Topic outbreak initial influence detection method based on small sample learning and data enhancement

Abstract

The invention belongs to the technical field of internet application, and relates to a topic outbreak initial influence detection method based on small sample learning and data enhancement, which comprises the following steps: carrying out feature extraction on a user relationship network and user basic information to obtain user feature representation; extracting domain features from user historical information, and obtaining user domain feature representation by using the extracted domain features; performing weighted fusion on the user feature representation and the user domain feature representation based on an attention mechanism; calculating a high-influence prototype vector and a low-influence prototype vector of the user according to the fused feature representation; determining the influence of the current user according to the distance between each user and the user high-influence prototype vector and the distance between each user and the user low-influence prototype vector and a set threshold value; the method can predict whether the behavior of the user at the initial stage of outbreak of a certain topic can promote the development of the topic, so that the trend and the influence range of the topic are predicted, and malignant propagation of the topic can be stopped or topic development can be properly promoted.

Description

Technical Field

[0001] The present invention belongs to the field of Internet application technology, and relates to the analysis of the influence of topic propagation on a social network platform, and in particular to a method for detecting the influence of a topic in the early stage of outbreak based on small sample learning and data enhancement. Background Art

[0002] As the scale of social networks continues to expand, the number of users is huge and the communication is frequent, forming a complex social network. With the development of the Internet, online social networks have become a part of people's daily life. In social networks, when hot events or topics appear, users tend to discuss them widely on the Internet, and relevant information will spread quickly to social networks. This information may promote the development of events or topics, but there may also be false and inflammatory news. If a key user processes the topic and publishes or forwards it, it may have an impact on the public and cause harm to society. In the early stage of the development of hot topics, that is, the peak period of information dissemination, a large number of users participate in discussions and share related content. If we can quickly and accurately trace the source of information and discover the influence of key users, this is crucial to understanding and guiding the development of hot topics.

[0003] Research on the influence of key users in the early stage of topic outbreak can provide important support for social media public opinion management. Online social networks themselves have two sides. While we can quickly enjoy this information and gain benefits through videos, live broadcasts, and publishing articles, some people use hot topics to process topics on the network and then cause negative effects. On the one hand, false news, rumors and other bad remarks will appear on social media, and they will get the attention of the public in this way. Because the spread of hot topics is time-sensitive, these cannot be intervened in time by manual methods. Therefore, the research on information tracing in hot topics is of great help for practical applications. Through the influence analysis of key users, social media platforms can better deal with public opinion risks, provide accurate information push, and effectively manage and guide users' discussions and behaviors. On the other hand, research on the influence of key users in the early stage of topic outbreak is also of great significance for the formulation of marketing and advertising strategies. The outbreak of hot topics is often accompanied by the participation and discussion of a large number of users, which also provides valuable opportunities for corporate and brand promotion. By accurately discovering influential key users, companies and brands can identify potential opinion leaders and brand promoters, and formulate personalized and targeted marketing and advertising strategies. This will help enhance brand awareness and influence, and increase users' awareness and willingness to purchase products or services. In this context, discovering key users has obvious practical significance and application value.

[0004] In recent years, researchers have conducted multi-faceted research on network structure, user attributes, and content characteristics in the field of user influence research on topic networks, and have achieved many excellent results. There are three main parts of research. The first is based on user behavior methods, using users' likes, forwarding, comments and other behaviors on social media to analyze users' behavior patterns and degree of interaction, so as to determine their influence. The second is to use machine learning and data mining techniques to extract features from social network data, and measure the influence strength between users by analyzing forwarding and like behaviors, and then use classification, regression and other methods to predict user influence. The third is the study of information dissemination in social networks, which is crucial to understanding user influence, including analyzing information dissemination paths, dissemination speed, and influence range, so as to discover important nodes. Overall, social network influence discovery involves the mutual influence between users, as well as the mutual influence between users and topic content and information.

[0005] Based on the existing research on social networks, we found that there are still some challenges in discovering the influence of users in the early stage of a topic outbreak:

[0006] 1. User-related data is sparse in the early stage of a topic outbreak. In the early stage of an outbreak, there is not much data and it contains a lot of invalid data that has no or little impact on the spread of the topic, which makes the amount of valid data even more scarce and thus hinders the discovery of user influence.

