Social platform guide topic key node tracing method

Abstract

The invention belongs to the technical field of internet application, and particularly relates to a social platform guide topic key node tracing method, which comprises the following steps: acquiring user data; preprocessing the user data; extracting attribute features of the preprocessed user data; processing the user data by adopting a hybrid random walk strategy to obtain a user hidden relationship; performing sentiment analysis on topic information in the user data to obtain user sentiment features; using a URE2vec algorithm to carry out vectorization representation on the user attribute features, the user hiding relationship and the user emotion features, and carrying out dimensionality reduction on a vectorization representation result to obtain hidden information; inputting the hidden information into a key node traceability model to obtain a topic propagation key node and a topic guide intention; according to the method, a hybrid random walk strategy is adopted, implicit friends with similar preferences are found out for the user through the hybrid random walk strategy by utilizing an explicit relationship and an implicit relationship, and data sparsity is relieved.

Description

Technical Field

[0001] The present invention belongs to the field of Internet application technology, and in particular relates to a key node tracing method for guiding topics on a social platform. Background Art

[0002] In today's digital age, social media has become an important platform for people to obtain information, share opinions and establish connections. With the help of the Internet, the speed and scope of topic information dissemination are faster and wider than before, and the impact is greater. Compared with traditional media such as television, newspapers and radio, social media platforms such as Weibo have obvious advantages. Users can update and share information in real time, and users can easily comment on and forward information, and interact with it such as liking. This increase in participation helps information spread more quickly and makes it easier for users to participate in various discussions. However, information overload and information credibility issues have become important challenges that need to be urgently addressed in the current social network environment. In the social media environment, the guidance of specific topics has a huge influence, which can guide users' attention and behavior, and have a profound impact on public opinion, the market and society.

[0003] In recent years, the research on discovering the influence of key nodes in topics has mainly focused on the following three aspects. First, it is based on the research of network structure, using graph theory methods to analyze the connection relationship between network structure and nodes, and using indicators such as node degree centrality to evaluate node influence. Second, it is based on the method of community discovery, by identifying the community structure, determining nodes with high influence. Third, it is based on user behavior research, through the analysis of user interaction behaviors such as likes, comments and reposts, to measure the strength of influence between users.

[0004] However, in the existing technology, there are still some challenges in tracing the key nodes of user propagation in guiding topics: 1. Hidden relationships among users in topics. The relationships between users in the topic network are complex and diverse. The implicit relationships among users will lead to the loss of node information, making it difficult to obtain accurate node representation, thus affecting the analysis and understanding of the topic network. How to discover and represent implicit relationships among users from the topic network structure is a problem. 2. The concealment of the intentions of the leaders in the topic space. Due to the hidden intentions in the guiding topics, the leaders often deliberately guide or conceal certain information to guide the public. How to mine the intentions of the leaders in the topic has also become a major challenge. 3. How to quantify the influence of users and mine the leaders behind them. Guiding topics are usually proposed by one or some users to achieve certain goals, thereby stimulating various behaviors among ordinary users. How to quantify the influence of key users and mine the leaders is also a major difficulty. Summary of the invention

[0005] In order to solve the above problems in the prior art, the present invention proposes a key node tracing method for guiding topics on a social platform, the method comprising:

[0006] Obtain user data, which includes basic user information, user historical behavior, and topic information; pre-process user data;

[0007] Extract the attribute features of the pre-processed user data; use the hybrid random walk strategy to process the user data to obtain the user hidden relationship; perform sentiment analysis on the topic information in the user data to obtain the user sentiment features;

[0008] The URE2vec algorithm is used to vectorize user attribute features, user hidden relationships, and user emotional features, and the vectorized representation results are reduced in dimension to obtain hidden information.

[0009] Input the hidden information into the key node tracing model to obtain the key nodes of topic propagation and the topic guiding intention.

