A Method for Predicting the Propagation Power of Information Mediation Based on Triple Competition Dynamics

By using a triplet competitive dynamics approach, we can identify and predict the evolution trend of information intermediary nodes' propagation power in directed social networks, thus solving the problem of decision lag in existing technologies and enabling real-time optimization of social platform delivery.

CN123118086APending Publication Date: 2026-11-17DATA SPACE RES INST
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611602585.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-16
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies cannot effectively identify and predict the evolution trend of the propagation power of information intermediary nodes in directed social networks, resulting in advertising traffic distribution and KOL placement decisions lagging behind changes in network structure.

Method used

Based on the competitive dynamics of triplet, a directed social network is constructed to identify information intermediary nodes and use graph neural networks to predict their propagation power decay and enhancement probabilities, and combine expansion behavior characteristics for dynamic prediction.

Benefits of technology

It enables dynamic prediction of the propagation power of information intermediary nodes, supports real-time delivery decisions on social platforms, reduces computational overhead, and improves the accuracy and efficiency of decision-making.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure FT_1
    Figure FT_1
  • Figure SMS_224
    Figure SMS_224
  • Figure SMS_257
    Figure SMS_257
Patent Text Reader

Abstract

This invention discloses a method for predicting the propagation power of information intermediaries based on triplet competition dynamics. It constructs a directed social network using user social behavior data collected from social platforms, with directed edges pointing from content producers to content consumers. The directed edges are weighted, aggregated, and their out-degrees are normalized to obtain normalized edge weights. Based on these normalized edge weights, directed communities are partitioned to obtain community tags for each node. Information intermediary nodes are identified based on their community tags, and their directed information intermediary triples are enumerated. Competitive motivation features are extracted for each triple, and embedding vectors for each node are obtained using a graph neural network. The closure probability of each triple within the prediction window is determined based on its features and embedding vectors. This method can dynamically predict the evolution trend of the propagation power of information intermediary nodes in a directed social network and can be applied to traffic distribution and KOL (Key Opinion Leader) budget allocation on social platforms.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of graph machine learning and social network analysis, and in particular to a method for predicting the propagation power of information mediators based on triplet competition dynamics. Background Technology

[0002] On social media platforms like Weibo and Twitter, the follow relationships between users are asymmetrical, with information flowing in the opposite direction from the followed to the follower. In these directed social networks, there exists a special type of node—one that is simultaneously followed by multiple unconnected user communities, thus acting as a conduit and amplifier of information between these communities. Content posted or forwarded by these nodes can cross community boundaries, reaching previously isolated user groups, resulting in higher information dissemination efficiency than ordinary nodes. In business scenarios such as advertising traffic distribution and brand KOL placement on social media platforms, platforms need to identify high-value traffic nodes with cross-community dissemination capabilities to optimize ad placement weight allocation, while brands need to select KOLs with rising influence from among multiple potential partners to improve the return on marketing investment.

[0003] Existing technologies primarily address these needs from two directions. The first type of method calculates node importance metrics based on network topology, including centrality measures such as degree centrality, betweenness centrality, and PageRank, as well as information propagation influence estimation based on independent cascade models or linear threshold models. The second type of method constructs a static scoring system based on user historical behavior data, using statistical indicators such as the number of followers, content interaction rate, and historical reposts to comprehensively score and rank opinion leaders. However, both of these methods use a snapshot of the network at the current or historical moment as the analysis object, outputting a static assessment of a node's importance or influence at a specific instant. When the social network structure changes over time—for example, when a new direct connection path emerges between two previously unconnected communities, or when new information intermediary nodes join the network—the cross-community propagation capabilities of previously high-influence nodes may be bypassed and rapidly decline. Evaluation systems relying on static indicators cannot capture this evolutionary trend, causing traffic distribution weights or KOL budget allocation decisions to lag behind actual changes in the network structure.

[0004] In recent years, research in the field of dynamic network embedding has begun to focus on the temporal evolution modeling of network topology. One representative method is learning dynamic network embeddings by modeling the triplet closure process. Its core idea is to treat the formation of new edges in the network as the cumulative result of triplet closure events, and use this mechanism as a self-supervised signal to train node embedding vectors. However, the design goals of this type of method differ fundamentally from the propagation power prediction task in social networks: its optimization objective is to learn a general node representation that can encode network evolution patterns, rather than directly predicting the direction of propagation power changes for specific types of nodes; it treats all nodes in the network equally in triplet closure modeling, without distinguishing the different roles of information intermediary nodes and ordinary nodes in structural evolution; its output is the node embedding vector itself, without involving any quantifiable propagation power prediction indicators, and cannot directly serve deployment decisions. Therefore, while this type of method has value in dynamic network modeling technology, it cannot solve the specific business problems faced by social platforms and brands.

