm{<span><span>k</span>k</span>}}<span><span>|</span>|</span>{\mathrm{<span><span>o</span>o</span>}}_{\mathrm{<span><span>i</span>i</span>}}<span><span>)</span>)</span>}{\sqrt{\mathrm{<span><span>d</span>d</span>}<span><span>(</span>(</span>{\mathrm{<span><span>o</span>o</span>}}_{\mathrm{<span><span>i</span>i</span>}}<span><span>)</span>)</span>}}<span><span>-</span>-</span>\frac{\mathrm{<span><span>P</span>P</span>}<span><span>(</span>(</span>{\mathrm{<span><span>z</span>z</span>}}_{\mathrm{<span><span>k</span>k</span>}}<span><span>|</span>|</span>{\mathrm{<span><span>o</span>o</span>}}_{\mathrm{<span><span>j</span>j</span>}}<span><span>)</span>)</span>}{\sqrt{\mathrm{<span><span>d</span>d</span>}<span><span>(</span>(</span>{\mathrm{<span><span>o</span>o</span>}}_{\mathrm{<span><span>j</span>j</span>}}<span><span>)</span>)</span>}}<span><span>)</span>)</span>$

$<span><span>=</span>=</span>\underset{\mathrm{<span><span>k</span>k</span>}<span><span>=</span>=</span>\mathrm{<span><span>1</span>1</span>}}{\overset{\mathrm{<span><span>K</span>K</span>}}{\mathrm{<span><span>\&Sigma;</span>\&Sigma;</span>}}}{\mathrm{<span><span>f</span>f</span>}}_{\mathrm{<span><span>l</span>l</span>}}^{\mathrm{<span><span>T</span>T</span>}}<span><span>(</span>(</span>\mathrm{<span><span>I</span>I</span>}<span><span>-</span>-</span>{\mathrm{<span><span>D</span>D.</span>}}_{\mathrm{<span><span>v</span>v</span>}}^{<span><span>-</span>-</span>\frac{\mathrm{<span><span>1</span>1</span>}}{\mathrm{<span><span>2</span>2</span>}}}{\mathrm{<span><span>HWD</span>HWD</span>}}_{\mathrm{<span><span>e</span>e</span>}}^{<span><span>-</span>-</span>\mathrm{<span><span>1</span>1</span>}}{\mathrm{<span><span>H</span>h</span>}}^{\mathrm{<span><span>T</span>T</span>}}{\mathrm{<span><span>D</span>D.</span>}}_{\mathrm{<span><span>v</span>v</span>}}^{<span><span>-</span>-</span>\frac{\mathrm{<span><span>1</span>1</span>}}{\mathrm{<span><span>2</span>2</span>}}}<span><span>)</span>)</span>{\mathrm{<span><span>f</span>f</span>}}_{\mathrm{<span><span>k</span>k</span>}}<span><span>-</span>-</span><span><span>-</span>-</span><span><span>-</span>-</span><span><span>(</span>(</span>\mathrm{<span><span>4</span>4</span>}<span><span>)</span>)</span>$

$<span><span>=</span>=</span>\underset{\mathrm{<span><span>k</span>k</span>}<span><span>=</span>=</span>\mathrm{<span><span>1</span>1</span>}}{\overset{\mathrm{<span><span>K</span>K</span>}}{\mathrm{<span><span>\&Sigma;</span>\&Sigma;</span>}}}{\mathrm{<span><span>f</span>f</span>}}_{\mathrm{<span><span>k</span>k</span>}}^{\mathrm{<span><span>T</span>T</span>}}\mathrm{<span><span>Lf</span>Lf</span>}$

我们采用泛化期望最大化算法来优化式(3)得到主题概率模型参数{P(w_j|z_k)，P(z_k|o_i)}。泛化期望最大化(generalized EM)算法包含期望步骤和最大化步骤迭代进行。在期望步骤内，基于当前的参数估计计算隐变量的后验概率；在最大化步骤内，优化似然目标式，更新参数得到一个更好的解。通过泛化最大化算法，我们得到每一个兴趣对象o_i的主题分布P(z_k|o_i)以及每一个主题z_k的词概率分布P(w_j|z_k)。在得到兴趣对象的主题分布后，我们通过聚合每个用户的兴趣对象得到每个用户的主题分布如下，We use the generalized expectation maximization algorithm to optimize formula (3) to obtain the topic probability model parameters {P(w_j |z_k ), P(z_k |o_i )}. The generalized expectation maximization (generalized EM) algorithm includes the expectation step and the maximization step iteratively. In the expectation step, the posterior probability of the hidden variable is calculated based on the current parameter estimation; in the maximization step, the likelihood objective formula is optimized, and the parameters are updated to obtain a better solution. Through the generalization maximization algorithm, we get the topic distribution P(z_k |o_i ) of each object of interest o_i and the word probability distribution P(w_j |z_k ) of each topic z_k . After obtaining the topic distribution of interest objects, we obtain the topic distribution of each user by aggregating the interest objects of each user as follows,

$\mathrm{<span><span>P</span>P</span>}<span><span>(</span>(</span>{\mathrm{<span><span>z</span>z</span>}}_{\mathrm{<span><span>k</span>k</span>}}<span><span>|</span>|</span>\mathrm{<span><span>u</span>u</span>}<span><span>)</span>)</span><span><span>=</span>=</span>\underset{{\mathrm{<span><span>o</span>o</span>}}_{\mathrm{<span><span>i</span>i</span>}}<span><span>\&Element;</span>\&Element;</span>{\mathrm{<span><span>O</span>o</span>}}_{\mathrm{<span><span>u</span>u</span>}}}{\mathrm{<span><span>\&Sigma;</span>\&Sigma;</span>}}\mathrm{<span><span>P</span>P</span>}<span><span>(</span>(</span>{\mathrm{<span><span>z</span>z</span>}}_{\mathrm{<span><span>k</span>k</span>}}<span><span>|</span>|</span>{\mathrm{<span><span>o</span>o</span>}}_{\mathrm{<span><span>i</span>i</span>}}<span><span>)</span>)</span>\mathrm{<span><span>P</span>P</span>}<span><span>(</span>(</span>{\mathrm{<span><span>o</span>o</span>}}_{\mathrm{<span><span>i</span>i</span>}}<span><span>|</span>|</span>\mathrm{<span><span>u</span>u</span>}<span><span>)</span>)</span><span><span>=</span>=</span>\underset{{\mathrm{<span><span>o</span>o</span>}}_{\mathrm{<span><span>i</span>i</span>}}<span><span>\&Element;</span>\&Element;</span>{\mathrm{<span><span>O</span>o</span>}}_{\mathrm{<span><span>u</span>u</span>}}}{\mathrm{<span><span>\&Sigma;</span>\&Sigma;</span>}}\frac{\mathrm{<span><span>P</span>P</span>}<span><span>(</span>(</span>{\mathrm{<span><span>z</span>z</span>}}_{\mathrm{<span><span>k</span>k</span>}}<span><span>|</span>|</span>{\mathrm{<span><span>o</span>o</span>}}_{\mathrm{<span><span>i</span>i</span>}}<span><span>)</span>)</span>}{<span><span>|</span>|</span>{\mathrm{<span><span>O</span>o</span>}}_{\mathrm{<span><span>u</span>u</span>}}<span><span>|</span>|</span>}<span><span>-</span>-</span><span><span>-</span>-</span><span><span>-</span>-</span><span><span>(</span>(</span>\mathrm{<span><span>5</span>5</span>}<span><span>)</span>)</span>$

