Topic burst early influence detection method based on small sample learning and data enhancement

Abstract

The application belongs to the technical field of internet application, and relates to a topic outbreak early stage influence detection method based on small sample learning and data enhancement, comprising: extracting features of 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; weighting and fusing the user feature representation and the user domain feature representation based on an attention mechanism; calculating a user high-influence prototype vector and a user low-influence prototype vector according to the fused feature representation; and determining the influence of a current user according to the size between the distance between each user and the user high-influence prototype vector and the user low-influence prototype vector and a set threshold value. The application can predict whether the behavior of a user at the early stage of a topic outbreak will promote the development of the topic, thereby predicting the trend and influence range of the topic, and can stop the vicious spread of the topic or appropriately promote the development of the topic.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of Internet application, and relates to topic propagation influence analysis of a social network platform, in particular to a topic outbreak initial influence detection method based on small sample learning and data enhancement. BACKGROUND

[0002] With the continuous expansion of social networks, the number of users is large and the communication is frequent, forming a complex social network. With the development of the Internet, online social networks have become part of people's daily life. In social networks, when a hot event or topic appears, users often discuss it extensively on the network, and related information will quickly spread on the social network. These information may drive the development of events or topics, but there may also be false and inciting news. If a key user processes and publishes or forwards the topic, it may have an impact on the public and cause harm to society. In the initial stage of the development of a hot topic, that is, the peak period of information dissemination, a large number of users participate in the discussion and share related content. If we can quickly and accurately trace the source of the information and discover the influence of key users. This is crucial for understanding and guiding the development of hot topics.

[0003] Research on key user influence discovery in the initial stage of topic outbreak can provide important support for social media public opinion management. Online social networks have two sides. On the one hand, we can quickly enjoy these information and obtain benefits through videos, live broadcasts, and publishing articles, etc. On the other hand, some people process and cause negative effects on topics by taking advantage of hot topics in the network. On the one hand, false news, rumors and other undesirable speech will appear on social media, and they will get the attention of the public through this way. Because the spread of hot topics has timeliness, these cannot be intervened in time by artificial means. Therefore, the research on information tracing in hot topics is of great help to practical application. Through the analysis of the influence of key users, social media platforms can better cope with public opinion risks, provide accurate information push, and effectively manage and guide user discussion and behavior. On the other hand, research on key user influence discovery in the initial 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 a large number of user participation and discussion, which also provides valuable opportunities for enterprise and brand promotion. By accurately discovering key users with influence, enterprises and brands can identify potential opinion leaders and brand promoters, and develop personalized and targeted marketing and advertising strategies. This will help to improve brand awareness and influence, and increase user 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 carried out multi-aspect research on network structure, user attributes and content characteristics in the field of topic network user influence research, and have achieved a lot of excellent results. There are three parts of research, one is based on user behavior method, using user's like, forward, comment and other behaviors on social media, analyzing user's behavior mode and interaction degree, so as to judge their influence size. Two is to use machine learning and data mining technology, extract features from social network data, and measure the influence strength between users by analyzing forwarding and liking behavior, and then use classification, regression and other methods to predict the influence of users. Three is the research on information dissemination in social network, which is very important to understand user influence, including analyzing information dissemination path, dissemination speed, influence range and other aspects, so as to find important nodes. Overall, social network influence discovery involves the mutual influence between users, and the mutual influence between users and topic content, information.

[0005] According to the existing research of social network, it is found that there are still some challenges in discovering the influence of users in the initial stage of topic outbreak:

[0006] 1. The data related to users in the initial stage of topic outbreak is sparse. In the initial stage of outbreak, there are not many data and many invalid data which have little or no influence on topic dissemination, which makes the amount of effective data become more scarce and hinders the discovery of user influence.

[0007] 2. The user and topic content feature space in the initial stage of topic outbreak is complex. In the topic dissemination space, user, message, network, relationship, behavior and other features have high dimensionality and complexity. This leads to different fields of user topic publishing, which brings difficulties to feature extraction and effective expression, thus reducing the accuracy of user influence discovery.

