A method for discovering the influence of key elements based on native-derived topic transfer learning

Abstract

The present invention belongs to the field of data mining technology and relates to a method for discovering the influence of key elements based on native-derivative topic transfer learning, including obtaining information including native topics and derived topics and related user information from an API interface provided by a social platform; constructing the early propagation network topology and propagation timing of derived topics, including using a joint distribution adaptive method to perform cross-domain feature adaptation on the content space of native topics and derived topics, and considering the sparsity of early data of derived topics, using an adversarial transfer learning method to compensate for the network structure; constructing a message-path-user ternary association graph of derived topics and performing cyclic iterative scoring to rank the influence of key elements of derived topics; the present invention can timely and accurately mine key elements in the early stage of the outbreak of derived topics, and the present invention can also be widely used in the precise placement of product advertisements, the discovery of important pathogenic genes, the prediction of popular research results, and the prevention of computer virus propagation.

Description

Technical Field

[0001] The present invention belongs to the field of data mining technology, relates to key element influence analysis, and in particular to a key element influence discovery method based on native-derived topic transfer learning. Background Art

[0002] Online social networks have become a crucial channel for spreading hot topics. Because they transcend the constraints of time and space, people can freely and conveniently publish and receive information. They also offer a multitude of dissemination methods, with a speed and reach unmatched by any other media product. Furthermore, their strong interactivity ensures communication and fosters user engagement. People use social networks to stay informed about global events, share real-life experiences, express opinions on social events, and maintain real-life relationships. Therefore, social networks, to a certain extent, reflect the true structure of social networks.

[0003] Some topics are highly correlated. When a hot topic breaks out, multiple related topics will randomly spawn at different times. People will consume multiple topics simultaneously, and related topics will influence each other, with the key elements in each topic having varying degrees of influence. This means that the influence of key elements in derived topics can be effectively predicted using data from the original topic. Traditional topic research has not considered the correlation between the original and derived topics, overlooked the interactions between multiple messages about hot topics and the diversity of message propagation paths, and lacked implicit influence calculations based on element correlations. Therefore, studying the influence of key elements in derived topics can quickly identify the source of information, facilitate public opinion monitoring, early warning, and analysis, minimize the impact of negative information, and maintain network security and social stability.

[0004] In recent years, research on discovering the influence of key elements in a topic has attracted widespread attention in academia, primarily focusing on methods based on node attributes and maximizing node influence. Node attribute-based methods are categorized into two levels: micro and macro. The micro level considers neighbor node information and information propagation paths, while the macro level considers the node's position within the global network structure, integrating network topology and user information to measure user social influence. Node influence maximization uses a subset of propagating nodes in the network as initial propagation sources, mining a set of nodes with a wide propagation range to maximize information dissemination. In recent years, numerous context-aware influence maximization problems have emerged, leveraging massive amounts of data to identify factors such as geographic location and interests.

[0005] Existing technologies have achieved some valuable results in discovering the influence of topic networks. However, they do not fully consider the differences and relevance between original topics and derived topics. Therefore, the research on the influence of topic nodes still has the following problems:

[0006] 1. The spatial distribution of derived topics is inconsistent with that of original topics. The subject matter of derived topics changes, and users have different areas of interest. How to reduce the data distribution difference between the source domain and the target domain is crucial to improving training results.

[0007] 2. How to more accurately construct the topology and time sequence of the early propagation network for derivative topics. Effective data is sparse in the early stages of a derivative topic outbreak, resulting in too little data that is truly helpful to the model, significantly impacting the model's data input.

