Social platform target user identification method for guiding random walk based on relation
By building a heterogeneous network on a social platform and using a random walk method of relationship-guided relationships, combining graph convolution networks and node-edge alternating convolution neural networks, the low KOL recognition efficiency and poor semantic information utilization in the existing technology are solved. The problem has achieved higher identification accuracy.
Patent Information
- Application Number
- CN202411402848.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-09
- Publication Date
- 2025-01-10
AI Technical Summary
The existing social platform KOL recognition method is inefficient and cannot effectively utilize high-order semantic information between users and content, resulting in poor recognition accuracy.
Using a random walk method based on relationship guidance, the characterization vectors are obtained by constructing a heterogeneous network, using graph convolution networks and node-edge alternating convolution neural networks, and combining distance decay factors and attention mechanisms to perform higher-order semantic information sampling and User embedding representation, thereby identifying the target users of the social platform.
It improves the accuracy of KOL recognition on social platform, can quantify users, content and their interactions, and classify and process relationships in social platforms, thereby better obtaining higher-order semantic information.
Smart Images
Figure CN119295247A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of social network data processing, and in particular to a method for identifying target users of a social platform based on relationship-guided random walk. Background Art
[0002] Key Opinion Leaders (KOL) refer to individuals or organizations that actively spread and share information on major social platforms and have a certain influence in the field. With the rapid rise of social platforms in real life such as Weibo and Bilibili, users on various social platforms will generate a large amount of interaction data such as attention, forwarding, and comments. These interaction data reveal the influence of users on each other. There are some special users on social platforms. These users are generally called KOLs. They are the core of the social network composed of the entire social platform and have the characteristics of wide information dissemination range and fast dissemination speed. Therefore, in the fields of public opinion control and advertising, it is particularly important to accurately and effectively identify these KOLs.
[0003] Existing methods for identifying KOLs on social platforms mainly include manual screening and methods based on social networks. Manual screening has the problems of low screening efficiency and heavy workload; methods based on social networks often only focus on the users themselves, while ignoring the semantic information contained in the relationship between users and content, and cannot fully reflect the influence of a certain user on other users. Therefore, methods based on social networks perform poorly in identifying KOLs on social platforms. At present, how to effectively extract and utilize high-order semantic information to improve the accuracy of KOL identification on social platforms is an important problem in the field of KOL identification today. Summary of the invention
[0004] To this end, the purpose of the present invention is to provide a method for identifying target users on social platforms based on relationship-guided random walks, so as to deeply explore the high-order semantic information between users and content on the social platform and improve the accuracy of KOL identification on the social platform.
[0005] A method for identifying target users on a social platform based on relationship-guided random walks, comprising the following steps:
[0006] Step S1, obtaining the interaction data of users and contents in the social platform, and constructing a heterogeneous network based on the relationship between users and contents;
[0007] Step S2, stratifying the heterogeneous network so that different types of users and content are in different network layers, and constructing initial features;
[0008] Step S3, inputting the constructed initial features into the graph convolutional network and the node-edge alternating convolutional neural network to obtain the representation vectors of users, content and the relationship between them;
[0009] Step S4, according to the obtained representation vector, use the relationship-guided random walk to perform high-order semantic information sampling to obtain a random walk sequence containing interactions of different network layers;
[0010] Step S5: In the random walk sequence, the distance decay factor and the attention mechanism are introduced through the semantic fusion component to obtain the user's embedded representation, and the target user of the social platform is identified based on the obtained user's embedded representation.
[0011] In the above-mentioned method for identifying target users on a social platform based on relationship-guided random walk, step S1 specifically includes:
[0012] Obtain social platform data, including users, content, and their interactions. Users include readers, authors, and video hosts. Content includes articles and videos. Their interactions include readers following authors and video hosts, readers reading articles and videos, and authors and video hosts publishing articles and videos. Construct a heterogeneous network G based on the relationship between users and content. HN =(V, E, A, P), where V represents the set of all users and content nodes, E represents the set of interactive relationships between users and content, A represents the set of user and content node types, and P represents the set of interactive relationship types between users and content.