[0008] 3. How to discover user influence in the initial stage of topic outbreak. The data sparsity and unstable data quality characteristics of the initial stage of topic outbreak make the data dissemination process in the initial stage of outbreak have certain flexibility. This increases the difficulty of discovering user influence. SUMMARY

[0009] In order to predict whether the user's behavior in the initial stage of a topic outbreak will promote the development of the topic, the present application proposes a topic outbreak initial influence detection method based on small sample learning and data enhancement, characterized in that it comprises the following steps:

[0010] Feature extraction is performed on the user relationship network and user basic information to obtain user feature representation;

[0011] Domain features are extracted from user historical information, and user domain feature representation is obtained by using the extracted domain features;

[0012] The user feature representation and the user field feature representation are weighted and fused based on an attention mechanism;

[0013] The user high-influence prototype vector and the user low-influence prototype vector are calculated according to the fused feature representation;

[0014] The influence of the current user is determined according to the size between the distance between each user and the user high-influence prototype vector and the user low-influence prototype vector and a set threshold. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 The flowchart of the topic burst initial influence detection method based on small sample learning and data enhancement of the present application is shown in the figure.

[0016] Figure 2 The data enhancement schematic diagram of the present application is shown in the figure.

[0017] Figure 3 The TUF2vec method schematic diagram of the present application is shown in the figure.

[0018] Figure 4 The classification model schematic diagram of the present application is shown in the figure. DETAILED DESCRIPTION

[0019] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0020] The present application provides a topic burst initial influence detection method based on small sample learning and data enhancement, characterized by comprising the following steps:

[0021] The user relationship network and the user basic information are subjected to feature extraction to obtain a user feature representation;

[0022] The field features are extracted from the user historical information, and the user field feature representation is obtained by using the extracted field features;

[0023] The user feature representation and the user field feature representation are weighted and fused based on an attention mechanism;

[0024] According to the fused feature representation, a user high-influence prototype vector and a user low-influence prototype vector are calculated;

[0025] According to the size between the distance between each user and the user high-influence prototype vector and the user low-influence prototype vector and a set threshold, the influence of the current user is determined.

[0026] As shown in the figure, the input of the present application is user basic attributes, user relationship networks and user historical field information, and the output of the model is whether the user will have an influence under the topic, so as to control the topic development as much as possible and provide a new opportunity for advertising marketing. Figure 1 The overall process of the present application is shown, which shows that the input of the present application is user basic attributes, user relationship networks and user historical field information, and the output of the model is whether the user will have an influence under the topic, so as to control the topic development as much as possible and provide a new opportunity for advertising marketing.

[0027] The present embodiment explains the present application scheme from three aspects of data acquisition, data enhancement and influence prediction, specifically including:

[0028] S1: Online data acquisition. The way to acquire data is to directly query the topic data in the database provided by the enterprise. What needs to be acquired here is the interaction of the user in the initial stage of the topic (such as comment, like, interaction and other data), the user historical behavior data existing in the history of the user and the relationship network between the user and the user. The historical behavior data includes the number of people being concerned in the historical field of the user, the sharing number of the user and the historical positioning label of the user. The basic information of the user and the user needs to be acquired, such as the number of fans and the number of attention, and then the data is simply processed.

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

[0030] S11: Acquire original data. The original data can be obtained by social network public API or directly downloading existing data source.

[0031] S12: Simple data cleaning. Usually, the acquired original data is unstructured and cannot be directly used for data analysis. Through simple data cleaning, most of the unstructured data can be structured. For example, deleting duplicate data, cleaning invalid information, etc.

[0032] S2: Data enhancement. Relevant attributes are extracted from user's own attributes and historical field data, topic content characteristics and user relationship network.

[0033] In the initial stage of the topic, there is often a problem of less user participation. At the same time, the user social media information, user relationship network data and user basic information in this period are relatively less and diversified, which leads to less effective information that can be obtained. Based on this problem, the data information is enhanced before the key user is found, and new sample data is generated.Figure 2 The specific process is as follows:

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

[0035] A variational autoencoder (VAE) is a structure of a probability distribution that maps input data to a distribution, denoted 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 can be described as: z is sampled from the prior distribution p θ (z) of the latent variable z, and x is sampled from the conditional probability distribution p θ (x|z) of x. Each sample has a specific Gaussian distribution, and the latent space Z of the variational autoencoder (VAE) is regarded as a mixture of infinite z. First, let x = UAttInfrom (u i ), and the 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, and a new likelihood function q φ (z|x) is introduced 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, and the distribution can be expressed as:

[0038] q φ (z|x) = N(μ, δ 2 ) (2)

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

[0040] Because the model has generation capability, in order to better express it in the latent space Z in the fitting space, each distribution N is made to tend to the standard Gaussian distribution and normal distribution N(0,1) as much as possible, and a reparameterization technique is added in the sampling process. In this model, sampling a z from N(μ, δ 2 ) is equivalent to sampling an ε from N(0,1), and then obtaining the latent variable z can be expressed as:

[0041] z = μ + ε × δ (3)

[0042] S22: Map the latent variable to the original data space.