[0007] 2. The feature space of users and topic content is complex in the early stage of topic outbreak. In the topic propagation space, the features of users, messages, networks, relationships, behaviors, etc. are highly dimensional and complex. This leads to different fields of user topic releases, which makes it difficult to extract features and express them effectively, thus reducing the accuracy of discovering user influence.

[0008] 3. How to discover user influence in the early stage of a topic outbreak. The sparse data and unstable data quality in the early stage of a topic outbreak make the data dissemination process in the early stage of the outbreak somewhat flexible. This increases the difficulty of discovering user influence. Summary of the invention

[0009] In order to predict whether the user's behavior at the initial stage of a topic outbreak will promote the development of the topic, the present invention proposes a method for detecting the influence of a topic outbreak at the initial stage based on small sample learning and data enhancement, which is characterized by comprising the following steps:

[0010] Extract features from the user relationship network and basic user information to obtain user feature representation;

[0011] Extract domain features from user historical information, and use the extracted domain features to obtain user domain feature representation;

[0012] The user feature representation and the user domain feature representation are weightedly fused based on the attention mechanism;

[0013] Calculate the user's high-influence prototype vector and low-influence prototype vector based on the fused feature representation;

[0014] The influence of the current user is determined based on the distance between each user and the user's high-influence prototype vector and low-influence prototype vector and the set threshold. From the perspective of the topic propagation cycle, the present invention analyzes the topic information at the beginning of the topic outbreak, and more accurately predicts the user influence at the beginning of the topic outbreak by extracting some of the user's own characteristics and domain characteristics at the beginning of the topic outbreak and combining them with a small sample classification method. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a flow chart of the method for detecting the influence of a topic in the early stage of outbreak based on small sample learning and data enhancement of the present invention;

[0016] Figure 2 This is a schematic diagram of data enhancement of the present invention;

[0017] Figure 3 Schematic diagram of the TUF2vec method of the present invention;

[0018] Figure 4 Schematic diagram of the classification model of the present invention. DETAILED DESCRIPTION

[0019] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0020] The present invention proposes a method for detecting the influence of a topic in the early stage of outbreak based on small sample learning and data enhancement, which is characterized by comprising the following steps:

[0021] Extract features from the user relationship network and basic user information to obtain user feature representation;

[0022] Extract domain features from user historical information, and use the extracted domain features to obtain user domain feature representation;

[0023] The user feature representation and the user domain feature representation are weightedly fused based on the attention mechanism;

[0024] Calculate the user's high-influence prototype vector and low-influence prototype vector based on the fused feature representation;

[0025] The influence of the current user is determined based on the distance between each user and the user's high-influence prototype vector and low-influence prototype vector and the set threshold.

[0026] like Figure 1 The figure shows the overall process of the present invention, indicating that the input of the present invention is the basic attributes of the user, the user relationship network and the user's historical domain information. After detection, the output of the model is whether the user will have an impact on the topic, thereby controlling the development of the topic as much as possible and providing new opportunities for advertising marketing.

[0027] This embodiment describes the solution of the present invention from three aspects: data acquisition, data enhancement, and influence prediction, which specifically include:

[0028] S1: Get data online. The way to get data is to directly query the topic data in the database provided by the enterprise. What needs to be obtained here is the interaction of users in the topic at the beginning of the topic outbreak (such as comments, likes, interactions, etc.), the user's historical behavior data in history, and the relationship network between users. The historical behavior data includes the number of people followed in the user's historical field, the number of user shares, and the user's historical location tags. The relationship between users requires obtaining basic information such as the number of users' fans and the number of followers, and then simply processing the data.

[0029] As an optional implementation method, this embodiment proposes the following steps of obtaining data and preprocessing:

[0030] S11: Obtaining raw data: Raw data can be obtained through social network public APIs or by directly downloading existing data sources.

[0031] S12: Simple data cleaning. The raw data usually obtained is unstructured and cannot be used directly for data analysis. Most unstructured data can be structured through simple data cleaning. For example, duplicate data can be deleted, invalid information can be cleaned, etc.

[0032] S2: Data enhancement. Extract relevant attributes from user attributes and historical domain data, topic content characteristics, and user relationship networks.

