Influential user prediction method based on information dissemination topic perception

Abstract

The present invention discloses an event topic-aware information propagation prediction method. In order to solve the problems of topic perception in information propagation models, heterogeneity of information propagation carrier networks, and poor practicality of current information propagation prediction methods, the present invention first designs an improved LDA model based on a neural network to realize a topic extraction model that is not confused and considers the text semantic context, thereby realizing topic perception of propagated information. In addition, in order to efficiently locate all potentially affected users from a complex heterogeneous network, the present invention utilizes a subgraph formed by locating users who are interested in the current topic in the network, and simulates the information propagation process by constructing a unified feature space and heterogeneous node relationships in the subgraph, thereby simulating the information propagation process of real scenes as much as possible, thereby improving the practicality of the model in terms of model efficiency and accuracy.

Description

Technical Field

[0001] The present invention belongs to information dissemination and prediction technology, and in particular to a dissemination impact user prediction technology. Background Art

[0002] The rapid development of the Internet has enabled the rapid spread of information on the Internet. Social platforms such as Weibo and WeChat, and e-commerce platforms such as Taobao and JD.com have become the main sources of information for people's daily lives. On the one hand, for positive information, such as product information, exploring its propagation rules can help businesses find potential consumers and achieve precision marketing; on the other hand, for the spread of negative information, such as rumors, studying the prediction of these information propagations can help network regulators find potential victims or users with bad intentions in advance. In recent years, more and more researchers have begun to focus on the study of information propagation rules. For example, Yang C, Tang J, Sun M, Cui G, Liu Z. "Multi-scale Information Diffusion Prediction with Reinforced Recurrent Networks". In IJCAI, pp. 4033-4039, 2019Aug10. proposed the use of recurrent neural networks (RNNs) to achieve multi-scale prediction of information propagation at the micro (propagation users) and macro (propagation scale) levels. Wang H, Yang C, Shi C. "Neural Information Diffusion Prediction with Topic-Aware Attention Network" in Proceedings of the 30th ACM International Conference on Information & Knowledge Management, pp. 1899-1908, 2021 Oct 26. A topic-aware information diffusion prediction model based on attention mechanism is proposed. Sun, Ling, Yuan Rao, Xiangbo Zhang, Yuqian Lan, and Shuanghe Yu. "MS-HGAT: Memory-enhanced Sequential Hypergraph Attention Network for Information Diffusion Prediction." (2022). Considering the global dependency of users, personalization, and the dynamics of message propagation, a memory-enhanced sequential hypergraph attention network is proposed to predict the next affected user of information propagation. However, the above scheme still has the following problems:

[0003] (1) The topics of information dissemination are rarely considered. Even if they are considered, no one has proposed a method for extracting clearer and more semantically representative topics.

[0004] (2) The particularity of information propagation in user networks. Past prediction methods have focused too much on the temporal dynamic characteristics of information propagation, but have ignored the carrier of information propagation, the network, and its own characteristics such as heterogeneity and sparsity of user interactions. This makes some simple sequence modeling methods, such as those based on RNN or attention mechanisms, insufficient to capture the laws of information propagation.

[0005] (3) Existing information propagation prediction methods have poor practicality. The representations of users and nodes in the network are not in the same feature space. This information propagation modeling method does not conform to the actual laws of information propagation and is therefore difficult to apply to actual scenarios. Summary of the invention

[0006] The present invention aims to address the problem that current information dissemination prediction models ignore the sparsity of the actual information dissemination network itself, and that users have different sensitivities to news on different topics. The technical problem to be solved is to provide a method for predicting the impact of information dissemination on users that can perceive the topic of the disseminated information and at the same time consider the heterogeneity of the information dissemination carrier network and has greater practicality.

[0007] The technical means adopted by the present invention to solve the above technical problems is a method for predicting users affected by information dissemination theme perception, comprising the following steps:

[0008] S1. News event topic extraction based on improved neural network topic model LDA:

[0009] Different users have different interests in different event topics, so we consider first selecting topic events that users may be interested in. We use an improved LDA as an event topic extraction model to extract news topics. Compared with existing topic-aware methods, although LDA is a simple bag-of-words model, it can extract text topics more accurately and quickly. Therefore, we intend to use a model similar to LDA to obtain the topics of news events.

