A method for predicting the spread of derivative topics on social networks based on transfer learning

Abstract

This invention belongs to the field of social network public opinion analysis, specifically involving a method for predicting the spread of derivative topics on social networks based on transfer learning. This method includes extracting relevant attributes from topic data, using the IT2vec representation learning algorithm to iteratively represent the original topic into derivative topics, employing a domain adaptation method based on TCA to adapt features between the original and derivative topics, obtaining general feature data; constructing a CNN prediction model based on subjective cognitive game theory, training the CNN prediction model using the general feature data, and testing the trained CNN prediction model using the iterative representation results; and using the tested CNN prediction model to predict the spread trend of the current derivative topics. This invention can effectively predict the spread trend of derivative topics, which is crucial for timely control of the further evolution of negative topics such as rumors and for correctly guiding online public opinion.

Description

Technical Field

[0001] This invention belongs to the field of social network public opinion analysis, and involves user behavior analysis, especially a method for predicting the spread of social network derivative topics based on transfer learning. Background Technology

[0002] With the rapid development of internet technology, online social networks such as Facebook, Twitter, Weibo, and WeChat have gradually permeated people's lives. These networks generate trending topics every moment, which are discussed and spread, and a single trending topic may spawn one or more derivative topics. Derivative topics differ from ordinary trending topics; they are usually formed through multiple rounds of dissemination of a primary trending topic. They are closely related to the trending topics that users previously followed and are new topics that are more likely to arouse user interest. This "fission-like" dissemination of information can easily lead to the widespread spread of negative news, affecting social harmony and stability. Therefore, understanding the dissemination patterns of derivative topics and analyzing user behavior within the context of their dissemination is crucial for public opinion control.

[0003] In recent years, a large number of scholars have conducted research in the field of social networks and achieved considerable success. However, research on the spread of derivative topics is relatively limited. Currently, scholars' research on the spread of derivative topics mainly focuses on two aspects: firstly, predicting derivative topics, primarily based on the topology of networks of common hot topics, user behavior under the topic, and textual information between users to study whether derivative topics will emerge; secondly, studying the evolution mechanism of derivative topics, exploring the ways and patterns of topic derivation from the perspective of topic spread, which can be achieved by simulating the topic derivation process using deep learning models or SIR models. Essentially, all of the above research is based on the derivative characteristics of topics, studying derivative topics based on the spread and changes of common hot topics to help relevant departments grasp the derivative effects of topics, understand the direction of derivative public opinion, and further adopt targeted public opinion guidance strategies.

[0004] Since derivative topics evolve from one or more trending topics through multiple rounds of dissemination, users who participated in the topic's spread tend to have a certain "cognition" of the subsequently emerging derivative topics, making them more likely to act on them. However, most current research on topic dissemination models fails to consider this aspect, leading to inaccurate conclusions. Wu et al., in their study "Public opinion dissemination with incomplete information on social network: a study based on the infectious diseases model and gametheory," proposed an information interference model. This model considers the interrelationships between topics in social networks and analyzes the impact of information on the interference and resistance to interference in highly clustered regular and random networks.

[0005] Based on current research on topic dissemination, the following challenges remain:

[0006] 1. Sparsity of effective user behavior data in the early stages of a derivative topic's emergence. Although the derivative topic space contains a large amount of data, effective user behavior data is very sparse, especially in the early stages of a derivative topic's emergence. This will inevitably affect the accuracy of the entire dissemination model.

[0007] 2. The complexity of the derivative topic dissemination space. In the derivative topic dissemination space, there are multiple complex relationships between entities such as users, messages, and networks, which poses a challenge to effectively extracting and simplifying features.

[0008] 3. The issue of users' accumulated knowledge about derivative topics. Before a derivative topic explodes, a similar preceding topic has already been circulating for a period of time, and users will accumulate a certain level of knowledge about this type of topic. This knowledge will inevitably affect users' behavior towards derivative topics, and how to quantify this cognitive influence is an urgent problem to be solved. Summary of the Invention

[0009] To address the aforementioned problems, this invention provides a method for predicting the spread of derivative topics on social networks based on transfer learning. Starting from the differences and connections between general trending topics and derivative topics, the method iteratively represents derivative topic data, fully considering the impact of topic spread. Simultaneously, it utilizes the concept of transfer learning to better predict the spread of derivative topics. This method includes the following steps:

[0010] S1. Obtain topic data through public dataset websites or by using public API interfaces provided by social networks. The topic dataset includes original topic data and derived topic data.

[0011] S2. Extract relevant attributes from the original topic data and the derived topic data respectively. The relevant attributes include the topic propagation network, user basic attributes, message propagation popularity and topic awareness. Construct the original topic propagation network and the derived topic propagation network through the topic propagation network.

[0012] S3. The IT2vec representation learning algorithm is used to iteratively represent the original topic propagation network to the derived topic propagation network, and the low-rank vectorized feature representation of the derived topic propagation network is obtained.

[0013] S4. Using a domain adaptation method based on TCA, feature adaptation is performed on the original topic propagation network and the derived topic propagation network to obtain general feature data;

[0014] S5. Construct a CNN prediction model based on subjective cognitive game theory, train the CNN prediction model using general feature data, and test the trained CNN prediction model using the low-rank vectorized feature representation of the derivative topic propagation network.

