Key user discovery method for subdivided domain and wide-area social contact

Abstract

The invention belongs to the technical field of internet application, and particularly relates to a key user discovery method oriented to subdivided fields and wide-area social contact, which comprises the following steps: acquiring user basic information, user historical behaviors and user social contact contents under a current topic, and extracting field information and social contact information from the user basic information, the user historical behaviors and the user social contact contents; calculating the domain recognition degree of one user to another user according to the domain information and the social information, and establishing a topic propagation network by taking the user as a node and the domain recognition degree as an edge weight; in the topic propagation network, walking based on a bias sampling strategy to obtain structural features of nodes; a two-layer graph convolutional neural network is established, the network is composed of a first graph convolutional layer, a Dropout layer and a second graph convolutional layer which are cascaded, and structural features of nodes are input into the network to predict whether the nodes have influence or not; according to the method, key users can be found in specific fields, key users with more influence are selected for customers to carry out product popularization, and the effectiveness of advertisement putting is improved.

Description

Technical Field

[0001] The present invention belongs to the field of Internet application technology, relates to e-commerce platform marketing data analysis, and in particular to a key user discovery method for segmented fields and wide-area social networking. Background Art

[0002] Influenced by digital technology and social platforms, people's purchasing methods have changed dramatically. In order to gain greater exposure and sales in the online market, many companies have begun to seek cooperation with influential key users and use their huge influence on social media to promote their products or services. These key users have successfully attracted a group of relevant target users by publishing high-quality content in specific fields or topics, and gradually accumulated a group of loyal fans. The cooperation between brands and key users has not only significantly increased the sales of goods, but also effectively improved brand awareness.

[0003] Combining existing research on e-commerce platform marketing data, we found that there are still the following problems in considering user influence from both the domain and wide-area perspectives:

[0004] 1. In view of the complexity of measuring user identification and the wide range of social scope, it is difficult to directly quantify users' domain recognition and social identification.

[0005] 2. The complexity of user communication behavior in marketing topics. There are many factors that affect user communication behavior in marketing topics. Not only should we consider the field to which the user belongs, but we should also combine the user's social comprehensive analysis to analyze the user's communication behavior. How to combine these two factors to more comprehensively mine node information is a difficult problem.

[0006] 3. Complexity of user domain information in the marketing topic domain. User domain information is complex and it is difficult to accurately determine which fields users are active in. Users in the marketing topic domain have domain relevance. On the basis of ensuring the original relationship of users, how to construct the association relationship of users in marketing topics is the key to subsequently using graph convolutional neural networks to discover user influence. Summary of the invention

[0007] In view of the problems still existing in the prior art, the present invention proposes a key user discovery method for segmented fields and wide-area social networking, which specifically includes the following steps:

[0008] Obtain basic user information, historical user behavior, and social content under the current topic, and extract domain information and social information from them;

[0009] Calculate the domain recognition of one user to another user based on domain information and social information, and establish a topic propagation network with users as nodes and domain recognition as edge weights;

[0010] In the topic propagation network, the structural features of the nodes are obtained by walking based on the biased sampling strategy;

[0011] Establish a two-layer graph convolutional neural network, which consists of a cascaded first graph convolutional layer, a dropout layer, and a second graph convolutional layer;

[0012] The structural features of the node are input into a two-layer graph convolutional neural network, which predicts whether the node is an influential node in the current topic.

[0013] The present invention starts from the user's domain information and behavior data, and uses information entropy to quantify the user's domain recognition and social identity; at the same time, using the quantification results, a new random walk strategy is set to incorporate the user's node's domain recognition and social identity to extract implicit structural features; finally, based on the user relationship network, a graph convolutional neural network is used to mine the user's influence in the topic space. In summary, the present invention can be applied to e-commerce platforms to more accurately analyze user influence, thereby planning more suitable candidates to make profits for the platform; in addition, it is also helpful to promote the company's products and services, enhance brand awareness and influence, and thus obtain good economic benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 This is a flow chart of a key user discovery method for segmented fields and wide-area social networking according to the present invention;

[0015] Figure 2 A schematic diagram showing hidden user relationships of the FSU2vec algorithm of the present invention;

[0016] Figure 3 Schematic diagram of the FS-GCN model of the present invention. DETAILED DESCRIPTION

[0017] 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.