[0033] In the early stage of a topic outbreak, we often face the problem of less user participation. At the same time, during this period, user social media information, user relationship network data, and user basic information are relatively small and diverse, resulting in less effective information that we can obtain. Based on this problem, this paper enhances this data information before finding key users to generate new sample data. Figure 2 The details are as follows:

[0034] S21: Map user basic information to latent space.

[0035] Variational Autoencoder (VAE) is a probability distribution structure that maps input data to a distribution, counted as p θ , where θ is a parameter. A latent variable z is sampled from the distribution, and then a new data x is sampled from z. In this embodiment, the latent variable representation can be described as: from the prior distribution p of the latent variable z θ (z) and sample z from the conditional probability distribution p of x θ (x|z). Each sample has its own Gaussian distribution, and the potential space Z of the variational autoencoder (VAE) is considered as a mixture of infinite z. First, let x = UAttInfrom(u i ), its probability distribution can be expressed as:

[0036] p θ (x) = ∫p θ (x|z)p θ (z)dz(1)

[0037] However, pθ(x) is difficult to calculate, so we introduce a new likelihood function q φ (z|x) to output the possible representation of x, which is equivalent to a probability encoder. Where φ represents the mean and variance parameters in the latent variable, the distribution can be expressed as:

[0038] q φ (z|x)=N(μ,δ 2 )(2)

[0039] Among them, (μ,δ 2 ) is a Gaussian distribution, μ is the mean of the Gaussian distribution, δ 2 is the variance of the Gaussian distribution.

[0040] Because the model has generative capabilities, in order to better express it in the latent space Z in the fitting space, each distribution N is made as close to the standard Gaussian distribution and the normal distribution N(0,1) as possible, and the heavy parameter technique is added in the sampling process. In this model, from N(μ,δ 2 ) is equivalent to sampling an ε from N(0,1), so the latent variable z can be expressed as:

[0041] z=μ+ε×δ(3)

[0042] S22: Mapping latent variables to the original data space.

[0043] This embodiment maps the latent variable z obtained in the hidden space back to the original data space to generate new samples, namely the decoder. The process of generating reconstructed data x' from the latent variable z is the process of solving the joint distribution of the reconstructed data x' and the latent variable z, namely:

[0044] p θ (x|z)(4)

[0045] Then, by measuring the difference between the generated samples and the original input data, and measuring the difference between the distribution of the latent variables and the prior distribution, the parameters of the encoder and decoder are calculated, and the model is updated to obtain better generated data.

[0046] In this embodiment, the difference between the distribution of the latent variables output in the encoder and the standard normal distribution is measured by KL divergence, and the difference between the original basic information of the user and the generated information in the decoder is calculated by the reconstruction loss function. In order to control the smoothness of the latent variables and improve the sample quality, we added L2 regularization in the calculation process, and its formula is expressed as follows:

[0047]

[0048] Among them, q φ (z|x) represents the latent variable distribution, sampling z~q from the latent space φ (z|x); d represents the dimension of the input data.

[0049] At the same time, one of the goals of the variational autoencoder is to minimize the reconstruction loss, that is, to measure the difference between the original data and the decoder-generated data, and to minimize the difference between the original data and the decoder-generated data through reconstruction loss. This paper uses cross entropy loss to measure and express it as:

[0050]

[0051] Among them, x i represents the original input data, Represents the output result of the decoder, and d represents the dimension of the input data.

[0052] According to the description in (6)(7), the loss function of the variational autoencoder can be expressed as:

[0053] L(x)=L recon +β·KL(q φ (z|x)||p θ (z))(7)

[0054] Among them, β represents a hyperparameter that can adjust the weight of KL divergence; p(z) represents the prior distribution of z.

[0055] It is necessary to constantly modify the KL divergence and reconstruction loss between the two to obtain an optimal learning result and minimize the difference in the regenerated data.

[0056] The learned encoder is used to generate new vector representations based on the representations of existing vector nodes, thereby expanding the dataset of node representations.

[0057] S3: Build a model. The present invention proposes a feature representation method of T(Topic)U(User)F(Field)2vec(to vector). First, the features between topics and user fields, and between users are constructed to represent the information. Then, the attention mechanism is introduced to fuse the vector representations of the two. Finally, the prototype network-based method is used to calculate the average value of the feature vector to obtain the prototype vector, analyze the influence of users in the early stage of the topic outbreak, and use this to study the development trend of the topic outbreak in the early stage.