[0013] In the above-mentioned method for identifying target users on a social platform based on relationship-guided random walk, step S2 specifically includes:
[0014] Step S2.1, analyze the types of user and content nodes in the heterogeneous network, stratify the heterogeneous network according to the node types of users and content, and obtain a hierarchical heterogeneous network Where n is the number of all user and content node types, Represents G VHHN Subnetworks at different layers in A 1 , A 2 , A n Indicates different node types;
[0015] Step S2.2, based on the obtained hierarchical heterogeneous network G VHHN Construct the initial feature vector and set the sub-network of each layer to be Among them, V l represents the node set of the lth layer, E l Represents the relationship set of the lth layer. In each sub-network, the Laplacian matrix of the node is extracted The adjacency matrix of the relationship And project them into the same feature space to obtain the initial features, including node v i The initial characteristics and relationship ij The initial characteristics
[0016] In the above-mentioned method for identifying target users on a social platform based on relationship-guided random walk, step S3 specifically includes:
[0017] Step S3.1, for node v i The initial characteristics The node representation vector is obtained using the graph convolutional network according to the following formula:
[0018]
[0019] in, Represents node v i The representation vector of
[0020] Step S3.2, for the relationship e ij The initial characteristics According to the following formula, the interaction function of the node-edge alternating convolutional neural network is used to aggregate the user or content node representations connected to the relationship and the initial features of the relationship itself to obtain the relationship representation vector:
[0021]
[0022] in, Represents node v i , node v j The representation vector of the relationship between Represents node v j The representation vector of .
[0023] In the above-mentioned social platform target user identification method based on relationship-guided random walk, step S4 specifically includes:
[0024] In the obtained hierarchical heterogeneous network G VHHN On the basis of , combined with the node representation vectors of users and content and the relationship representation vectors between users and content, a relationship-guided random walk is used to sample high-order semantic information, and a random walk sequence containing interactions at different network layers is obtained, as shown in the following formula:
[0025]
[0026] Among them, V Sequence represents a set of random walk sequences involving interactions between different network layers, Indicates V Sequence The first, second, and nth random walk sequences in H e Represents the set of relationship representation vectors between all users and content, H vRepresents the set of representation vectors of all users and content nodes, RGRW represents the relationship-guided random walk, which can distinguish the differences between different types of relationships, and then handle different types of relationships respectively, guide the direction of random walk, and explore the high-order semantic information between users and content on the social platform. In the relationship-guided random walk process, the calculation of the transition probability is shown as follows:
[0027]
[0028] in, Represents the relationship r from node v i To node v j The random walk transition probability of Represents node v k The representation vector of Represents node v i , node v k The characterization vector of the relationship between them, || represents the concatenation of vectors.
[0029] In the above-mentioned social platform target user identification method based on relationship-guided random walk, step S5 specifically includes:
[0030] Step S5.1, based on the representation vectors of users and content nodes, the distance decay factor and attention mechanism are introduced through the semantic fusion component to aggregate all users and content nodes in the random walk sequence, as well as the high-order semantic information contained between them, to obtain the embedding vector of the user node. The calculation formula of the distance decay factor is:
[0031] δ t = exp(-λ·t);
[0032] Among them, δ t For node v t For node v i The decay factor, λ is a hyperparameter, and t represents the distance between the current node and the starting node of the random walk sequence;
[0033] The weight calculation in the attention mechanism is as follows:
[0034]
[0035] Among them, α it represents the attention weight of node vt to node vi, Represents the representation vector of node vt, δ k For node v k For node v i The decay factor of , T represents the total number of random walk steps;
[0036] The process of aggregating user node features containing high-level semantic information is as follows:
[0037]
[0038] in, Represents the node v after aggregation i The embedding vector of node v i is the number of random walk sequences of the starting node, Represents node v t In the random walk sequence User node representation vector;
[0039] Step S5.2, use a multi-layer perceptron to build a social platform target user identification model, take the embedding vector of the user node as the input of the model, and output the importance score of the user node in the social platform, so as to obtain the target user sequence. The social platform target user identification model satisfies the following formula:
[0040]
[0041] Among them, W m and b m represents the weights and biases of the multilayer perceptron, Represents node v i The importance score of
[0042] The recognition model training uses the MSE loss function, as shown below:
[0043]
[0044] Among them, L MSE represents the loss value, N represents the total number of user nodes, y i Represents node v i The actual importance score of Represents the recognition model prediction node v i The importance score of .