[0008] 3. How to quantify the influence of key elements of derivative topics. In the early stages of a derivative topic's emergence, content changes and effective data is scarce. The key elements in the "original topic-derivative topic" influence each other, making it difficult to directly quantify the influence of key elements of the derivative topic. Summary of the Invention

[0009] To address the above issues, the present invention proposes a method for discovering the influence of key elements based on native-derived topic transfer learning, which specifically includes the following steps:

[0010] Obtain information including original topics and derived topics as well as related user information from the API interface provided by the social platform;

[0011] Construct the topological structure and propagation time sequence of the early propagation network of derivative topics, including using the joint distribution adaptation method to adapt the cross-domain features of the content space of the original topic and derivative topics, and considering the sparse data of the early derivative topics, using the adversarial transfer learning method to compensate for the network structure;

[0012] Construct a message-path-user ternary association graph of derived topics and perform iterative scoring to rank the influence of key elements of derived topics.

[0013] Furthermore, the information of native topics and derived topics includes the topic content, user ID, user reposts and number, and user comment content and number of the native topics and derived topics; relevant user information includes user basic attributes and user behavior information. User basic attributes include user interests and hobbies, user areas of expertise, user tag information, user friend relationships and number. User behavior information includes user reposts, user comments, user friend reposts, and user historical published content.

[0014] Furthermore, before using the joint distribution adaptation method to perform cross-domain feature adaptation on the content space of the original topic and the derived topic, the information of the original topic and the derived topic is preprocessed, specifically including:

[0015] Extract the topic propagation network of original topics and derived topics, that is, extract the relevant Weibo content posted by all users, and obtain the content text of the original topics and derived topics posted by users respectively;

[0016] Extract the one-month historical postings of original topics, derived topics, overlapping users of original and derived topics, and potential users of derived topics, and obtain the text content of the content posted by original topic users and potential users, as well as the text content of the content posted by derived topic users and potential users;

[0017] Perform Chinese word segmentation and part-of-speech tagging on each text content, remove useless words and stop words, and obtain candidate keywords for the topic;

[0018] Use the TF-IDF algorithm to calculate the weight of each candidate keyword, extract the central keywords of the topic, eliminate noise, and obtain the main keyword sequences of the original topic and derived topics;

[0019] Use the Doc2vec algorithm to learn the keyword sequence and obtain the native topic content feature vector T S , derived topic content feature vector T D , native topic users and potential user content feature vector U S and the derived topic user and potential user content feature vector U D .

[0020] Furthermore, the optimization objective of the cross-domain feature adaptation process of the original topic and derived topic content space using the joint distribution adaptation method is expressed as:

[0021]

[0022] Among them, X represents the data of the original topic and the derived topic; C represents the number of sample categories; M c represents the maximum mean difference matrix between classes; λ represents the regularization parameter; I represents the identity matrix; W represents the shared features of the content space of the original topic and the derived topic; Θ is the Lagrange multiplier, and H is the center matrix.

[0023] Furthermore, considering the sparseness of early data in derived topics, the adversarial transfer learning method is used to compensate for the topic network structure. The specific steps include:

[0024] Construct the structural feature vector of each node. The structural feature vector of a node is composed of the node structure and node influence of the node;

[0025] Construct a feature extractor G f , label classifier G y and domain discriminator G dAdversarial transfer learning network;

[0026] The feature extractor extracts features from the structural feature vector of the node, the label classifier classifies based on the extracted features, and the domain discriminator determines whether the extracted features are from the original topic or the derived topic.

[0027] During the compensation process, the feature extractor minimizes the label classifier error and maximizes the classification error of the domain discriminator.

[0028] Furthermore, the structural feature vector of a node is composed of the node structure and node influence of the node. The structural feature vector of a node includes:

[0029] When obtaining the node structure of a node, the SDNE algorithm is used to perform representation learning on the network structure to obtain the global structural features and local structural features of the node. The global structural features and local structural features of the node are used as the structural features of the node;

[0030] The influence of a node is expressed as:

[0031] Inf(w i )=γNum[foll(w i )]+Act(w i )

[0032]

[0033] Active(w i )=Num[orig(w i )]+α*Num[retw(w i )]+β*Num[com(w i )]

[0034] Among them, γ represents the weight coefficient; Active(w i ) represents user w i Active index; Active(w max ) represents the most active user w max Active index, Active(w min ) represents the least active user w min Activity index; Num[orig(w i )] represents user w i The number of microblogs published within one month before the outbreak of the derivative topic, Num[retw(w i )] and Num[com(w i )] represents user w iThe number of reposted Weibo posts and comments within one month before the outbreak of the derivative topic; α and β are reinforcement factors, and α∈[0,1] and β∈[0,0.5].