[0010] Specifically, suppose we finally output K topics, the topic set is represented by T, and a total of M news text topics need to be considered. The optimization goal of the event topic extraction model is J i as follows:

[0011]

[0012] Among them, Perplexity (news i ) represents the i-th news in the data set iThe topic perplexity of the topic is an indicator of the existing LDA model, z i,j Represents the model for news news i The word w appears in j The estimated probability of z i,j+1 Indicates news and w j The adjacent word w j+1 The probability of Represents two adjacent words w j With w j+1 The loss caused by the co-occurrence probability is referred to as co-occurrence loss in the embodiment. The last term of the model represents the mutual information I(a, b) between any two topics a and b in the final set of all topics. In the training process of the event topic extraction model, we minimize the above J i To ensure that the model obtains the minimum topic confusion and ensures that the context information is retained and each topic brings as much more information as possible.

[0013] The improved LDA ensures that the final extracted topic focuses on contextual information by ensuring the co-occurrence probability of two adjacent words, thereby solving the problem that there is actually a semantic association between text contexts, but traditional LDA assumes that each word is independent of each other. In addition, in addition to the optimization of the perplexity in traditional LDA, the model also distinguishes different topics by minimizing the mutual information of different topic distributions, thereby solving the problem that the extraction results of traditional topic extraction models are confused among different topics, that is, the same words are often mixed in different topics.

[0014] The training steps of the event topic extraction model are as follows:

[0015] (1) Initialize hyperparameters α, β;

[0016] (2) Use parameter α to randomly initialize the word distribution TW of K topics, Where d represents the number of all words in the entire dataset. Any row of TW is represented as tw, tw~Dir(α), where Dir(α) represents the Dirichlet distribution. TW is a parameter that the deep model needs to learn and update.

[0017] (3) The event topic extraction model uses the news text news of each training input i Get the topic distribution estimate DT and the word distribution TW of the topic, and repeat this step until the target J is optimized. i Stop the event topic extraction model training when the basic convergence occurs; specifically: for each training news text news i , calculate the frequency of occurrence of each word ρ i , And use β to randomly initialize a current news text newsi The text-topic distribution θ i , Then estimate the probability of each word appearing under the current topic-word and text-topic distribution, and obtain the model's prediction of news news. i The frequency of all words in is estimated to be Z i =θ i ×TW, Estimate the current perplexity, co-occurrence loss and mutual information. According to the optimization goal J of the event topic model i Use stochastic gradient descent to optimize the model parameters. i As the i-th row of the probability distribution Z of each word appearing in each document in the entire dataset, Z = DT × TW, where DT is the estimated topic distribution. i The frequency of all words in is estimated to be Z i =θ i ×TW. Z i is the i-th row of Z, θ i is the i-th row of DT.

[0018] (4) Repeat the previous step until the target J is optimized. i Stop model training when basic convergence occurs.

[0019] (5) The output model estimates the topic distribution DT of all news texts and the word distribution TW of each topic.

[0020] The topic event propagation prediction model is implemented through the following steps:

[0021] S2. Construct a heterogeneous information propagation graph based on event topics. On the one hand, this graph excludes the influence of irrelevant users in the current network, and on the other hand, it locks all users who may be interested in the current topic into a smaller subgraph.

[0022] Specifically:

[0023] (1) Construct an event topic heterogeneous graph G = (V, E), which contains three types of nodes: user node set U, news node R, and topic node T. The event topic heterogeneous graph G contains two types of edges: the edge between the user and the news node, which indicates that the user has interacted with the news, and the weight of the edge is the number of interactions; the edge between the news and the topic node, which indicates that the current news belongs to the topic, and the weight of this type of edge is the value of the corresponding position in the DT matrix.

[0024] (2) News i , assuming that the topic distribution of the topic extraction module is θ i , we take the topic with the highest median value in the distribution as the main topic of the current event propagation, for example, we choose θi The top 10 largest topics in the topic are taken as the main topics of the current event propagation, and this topic set is denoted as T * .

[0025] (3) In G, T * The topic in is the seed node sampling all the neighbor news nodes of the seed node, represented as a set R * The user nodes related to these news are represented as a set U * , thus forming the subgraph G of the current topic * , G * The user nodes included in are the candidate users U for information dissemination. * , G * The corresponding adjacency matrix serves as an input to the graph convolutional neural network GCN.

[0026] S3. Obtain the representation of various nodes of the heterogeneous graph in S2 in the same latent space.