[0045] Beneficial effects of the present invention:
[0046] The present invention models the social platform into a heterogeneous network and uses the processing method in the field of heterogeneous networks to quantitatively represent the users, content and their interactions in the social platform, and then identify KOLs through the influence of users. By stratifying the heterogeneous network, the various coupled relationships in the social platform can be classified and processed, and the high-order semantic information hidden in the social platform can be better obtained through relationship-guided random walks, thereby improving the accuracy of KOL identification on the social platform. BRIEF DESCRIPTION OF THE DRAWINGS
[0047] The above and / or additional aspects and advantages of the embodiments of the present invention will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which:
[0048] Figure 1 A schematic diagram of the flow of a method for identifying target users on a social platform based on relationship-guided random walks provided by the present invention. DETAILED DESCRIPTION
[0049] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are 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.
[0050] See also Figure 1 The present invention provides a method for identifying target users of a social platform based on relationship-guided random walk, wherein the target user of the social platform is a KOL. The method comprises steps S1 to S5:
[0051] Step S1, obtaining the interaction data of users and contents in the social platform, and constructing a heterogeneous network based on the relationship between users and contents.
[0052] Wherein, step S1 specifically includes:
[0053] Obtain social platform data, including users, content, and their interactions. Users include readers, authors, and video hosts. Content includes articles and videos. Their interactions include readers following authors and video hosts, readers reading articles and videos, and authors and video hosts publishing articles and videos. Construct a heterogeneous network G based on the relationship between users and content. HN =(V, E, A, P), where V represents the set of all users and content nodes, E represents the set of interactive relationships between users and content, A represents the set of user and content node types, and P represents the set of interactive relationship types between users and content.
[0054] Step S2, stratify the heterogeneous network so that different types of users and content are in different network layers, and construct initial features.
[0055] Wherein, step S2 specifically includes:
[0056] Step S2.1, analyze the types of user and content nodes in the heterogeneous network, stratify the heterogeneous network according to the node types of users and content, and obtain a hierarchical heterogeneous network Where n is the number of all user and content node types, Represents G VHHN Subnetworks at different layers in A 1 , A 2 , A n Indicates different node types;
[0057] Step S2.2, based on the obtained hierarchical heterogeneous network G VHHN Construct the initial feature vector and set the sub-network of each layer to be Among them, V l represents the node set of the lth layer, E l Represents the relationship set of the lth layer. In each sub-network, the Laplacian matrix of the node is extracted The adjacency matrix of the relationship And project them into the same feature space to obtain the initial features, including node v i The initial characteristics and relationship ij The initial characteristics
[0058] In step S3, the constructed initial features are input into the graph convolutional network and the node-edge alternating convolutional neural network to obtain the representation vectors of users, content and the relationships between them.
[0059] Wherein, step S3 specifically includes:
[0060] Step S3.1, for node v i The initial characteristics The node representation vector is obtained using the graph convolutional network according to the following formula:
[0061]
[0062] in, Represents node v i The representation vector of
[0063] Step S3.2, for the relationship e ij The initial characteristics According to the following formula, the interaction function of the node-edge alternating convolutional neural network is used to aggregate the user or content node representations connected to the relationship and the initial features of the relationship itself to obtain the relationship representation vector:
[0064]
[0065] in, Represents node v i , node v j The representation vector of the relationship between Represents node v j The representation vector of .
[0066] Step S4: According to the obtained representation vector, a relation-guided random walk is used to sample high-order semantic information to obtain a random walk sequence containing interactions between different network layers.
[0067] Wherein, step S4 specifically includes:
[0068] In the obtained hierarchical heterogeneous network G VHHN On the basis of , combined with the node representation vectors of users and content and the relationship representation vectors between users and content, a relationship-guided random walk is used to sample high-order semantic information, and a random walk sequence containing interactions at different network layers is obtained, as shown in the following formula:
[0069]
[0070] Among them, V Sequence represents a set of random walk sequences involving interactions between different network layers, Indicates V Sequence The first, second, and nth random walk sequences in H e Represents the set of relationship representation vectors between all users and content, H v Represents the set of representation vectors of all users and content nodes, RGRW represents the relationship-guided random walk, which can distinguish the differences between different types of relationships, and then handle different types of relationships respectively, guide the direction of random walk, and explore the high-order semantic information between users and content on the social platform. In the relationship-guided random walk process, the calculation of the transition probability is shown as follows:
[0071]
[0072] in, Represents the relationship r from node v i To node v j The random walk transition probability of Represents node v k The representation vector of Represents node v i , node v k The characterization vector of the relationship between them, || represents the concatenation of vectors.
