Key opinion leader identification method based on double-layer graph representation network

Abstract

The invention discloses a key opinion leader identification method based on a double-layer graph representation network, and belongs to the field of data mining. The method comprises the following steps: firstly, carrying out graph sampling on a social network by utilizing TLGNet, and respectively inputting a news graph and a user graph into a news feature transmission module and a user feature extraction module; in the feature transmission process of each layer, the features of the news layer are transmitted and fused into the user layer, and after multi-layer feature fusion is carried out, a KOL recognition task is completed. According to the method, through depth information fusion of the double-layer graph, the KOL identification efficiency and precision are effectively improved.

Description

Technical Field

[0001] The present invention belongs to the field of data mining, and particularly relates to a method for identifying key opinion leaders based on a deep learning double-layer graph representation network. Background Art

[0002] With the rapid development of social media, the diversity and complexity of information have brought unprecedented challenges. Social media platforms are filled with a large amount of multimodal information, and this information forms a complex dynamic graph network through user interactions. Due to the real-time nature and diversity of this data, traditional analysis methods face higher processing difficulties and require efficient analysis methods to identify key opinion leaders (KOLs). A KOL refers to an individual or organization with extensive influence and high trust in a specific field or industry, and their opinions and suggestions can significantly influence the behavior and decisions of others. The KOL identification task is usually a node classification problem, aiming to identify nodes with high influence in a specific field by analyzing and mining the content on social media, blogs, news, and other platforms.

[0003] Existing KOL identification methods can be roughly divided into three categories: traditional mathematical modeling methods, machine learning methods, and deep learning methods. Traditional mathematical modeling methods, such as the PageRank algorithm and centrality measurement methods, usually rely on the node relationships in the graph structure and identify KOLs by calculating the centrality of nodes. Although these methods perform well in small-scale social networks, they often have problems of high computational complexity and low efficiency when facing complex dynamic graph data. Machine learning methods, such as decision trees, support vector machines, and collaborative filtering, attempt to automatically identify KOLs by learning patterns from data. However, the performance of these methods highly depends on feature selection and engineering, and different data sets often require different features, which increases the workload of manual tuning. With the rise of deep learning, the limitations of traditional methods have been alleviated, especially when dealing with large-scale and complex data, the automatic feature extraction and pattern recognition capabilities of deep learning models have shown significant advantages.

[0004] To address the deficiencies of existing methods, this study proposes a new node classification model based on a user-news two-layer graph - TLGNet (Two-Layer Graph Network). This model learns the propagation characteristics of news in the news layer, learns the interaction characteristics of users in the user layer, and integrates news features into user signals through inter-layer information transmission in the graph. In this way, TLGNet can effectively combine news content and user behavior, and extract feature information crucial for KOL identification. This two-layer graph model can not only improve the accuracy of the recognition results, but also enhance the model's adaptability to dynamic and multimodal data. Through the transmission of inter-layer information, TLGNet achieves effective information fusion between the news layer and the user layer, thus providing an innovative and more efficient KOL identification method with strong practicality and scalability. Summary of the Invention

[0005] The present invention proposes a node classification model based on a user-news two-layer graph, called TLGNet (Two-Layer Graph Network), for effectively identifying key opinion leaders (KOLs) in social media. This model learns the propagation characteristics of news in the news layer, learns the interaction characteristics of users in the user layer, and through inter-layer information transmission in the graph, integrates news features with high credibility into user signals, thereby improving the accuracy of KOL identification. Specifically, TLGNet first performs graph sampling on the social network, and inputs the news graph and the user graph into the news feature transmission module and the user feature extraction module respectively. During the feature transmission process of each layer, the features of the news layer are transmitted and fused into the user layer. After multiple layers of feature fusion, the KOL identification task is completed. This method effectively improves the efficiency and accuracy of KOL identification through the deep information fusion of the two-layer graph.

[0006] The technical solution of the present invention:

[0007] A method for identifying key opinion leaders based on a two-layer graph representation network, and the steps for implementing the key opinion leader identification method using the TLGNet model are as follows:

[0008] Step 1: Two-layer graph construction

[0009] Step 1.1: Construction of the news layer:

[0010] In the news layer, each node represents a news item, and each node contains the feature information of the news. The feature information of the news includes the title, publication time, and content vector. The edges in the news layer represent the relationships between news items, including the reposting relationship: which represents the chain relationship of a news item being reposted by users, helping to track the propagation path of the news.