S3、主题敏感影响力排序S3, Topic Sensitive Influence Sorting

在得到每一个用户和兴趣对象的主题分布后，该步骤S3进行主题影响力排序，采用相似性传播模型在超图上的用户和兴趣对象及相互之间的超边传播影响力，直到稳态，然后排序可得到特定主题下的有影响力的用户。After obtaining the topic distribution of each user and object of interest, this step S3 sorts the topic influence, and uses the similarity propagation model to spread the influence on the users and objects of interest on the hypergraph and the hyperedge between them until the steady state , and then sort to get influential users under a specific topic.

图5是影响力消息传播示意图。如图5所示，主题影响力在超图的用户结点和兴趣对象结点，和之间的超边传播并直到收敛。Fig. 5 is a schematic diagram of influence message dissemination. As shown in Figure 5, the topic influence propagates in the user nodes and interest object nodes of the hypergraph, and the hyperedges between them until convergence.

如前所述，异质超边分为两种，一种是拥有者-多个兴趣对象-单一用户(owner-OIs-user)的超边，另一种是拥有者-单一兴趣对象-多个用户。As mentioned earlier, there are two types of heterogeneous hyperedges, one is owner-OIs-user hyperedge, and the other is owner-OIs-multiple users.

首先，在拥有者-多个兴趣对象-单一用户的超边学习用户的主题敏感影响力。以

表示用户u_i在主题k上的影响力得分，

表示兴趣对象o_p在主题k上的影响力得分。首先计算两个用户u_i，u_j的主题相似度f^k(i，j)如下：First, the user's topic-sensitive influence is learned on the hyperedge of owner-multiple interest objects-single user. by

Indicates the influence score of user u_i on topic k,

Indicates the influence score of object of interest o_p on topic k. First calculate the topic similarity f^k (i, j) of two users u_i , u_j as follows:

${\mathrm{<span><span>f</span>f</span>}}^{\mathrm{<span><span>k</span>k</span>}}<span><span>(</span>(</span>\mathrm{<span><span>i</span>i</span>}<span><span>,</span>,</span>\mathrm{<span><span>j</span>j</span>}<span><span>)</span>)</span><span><span>=</span>=</span>\mathrm{<span><span>log</span>log</span>}\frac{\mathrm{<span><span>g</span>g</span>}<span><span>(</span>(</span>{\mathrm{<span><span>u</span>u</span>}}_{\mathrm{<span><span>i</span>i</span>}}<span><span>,</span>,</span>{\mathrm{<span><span>u</span>u</span>}}_{\mathrm{<span><span>j</span>j</span>}}<span><span>,</span>,</span>\mathrm{<span><span>k</span>k</span>}<span><span>)</span>)</span>}{\underset{\mathrm{<span><span>z</span>z</span>}<span><span>\&Element;</span>\&Element;</span>\mathrm{<span><span>S</span>S</span>}}{\mathrm{<span><span>\&Sigma;</span>\&Sigma;</span>}}\mathrm{<span><span>g</span>g</span>}<span><span>(</span>(</span>{\mathrm{<span><span>u</span>u</span>}}_{\mathrm{<span><span>i</span>i</span>}}<span><span>,</span>,</span>{\mathrm{<span><span>u</span>u</span>}}_{\mathrm{<span><span>z</span>z</span>}}<span><span>,</span>,</span>\mathrm{<span><span>k</span>k</span>}<span><span>)</span>)</span>}<span><span>,</span>,</span>\mathrm{<span><span>g</span>g</span>}<span><span>(</span>(</span>{\mathrm{<span><span>u</span>u</span>}}_{\mathrm{<span><span>i</span>i</span>}}<span><span>,</span>,</span>{\mathrm{<span><span>u</span>u</span>}}_{\mathrm{<span><span>j</span>j</span>}}<span><span>,</span>,</span>\mathrm{<span><span>k</span>k</span>}<span><span>)</span>)</span><span><span>=</span>=</span>\underset{{\mathrm{<span><span>o</span>o</span>}}_{\mathrm{<span><span>p</span>p</span>}}<span><span>\&Element;</span>\&Element;</span>{\mathrm{<span><span>O</span>o</span>}}_{{\mathrm{<span><span>u</span>u</span>}}_{\mathrm{<span><span>j</span>j</span>}}}}{\mathrm{<span><span>\&Sigma;</span>\&Sigma;</span>}}{\mathrm{<span><span>s</span>the\; s</span>}}^{\mathrm{<span><span>k</span>k</span>}}<span><span>(</span>(</span>{\mathrm{<span><span>o</span>o</span>}}_{\mathrm{<span><span>p</span>p</span>}}<span><span>)</span>)</span><span><span>-</span>-</span><span><span>-</span>-</span><span><span>-</span>-</span><span><span>(</span>(</span>\mathrm{<span><span>6</span>6</span>}<span><span>)</span>)</span>$

主题相似度f^k(i，j)用于在相似性传播中计算用户的影响力

引入两组变量

和

表示影响力消息，r^k(i，j)由用户结点u_i发至u_j，表示从用户u_i的角度看，他同意用户u_j在主题k上影响他的程度；a^k(i，j)由用户结点u_j发至u_i，表示从用户u_j的角度看，他认为他对用户u_i在主题k上影响的程度。Topic similarity f^k (i, j) is used to calculate user influence in similarity propagation

Introduce two sets of variables

and

Indicates the influence message, r^k (i, j) is sent by user node u_i to u_j , indicating that from the perspective of user u_i , he agrees with the extent to which user u_j influences him on topic k; a^k (i , j) is sent from user node u_j to u_i , indicating from the perspective of user u_j , he thinks the extent of his influence on user u_i on topic k.