[0073] Step S5: In the random walk sequence, the distance decay factor and the attention mechanism are introduced through the semantic fusion component to obtain the user's embedded representation, and the target user of the social platform is identified based on the obtained user's embedded representation.
[0074] Wherein, step S5 specifically includes:
[0075] Step S5.1, based on the representation vectors of users and content nodes, the distance decay factor and attention mechanism are introduced through the semantic fusion component to aggregate all users and content nodes in the random walk sequence, as well as the high-order semantic information contained between them, to obtain the embedding vector of the user node. The calculation formula of the distance decay factor is:
[0076] δ t = exp(-λ·t);
[0077] Among them, δ t For node v t For node v i The decay factor, λ is a hyperparameter, and t represents the distance between the current node and the starting node of the random walk sequence;
[0078] The weight calculation in the attention mechanism is as follows:
[0079]
[0080] Among them, α it Represents node v t For node v i The attention weight, Represents node v t The characterization vector, δ k For node v k For node v i The decay factor of , T represents the total number of random walk steps;
[0081] The process of aggregating user node features containing high-level semantic information is as follows:
[0082]
[0083] in, Represents the node v after aggregation i The embedding vector of node v i is the number of random walk sequences of the starting node, Represents node v t In the random walk sequence User node representation vector;
[0084] Step S5.2, use a multi-layer perceptron to build a social platform target user identification model, take the embedding vector of the user node as the input of the model, and output the importance score of the user node in the social platform, so as to obtain the target user sequence. The social platform target user identification model satisfies the following formula:
[0085]
[0086] Among them, W mand b m represents the weights and biases of the multilayer perceptron, Represents node v i The importance score of the user nodes is calculated; the user nodes are arranged according to the importance score, and the user nodes with higher rankings are more likely to be KOLs.
[0087] The recognition model training uses the MSE loss function, as shown below:
[0088]
[0089] Among them, L MSE represents the loss value, N represents the total number of user nodes, y i Represents node v i The actual importance score of Represents the recognition model prediction node v i The importance score of .
[0090] According to the method for identifying target users of social platforms based on relationship-guided random walks provided by the present invention, by modeling the social platform into a heterogeneous network and using the processing method in the field of heterogeneous networks, the users, content and their interactions in the social platform can be quantitatively represented, and then KOLs can be identified through the influence of users. By stratifying the heterogeneous network, various coupled relationships in the social platform can be classified and processed, and high-order semantic information hidden in the social platform can be better obtained through relationship-guided random walks, thereby improving the accuracy of KOL identification on the social platform.
[0091] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.
[0092] Although the 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 claims and their equivalents.
Claims
1. A method for identifying target users on a social platform based on relationship-guided random walk, characterized in that: The following steps are involved: Step S1, obtaining the interaction data of users and contents in the social platform, and constructing a heterogeneous network based on the relationship between users and contents; Step S2, stratifying the heterogeneous network so that different types of users and content are in different network layers, and constructing initial features; Step S3, inputting the constructed initial features into the graph convolutional network and the node-edge alternating convolutional neural network to obtain the representation vectors of users, content and the relationship between them; Step S4, according to the obtained representation vector, use the relationship-guided random walk to perform high-order semantic information sampling to obtain a random walk sequence containing interactions of different network layers; Step S5: In the random walk sequence, the distance decay factor and the attention mechanism are introduced through the semantic fusion component to obtain the user's embedded representation, and the target user of the social platform is identified based on the obtained user's embedded representation.
2. The method for identifying target users on a social platform based on relationship-guided random walk according to claim 1, characterized in that: Step S1 specifically includes: Obtain social platform data, including users, content, and their interactions. Users include readers, authors, and video hosts. Content includes articles and videos. Their interactions include readers following authors and video hosts, readers reading articles and videos, and authors and video hosts publishing articles and videos. Construct a heterogeneous network G based on the relationship between users and content. HN =(V, E, A, P), where V represents the set of all users and content nodes, E represents the set of interactive relationships between users and content, A represents the set of user and content node types, and P represents the set of interactive relationship types between users and content.