[0035] The present invention can timely and accurately mine the key elements in the early stage of the outbreak of derivative topics. In addition, the present invention can also be widely used in accurately placing product advertisements, discovering important pathogenic genes, predicting popular research results, and preventing the spread of computer viruses. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is a model diagram of the key element influence discovery method based on native-derived topic transfer learning in the present invention;

[0037] Figure 2 Discover examples of the influence of key elements of the topic derived from this invention;

[0038] Figure 3 Adapting the joint distribution of native-derived topic content space in the present invention;

[0039] Figure 4 This is the adversarial migration of the original-derived topic network structure in the present invention. DETAILED DESCRIPTION

[0040] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. 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 making creative efforts are within the scope of protection of the present invention.

[0041] The present invention proposes a method for discovering the influence of key elements based on native-derived topic transfer learning, which specifically includes the following steps:

[0042] Obtain information including original topics and derived topics as well as related user information from the API interface provided by the social platform;

[0043] Construct the topological structure and propagation time sequence of the early propagation network of derivative topics, including using the joint distribution adaptation method to adapt the cross-domain features of the content space of the original topic and derivative topics, and considering the sparse data of the early derivative topics, using the adversarial transfer learning method to compensate for the network structure;

[0044] Construct a message-path-user ternary association graph of derived topics and perform iterative scoring to rank the influence of key elements of derived topics.

[0045] like Figure 1The figure shows the overall framework of the present invention, indicating that the input of the present invention is the network data of the original topic and the derived topic, and the output after the key element influence discovery model is the key element of the derived topic.

[0046] The present invention comprises the following three steps:

[0047] S1: Obtain data. Data can be obtained from public data websites or through mature social network public APIs. The data obtained mainly includes three parts: topic information data, user basic attributes, and user behavior information. Topic information data includes the topic content of original and derived topics, user ID, user forwarding and number, and user comment content and number; user basic attributes include user interests and hobbies, user expertise, user tag information, user friend relationships and number; user behavior information includes the number of user forwarding, user comments, the number of user friend forwarding, and the user's historical postings.

[0048] S2: Construct the early propagation network topology and propagation time sequence of derived topics. Based on the acquired data, extract the content space and network structure characteristics of the original and derived topics, and use joint distribution adaptation and adversarial transfer to perform transfer learning on the original topic data space.

[0049] S3: Quantify the influence of key elements. Construct a ternary relationship diagram of "message-path-user" and iteratively score and rank the influence of key elements.

[0050] The above step S1 obtains the data source, which is mainly divided into the following two steps.

[0051] S11: Obtaining raw data: Raw data can be obtained through social network public APIs or public data websites.

[0052] S12: Simple data cleaning. The raw data typically obtained is unstructured and cannot be directly used for data analysis. Simple data cleaning can structure most unstructured data, eliminating outliers or null values and reducing the inconvenience caused by subsequent calculations.

[0053] S13: Data storage. After simple data clarification, the data needs to be stored in a database. The data is further standardized through the table structure. The database can also greatly improve the efficiency of data retrieval and the mapping of relationships between tables.