[0005] Furthermore, existing methods for identifying nodes that act as information bridges in directed social networks primarily draw upon structural hole theory and its derivative algorithms. However, these methods are mainly based on undirected graph design and do not distinguish between the different semantics of information inflow and outflow directions. In practical applications, some methods symmetrically convert the directed network into an undirected network before calculating the structural hole index. This approach loses the crucial dimension of information propagation direction. For example, a node being followed by multiple communities (in-degree structural hole) and a node actively forwarding content to multiple communities (out-degree structural hole) are two structures with drastically different information propagation mechanisms, but they cannot be distinguished after symmetry conversion. This loss of directional information results in the identified nodes lacking clear propagation semantic support, making it difficult to accurately reflect the actual bridging role of nodes in directed information flows.

[0006] In summary, existing technologies generally rely on static indicators and lack the ability to predict the evolution trend of propagation power. Although existing dynamic network embedding methods involve triple closure processes, their objective functions, node role differentiation, and application outputs are completely different from the prediction task focused on in this application. Furthermore, there is a lack of dedicated methods for identifying information intermediary nodes in directed networks. Therefore, how to independently predict the two evolution trends of propagation power decline and enhancement for the special role of information intermediary nodes based on the evolution mechanism of directed social networks, and directly map the prediction results into resource allocation parameters that can be used by advertising systems and delivery decision-making systems, is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0007] To address the technical problems existing in the background art, this invention proposes a method for predicting the propagation power of information mediators based on the competitive dynamics of triplet.

[0008] The information mediation propagation power prediction method based on triplet competition dynamics proposed in this invention includes the following steps: We acquire user social behavior data collected from social platforms and construct a directed social network. The nodes of the directed social network are users, and the directed edges point from content producers to content consumers. We perform weighted aggregation and out-degree normalization on the directed edges to obtain normalized edge weights. Directed community partitioning is performed on the directed social network based on normalized edge weights to obtain the community labels of each node; Identify information intermediary nodes based on community tags and enumerate their directed information intermediary triples. Each triple consists of two consecutive directed edges from the source node to the intermediary node and then to the audience node. For each triple, features are extracted to quantify the competitive motivation of the audience node to bypass the intermediary node and directly connect to the source node, and the embedding vector of each node is obtained through a graph neural network. Based on the features and embedding vectors of each triplet, the triplet is processed by the first multilayer perceptron and the second multilayer perceptron respectively to determine the closure probability of each triplet within the prediction window; Aggregate the closure probabilities of all triples of each intermediary node to obtain the propagation decay probability of each intermediary node; The expansion behavior features of each intermediary node are obtained, and the propagation power enhancement probability of each intermediary node is predicted by combining the embedding vector of each intermediary node. The expansion behavior features are used to quantify the trend of each intermediary node establishing new cross-community connections within the observation window. The probability of transmission decay and the probability of transmission enhancement are output independently as bi-objective prediction results.

[0009] Preferably, the user social behavior data includes behavior-associated users, behavior types, and behavior timestamps, wherein behavior-associated users include content consumers and content producers, and behavior types include following, forwarding, commenting, and liking; The construction of the directed social network specifically includes: With users as nodes and social behaviors between users as directed edges, the directed edges point from content producers to content consumers. When a content consumer follows a content producer, a directed edge points from the content producer to the content consumer; when a content consumer shares content produced by a content producer, a directed edge points from the content producer to the content consumer; when a content consumer comments on content produced by a content producer, a directed edge points from the content producer to the content consumer; when a content consumer likes content produced by a content producer, a directed edge points from the content producer to the content consumer. For behavior data related to following relationships, if the data collected by the social platform cannot obtain the specific timestamp of establishing the following relationship, then the time when the following relationship was first collected shall be used as the behavior timestamp of the directed edge.