[0015] S6. Use a tested CNN prediction model to predict the spread trend of current derivative topics, and obtain the user's cognitive influence and the user group behavior under the derivative topics at the next moment.

[0016] Furthermore, relevant attributes are extracted from the topic data and represented as follows:

[0017] Topic dissemination network:

[0018]

[0019] U represents the topic propagation network at time t. t This represents the set of users participating in the topic at time t. U t The set of relational edges between users, where each relational edge is directed;

[0020] User basic properties UserProp(u i ):

[0021] UserProp(u i ) = [age(u i ),sex(u i ),follow(u i ),fans(u i )]

[0022] age(u i ) indicates the user u who participated in the topic. i age, sex(u i) indicates the user u who participated in the topic. i The gender, follow(u i ) indicates the user u who participated in the topic. i Number of followers, fans(u i ) indicates the user u who participated in the topic. i The number of fans;

[0023] Message spread popularity MsgHeat(m):

[0024]

[0025] Num[forward(m)] represents the number of forwards of the m-th original message published in the topic, and Num[fans(u i )] represents the number of followers of the original user who created the original message m, and ω and For adjustable hyperparameters, Let t be a time decay function, λ be the heat decay factor, and t be a time decay function. m0 Let t be the publication time of the original message m, and t be the current time.

[0026] UserAware (u) i ):

[0027]

[0028] Num[original(u i )] for user u i The total number of original messages in the topic, Num[forword(u i )] represents user u i The total number of forwarded messages in the topic; Num[original(U N )] represents user u i User group U N The total number of all original messages in the topic, Num[forword(U N )] represents user u i User group U N The total number of all forwarded messages, where α and β are adjustable hyperparameters.

[0029] Furthermore, the specific process of step S3 includes:

[0030] S31. Obtain the basic user attributes of all users in the original topic propagation network, and calculate the edge weights between user nodes in the original topic propagation network based on the user interaction degree and the similarity of user basic attributes.

[0031] S32. Set the node traversal strategy by setting the edge weights between user nodes in the original topic propagation network, randomly give a node feature mapping function, and obtain the node feature mapping function of the original topic propagation network. The first iteration is completed.

[0032] S33. Obtain the basic user attributes of all users in the derivative topic propagation network, and obtain the edge weights between user nodes in the derivative topic propagation network based on the interaction degree between user nodes in the original topic propagation network, the user interaction degree in the derivative topic propagation network, and the similarity of user basic attributes.

[0033] S34. Combine the edge weights between user nodes in the derivative topic propagation network to set a new node traversal strategy. Based on the node feature mapping function of the original topic propagation network obtained in S32, a new node feature mapping function for training is obtained, and the node feature mapping function of the derivative topic propagation network is obtained. The new feature representation of the derivative topic propagation network is further obtained through the node feature mapping function of the derivative topic propagation network.

[0034] Furthermore, the node walking observation probability in the node walking strategy is expressed as:

[0035]

[0036]

[0037] Wherein, P(u j |u i ) represents the current user node u i The next-hop user node is u j The probability of i ≠ j, and α p,q (u i ,u j W(u) represents the walking bias. i ,u j ) represents user node u i and user node u j The edge weights between u and z are given by z, which represents the normalization constant, and E represents the set of edges in the topic propagation network. i ,u j ∈E represents user node u i and user node u j There are connections between them. Represents user node u i To user node u j The distance is p,q, where p and q represent the biasing coefficients.

[0038] Furthermore, user node u i and user node u j The formula for calculating the edge weight between them is:

[0039]

[0040]

[0041]

[0042] in, This represents the correlation coefficient of basic user attributes, Multual(u i ,u j ) represents user node u i and user node u j User interaction, Mutual′(u i ,u j ) represents user node u in the topic propagation network obtained in the previous iteration. i ,u j The degree of interaction between them, if the topic propagation network is the original topic propagation network, then Mutual′(u i ,u j `interact(u) = 0`, where `μ` is an adjustable parameter used to better adjust the impact of the interaction degree between nodes during the propagation of the original topic on the propagation of derived topics. i (u) i ,u j ) represents user node u i Relative to user node u j Interactivity level, l i,j Represents user node u i and user node u j The relational layer number, M represents the user node u j The total number of original messages published in the topic.

[0043]

[0044] Let t be the time decay function, where t represents the current time. ma Represents user node u i For user node u j The time when action a is performed on the m-th message.

[0045] Furthermore, in step S4, a domain adaptation method based on TCA is used to perform data migration between the original topic propagation network and the derived topic propagation network, resulting in a nonlinear feature mapping function to achieve feature adaptation between the two topics. The calculation formula for the nonlinear feature mapping function is as follows:

[0046]

[0047] Dist(X′ s,X′ t ) represents the L2 norm between the original topic feature set and the derived topic feature set, given a feature mapping function, in order to minimize Dist(X′). s ,X′ t ) Optimize φ() and H, X′ for the objective. s Let X′ represent the original set of topic features after the feature mapping function. t Let φ() represent the set of derived topic features after passing through the feature mapping function, and x represent the feature mapping function. sr Let x represent the r-th feature in the original topic feature set. tb Let R represent the b-th feature in the derived topic feature set, R represent the number of features in the original topic feature set, B represent the number of features in the derived topic feature set, and H represent the latent feature representation from the original topic feature set to the derived topic feature set.

