A twitter platform account classification method based on knowledge path guidance

Abstract

The application discloses a kind of based on knowledge path guide's tweet platform account classification method, first constructs the social network knowledge graph for tweet platform, then proposes the account similarity index based on knowledge path, calculates the similarity of tweet account based on knowledge path, finally constructs the graph convolution network account classification model guided by knowledge path, completes tweet platform account classification.The method of the application proposes social network knowledge path through the constructed social network knowledge graph of tweet platform, effectively measures the correlation similarity between accounts in the scene of social network, constructs the account similarity index based on knowledge path, innovatively measures the contribution of correlation entity in knowledge path to account similarity, is more consistent with the classification of tweet platform account than existing similarity index, and the tweet platform account classification model guided by knowledge path, the similarity based on knowledge path is used as the weight of feature transmission between accounts, and more accurate account classification effect can be achieved.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of Internet and mobile communication technology, and particularly relates to a Twitter platform account classification method based on knowledge path guidance. BACKGROUND

[0002] A social network refers to an online platform based on Internet and mobile communication technology, connecting people and people, allowing users to establish, share, exchange and expand social relationships in a virtual community. Social networks have attracted widespread attention due to their rich information, and modeling and analyzing these large amounts of heterogeneous information have become a research hotspot.

[0003] Social network analysis is a method of studying the structure, relationships and interaction patterns of social networks. Through social network analysis, hidden information and relationships in social networks can be revealed, helping people better understand the operation and influence of social networks, which is of great significance to individuals, organizations and society. Twitter is a very popular social media platform, and Twitter platform users will post a large number of tweets, which contain a variety of rich text and image heterogeneous information. Twitter platform account classification is an important part of social network analysis research and a research hotspot in the fields of sociology, information science, etc. The purpose of Twitter platform account classification is to classify different accounts into different categories, which helps to understand the characteristics of different categories of accounts and provides a basis for subsequent social network management, recommendation systems, etc. to facilitate content management and advertising of social networks. For the classification problem of Twitter platform accounts, there are currently two main methods: one is based on the content characteristics of the account itself, such as the account's personal introduction and published content; the other is based on the interaction relationship of the account, including the attention relationship network between accounts, likes, tweet retweets, comment interactions, etc.

[0004] However, social networks also contain a large amount of heterogeneous information, such as the interests of the account reflected in the published tweets, the keywords in the tweets that highlight the focus of the tweets, and the events formed by similar topic tweets that reflect the focus of the account, which can help improve the relevance analysis between accounts. However, the existing research has little use of this information, so the relevance calculation of the account is not sufficient enough, and some accuracy will be lost in the account classification problem.

[0005] Combining the related research status at home and abroad can find that the existing method considers single and insufficient information dimension and lacks explainability of the result, and it is difficult to solve the problem of low accuracy of account classification caused by incomplete and uncertain social network data. In the related research of account classification on the Twitter platform, the feature extraction of the account is mostly based on the content information of the account itself, such as the published invention information and the related attribute information of the personal document, or based on the interaction relationship between accounts such as attention, like, forwarding, etc., but the mining of the multi-heterogeneous information contained in the social network is insufficient. SUMMARY

[0006] To solve the above technical problems, the application provides a Twitter platform account classification method based on knowledge path guidance, which fully utilizes the rich heterogeneous information on the social platform, mines the potential complex multi-dimensional correlation between social accounts, and proposes the concept of social network knowledge path based on the constructed complete social network knowledge graph, to effectively measure the correlation similarity between accounts in the social network scene, and further proposes the knowledge path guided Twitter platform account classification method.

[0007] The technical scheme adopted by the application is: a Twitter platform account classification method based on knowledge path guidance, the specific steps are as follows:

[0008] S1, constructing a social network knowledge graph of the Twitter platform;

[0009] S2, calculating the Twitter account similarity based on the knowledge path based on step S1;

[0010] S3, constructing a knowledge path guided graph convolution network account classification model to complete the Twitter platform account classification.