[0011] Step 1.2: Construction of the user layer:

[0012] At the user layer, each node represents a user, and each node also contains the user's feature information, which includes user ID, activity, and influence. The edges in the user layer represent the relationships between users, including:

[0013] Follow relationship: It represents the social connection where user A follows user B. The follow relationship can reflect the potential influence propagation path between users.

[0014] Interaction relationship: It represents the interaction behavior between user A and user B, such as comments and likes. The interaction relationship helps to evaluate the interaction frequency and intensity between users.

[0015] Construction of cross-layer connections:

[0016] The edges between the news layer and the user layer represent the interaction behaviors between users and news. Through these edges, the activity and influence of users in the social network can be captured. Specifically, it includes the following types of edges:

[0017] Publication edge: A user publishes news, indicating the user's contribution to content creation.

[0018] Sharing edge: A user shares news, reflecting the spread of the news in the user network.

[0019] Comment edge: A user comments on news, indicating the user's feedback and discussion on the news content.

[0020] Like edge: A user likes news, reflecting the user's recognition and support for the news content.

[0021] By constructing the nodes and edges of the news layer and the user layer, as well as the cross-layer connection edges, a comprehensive two-layer graph structure can be formed. The two-layer graph structure can fully capture the multiple relationships between news and users in the social network, thereby providing rich information support for the identification of KOLs.

[0022] Step 2: Construct the TLGNet model

[0023] The TLGNet model includes four modules: a graph sampling module, a news propagation layer module, a user interaction layer module, and a final classification module;

[0024] Step 2.1: The graph sampling module performs graph sampling on the two-layer graph formed in Step 1, and randomly selects r root nodes from the news node set and the user node set respectively as the initial sampling node set.

[0025] Step 2.2: Define a random walk depth of h and sample nodes through the walk mechanism. Each newly walked node is added to the current sampling node set, thereby forming the final user-news subgraph, which contains the news graph Gt , user graph G u and the connection edge E between the user and the news (t,u) .

[0026] Step 2.3: Input G t and G u into the news dissemination layer module and the user interaction layer module respectively, and use the graph convolutional neural network GCN as an encoder to learn the latent representation to obtain the feature representation. The specific process is as follows:

[0027]

[0028] Among them, and respectively represent the feature representations of the news graph and the user graph in the (l - 1)-th layer network, σ is the activation function, and W l is the weight matrix of the l-th layer;

[0029] Step 2.4: After obtaining the feature representations of the news graph and the user graph in the l-th layer, update the feature representation of the user graph by weighted fusion of the two feature representations, and then use the fused feature representation of the user graph and the feature representation of the news graph before fusion as the input of the next layer network. The process of weighted fusion is:

[0030]

[0031] Among them, α is a hyperparameter, and FFN(*) represents a shallow feed-forward neural network.

[0032] Step 2.5: After feature aggregation of the multi-layer networks in the news dissemination layer module and the user interaction layer module, obtain the final feature representation F u of the user layer. The final classification module is used to perform KOL recognition using F u user layer embedding. The model is trained by supervised learning. The specific method is to use the label (i.e., the mark of KOL or non-KOL) as the supervision signal to optimize the loss function, so that the model can correctly identify users with high influence.

[0033] Advantages of the present invention: In TLGNet, KOL recognition combines the information embedding of the news layer, and uses the features learned from the news dissemination layer and the user interaction layer to form a unique user-aware news embedding. This fusion method ensures that KOL recognition not only depends on user behavior features, but also fully considers the complex interaction relationship between the user and the news, thereby improving the accuracy and robustness of recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 is the overall flowchart of the present invention.

[0035] Figure 2 This is the overall structure diagram of the TLGNet of the present invention. Detailed implementation manners

[0036] The evaluation metrics used in the present invention are specifically as follows:

[0037] To evaluate the overall classification effect of the KOL recognition task, the present invention adopts two widely used evaluation metrics: accuracy (Accuracy, Acc) and F1 score (F1). These metrics help measure the balance between the overall recognition performance, precision, and recall of the model, and are particularly significant in dealing with scenarios of class imbalance. In the experiment, TLGNet has achieved excellent performance in these metrics, demonstrating its effectiveness in the KOL recognition task.

[0038] The TLGNet model of the present invention is implemented based on the PyTorch framework, and the loss function used is the cross-entropy loss function with a regularization penalty term (Cross-Entropy Loss). During the training process, to avoid overfitting, the coefficient λ of the regularization penalty term is set to 0.0004. In terms of the optimizer, the widely used Adam optimizer in the current deep learning field is adopted, the learning rate is set to 0.001, and the weight decay rate (for L2 regularization) is set to 0.00015. The number of training iterations is set to 200 rounds to ensure that the model can fully learn the features in the data. For the graph convolutional network (GCN), the networks of both the news layer and the user layer are set to 5 layers to further improve the model's expressive ability on graph-structured data.

