Cross-platform propagation model construction method based on multiple factors
Through multi-factor modeling and cascading models, the problem that existing technology is difficult to explain the reasons for information dissemination and cannot combine with platform differentiation is solved, and information dissemination simulation is achieved that is closer to reality, which improves the effectiveness and reliability of the model. Research and application of multi-social networks.
Patent Information
- Application Number
- CN202411170593.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-26
- Publication Date
- 2024-11-26
AI Technical Summary
The existing communication models are difficult to explain the reasons for information dissemination, cannot combine the differentiation of social media platforms and external factors, cannot effectively model information dissemination in multiple social networks, and it is difficult to support research needs in social computing fields such as public opinion guidance and information suppression.
Using a multi-factor cross-platform communication model construction method, the user's post generation modeling, personal preference factor modeling, social relationship factor modeling, platform environmental factor modeling and external influencing factor modeling, and combining multiple factors to construct information dissemination Cascading model to realize cross-platform propagation simulation.
It improves the effectiveness and reliability of the communication model, can be closer to the actual information dissemination process, supports the implementation and effectiveness verification of cross-domain communication control strategies, enhances the explanatory nature of the model, and is suitable for research in complex networks and social computing fields.
Smart Images

Figure CN119026477A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to a method for constructing a cross-platform communication model based on multiple factors, and belongs to the technical field of social network information communication simulation. Background Art
[0002] The rapid development of the Internet has made online social networks an important part of people's daily lives. Massive amounts of information are widely disseminated through online social networks. Understanding the laws of information dissemination on the Internet is of great significance to various practical applications and information governance in the Internet environment. How to study the laws of information dissemination on social networks has always faced huge challenges and has received widespread attention.
[0003] In recent years, online social networks have tended to diversify to meet people's usage needs, and social networks with different services have emerged on the Internet. Relevant surveys show that there are a large number of information dissemination phenomena between social networks. Based on multiple social networks, the study of users' information dissemination behavior can more deeply analyze the behavioral logic of users' participation in information diffusion and explore the potential motivations of users to use different social network platforms. However, there are relatively few studies on information dissemination in multiple social network scenarios. Compared with a single social network, the spread of information between multiple social networks has directional and preference characteristics. The directional characteristic is the extension of the dimension of information dissemination direction. In the multi-social network scenario, information can spread from the source network to the target network. Preference means that information in the multi-social network scenario may tend to spread to certain networks based on the theme content characteristics of different social network platforms. The research content of this patent will focus on the concept of information fusion, based on the expansion of feature dimensions based on the information fusion of each social network, and modeling the social network environment itself as a feature dimension.
[0004] In the field of social computing, if you want to study the development of social public opinion and analyze and predict the spread of online information, you first need to profile individual users in social networks, extract features from massive user data, and build connections and interactions between users based on this, so as to achieve the topology of social networks from points to lines and then to surfaces. Based on network topology, the communication process is as close to the real world as possible, and the study of communication dynamics based on this is of great practical significance for public opinion analysis and behavior identification on online networks, and is an important research direction in the field of social governance.
[0005] For traditional communication models, such as infectious disease models, linear threshold models, independent cascade models, etc., these models can well model the spread of information in social networks and are simple and easy to use. However, it is difficult to further explain the reasons for the spread of information, and it is impossible to combine cross-platform communication that is more in line with reality. There is also a lack of effective consideration of the interaction characteristics between users themselves, and it is difficult to make targeted adjustments to the platform differentiation of social media and external offline factors that closely interact with online communication. The simulation of information dissemination in the real world is not sufficient, and it is difficult to support subsequent research in social computing fields such as public opinion guidance and information suppression. Therefore, it is necessary to design a new communication modeling method to meet actual simulation needs. Summary of the invention
[0006] The purpose of the present invention is to provide a method for constructing a cross-platform communication model based on multiple factors to address the above-mentioned shortcomings, integrating multiple social computing and artificial intelligence technologies to obtain a systematic method that solves the current difficulties in modeling information-driven multi-social networks.