[0054] The above step S2 constructs the topological structure and propagation time sequence of the early propagation network of derived topics. It mainly consists of the following two steps. In view of the inconsistent spatial distribution of the content of derived topics and original topics, a joint distribution adaptation method is used to perform cross-domain feature adaptation on the content space of original topics and derived topics. Taking into account the sparse data of early derived topics, an adversarial transfer learning method is used to compensate for the network structure. The details are as follows:

[0055] S21: Content Space Adaptation

[0056] S211: Topic Content and User Content

[0057] On social networks, the users participating in primary and secondary topics vary somewhat. This is due to changes in topic content and differing user interests, manifesting as inconsistent data distribution in the data space. Focusing on topic content, we extract the topic propagation network for primary and secondary topics, extracting all relevant Weibo posts from users into two articles. Focusing on user content, we extract a month's worth of historical posts for primary and secondary topics, as well as users who overlap between primary and secondary topics and potential users of secondary topics.

[0058] First, we preprocess the text content. Each article undergoes Chinese word segmentation and part-of-speech tagging, removing useless parts of speech and stop words to obtain candidate keywords for the topic. Next, we use the TF-IDF algorithm to calculate the weight of each candidate keyword, extract the central keywords of the topic, eliminate noise, and obtain the main keyword sequences for the original topic and derived topics. Finally, we use the Doc2vec algorithm to learn the output feature vectors:

[0059] Original topic content feature vector T S (1)

[0060] Derived topic content feature vector T D (2)

[0061] Native topic user (and potential user) content feature vector U S (3)

[0062] Derived topic user (and potential user) content feature vector U D (4)

[0063] S212: Joint Distribution Adaptation

[0064] Assume X S =X T , Y S =Y T , P(X s )≠P(X t ), Q(Y s |Xs )≠Q(Y t |X t ). Find a transformation W so that the difference between the marginal distribution and conditional distribution of the original topic and the derived topic data space after the transformation is significantly reduced, such as Figure 3 As shown, the classifier in the original topic can be used to predict user behavior in the derived topic.

[0065] According to the feature vectors of topic content and user content obtained above, assuming that the original topic X S =[T S ,U S ], derivative topic X T =[T T ,U T ], then the source domain D S ={X s ,P(x s )}, where x s ∈X s , the source domain task is T S ={Y s ,f(x s )}, where y s ∈Y s . f(x s ) is the classifier on the source domain. The target domain D T ={X T ,P(x t )}, where x t ∈X t , the target domain task is X T Predict Y T .

[0066] First, reduce the difference in marginal probability distribution. Make P(W T x s )=P(W T x t ) as close as possible, and the maximum mean difference (MMD) distance is used to minimize the maximum mean difference between the original topic and the derived topic. s and n t Represents the number of samples of original topics and derived topics respectively.

[0067]

[0068] By introducing the kernel method, (1) becomes the following formula:

[0069] tr(W T XM0X T W)(2)

[0070] in,

[0071]

[0072] Among them, M o Represents the maximum mean difference matrix between the original topic and the derived topic; n s Indicates the number of samples of the original topic; D s Indicates the original topic; n t Indicates the number of samples of derived topics; D t Indicates a derivative topic.

[0073] Secondly, adapt the conditional probability of the original topic and the derived topic. s ,Y s ) Train a simple classifier to directly predict the derived topic based on the pseudo label Calculate. Then the maximum mean difference MMD distance between classes is

[0074]

[0075] in, and Represent the number of samples from category c in the original topic and the derived topic respectively. Similarly, the kernel method is introduced to obtain:

[0076] tr(W T XM c X T W)(5)

[0077] in,

[0078]

[0079] Among them, M c Represents the maximum mean difference matrix between classes; Indicates the number of samples of category c in the original topic; The cth category represents the original topic; Indicates the number of samples of category c in the derived topic; Represents the cth category of derived topics.

[0080] Finally, the two distances are combined to obtain an overall optimization goal:

[0081]

[0082] Where X represents the combined data of the original topic and derived topics, C represents the number of sample categories, λ represents the regularization parameter, Θ represents the Lagrange multiplier, H represents the center matrix, and I represents the identity matrix. The W transformation is the shared feature matrix of the content space of the original topic and derived topics, which is used to predict whether the data in the derived topic should be forwarded.