通过在拥有者-多个兴趣对象-单一用户的超边传递消息，更新用户影响力得分如下：By passing messages on the hyperedge of the owner-multiple interest objects-single user, the user influence score is updated as follows:

${\mathrm{<span><span>r</span>r</span>}}^{\mathrm{<span><span>k</span>k</span>}}<span><span>(</span>(</span>\mathrm{<span><span>i</span>i</span>}<span><span>,</span>,</span>\mathrm{<span><span>j</span>j</span>}<span><span>)</span>)</span><span><span>\&LeftArrow;</span>\&LeftArrow;</span>{\mathrm{<span><span>f</span>f</span>}}^{\mathrm{<span><span>k</span>k</span>}}<span><span>(</span>(</span>\mathrm{<span><span>i</span>i</span>}<span><span>,</span>,</span>\mathrm{<span><span>j</span>j</span>}<span><span>)</span>)</span><span><span>-</span>-</span>\underset{\mathrm{<span><span>z</span>z</span>}<span><span>\&Element;</span>\&Element;</span>\mathrm{<span><span>S</span>S</span>}<span><span>(</span>(</span>\mathrm{<span><span>i</span>i</span>}<span><span>)</span>)</span>}{\mathrm{<span><span>max</span>max</span>}}<span><span>\{</span>\{</span>{\mathrm{<span><span>f</span>f</span>}}^{\mathrm{<span><span>k</span>k</span>}}<span><span>(</span>(</span>\mathrm{<span><span>i</span>i</span>}<span><span>,</span>,</span>\mathrm{<span><span>z</span>z</span>}<span><span>)</span>)</span><span><span>+</span>+</span>{\mathrm{<span><span>a</span>a</span>}}^{\mathrm{<span><span>k</span>k</span>}}<span><span>(</span>(</span>\mathrm{<span><span>i</span>i</span>}<span><span>,</span>,</span>\mathrm{<span><span>z</span>z</span>}<span><span>)</span>)</span><span><span>\}</span>\}</span>$

${\mathrm{<span><span>a</span>a</span>}}^{\mathrm{<span><span>k</span>k</span>}}<span><span>(</span>(</span>\mathrm{<span><span>i</span>i</span>}<span><span>,</span>,</span>\mathrm{<span><span>j</span>j</span>}<span><span>)</span>)</span><span><span>\&LeftArrow;</span>\&LeftArrow;</span>\mathrm{<span><span>min</span>min</span>}<span><span>\{</span>\{</span>\mathrm{<span><span>0</span>0</span>}<span><span>,</span>,</span>{\mathrm{<span><span>r</span>r</span>}}^{\mathrm{<span><span>k</span>k</span>}}<span><span>(</span>(</span>\mathrm{<span><span>j</span>j</span>}<span><span>,</span>,</span>\mathrm{<span><span>j</span>j</span>}<span><span>)</span>)</span><span><span>+</span>+</span>\underset{\mathrm{<span><span>z</span>z</span>}<span><span>\&NotElement;</span>\&NotElement;</span><span><span>\{</span>\{</span>\mathrm{<span><span>i</span>i</span>}<span><span>,</span>,</span>\mathrm{<span><span>j</span>j</span>}<span><span>\}</span>\}</span>}{\mathrm{<span><span>\&Sigma;</span>\&Sigma;</span>}}\mathrm{<span><span>max</span>max</span>}<span><span>\{</span>\{</span>\mathrm{<span><span>0</span>0</span>}<span><span>,</span>,</span>{\mathrm{<span><span>r</span>r</span>}}^{\mathrm{<span><span>k</span>k</span>}}<span><span>(</span>(</span>\mathrm{<span><span>z</span>z</span>}<span><span>,</span>,</span>\mathrm{<span><span>j</span>j</span>}<span><span>)</span>)</span><span><span>\}</span>\}</span><span><span>\}</span>\}</span><span><span>-</span>-</span><span><span>-</span>-</span><span><span>-</span>-</span><span><span>(</span>(</span>\mathrm{<span><span>7</span>7</span>}<span><span>)</span>)</span>$

${\mathrm{<span><span>a</span>a</span>}}^{\mathrm{<span><span>k</span>k</span>}}<span><span>(</span>(</span>\mathrm{<span><span>j</span>j</span>}<span><span>,</span>,</span>\mathrm{<span><span>j</span>j</span>}<span><span>)</span>)</span><span><span>\&LeftArrow;</span>\&LeftArrow;</span>\underset{{\mathrm{<span><span>i</span>i</span>}}^{<span><span>\&prime;</span>\&prime;</span>}<span><span>\&NotEqual;</span>\&NotEqual;</span>\mathrm{<span><span>j</span>j</span>}}{\mathrm{<span><span>\&Sigma;</span>\&Sigma;</span>}}\mathrm{<span><span>max</span>max</span>}<span><span>\{</span>\{</span>\mathrm{<span><span>0</span>0</span>}<span><span>,</span>,</span>{\mathrm{<span><span>r</span>r</span>}}^{\mathrm{<span><span>k</span>k</span>}}<span><span>(</span>(</span>{\mathrm{<span><span>i</span>i</span>}}^{<span><span>\&prime;</span>\&prime;</span>}<span><span>,</span>,</span>\mathrm{<span><span>j</span>j</span>}<span><span>)</span>)</span><span><span>\}</span>\}</span>$

其中S(i)表示用户u_i所感兴趣的兴趣对象的拥有用户的集合。通过上面相似传播算法我们得到稳态的

和

基于此，我们定义用户之间的影响力为：用q^k(i，j)表示用户u_j对用户u_i在主题k上的影响力，计算如下：Where S(i) represents the set of users who own the object of interest that user u_i is interested in. Through the above similarity propagation algorithm we get the steady-state

and

Based on this, we define the influence between users as: use q^k (i, j) to represent the influence of user u_j on user u_i on topic k, calculated as follows:

${\mathrm{<span><span>q</span>q</span>}}^{\mathrm{<span><span>k</span>k</span>}}<span><span>(</span>(</span>\mathrm{<span><span>i</span>i</span>}<span><span>,</span>,</span>\mathrm{<span><span>j</span>j</span>}<span><span>)</span>)</span><span><span>=</span>=</span>\frac{\mathrm{<span><span>1</span>1</span>}}{\mathrm{<span><span>1</span>1</span>}<span><span>+</span>+</span>{\mathrm{<span><span>e</span>e</span>}}^{<span><span>-</span>-</span><span><span>(</span>(</span>{\mathrm{<span><span>r</span>r</span>}}^{\mathrm{<span><span>k</span>k</span>}}<span><span>(</span>(</span>\mathrm{<span><span>i</span>i</span>}<span><span>,</span>,</span>\mathrm{<span><span>j</span>j</span>}<span><span>)</span>)</span><span><span>+</span>+</span>{\mathrm{<span><span>a</span>a</span>}}^{\mathrm{<span><span>k</span>k</span>}}<span><span>(</span>(</span>\mathrm{<span><span>i</span>i</span>}<span><span>,</span>,</span>\mathrm{<span><span>j</span>j</span>}<span><span>)</span>)</span><span><span>)</span>)</span>}}<span><span>-</span>-</span><span><span>-</span>-</span><span><span>-</span>-</span><span><span>(</span>(</span>\mathrm{<span><span>8</span>8</span>}<span><span>)</span>)</span>$

然后我们计算用户在主题的全局影响力：Then we calculate the user's global influence on the topic:

${\mathrm{<span><span>s</span>the\; s</span>}}^{\mathrm{<span><span>k</span>k</span>}}<span><span>(</span>(</span>\mathrm{<span><span>j</span>j</span>}<span><span>)</span>)</span><span><span>=</span>=</span>\mathrm{<span><span>\&eta;</span>\&eta;</span>}\underset{\mathrm{<span><span>i</span>i</span>}<span><span>:</span>:</span>\mathrm{<span><span>i</span>i</span>}<span><span>\&RightArrow;</span>\&Right\; Arrow;</span>\mathrm{<span><span>j</span>j</span>}}{\mathrm{<span><span>\&Sigma;</span>\&Sigma;</span>}}{\mathrm{<span><span>s</span>the\; s</span>}}^{\mathrm{<span><span>k</span>k</span>}}<span><span>(</span>(</span>\mathrm{<span><span>i</span>i</span>}<span><span>)</span>)</span>{\mathrm{<span><span>p</span>p</span>}}^{\mathrm{<span><span>k</span>k</span>}}<span><span>(</span>(</span>\mathrm{<span><span>j</span>j</span>}<span><span>|</span>|</span>\mathrm{<span><span>i</span>i</span>}<span><span>)</span>)</span><span><span>+</span>+</span><span><span>(</span>(</span>\mathrm{<span><span>1</span>1</span>}<span><span>-</span>-</span>\mathrm{<span><span>\&eta;</span>\&eta;</span>}<span><span>)</span>)</span>{\mathrm{<span><span>v</span>v</span>}}_{\mathrm{<span><span>j</span>j</span>}}^{\mathrm{<span><span>k</span>k</span>}}<span><span>-</span>-</span><span><span>-</span>-</span><span><span>-</span>-</span><span><span>(</span>(</span>\mathrm{<span><span>9</span>9</span>}<span><span>)</span>)</span>$

η是一个控制参数。p^k(j|i)表示用户u_i到用户u_j的转移概率如下计算：η is a control parameter. p^k (j|i) means the transition probability from user u_i to user u_j is calculated as follows:

${\mathrm{<span><span>p</span>p</span>}}^{\mathrm{<span><span>k</span>k</span>}}<span><span>(</span>(</span>\mathrm{<span><span>j</span>j</span>}<span><span>|</span>|</span>\mathrm{<span><span>i</span>i</span>}<span><span>)</span>)</span><span><span>=</span>=</span>\frac{{\mathrm{<span><span>q</span>q</span>}}^{\mathrm{<span><span>k</span>k</span>}}<span><span>(</span>(</span>\mathrm{<span><span>i</span>i</span>}<span><span>,</span>,</span>\mathrm{<span><span>j</span>j</span>}<span><span>)</span>)</span>}{\underset{{\mathrm{<span><span>j</span>j</span>}}^{<span><span>\&prime;</span>\&prime;</span>}<span><span>:</span>:</span>\mathrm{<span><span>i</span>i</span>}<span><span>\&RightArrow;</span>\&Right\; Arrow;</span>{\mathrm{<span><span>j</span>j</span>}}^{<span><span>\&prime;</span>\&prime;</span>}}{\mathrm{<span><span>\&Sigma;</span>\&Sigma;</span>}}{\mathrm{<span><span>q</span>q</span>}}^{\mathrm{<span><span>k</span>k</span>}}<span><span>(</span>(</span>\mathrm{<span><span>i</span>i</span>}<span><span>,</span>,</span>{\mathrm{<span><span>j</span>j</span>}}^{<span><span>\&prime;</span>\&prime;</span>}<span><span>)</span>)</span>}<span><span>-</span>-</span><span><span>-</span>-</span><span><span>-</span>-</span><span><span>(</span>(</span>\mathrm{<span><span>10</span>10</span>}<span><span>)</span>)</span>$

是用户u_j的初始化在主题k的影响力，

对式(9)求解得到稳态的用户影响力为

is the initial influence of user u_j on topic k,

Solving equation (9), the steady-state user influence is obtained as

${\mathrm{<span><span>s</span>the\; s</span>}}_{\mathrm{<span><span>u</span>u</span>}}^{\mathrm{<span><span>k</span>k</span>}}<span><span>=</span>=</span><span><span>(</span>(</span>\mathrm{<span><span>1</span>1</span>}<span><span>-</span>-</span>\mathrm{<span><span>\&eta;</span>\&eta;</span>}<span><span>)</span>)</span>{<span><span>(</span>(</span>\mathrm{<span><span>I</span>I</span>}<span><span>-</span>-</span>{\mathrm{<span><span>\&eta;Q</span>\&eta;Q</span>}}_{\mathrm{<span><span>k</span>k</span>}}<span><span>)</span>)</span>}^{<span><span>-</span>-</span>\mathrm{<span><span>1</span>1</span>}}{\mathrm{<span><span>v</span>v</span>}}^{\mathrm{<span><span>k</span>k</span>}}<span><span>-</span>-</span><span><span>-</span>-</span><span><span>-</span>-</span><span><span>(</span>(</span>\mathrm{<span><span>11</span>11</span>}<span><span>)</span>)</span>$