[0007] The main process of the present invention includes user post generation modeling, personal preference factor modeling, social relationship factor modeling, platform environment factor modeling, external influence factor modeling, information propagation cascade joint density function, user social network selection modeling and construction of information propagation cascade model.
[0008] First, the data collected by the social network platform is cleaned, mainly collecting various types of data including user static attributes, including nicknames, addresses, introductions, number of fans, etc.; user posting history, including self-posting and forwarded tweets; user interaction data, including tweet forwarding data. Then, the personal preference factors, social relationship factors, platform environment factors and external influencing factors are modeled respectively through the existing data, and the user's social network selection model is constructed by combining multiple factors, and a cross-platform communication model is realized based on this.
[0009] In order to achieve the above object, the technical solution of the present invention is as follows: a method for constructing a cross-platform communication model based on multiple factors, the method is as follows:
[0010] S1. Train the model by cleaning, integrating, and labeling, and preprocess the data required for calculation;
[0011] S2. Build the original topology of the cross-platform transmission network based on the existing data;
[0012] S3, modeling the user's personal preference factors, social relationship factors, platform environment factors and external influencing factors respectively;
[0013] S4, relying on the multi-factor modeling structure to train the post prediction model (i.e., social platform selection model) to predict the information transmission in social networks;
[0014] S5, combining the two-stage cross-domain propagation algorithm to build a propagation network of multiple social networks;
[0015] S6. Relying on the multi-factor communication model, simulate the information dissemination process on multiple platforms.
[0016] The specific data processing process of step S1 includes the following steps:
[0017] S1-1, starting from the source platform network root node, expand the user ID of the network topology based on the interaction data (forward, like);
[0018] S1-2, starting from the overlapping users of the source platform, according to the cross-domain accounts of the user on the target social network platform, the user ID of the network topology is expanded according to the interaction data (forwards, likes);
[0019] S1-3. Based on the collected user ID, crawl the user's static attributes on the corresponding platform, including screen_name, location, description, and friends_count;
[0020] S1-4, crawl the detailed tweet data of the user on the corresponding platform according to the collected user ID;
[0021] S1-5, crawling user interaction data, including the interaction edges between users generated by forwarding tweets and liking them;
[0022] S1-6. Clean the above data and organize it into a data set that is easy to call.
[0023] Step S2 of constructing the original network topology includes the following steps:
[0024] S2-1. Based on the crawled network edges, start from the root node and add a one-hop node that affects the root node upwards;
[0025] S2-2, starting from the root node, add the one-hop node affected by the root node downward;
[0026] S2-3, starting from one node, continue to join the two-hop nodes affected by the one-hop node;
[0027] S2-4, connect overlapping user pairs across platforms and add other potential cross-domain users.
[0028] Step S3 models the user's personal preference factors, social relationship factors, platform environment factors, and external influencing factors respectively, and includes the following steps:
[0029] S3-1, use the BERT model to calculate the similarity between the historical posting data of the user to be activated and the propagated tweets;
[0030] S3-2, use the BERT model to calculate the similarity between the static attributes of the user to be activated and the static attributes of the user who spreads the tweet, including nickname, address, introduction, number of fans, etc.;
[0031] S3-3. Count the out-degree and in-degree of nodes in the network topology as the activation threshold for calculating the susceptibility and influence of nodes;
[0032] S3-4, calculate the similarity between the main hot events of the social network platform and the original posts;
[0033] S3-5. Calculate the similarity between the active user groups on the platform and the nodes to be activated.
[0034] Step S4 relies on the multi-factor modeling structure to train the post prediction model (i.e., the social platform selection model), including the following steps:
[0035] S4-1. Through the edge data of the network topology, 10,000 user pairs that have interacted are selected as positive examples of the training set;
[0036] S4-2, select 10,000 pairs of users who have never interacted as counterexamples of the training set, and limit the out-degree of a single user to less than 50;
[0037] S4-3, calculate the similarity of the tweet list of the user pair as the training set (at least five tweets) through BERT;
[0038] S4-4, use BERT to calculate the static attribute similarity (nickname, user name, address, number of fans) of the user pairs used as training sets;
[0039] S4-5, use the lgb model (a type of decision tree model) to perform training based on the training set;
[0040] S4-6. Adjust the model parameters to improve the accuracy on the validation set.
