Event propagation popularity prediction method based on graph neural network

Abstract

The present invention discloses a method for predicting the popularity of event propagation based on a graph neural network. For each event text in an event propagation data sample, the feature vector of each user in the social network in the propagation of the event text is determined, and the temporal feature cascade graph corresponding to the event text is determined. A dual graph neural network is used to obtain the temporal user feature sequence of the event text based on the user feature vector and the temporal feature cascade graph. An event propagation popularity prediction model including a temporal convolutional neural network and a summation and pooling module is constructed. The event propagation popularity prediction model is trained using the temporal user feature sequence and popularity value of the event text in the event propagation data sample. The popularity value of the event currently propagating in the social network is predicted using the event propagation popularity prediction model. The present invention combines a graph neural network and a temporal convolutional neural network to improve the performance of event propagation popularity prediction.

Description

Technical Field

[0001] The present invention belongs to the field of network public opinion monitoring and defense technology, and more specifically, relates to a method for predicting the popularity of event propagation based on graph neural networks. Background Art

[0002] We are currently in a modern information society with a huge explosion of various types of data and information and the rapid development of social platforms. Social media has replaced traditional media as the main channel for people to obtain and publish information. How to efficiently obtain the required hot information from short text messages and monitor the spread of events, so as to better help the public understand current hot events and their directions, the event propagation model plays a vital role in this.

[0003] With the development of artificial intelligence, the use of deep learning to analyze online public opinion events has become a trend. Event spread popularity prediction is also known as event evolution trend prediction. When an event is exposed in a social network, online media and netizens discuss the event on social media, which will affect the popularity of the event in reality. In addition, when social users exchange information with each other, they will also have an impact and be influenced by others. Regarding the research on event evolution trend prediction, previous work is mainly divided into dynamic models and intelligent learning models. Among the learning models based on dynamic models, its main features are: (1) predicting the dynamics of topic reading through a modified epidemic model; (2) considering public contact and Weibo participation, a susceptible reading and forwarding immunity (SRFI) model is proposed to predict the overall Weibo event popularity trend at each stage. Among models based on intelligent learning, the main features are as follows: (1) Traditional logistic regression and Bayesian algorithms are commonly used to predict the popularity of online public opinion; (2) Driven by deep learning, neural networks such as BP and LSTM are also used to predict the trend of Weibo event propagation, so as to more accurately grasp the message increment of Weibo events and effectively judge the development trend of events; these methods have certain limitations in capturing time series features and the influence between cascade data users. Monitoring the popularity of event propagation through automated methods, and thus analyzing and monitoring the transmission of event information, will be more helpful in controlling the direction of online public opinion.

[0004] However, existing deep learning methods also face many problems: (1) Traditional models only consider the local structural features of events and are insufficient in capturing the development information of events; (2) The characteristics of Chinese short text data are difficult to express, which will lead to feature loss; (3) The single use of feature encoding and time series methods makes it difficult to capture the cascading effect of event propagation. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the existing technology and provide an event propagation popularity prediction method based on graph neural network, which combines graph neural network and temporal convolutional neural network to enhance the feature extraction ability of social network users and events and improve the performance of event propagation popularity prediction.

[0006] To achieve the above-mentioned object, the event propagation popularity prediction method based on graph neural network of the present invention includes the following steps:

[0007] S1: Get the users included in the social network, record the number of users as N, determine the network topology of the social network based on the mutual attention of users, then get the user information of each user, extract the preset user attribute features from the user information, and form the user attribute feature vector x n,1 , n=1,2,…,N;

[0008] S2: Extract several event propagation data samples from the social network. Each event propagation data sample includes the event text propagated within a preset propagation time T and the users who participated in the propagation.