[0027] Specifically:

[0028] (1) Obtain news-related word representations. First, use the open source BERT pre-trained word vectors to obtain the representations of all words in the dictionary;

[0029] (2) Then, the distribution relationship between words and topics is used to obtain the semantic representation of each topic; for each topic, the top-ranked words are selected as representative words, and the pre-trained word vectors of the representative words are aggregated during the aggregation process, and the result is used as the representation of the topic; in this embodiment, the top 15 words are selected as representative words;

[0030] (3) Then, the distribution probability between topics and news, as well as the distribution of interaction relations between news and users, are used to obtain the semantic representation of news and users and output to the other input end of GCN;

[0031] S4. Based on the graph convolutional neural network GCN, simulate the propagation process of the topic event in the network and obtain the high-order representation of the target news and candidate users: GCN takes the heterogeneous subgraph network topology obtained in step S2 and the node features of various nodes in the same latent space obtained in step S3 as input, and outputs the high-order representation h of the target news and candidate users. Target and H * ,

[0032] S5. Calculate the similarity s between the target news and the high-order representations of the candidate users by finding the inner product, and sort the candidate users based on this, and select the top c users with the highest scores as the prediction results. The method for calculating the similarity between high-order representations is a common GCN technology and will not be described here.

[0033] Through steps S2-S5, the information propagation impact user prediction problem is converted into a multi-classification problem. The topic event propagation prediction model uses cross entropy loss as the optimization target and uses stochastic gradient descent SGD to update the model parameters.

[0034] The present invention solves the problem (1) raised by the background technology through step S1, solves the problem (2) raised by the background technology through step S2, and solves the problem (3) raised by the background technology through step S3, so as to achieve the prediction of the influence of the propagation of information topic perception on users in a complex network environment, and simulate the information propagation in the real scene as much as possible to ensure the prediction accuracy.

[0035] The beneficial effects of the present invention are as follows: a new topic extraction method is introduced that does not confuse topics and considers contextual semantic associations, and the method is used to realize information topic-aware propagation prediction, which has the following advantages:

[0036] 1. The present invention proposes a method for predicting users affected by information that can perceive the subject of information dissemination. This method locates the scope of information dissemination in a smaller group of candidate users who are interested in the same subject, and predicts the users who may be affected from these candidate users. Compared with the previous information dissemination prediction model, this method has higher efficiency and practicality.

[0037] 2. The topic extraction model proposed in the present invention is an improved LDA model based on neural network. Compared with the previous topic extraction model, it can extract clearer, more representative topics and keywords that take into account the text context.

[0038] 3. The present invention models the network carrier of information dissemination as a heterogeneous information network containing users, news, and topics, which is more in line with the actual situation of information dissemination. Moreover, unlike previous information dissemination prediction models, we convert the node features in the heterogeneous network into the same feature space, which ensures that in the simulation process of information dissemination, it is as relevant as possible to the information itself being disseminated, avoiding the integration of more irrelevant noise. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 The information topic-aware propagation impact user prediction framework provided for the embodiments. DETAILED DESCRIPTION

[0040] In real life, the information we receive is ever-changing. Information operators want to know which users will be real consumers; network regulators want to know which network users will become victims of current bad information.

[0041] The embodiment designs a prediction model of information topic perception and the impact of information dissemination on users, based on the real law of simulated information dissemination to predict the affected users in advance. The prediction model includes the following: Figure 1 As shown, it includes an event topic extraction model and a topic event propagation prediction model. The training of the event topic extraction model is implemented through step 1, and the training of the topic event propagation prediction model is implemented through steps 2-4.

[0042] The steps are detailed below:

[0043] 1. Event topic extraction

[0044] Different users have different interests in the topics of different events, so we consider first selecting the topic events that users may be interested in. We use an improved LDA event topic extraction model to extract the topics of news. Compared with existing topic-aware methods, although LDA is a simple bag-of-words model, it can extract the topics of texts more accurately and quickly. Therefore, we intend to use a model similar to LDA to obtain the topics of news events. In the traditional LDA model, the document-topic distribution and topic-word distribution are adjusted iteratively to make their values ​​close to the true frequency distribution of words in each document. In this way, there are two disadvantages:

[0045] (1) There is confusion among the topics, that is, the same words are often mixed in different topics;

[0046] (2) There is actually a semantic connection between contexts, but traditional LDA assumes that each word is independent of each other.