[0041] The specific operation of S5 combining the two-stage cross-domain propagation algorithm to build a multi-social network propagation network includes the following steps:
[0042] S5-1, processing overlapping user pairs;
[0043] S5-2. Process cross-domain users to generate a candidate list.
[0044] The specific operation of step S5-1 includes the following steps:
[0045] S5-1-1, constructing a user pair according to a mapping between a source social network user account nickname and a target social network user account nickname;
[0046] S5-1-2. Construct an ID mapping dictionary from the source platform to the target platform based on the user.
[0047] The specific operation of step S5-2 includes the following steps:
[0048] S5-2-1. Generate a candidate list of potential overlapping accounts of cross-domain users of non-overlapping users on the corresponding platform;
[0049] S5-2-2, calculating the similarity between the cross-domain user and the potential account;
[0050] S5-2-3. Probability normalization of the candidate list is performed according to the similarity.
[0051] Step S6 simulates the process of information propagation on multiple platforms based on a multi-factor propagation model, including the following steps:
[0052] S6-1, select the initial node according to a specific algorithm;
[0053] S6-2, take the initial node as the activation node and start propagation;
[0054] S6-2. Get the final propagation result and save it in a file.
[0055] The specific operation of step S6-2 includes the following steps:
[0056] S6-2-1, adding the initial node to the sequence to be activated;
[0057] S6-2-2, traverse the sequence to be activated in turn, try to activate the next hop node, and add the successfully activated node to the list of nodes to be activated in the next round;
[0058] S6-2-3, proceed in sequence until convergence. In the last round, no new nodes to be activated are generated, and the propagation simulation ends.
[0059] The specific operation of step S6-2-2 includes the following steps:
[0060] S6-2-2-1: Determine whether the node is an overlapping user. If so, directly activate the account of the target platform;
[0061] S6-2-2-2: Determine whether the node is a cross-domain user. If so, generate a random number [0,1], select the target platform account as the overlapping user pair of the original account according to the candidate list where the random number is located, and activate it directly;
[0062] S6-2-2-3: Traverse neighbor nodes and generate activation probability according to the post prediction model. If the activation probability is greater than the activation threshold, it is considered as the activated target node, otherwise the activation fails.
[0063] Beneficial effects: The present invention can simulate the information propagation in the context of multiple social networks, and further support the specific implementation and effect verification of cross-domain propagation control strategies. For the research in the field of complex networks and social computing, the present invention has a very broad application prospect in large-scale cross-domain scenarios. It can integrate different platforms and different types of data to achieve more efficient and stable information propagation simulation. Compared with the existing technical solutions, the cross-social media propagation network is closer to reality; at the same time, the introduction of multi-dimensional factors including individual user factors and platform factors significantly improves the effectiveness and reliability of the propagation model, and can also support the propagation simulation under data constraints, and also improves the interpretability of the model. For the application of this technology in real-world scenarios, for example, on social media platforms, the model can integrate multi-platform data to provide a unified information flow experience and personalized recommendations; in the field of e-commerce, it can support multi-channel marketing and optimize supply chain management; in the field of news dissemination, the model helps content distribution, false information detection and customized content push, etc.; for public opinion control, the model can predict the propagation path of hot information on social networks, and conduct expected verification of the control and guidance strategies for planned implementation, and feedback the results to support the next step of delivery. BRIEF DESCRIPTION OF THE DRAWINGS
[0064] Figure 1 It is a schematic diagram of the overall process of the present invention;
[0065] Figure 2 Schematic diagram of the cross-domain communication process. DETAILED DESCRIPTION
[0066] Example: Refer to the attached Figure 1 In this embodiment, a method for constructing a cross-platform communication model based on multiple factors is described as follows:
[0067] Train the model by cleaning, integrating, and labeling, and preprocess the data required for calculation
[0068] Starting from the root user, crawl the user IDs that interact with the user;
[0069] For users who actively interact with the root user (like, forward), it is used as the previous hop node; for users who actively interact with the root user, it is used as the next hop node;
[0070] Continue to traverse the next hop node downward to find the second hop node;
[0071] Select user static attributes according to the platform, and select multiple attribute data including nickname, address, introduction, number of fans, etc. for crawling.