[0011] Further, the step S1 is specifically as follows:

[0012] The social network knowledge graph of the Twitter platform is constructed by using top-down knowledge graph, first constructing the mode layer of the knowledge graph, then extracting entities and relationships from the structured and unstructured data on the Twitter platform according to the mode layer, and completing the construction of the data layer.

[0013] The mode layer includes ontology category and attribute definition and semantic relationship definition.

[0014] Further, the step S2 is specifically as follows:

[0015] KPath represents the social network knowledge path proposed on the basis of the meta-path in the social network scene, and the expression is as follows:

[0016]

[0017] Wherein, A1, A2…∈A represents a knowledge path meta node, indicating a class of nodes with the same properties; and R1, R2…∈R represents a knowledge path meta relationship, indicating a class of edge relationships with the same properties.

[0018] A kind of social account similarity measurement method based on meta path is set, as follows:

[0019] KPathSim represents the similarity of knowledge path, the KPathSim similarity between account x and account y is calculated, and the expression is as follows:

[0020]

[0021] Wherein, M represents a set of knowledge path patterns defined according to prior knowledge, P represents a pattern in M, p represents a specific social network knowledge path under this pattern, C p Indicates the number of links of the central entity in the knowledge graph in the knowledge path.

[0022] Further, the step S3 is as follows:

[0023] A knowledge path guided graph convolution network account classification model is constructed, the initial features of the account and the weighted account relationship network are input, the account prediction result is obtained after inputting two layers of graph convolution network, and the model adopts cross entropy loss function.

[0024] Wherein, the initial features of the account are randomly generated; the initial unweighted account relationship network is constructed through the attention relationship of the account, and then the edge weight is given by calculating the account knowledge path similarity to obtain the weighted account relationship network.

[0025] The account similarity KPathSim based on knowledge path is used as the weight of message passing in GCN, and the final node representation is as follows:

[0026]

[0027]

[0028] Wherein, Indicates the feature of node i in the lth layer, W (l) Indicates the value of the learnable parameter matrix in the lth layer, Indicates the neighbor node set of node i, e ji Indicates the weight of node j connected with node i, M represents a set of knowledge path patterns defined according to prior knowledge, P represents a pattern in M, p represents a specific social network knowledge path under this pattern, C p Indicates the number of links of the central entity in the knowledge graph in the knowledge path instance, and the activation function σ adopts ReLU.

[0029] Then, the correlation similarity between accounts is mined by using the information of the knowledge graph, and M is defined as follows:

[0030]

[0031] Wherein, U represents an account node, L represents a location node, T represents a tweet node, K represents a keyword node, E represents an event node, and H represents a hashtag node. In the relationship type, B represents belonging, F represents following, Q represents publishing, and I represents containing.

[0032] In the above knowledge path mode, two accounts belong to the same location, two accounts follow the same account, two accounts publish the same tweet containing the same keyword, two accounts publish the same tweet related to the same event, two accounts publish the same tweet containing the same hashtag.

[0033] Based on steps S1-S3, the account prediction result is obtained by the knowledge path guided graph convolution network account classification model, and the twitter platform account classification is completed.

[0034] The method of the present application first constructs a social network knowledge graph for the twitter platform, then proposes an account similarity index based on a knowledge path, calculates the twitter account similarity based on the knowledge path, and finally constructs a knowledge path guided graph convolution network account classification model to complete the twitter platform account classification. The method of the present application proposes a social network knowledge path through the constructed twitter platform social network knowledge graph, effectively measures the correlation similarity between accounts in the social network scene, constructs an account similarity index based on the knowledge path, and innovatively measures the contribution of the correlation entity in the knowledge path to the account similarity. The existing similarity index is more consistent with the classification of the twitter platform account, and the knowledge path guided twitter platform account classification model takes the similarity based on the knowledge path as the weight of the feature transmission between accounts, and can achieve more accurate account classification effect. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 The flowchart of the twitter platform account classification method based on the knowledge path guidance of the present application. DETAILED DESCRIPTION

[0036] The present application will be further described below in combination with the drawings and examples.