[0009] For each event text, the following method is used to determine the feature vector of each user in the social network in the spread of the event text:

[0010] Extract the text feature vector for each event text. If the user participates in the dissemination of the event text, the text feature of the event text is used as the text feature vector x of the user. n,2 If the user does not participate in the spread of the event text, then the text feature vector x n,2 Set to the default value, and concatenate the user's attribute feature vector and text feature vector to obtain the user feature vector X n =[x n,1 ,x n,2 ];

[0011] For each event text, the following method is used to determine the temporal feature cascade graph corresponding to the event text:

[0012] Divide the event propagation time T into S time periods t s , s=1,2,…,S, the value of S is set according to actual needs; for each event text, generate a feature cascade graph in each time interval, the specific method is as follows:

[0013] The number of event texts propagated in the event propagation data is M, and the number of event texts propagated in time period t is obtained. s The set of users who participated in the dissemination of the mth event text m=1,2,…,M, user set The users in the cascade graph are used as features The node, the corresponding adjacency matrix Chinese elements Determined using the following formula:

[0014]

[0015] in, Represents a user collection Is there a directed edge from user i to user j? i→j means user j forwarded the mth event text posted by user i, and i=j means user i’s self-connection.

[0016] For each event propagation data, the popularity of each event text in each time period t is determined according to the preset popularity evaluation method. s Popularity value within

[0017] S3: Construct a state propagation graph neural network and an influence propagation graph neural network. Each graph neural network includes all users in the entire social network. The number of layers of both graph neural networks is K. At the kth layer of the state propagation graph neural network, the aggregation of the expected state influence received by each user v from all users in its neighbor set N(v) is expressed as follows:

[0018]

[0019] Where k = 1, 2, ..., K, represents the expected activation influence that user v receives from all users in its neighbor set N(v) when it is at level k, represents the activation state of user v’s neighbor user u at the kth layer, p v Represents the preset self-activation parameters of user v, and They represent the state representation vectors of user v and user u at the kth layer respectively, and StateGNN() represents the preset state gating function;

[0020] At the kth layer of the influence propagation graph neural network, the aggregate of the expected diffusion influence received by each user v from all users in its neighbor set N(v) is expressed as follows:

[0021]

[0022] in, represents the expected diffusion influence that user v receives from all users in its neighbor set N(v) when it is at level k, and a uv represents the attention weight from user u to user v, σ() represents the preset activation function, and Represent the influence representation vectors of user u and user v at the kth layer respectively;

[0023] After K layers of propagation, the expected activation influence received by user v from all users in its neighbor set N(v) at the Kth layer is and expected diffusion impact Fusion is performed to obtain the user's final feature representation vector h v , the fusion calculation formula is as follows:

[0024]

[0025] Where W represents the preset weight, σ′() represents the preset activation function;

[0026] S4: For each event text, its corresponding user feature vector is used as the initial feature representation vector of each user in the dual graph neural network, and then the corresponding user feature vector is used in each time period t according to each event text. s Feature cascade graph within The feature representation vector of each user in the social network is obtained through the dual graph neural network, and the user feature representation matrix of the entire social network is constructed. The user feature representation matrix of each event text in S time periods Constructing a temporal user feature sequence to obtain the event text;

[0027] S5: Construct an event propagation popularity prediction model based on a temporal convolutional neural network, which includes a temporal convolutional neural network and a summing and pooling module. The temporal convolutional neural network is used to extract temporal features from the temporal user feature sequence of the event text. The summing and pooling module performs summing and pooling operations on the temporal features to obtain the popularity value of the event text.

[0028] The temporal user feature sequence of each event text obtained in step S4 is used as the input of the event propagation popularity prediction model, and its popularity value As the desired output, the event propagation popularity prediction model is trained;

[0029] S6: When it is necessary to predict the popularity of an event that is currently being spread on the social network, first collect the event texts spread within the preset spread time T and the users who participated in the spread. For each event text, the same method as in step S2 is used to obtain the feature vector of each user in the social network in the spread of the event text and the temporal feature cascade graph corresponding to the event text. The method in step S4 is used to obtain the temporal user feature sequence of each event text. The temporal user feature sequence is input into the trained event spread popularity prediction model to obtain the predicted popularity value of the event text. Finally, the popularity value of the event is obtained by summing up the popularity values ​​of all event texts of the event.