[0072] Use regular matching to clean data. Rely on Python and regular expressions to clean user attributes of garbled characters, meaningless emoticons, meaningless test English alphanumeric strings, etc.
[0073] 2) Construct the original network topology. Add the root node, previous hop node, next hop node, second hop node and overlapping user pairs to the network topology to form a connected dual social network topology.
[0074] 3) Modeling of user’s personal preference factors, social relationship factors, platform environment factors and external influencing factors.
[0075] 3-1) For modeling of personal preference factors, given a user posting event e and a user's historical posting event e', this study uses sim(c,c') to represent the similarity of the content of the two posting events. As time t passes, this effect will weaken. In the model, sim(c,c') is combined with the decay function f(tt',γ ins ) to reflect this change. Therefore, the intensity function under the influence of this factor is modeled as:
[0076]
[0077] Among them, sim(c′,c) represents the similarity between the content c′ and c. f(tt′,γ ins ) is the attenuation factor function.
[0078] 3-2) For the modeling of social relationship factors, given a user posting event e and a user posting event e′ under a given network s, the probability of a user posting event e is related to the influence of the user he follows and his own susceptibility. This study uses influence a u′ and susceptibility u Multiply to express this relationship, and also introduce the attenuation function f(tt′,γ soc ). Therefore, the intensity function under this factor can be modeled as:
[0079]
[0080] Among them, a u′ is the influence of user u, r u is the susceptibility of user u. f(tt′,γ soc ) is the attenuation factor function.
[0081] 3-3) For the modeling of platform environment factors, the intensity function based on the platform heat factor is modeled as:
[0082]
[0083] Among them, k S represents the influence of social network s, r u For user susceptibility, pop S C (t) is the association heat of social network s to the user's to-be-published content c at time t.
[0084] 3-4) For the modeling of external influencing factors, the intensity function under the external influencing factors is modeled as:
[0085] λ ext (u, c, t) = k0r u M c (t)
[0086] Among them, k O Represents the influence of external factors, r u is the user's susceptibility, M C (t) is the relevance level to the content c to be published by the user at time t.
[0087] 4) For the training and prediction of the post prediction model, use models such as LightGBM, bert-base-nli-mean-tokens;
[0088] 4-1) Extract the edge data in the existing network topology, and sort out the required user tweet list and static attribute dictionary according to the user ID;
[0089] 4-2) According to the edges, the user ID pairs that have interacted are sorted out as positive examples in the training set (userid_1, userid_2), where the first ID of the tuple is the parent node, that is, the activated node in the interaction.
[0090] 4-3) Sort out the user ID pairs that do not generate edges as counterexamples in the training set (userid_1, userid_2).
[0091] 4-4) Use the open source bert-base-nli-mean-tokens model to calculate the static attributes of both user links and the similarity between tweet lists, which constitute the input features of the post prediction model.
[0092] 4-5) Use numpy and Kfold in sklearn to divide the training set and validation set, and use lightgbm to import the lgb model.
[0093] 4-6) Use the training set to train the LightGBM model, use each similarity as a reference feature during training, and use binary classification as the classification method.
[0094] 4-7) Adjust parameters such as 'gbdt', 'rf', and learning rate to improve the accuracy of the model on the training set and validation set while preventing overfitting.
[0095] 4-8) Select the best model and save it as a txt file through model.save, and call it when calculating the propagation probability later.
[0096] 5) Process overlapping user pairs, determine the source platform and target platform of the propagation process, use the user ID of the source platform as the key of the mapping dictionary, and the user ID of the target platform as the value;
[0097] 5-1) For potential cross-domain users, a list of potential cross-domain user accounts on the target platform for each source platform social network account is obtained through a cross-domain user matching system based on embedding correction;
[0098] 5-2) Perform normalization based on the similarity between cross-domain users to obtain a mapping list from each source platform to the target platform, where the users on each target platform correspond to a specific interval on [0,1];
[0099] 6) Determine the cross-domain propagation simulation process from the source platform to the target platform. First, it is necessary to calculate the activation probability between nodes in the network topology through the publication prediction model, determine the initial activation node through a specific algorithm, and then start from the initial node and try to activate neighbor nodes in an independent cascade manner, and record the node IDs activated in each round. Finally, when the entire propagation process converges, that is, no new activated nodes are generated, the propagation is completed and the result graph is output.