[0083] S22: Network Structure Adversarial Transfer

[0084] S221: Local and global structure of nodes

[0085] In the early stages of a derivative topic outbreak, data is sparse, so we can learn the network topology and propagation time sequence of the original topic. We use the SDNE algorithm to learn the representation of the network structure, maintaining the first-order and second-order similarities of the network and preserving the local and global features of the nodes. SDNE uses a semi-supervised deep model based on an opportunistic autoencoder for network embedding, where:

[0086] The input of the model is the adjacency matrix x of the i-th node i =s i , S={s1,s2,…,s n} is the adjacency matrix of the network, and the output is the reconstructed adjacency matrix The error between the input and reconstruction matrix is continuously reduced until it is less than the set threshold to obtain the global structural features.

[0087] The second-order similarity loss function is Among them, ⊙ represents the basic product, If s i,j = 0, then b i,j =1, otherwise b i,j =β>1.

[0088] For each layer is the local feature of the node, Among them, W (k) is the weight matrix of the kth layer, b (k) is the bias term of the kth layer.

[0089] The first-order similarity loss function is in, represents the hidden representation of node i in layer k, s i,j Indicates whether the i-th node is connected to the j-th node. If there is an edge, it is 1, otherwise it is 0. n represents the number of edges between the two nodes.

[0090] The minimum optimization objective function after the combination is: L min =L 2nd +λL 1st +vL reg , where L reg is the regularization term, λ is the parameter that controls the first-order loss, and v is the parameter that controls the regularization term.

[0091] Finally, the training model is used to obtain the minimum L minAfter that, the output is the user's network structure feature vector expressed as: S = n × F a Among them, n is the number of network nodes, that is, the number of users under the topic, F a is the structural feature representation vector corresponding to the user.

[0092] S222: Node Influence

[0093] Even if two nodes have similar local and global structures, their influence on other nodes may be different, which is related to the influence of the node itself. Assume that the influence of a node is determined by the number of user friends and user activity.

[0094] Act(w i ) is used to describe user w i Generally speaking, the higher the user's activity, the greater the impact on other users. The user's activity Act(w i ) is defined as follows:

[0095]

[0096] Among them, Active(w i ) refers to user w i The activity index, w max and w min Represent the most active user and the least active user respectively; user w i The activity index is expressed as:

[0097] Active(w i )=Num[orig(w i )]+α*Num[retw(w i )]+β*Num[com(w i )]

[0098] Among them, γ represents the weight coefficient; Num[orig(w i )] represents user w i The number of microblogs published within one month before the outbreak of the derivative topic, Num[retw(w i )] and Num[com(w i )] represents user w i The number of forwarded Weibo posts and comments within one month before the outbreak of the derivative topic. Generally, the number of forwarded Weibo posts of a user can better reflect the influence of a user on others. Therefore, α and β are defined as reinforcement factors, and α∈[0,1] and β∈[0,0.5].

[0099] The number of user friends refers to the number of people who follow this user. i)], the influence of the node is expressed as: Inf(w i )=γNum[foll(w i )]+Act(w i ), γ represents the weight coefficient.

[0100] S223: Adversarial Transfer Learning

[0101] During the topic propagation process, due to the sparse data in the early stage of the outbreak of derived topics, it is necessary to compensate for the derived topic data. The network topology of the original topic is related to the network topology of the derived topic. By combining transfer learning with deep learning, we can explore the potential relationship between the original and derived topics in the data and extract shared features, so that the domain discriminator cannot distinguish between the data from the original topic and the derived topic. Figure 4 While achieving better feature representation capabilities of the deep learning model, it also solves the problem of changes in the derived topic network structure to a certain extent and accurately enriches the derived topic data.

[0102] Assume that the original topic dataset is in Represents the native topic feature vector, including the potential elements in the native topic, Represents the original topic element label. The derived topic dataset is in represents the derived topic feature vector, Represents the derived topic element label. Each structural feature vector consists of node structure and node influence.