[0010] Beneficial effects of the present invention:

[0011] The present invention constructs a hybrid random walk strategy, which uses the hybrid random walk strategy and utilizes explicit and implicit relationships to find implicit friends with similar preferences for users, thereby alleviating data sparsity. The present invention adopts the URE2vec algorithm to construct hidden information between user relationships and emotions, and uses the user relationship-emotion feature classification model RE-GCN to analyze the influence of user behavior and emotion on the topic propagation process, thereby accurately identifying key nodes. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 It is the overall flow chart of the present invention;

[0013] Figure 2 A schematic diagram of a hybrid random walk of the present invention;

[0014] Figure 3 Schematic diagram of RE-GCN of the present invention. DETAILED DESCRIPTION

[0015] 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.

[0016] A method for tracing the key nodes that guide topics on social platforms, such as Figure 1As shown, the method includes: obtaining user data, which includes user basic information, user historical behavior and topic information; preprocessing the user data; extracting attribute features of the preprocessed user data; using a hybrid random walk strategy to process the user data to obtain user hidden relationships; performing sentiment analysis on the topic information in the user data to obtain user sentiment features; using the URE2vec algorithm to vectorize user attribute features, user hidden relationships and user sentiment features, and reducing the dimension of the vectorized representation results to obtain hidden information; inputting the hidden information into a key node tracing model to obtain key nodes for topic propagation and topic guiding intentions.

[0017] like Figure 1 The overall process of the present invention is shown, indicating that the input of the present invention is the basic information of users on the social platform, the historical behavior data of users and the basic information of topics, and the output after the traceability model is which are the key propagation nodes. The specific steps include:

[0018] S1: Get data online. The way to get data is to access the Twitter and Weibo datasets. What needs to be obtained here is the basic information of the users participating in the guiding topic, including interests, hobbies, concerns, fans, etc., the historical behavior data of users in the topic (such as browsing, collecting, liking, commenting, and forwarding, etc.), and the content information of the topic itself.

[0019] S2: Extract relevant attributes. By designing a dual-graph hybrid random walk algorithm to mine the potential relationships and hidden information of users in the topic, a more detailed network propagation topology is obtained. Then, the graph attention network is introduced to mine and guide the user emotions in the topic, and the user's emotional feature representation is obtained.

[0020] S3: Build a model. We designed a URE2vec algorithm to represent user relationships and sentiment information as vectors, and then used a graph convolutional neural network to build a guided activity-user relationship sentiment classification model RE-GCN. We analyzed the influence of user nodes on topic propagation during the guided activity propagation process, and classified them accordingly to find the key propagation nodes.

[0021] In this embodiment, obtaining data includes: obtaining raw data. Raw data can be obtained through a social network public API or by directly downloading an existing data source. Simple data cleaning. Usually, the raw data obtained is unstructured and cannot be directly used for data analysis. Most unstructured data can be structured through simple data cleaning. For example, duplicate data can be deleted, invalid information can be cleaned, etc.

[0022] In this embodiment, the attribute characteristics of the pre-processed user data are extracted: in a topic network with guidance, the guide usually has his own purpose and emotional tendency, and uses the hiddenness of user relationships in the network and the hiddenness of the guidance intention to guide the influence. Based on this, this paper extracts the factors that influence the topic from two aspects: user potential relationships and emotional characteristics, as follows:

[0023] S21: extracting user potential relationships;

[0024] The process is mainly divided into three stages: Figure 2 As shown in the figure. In the first stage, a user-user relationship graph is constructed based on the basic attribute characteristics of users and the relationship network structure. In the second stage, a user-topic-user relationship hypergraph is constructed based on the user's historical behavior and topic content information. In the third stage, random walks are performed on the ordinary graph and hypergraph in combination with user similarity to obtain the node sequence of the user's complex relationship.

[0025] In a topic network, the similarity between users can effectively measure the relationship between users. Similar behaviors between users include following, sharing, replying, and mentioning. Generally speaking, the more frequent the interaction between users or the more similar their behaviors are, the higher the similarity between them. In order to calculate user similarity, different weight coefficients are assigned to different user behaviors. Therefore, the definition of user similarity is as follows:

[0026]

[0027] Among them, im(u i ,u j ) represents user u i and user u j The similarity between k (u i ,u j )] represents user u i and user u j The total number of k interactions, Num[Link k u i ] and Num[Link k u j ] represents user u i and u j The total number of interactions with all users in the network under k behaviors. k represents the interaction weight of the kth interaction mode. In the interaction behavior mode, the mention relationship is often closer than the forwarding relationship. k The value is larger.