[0100] 6-1) Traverse the edges between the network topologies to be propagated;
[0101] 6-2) Sort out the static attributes (nickname, address, introduction, number of fans) and tweet list required to calculate the activation probability based on the parent-child node IDs of the edges;
[0102] 6-3) Use the bert-base-nli-mean-tokens model to calculate the similarity of each attribute between the parent and child nodes;
[0103] 6-4) Use the trained post prediction model to predict the activation probability between nodes;
[0104] 6-5) Use a specific algorithm to select the initial node of the activation process;
[0105] 6-6) Start the propagation simulation from the initial activation node, refer to the attached Figure 2 .
[0106] 6-6-1) Add the initial node to the sequence to be activated;
[0107] 6-6-2) Traverse each node in the sequence to be activated in this round in turn;
[0108] 6-6-2-1) Traverse the neighbor nodes of each node. If the calculated activation probability is greater than the activation threshold, add the neighbor node to the next round of nodes to be activated;
[0109] 6-6-2-2) Determine whether the node is in an overlapping user pair. If so, directly add the account node of the node on the corresponding platform to the next round of nodes to be activated;
[0110] 6-6-2-3) Determine whether the node is a cross-domain user of a non-overlapping user. If so, generate a random number between [0, 1] and add the user node of the target platform corresponding to the interval where the number is located to the next round of nodes to be activated;
[0111] 6-6-3) When the traversal of the nodes to be activated in this round is completed, the next round of traversal begins, and the original platform node ID and the target platform node ID activated in each round are recorded;
[0112] 6-6-4) When there are no nodes to be activated in the next round, the propagation converges and the final result is output.
[0113] It should be noted that the above embodiments are not intended to limit the protection scope of the present invention, and equivalent changes or substitutions made on the basis of the above technical solutions all fall within the protection scope of the claims of the present invention.
Claims
1. A method for constructing a cross-platform communication model based on multiple factors, characterized in that: The steps include: S1. Train the model by cleaning, integrating, and labeling, and preprocess the data required for calculation; S2. Build the original topology of the cross-platform transmission network based on the existing data; S3, modeling the user's personal preference factors, social relationship factors, platform environment factors and external influencing factors respectively; S4, relying on the multi-factor modeling structure to train the post prediction model (i.e., social platform selection model) to predict the information transmission in social networks; S5, combining the two-stage cross-domain propagation algorithm to build a propagation network of multiple social networks; S6. Relying on the multi-factor communication model, simulate the information dissemination process on multiple platforms.
2. The method for constructing a cross-platform communication model based on multiple factors according to claim 1 is characterized in that: The specific data processing process of step S1 includes the following steps: S1-1, starting from the source platform network root node, expand the user ID of the network topology based on the interaction data (forward, like); S1-2, starting from the overlapping users of the source platform, according to the cross-domain accounts of the user on the target social network platform, the user ID of the network topology is expanded according to the interaction data (forwards, likes); S1-3. Based on the collected user ID, crawl the user's static attributes on the corresponding platform, including screen_name, location, description, and friends_count; S1-4, crawl the detailed tweet data of the user on the corresponding platform according to the collected user ID; S1-5, crawling user interaction data, including the interaction edges between users generated by forwarding tweets and liking them; S1-6. Clean the above data and organize it into a data set that is easy to call.
3. The method for constructing a cross-platform communication model based on multiple factors according to claim 1, characterized in that: Step S2 of constructing the original network topology includes the following steps: S2-1. Based on the crawled network edges, start from the root node and add a one-hop node that affects the root node upwards; S2-2, starting from the root node, add the one-hop node affected by the root node downward; S2-3, starting from one node, continue to join the two-hop nodes affected by the one-hop node; S2-4, connect overlapping user pairs across platforms and add other potential cross-domain users.