[0103] The adversarial transfer learning network consists of the feature extractor G f , label classifier G y and domain discriminator G d They are composed of multiple layers of composite functions. For the feature extractor, its output is expressed as:

[0104] G f (x i ;W,b)=sigm(Wx i +b);

[0105] Among them, x i represents the i-th sample of the input, W is the weight term of the feature extractor, b is the bias term of the feature extractor, and sigm() is the activation function; G f (x i ;W,b)(abbreviated as G f (x i ))Extract a new feature space.

[0106] For the label classifier, its output is expressed as:

[0107] G y (G f (x i ),V,c)=softmax(VG f (x i )+c);

[0108] Among them, V represents, c represents, G y (G f (x i ),V,c) represents the predicted label, and the loss of the label classifier is:

[0109] L y (G y (G f (x i )),y i )=-logG y (G f (x i ))yi;

[0110] Among them, x i and y i are the input i-th sample and label, and the loss L of the label classifier y (G y (G f (x i )),y i ) is abbreviated as L y .

[0111] The domain discriminator and the label classifier share input features, and the output is the source of the sample, that is, whether the sample comes from the original topic or the derived topic. The domain discriminator loss is:

[0112] L d (G d (G f (x i )),d i )=-d i logG d (G f (x i ))-(1-d i )logG d (G f (x i ));

[0113] Among them, d i is the domain category label of the i-th sample, and the domain discriminator loss L d (G d (G f (x i )),d i) is abbreviated as L d The feature extractor minimizes the error of the label classifier and continuously improves the classification of the features. At the same time, the feature extractor maximizes the classification error of the domain discriminator, making the features domain invariant.

[0114] In order to minimize the overall loss function, a gradient reversal layer is added between the feature extractor and the domain discriminator, which remains unchanged during forward propagation and is automatically reversed during backward propagation. The total loss function L = L y -ηL d η represents the regularization parameter of the adversarial transfer learning model. The label classification loss and domain classification loss synergistically optimize the feature extractor to find domain-invariant features V. This allows the model trained on the original topic to be transferred to the derived topics. It predicts whether the data in the derived topics will be forwarded, effectively enhancing the construction of the early propagation network topology and propagation time sequence of the derived topics.

[0115] Quantifying the influence of key elements mainly includes three steps, namely, constructing a ternary association graph of derived topics, calculating the transition probability matrix, and scoring through iterative loops, so as to timely and accurately discover the influence of key elements of derived topics.

[0116] S31: Constructing a ternary association graph of the derived topic "message-path-user"

[0117] Users, paths, and messages are key elements in the network. By constructing a ternary association graph, we can discover the correlation between them and the influence of key elements. The derived topic ternary graph can be represented as M is the message set, P is the path set, and U is the user set. and is the transition probability matrix between messages and paths, and is the transition probability matrix between paths and users.

[0118] S32: Calculate the transition probability matrix

[0119] Different types of key elements, such as users, paths, and messages, have their own characteristics and are interrelated. Different key elements contain rich information, and different relationships can be used to learn the feature representation of each node. The knowledge representation transD algorithm is used to unify the vector representation of nodes. The mapping matrix of the relationship is simplified to the product of two vectors, and the feature representation of each node in the unified vector space is obtained by mapping it to the relationship space through a dynamic mapping matrix. The topic network is then reconstructed using knowledge representation. At this point, the similarity between the message vector and the path vector, and between the path vector and the user vector can be calculated, and the transition probability matrices of message-path, path-message, user-path, and path-user can be obtained:

[0120]

[0121]

[0122]

[0123]

[0124] Among them, sim(m i ,p j ) represents the similarity matrix between the i-th message vector and the j-th path vector; sim(p j ,u k ) represents the similarity matrix between the j-th path vector and the i-th user vector; p j ∈N(m i ) represents the path type node that has a transfer relationship with the i-th message; m i ∈N(p j ) represents the message type node that has a transfer relationship with the jth path; u k ∈N(p j ) represents the user type node that has a transfer relationship with the jth path; p j ∈N(u k ) represents the path type node that has a transfer relationship with the k-th user.