[0030] The present invention is an event propagation popularity prediction method based on a graph neural network. For each event text in an event propagation data sample, the feature vector of each user in the social network in the propagation of the event text is determined, and the temporal feature cascade graph corresponding to the event text is determined. A dual graph neural network is used to obtain the temporal user feature sequence of the event text according to the user feature vector and the temporal feature cascade graph. An event propagation popularity prediction model including a temporal convolutional neural network and a summation and pooling module is constructed. The event propagation popularity prediction model is trained using the temporal user feature sequence and popularity value of the event text in the event propagation data sample. For events being propagated in the social network, the popularity value is predicted using the event propagation popularity prediction model.

[0031] The present invention has the following beneficial effects:

[0032] 1) The present invention uses a time series feature cascade graph to process event propagation data, which can reflect the cascade propagation characteristics of event propagation and more accurately characterize the characteristics of event propagation;

[0033] 2) The dual graph neural network used in this invention uses an influence graph neural network to capture the state and influence of information dissemination, and a state graph neural network to activate the cascading effect between inactive users, thereby more accurately capturing the cascading effect during event propagation and improving the accuracy of event spread popularity prediction;

[0034] 3) The present invention uses a temporal convolutional neural network to describe the temporal characteristics of the event cascade propagation process, solves the problem of lack of temporal characteristics of event propagation, and further improves the accuracy of event propagation popularity prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 This is a flowchart of a specific implementation of the method for predicting event propagation popularity based on graph neural network of the present invention;

[0036] Figure 2 This is an example of a feature cascade graph;

[0037] Figure 3 1 is a comparison chart of the training time of the model of the present invention and the model of the comparative method in this embodiment;

[0038] Figure 4 3 is a comparison chart of the training loss of the model of the present invention and the comparison method model in this embodiment. DETAILED DESCRIPTION

[0039] The following describes the specific embodiments of the present invention in conjunction with the accompanying drawings so that those skilled in the art can better understand the present invention. It should be noted that in the following description, when detailed descriptions of known functions and designs may dilute the main content of the present invention, such descriptions will be omitted here.

[0040] Example

[0041] Figure 1 This is a flowchart of a specific implementation of the event propagation popularity prediction method based on graph neural network of the present invention. Figure 1 As shown, the specific steps of the event propagation popularity prediction method based on graph neural network of the present invention include:

[0042] S101: Obtaining social network data:

[0043] Get the users included in the social network, record the number of users as N, determine the network topology of the social network based on the mutual attention of users, then obtain the user information of each user, extract the preset user attribute features from the user information, and form the user attribute feature vector x n,1 , n = 1, 2, ..., N. User attribute features can include gender, age, level, and activity level, which can be determined based on actual circumstances. In practical applications, users can also be screened, with pre-set criteria used to remove abnormal users to improve the accuracy of popularity prediction.

[0044] S102: Obtain event propagation data sample:

[0045] Extract several event propagation data samples from social networks. Each event propagation data sample includes the event text and the users who participated in the propagation within a preset propagation time T. Generally speaking, the collected raw event propagation data can be preprocessed based on actual conditions. For example, only event text with a high number of forwarding times can be retained.

[0046] For each event text, the following method is used to determine the feature vector of each user in the social network in the spread of the event text:

[0047] Extract the text feature vector for each event text. If the user participates in the dissemination of the event text, the text feature of the event text is used as the text feature vector x of the user. n,2 If the user does not participate in the spread of the event text, then the text feature vector x n,2 Set to the default value, and concatenate the user's attribute feature vector and text feature vector to obtain the user feature vector X n =[x n,1 ,x n,2 ].

[0048] The text feature extraction method of the event text can be set according to the actual situation. In this embodiment, the text features of the event text are extracted by the following method: first, each sentence in the event text is segmented, and then redundant information is removed, such as stop words and symbols, to obtain the feature word set of the sentence, and the Word2Vec algorithm is used to obtain the word vector of each feature word in the feature word set, and the word vectors of each feature word corresponding to the sentence are spliced ​​to obtain a sentence vector, and the sentence vectors of the event text are spliced ​​to obtain a text feature vector.