[0037] As Figure 1As shown, a knowledge path guide-based Twitter platform account classification method flowchart of the present application has the following specific steps:

[0038] S1, constructing a social network knowledge graph of the Twitter platform;

[0039] S2, calculating a knowledge path-based Twitter account similarity based on step S1;

[0040] S3, constructing a knowledge path guide-based graph convolution network account classification model to complete the Twitter platform account classification.

[0041] In the present embodiment, the step S1 is specifically as follows:

[0042] The knowledge graph is logically divided into a schema layer and a data layer, the schema layer is a conceptual model and a logical basis of the knowledge graph, and it standardizes and constrains the data layer. The social network knowledge graph of the Twitter platform in the present embodiment adopts a top-down knowledge graph construction method, that is, the schema layer of the knowledge graph is first constructed, the schema layer contains ontology categories and attribute definitions and semantic relationship definitions as shown in Table 1 and Table 2, then entities and relationships are extracted from structured and unstructured data on the Twitter platform according to the schema layer to complete the construction of the data layer.

[0043] Table 1

[0044]

[0045] Table 2

[0046]

[0047] In the present embodiment, the step S2 is specifically as follows:

[0048] A meta-path refers to a path between two entities in a network under a given node and relationship type mode, and its form is as follows:

[0049]

[0050] Wherein, A1', A2'…∈A' represents a meta-node, which represents a node with the same property; and R1', R2'…∈R' represents a meta-relationship, which represents a connection relationship with the same property.

[0051] According to the definition of the meta-path, the meta-path can contain various types of entities and relationships, so the semantics expressed by each meta-path is different. For the account classification task of the embodiment, and in combination with the analysis of the content of the social network knowledge graph of the Twitter platform constructed in step S1, it can be found that for the social network knowledge graph of the Twitter platform, a special type of meta-path can explicitly express a similarity association between accounts. In this meta-path, the types of head and tail entities and the types of relationships are always symmetrical, and the path length is odd, which can explicitly express a semantic association relationship between accounts in different types of entities. Therefore, for the social network scenario, the embodiment further proposes the concept of social network knowledge path based on the meta-path. KPath represents a social network knowledge path proposed based on the meta-path in the social network scenario, and the expression is as follows:

[0052]

[0053] Wherein, A1, A2…∈A represents a knowledge path meta-node, which represents a class of nodes with the same properties; and R1, R2…∈R represents a knowledge path meta-relationship, which represents a class of edge relationships with the same properties.

[0054] The length of the knowledge path is always odd, and the types of head and tail entities and the types of relationships are always symmetrical, which means that the central entity of the knowledge path can reflect the similarity association between the head and tail entities. For example, a typical knowledge path Can explicitly express that the tweets published by the two accounts belong to the same event, which to some extent reflects the similarity of the content published by the accounts. With the concept of social network knowledge path, the embodiment can further define the similarity between social network accounts.

[0055] The embodiment proposes an innovative social account similarity measurement method based on meta-path, called knowledge path similarity KPathSim, which can capture subtle similarity associations between peer entities.

[0056] The KPathSim similarity between account x and account y is calculated, and the expression is as follows:

[0057]

[0058] Wherein, M represents a set of knowledge path patterns defined according to prior knowledge, P represents a pattern in M, p represents a specific social network knowledge path under this pattern, C p represents the number of links of the central entity in the knowledge graph.

[0059] The definition of KPathSim similarity refers to the idea of the RA (Resource Allocation) index, and the RA index is improved on the basis of the AA (Adamic-Adar) index. Therefore, the embodiment further discusses the concepts of the two indexes.