对于基于用户影响力和拥有者-单一兴趣对象-多个用户的超边，我们计算每一个兴趣对象的影响力如下：For hyperedges based on user influence and owner-single interest object-multiple users, we calculate the influence of each interest object as follows:

${\mathrm{<span><span>s</span>the\; s</span>}}_{\mathrm{<span><span>o</span>o</span>}}^{\mathrm{<span><span>k</span>k</span>}}<span><span>(</span>(</span>{\mathrm{<span><span>o</span>o</span>}}_{\mathrm{<span><span>p</span>p</span>}}<span><span>)</span>)</span><span><span>=</span>=</span>\mathrm{<span><span>P</span>P</span>}<span><span>(</span>(</span>{\mathrm{<span><span>z</span>z</span>}}_{\mathrm{<span><span>k</span>k</span>}}<span><span>|</span>|</span>{\mathrm{<span><span>o</span>o</span>}}_{\mathrm{<span><span>p</span>p</span>}}<span><span>)</span>)</span><span><span>(</span>(</span>\mathrm{<span><span>\&beta;</span>\&beta;</span>}\underset{\mathrm{<span><span>z</span>z</span>}<span><span>=</span>=</span>\mathrm{<span><span>1</span>1</span>}}{\overset{\mathrm{<span><span>C</span>C</span>}}{\mathrm{<span><span>\&Sigma;</span>\&Sigma;</span>}}}{\mathrm{<span><span>s</span>the\; s</span>}}^{\mathrm{<span><span>k</span>k</span>}}<span><span>(</span>(</span>{\mathrm{<span><span>u</span>u</span>}}_{\mathrm{<span><span>z</span>z</span>}}<span><span>)</span>)</span><span><span>+</span>+</span><span><span>(</span>(</span>\mathrm{<span><span>1</span>1</span>}<span><span>-</span>-</span>\mathrm{<span><span>\&beta;</span>\&beta;</span>}<span><span>)</span>)</span>{\mathrm{<span><span>s</span>the\; s</span>}}^{\mathrm{<span><span>k</span>k</span>}}<span><span>(</span>(</span>{\mathrm{<span><span>u</span>u</span>}}_{\mathrm{<span><span>h</span>h</span>}}<span><span>)</span>)</span><span><span>)</span>)</span><span><span>-</span>-</span><span><span>-</span>-</span><span><span>-</span>-</span><span><span>(</span>(</span>\mathrm{<span><span>12</span>12</span>}<span><span>)</span>)</span>$

其中C是对兴趣对象o_p产生兴趣的用户。β是一个控制参数。where C is the user who is interested in the object of interest o_p . β is a control parameter.

对用户和兴趣对象的影响力迭代更新，直到达到稳态，最终我们得到每个用户和兴趣对象的主题敏感的影响力

The influence of users and objects of interest is updated iteratively until a steady state is reached, and finally we get the topic-sensitive influence of each user and object of interest

具体实施例：Specific examples:

为了评估本发明，本发明的一个实施例基于兴趣的社区媒体图片分享网站Flickr的API接口，抓取了2,314个用户，和总共556942张图片。图6a显示了本发明的方法筛选的在四个主题有影响力的代表性用户。每个主题用超图正则化的主题概率模型学习得到，用最显著的词表示，词越相关重要，它的字体就越大。从图6a可以看出四个主题分别与花、女孩、城市、海滩景色相关。每个主题下，影响力的分最高的用户和他的在该主题下影响力分数最高的图片罗列出，从图中可以看出，主题词汇与图片相关。图6b显示了本发明的方法筛选的在四个主题有影响力的代表性图片。每个主题下影响力分数最高的图片罗列出来。四个主题分别与天空云、暗场景、肖像、海边日出有关，图片视觉内容与文字保持很好的一致性解释。In order to evaluate the present invention, an embodiment of the present invention is based on the API interface of Flickr, an interest-based social media photo sharing website, and 2,314 users, and a total of 556,942 pictures were captured. Figure 6a shows representative users who are influential in four topics screened by the method of the present invention. Each topic is learned with a hypergraph regularized topic probability model, represented by the most prominent word, the more relevant and important the word, the larger its font. It can be seen from Figure 6a that the four themes are respectively related to flowers, girls, cities, and beach scenes. Under each topic, the user with the highest influence score and his picture with the highest influence score under the topic are listed. It can be seen from the figure that the topic vocabulary is related to the picture. Fig. 6b shows representative images screened by the method of the present invention that are influential in four subjects. The images with the highest impact scores under each topic are listed. The four themes are respectively related to sky clouds, dark scenes, portraits, and seaside sunrises. The visual content of the pictures and the text maintain a good consistency in interpretation.

以上所述的具体实施例，对本发明的目的、技术方案和有益效果进行了进一步详细说明，应理解的是，以上所述仅为本发明的具体实施例而已，并不用于限制本发明，凡在本发明的精神和原则之内，所做的任何修改、等同替换、改进等，均应包含在本发明的保护范围之内。The specific embodiments described above have further described the purpose, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above descriptions are only specific embodiments of the present invention, and are not intended to limit the present invention. Within the spirit and principles of the present invention, any modifications, equivalent replacements, improvements, etc., shall be included in the protection scope of the present invention.

Claims (8)