[0010] Preferably, the weighted aggregation specifically involves: setting weight coefficients for follow, forward, comment, and like behaviors respectively, and using the sum of the products of the occurrence frequency of each behavior type on the same edge and the corresponding weight coefficient as the aggregation weight of the directed edge; the out-degree normalization specifically involves: dividing the aggregation weight of each out-edge of each node by the sum of the aggregation weights of all out-edges of that node, so that the sum of the normalized edge weights of each out-edge of that node is 1.

[0011] Preferably, the step of performing directed community partitioning on the directed social network based on normalized edge weights to obtain the community label of each node specifically includes: The directed social network is converted into an undirected projection, and the Louvain algorithm is run on the undirected projection to obtain the first community label of each node. Run the strongly connected component algorithm on a directed social network to obtain the strongly connected component labels of each node; For each community corresponding to the first community label, if the community contains multiple strongly connected components, the ratio of the average directed edge density between these strongly connected components to the average directed edge density within these strongly connected components is calculated. When this ratio is lower than a preset self-consistency threshold, the community is split into multiple sub-communities according to the boundaries of the strongly connected components, and the split sub-communities are used as the updated community labels of each node. For communities that do not meet the splitting conditions, their community labels remain unchanged.

[0012] Preferably, the directed community partitioning further includes: When the number of nodes in a strongly connected component is less than the preset minimum community size threshold, the strongly connected component is not treated as an independent community. Instead, the directed edge density between the strongly connected component and other adjacent communities is calculated, and the component is merged into the adjacent community with the highest directed edge density. When the bidirectional directed edge density across strongly connected components between two different first communities exceeds a preset merging threshold, the two first communities are merged into the same community, and the merged community is used as the community label for each node within that community.

[0013] Preferably, the source node belongs to a first community, and the audience node belongs to a second community; the first community and the second community are different communities. The step of identifying information intermediary nodes based on community tags and enumerating their directed information intermediary triples specifically includes: The directed edge density matrix between communities is calculated based on normalized edge weights. The elements in the directed edge density matrix between communities are the directed edge density from the first community to the second community, which is equal to the sum of all normalized edge weights from the first community to the second community divided by the product of the number of nodes in the first community and the number of nodes in the second community. Information intermediary nodes are identified based on the directed edge density matrix between communities and a preset community isolation threshold. For any node, if the node does not belong to the first community or the second community, and there are directed edges from nodes in the first community to the node and from the node to nodes in the second community, and the density of bidirectional directed edges between the first community and the second community is lower than the community isolation threshold, then the node is marked as an information intermediary node. Enumerate the directed information mediation triples for each information mediation node. For a node marked as an information mediation node, enumerate all triples that satisfy the following conditions: the source node belongs to the first community, the audience node belongs to the second community, there exists a directed edge from the source node to the mediation node and a directed edge from the mediation node to the audience node, and there is no directed edge from the source node directly to the audience node. All triples that satisfy the above conditions constitute the set of directed information mediation triples for that mediation node.

[0014] Preferably, the features include information complementarity, indirect perception, and connection strength; the extraction of features for each triple to quantify the competitive motivation of the audience node to bypass the intermediary node and directly connect to the source node specifically includes: Extract the information complementarity between the community to which the source node belongs and the community to which the audience node belongs. The information complementarity is the quotient obtained by subtracting the size of the intersection of the sets of objects of interest of the first community and the second community from the size of the union. Extract the structural dependence degree of the intermediary node on the content of the first community. The structural dependence degree is the quotient obtained by dividing the sum of the normalized edge weights from the first community in the incoming edges of the intermediary node by the sum of the normalized edge weights of all incoming edges of the intermediary node. Extract the coverage of the intermediary node to the second community. The coverage is the quotient obtained by dividing the number of users in the second community who pay attention to the intermediary node by the total number of people in the second community. The product of the degree of structural dependence and the degree of coverage is used as the indirect perception. Extract the connection strength between the intermediary node and the source node and the audience node respectively. The connection strength is the average of the normalized edge weight from the source node to the intermediary node and the normalized edge weight from the intermediary node to the audience node. Information complementarity, indirect perception, and connection strength are used as features to quantify competitive motivation.