[0125] S33: Iterative scoring of key element influence

[0126] If a node is considered important, it is easily accessible to other nodes and has a high authority value. This means that the node is highly likely to be accessed by other nodes, and its messages are more likely to spread within the network. Therefore, we can draw on the idea of point cross-iterative scoring to consider the mutual influence relationships in the topic ternary association graph and calculate the final influence score of the key element node.

[0127] First, an initial value is assigned to each node in the three-part graph, and then an update is performed after each iteration. Let X, Y, and Z be the score vectors of the message, path, and user, respectively, representing the popularity of the message, the importance of the propagation path in the network, and the criticality of the user in the network. Define X (0) ,Y (0) ,Z (0) is the initialization score vector for the three sets of messages, paths and users.

[0128]

[0129]

[0130]

[0131] Among them, α i Indicates message M i The total number of participating users, β j Represents path p j The number of users included, r k Represents user u k Node influence.

[0132] After obtaining the initial influence value and transfer probability matrix of each element node, perform forward and reverse iterative scoring. The forward iterative calculation formula is as follows:

[0133]

[0134]

[0135] The reverse iterative calculation formula is as follows:

[0136]

[0137]

[0138] Where μ is the damping coefficient and ε is the threshold. The iteration ends when ‖X′-X‖+‖Y′-Y‖+‖Z′-Z‖<ε. After each iteration, X, Y, and Z need to be normalized.

[0139] While 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 these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for discovering the influence of key elements based on native-derived topic transfer learning, characterized by: The following steps are involved: Obtain information including original topics and derived topics as well as related user information from the API interface provided by the social platform; Construct the topological structure and propagation time sequence of the early propagation network of derived topics, including using the joint distribution adaptation method to adapt the cross-domain features of the original topic and derived topic content space. The optimization goal of this process is expressed as: Among them, X represents the data of the original topic and the derived topic; C represents the number of sample categories; M c represents the maximum mean difference matrix between classes; λ represents the regularization parameter; I represents the identity matrix; W represents the shared features of the original topic and derived topic content space; Θ is the Lagrange multiplier, and H is the center matrix; Considering the sparseness of early data in the derived topic, the adversarial transfer learning method is used to compensate for the network structure. The specific steps include: Construct the structural feature vector of each node. The structural feature vector of a node is composed of the node structure and node influence of the node; Construct a feature extractor G f , label classifier G y and domain discriminator G d Adversarial transfer learning network; The feature extractor extracts features from the structural feature vector of the node, the label classifier classifies based on the extracted features, and the domain discriminator determines whether the extracted features are from the original topic or the derived topic. During the compensation process, the feature extractor minimizes the label classifier error and maximizes the classification error of the domain discriminator; Construct the message-path-user ternary association graph of the derived topic, expressed as: The transition probability matrix from user to path is expressed as: The transition probability matrix from path to user is expressed as: The transition probability matrix from path to user is expressed as: is the message-to-path transition probability matrix, expressed as: Among them, U is the user set, P is the path set, and M is the message set; sim(m i ,p j ) represents the similarity matrix between the i-th message vector and the j-th path vector; N(m i ) represents the set of path type nodes that have a transfer relationship with the i-th message; sim(p j ,m i ) represents the similarity matrix between the j-th path vector and the i-th message vector; N(p j ) represents the set of user type nodes that have a transfer relationship with the jth path; sim(p j ,u k ) represents the similarity matrix between the j-th path vector and the i-th user vector; N(u k ) represents the set of path type nodes that have a transfer relationship with the k-th user; Then, we conduct iterative scoring to rank the influence of key elements of derived topics.