Translated fromChinese

1.一种在社交媒体网络上自动筛选有影响力用户的方法，其特征在于，该方法包括步骤如下：1. A method for automatically screening influential users on a social media network, characterized in that the method comprises steps as follows:步骤S1：利用超图模型为兴趣社交媒体网络中的用户、兴趣对象及其相互作用关系建模；Step S1: Use the hypergraph model to model users, interest objects and their interaction relationships in the interest social media network;步骤S2：采用超图约束的正则化主题概率模型，利用兴趣对象的内容信息和内容信息之间的相似性关系作为约束，自动学习得到隐含的兴趣主题；Step S2: Using the regularized topic probability model constrained by the hypergraph, using the content information of the object of interest and the similarity relationship between the content information as constraints, automatically learn to obtain the hidden topic of interest;步骤S3：对每个用户和兴趣对象进行主题影响力排序，采用相似性传播模型及在超图上的用户和兴趣对象及相互之间的超边传播主题影响力，直到稳态，然后排序可得到特定主题下的有影响力的用户。Step S3: Sorting the subject influence of each user and interest object, using the similarity propagation model and users and interest objects on the hypergraph and the hyperedge propagation topic influence between them, until the steady state, and then sorting can Get influential users on a specific topic.2.根据权利要求1所述的在社交媒体网络上自动筛选有影响力用户的方法，其特征在于，所述步骤S1包括：用超图节点表示社交媒体网络中的用户和兴趣对象，用同质超边表示兴趣对象之间的内容相似性，用异质超边表示用户和兴趣对象之间的高阶社交链接关系。2. the method for automatic screening influential user on social media network according to claim 1, is characterized in that, described step S1 comprises: represent the user and interest object in social media network with hypergraph node, use same The qualitative hyperedge represents the content similarity between interest objects, and the heterogeneous hyperedge represents the high-order social link relationship between users and interest objects.3.根据权利要求2所述的在社交媒体网络上自动筛选有影响力用户的方法，其特征在于，所述兴趣对象之间的内容相似性包括视觉内容相似性和文本内容相似性，并且，3. The method for automatically screening influential users on a social media network according to claim 2, wherein the content similarity between the objects of interest comprises visual content similarity and text content similarity, and,用于表示视觉内容相似性的超边构建步骤为：对于每一个兴趣对象，找到其K个最近邻的兴趣对象，然后用一条同质超边连接这些节点，并且权重设为1；The hyperedge construction steps used to represent the similarity of visual content are: for each object of interest, find its K nearest neighbor objects of interest, and then connect these nodes with a homogeneous hyperedge, and set the weight to 1;用于表示文本内容相似性的超边构建步骤为：首先从所有兴趣对象的文本元数据抽取一个词典，然后对于每一个词，为所有包含该词的兴趣对象建立一条超边，并且权重设为1。The hyperedge construction steps used to represent the similarity of text content are as follows: firstly, a dictionary is extracted from the text metadata of all interest objects, and then for each word, a hyperedge is established for all interest objects containing the word, and the weight is set to 1.4.根据权利要求2所述的在社交媒体网络上自动筛选有影响力用户的方法，其特征在于，所述异质超边包括：4. The method for automatically screening influential users on a social media network according to claim 2, wherein the heterogeneous hyperedge comprises:拥有者-多个兴趣对象-单一用户的超边，该超边的权重设为1；The owner-multiple interest objects-single user hyperedge, the weight of the hyperedge is set to 1;拥有者-单一兴趣对象-多个用户的超边，该超边的权重设为1。Owner-single interest object-multiple users hyperedge, the weight of the hyperedge is set to 1.5.根据权利要求1所述的在社交媒体网络上自动筛选有影响力用户的方法，其特征在于，所述超图约束正则化的主题概率模型为：对象在主题语义空间的分布保持局部相似性。计算兴趣对象o的主题分布p(z|o)与语义空间主题p(w|z)公式如下：5. The method for automatically screening influential users on a social media network according to claim 1, wherein the topic probability model of the hypergraph constraint regularization is: the distribution of objects in the topic semantic space remains locally similar sex. The formula for calculating the topic distribution p(z|o) and semantic space topic p(w|z) of the object of interest o is as follows:${\mathrm{<span><span>L</span>L</span>}}^{<span><span>\&prime;</span>\&prime;</span>}<span><span>=</span>=</span>\underset{\mathrm{<span><span>i</span>i</span>}<span><span>=</span>=</span>\mathrm{<span><span>1</span>1</span>}}{\overset{\mathrm{<span><span>N</span>N</span>}}{\mathrm{<span><span>\&Sigma;</span>\&Sigma;</span>}}}\underset{\mathrm{<span><span>j</span>j</span>}<span><span>=</span>=</span>\mathrm{<span><span>1</span>1</span>}}{\overset{\mathrm{<span><span>M</span>m</span>}}{\mathrm{<span><span>\&Sigma;</span>\&Sigma;</span>}}}\mathrm{<span><span>n</span>no</span>}<span><span>(</span>(</span>{\mathrm{<span><span>o</span>o</span>}}_{\mathrm{<span><span>i</span>i</span>}}<span><span>,</span>,</span>{\mathrm{<span><span>w</span>w</span>}}_{\mathrm{<span><span>j</span>j</span>}}<span><span>)</span>)</span>\mathrm{<span><span>log</span>log</span>}\underset{\mathrm{<span><span>k</span>k</span>}<span><span>=</span>=</span>\mathrm{<span><span>1</span>1</span>}}{\overset{\mathrm{<span><span>K</span>K</span>}}{\mathrm{<span><span>\&Sigma;</span>\&Sigma;</span>}}}\mathrm{<span><span>P</span>P</span>}<span><span>(</span>(</span>{\mathrm{<span><span>w</span>w</span>}}_{\mathrm{<span><span>j</span>j</span>}}<span><span>|</span>|</span>{\mathrm{<span><span>z</span>z</span>}}_{\mathrm{<span><span>k</span>k</span>}}<span><span>)</span>)</span>\mathrm{<span><span>P</span>P</span>}<span><span>(</span>(</span>{\mathrm{<span><span>z</span>z</span>}}_{\mathrm{<span><span>k</span>k</span>}}<span><span>|</span>|</span>{\mathrm{<span><span>o</span>o</span>}}_{\mathrm{<span><span>i</span>i</span>}}<span><span>)</span>)</span>$通过优化上式可以求出p(z|o)和p(w|z)。其中，N是兴趣对象数目，M是词的总数目，K是隐含主题的个数。n(o_i，w_j)是词和兴趣对象的共生数目。上式表示的意思时通过优化似然求解参数。By optimizing the above formula, p(z|o) and p(w|z) can be obtained. Among them, N is the number of interest objects, M is the total number of words, and K is the number of hidden topics. n(o_i , w_j ) is the co-occurrence number of words and interest objects. The meaning expressed by the above formula is to solve the parameters by optimizing the likelihood.6.根据权利要求1所述的在社交媒体网络上自动筛选有影响力用户的方法，其特征在于，所述步骤S3包括：6. The method for automatically screening influential users on a social media network according to claim 1, wherein said step S3 comprises:基于拥有者-多个兴趣对象-单一用户的超边，计算用户影响力；Calculate user influence based on the hyperedge of owner-multiple interest objects-single user;基于用户影响力和拥有者-单一兴趣对象-多个用户的超边，计算每一个兴趣对象影响力；Calculate the influence of each interest object based on the user influence and the hyperedge of the owner-single interest object-multiple users;对用户影响力和兴趣对象影响力迭代更新，直到达到稳态，得到每个用户和兴趣对象的影响力。Iteratively update the influence of users and objects of interest until a steady state is reached, and the influence of each user and object of interest is obtained.7.根据权利要求6所述的在社交媒体网络上自动筛选有影响力用户的方法，其特征在于，所述步骤S3的计算用户影响力的步骤包括：7. The method for automatically screening influential users on a social media network according to claim 6, wherein the step of calculating user influence of the step S3 comprises:通过下式计算两个用户u_i，u_j的主题相似度f^k(i，j)：The topic similarity f^k (i, j) of two users u_i , u_j is calculated by the following formula:$f^k(i,j)=\log \frac{g(u_i,u_j,k)}{\underset{z\&Element;S}{\mathrm{\&Sigma;}}g(u_i,u_z,k)},$$g(u_i,u_j,k)=\underset{o_p\&Element;O_{u_j}}{\mathrm{\&Sigma;}}{s}^k\left(o_p\right),$其中，变量