[0058] In the process of the present invention performing influence detection, if Figure 1 , divided into three stages, namely:

[0059] In the first stage, considering the influence of the user's historical field and the user's relationship network on the spread of topics, the TUF2vec algorithm is proposed. Its process can be seen Figure 3 ,First, the interaction between user nodes and user ,relationship networks is fully considered to achieve the representation of user ,features. Then, the domain feature representation of the user is ,mined in the historical behavior information;

[0060] In the second stage, the user relationship features and user historical domain features obtained at the beginning of the outbreak are combined, and the vector representations of the two are fused using the attention mechanism;

[0061] In the third stage, a key element influence discovery model based on small samples was proposed to analyze the influence of user participation behavior in the early stage of topic outbreak, and to detect potential opinion leaders or topic developments based on the above information.

[0062] The user relationship network and basic user information are extracted to obtain user feature representation. u = {U u ,E u} and user basic attribute information set A = {(UAtt,u i |u i∈U}. The feature representation of users is obtained through the correlation between related behaviors and attributes between users. The interactive behaviors between users and the similarity of interests between users are analyzed to determine the edge weights between users, so as to more accurately describe the relationship between users and obtain the node sequence related to users.

[0063] In order to better explore the results of learning representation between nodes, the transition probability between nodes is specified as:

[0064] P(u i ,u j )=W(u i ,u j )×p_q_ratio(u i ,u j )(8)

[0065] Among them, W(u i ,u j ) is the node u i To node u j The weight, p_q_ratio(u i ,u j ) represents the parameters defined according to the characteristics between nodes, which are defined as:

[0066]

[0067] Among them, d(u i ,u j ) represents node u i and Festival j The distance between them, p and q are two hyperparameters; d(u i ,u j )=1, it means node u i and node u j The first-order neighbor of i ,u j )=2, it means node u i and node u j The second-order neighbor of d(u i ,u j )=0, it means node u i Equal to node u j .

[0068] Collect interaction information between users, including the number of words in a user's comments, number of interactions, and number of shares under a topic, and express the degree of interaction between users as:

[0069] Interaction(u i ,u j )=∑w k×NumInter(u i ,u j )(10)

[0070] Among them, w k Indicates the weight coefficient of different interaction methods, which is used to balance the importance of different interaction methods. k (u i ,u j ) represents user u i and user u j The number of interactions between users for this user mode, including u i To u j , and u j To u i Then, the minimum-maximum normalization method is used to map the interaction between users to [0,1], making the data easier to process. The formula can be expressed as:

[0071]

[0072] If there is a following relationship between users, then there is a high probability that the users will have similar interests and hobbies. If the following relationship is connected to establish a user relationship network, the interest similarity can be used as part of the user relationship network weight. The interest similarity between users is expressed as:

[0073]

[0074] Among them, SimilarU i Represents user u i Interest tag collection, | SimilarU i |Signing interest tag collection SimilarU i The number of labels in , A∩B means finding the intersection of set A and set B. The value range of interest similarity is [0,1], where 1 means the interests are exactly the same and 0 means completely different interests.

[0075] In summary, node u i and u j The weight is expressed as:

[0076]

[0077] Where Z is the normalization coefficient; this embodiment uses random walk to generate the context of the node. The TUF2vec method is used to extract features from the user relationship network and the user basic information, and the user node is vectorized and embedded into a low-dimensional space. That is, starting from each node in the graph, a random walk is performed according to the calculated transition probability, and its output is expressed as:

[0078] ListU=[n1,n2,n3,…,n i-1 ,n i ](14)

[0079] ListU represents the vector matrix formed by the user relationship network, which is used as the user feature representation of the present invention. i Represents the vector representation of the i-th user.

[0080] In the process of topic propagation, in order to better discover the influence of key elements, in the early stage of topic outbreak, sparse topic data cannot support accurate discovery of user influence. The present invention has found that the historical fields in the user's historical information have a certain influence on the spread of topics. The higher the similarity between a user's own field and his topic, the more attention the user will get in the field, which proves that the topic is more likely to spread, and the user may be a key element. Therefore, this embodiment mines the relevant field information existing in the user's history from the user's historical information, extracts field features, and obtains relevant information that can be added to the user's historical field to represent the user's field features.