[0049] For each event text, the following method is used to determine the temporal feature cascade graph corresponding to the event text:

[0050] In order to observe the time series characteristics of event propagation from a macro perspective, the present invention divides the event propagation data into fine-grained time segments, so as to effectively characterize the microscopic time characteristics of cascade growth. At the same time, in order to capture the cascade effect in the event propagation process, a feature cascade graph is constructed based on the user's forwarding relationship of the event text. Therefore, the present invention divides the event propagation duration T into S time periods t s , s=1,2,…,S, where the value of S is set according to actual needs. For each event text, a feature cascade graph is generated within each time interval. The specific method is as follows:

[0051] The number of event texts propagated in the event propagation data is M, and the number of event texts propagated in time period t is obtained. s The set of users who participated in spreading the mth event text m=1,2,…,M, user set The users in the cascade graph are used as features The node, the corresponding adjacency matrix Chinese elements Determined using the following formula:

[0052]

[0053] in, Represents a user collection Is there a directed edge from user i to user j in , i→j means that user j forwarded the mth event text published by user i, and i=j means the self-connection of user i.

[0054] Figure 2 is an example of a feature cascade graph. Figure 2 As shown in Figure 2, the adjacency matrix can be used to represent the feature cascade graph, thereby showing the propagation process of event text.

[0055] For each event propagation data, the popularity of each event text in each time period t is determined according to the preset popularity evaluation method.s Popularity value within In this embodiment, the forwarding volume, likes and comments of the event text are used as its corresponding popularity evaluation index, and the event text in the time period t s The three evaluation indicators are weighted averaged according to the preset weights, and the obtained results are used as the event text in time period t s Popularity value within.

[0056] S103: Building a Graph Neural Network:

[0057] Graph neural networks are linkist models that rely on information transfer between nodes in a graph to capture dependencies. Graph neural networks, as an important offshoot of graph networks, are a class of deep learning-based methods for processing graph information. Due to their excellent performance and interpretability, graph neural networks have recently become a widely used graph analysis method. Recursive neighborhood aggregation is widely used in graph neural networks, where each node leverages information from its aggregated neighborhood to modify its features. After K iterations of aggregation, the updated representation of each node captures information about the structure and representation of its K-hop neighborhood. The problem of predicting network popularity through event propagation emphasizes the role of the network, namely, the interactions between early active users and potential active users, or even between potential active users themselves. Predicting the cascading effects of network interactions is crucial for accurately predicting the future popularity of online content. In order to effectively capture the cascading effects between these users, the present invention proposes a dual graph neural network to mine this information, including a state propagation graph neural network and an influence propagation graph neural network. The state propagation graph neural network is mainly used to activate the cascading effects between inactive users, and the influence propagation graph neural network is mainly used to capture the state and influence in the information propagation process, thereby effectively capturing the current information and the information propagation of users from their upper layer and the influence of upper-level users.

[0058] The state propagation graph neural network is used to simulate the activation of each user during the cascade effect. Specifically, for a target user v, let its neighbor set be N(v). Since user v is influenced by the active users in its neighbor set N(v), the present invention uses the state propagation graph neural network to model the activity state of each user and then extract the user's state feature representation vector.

[0059] The influence propagation graph neural network is used to simulate the diffusion of interpersonal influence in social networks. Specifically, each user is associated with an influence representation feature vector. The influence representation feature vector of the activated user further spreads to other users as the network structure changes. This diffusion is achieved through the neighborhood aggregation and state mechanism of the graph neural network, resulting in the influence feature representation vector of each user.

[0060] The present invention extracts the feature representation vector of each user by combining the state propagation graph neural network and the influence propagation graph neural network. The specific method is as follows:

[0061] A dual graph neural network is constructed, including a state propagation graph neural network and an influence propagation graph neural network. The nodes of each graph neural network contain all users in the entire social network, and the number of layers of both graph neural networks is K. At the kth layer of the state propagation graph neural network, the aggregate of the expected state influence received by each user v from all users in its neighbor set N(v) is expressed as follows:

[0062]

[0063] Where k = 1, 2, ..., K, represents the expected activation influence that user v receives from all users in its neighbor set N(v) when it is at level k, represents the activation state of user v’s neighbor user u at the kth layer, p v represents the preset self-activation parameter of user v, which indicates the probability that user v is activated by means other than neighbor nodes. and Respectively represent the state representation vectors of user v and user u at the kth layer, and StateGNN() represents the preset state gating function used to calculate the influence between two users. In this embodiment, the state gating function The calculation formula is as follows:

[0064]

[0065] Among them, β (k) 、W (k) They represent two preset weight vectors of the kth layer, and || represents the concatenation of the two vectors.