和

表示影响力消息，r^k(i，j)由用户结点u_i发至u_j，表示从用户u_i的角度看，他同意用户u_j在主题k上影响他的程度；a^k(i，j)由用户结点u_j发至u_i，表示从用户u_j的角度看，他认为他对用户u_i在主题k上影响的程度；$f^k(i,j)=\log \frac{g(u_i,u_j,k)}{\underset{z\&Element;S}{\mathrm{\&Sigma;}}g(u_i,u_z,k)},$$g(u_i,u_j,k)=\underset{o_p\&Element;o_{u_j}}{\mathrm{\&Sigma;}}{\mathrm{the\; s}}^k\left(o_p\right),$ where the variable

and

Indicates the influence message, r^k (i, j) is sent by user node u_i to u_j , indicating that from the perspective of user u_i , he agrees with the extent to which user u_j influences him on topic k; a^k (i , j) is sent from user node u_j to u_i , indicating from the perspective of user u_j , he thinks the extent of his influence on user u_i on topic k;通过在拥有者-多个兴趣对象-单一用户的超边传递消息，更新用户影响力得分如下：By passing messages on the hyperedge of the owner-multiple interest objects-single user, the user influence score is updated as follows:${\mathrm{<span><span>r</span>r</span>}}^{\mathrm{<span><span>k</span>k</span>}}<span><span>(</span>(</span>\mathrm{<span><span>i</span>i</span>}<span><span>,</span>,</span>\mathrm{<span><span>j</span>j</span>}<span><span>)</span>)</span><span><span>\&LeftArrow;</span>\&LeftArrow;</span>{\mathrm{<span><span>f</span>f</span>}}^{\mathrm{<span><span>k</span>k</span>}}<span><span>(</span>(</span>\mathrm{<span><span>i</span>i</span>}<span><span>,</span>,</span>\mathrm{<span><span>j</span>j</span>}<span><span>)</span>)</span><span><span>-</span>-</span>\underset{\mathrm{<span><span>z</span>z</span>}<span><span>\&Element;</span>\&Element;</span>\mathrm{<span><span>S</span>S</span>}<span><span>(</span>(</span>\mathrm{<span><span>i</span>i</span>}<span><span>)</span>)</span>}{\mathrm{<span><span>max</span>max</span>}}<span><span>\{</span>\{</span>{\mathrm{<span><span>f</span>f</span>}}^{\mathrm{<span><span>k</span>k</span>}}<span><span>(</span>(</span>\mathrm{<span><span>i</span>i</span>}<span><span>,</span>,</span>\mathrm{<span><span>z</span>z</span>}<span><span>)</span>)</span><span><span>+</span>+</span>{\mathrm{<span><span>a</span>a</span>}}^{\mathrm{<span><span>k</span>k</span>}}<span><span>(</span>(</span>\mathrm{<span><span>i</span>i</span>}<span><span>,</span>,</span>\mathrm{<span><span>z</span>z</span>}<span><span>)</span>)</span><span><span>\}</span>\}</span>$${\mathrm{<span><span>a</span>a</span>}}^{\mathrm{<span><span>k</span>k</span>}}<span><span>(</span>(</span>\mathrm{<span><span>i</span>i</span>}<span><span>,</span>,</span>\mathrm{<span><span>j</span>j</span>}<span><span>)</span>)</span><span><span>\&LeftArrow;</span>\&LeftArrow;</span>\mathrm{<span><span>min</span>min</span>}<span><span>\{</span>\{</span>\mathrm{<span><span>0</span>0</span>}<span><span>,</span>,</span>{\mathrm{<span><span>r</span>r</span>}}^{\mathrm{<span><span>k</span>k</span>}}<span><span>(</span>(</span>\mathrm{<span><span>j</span>j</span>}<span><span>,</span>,</span>\mathrm{<span><span>j</span>j</span>}<span><span>)</span>)</span><span><span>+</span>+</span>\underset{\mathrm{<span><span>z</span>z</span>}<span><span>\&NotElement;</span>\&NotElement;</span><span><span>\{</span>\{</span>\mathrm{<span><span>i</span>i</span>}<span><span>,</span>,</span>\mathrm{<span><span>j</span>j</span>}<span><span>\}</span>\}</span>}{\mathrm{<span><span>\&Sigma;</span>\&Sigma;</span>}}\mathrm{<span><span>max</span>max</span>}<span><span>\{</span>\{</span>\mathrm{<span><span>0</span>0</span>}<span><span>,</span>,</span>{\mathrm{<span><span>r</span>r</span>}}^{\mathrm{<span><span>k</span>k</span>}}<span><span>(</span>(</span>\mathrm{<span><span>z</span>z</span>}<span><span>,</span>,</span>\mathrm{<span><span>j</span>j</span>}<span><span>)</span>)</span><span><span>\}</span>\}</span><span><span>\}</span>\}</span><span><span>,</span>,</span>$${\mathrm{<span><span>a</span>a</span>}}^{\mathrm{<span><span>k</span>k</span>}}<span><span>(</span>(</span>\mathrm{<span><span>j</span>j</span>}<span><span>,</span>,</span>\mathrm{<span><span>j</span>j</span>}<span><span>)</span>)</span><span><span>\&LeftArrow;</span>\&LeftArrow;</span>\underset{{\mathrm{<span><span>i</span>i</span>}}^{<span><span>\&prime;</span>\&prime;</span>}<span><span>\&NotEqual;</span>\&NotEqual;</span>\mathrm{<span><span>j</span>j</span>}}{\mathrm{<span><span>\&Sigma;</span>\&Sigma;</span>}}\mathrm{<span><span>max</span>max</span>}<span><span>\{</span>\{</span>\mathrm{<span><span>0</span>0</span>}<span><span>,</span>,</span>{\mathrm{<span><span>r</span>r</span>}}^{\mathrm{<span><span>k</span>k</span>}}<span><span>(</span>(</span>{\mathrm{<span><span>i</span>i</span>}}^{<span><span>\&prime;</span>\&prime;</span>}<span><span>,</span>,</span>\mathrm{<span><span>j</span>j</span>}<span><span>)</span>)</span><span><span>\}</span>\}</span>$其中S(i)表示用户u_i所感兴趣的兴趣对象的拥有用户的集合。通过上面相似传播算法我们得到稳态的