[0048] Furthermore, a belief function is constructed based on subjective cognitive game theory, and information entropy is used to measure the user's cognitive level, where:

[0049]

[0050]

[0051] Bef(u i ,t) represents user node u i The belief function at time t, where a0, a1, and a2 represent the importance coefficients of message influence, T(t) represents the topic popularity at time t, n represents the total number of user nodes, and M represents the number of user nodes u. j The total number of original messages published in the topic.

[0052] The beneficial effects of this invention are:

[0053] Although the derivative topic space contains a large amount of data, effective user behavior data is very sparse, especially in the early stages of a derivative topic's emergence, which inevitably affects the accuracy of the entire propagation model. This invention designs a sparse data compensation method for derivative topics based on transfer learning. To address the sparsity of user behavior data in the early stages of a derivative topic's emergence, preceding topics with richer user behavior data are used as the data transfer source to alleviate the data sparsity problem. Simultaneously, a domain adaptation method based on TCA is introduced to achieve feature adaptation between preceding topic data and derivative topic data, further improving the accuracy of the derivative topic propagation model.

[0054] In the context of derivative topic propagation, multiple complex relationships exist between entities such as users, messages, and networks, posing a challenge to effectively extracting and simplifying features. This invention proposes an iterative topic representation method, IT2vec, based on predecessor topic representation. Addressing the complexity of the derivative topic feature space, and fully considering the connections and differences between derivative and predecessor topics, it iteratively represents both types of topic spaces from the perspectives of predecessor and derivative topics, thus enriching the low-rank vectorization of the derivative topic feature space.

[0055] Before a derivative topic explodes, a similar preceding topic has already been circulating for a period of time, allowing users to accumulate some knowledge about this type of topic. This knowledge inevitably influences users' behavior regarding the derivative topic. This invention addresses the issue of users' accumulated knowledge about preceding topics before a derivative topic explodes by introducing subjective game theory to construct user cognitive influence and integrate it into the characteristics of the derivative topic. Finally, a CNN-based derivative topic propagation prediction model is constructed, which can accurately and effectively predict user group behavior at the next moment of the derivative topic's emergence. Attached Figure Description

[0056] Figure 1 This is a flowchart of the social network-derived topic propagation prediction method of the present invention;

[0057] Figure 2 This is a flowchart illustrating the iterative representation of the original topic and derived topics in an embodiment of the present invention.

[0058] Figure 3 A flowchart illustrating the adaptation of original and derived topic features in embodiments of the present invention. Detailed Implementation

[0059] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0060] This invention provides a method for predicting the spread of derivative topics on social networks based on transfer learning, comprising three parts: user feature extraction, feature transfer, and model building.

[0061] First, user features are extracted, specifically by obtaining topic data from public datasets, including basic information, network relationship information, and behavioral information (likes, follows, and reposts) of participating users, as well as text information under the topic; relevant attributes related to the topic are extracted from the topic data; and the IT2vec (Iterative Topic to vector) representation learning algorithm is proposed to iteratively represent the topic data, thereby vectorizing the derived topic feature space into a low-rank vector.

[0062] Based on the relevant attributes in different topic data obtained from the user feature extraction part, a general feature mapping is learned in the feature transfer part, thereby realizing feature adaptation between different topic data and completing feature transfer between topic data; since the derivative topic and its predecessor hot topic have a large similarity, the rich predecessor hot topic data can be used to train a model for predicting the spread of derivative topics, and then the derivative topic data can be used to test and adjust the model.

[0063] In the model building section, a derivative topic propagation prediction model is built based on subjective cognitive game theory and CNN to predict the propagation of derivative topics in the next moment.

[0064] In one embodiment, the process of a method for predicting the spread of social network-derived topics based on transfer learning is as follows: Figure 1 As shown, by inputting relevant data on trending topics and their derivative topics, the spread trend of derivative topics can be predicted, specifically including:

[0065] S1. Obtain topic data through public dataset websites or by using public API interfaces provided by social networks. The topic dataset includes original topic data and derived topic data.

[0066] S2. Extract relevant attributes from the original topic data and the derived topic data respectively, and construct the original topic propagation network and the derived topic propagation network;

[0067] Specifically, relevant attributes are extracted from the original topic data to obtain the topic propagation network, basic user attributes, message propagation popularity, and topic awareness of the original topic; relevant attributes are extracted from the derived topic data to obtain the topic propagation network, basic user attributes, message propagation popularity, and topic awareness of the derived topic.

[0068] S3. The IT2vec representation learning algorithm is used to iteratively represent the original topic propagation network to the derived topic propagation network, and the low-rank vectorized feature representation of the derived topic propagation network is obtained.