[0081] This embodiment will use user history information and propagation topics to represent it. First, a user-domain graph structure is constructed, and the user's browsing history and historical interaction behavior on a topic and its related topics are used as the weight of the transition probability in the TUF2vec method, and then the method is used to obtain the feature representation of the user's historical domain.

[0082] The weights are expressed as:

[0083]

[0084] Among them, u i represents the i-th user (i.e., user u in the present invention) i , node u i ), d j Indicates the jth field. browseHis(u i ,d j ) represents node u i In the field j The number of browsing records, interCount(u i ,d j ) represents node u i In the field j The historical interaction behaviors in the domain include the number of times a user likes and shares a topic in the domain.

[0085] Finally, this embodiment can obtain the domain feature representation formed under the TUF2vec method as follows:

[0086] ListF=[m1,m2,m3,…,m i-1 ,m i ](16)

[0087] Among them, ListF represents the vector matrix formed by the user-domain network structure, m i Represents the domain feature vector representation of the i-th user.

[0088] By analyzing the feature representations from different angles of user influence detection, we obtain user feature representation and domain feature representation. Considering that the attention mechanism can ignore features that have little impact on the results, the attention mechanism is first used to fuse user node features and domain node features in the influence discovery process.

[0089] First, according to the specific implementation of the attention mechanism, the user features are calculated and the user features and user domain features are weighted averaged. The formula is expressed as:

[0090] ListV=α×ListU+(1-α)×ListF(17)

[0091] Among them, ListV represents the fused feature representation vector, and α represents the attention weight.

[0092] Using the softmax function to calculate the attention weight and process the attention score, the formula can be expressed as:

[0093] α=softmax(score(ListU,ListF))(18)

[0094] Among them, score(ListU, ListF) is the attention score, which can be used to measure the correlation between user features and domain features. In this embodiment, the attention score is calculated by using the dot product attention method, that is, the domain feature representation ListF is used as the key vector and the user feature representation ListU formed by the user relationship network is used as the value vector, which is expressed as The superscript T indicates the transpose of a vector or matrix. k Represents the dimension of the key vector, that is, the dimension of ListF.

[0095] At the same time, if Figure 4As shown in the figure, considering that there are fewer user nodes in the early stage of the outbreak, we use a prototype-based small sample classification method to divide the influence of users into high influence and low influence. The prototype vector for calculating influence is obtained by calculating the average value of the feature representation vectors of users in the same category. The prototype vector of high influence is represented as P1, and the prototype vector of low influence is represented as P2. For each influence category, the average value of the feature representation vectors of all users in the category is calculated as the prototype vector. The formula of the prototype vector after the average value is calculated is expressed as follows:

[0096]

[0097] Among them, N1 is the number of users in the high-influence category, N2 is the number of users in the low-influence category; ListV is the fused feature representation, ListV i 1 ListV represents the fused feature representation of the i-th user in the user set of high-influence categories. i 2 Represents the fused feature representation of the i-th user in the user set of the low-influence category.

[0098] For the user to be classified, calculate the distance between the corresponding fusion feature representation vector ListV and the prototype vector of each influence category. Calculate the distance dist between the user vector x to be classified and the high influence and low influence respectively, and express the distance with high influence as dist1, and the distance with low influence as dist2. Assign the user to the influence category with the closest distance. The present invention sets a threshold threshold to determine the size of the influence, and its formula can be expressed as:

[0099]

[0100] Among them, dist is calculated using the Euclidean distance, dist1 is the distance between the user to be classified and the prototype vector of the high-influence category, dist2 is the distance between the user to be classified and the prototype vector of the low-influence category, and threshold is the threshold. If the distance between the user to be classified and all influence categories is greater than the threshold, it will be classified as unknown. HihgInfluence means high influence, and LowInfluence means low influence.

[0101] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for detecting the influence of a topic in the early stage of outbreak based on small sample learning and data enhancement, characterized in that: The following steps are involved: Extract features from the user relationship network and basic user information to obtain user feature representation; Extract domain features from user historical information, and use the extracted domain features to obtain user domain feature representation; Perform weighted fusion of user feature representation and user domain feature representation based on the attention mechanism; Calculate the user's high-influence prototype vector and low-influence prototype vector based on the fused feature representation; The influence of the current user is determined based on the distance between each user and the user's high-influence prototype vector and low-influence prototype vector and the set threshold.