和

其中S(i)表示用户u_i所感兴趣的兴趣对象的拥有用户的集合；Where S(i) represents the set of users who own the object of interest that user u_i is interested in. Through the above similarity propagation algorithm we get the steady-state

and

Among them, S(i) represents the collection of users who own the object of interest that user u_i is interested in;定义用户之间的影响力为：用q^k(i，j)表示用户u_j对用户u_i在主题k上的影响力：Define the influence between users as: use q^k (i, j) to represent the influence of user u_j on user u_i on topic k:${\mathrm{<span><span>q</span>q</span>}}^{\mathrm{<span><span>k</span>k</span>}}<span><span>(</span>(</span>\mathrm{<span><span>i</span>i</span>}<span><span>,</span>,</span>\mathrm{<span><span>j</span>j</span>}<span><span>)</span>)</span><span><span>=</span>=</span>\frac{\mathrm{<span><span>1</span>1</span>}}{\mathrm{<span><span>1</span>1</span>}<span><span>+</span>+</span>{\mathrm{<span><span>e</span>e</span>}}^{<span><span>-</span>-</span><span><span>(</span>(</span>{\mathrm{<span><span>r</span>r</span>}}^{\mathrm{<span><span>k</span>k</span>}}<span><span>(</span>(</span>\mathrm{<span><span>i</span>i</span>}<span><span>,</span>,</span>\mathrm{<span><span>j</span>j</span>}<span><span>)</span>)</span><span><span>+</span>+</span>{\mathrm{<span><span>a</span>a</span>}}^{\mathrm{<span><span>k</span>k</span>}}<span><span>(</span>(</span>\mathrm{<span><span>i</span>i</span>}<span><span>,</span>,</span>\mathrm{<span><span>j</span>j</span>}<span><span>)</span>)</span><span><span>)</span>)</span>}}<span><span>;</span>;</span>$计算用户在主题的全局影响力：Calculate the global influence of a user on a topic:$s^k\left(j\right)=\mathrm{\&eta;}\underset{i:i\&RightArrow;j}{\mathrm{\&Sigma;}}{s}^k\left(i\right)p^k\left(j\right|i)+(1-\mathrm{\&eta;})v_j^k,$其中η是一个控制参数，${\mathrm{the\; s}}^k\left(j\right)=\mathrm{\&eta;}\underset{i:i\&Right\; Arrow;j}{\mathrm{\&Sigma;}}{\mathrm{the\; s}}^k\left(i\right)p^k\left(j\right|i)+(1-\mathrm{\&eta;})v_j^k,$ where η is a control parameter,p^k(j|i)表示用户u_i到用户u_j的转移概率，且p^k (j|i) represents the transition probability from user u_i to user u_j , and

其中

是用户u_j的初始化在主题k的影响力，$v_j^k=\underset{o_i\&Element;O_{v_j}}{\mathrm{\&Sigma;}}P\left(z_k\right|o_i);$

in

is the initial influence of user u_j on topic k,$v_j^k=\underset{o_i\&Element;o_{v_j}}{\mathrm{\&Sigma;}}P\left(z_k\right|o_i);$对式$(s^k\left(j\right)=\mathrm{\&eta;}\underset{i:i\&RightArrow;j}{\mathrm{\&Sigma;}}{s}^k\left(i\right)p^k\left(j\right|i)+(1-\mathrm{\&eta;})v_j^k$求解得到稳态的用户影响力为Pair$({\mathrm{the\; s}}^k\left(j\right)=\mathrm{\&eta;}\underset{i:i\&Right\; Arrow;j}{\mathrm{\&Sigma;}}{\mathrm{the\; s}}^k\left(i\right)p^k\left(j\right|i)+(1-\mathrm{\&eta;})v_j^k$ Solving the steady-state user influence is${\mathrm{<span><span>s</span>the\; s</span>}}_{\mathrm{<span><span>u</span>u</span>}}^{\mathrm{<span><span>k</span>k</span>}}<span><span>=</span>=</span><span><span>(</span>(</span>\mathrm{<span><span>1</span>1</span>}<span><span>-</span>-</span>\mathrm{<span><span>\&eta;</span>\&eta;</span>}<span><span>)</span>)</span>{<span><span>(</span>(</span>\mathrm{<span><span>I</span>I</span>}<span><span>-</span>-</span>{\mathrm{<span><span>\&eta;Q</span>\&eta;Q</span>}}_{\mathrm{<span><span>k</span>k</span>}}<span><span>)</span>)</span>}^{<span><span>-</span>-</span>\mathrm{<span><span>1</span>1</span>}}{\mathrm{<span><span>v</span>v</span>}}^{\mathrm{<span><span>k</span>k</span>}}<span><span>.</span>.</span>$8.根据权利要求6所述的在社交媒体网络上自动筛选有影响力用户的方法，其特征在于，所述步骤S3中计算兴趣对象影响力的公式为：8. The method for automatically screening influential users on a social media network according to claim 6, wherein the formula for calculating the influence of the object of interest in the step S3 is:利用$s_o^k\left(o_p\right)=P\left(z_k\right|o_p)(\mathrm{\&beta;}\underset{z=1}{\overset{C}{\mathrm{\&Sigma;}}}{s}^k\left(u_z\right)+(1-\mathrm{\&beta;})s^k\left(u_h\right)),$其中C是对兴趣对象o_p产生兴趣的用户，β是一个控制参数。use${\mathrm{the\; s}}_o^k\left(o_p\right)=P\left(z_k\right|o_p)(\mathrm{\&beta;}\underset{z=1}{\overset{C}{\mathrm{\&Sigma;}}}{\mathrm{the\; s}}^k\left(u_z\right)+(1-\mathrm{\&beta;}){\mathrm{the\; s}}^k\left(u_h\right)),$ where C is the user who is interested in the object of interest o_p , and β is a control parameter.

Images