2. The method for discovering the influence of key elements based on native-derived topic transfer learning according to claim 1 is characterized in that: Information on native topics and derived topics includes the topic content, user ID, user reposts and number, and user comment content and number; relevant user information includes user basic attributes and user behavior information. User basic attributes include user interests and hobbies, user areas of expertise, user tag information, user friend relationships and number. User behavior information includes user reposts, user comments, number of reposts by user friends, and user historical published content.

3. The method for discovering the influence of key elements based on native-derived topic transfer learning according to claim 1 is characterized in that: Before using the joint distribution adaptation method to perform cross-domain feature adaptation on the content space of the original topic and the derived topic, the information of the original topic and the derived topic is preprocessed, specifically including: Extract the topic propagation network of original topics and derived topics, that is, extract the relevant Weibo content posted by all users, and obtain the content text of the original topics and derived topics posted by users respectively; Extract the one-month historical postings of original topics, derived topics, overlapping users of original and derived topics, and potential users of derived topics, and obtain the text content of the content posted by original topic users and potential users, as well as the text content of the content posted by derived topic users and potential users; Perform Chinese word segmentation and part-of-speech tagging on each text content, remove useless words and stop words, and obtain candidate keywords for the topic; Use the TF-IDF algorithm to calculate the weight of each candidate keyword, extract the central keywords of the topic, eliminate noise, and obtain the main keyword sequences of the original topic and derived topics; Use the Doc2vec algorithm to learn the keyword sequence and obtain the native topic content feature vector T S , derived topic content feature vector T D , native topic users and potential user content feature vector U S and the derived topic user and potential user content feature vector U D .

4. The method for discovering the influence of key elements based on native-derived topic transfer learning according to claim 1 is characterized in that: When c = 0, the maximum mean difference matrix M0 between classes is expressed as: Among them, (M0) ij The maximum mean difference matrix M0 of the original topic and the derived topic in row i and column j; n s Indicates the number of samples of the original topic; D s Indicates the original topic; n t Indicates the number of samples of derived topics; D t Indicates a derived topic; x i 、x j They represent the i-th data and j-th data in the data X that is the combination of the original topic and the derived topic respectively.

5. The method for discovering the influence of key elements based on native-derived topic transfer learning according to claim 1 is characterized in that: The structural feature vector of a node is composed of the node structure and node influence of the node. The structural feature vector of a node includes: When obtaining the node structure of a node, the SDNE algorithm is used to perform representation learning on the network structure to obtain the global structural features and local structural features of the node. The global structural features and local structural features of the node are used as the structural features of the node; The influence of a node is expressed as: Inf(w i )=γNum[foll(w i )]+Act(w i ) Active(w i )=Num[orig(w i )]+α*Num[retw(w i )]+β*Num[com(w i )] Among them, γ represents the weight coefficient; Active(w i ) represents user w i Active index; Active(w max ) represents the most active user w max Active index, Active(w min ) represents the least active user w min Activity index; Num[orig(w i )] represents user w i The number of microblogs published within one month before the outbreak of the derivative topic, Num[retw(w i )] and Num[com(w i )] represents user w i The number of reposted Weibo posts and comments within one month before the outbreak of the derivative topic; α and β are reinforcement factors, and α∈[0,1] and β∈[0,0.5].

6. The method for discovering the influence of key elements based on native-derived topic transfer learning according to claim 1, characterized in that: The process of loop iterative scoring includes forward iterative calculation and reverse iterative calculation, specifically including: The forward iteration calculation formula is as follows: The reverse iterative calculation formula is as follows: When ‖X′-X‖+‖Y′-Y‖+‖Z′-Z‖<ε, the iteration ends. After each iteration, the score vectors X, Y, and Z of the message, path, and user need to be normalized. Among them, Y′ represents the score of the current iteration path; X represents the score of the previous iteration message; μ represents the damping coefficient; Y represents the score of the previous iteration path; Z′ represents the score of the current iteration user; Z represents the score of the previous iteration user; X′ represents the score of the current iteration message; ε represents the threshold.

Images