2. The method for detecting the influence of a topic at the early stage of outbreak based on small sample learning and data enhancement according to claim 1 is characterized in that: Extracting features from the user relationship network and basic user information to obtain user feature representation specifically includes the following steps: Establish the transition probability between nodes according to the weight relationship between nodes and the characteristics between nodes; Starting from each node, a random walk is performed according to the transition probability to obtain the context representation of the corresponding node; The obtained context representation is vectorized using the TUF2vec method to obtain the vector representation of the corresponding node.

3. The method for detecting the influence of a topic at the early stage of outbreak based on small sample learning and data enhancement according to claim 1 is characterized in that: The calculation of the transition probability between nodes includes: P(u i ,u j )=W(u i ,u j )×p_q_ratio(u i ,u j ) Among them, P(u i ,u j ) represents node u i To node u j The transition probability W(u i ,u j ) represents node u i To node u j The weight of p_q_ratio(u i ,u j ) represents node u i With node u j The characteristic parameters between .

4. The method for detecting the influence of a topic at the early stage of outbreak based on small sample learning and data enhancement according to claim 3 is characterized in that: Node u i With node u j The characteristic parameter p_q_ratio(u i ,u j ) is expressed as: Among them, d(u i ,u j ) represents node u i and node u j The distance between them, p and q are two hyperparameters; when d(u i ,u j )=1, it means node u i and node u j The first-order neighbor of i ,u j )=2, it means node u i and node u j The second-order neighbor of d(u i ,u j )=0, it means node u i Equal to node u j .

5. The method for detecting the influence of a topic at the early stage of outbreak based on small sample learning and data enhancement according to claim 3 is characterized in that: Node u i To node u j The weight W(u i ,u j ) calculation includes: Interaction i ,in j )=∑w k ×NumInter(in i ,in j ) Among them, Z is the normalization coefficient; Similarity(u i ,u j ) represents node u i With node u j Similarity of interests between i Represents node u i Interest tag collection, SimilarU j Represents node u j Interest tag collection, | SimilarU i |Signing interest tag collection SimilarU i The number of labels in the set A∩B, which means finding the intersection of set A and set B; Interaction normal (u i ,u j ) is the normalized node u i With node u j Interaction min is the minimum number of activities between nodes, Interaction max is the maximum number of activities between nodes, Interaction (ui,uj) For node u i With node u j Number of interactions; NunInter k (u i ,u j ) represents node u i With node u j The number of interactions using interaction method k, w k Represents the weight coefficient of interaction mode k.

6. The method for detecting the influence of a topic at the early stage of outbreak based on small sample learning and data enhancement according to claim 1 is characterized in that: The process of extracting domain features from user historical information and obtaining domain feature representation using the extracted domain features includes: constructing a user-domain graph structure, using the user's browsing history and historical interaction behaviors on a topic and its related topics as the weight of the transfer probability in the TUF2vec method, and obtaining the feature representation of the user's historical domain using the TUF2vec method.

7. The method for detecting the influence of a topic at the early stage of outbreak based on small sample learning and data enhancement according to claim 6 is characterized in that: The weight calculation of the transition probability in the TUF2vec method includes: Among them, W(u i ,d j ) represents node u i For field d j The weight of; Z is the normalization coefficient; browseHis(u i ,d j ) represents node u i In the field j The number of browsing records; interCount(u i ,d j ) represents node u i In the field j The historical interaction behavior in .

8. The method for detecting the influence of a topic at the early stage of outbreak based on small sample learning and data enhancement according to claim 1 is characterized in that: The process of weighted fusion of user feature representation and user domain feature representation based on the attention mechanism includes: ListV=α×ListU+(1-α)×ListF α=softmax(score(ListU,ListF)) Among them, ListV represents the fused feature representation vector, α represents the attention weight; ListU is the user feature representation; ListF is the user domain feature representation; score(ListU, ListF) is the attention score, d k Indicates the dimension of the key vector.

9. The method for detecting the influence of a topic at the early stage of outbreak based on small sample learning and data enhancement according to claim 1 is characterized in that: When calculating the user's high-influence prototype vector and low-influence prototype vector, the cumulative value of the feature representation after fusion of the classified users is used as the prototype vector of the influence of this type.

Images