[0069] Specifically, the IT2vec representation learning method is used to process the original topic propagation network, mapping each user node in the network to a feature vector to form the original topic user feature matrix. Similarly, the derivative topic propagation network is processed to obtain the derivative topic user feature matrix. At the same time, the IT2vec representation learning algorithm is used to iteratively represent the original topic user feature matrix to the derivative topic user feature matrix, thereby making the feature space of the derivative topic propagation network low-rank vectorized.

[0070] S4. Using a domain adaptation method based on TCA, feature adaptation is performed on the original topic propagation network and the derived topic propagation network to obtain general feature data;

[0071] S5. Construct a CNN prediction model based on subjective cognitive game theory, train the CNN prediction model using general feature data, and test the trained CNN prediction model using the low-rank vectorized feature representation of the derivative topic propagation network.

[0072] S6. Use a tested CNN prediction model to predict the spread trend of current derivative topics, and obtain the user's cognitive influence and the user group behavior under the derivative topics at the next moment.

[0073] In one embodiment, based on the topic data (original topic data or derived topic data), further operations and mining are performed on user, message, and behavior data to obtain more detailed user-related expressions, that is, to obtain the relevant attributes of the topic data, represented as:

[0074] Topic dissemination network:

[0075]

[0076] in, U represents the topic propagation network at time t. t This represents the set of users participating in the topic at time t. U t The set of relational edges between users, where each relational edge is directed, and if there exists... This indicates that user u i Followed user u j That is, user u i User u j _ fans.

[0077] User basic properties UserProp(u i ):

[0078] UserProp(u i ) = [age(u i ),sex(u i ),follow(ui ),fans(u i )]

[0079] age(u i ) indicates the user u who participated in the topic. i age, sex(u i ) indicates the user u who participated in the topic. i The gender, follow(u i ) indicates the user u who participated in the topic. i Number of followers, fans(u i ) indicates the user u who participated in the topic. i The number of followers, user age, user gender, number of followers, and other attributes all have a certain impact on whether a user will participate in the dissemination of a topic.

[0080] Message spread popularity MsgHeat(m):

[0081]

[0082] Num[forward(m)] represents the number of forwards of the m-th original message published in the topic, and Num[fans(u i )] represents the number of followers of the original user who created the original message m, and ω and It is an adjustable hyperparameter, and ω, To mitigate the difference in impact between the number of reposts of an original message and the number of followers of the user who posted the original message, and because the popularity of a message gradually decreases over time, this embodiment introduces a time decay function. This simulates the change in message popularity over time, where λ is the popularity decay factor. The value of λ varies depending on the topic and is derived by fitting real data. m0 t represents the publication time of the original message m, and t represents the current time. Generally, the higher the popularity of a message, the more likely it is to attract other users to take action on it.

[0083] UserAware (u) i ):

[0084]

[0085] Num[original(u i )] for user u i The total number of original messages in the topic, Num[forword(u i )] represents user u i The total number of forwarded messages in the topic; Num[original(U N)] represents user u i User group U N The total number of all original messages in the topic, Num[forword(U N )] represents user u i User group U N The total number of all forwarded messages, where α and β are adjustable hyperparameters used to better fit the impact of original messages and forwarded messages on users' topic perception.

[0086] In one embodiment, since derivative topics evolve from common hot topics through multiple propagations, and are closely related to their predecessor topics in terms of both structure and content, the IT2vec representation learning algorithm is proposed from the perspective of the connection and differences between the original topic and the derivative topic. This algorithm iteratively represents the derivative topic from the original topic, and vectorizes the feature space of the derivative topic into a low-rank vector. Specifically, this includes:

[0087] S11. Obtain the basic user attributes of all users in the original topic propagation network, and calculate the edge weights between user nodes in the original topic propagation network based on the user interaction degree and the similarity of user basic attributes.

[0088] S12. Set the node traversal strategy by adjusting the edge weights between user nodes in the original topic propagation network; specifically including: randomly setting the node feature mapping function, and randomly selecting user node u in the original topic propagation network. a As the starting node, the edge weights between user nodes are used to calculate the probability of observation from user node u using the node walk observation formula. a The probability set of other user nodes that wander to its maximum depth range; select a group of user nodes u from this probability set. a User nodes with high relevance form a neighborhood user node set; each user node in the original topic propagation network is used as a starting node to perform the above operation once, resulting in a corresponding neighborhood user node set; using these neighborhood user node sets as training datasets, and then using a given objective function, the node feature mapping function is trained through backpropagation and gradient descent to obtain the node feature mapping function of the original topic propagation network, completing the first iteration; where, the maximum depth refers to the depth limit when a user node randomly walks to select neighboring user nodes, in this embodiment, the maximum depth is set to 2, that is, the starting node can walk to its two-hop neighboring nodes; when selecting a neighborhood user node set, the number of nodes in the neighborhood user node set is half the number of nodes within the maximum depth range of its corresponding starting node.

[0089] S13. Obtain the basic user attributes of all users in the derivative topic propagation network, and obtain the edge weights between user nodes in the derivative topic propagation network based on the interaction degree between user nodes in the original topic propagation network, the user interaction degree in the derivative topic propagation network, and the similarity of user basic attributes.