[0060] The idea of the AA index is that when two nodes are connected to nodes with low similarity, the similarity between them is greater than when they are connected to nodes with high similarity. This is easy to understand. For example, on the social platform Twitter, most people are interested in experts or celebrities in a certain field. If two people are interested in a person with few fans, it means that they may have similar interests or social circles, so their similarity is higher. Therefore, they are likely to be the same type of people. The embodiment analyzes and finds that the average degree of keyword entities and event entities and the like in the social network scenario is very large. Therefore, the embodiment decides to adopt the idea of the RA index and use the form of 1 / C p to decrease the contribution value of the center entity pair of the social network knowledge path to the account similarity, so as to be closer to the real scenario.

[0061] In the embodiment, the step S3 is specifically as follows:

[0062] The graph convolution network (GCN) has shown strong performance in many tasks because the GCN has the advantage over some existing network node representation learning methods in that it not only utilizes the information of the nodes themselves, but also fully utilizes the network structure information of the graph, which enables it to obtain better node representation.

[0063] However, the GCN has two problems when applied to node classification problems. The first problem is that the GCN does not distinguish the difference in similarity between nodes and regards all nodes as equal status. The second problem is that the GCN only considers the network structure factor when doing message passing, that is, only considers the degree information of the nodes to determine the weight of the message passing of the neighbor nodes of a node, and does not consider the semantic correlation between the nodes. Therefore, when the message is passed, the features of the nodes with high similarity are not sufficiently transmitted to each other, and the features of the nodes with low similarity interfere with each other, resulting in that the features of the nodes of different categories are not clearly distinguished, and the classification effect is not good.

[0064] In the message passing process of the GCN, the similarity of the accounts is considered and regularized to finally determine the weight of the node message passing.

[0065] In the conventional graph neural network, the feature representation of a node is as follows:

[0066]

[0067] wherein, represents the feature of node i at the l-th layer, W (l) represents the value of the learnable parameter matrix at the l-th layer, represents the set of neighbor nodes of node i, c ji represents the product of the square root of the node degree, i.e. σ represents an activation function.

[0068] If the weight of the given edge is given, the weighted graph convolutional network node representation is as follows:

[0069]

[0070] wherein, e ji represents the weight of the connection between node j and node i; regularization is performed to prevent gradient explosion, and then:

[0071]

[0072] In this embodiment, the knowledge path-based account similarity KPathSim is taken as the weight of message passing in GCN, and the final node representation is as follows:

[0073]

[0074]

[0075] wherein, M represents a set of knowledge path patterns defined according to prior knowledge, P represents one of the patterns, p represents a specific social network knowledge path under this pattern, C p represents the number of links of the central entity in the knowledge graph in the knowledge path instance, and the activation function σ adopts ReLU.

[0076] In order to fully utilize the information of the knowledge graph to mine the association similarity between accounts, the embodiment defines M as follows:

[0077]

[0078] wherein, U represents an account node, L represents a location node, T represents a tweet node, K represents a keyword node, E represents an event node, and H represents a hashtag node. In the relationship type, B represents belonging, F represents following, Q represents publishing, and I represents containing.

[0079] In the above knowledge path pattern, represents that two accounts belong to the same location, represents that two accounts follow a certain account, representing that a certain tweet published by two accounts contains the same keyword, representing that a certain tweet published by two accounts is associated with the same event, representing that a certain tweet published by two accounts contains the same hashtag.

[0080] Based on steps S1-S3, the account prediction result is obtained through the knowledge path guided graph convolution network account classification model, and the twitter platform account classification is completed.