[0028] According to the user's basic attribute characteristics Fea(u i) Construct a user-user relationship general graph Gg, and the basic attribute characteristics of users are as follows:

[0029] Fea(u i )=[fans(u i ),follows(u i ),field(u i )]

[0030] Among them, fans(u i ) represents user u i fans, follows(u i ) represents user u i attention, filed(u i ) represents user u i The authentication field of this field may be empty. These basic characteristics have a certain influence on users in the process of topic propagation.

[0031] According to the user-topic rating matrix R N×M Constructing the hypergraph G H , taking user u as a super-point and topic t as a hyper-edge, and designing a weight calculation method for super-point and hyper-edge pairs according to the scenario of social recommendation, the scoring formula ImplicitScore(u,i) is defined as follows:

[0032]

[0033] Among them, the matrix R N×M The element R u,t represents the rating of user u on topic t. The rating value can reflect the user's interest in the topic. Two users with high ratings on the same topic have similar preferences. Such a pair of users are potential implicit friends and have a higher frequency of co-occurrence in the random walk sequence.

[0034] In a normal graph, from the current node u i Starting with probability P ij Jump to another user node u j , where P ij With edge weight w ij Proportional to, defined as follows:

[0035]

[0036] Among them, δ(u i ) indicates that the i A collection of directly connected user nodes.

[0037] In the hypergraph, when the random walk reaches a user node u i When you select to jump to the iAny related topic node t j Then jump from the topic node to the node with t j Related other user nodes, jump probability The definition is as follows:

[0038]

[0039] Among them, δ t (u i ) indicates that the topic node is connected to user u i The set of connected user nodes, represents the hyperedge weight.

[0040] In order to utilize both explicit and implicit user relationships, the MixRandom strategy combines the hypergraph and the ordinary graph for random walks. Before transferring from the current node to the next node, a probability parameter α (0<α<1) needs to be introduced to control whether the transfer process is performed on the hypergraph or the ordinary graph, and the ɑ value is determined by the current user u i and the next candidate user u j The specific steps of hybrid random walk are as follows:

[0041] Set a similarity threshold θ, if the user similarity Sim(u i ,u j )>θ, it means that the user similarity is high. The next step is to perform random walks on the ordinary graph to explore the social relationships of similar users. On the contrary, it means that the user similarity is low. We can consider more features such as user behavior to mine implicit relationships. The next step is to perform random walks on the hypergraph. If no new nodes are found during the walk, the termination condition is reached.

[0042] By hybrid random walk, we can get the user relationship node representation F u =[u1,u2,u3,...,u k ], where u k is the node sequence of user k's random walk, indicating the users related to k in the network.

[0043] S22: extracting user emotion features;

[0044] The process is mainly divided into three stages. In the first stage, by analyzing the topic content, the grammatical distance and relative distance between aspects and context words are calculated, so as to achieve effective use of position information. In the second stage, a syntactic dependency tree and a graph attention network are constructed to achieve the fusion of syntactic dependency information. In the third stage, an interactive attention model between aspects and context words is established to achieve coordinated optimization between the two.

[0045] The syntactic dependency tree of the sentence is generated by the dependency parser in the Spacy text processing library. The grammatical distance can be obtained from the syntactic dependency tree. Define d i Aspect word w a and the context word w i The grammatical distance between them, the grammatical distance feature calculation formula is:

[0046]

[0047] According to the above expression, we can get the grammatical distance feature vector S between the aspect and the context word in the sentence = (s1, s2, ..., s n ), where d max Indicates the maximum path length, s i Represents the grammatical distance feature between the aspect and the context word, a represents the position index of the aspect word in the sentence, and there are n words and m words in the sentence. The relative distance can be obtained by the path length of the aspect and the context word in the sentence. The relative distance calculation formula is:

[0048]

[0049] The relative distance feature vector R = (r1, r2, ..., r n ), where r i Relative distance feature between aspects and context words.