[0090] S14. Set a new node walking strategy by adjusting the edge weights between user nodes in the derivative topic propagation network. Specifically, this includes: calculating the set of neighboring user nodes of each user node in the derivative topic propagation network as training data; training the node feature mapping function of the original topic propagation network through backpropagation and gradient descent based on the node feature mapping function of the original topic propagation network obtained in S12 and the given objective function to obtain the node feature mapping function of the derivative topic propagation network; and further obtaining a new feature representation of the derivative topic propagation network through the node feature mapping function of the derivative topic propagation network. The second iteration is then completed.

[0091] Specifically, in the first and second iterations, a set of neighboring user nodes is used for training each time, and multiple training sessions are conducted in both iterations.

[0092] Specifically, during the Nth training iteration of the second iteration, the node feature mapping function obtained from the previous N-1 training iterations is used to select user nodes u that did not participate in the previous N-1 training iterations. b As the starting node, the edge weights between user nodes in the propagation network and the node traversal observation probability formula are used to calculate the value from user node u. b The probability set of other user nodes that wander to its maximum depth range; select a group of user nodes u from this probability set. b User nodes with high relevance form a set of neighboring user nodes, and user node u is calculated. b The log probability P(Ns(u) b )|f(u b By combining the objective function with backpropagation and gradient descent, the parameters of the node feature mapping function trained N-1 times are adjusted to obtain the node feature mapping function trained N times.

[0093] Specifically, the objective function is defined as:

[0094]

[0095] Wherein, Ns(u i ) represents user node u i The set of neighboring user nodes in the network, f(u i ) represents user node u i The node feature mapping function, logP(Ns(u i )|f(u i)) represents the feature mapping function f(u) for a given node. i Under the condition of ), observe user node u i Simultaneously observe its neighboring user node set Ns(u i The logarithmic probability of P(Ns(u)). Assuming the probability of observing a neighboring user node is independent of the probability of observing other neighboring user nodes, and modeling the conditional probability of a user node and each of its source neighbor pairs as a softmax unit, we quantify it using the dot product of features, and further... i )|f(u i )) is represented as:

[0096]

[0097]

[0098] in Represents user node u i and user nodes in its neighboring user node set dot product of features Represents user node u i , The relationships (direct or indirect) that exist in the topic propagation network E.

[0099] Specifically, in this embodiment, edge weights between user nodes in the topic propagation network are first constructed based on user basic attributes. Then, a node traversal strategy is set based on these edge weights. The node traversal observation probability of the current user node in the node traversal strategy is expressed as:

[0100]

[0101] Wherein, P(u j |u i ) represents the current user node u i The next-hop user node is u j The probability of i ≠ j, and α p,q (u i ,u j W(u) represents the walking bias. i ,u j ) represents user node u i and user node u j The edge weights between u and z are given by z, which represents the normalization constant, and E represents the set of edges in the topic propagation network. i ,u j ∈E represents user node u i and user node u j There are connections between them.

[0102] The wander bias is defined as:

[0103]

[0104] in, Represents user node u i To user node u j The distance is p,q, where p and q represent the biasing coefficients.

[0105] In this invention, it is assumed that the correlation and interaction level of basic user attributes between user nodes are related to the edge weights between user nodes. Therefore, user node u i and user node u j The formula for calculating the edge weight between them is:

[0106]

[0107] in, The correlation coefficient of user basic attributes is represented by mapping the user basic attributes of each user node to the feature vector space to obtain their respective user attribute vectors. The cosine similarity algorithm is then used to calculate the similarity between user attribute vectors, thus obtaining the correlation coefficient of user basic attributes between every two user nodes; Multidimensional(u i ,u j ) represents user node u i and user node u j User interaction score, calculated based on the interaction records between user nodes, is expressed as:

[0108]

[0109]

[0110] Since message publishing and user behavior are directional, calculating user interaction degree requires considering who interacts with whom, and the interaction degree is relative to whom. i (u) i ,u j ) represents user node u i Relative to user node u j Interactivity level, l i,j Represents user node u i and user node u j The number of relational layers, if user node u i Focus directly on user node u j Then l i,j =1; if user node u i Directly followed user node u k However, no attention was paid to user node u. j But user node uk I followed u directly j Then l i,j =2, and so on. M represents user node u j The total number of original messages published in the topic is subject to a time decay function, as interactions between user nodes are affected by time. t represents the current time, t ma This indicates the time when the m-th message is propagated.

[0111] I ma The behavior coefficient is defined in this embodiment as:

[0112]

[0113] Where 'a' represents user behavior, a=1 represents forwarding, a=2 represents commenting, and a=3 represents liking.

[0114] Specifically, such as Figure 2 As shown, the original topic propagation network includes user nodes 1, 2, 3, 4, and 5, with the following connections: Figure 2 As shown, the basic user attributes of user nodes are obtained, where user nodes 1 and 4 follow each other, and user nodes 2, 3, and 5 have the same IP address. The edge weights between user nodes in the original topic propagation network are calculated based on user interaction degree and similarity of basic user attributes.