[0039] The experimental operating environment of the present invention is the Ubuntu 20.04 operating system, and the CPU uses an AMD EPYC 7763 64-core processor, which has powerful computing capabilities. In addition, 1 NVIDIA RTX 4090 graphics card is used in the experiment, and each graphics card is equipped with 24GB of video memory.

Claims

1. A key opinion leader identification method based on a double-layer graph representation network, characterized in that, The steps to implement the key opinion leader identification method using the TLGNet model are as follows: Step 1: Construction of a two-layer graph: including the construction of the news layer, the construction of the user layer, and the construction of cross-layer connections; Step 2: Construct the TLGNet model. The TLGNet model includes four modules: a graph sampling module, a news propagation layer module, a user interaction layer module, and a final classification module. The specific implementation steps are as follows: Step 2.1: The graph sampling module performs graph sampling on the two-layer graph formed in Step 1, and randomly selects r root nodes from the news node set and the user node set respectively as the initial sampling node set; Step 2.2: Define a random walk depth of h and sample nodes through the walk mechanism; each newly visited node is added to the current set of sampled nodes, thus forming the final user-news subgraph, which includes the news graph G t , the user graph G u , and the connection edges E between users and news (t,u) ; Step 2.

3. Respectively input G t and G u into the news dissemination layer module and the user interaction layer module, and use the graph convolutional neural network GCN as an encoder to learn the latent representation to obtain the feature representation. The specific process is as follows: Among them, and respectively represent the feature representations of the news graph and the user graph in the (l - 1)-th layer network, σ is the activation function, and W l is the weight matrix of the l-th layer; After obtaining the feature representations of the news graph and the user graph at the l-th layer, the feature representation of the user graph is updated by weighted fusion of the two feature representations, and then the fused feature representation of the user graph and the feature representation of the news graph before fusion are used as the input of the next layer network. The process of weighted fusion is: where α is a hyperparameter, and FFN(*) represents a shallow feed-forward neural network; Step 2.5: After the feature aggregation of multiple layers of networks in the news dissemination layer module and the user interaction layer module, the final feature representation F of the user layer is obtained. u , and the final classification module is used to identify KOLs by using F u for user layer embedding; the model is trained by means of supervised learning.

2. The key opinion leader identification method based on a double-layer graph representation network according to claim 1, wherein The specific method of training through supervised learning is to use the labeled tags of KOL or non-KOL as the supervision signal to optimize the loss function, so that the model can correctly identify users with high influence.

3. The key opinion leader identification method based on a double-layer graph representation network according to claim 1, wherein Construction of the news layer: In the news layer, each node represents a news article, and each node contains the feature information of the news. The feature information of the news includes the title, publication time, and content vector. The edges in the news layer represent the relationships between news articles, including the forwarding relationship: which represents the chain relationship of a news article being forwarded by users and helps to track the propagation path of the news.

4. A method for identifying key opinion leaders based on a double-layer graph representation network according to claim 1, characterized in that Construction of the user layer: In the user layer, each node represents a user, and each node also contains the feature information of the user. The feature information of the user includes the user ID, activity level, and influence; The edges in the user layer represent the relationships between users, including: Follow relationship: which represents the social connection where user A follows user B, and the follow relationship can reflect the potential influence propagation path between users; Interaction relationship: which represents the interaction behavior between user A and user B, and the interaction relationship helps to evaluate the interaction frequency and intensity between users.

5. The key opinion leader identification method based on a double-layer graph representation network according to claim 1, wherein Construction of cross-layer connections: The edges between the news layer and the user layer represent the interaction behavior between users and news articles. Through these edges, the activity level and influence of users in the social network can be captured; Specifically, it includes the following types of edges: Publication edge: A user publishes a news article, indicating the user's contribution in content creation; Sharing edge: A user shares a news article, reflecting the propagation of the news article in the user network; Comment edge: A user comments on a news article, indicating the user's feedback and discussion on the news content; Like edge: A user likes a news article, reflecting the user's recognition and support for the news content; By constructing the nodes and edges of the news layer and the user layer, as well as the edges of cross-layer connections, a comprehensive two-layer graph structure can be formed; The two-layer graph structure can fully capture the multiple relationships between news articles and users in the social network, thereby providing rich information support for the identification of KOLs.

Images