[0050] Combining the grammatical distance feature and the relative distance feature to generate a new distance feature representation is:

[0051]

[0052] Among them, p i Represents the distance feature between the aspect and the context word in the sentence. Finally, by introducing the distance feature, the hidden state of the sentence is updated. The newly generated sentence feature representation contains the semantic information and position information of the word. The update formula is:

[0053] h i =p i h i

[0054] The GAT layer updates the node representation through a multi-head attention mechanism to capture the syntactic dependencies between words.

[0055]

[0056] Among them, || represents the concatenation operation, N(i) represents the adjacent nodes of node i, represents the weight coefficient calculated by the kth attention head of the lth layer GAT, W lk and is the weight parameter, represents the hidden state vector of the adjacent node, and LeakyReLU represents the activation function. The output of the k-layer GATs is

[0057] The interaction attention weights β and Φ between aspect word a and context word c are calculated as:

[0058]

[0059]

[0060] Among them, a i and c i is a trainable parameter vector.

[0061] Normalize the weights through the softmax function:

[0062]

[0063] Weighted Representation of Context and Aspects:

[0064]

[0065] Finally, we get the concatenation of aspect and context word representations:

[0066] F e =[c;a]

[0067] In this embodiment, inputting hidden information into the key node tracing model includes: The model is mainly divided into two stages: hidden information representation, generalized behavior discovery and model algorithm design. In the first stage, starting from the complex relationship between user behavior and topic content on the social platform, the URE2vec algorithm is proposed, and the entity is vectorized using knowledge representation and embedded into a low-dimensional dense vector space to obtain a hidden information network. In the second stage, a graph convolutional network classification model RE-GCN based on user relationship and sentiment information is established to analyze the influence of topic activities on users, and the key nodes that affect the spread of topics are discovered by combining the above information.

[0068] Specifically include:

[0069] S31: Hidden information representation: The URE2vec algorithm first uses the knowledge representation learning algorithm to vectorize the user-relationship-emotion feature triples to obtain a vectorized user set. Then, the user set is reduced in dimension, and the users are embedded in a low-dimensional dense vector space to obtain a hidden information network.

[0070] S32: Guided topic key node discovery model based on RE-GCN;

[0071] In the existing research on node influence, few studies have simultaneously considered the complex social relationships and emotional tendencies between users during the topic propagation process. This paper combines the hidden relationships and emotional features mentioned above, uses the attention mechanism to fuse the node vector representations of the two aspects of the network, and proposes an influence discovery model based on hidden relationships and emotional intentions. Figure 3 As shown in Figure 2, this paper defines the final user influence discovery as a binary classification problem, that is, whether the user is influential.

[0072] Through the output of S21 and S22, we can get the vector representation of user relationships in the topic network and the sentiment feature vector representation. The URE2vec method is used to extract information from the topic network and embed it into a low-dimensional dense vector space. The output of the URW2vec algorithm is expressed as:

[0073] F=[f1,f2,f3,...,f k ]

[0074] Among them, F represents the relationship between users in the topic network and the user's emotional feature vector matrix, f k Represents the vector representation of the k-th user node.

[0075] Since social networks are typical non-Euclidean structured data, traditional discrete convolution cannot process them properly. Graph convolutional neural networks (GCN) start from the framework of graph theory and implement convolution operations on graphs. Therefore, the present invention selects GCN to process the guiding topic data, and takes into account the complex relationship and emotional guidance characteristics of guiding topics in communication, and finally proposes a key node discovery model for relational emotional information communication based on RE-GCN. The task goal of this application is to discover the key user nodes in the topic, and to determine whether the user is a key user node by judging the size of the user's influence in the network, which can be converted into a binary classification task. The input of the model in this paper is as follows: feature matrix X = N × F, where N represents the number of user nodes in the topic propagation network, and F is the input feature dimension of each node, and the feature is the information processed by URE2vec. Adjacency matrix under topic propagation.