4. The method for constructing a cross-platform communication model based on multiple factors according to claim 1, characterized in that: Step S3 models the user's personal preference factors, social relationship factors, platform environment factors, and external influencing factors respectively, and includes the following steps: S3-1, use the BERT model to calculate the similarity between the historical posting data of the user to be activated and the propagated tweets; S3-2, use the BERT model to calculate the similarity between the static attributes of the user to be activated and the static attributes of the user who spreads the tweet, including nickname, address, introduction, and number of fans; S3-3. Count the out-degree and in-degree of nodes in the network topology as the activation threshold for calculating the susceptibility and influence of nodes; S3-4, calculate the similarity between the main hot events of the social network platform and the original posts; S3-5. Calculate the similarity between the active user groups on the platform and the nodes to be activated (posting topics and static attributes).
5. The method for constructing a cross-platform communication model based on multiple factors according to claim 1, characterized in that: Step S4 relies on the multi-factor modeling structure to train the post prediction model (i.e., the social platform selection model), including the following steps: S4-1. Through the edge data of the network topology, 10,000 user pairs that have interacted are selected as positive examples of the training set; S4-2, select 10,000 pairs of users who have never interacted as counterexamples of the training set, and limit the out-degree of a single user to less than 50; S4-3, calculate the similarity of the tweet list of the user pair as the training set (at least five tweets) through the BERT model; S4-4, calculate the static attribute similarity (nickname, user name, address, number of fans) of the user pairs used as training sets through the BERT model; S4-5, use the lgb model (a type of decision tree model) to perform training based on the training set; S4-6. Adjust the model parameters to improve the accuracy on the validation set.
6. The method for constructing a cross-platform communication model based on multiple factors according to claim 1, characterized in that: The specific operation of S5 combining the two-stage cross-domain propagation algorithm to build a multi-social network propagation network includes the following steps: S5-1, processing overlapping user pairs; S5-2, processing cross-domain users to generate candidate lists, The specific operation of step S5-1 includes the following steps: S5-1-1, constructing a user pair according to a mapping between a source social network user account nickname and a target social network user account nickname; S5-1-2, construct an ID mapping dictionary from the source platform to the target platform according to the user; The specific operation of step S5-2 includes the following steps: S5-2-1. Generate a candidate list of potential overlapping accounts of cross-domain users of non-overlapping users on the corresponding platform; S5-2-2, calculating the similarity between the cross-domain user and the potential account; S5-2-3. Probability normalization of the candidate list is performed according to the similarity.
7. The method for constructing a cross-platform communication model based on multiple factors according to claim 1, characterized in that: Step S6 simulates the process of information propagation on multiple platforms based on a multi-factor propagation model, including the following steps: S6-1, select the initial node according to a specific algorithm; S6-2, take the initial node as the activation node and start propagation; S6-2. Get the final propagation result and save it in a file.
8. The method for constructing a cross-platform communication model based on multiple factors according to claim 1, characterized in that: The specific operation of step S6-2 multi-factor propagation includes the following steps: S6-2-1, adding the initial node to the sequence to be activated; S6-2-2, traverse the sequence to be activated in turn, try to activate the next hop node, and add the successfully activated node to the list of nodes to be activated in the next round; S6-2-3, proceed in sequence until convergence. In the last round, no new nodes to be activated are generated, and the propagation simulation ends.
9. The method for constructing a cross-platform communication model based on multiple factors according to claim 8, characterized in that: The specific operation of the node activation process in step S6-2-2 includes the following steps: S6-2-2-1: Determine whether the node is an overlapping user. If so, directly activate the account of the target platform; S6-2-2-2: Determine whether the node is a cross-domain user. If so, generate a random number [0,1], select the target platform account as the overlapping user pair of the original account according to the candidate list where the random number is located, and activate it directly; S6-2-2-3: Traverse neighbor nodes and generate activation probability according to the post prediction model. If the activation probability is greater than the activation threshold, it is considered as the activated target node, otherwise the activation fails.