[0018] The present invention proposes a key user discovery method for segmented fields and wide-area social networking, which specifically includes the following steps:

[0019] Obtain basic user information, historical user behavior, and social content under the current topic, and extract domain information and social information from them;

[0020] Calculate the domain recognition of one user to another user based on domain information and social information, and establish a topic propagation network with users as nodes and domain recognition as edge weights;

[0021] In the topic propagation network, the structural features of the nodes are obtained by walking based on the biased sampling strategy;

[0022] Establish a two-layer graph convolutional neural network, which consists of a cascaded first graph convolutional layer, a dropout layer, and a second graph convolutional layer;

[0023] The structural features of the node are input into a two-layer graph convolutional neural network, which predicts whether the node is an influential node in the current topic.

[0024] like Figure 1 The figure shows the overall process of the present invention, indicating that the input of the present invention is the basic information of users on the e-commerce platform, the historical behavior data of users and the social content of users, and the output after the discovery model is whether the users participating in the marketing topic are influential.

[0025] In this embodiment, a key user discovery method for segmented fields and wide-area social networking can be divided into three parts, specifically including:

[0026] S1: Get data online.

[0027] The way to obtain data is to directly query the real-time marketing data in the database provided by the enterprise. What needs to be obtained here is the basic information of users participating in the marketing topic and the historical behavior data of users participating in the marketing topic; in addition, it is also necessary to obtain the social content of users participating in the marketing topic and pre-process the data. The specific steps include:

[0028] S11: Get the original data. You can get the original data through the public API of social networks or by directly downloading existing data sources.

[0029] S12: Simple data cleaning. The raw data usually obtained are unstructured and cannot be used directly for data analysis. Most unstructured data can be structured through simple data cleaning, for example, deleting duplicate data, cleaning invalid information, etc.

[0030] S2: Extract relevant attributes.

[0031] User-related attributes, including domain information and social information, are extracted based on the user's basic information, user historical behavior, and user social content obtained from the marketing topics.

[0032] In e-commerce platforms, users' dissemination behavior of marketing topics is affected by many factors, such as the influence of domain similarity between users and the interaction rate between users and their friends on user dissemination behavior. Based on this, this embodiment first extracts factors that affect user behavior, and extracts them from two aspects: domain and wide area. The details are as follows:

[0033] S21: Extracting Domain Factors

[0034] S211: Domain Similarity

[0035] Domain similarity is defined as the similarity or correlation between the domains to which two users belong. Keywords are extracted from user social content, and the Jaccard index is used to describe the domain similarity between two users. The larger the Jaccard index, the greater the domain correlation between the two users. Conversely, the smaller the Jaccard index, the smaller the domain correlation between the two users. Field(u1) and Field(u2) represent the domain keywords of user u1 and user u2 respectively. In this embodiment, the domain similarity is defined as:

[0036]

[0037] Among them, Sim(u1,u2) is the domain similarity between user node u1 and user node u2; ∩ means finding the intersection, ∪ means finding the union, and |·| means finding the number of elements in the set.

[0038] S22: Extracting social factors

[0039] S221: User interaction rate

[0040] The user interaction rate indicates the frequency of interaction between a user and his friends. Interaction includes forwarding, commenting, and liking. It is defined as:

[0041] Mutual(ui,x)=(Num[interact(ui,x)] 2 / Num[orig(x)]) 13 (2)

[0042] Among them, Num[interact(u i ,x)] is the total number of interactions of user node x forwarding, commenting and liking the topic published by user node ui, Num[orig(x)] is the total number of interactions of user node u before the topic spread i The number of original content published.

[0043] S222: User attention rate, user forwarding rate, user comment rate, user like rate

[0044] The user attention rate, forwarding rate, comment rate and like rate can reflect the user's recognition to a certain extent, and it is helpful to discover the user's influence. Generally speaking, the higher the user's attention rate, forwarding rate, comment rate and like rate, the more people recognize the content posted by the user. Therefore, this embodiment defines the user attention rate, user forwarding rate, user comment rate and user like rate as:

[0045] Ra(ui)=(Num[attn(ui)] 2 / Num[orig(ui)]) 13 (3)

[0046] Rf(ui)=(Num[frwd(ui)] 2 / Num[orig(ui)]) 13 (4)

[0047] Rc(ui)=(Num[coms(ui)] 2 / Num[orig(ui)]) 13 (5)

[0048] Rl(ui)=(Num[like(ui)] 2 / Num[orig(ui)]) 13 (6)

[0049] Among them, R f (u i ) is the user node u i The forwarding rate, R c (u i ) is the user node u i The comment rate, R l (u i ) is the like rate of user node ui, R a (u i ) is the attention rate of user node ui; Num[frwd(u i )] represents the number of reposts of the content published by user node ui before the topic spreads; Num[orig(u i )] represents the number of original content published by user node ui before the topic spreads; Num[coms(u i )] represents the number of comments on the content published by user node ui before the topic spreads; Num[like(u i )] represents the number of likes for the content published by user node ui before the topic spreads; Num[attn(u i )] represents the number of users who follow the user node ui.