[0115] A node feature mapping function is randomly set, and the set of neighboring user nodes for each user node in the original topic propagation network is calculated as training data. This node feature mapping function is trained multiple times using a given objective function to obtain the node feature mapping function of the original topic propagation network. In one training iteration, user node 4 in the original topic propagation network is selected as the starting node. Using the edge weights between user nodes in the original topic propagation network and the given node walk observation probability formula, the probability of user node 3 walking to node 2 is 0.5, the probability of user node 1 walking to node 3 is 0.3, and the probability of user node 4 walking to node 1 is 0.8. The probability set of users walking from user node 4 to other user nodes within its maximum depth range is calculated, including the probability of node 4 walking to node 3 (0.24), the probability of user node 4 walking to node 5 (0.2), and the probability of user node 4 walking to node 1 (0.8). From this probability set, a group of user nodes with high correlation to user node 4 are selected to form a set of neighboring user nodes, such as... Figure 2 As shown, the set of neighboring user nodes includes user node 1 and user node 3; based on the node feature mapping function trained in the previous training, the set of neighboring user nodes of user node 4 is used to optimize the objective function;

[0116] The derivative topic propagation network includes user nodes 1, 2, 3, 6, and 7, with connections as follows: Figure 2 As shown, there are overlapping users between the derived topic propagation network and the original topic propagation network, namely user nodes 1, 2, and 3. The basic user attributes of the user nodes are obtained. Among them, the ages of user nodes 3, 6, and 7 are similar, and the IP addresses of user nodes 2 and 3 are the same. Considering that user nodes 2 and 3, as well as user nodes 1 and 3, have certain interactions in the original topic propagation network, the edge weights between user nodes in the derived topic propagation network are calculated by combining the user interaction degree and the similarity of user basic attributes.

[0117] The set of neighboring user nodes for each user node in the derivative topic propagation network is calculated as training data. The node feature mapping function of the original topic propagation network is trained multiple times using a given objective function to obtain the node feature mapping function of the original topic propagation network. In one training iteration, user node 3 is selected as the starting node. Based on the edge weights between nodes, the probability of user node 3 moving to node 2 is 0.9, the probability of user node 3 moving to node 1 is 0.4, the probability of user node 3 moving to node 7 is 0.5, and the probability of user node 7 moving to node 6 is 0.2. The probability set of user node 3 moving to other user nodes within its maximum depth range is calculated. From this probability set, a group of user nodes with high correlation to user node 3 is selected to form the neighboring user node set. For example... Figure 2 As shown, the set of neighboring user nodes includes user node 2 and user node 7; based on the node feature mapping function trained in the previous training, the set of neighboring user nodes of user node 3 is used to optimize the objective function.

[0118] The final result can inherit the structural information in the original topic propagation network well, while also taking into full account the new propagation characteristics in the neighborhood of derived topics.

[0119] The traditional view is that training and testing data should come from the same dataset. However, this is almost impossible to achieve for derivative topics in their early stages of a trend, as sparse derivative topic data makes it difficult to train an effective propagation model. For original trending topics that have spread multiple times, this seems to be less of a challenge. Furthermore, because the data structure of derivative topics is highly similar to that of original trending topics, the abundant user data from the original trending topics can effectively compensate for the scarcity of effective data in derivative topics.

[0120] In one embodiment, a TCA-based method is introduced to learn the similarity of data features between different domains, enabling cross-domain data transfer between two types of topics. The two domains include relatively abundant and already labeled original topic-domain data: Where x sr ∈X sThe input is the low-rank vectorized feature representation of the original topic propagation network obtained through the IT2vec representation learning algorithm, y. sr ∈Y represents the corresponding output, i.e., the user's behavior towards the topic; and the relatively sparse, unlabeled derivative topic data: Where x tb ∈X t As input. Using P(X) s ) and Q(X t ) represent the data distribution of the original topic data and the derived topic data, respectively, where X s X represents the source domain feature set, which in this embodiment is also represented as the original topic feature set after the nodes of the original topic propagation network have been mapped by the node feature mapping function. t This represents the target domain feature set, which in this embodiment is also represented as the derivative topic feature set after the nodes of the derivative topic propagation network have been mapped by the node feature mapping function. The purpose of this embodiment is to... s ) and Q(X t Based on this, and combined with a TCA-based domain adaptation method, a domain-adaptive X is found. s With X t A common latent representation H exists between the two domains, which retains common configuration information after transformation. That is, there exists a non-linear feature mapping function φ:X→H, where X′ s ={x′ sr}={φ(x sr )},X′ t ={x′ tb}={φ(x tb We expect the final P(X′) to be... s )=Q(X′ t Therefore:

[0121]

[0122] We only need to minimize the distance between the two distributions to obtain an optimal nonlinear feature mapping function φ, such that in P(X) s )≠Q(X t In the case of ), P(Y) is obtained through the feature mapping function φ. s |X′ s )=P(Y t |X′ t Subsequently, a TCA-based eigenvalue decomposition algorithm is used to learn the optimal solution of the nonlinear feature mapping function φ.

[0123] Specifically, the process of training a prediction model by performing feature transfer is as follows: Figure 3 As shown:

[0124] First, the domain adaptation method based on TCA continuously learns the relationship and differences between source domain features and target domain features to obtain the feature mapping function;

[0125] Then, by utilizing the richer source domain features, a feature mapping function is used to obtain the general features after feature adaptation. These features are then used to train the subsequent prediction model, which involves performing processes such as convolution, pooling, and backpropagation of fully connected layers.