[0066] At the kth layer of the influence propagation graph neural network, the aggregate of the expected diffusion influence received by each user v from all users in its neighbor set N(v) is expressed as follows:

[0067]

[0068] in, represents the expected diffusion influence that user v receives from all users in its neighbor set N(v) when it is at level k, It represents the influence representation vector of user v’s neighbor user u at the kth layer, and InfluGNN() represents the preset influence gating function.

[0069] In order to obtain the user's influence feature representation vector more accurately, the present invention introduces a self-attention mechanism, that is, the influence propagation graph neural network adopts a graph attention network (Graph Attention Networks, GAT). According to the graph attention mechanism, the feature representation of a node in a graph comes from the information aggregation of each neighboring node, and the final node representation is obtained by continuously learning the information of the neighborhood. However, in the computational engineering of GAT, the attention it expresses is very limited, and the score ranking of attention is an unconditional query node, so that when GAT performs model learning, it will be limited to a certain extent by the performance of attention. In order to solve this problem, the present invention improves the attention coefficient of the GAT model. The main improvement is that when performing attention learning, the features of the user node are also learned, so that the feature information based on the neighboring nodes can be learned more effectively. At this time, the aggregation of the expected diffusion influence received by each user v from all users in its neighbor set N(v) is expressed by the following formula:

[0070]

[0071] Among them, a uv represents the attention weight from user u to user v, σ() represents the preset activation function, and Represent the influence representation vectors of user u and user v at the kth layer respectively.

[0072] After K layers of propagation, the expected activation influence received by user v from all users in its neighbor set N(v) at the Kth layer is and expected diffusion impact Fusion is performed to obtain the user's final feature representation vector h v , the fusion calculation formula is as follows:

[0073]

[0074] Where W represents the preset weight, and σ′() represents the preset activation function.

[0075] After K layers of propagation, the expected activation influence received by user v from all users in its neighbor set N(v) at the Kth layer is and expected diffusion impact The final feature representation vector of the user is obtained by fusion.

[0076] S104: Extracting the temporal user feature sequence of the event text:

[0077] For each event text, its corresponding user feature vector is used as the initial feature representation vector of each user in the dual graph neural network, and then the corresponding user feature vector is used in each time period t according to each event text. s Feature cascade graph within The feature representation vector of each user in the social network is obtained through the dual graph neural network, and the user feature representation matrix of the entire social network is constructed. The user feature representation matrix of each event text in S time periods A time-series user feature sequence is constructed to obtain the event text.

[0078] S105: Build and train an event propagation popularity prediction model:

[0079] In order to capture the time series information in the cascade propagation process, the present invention adopts a temporal convolutional neural network (TCN) to learn and represent the time information in the event propagation process, and constructs an event propagation popularity prediction model based on the temporal convolutional neural network. In addition to the temporal convolutional neural network (TCN), the model also includes a sum pooling module, wherein the temporal convolutional neural network is used to extract temporal features from the temporal user feature sequence of the event text, and the sum pooling module performs sum pooling operations on the temporal features to obtain the popularity value of the event text.

[0080] The temporal user feature sequence of each event text obtained in step S104 is used as the input of the event propagation popularity prediction model, and its popularity value As the expected output, the event propagation popularity prediction model is trained.

[0081] During the training process, the mean relative squared error (MRSE) loss is used as the training loss in this embodiment. It is robust to outliers, smoothness, and differentiability, and can achieve better training results.

[0082] S106: Event propagation popularity prediction:

[0083] When predicting the popularity of an event currently spreading on a social network, the first step is to collect the event text and the users involved in the event text within a preset transmission time T. For each event text, the same method as in step S102 is used to obtain the feature vector of each user on the social network in the event text transmission and the corresponding time series feature cascade graph of the event text. The method in step S104 is then used to obtain the time series user feature sequence for each event text. This time series user feature sequence is then input into the trained event transmission popularity prediction model to obtain the predicted popularity value of the event text. Finally, the popularity value of the event is obtained by summing the popularity values ​​of all event texts.