3. The method for identifying target users on a social platform based on relationship-guided random walk according to claim 2, characterized in that: Step S2 specifically includes: Step S2.1, analyze the types of user and content nodes in the heterogeneous network, stratify the heterogeneous network according to the node types of users and content, and obtain a hierarchical heterogeneous network Where n is the number of all user and content node types, Represents G VHHN Subnetworks on different layers, A1, A2, A n Indicates different node types; Step S2.2, based on the obtained hierarchical heterogeneous network G VHHN Construct the initial feature vector and set the sub-network of each layer to be Among them, V l represents the node set of the lth layer, E l Represents the relationship set of the lth layer. In each sub-network, the Laplacian matrix of the node is extracted The adjacency matrix of the relationship And project them into the same feature space to obtain the initial features, including node v i The initial characteristics and relationship ij The initial characteristics 4. The method for identifying target users on a social platform based on relationship-guided random walk according to claim 3, characterized in that: Step S3 specifically includes: Step S3.1, for node v i The initial characteristics The node representation vector is obtained using the graph convolutional network according to the following formula: in, Represents node v i The representation vector of Step S3.2, for the relationship e ij The initial characteristics According to the following formula, the interaction function of the node-edge alternating convolutional neural network is used to aggregate the user or content node representations connected to the relationship and the initial features of the relationship itself to obtain the relationship representation vector: in, Represents node v i , node v j The representation vector of the relationship between Represents node v j The representation vector of .
5. The method for identifying target users on a social platform based on relationship-guided random walk according to claim 4, characterized in that: Step S4 specifically includes: In the obtained hierarchical heterogeneous network G VHHN On the basis of , combined with the node representation vectors of users and content and the relationship representation vectors between users and content, a relationship-guided random walk is used to sample high-order semantic information, and a random walk sequence containing interactions at different network layers is obtained, as shown in the following formula: Among them, V Sequence represents a set of random walk sequences involving interactions between different network layers, Indicates V Sequence The first, second, and nth random walk sequences in H e Represents the set of relationship representation vectors between all users and content, H v Represents the set of representation vectors of all users and content nodes, RGRW represents the relationship-guided random walk, which can distinguish the differences between different types of relationships, and then handle different types of relationships respectively, guide the direction of random walk, and explore the high-order semantic information between users and content on the social platform. In the relationship-guided random walk process, the calculation of the transition probability is shown as follows: in, Represents the relationship r from node v i To node v j The random walk transition probability of Represents node v k The representation vector of Represents node v i , node v k The characterization vector of the relationship between them, || represents the concatenation of vectors.
6. The method for identifying target users on a social platform based on relationship-guided random walk according to claim 5, characterized in that: Step S5 specifically includes: Step S5.1, based on the representation vectors of users and content nodes, the distance decay factor and attention mechanism are introduced through the semantic fusion component to aggregate all users and content nodes in the random walk sequence, as well as the high-order semantic information contained between them, to obtain the embedding vector of the user node. The calculation formula of the distance decay factor is: δ t =exp(-λ·t). Among them, δ t For node v t For node v i The decay factor, λ is a hyperparameter, and t represents the distance between the current node and the starting node of the random walk sequence; The weight calculation in the attention mechanism is as follows: Among them, α it Represents node v t For node v i The attention weight, Represents node v t The characterization vector, δ k For node v k For node v i The decay factor of , T represents the total number of random walk steps; The process of aggregating user node features containing high-level semantic information is as follows: in, Represents the node v after aggregation i The embedding vector of node v i is the number of random walk sequences of the starting node, Represents node v t In the random walk sequence User node representation vector; Step S5.2, use a multi-layer perceptron to build a social platform target user identification model, take the embedding vector of the user node as the input of the model, and output the importance score of the user node in the social platform, so as to obtain the target user sequence. The social platform target user identification model satisfies the following formula: Among them, W m and b m represents the weights and biases of the multilayer perceptron, Represents node v i The importance score of The recognition model training uses the MSE loss function, as shown below: Among them, L MSE represents the loss value, N represents the total number of user nodes, y i Represents node v i The actual importance score of Represents the recognition model prediction node v i The importance score of .