[0081] To sum up, the method of the present application makes full use of the rich heterogeneous information on the social platform, mines the potential complex multi-dimensional correlation between social accounts, proposes the concept of social network knowledge path on the basis of the constructed complete social network knowledge graph, effectively measures the correlation similarity between accounts in the social network scene, and further proposes a knowledge path guided twitter platform account classification method. The method of the present application has better account classification effect than the existing method due to its advantage in account correlation similarity calculation. In order to overcome the shortcomings of the traditional graph convolution network which regards each node as the same and only considers the network structure characteristics when transmitting messages, the method of the present application proposes a knowledge path guided twitter platform account classification model, takes the similarity based on the knowledge path as the weight of the feature transmission between accounts, and achieves more accurate account classification effect.

[0082] Those skilled in the art will realize that the embodiments described herein are for the purpose of helping the reader understand the principles of the present application and should be understood as not limiting the scope of protection of the present application to such specific statements and embodiments. The present application can have various modifications and changes for those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application should be included in the scope of protection of the claims of the present application.

Claims

1. A tweet platform account classification method based on knowledge path guidance, the specific steps are as follows: S1, constructing a social network knowledge graph of the tweet platform; S2, based on step S1, calculating the similarity of the tweet account based on the knowledge path; S3, constructing a knowledge path guided graph convolution network account classification model, completing the classification of the tweet platform account; The step S3 is specifically as follows: Constructing a knowledge path guided graph convolution network account classification model, inputting the initial features of the account and the weighted account relationship network, obtaining the account prediction result after inputting the two-layer graph convolution network, and the model adopts the cross-entropy loss function; wherein The initial features of the account are randomly generated; the initial unweighted account relationship network is constructed through the attention relationship of the account, and then the edge weight is obtained through the calculation of the account knowledge path similarity to obtain the weighted account relationship network; Knowledge path based account similarity As the weight of message passing in GCN, the final node representation is given as follows: （1）； （2）； in, Represents a node In the Features of the layer The learnable parameter matrix is ​​represented in the th... The value of the layer, Represents a node The set of neighboring nodes, Represents a node With nodes Connection weights, This represents the set of knowledge path patterns defined based on prior knowledge. express One of the patterns, This represents a specific social network knowledge path within this model. This represents the number of links in the knowledge graph to the central entity in this knowledge path instance, and the activation function. Use ReLU; Then, the correlation similarity between accounts is defined by information mining of the knowledge graph As follows: （3）； wherein, define represents an account node, represents a location node, represents a tweet node, represents a keyword node, represents an event node, represents a hashtag node; in the relation type, represents belongs to, represents follows, represents posts, represents contains; In the knowledge path pattern above, indicates that two accounts belong to the same location, indicates that two accounts follow the same account, indicates that two accounts post a tweet containing the same keyword, indicates that two accounts post a tweet relating to the same event, indicates that two accounts post a tweet containing the same hashtag. Based on steps S1-S3, the account prediction result is obtained through the knowledge path guided graph convolution network account classification model, and the classification of the tweet platform account is completed.

2. The method of claim 1, wherein, The step S1 is specifically as follows: The social network knowledge graph of the tweet platform is constructed by adopting the top-down knowledge graph, first constructing the mode layer of the knowledge graph, then extracting entities and relationships from the structured and unstructured data on the tweet platform according to the mode layer, and completing the construction of the data layer; The mode layer includes ontology category, attribute definition and semantic relationship definition. 3.The method of claim 1, wherein, The step S2 is specifically as follows: A social network knowledge path is proposed based on a meta-path in a social network scenario, and an expression is as follows: （4）； wherein, represents a knowledge path meta-node, representing a class of nodes with the same properties; and represents a knowledge path meta-relation, representing a class of edge relations with the same properties; A social account similarity measurement method based on meta path is set, which is specifically as follows: represents the similarity of knowledge paths, and the similarity between the account and the account is calculated as follows: The expression is as follows: （5）； wherein, represents a set of knowledge path patterns defined according to prior knowledge, represents one pattern, represents a specific social network knowledge path under this pattern, represents the number of links of the central entity in the knowledge graph in this knowledge path.

Images