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

[0044] p θ (x|z)(4)

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

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

[0047]

[0048] Where q φ (z|x) represents the distribution of the latent variable, and z is sampled from the latent space z ~ q 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 data generated by the decoder, and to minimize the difference between the original data and the data generated by the decoder through the reconstruction loss. Cross-entropy loss is used to measure and is expressed as:

[0050]

[0051] Where 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 of (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] Where β represents a hyperparameter that can adjust the weight of the KL divergence; p(z) represents the prior distribution of z.

[0055] The size between the KL divergence and the reconstruction loss needs to be constantly modified to obtain an optimal learning result, so that the difference of the regenerated data is minimized.

[0056] The completed learning encoder is used to generate a new vector representation according to the existing vector node representation, and the data set of node representation is expanded.

[0057] S3: Model establishment. The application proposes a feature representation method of T (Topic) U (User) F (Field) 2vec (to vector), first constructs the features between the topic and the user field, the user and the user, and performs feature representation on the information, then introduces an attention mechanism to fuse the vector representations of the two, finally uses a prototype network-based method to calculate the average value of the feature vectors to obtain a prototype vector, analyzes the influence of the user in the initial stage of the topic outbreak, and studies the development trend in the initial stage of the topic outbreak.

[0058] In the process of influence detection of the application, Figure 1 , is divided into three stages, namely:

[0059] The first stage, considering the influence of the user history field and the user relationship network on the topic propagation, proposes a TUF2vec algorithm, and the flowchart can be seen Figure 3 , first, fully consider the interaction between the user node and the user relationship network, realize the representation of the user features, then in the historical behavior information, mine the field feature representation of the user;

[0060] The second stage, combining the user relationship features obtained in the initial stage of the outbreak and the user historical field features, uses an attention mechanism to fuse the vector representations of the two;

[0061] The third stage, a key element influence discovery model based on a small sample is proposed, the influence of the user's participation behavior in the initial stage of the topic outbreak is analyzed, and the potential opinion leaders or topic development are detected by comprehensively analyzing the above information.

[0062] The user relationship network and the user basic information are extracted to obtain the user feature representation. In the process of topic propagation, the user relationship network RNet u ={U u ,E u} and the user basic attribute information set A={(UAtt,u i |u i∈U}. The feature representation of users is obtained by the correlation of the behaviors between users and the correlation of the attributes between users. The edge weight between users is judged by analyzing the interaction behavior between users and the interest similarity 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 defined as:

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

[0065] Where W(u i ,u j ) is the weight of node u i to node u j , and p_q_ratio(u i ,u j ) represents a parameter defined according to the characteristics between nodes, which is defined as:

[0066]

[0067] Where d(u i ,u j ) represents the distance between node u i and node u j , and p and q are two hyperparameters; when d(u i ,u j ) = 1, it represents the first-order neighbor of node u i and node u j ; when d(u i ,u j ) = 2, it represents the second-order neighbor of node u i and node u j ; when d(u i ,u j ) = 0, it represents that node u i is equal to node u j .

[0068] The interaction information between users is collected, including the number of comments, the number of interactions and the number of shares of a user under a topic, and the interaction between users and users is expressed as:

[0069] Interaction(u i ,u j )=∑w kNumInter(u i , u j )

[0070] where w k represents the weight coefficient of different interaction modes, used to balance the importance of different interaction modes, NumInter k (u i , u j ) represents the number of interactions between user u i and user u j for this user mode, including u i to u j , and u j to u i . Then use the minimum-maximum normalization method to map the interaction between users to [0, 1], so that the data is better processed. Its formula can be expressed as:

[0071]

[0072] If there is a follow relationship between users, the users are likely to have similar interests and hobbies. If the contact follow relationship is established to build a user relationship network, the interest similarity between users can be used as part of the weight of the user relationship network. The interest similarity between users is represented as:

[0073]

[0074] where SimilarU i represents the interest label set of user u i , |SimilarU i | represents the number of labels in the interest label set SimilarU i , and A∩B represents the intersection of sets A and B. The value range of interest similarity is [0, 1], where 1 represents complete interest, and 0 represents complete difference.