[0047] In order to avoid the above situation, we reconstructed an event topic extraction model based on deep learning. In our model, by ensuring the co-occurrence probability of two adjacent words, the final extracted topic focuses on contextual information, thereby solving problem (2) in traditional LDA. In addition, the convergence condition in the deep model is not only the perplexity in traditional LDA, but we achieve the distinction between different topics by minimizing the mutual information of different topic distributions, thereby solving problem (1). Assuming that we finally output K topics, the topic set is represented as T, and a total of M news text topics need to be considered. The optimization goal of our event topic model is as follows:

[0048]

[0049] Among them, Perplexity (news i ) indicates the current news i The topic perplexity, z i,j Represents the model for news news i The word w appears in j The estimated probability of zi,j+1 Indicates news and w j The adjacent word w j+1 The second term in the above formula represents the loss caused by the co-occurrence probability of two adjacent words, referred to as co-occurrence loss. The last term of the model represents the mutual information I(a,b) between any two topics a and b in the final set of all topics. During the training of the model, we minimize the above J i To ensure that the model obtains the least topic perplexity and retains the context information and each topic brings as much more information as possible. Our topic extraction model performs the following steps:

[0050] (1) Initialize hyperparameters α, β.

[0051] (2) Use parameter α to randomly initialize the word distribution TW of K topics, Where d represents the number of all words in the entire dataset. Any row of TW is represented as tw, tw~Dir(α), where Dir(α) represents the Dirichlet distribution. TW is a parameter that the deep model needs to learn and update.

[0052] (3) Perform the following steps for each training text:

[0053] a. Calculate the number of words in the entire data dictionary that appear in the current news text news i The real frequency ρ in i ,

[0054] b. Use β to randomly initialize a current news text news i The text-topic distribution θ i , Assume that the topic distribution of all texts is represented as DT, The i-th row of DT is θ i DT is the parameter that the deep model needs to learn and update.

[0055] c. Estimate the probability of occurrence of each word. The model is used for current news news. i The word w in j The estimated probability is z i,j . Assume that the estimated probability distribution of each word appearing in each document in the entire data set is Z, z i,j represents the value of the jth column of the i-th row of Z, The calculation method of Z is: Z = DT × TW. During the model training process, the model is used to train news i The frequency of all words in is estimated to be Z i =θ i ×TW. Z iis the i-th row of Z,

[0056] d. Calculate the perplexity of the current model estimate Perplexity (news i ), the input of this part is the real word distribution ρ of the text i , and the topic model estimates Z i .

[0057] e. From news i Sampling statistics of each possible co-occurring word pair. In fact, we can ensure that the learned topic takes into account the contextual semantics by maximizing the probability of co-occurring word pairs. In order to obtain a stable co-occurrence probability, we use z i, j logz i,j+1 Represents the probability of two co-occurring words. Maximizing this value minimizes -z i,j logz i,j+1 .

[0058] f. In order to ensure the diversity of each topic, we consider minimizing the mutual information I(a,b) of the distribution of each topic, where a and b are any topic pairs from the topic set T.

[0059] g. Calculate the model optimization target J based on d, e, and f i And use stochastic gradient descent to optimize the model parameters TW and DT.

[0060] h. Stop model training when the overall perplexity of the model basically converges.

[0061] (4) The output model estimates the topic distribution DT of all news texts and the word distribution TW of each topic.

[0062] After the above steps, relatively credible topics of news texts can be obtained.

[0063] 2. Target user positioning

[0064] We designed this module to find all users who may be interested in the current topic and exclude the influence of irrelevant users in the current network. Taking the spread of bad information on social networks as an example, searching the entire network for users who may be affected will inevitably affect efficiency. Therefore, we designed a simpler way to limit the user group that may spread the event by the event topic, and study the specific rules of event spread in this small user group. Specifically, our model performs the following operations:

[0065] (1) Construct an event topic heterogeneous graph G = (V, E), which contains three types of nodes: user node set U, news node R, and topic node T. There are two types of edges in this graph: one is the edge from user node to news node, which indicates that the user has interacted with the news, and the weight of the edge is the number of interactions; the other is the edge from news to topic, which indicates that the current news belongs to the topic, and the weight of this type of edge is the value of the corresponding position in the DT matrix. In order to avoid the influence of irrelevant topics, we set the top 10 topics in the topic distribution of the news as the topic to which the current news belongs. For the convenience of expression, we use E U-R represents the edge between user and news nodes, E R-T Represents the edge between news and topic nodes. In this graph, there are the following relations: V = U ∪ R ∪ T, E = E U-R ∪E R-T .