[0126] Finally, the trained prediction model can be used to make predictions on target domain data that is sparser.

[0127] User cognitive influence can be understood as the subjective impact of a topic's evolution on a user's participation, which, to some extent, guides the user's subsequent behavior. The term "subjectivism" means using probability and other numerical functions designed to represent uncertainty as tools to quantify an individual's beliefs about an event, without necessarily considering their potential randomness or repeatability. This invention quantifies changes in user cognition during topic dissemination by setting a belief function, and the user's next action can be viewed as a subjective cognitive game between the user and their friends.

[0128] Unlike traditional game theory, the strategies in subjective game theory are not predetermined. Participants decide their strategies through subjective learning. This is somewhat similar to the scenario studied in this invention, as derivative topics evolve from one or more hot topics through multiple rounds of dissemination. Users have some subjective understanding of these derivative topics, making their dissemination more flexible. Users' actions are not limited to simple, given strategies like forwarding or not forwarding; they can evolve further. Therefore, we cannot measure a user's cognitive influence simply by calculating the payoff of different strategies based on a given game theory framework. This invention constructs a belief function based on subjective game theory to calculate the trend of user belief changes. Finally, it uses information entropy theory combined with the belief function to induce user beliefs and further measure the user's cognitive level.

[0129] Common sense tells us that a user's belief change at time t is related to the message's popularity and the user's own awareness of the topic. Based on this, this invention constructs a belief function for the user at time t as follows:

[0130]

[0131] Where T(t) represents the popularity of the topic at time t, and a0, a1, a2 are preset coefficients that can truly reflect the importance of different features to the influence of the message. In reality, the spread of the message is affected by multiple factors. This paper uses a multivariate linear equation to fuse the features that affect the spread of derivative topics in order to achieve a better fit to the real situation.

[0132] Finally, due to the discrepancy in the amount of information delivered to users during the dissemination of a topic, user cognition fluctuates. Therefore, this invention introduces information entropy theory to quantify the uncertain factor of variability in user cognition during topic dissemination, thereby obtaining the user's cognitive influence at time t:

[0133]

[0134] This invention combines user features derived from iterative topic representation with a CNN-based adaptive group behavior prediction model based on topic evolution. The prediction of group behavior is defined as a binary classification problem. Using the network-wide feature matrix of derived topics for time period t, it predicts whether potential users in time period t+1 will participate in the spread of derived topics.

[0135] The input to the group behavior prediction model is the feature matrix of derived topics over time period t. The model consists of three parts: convolutional layers, pooling layers, and fully connected layers. First, the first layer is a convolutional layer, using a dual-channel convolutional kernel matrix. Perform convolution operation:

[0136]

[0137] Where Relu() represents a non-linear activation function, w i Represents the weight matrix. Let b be the derived topic feature vector corresponding to the i-th channel of the convolutional layer. i x is the bias value. conv This represents the output of the convolution operation.

[0138] Max pooling (maxPool(x)) is performed on the topic feature information after convolution. conv The pooling layer then aggregates the local features extracted by the convolutional layer. Finally, the output of the pooling layer is used as the input to the fully connected layer for further processing. The propagation behavior of the user group V at time t+1 is then derived using the following formula:

[0139]

[0140]

[0141] Where p(y=i|θ,h) represents the probability that users in the group participate in the dissemination of hot topics at time t+1, i takes the value 0 or 1, θ is the weight matrix, and h represents the output of the fully connected layer.

[0142] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "setting," "connection," "fixing," "rotation," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal connection of two components or the interaction between two components. Unless otherwise explicitly limited, those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0143] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the spread of derivative topics on social networks based on transfer learning, characterized in that, Includes the following steps: S1. Obtain topic data through public dataset websites or by utilizing public APIs provided by social networks. The topic data includes original topic data and derived topic data. S2. Extract relevant attributes from the original topic data and the derived topic data respectively. The relevant attributes include the topic propagation network, user basic attributes, message propagation popularity and topic awareness. Construct the original topic propagation network and the derived topic propagation network through the topic propagation network. S3. The IT2vec representation learning algorithm is used to iteratively represent the original topic propagation network to the derived topic propagation network, and the low-rank vectorized feature representation of the derived topic propagation network is obtained. S4. Using a domain adaptation method based on TCA, feature adaptation is performed on the original topic propagation network and the derived topic propagation network to obtain general feature data; In step S4, a domain adaptation method based on TCA is used to perform data migration between the original topic propagation network and the derived topic propagation network, resulting in a nonlinear feature mapping function to achieve feature adaptation between the two topics. The calculation formula of the nonlinear feature mapping function is as follows: Dist(X′ s ,X′ t ) represents the L2 norm between the original topic feature set and the derived topic feature set, given a feature mapping function, in order to minimize Dist(X′). s ,X′ t ) Optimize φ() and H, X′ for the objective. s Let X′ represent the original set of topic features after the feature mapping function. t Let φ() represent the set of derived topic features after passing through the feature mapping function, and x represent the feature mapping function. sr Let x represent the r-th feature in the original topic feature set. tb Let R represent the b-th feature in the derived topic feature set, R represent the number of features in the original topic feature set, B represent the number of features in the derived topic feature set, and H represent the latent feature representation from the original topic feature set to the derived topic feature set. S5. Construct a CNN prediction model based on subjective cognitive game theory, train the CNN prediction model using general feature data, and test the trained CNN prediction model using the low-rank vectorized feature representation of the derivative topic propagation network. S6. Use a tested CNN prediction model to predict the spread trend of the current derivative topic and obtain the spread trend of the derivative topic at the next moment.