[0015] Preferably, obtaining the embedding vector of each node through the graph neural network specifically involves: For each node, an initial embedding vector is constructed. The initial embedding vector shall contain at least the node’s in-degree, out-degree, one-hot encoding of its community, number of active days within the observation window, and number of published messages. In each layer of the graph neural network, messages are aggregated along the in-degree neighbor direction and the out-degree neighbor direction, respectively. The in-degree neighbor is other nodes that are followed by this node, and the out-degree neighbor is other nodes that follow this node. The message in each direction is calculated by the message function based on the dynamic edge features of the connection between the current embedding vector of the neighbor node and the node. The in-degree direction aggregation result, the out-degree direction aggregation result, and the current embedding vector of the node are concatenated and then transformed by the Message Passing Perceptron (MLP) to obtain the embedding vector of the next layer of the node. The dynamic edge features are calculated based on normalized edge weights and behavior timestamps in an exponential decay manner. The exponential decay is an exponential decay function with a base of the natural constant and an exponent of a negative learnable decay rate multiplied by the difference between the current time and the most recent behavior timestamp.

[0016] Preferably, the step of determining the closure probability of each triple within the prediction window by processing its features and embedding vectors through a first multilayer perceptron and a second multilayer perceptron, respectively, specifically includes: The embedding vectors of the audience node, the source node, the intermediary node, the information complementarity, the indirect perception, and the connection strength are concatenated and input into the first multilayer perceptron to obtain the approximate competitive payoff value of each triple. The approximate competitive payoff value is used to characterize the payoff of the audience node to bypass the intermediary node and directly connect to the source node. The embedding vector of the intermediary node is input into the second multilayer perceptron and mapped to obtain the adaptive closure threshold of the intermediary node. The adaptive closure threshold is used to characterize the different benefit thresholds that different intermediary nodes need to reach to transform competitive benefits into actual closure behavior. The difference between the approximate competitive payoff and the adaptive closure threshold is mapped using the sigmoid function to obtain the closure probability of the triple within the prediction window. The aggregation of the closure probabilities of all triples of each intermediary node to obtain the propagation decay probability of each intermediary node specifically includes: For each second community, the closure probability of all triples in the second community is compared, and the maximum value is taken as the bypass probability of the second community. The bypass probability is used to characterize the trend strength of the second community as a whole bypassing the intermediary node and directly accessing the community information source to which the source node belongs. Calculate the complement of the bypass probability of each second community, multiply the complements of all bypass probabilities of the second community together, and then subtract the result of the multiplication by one to obtain the propagation power decay probability of the intermediary node. The propagation power decay probability is used to characterize the overall probability that the intermediary node has at least one cross-community information channel bypassed within the prediction window, thus causing its propagation power to decay.

[0017] Preferably, the step of obtaining the expansion behavior features of each intermediary node, and combining the embedding vectors of each intermediary node to predict the probability of enhanced propagation power of each intermediary node, wherein the expansion behavior features are used to quantify the trend of each intermediary node establishing new cross-community connections within the observation window, specifically includes: Obtain the expansion behavior feature vector of each intermediary node. The expansion behavior feature vector includes the intensity of expansion to new audiences, the breadth of community coverage, the expansion speed of acquiring new content sources, and historical expansion records. The strength of expansion to new audiences is the quotient of the number of new outgoing edges of the intermediary node within the observation window divided by the length of the observation window; the breadth of community coverage is the percentage of all outgoing edges of the intermediary node pointing to new community nodes that have not been connected in the past; the expansion speed of acquiring new content sources is the quotient of the number of incoming edges of the intermediary node from new communities that have not been contacted in the past within the observation window divided by the length of the observation window; the historical expansion record is a binary indicator of whether the current set of directed information intermediary triples has added at least one directed information intermediary triple compared to the previous observation window; The embedding vectors of each intermediary node are concatenated with the feature vectors of the expansion behavior. The concatenated vectors are then mapped through a third multilayer perceptron and activated by the sigmoid function to obtain the propagation enhancement probability of each intermediary node within the prediction window.