[0066] (2) News i , assuming that the topic distribution of the topic extraction module is θ i , we take the top 10 topics with the largest median value of the distribution as the main topics of the current event propagation, and denote this topic set as T * .

[0067] (3) In G, T * The topic in is the seed node sampling all the neighbor news nodes of the seed node, represented as a set R * and the user nodes related to these news are represented as a set U * , thus forming the subgraph G of the current topic * , G * The user nodes included in are the candidate users U for information dissemination. * The weight of each edge is represented by the adjacency matrix A. For any edge e p，q , and the corresponding edge weights are as follows:

[0068]

[0069] Among them, A p，q Represents the weight of the edge between any node p and node q, which is the value of the corresponding position in A.

[0070] There are two types of edges. p，q ∈E R-T When the edge e p，q The corresponding weight is the value corresponding to topic a in the topic distribution of news r in the DT matrix calculated in module 1, θ r→a represents the weight of the edge from news r to topic a; when e p,q ∈E U-R When the edge e p,qThe corresponding weight is the number of interactions between user u and news r, α u→r Represents the weight of the edge from user u to news r.

[0071] 3. Prediction of the spread of topic events

[0072] In order to find specific propagation users from candidate users, in the topic event propagation prediction module, the present invention uses a GCN network to simulate the propagation process of events among interested users. The execution process of this part is as follows:

[0073] (1) Obtain word representations related to news. News and other information are mainly conveyed in text, so understanding information is the key to understanding the law of information dissemination among the masses. We first obtain the words contained in all news texts, and the dictionary is represented as the set To. Use the open source BERT pre-trained word vector to obtain the representation of all words in the dictionary.

[0074] (2) Obtain the representation of various nodes in the same latent space. i We still assume that the set of 10 related topics is T * i For each topic, the top 15 words are selected as representative words, and the word set is represented as W * i For G * The initial representation of the three types of nodes, namely user, topic and news, is as follows:

[0075]

[0076] h r =∑ a∈neigber(r) θ r→a h a (3)

[0077] h u =∑ r∈neigber(u) α u→r h r (4)

[0078] where h a ,h r ,h u Respectively represent the topic type node a, a∈T * , news type node r, r∈R * , user type node u, u∈U * Formula (2) indicates that in the current subgraph G * In the example, the representation h of subject a a are the keywords w contained in the corresponding topic in TW j The weighted sum of the representations of tw a→jRepresents the topic a and word w in TW j The corresponding value indicates the keyword w in the current topic j The importance of. Indicates the previous step, for word w j In formulas (3) and (4), a∈neigber(r) represents the * The middle node a is the topic type neighbor node of the news node r. Similarly, r∈neigber(u) represents the news neighbor node of the user node u. As shown in formula (3), for the news node r, it represents h r G * The weighted sum of the topic nodes adjacent to the r node. r→a represents the value corresponding to topic a in the topic distribution of news r in DT calculated in module 1, which also expresses the proportion of topic a in the current news. Similarly, formula (4) expresses the value of user u whose representation h u is the weighted sum of the representations of its neighboring news nodes. u→r represents the interaction weight between user u and news r counted in module 2. Using the above formulas (2), (3), and (4), we can obtain the unified representation h of various nodes in the same latent space, i.e., the news semantic space. a ,h r ,h u ,h a ,hr, For the convenience of representation, we will unify all node representations as H. Among them, |U * |,|R * |,|T * | respectively represent graph G * The number of user, news, and topic nodes in the dataset. Hidden represents the dimension of the latent space.

[0079] (3) Then we use a two-layer GCN to simulate the propagation process of topic events in the network. The input of GCN is G * The corresponding adjacency matrix A and the unified representation of each node H. GCN outputs the high-order representation of each node through learning. We represent the embedding matrix corresponding to each user node after GCN learning as H * , The embedding vector corresponding to the target news after GCN learning is represented as h Target ,

[0080] (4) Prediction of potentially affected users. Suppose we need to know who are the c users who may be affected by the current event. In this step, we will use the high-order representation obtained in the previous step to predict the users that may be affected by the current event. *Select c most likely users from . The selection method is as follows:

[0081] a. Calculate the similarity s between the high-order representation of the target news and the high-order representation of all candidate users. The calculation method is to find the inner product, that is, calculate s = h Target ⊙H * T ,

[0082] b. Then we use softmax to normalize the similarity score. The processed similarity score is expressed as |s|, that is, |s| = softmax(s).