[0075] Therefore, the weight of node u i and u j is represented as:

[0076]

[0077] where Z is a 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 user basic information, and the user node is vectorized. Embedded in 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 the output is represented as:

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

[0079] wherein, ListU represents a vector matrix formed by a user relationship network, as a user feature representation of the present application, n i represents a vector representation of the i-th user.

[0080] In the topic propagation process, in order to better discover the influence of the key element, at the initial stage of the topic outbreak, sparse topic data cannot support accurate user influence discovery. The present application research finds that the historical field in the user historical information has a certain influence on the topic propagation. The higher the similarity between the field in which a user is located and the topic, the more attention the user obtains in the field, which proves that the topic is more likely to be propagated, and the user is likely to be a key element. Therefore, the present embodiment mines the related field information existing in the user history from the user historical information, extracts the field features, and thereby obtains the related information of the field in which the user is located, which is used for field feature representation of the user.

[0081] The present embodiment uses user historical information and propagation topics to represent the user. First, a user-field graph structure is constructed, the browsing records and historical interaction behaviors of the user for a topic and 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 historical field of the user.

[0082] The weight is represented as:

[0083]

[0084] wherein, u i represents the i-th user (i.e. the user u i in the present application), node u i , d j represents the j-th field. browseHis(u i , d j ) represents the browsing record number of node u i in field d j , and interCount(u i , d j ) represents the historical interaction behavior of node u i in field d j , including the like number, sharing number and other behaviors of the user in the field.

[0085] Finally, the present embodiment can obtain the field feature representation formed under the TUF2vec method as:

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

[0087] wherein, ListF represents a vector matrix formed by the user-field network structure, m i represents the field feature vector representation of the i-th user.

[0088] By analyzing the user influence from different angles, the user feature representation and the field feature representation are obtained, and considering that the attention mechanism can ignore the features with less influence on the result, the attention mechanism is used to fuse the user node features and the field 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 the user field features are weighted and averaged. The formula is represented as:

[0090] ListV = a x ListU + (1-a) x ListF (17)

[0091] wherein, ListV represents the fused feature representation vector, and a represents the attention weight.

[0092] The attention weight is calculated using the softmax function, and the attention score is processed, which can be represented as:

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

[0094] wherein, score(ListU, ListF) is the attention score, which can be used to measure the correlation between the user features and the field features. In this embodiment, the dot product attention is used to calculate the attention score, that is, the field feature representation ListF is taken as the key vector, and the user feature representation ListU formed by the user relationship network is taken as the value vector, which is represented as The superscript T represents the transpose of the vector or matrix, d k represents the dimension of the key vector, that is, the dimension of ListF.

[0095] At the same time, as Figure 4 ​As shown, considering the initial stage of the outbreak with fewer user nodes, a small sample classification method based on prototype is used, and the influence of the user is divided into high influence and low influence. The prototype vector of influence is obtained by calculating the average value of the feature representation vector of the same category user. 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 vector of all users in this category is calculated as the prototype vector, and the formula of the prototype vector after average value calculation is as follows:

[0096]

[0097] Wherein, 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 the high influence category, ListV i 2 ListV 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, the distance between the corresponding fused feature representation vector ListV and each influence category prototype vector is calculated. The distance dist between the to-be-classified user vector x and high influence and low influence is calculated, and the distance between the to-be-classified user vector x and high influence is represented as dist1, and the distance between the to-be-classified user vector x and low influence is represented as dist2. The user is assigned to the influence category with the closest distance. The present application sets a threshold value threshold to determine the influence size, and the formula can be represented as:

[0099]

[0100] Wherein, dist uses the Euclidean distance calculation, dist1 is the distance between the to-be-classified user and the prototype vector of the high influence category, dist2 is the distance between the to-be-classified user and the prototype vector of the low influence category, and threshold is the threshold value. If the distance between the to-be-classified user and all influence categories is greater than the threshold value, it will be classified as Unknown, HihgInfluence represents high influence, and LowInfluence represents low influence.