[0084] In order to better illustrate the technical effect of the present invention, a specific data set is used to experimentally verify the present invention. The data set in this experimental verification adopts the crawled Weibo data set, which contains 10,814 event information, involving about 300,000 users. 1,000 events were extracted from the Weibo data set to predict the popularity of the events, of which 800 events were used as training sets and 200 events were used as test sets. For each event, the number of reposts, likes and comments in the corresponding time period during the propagation process is used as its corresponding popularity evaluation index in this embodiment. Three propagation time lengths were selected for this experimental verification, namely 1 hour, 4 hours and one day, and popularity prediction experiments were carried out respectively.

[0085] Since the present invention is a popularity prediction method based on time series, the following methods are selected as comparison methods among the current popularity prediction methods:

[0086] SEISMIC: SEISMIC is a representative method that attempts to capture the cascading effect. It is an implementation of the Hawkes self-excitation point process, which estimates or approximates the impact of the cascading effect of each generation by the average number of followers of a user.

[0087] DeepCas: DeepCas is one of the most advanced deep representation learning methods currently used for network popularity prediction. It captures the relationships between cascade graphs from an end-to-end perspective. It uses random walks to represent information cascade graphs as ordered sets. To represent each sequence, DeepCas embeds nodes in a recurrent neural network and uses an attention mechanism to reassemble the sequences, thereby capturing the important correlations between information cascade graphs.

[0088] DeepCas is a state-of-the-art deep representation learning method for network-aware popularity prediction. It learns the representation of cascade graphs in an end-to-end manner. Therefore, it can describe the information cascade graph as an ordered set using random walks, and then obtain a representation of each sequence by using node and recurrent neural network embeddings. Attention mechanisms are further used to assemble the representation of the information cascade graph in the sequence.

[0089] CoupledGNN: CoupledGNN leverages early adapters and social networks for popularity prediction. To explicitly capture cascading effects, it models the activation state of a target user and the influence of their neighbors. A couplergGNN is proposed, which uses two coupled graph neural networks to capture the interplay between node activation states and the spread of influence. By stacking graph neural network layers, cascading effects along the network are continuously captured. This approach also leverages graph neural networks for popularity prediction.

[0090] GCN-Net: GCN-Net is implemented based on the GCN graph neural network. Its main feature is to capture the influence of state propagation and influence propagation, and to perform message transmission and domain information aggregation through GCN.

[0091] GCNv2-Net: In GCNv2-Net, GCNv2 is an optimized model based on the GCN model. It uses CNN to learn the extraction of feature points and descriptors, and uses RGB-D depth information and the true relative position of the camera to perform supervised learning through 3D-2D projection relationships.

[0092] GAT-Net: GAT-Net is an implementation based on the GAT graph attention network, which captures the diffusion process and mutual influence between nodes for state propagation and influence propagation.

[0093] GATv2-Net: GATv2-Net is an optimized version of GAT, GATv2, where each node attends to its neighbors and gives its own representation of the query. However, GAT computes a very limited type of attention: the attention score is unconditional on the ranking of query nodes. Because GAT uses a static attention mechanism, there are simple graph problems that GAT cannot express: in controlled problems, static attention prevents GAT from fitting the training data well. To address this limitation, we introduce a simple fix and propose GATv2: a more expressive dynamic graph attention variant of GAT.

[0094] To ensure the accuracy and fairness of the experiment, this experiment used mean relative squared error (MRSE), median relative squared error (mRSE), and mean absolute percentage error (MAPE) to evaluate the effectiveness of the experiment and conduct correlation analysis. Table 1 is a comparison table of the event propagation popularity experimental results of the present invention and the comparative method in this embodiment.

[0095]

[0096] Table 1

[0097] As shown in Table 1, the experimental results of the present invention are generally better than those of the comparison methods. For SEISMIC, since it predicts the infinite popularity information of certain events, only mRSE is used as an indicator for fair comparison. From the experimental results, it can be seen that SEISMIC does not perform well in predicting the popularity of Weibo events in event propagation. Since it only estimates the impact of each generation of cascading effects and estimates them according to the average number of fans, it is easy to deviate from the complex and real situation, so the prediction ability is limited. The experimental results of DeepCas are better than feature-based methods, indicating that deep learning-based methods can effectively and automatically learn the propagation influence in the cascade propagation process, and heuristic learning is better than models based on prior knowledge.