[0076] In this application, a two-layer graph convolutional neural network with an intermediate Dropout layer is used as the key node classification model. First, the weights and biases are randomly initialized, then X is multiplied by W and the bias is added, and then Multiply. Then, use the ReLU function as the activation function of this layer, and perform the Dropout operation during model training. Finally, use the softmax activation function to represent the convolution output as the probability value of different node classifications. The specific formula is:

[0077]

[0078] Among them, W i is the weight matrix corresponding to the i-th layer network in the graph convolutional network.

[0079]

[0080] Since this article discusses a binary classification problem, let the model output Z = C (a, b | u i ), the specific definition is as follows:

[0081]

[0082] If Y=1, then user u is judged i is the key node with influence in the topic; if Y = 0, then user u i It is a key node that has no influence in the topic.

[0083] The above embodiments further illustrate the purpose, technical solutions and advantages of the present invention in detail. It should be understood that the above embodiments are only preferred implementation modes of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made to the present invention within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for tracing the key nodes of a topic guided by a social platform, characterized in that: include: Obtain user data, which includes basic user information, user historical behavior, and topic information; Preprocess user data; Extract the attribute features of the pre-processed user data; use the hybrid random walk strategy to process the user data to obtain the user's hidden relationship; perform sentiment analysis on the topic information in the user data to obtain the user's sentiment features; The URE2vec algorithm is used to vectorize user attribute features, user hidden relationships, and user emotional features, and the vectorized representation results are reduced in dimension to obtain hidden information. Input the hidden information into the key node tracing model to obtain the key nodes of topic propagation and the topic guiding intention.

2. A method for tracing the key nodes of a topic guided by a social platform according to claim 1, characterized in that: Preprocessing user data includes: obtaining original data, deleting duplicate data and invalid information in the original data, and obtaining structured data.

3. The key node tracing method for guiding topics on a social platform according to claim 1, characterized in that: The attribute characteristics of user data include: user behavior characteristics, social relationship characteristics, and topic content characteristics.

4. The key node tracing method for guiding topics on a social platform according to claim 1, characterized in that: The hybrid random walk strategy is used to process user data, including: constructing a user-user relationship graph based on user attribute characteristics and relationship network structure; constructing a user-topic-user relationship hypergraph based on the user-user relationship graph, user historical behavior and topic information; calculating the similarity between users, and using a hybrid random walk strategy to perform random walks on the user-user relationship graph and the user-topic-user relationship hypergraph based on the similarity to obtain a node sequence of complex user relationships.

5. The method for tracing the key nodes of a topic guided by a social platform according to claim 1, characterized in that: The sentiment analysis of topic information in user data includes: calculating the grammatical distance and relative distance between aspects and context words in topic information; constructing a syntactic dependency tree and a graph attention network, and fusing the grammatical distance of aspect words and the grammatical distance of context words through the syntactic dependency tree and the graph attention network to obtain a grammatical distance feature; fusing the relative distance of aspect words and the relative distance of context words using the syntactic dependency tree and the graph attention network to obtain a relative distance feature; fusing the grammatical distance feature and the relative distance feature to obtain a new distance feature; constructing an interactive attention model between aspects and context words; and inputting the new distance feature into the interactive attention model between aspects and context words to obtain a user sentiment feature.

6. The key node tracing method for guiding topics on a social platform according to claim 1, characterized in that: The URE2vec algorithm is used to vectorize user attribute features, user hidden relationships and user emotional features, including: converting user attribute features, user hidden relationships and user emotional features into user-relationship-emotion feature triplets; using a knowledge learning algorithm to vectorize the user-relationship-emotion feature triplets; and embedding the vectorized representation into a low-dimensional dense vector space to obtain hidden information.

7. The method for tracing the key nodes of a topic guided by a social platform according to claim 1, characterized in that: The key node traceability model adopts the RE-GCN network model.

8. A method for tracing the key nodes of a topic guided by a social platform according to claim 7, characterized in that: The RE-GCN network model processes hidden information by: inputting user information into the RE-GCN network model, extracting node features, relationship features, and sentiment features through multi-layer graph convolution operations; fusing the extracted node features, relationship features, and sentiment features, and embedding them into a low-dimensional dense vector space; and classifying the nodes into key nodes and ordinary nodes based on their feature representation.

Images