[0101] Although the embodiments of the present application have been shown and described, it can be understood by those skilled in the art that various changes, modifications, replacements and variations can be made to the embodiments without departing from the principles and spirits of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for detecting influence in early stage of topic burst based on small sample learning and data augmentation, characterized in that, The method comprises the following steps: The user relationship network and the user basic information are subjected to feature extraction to obtain a user feature representation, specifically including: The transition probability between nodes is established according to the weight relationship between nodes and the characteristics between nodes; Random walk is performed according to the transition probability from each node to obtain the context representation of the corresponding node; The TUF2vec method is adopted to vectorize the obtained context representation to obtain the vector representation of the corresponding node; The field features are extracted from the user historical information, and the extracted field features are used to obtain a user field feature representation, including: a user-field graph structure is constructed, the browsing records and historical interaction behaviors of the user for a topic and related topics are taken as the weight of the transition probability in the TUF2vec method, and the TUF2vec method is used to obtain the feature representation of the user historical field, and the weight calculation of the transition probability in the TUF2vec method includes: wherein W(u i ,d j ) represents the weight of node u i in the field d j ; Z is a normalization coefficient; browseHis(u i ,d j ) represents the number of browsing records of node u i in the field d j ; interCount(u i ,d j ) represents the historical interaction behavior of node u i in the field d ; The user feature representation and the user field feature representation are weighted and fused based on an attention mechanism; The user high-influence prototype vector and the user low-influence prototype vector are calculated according to the fused feature representation; The influence of the current user is determined according to the size between the distance between each user and the user high-influence prototype vector and the user low-influence prototype vector and a set threshold.

2. The method for detecting early influence of a topic burst based on small sample learning and data augmentation according to claim 1, characterized in that, The calculation of the transition probability between nodes includes: P(u i ,u j ) = W(u i ,u j ) x p_q_ratio(u i ,u j ) where P(u i ,u j ) represents the transition probability from node u i to node u j ; W(u i ,u j ) represents the weight from node u i to node u j ; and p_q_ratio(u i ,u j ) represents the characteristic parameter between node u i and node u j , and is represented as: where d(u i ,u j ) denotes the distance between node u i and node u j , and p and q are two hyperparameters; when d(u i ,u j )=1, it represents the first-order neighbor of node u i and node u j ; when d(u i ,u j )=2, it represents the second-order neighbor of node u i and node u j ; when d(u i ,u j )=0, it represents that node u i is equal to node u j .

3. The method for detecting early influence of a topic burst based on small sample learning and data augmentation according to claim 2, characterized in that, Node u i To node u j The calculation of the weight W(u i ,u j ) comprises: wherein Z is a normalization coefficient; Similarity(u i ,u j ) represents the interest similarity between node u i and node u j , SimilarU i represents the interest label set of node u i , SimilarU j represents the interest label set of node u j , |SimilarU i | represents the number of labels in the interest label set SimilarU i , represents the intersection of set A and set B; Interaction normal (u i ,u j ) is the normalized interaction frequency between node u i and node u j , Interaction min is the minimum activity frequency between nodes, Interaction max is the maximum activity frequency between nodes, Interaction (ui,uj) is the interaction frequency between node u i and node u j ; NunInter k (u i ,u j ) represents the interaction frequency of node u i and node u j in interaction mode k, w k represents the weight coefficient of interaction mode k.

4. The method for detecting early influence of a topic burst based on small sample learning and data augmentation according to claim 1, characterized in that, The process of weighting and fusing the user feature representation and the user field feature representation based on the attention mechanism includes: ListV=α×ListU+(1-α) ×ListF α=softmax(score(ListU, ListF)) wherein ListV represents the fused feature representation vector, a represents the attention weight; ListU is the user feature representation; ListF is the user field feature representation; score(ListU, ListF) is the attention score, d k represents the dimension of the key vector.

5. The method for detecting early influence of topic burst based on small sample learning and data augmentation according to claim 1, characterized in that, When the user high-influence prototype vector and the user low-influence prototype vector are calculated, the average value of the fused feature representation of the classified user is taken as the influence prototype vector of the type.

Images