[0098] Compared with the graph-based methods GCN and GAT, the graph-based methods using the machine learning method SEISMIC and the deep learning method DeepCas have better experimental results in terms of overall capture of cascade effects. From the traditional graph convolution algorithm (GCN-Net, GCNv2-Net) and the method based on the graph attention mechanism (GAT-Net, GATv2-Net), it can be seen that the simple use of graph convolution to capture domain information in the event propagation process is still relatively simple. For the capture of the influence between some domains, attention can also effectively capture some information from it. Compared with CoupledGNN and the present invention, although CoupledGNN can also well simulate the influence transfer and state changes between users, the cascade effect between them is also well reflected. However, CoupledGNN is a manifestation of organizing popularity, and does not effectively capture time series information, resulting in the experimental effect being not as good as the present invention.

[0099] In order to compare and analyze the training loss effect and time performance of the model of the present invention, the model structure in the comparison method is used for comparison. Figure 3 This is a comparison chart of the training time of the model of the present invention and the comparison method model in this embodiment. Figure 3As shown, due to the three-layer network structure of the model of the present invention, and each layer of the network captures state information, cascade influence and time series information respectively, the model has better fitting ability. It can be seen that the PTG-Net network has lower overall training time and better performance. During training, the SEISMIC, GCN and GCNv2-Net networks have greater differences in corresponding time changes as the data set is divided, which shows that the model is relatively unstable and will show obvious differences in performance as the structure of the data changes. The overall stability of GAT-Net and GATv2-Net is not much different from that of the present invention, but it is more time-consuming and has higher performance.

[0100] In order to ensure the fairness and rationality of the experiment, the same loss function, learning rate and training batch are used for the model of the present invention and the comparison method model. Figure 4 This is a comparison chart of the training loss of the model of the present invention and the model of the comparative method in this embodiment. Figure 4 As shown, as the model trains, the CoupledGNN model loss converges significantly slower than other models, starting with initial training and converging. Loss values ​​fluctuate more significantly across different datasets. However, the algorithm presented in this paper shows little change in loss across the three time intervals, while converging faster. Therefore, we can infer that the present invention demonstrates better model robustness across different datasets.

[0101] Finally, we experimentally verified the improvements to the attention coefficient of the propagation graph neural network proposed in this invention. To test the proposed model, we used the PTG-Netv1 model, which uses only GCN, and the PTG-Netv2 model, which incorporates a graph attention layer, as comparison models, and compared them with the dual graph neural network model proposed in this example. Table 2 compares the experimental results of event propagation popularity for the proposed model and the two comparison models.

[0102]

[0103] Table 2

[0104] As shown in Table 2, the model of the present invention optimizes the ability to capture important event information in the event propagation process by improving the attention coefficient, achieving better prediction effect of event propagation popularity.

[0105] Although the above describes the illustrative specific embodiments of the present invention to facilitate understanding of the present invention by those skilled in the art, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations using the concepts of the present invention are protected.