2. The method for predicting the spread of social network derivative topics based on transfer learning according to claim 1, characterized in that, Relevant attributes extracted from topic data are represented as follows: Topic dissemination network: U represents the topic propagation network at time t. t This represents the set of users participating in the topic at time t. U t The set of relational edges between users, where each relational edge is directed; User basic properties UserProp(u i ): UserProp(u i )=[age(u i ),sex(u i ),follow(u i ),fans(u i )] age(u i ) indicates the user u who participated in the topic. i age, sex(u i ) indicates the user u who participated in the topic. i The gender, follow(u i ) indicates the user u who participated in the topic. i Number of followers, fans(u i ) indicates the user u who participated in the topic. i The number of fans; Message spread popularity MsgHeat(m): Num[forward(m)] represents the number of forwards of the m-th original message published in the topic, and Num[fans(u i )] represents the number of followers of the original user who created the original message m, and ω and For adjustable hyperparameters, Let t be a time decay function, λ be the heat decay factor, and t be a time decay function. m0 Let t be the publication time of the original message m, and t be the current time. UserAware (u) i ): Num[original(u i )] for user u i The total number of original messages in the topic, Num[forword(u i )] represents user u i The total number of forwarded messages in the topic; Num[original(U N )] represents user u i User group U N The total number of all original messages in the topic, Num[forword(U N )] represents user u i User group U N The total number of all forwarded messages, where α and β are adjustable hyperparameters.

3. The method for predicting the spread of social network derivative topics based on transfer learning according to claim 1, characterized in that, The specific process of step S3 includes: S31. Obtain the basic user attributes of all users in the original topic propagation network, and calculate the edge weights between user nodes in the original topic propagation network based on the user interaction degree and the similarity of user basic attributes. S32. By setting the node traversal strategy through the edge weights between user nodes in the original topic propagation network, the node feature mapping function of the original topic propagation network is obtained, and the first iteration is completed; S33. Obtain the basic user attributes of all users in the derivative topic propagation network, and obtain the edge weights between user nodes in the derivative topic propagation network based on the interaction degree between user nodes in the original topic propagation network, the user interaction degree in the derivative topic propagation network, and the similarity of user basic attributes. S34. Combine the edge weights between user nodes in the derivative topic propagation network to set a new node traversal strategy. Based on the node feature mapping function of the original topic propagation network obtained in S32, a new node feature mapping function for training is obtained, and the node feature mapping function of the derivative topic propagation network is obtained. The new feature representation of the derivative topic propagation network is further obtained through the node feature mapping function of the derivative topic propagation network.

4. The method for predicting the spread of social network derivative topics based on transfer learning according to claim 3, characterized in that, The node walk observation probability in the node walk strategy is expressed as: Wherein, P(u j |u i ) represents the current user node u i The next-hop user node is u j The probability of i ≠ j, and α p,q (u i ,u j W(u) represents the walking bias. i ,u j ) represents user node u i and user node u j The edge weights between u and z are given by z, which represents the normalization constant, and E represents the set of edges in the topic propagation network. i ,u j ∈E represents user node u i and user node u j There are connections between them. Represents user node u i To user node u j The distance is p,q, where p and q represent the biasing coefficients.

5. The method for predicting the spread of social network derivative topics based on transfer learning according to claim 4, characterized in that, User node u i and user node u j The formula for calculating the edge weight between them is: in, This represents the correlation coefficient of basic user attributes, Multual(u i ,u j ) represents user node u i and user node u j User interaction, Mutual′(u i ,u j ) represents user node u in the topic propagation network obtained in the previous iteration. i ,u j The degree of interaction between them, if the topic propagation network is the original topic propagation network, then Mutual′(u i ,u j ) = 0, μ is an adjustable parameter, interact(u i (u) i ,u j ) represents user node u i Relative to user node u j Interactivity level, l i,j Represents user node u i and user node u j The relational layer number, M represents the user node u j The total number of original messages published in the topic. in, Let t be the time decay function, where t represents the current time. ma Represents user node u i For user node u j The time when action a is performed on the m-th message.

6. The method for predicting the spread of social network derivative topics based on transfer learning according to claim 2, characterized in that, A belief function is constructed based on subjective cognitive game theory, and information entropy is used to measure the user's cognitive level, where: Bef(u i ,t) represents user node u i The belief function at time t, where a0, a1, and a2 represent the importance coefficients of message influence, T(t) represents the topic popularity at time t, n represents the total number of user nodes, and M represents the number of users u. i The total number of messages published under the derivative topic at time t.

Images