[0050] S3: Build the model.

[0051] In the process of model building, we first use information entropy to quantify the user's domain recognition and social identity; secondly, we propose the FSU2vec algorithm based on domain recognition and social identity to extract implicit structural features; finally, we use convolutional neural networks to build a marketing topic user influence discovery model FS-GCN to determine whether the users participating in the marketing topic are influential. In summary, model building is mainly divided into three stages, such as Figure 1 , specifically including:

[0052] In the first stage, starting from the user's domain information and behavior data, information entropy is used to quantify the user's domain recognition and social identity;

[0053] In the second stage, we use the quantization results to set a new bias sampling strategy and propose a walking algorithm FSU2vec, which incorporates the domain recognition and social identity of user nodes to extract implicit structural features.

[0054] In the third stage, based on the user relationship network, a key user influence discovery model FS-GCN for marketing topics in segmented fields and wide-area social networking is proposed to study the user influence in the marketing topic space.

[0055] In marketing topics, since different users have their own fields, users will selectively spread topics based on their own fields. Therefore, the field factor is a major factor that determines the influence of users in the marketing topic network. At the same time, the user's social factors such as user forwarding rate, user comment rate and user like rate also determine the size of the influence. In order to better measure the recognition, the field factor and the wide-area factor are combined, and the user's field recognition is defined using information entropy as:

[0056]

[0057] Among them, F f (u i ) is the user node u i The average domain similarity with all users who have forwarded their content, F c (u i ) is the average domain similarity between the user node ui and all users who have commented on its content, F l (u i ) is the average domain similarity between user node ui and all users who have liked its content.

[0058] The average domain similarity between all commenting users and all liking users is defined as follows:

[0059]

[0060] Among them, Sim(u i ,u k ) is the user node u i With user node u k The domain similarity between f (u i ) represents the set of users who have forwarded the content published by user node ui, C c (u i ) represents the set of users who have commented on the content published by user node ui, C l (u i ) represents the set of users who have liked the content published by user node ui.

[0061] Combine the user interaction rate and the user's recognition in the field of the entire network to define user u j For user u i The recognition in the field is:

[0062]

[0063] Among them, Mutual(u i ,x) is the interaction frequency between user node ui and user node x.

[0064] Social identity is determined by two factors: the strength of social relations and the scope of social influence. Combining these two factors, the present invention uses information entropy to construct the user u i The social recognition is:

[0065]

[0066] S32: Implicit Structure Representation

[0067] First, we define the random walk strategy of the FSU2vec algorithm by combining domain recognition and social recognition to optimize the FSU2vec algorithm. Then, we use the optimized FSU2vec algorithm to extract the hidden information of the topic network and obtain the hidden feature vector matrix of the topic network. The specific process is as follows: Figure 2 shown.

[0068] The edges in the network are the relationships between users, and the edge weights are assigned based on the domain recognition between users. The edge weights are defined as follows:

[0069] w(u i ,x)=Rec field (u i ,x) (11)

[0070] The formula represents the user u i The edge weight with the next user x, Rec field (u i,x) represents the relationship between user x and user u i recognition in the field.

[0071] This paper proposes a new biased sampling strategy and a walking algorithm FSU2vec. In the communication network, domain recognition and social recognition are used to extract multiple user sequences of equal length, and the structural features of the nodes are discovered with the help of the sequences to generate structural feature vectors. i-1 Represents the previous user node, u i Represents the current node, the transition probability of the next node x is affected by the weight value and is defined as:

[0072]

[0073] Among them, P(xu i ,u i-1 ) represents the transition probability of the next user node x under the premise that the previous user node is ui-1 and the current user node is ui; w(u i ,x) represents the edge weight between the current user node ui and the next user node x; α(u i-1 ,u i ,x) is the bias for controlling the walking strategy; (u i-1 ,x)∈E means the previous user node is u i-1 There is an edge relationship between it and the next user node x, and this edge relationship belongs to the edge relationship set E in the topic propagation network; z is the normalization coefficient, which controls the transition probability within the range of (0,1).

[0074] In the marketing topic communication network, each user has different social identity, so different users have different social influence. Combined with the influence of social identity, the bias α(u i-1 ,u i ,x) as shown in the formula:

[0075]

[0076] Among them, p is a hyperparameter that controls the probability of visiting the previous node. This strategy considers the social recognition of the next node; Rec social (x) is the social recognition of user node x, Rec social (u i-1 ) is the social identity of user node ui-1; q is a hyperparameter that controls the search strategy, such as whether to use a strategy based on depth traversal or a strategy based on breadth search; d(u i-1 ,x) represents the shortest path distance between user node ui-1 and user node x. i-1 ,x)=0, which means that user node ui-1 and user node x are the same node.