Claims

1. A method for predicting event propagation popularity based on graph neural network, characterized in that: The following steps are involved: S1: Get the users included in the social network, record the number of users as N, determine the network topology of the social network based on the mutual attention of users, then get the user information of each user, extract the preset user attribute features from the user information, and form the user attribute feature vector x n,1 , n=1,2,…,N; S2: Extract several event propagation data samples from the social network. Each event propagation data sample includes the event text propagated within a preset propagation time T and the users who participated in the propagation. For each event text, the following method is used to determine the feature vector of each user in the social network in the spread of the event text: Extract the text feature vector for each event text. If the user participates in the dissemination of the event text, the text feature of the event text is used as the text feature vector x of the user. n,2 If the user does not participate in the spread of the event text, then the text feature vector x n,2 Set to the default value, and concatenate the user's attribute feature vector and text feature vector to obtain the user feature vector X n =[x n,1 ,x n,2 ]; For each event text, the following method is used to determine the temporal feature cascade graph corresponding to the event text: Divide the event propagation time T into S time periods t s , s=1,2,…,S, the value of S is set according to actual needs; for each event text, generate a feature cascade graph in each time interval, the specific method is as follows: The number of event texts propagated in the event propagation data is M, and the number of event texts propagated in time period t is obtained. s The set of users V who participated in the dissemination of the mth event text s m , m=1,2,…,M, user set The users in the cascade graph are used as features The node, the corresponding adjacency matrix Chinese elements Determined using the following formula: in, Represents the user set V s m Is there a directed edge from user i to user j? i→j means user j forwarded the mth event text posted by user i, and i=j means user i’s self-connection. For each event propagation data, the popularity of each event text in each time period t is determined according to the preset popularity evaluation method. s Popularity value within S3: Construct a state propagation graph neural network and an influence propagation graph neural network. Each graph neural network includes all users in the entire social network. The number of layers of both graph neural networks is K. At the kth layer of the state propagation graph neural network, the aggregation of the expected state influence received by each user v from all users in its neighbor set N(v) is expressed as follows: Where k = 1, 2, ..., K, represents the expected activation influence that user v receives from all users in its neighbor set N(v) when it is at level k, represents the activation state of user v’s neighbor user u at the kth layer, p v Represents the preset self-activation parameters of user v, and They represent the state representation vectors of user v and user u at the kth layer respectively, and StateGNN() represents the preset state gating function; At the kth layer of the influence propagation graph neural network, the aggregate of the expected diffusion influence received by each user v from all users in its neighbor set N(v) is expressed as follows: Among them, a uv represents the attention weight from user u to user v, σ() represents the preset activation function, and Represent the influence representation vectors of user u and user v at the kth layer, W (k) Represents the preset weight vector of the kth layer; After K layers of propagation, the expected activation influence received by user v from all users in its neighbor set N(v) at the Kth layer is and expected diffusion impact Fusion is performed to obtain the user's final feature representation vector h v , the fusion calculation formula is as follows: Where W represents the preset weight, σ′() represents the preset activation function; S4: For each event text, its corresponding user feature vector is used as the initial feature representation vector of each user in the dual graph neural network, and then the corresponding user feature vector is used in each time period t according to each event text. s Feature cascade graph within The feature representation vector of each user in the social network is obtained through the dual graph neural network, and the user feature representation matrix F of the entire social network is constructed. s m , the user feature representation matrix F of each event text in S time periods s m Constructing a temporal user feature sequence to obtain the event text; S5: Construct an event propagation popularity prediction model based on a temporal convolutional neural network, which includes a temporal convolutional neural network and a summing and pooling module. The temporal convolutional neural network is used to extract temporal features from the temporal user feature sequence of the event text. The summing and pooling module performs summing and pooling operations on the temporal features to obtain the popularity value of the event text. The temporal user feature sequence of each event text obtained in step S4 is used as the input of the event propagation popularity prediction model, and its popularity value As the desired output, the event propagation popularity prediction model is trained; S6: When it is necessary to predict the popularity of an event that is currently being spread on the social network, first collect the event texts spread within the preset spread time T and the users who participated in the spread. For each event text, the same method as in step S2 is used to obtain the feature vector of each user in the social network in the spread of the event text and the temporal feature cascade graph corresponding to the event text. The method in step S4 is used to obtain the temporal user feature sequence of each event text. The temporal user feature sequence is input into the trained event spread popularity prediction model to obtain the predicted popularity value of the event text. Finally, the popularity value of the event is obtained by summing up the popularity values ​​of all event texts of the event.

2. The method for predicting event spread popularity according to claim 1, characterized in that: The method for calculating the popularity value of the event text in step S2 is: using the forwarding volume, number of likes and number of comments of the event text as its corresponding popularity evaluation index, and calculating the popularity of the event text in time period t s The three evaluation indicators are weighted averaged according to the preset weights, and the obtained results are used as the event text in time period t s Popularity value within.

3. The method for predicting event spread popularity according to claim 1, characterized in that: The state gating function in step S3 The calculation formula is as follows: Among them, β (k) 、W (k) They represent the two preset weight vectors of the kth layer respectively, and || represents the concatenation of the two vectors.

Images