[0083] c. After getting the scores of each user, sort |s| and take the top c users with the highest scores as the prediction results. The predicted users are represented as a set

[0084] 4. Model Optimization

[0085] In this module, we transform the problem of information propagation affecting user prediction into a multi-classification problem. The model uses cross entropy loss as the optimization objective and uses SGD to update the model parameters.

[0086] It can be seen from the above embodiments that the information topic-aware method of the present invention realizes the prediction of the impact of information dissemination on users through topic awareness, thereby improving the efficiency, practicality and accuracy of information dissemination modeling. Specifically, a neural network-based topic extraction model is designed in the present invention to achieve contextual awareness and clearer text topic extraction through mutual information constraints of context and topic distribution. In addition, in order to simulate the real information dissemination process as much as possible, the present invention models the information dissemination carrier as a heterogeneous information network that is only related to the topic, and utilizes the keywords contained in the disseminated information to convert the nodes in the heterogeneous information network to the same semantic space. The above scheme improves the efficiency, accuracy and practicality of the prediction model.

Claims

1. An event-theme-aware information diffusion prediction method, characterized in that: include: S1. Training event topic extraction model: 1) Construct an event topic extraction model as an improved topic model LDA, using the topic perplexity of the news, the loss caused by the co-occurrence probability of two adjacent words in the news, and the mutual information between topics to construct the model optimization target; 2) Initialize the hyperparameters of the training event topic extraction model, the word distribution TW of the topic, and the current news text news i The text-topic distribution of 3) The event topic extraction model uses the news text news of each training input i Get the topic distribution estimate DT and the word distribution TW of the topic, and repeat this step until the target J is optimized. i Stop event topic extraction model training when basic convergence occurs; 4) The trained event topic extraction model receives the input news text and outputs the topic distribution estimate DT and the word distribution TW of the topic; The training of the topic event propagation prediction model is completed through steps S2-S5: S2. Construct heterogeneous information propagation graph based on event topics: 1) Construct an event topic heterogeneous graph. The event topic heterogeneous graph contains three types of nodes: user node set U, news node R, and topic node T; it contains two types of edges: the edge between the user and the news node, the weight of the edge is the number of interactions between the user and the news, and the edge between the news and the topic node, the weight of this type of edge is the value of the corresponding topic position in DT; 2) For news text news i , whose text topic distribution is θ i , choose θ i The largest number of topics in the previous set forms the topic set T * ; 3) In G, T * The topic in is the seed node sampling all the neighbor news nodes of the seed node as the set R * And the user nodes related to these news as set U * , using the set T * , set R * and the set U * The subgraph G that constitutes the current topic * , G * The corresponding adjacency matrix is ​​used as an input to the graph convolutional neural network GCN; S3. Get subgraph G * The representations of various nodes in the same latent space are used as node features and output to the other input end of GCN; S4, GCN receives subgraph G * and node features, outputting high-level representations h of target news and candidate users Target and H * ; S5. Calculate the similarity between the target news and the high-order representations of the candidate users by taking the inner product method, and sort the candidate users based on this, and select a preset number of users with the highest scores as the prediction results.

2. The method according to claim 1, characterized in that: The model optimization objective J used in S1 i as follows: Among them, Perplexity (news i ) represents the i-th news in the data set i The topic perplexity, z i,j Represents the model for news news i The word w appears in j The estimated probability, z i,j+1 Indicates news and w j The adjacent word w j+1 The probability of the second optimization objective represents the probability of two adjacent words w j With w j+1 The last item of the optimization objective represents the mutual information I(a,b) between any two topics a and b in the final set of all topics, where T is the topic set.

3. The method according to claim 1, characterized in that: Step 3) in S1 is specifically: For each training news text news of the input event topic extraction model i , calculate the frequency of occurrence of each word, estimate the probability of each word appearing under the current topic-word and text topic distribution, and obtain the news news i The frequency estimate Z of all words in i =θ i ×TW,Z i As the i-th row of the probability distribution Z of each word appearing in each document in the entire data set, the probability distribution Z of all words appearing in each document is obtained, Z = DT × TW, to obtain the topic distribution estimate DT and the word distribution TW of the topic, and repeat this step until the optimization target J is reached. i Stop event topic extraction model training when basic convergence; θ i is the i-th row of DT.

4. The method according to claim 1, characterized in that: The cross entropy loss is used as the optimization target in the training structure of the topic event propagation prediction model, and stochastic gradient descent SGD is used to update the parameters of the model.

Images