[0077] Taking node u as the starting point, the walking strategy proposed in the present invention is used to walk to obtain a node sequence of a set length, and the obtained node sequence is vectorized to obtain the structural features of the node. The structural features of the node are optimized by the FSU2vec method according to the obtained structural features of the node, that is, the structural features of the node are optimized with the goal of maximizing the sum of the probabilities of the nodes appearing in the node set obtained by the walking strategy. The optimization function is defined as shown in the formula:

[0078]

[0079] Among them, F(u) is the communication node under the marketing topic The embedding vector N S (u) is the neighboring node sampled by the walking strategy proposed by the present invention, Pr(N S (u)F(u)) is the probability of the neighboring nodes of node u appearing, as shown in the formula:

[0080]

[0081] Finally, the structural eigenvector matrix is:

[0082] X=N·F(u) (16)

[0083] Among them, N represents the number of marketing topic propagation nodes, and F(u) represents the structural feature vector of a single user node.

[0084] S33: User influence discovery model FS-GCN

[0085] Taking advantage of the processing advantages of graph convolutional networks on non-Euclidean structured data such as social networks, and considering the influence of segmented fields and wide-area social networks in communication, this paper proposes a marketing topic key user influence discovery model FS-GCN for segmented fields and wide-area social networks. The purpose of the discovery model is to determine whether the users participating in the marketing topic are influential, so it can be converted into a binary classification task.

[0086] like Figure 3 As shown in FIG. 1 , the present invention constructs a two-layer graph convolutional neural network with a Dropout layer as a key user influence discovery model. In a GCN layer, the weights and biases are randomly initialized, and the normalized symmetric matrix is ​​calculated. The RELU function is used as the activation function of this layer, and the Dropout operation is performed during model training. This layer of the neural network can be represented by a nonlinear function:

[0087]

[0088] Among them, H (i)represents the feature vector output by the i-th layer graph convolutional neural network, is a symmetric matrix, expressed as D is the diagonal matrix of matrix A and D = ∑ j A ij , A ij is the element in the i-th row and j-th column of matrix A. Matrix A is expressed as A=A+I, where A is the adjacency matrix of the topic propagation network, I is the unit matrix, and W (i) Represents the weight matrix of the i-th layer graph convolutional neural network. In this embodiment, the dimension of matrix A is N×N, and the dimension of the structural feature X of the user node is N×C, where N represents the number of marketing topic propagation nodes, and C represents the vector length of a user structural feature.

[0089] Combined with the formula, the SoftMax activation function is added to the two-layer neural network to represent the convolutional network as the probability values ​​of different node classifications. The graph convolutional network classification model can be expressed as:

[0090]

[0091] Since the influence discovery problem studied in this paper is a binary classification problem, the model output is expressed as P(i,nu i ), which is defined as follows:

[0092]

[0093] Among them, Y = 1 represents the current user u i is judged as an influential user, and Y = 0 means that the current user u i Deemed as a non-influential user.

[0094] 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 key user discovery method for segmented fields and wide-area social networking, characterized in that: The specific steps include: Obtain basic user information, historical user behavior, and social content under the current topic, and extract domain information and social information from them; Calculate the domain recognition of one user to another user based on domain information and social information, and establish a topic propagation network with users as nodes and domain recognition as edge weights; In the topic propagation network, the structural features of the nodes are obtained by walking based on the biased sampling strategy; Establish a two-layer graph convolutional neural network, which consists of a cascaded first graph convolutional layer, a dropout layer, and a second graph convolutional layer; The structural features of the node are input into a two-layer graph convolutional neural network, which predicts whether the node is an influential node in the current topic.

2. A method for discovering key users in segmented fields and wide-area social networking according to claim 1, characterized in that: When walking based on the biased sampling strategy, the transition probability of the next node is expressed as: Among them, P(xu i ,u i-1 ) represents the transition probability of the next user node x under the premise that the previous user node is ui-1 and the current user node is ui; w(u i ,x) represents the edge weight between the current user node ui and the next user node x; α(u i-1 ,u i ,x) is the bias for controlling the walking strategy; (u i-1 ,x)∈E means the previous user node is u i-1 There is an edge relationship between it and the next user node x, and this edge relationship belongs to the edge relationship set E in the topic propagation network; z is the normalization coefficient.

3. A method for discovering key users in segmented fields and wide-area social networking according to claim 2, characterized in that: If the previous user node is u i-1 、The current user node is u i , the next user node is x, then the bias α(u i-1 ,u i ,x) is expressed as: Among them, p is a hyperparameter that controls the probability of visiting the previous node; Rec social (x) is the social recognition of user node x, Rec social (u i-1 ) is the social identity of user node ui-1; q is the hyperparameter controlling the search strategy; d(u i-1 ,x) represents the shortest path distance between user node ui-1 and user node x. i-1 ,x)=0, which means that user node ui-1 and user node x are the same node.

4. A method for discovering key users in segmented fields and wide-area social networking according to claim 3, characterized in that: The social recognition of user node ui Rec social (u i ) is expressed as: Among them, R f (u i ) is the user node u i The forwarding rate, R c (u i ) is the user node u i The comment rate, R l (u i ) is the like rate of user node ui, R a (u i ) is the attention rate of user node ui.

5. The key user discovery method for segmented fields and wide-area social networking according to claim 2, characterized in that: The edge weight w(u i ,x) is expressed as: Among them, Mutual(u i ,x) is the interaction frequency between user node ui and user node x, Sim(u i ,x) is the user node u i The domain similarity between the user node x, Rec field (u i ) is the user node u i Recognition in the field, Rec field (x) is the domain recognition of user node x.

6. The key user discovery method for segmented fields and wide-area social networking according to claim 5, characterized in that: The interaction frequency between user node ui and user node x is Mutual(u i ,x) is expressed as: Mutual(ui,x)=(Num[interact(ui,x)] 2 / Num[orig(x)]) 13 Among them, Num[interact(u i ,x)] is the total number of interactions of user node x forwarding, commenting and liking the topic posted by user node ui, Num[orig(x)] is the total number of interactions of user node u before the topic propagation i The number of original content published.

7. The key user discovery method for segmented fields and wide-area social networking according to claim 5, characterized in that: User node u i Field recognition field (u i ) is expressed as: Among them, R f (u i ) is the user node u i The forwarding rate, R c (u i ) is the user node u i The comment rate, R l (u i ) is the user node u i Like rate; F f (u i ) is the user node u i The average domain similarity with all users who have forwarded their content, F c (u i ) is the average domain similarity between the user node ui and all users who have commented on its content, F l (u i ) is the average domain similarity between user node ui and all users who have liked its content.

8. A method for discovering key users in segmented fields and wide-area social networking according to claim 4 or 7, characterized in that: User node u i The forwarding rate R f (u i ) is expressed as: Rf(ui)=(Num[frwd(ui)] 2 / Number[orig(ui)]) 13 The comment rate R of user node ui c (u i ) is expressed as: Rc(ui)=(Num[coms(ui)] 2 / Number[orig(ui)]) 13 The like rate R of user node ui l (u i ) is expressed as: Rl(ui)=(Num[like(ui)] 2 / Num[orig(ui)]) 13 The attention rate R of user node ui a (u i ) is expressed as: Ra(ui)=(Num[attn(ui)] 2 / Number[orig(ui)]) 13 Among them, Num[frwd(u i )] represents the number of reposts of the content published by user node ui before the topic spreads; Num[orig(u i )] represents the number of original content published by user node ui before the topic spreads; Num[coms(u i )] represents the number of comments on the content published by user node ui before the topic spreads; Num[like(u i )] represents the number of likes for the content published by user node ui before the topic spreads; Num[attn(u i )] represents the number of users who follow the user node ui.

9. A method for discovering key users in segmented fields and wide-area social networking according to claim 4 or 7, characterized in that: The calculation of the domain similarity between two user nodes includes: extracting domain keywords from the social content of the user nodes, and taking the Jaccard index of the domain keyword sets of the two users as the domain similarity between the two user nodes.

10. The key user discovery method for segmented fields and wide-area social networking according to claim 1, characterized in that: The data processing process of the two-layer graph convolutional neural network is expressed as: Among them, H (1) represents the output of the first graph convolutional layer, ReLU(·) represents the RELU function; X represents the structural features of the user node; is a symmetric matrix, expressed as D is the diagonal matrix of matrix A and D = ∑ j A ij , A ij is the element in the i-th row and j-th column of matrix A. Matrix A is expressed as A=A+I, where A is the adjacency matrix of the topic propagation network and I is the unit matrix; W (0) is the learnable weight matrix of the first graph convolutional layer, W (1) is the learnable weight matrix of the second graph convolutional layer; softmax(·) is the SoftMax activation function; Z is the output of the second graph convolutional layer.

Images