[0018] This invention proposes an information intermediary propagation power prediction method based on triplet competition dynamics. By constructing a directed social network and identifying information intermediary nodes, it models the evolution trend of the propagation power of these nodes based on triplet competition dynamics. This dynamically predicts the decline and enhancement directions of a node's cross-community propagation ability, overcoming the shortcomings of existing technologies that rely on static indicators and lead to delayed deployment decisions. By independently outputting two prediction scores—decline probability and enhancement probability—it corresponds to the competition mechanism where existing cross-community channels are bypassed and the expansion mechanism formed by newly added cross-community channels, respectively, ensuring a clear semantic separation between the predictions of the two evolutionary modes. Furthermore, by using a graph neural network to approximate the triplet competition payoff function, it reduces computational complexity in large-scale networks. The overhead is low, which can support the real-time or near real-time inference needs of social platform advertising systems for massive nodes; and this application can run only based on user behavior interaction sequences, without the need for semantic understanding or annotation of content, resulting in low data acquisition costs and easy migration and deployment between different social platforms; on this basis, the prediction results are further mapped into two types of resource allocation interfaces: traffic distribution weights that can be directly called by the platform and budget allocation coefficients that can be directly called by the brand, forming a complete closed loop from network structure evolution perception to resource allocation decision-making; at the same time, this application is specifically designed for unidirectional directed social networks, and identifies information intermediary nodes by distinguishing the different propagation semantics of the in-degree direction and the out-degree direction, overcoming the defect of existing structural hole algorithms that cause the loss of directional information due to the symmetric processing of directed networks. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the workflow of the information mediation propagation force prediction method based on triplet competition dynamics proposed in this invention. Detailed Implementation

[0020] Reference Figure 1 The information mediation propagation power prediction method based on triplet competition dynamics proposed in this invention includes the following steps: S1. Obtain user social behavior data collected from social platforms and construct a directed social network. The nodes of the directed social network are users, and the directed edges point from content producers to content consumers. The directed edges are weighted and aggregated and their out-degrees are normalized to obtain normalized edge weights.

[0021] In this embodiment, user social behavior data includes behavior-associated users, behavior types, and behavior timestamps. Behavior-associated users include content consumers and content producers, and behavior types include following, forwarding, commenting, and liking. Building a directed social network specifically includes: With users as nodes and social behaviors between users as directed edges, the directed edges point from content producers to content consumers. When a content consumer follows a content producer, a directed edge points from the content producer to the content consumer; when a content consumer shares content produced by a content producer, a directed edge points from the content producer to the content consumer; when a content consumer comments on content produced by a content producer, a directed edge points from the content producer to the content consumer; when a content consumer likes content produced by a content producer, a directed edge points from the content producer to the content consumer. For behavior data related to following relationships, if the data collected by the social platform cannot obtain the specific timestamp of establishing the following relationship, then the time when the following relationship was first collected shall be used as the behavior timestamp of the directed edge.

[0022] Specifically, weighted aggregation is as follows: weight coefficients are set for each behavior such as following, forwarding, commenting, and liking. The sum of the products of the occurrence frequency of each behavior type on the same edge and the corresponding weight coefficient is used as the aggregation weight of the directed edge. Out-degree normalization is as follows: the aggregation weight of each out-edge of each node is divided by the sum of the aggregation weights of all out-edges of that node, so that the sum of the normalized edge weights of each out-edge of that node is 1.

[0023] Specifically, the input consists of user social behavior data collected from social media platforms, and each record can be in the following format: ,in, For users associated with social behaviors, They are content consumers. It is a content producer. For behavior type (follow / share / comment / like). A timestamp generated for the action.

[0024] For follow-related behavioral data, since social media platforms typically cannot obtain the specific timestamp of establishing a follow relationship, it is uniformly set as the time when the follow relationship was first collected. A directed social network is constructed based on user behavior data, with users as nodes and edges aligned with the information flow direction (i.e., directed edges point from content producers to content consumers), as detailed below: focus on The direction of the side is The semantics are information from Flow direction ,Right now Published content, See.

[0025] Forward The content, with the side direction as The semantic meaning is I saw The content and dissemination, information from Flow direction .

[0026] Comment The content, with the side direction as The semantic meaning is They are content consumers. They are content producers, similar to those who forward content.

[0027] Like The content, with the side direction as The semantic meaning is Consumption The content.

[0028] The act of liking itself does not generate content dissemination, but it reflects the intensity of content attention among users and is included in the edge weight calculation with a low weight; in implementation, it can be decided whether to include the liking edge based on the characteristics of the platform.

[0029] For cases where there are multiple edges between the same pair of nodes, different edge weights are assigned to different behavior types: in, express right The number of times it has been followed (usually 0 or 1); express Forward The number of times the content is displayed; for Comment The number of times the content is displayed; for Like The number of times the content is displayed; The weighting coefficients are used to determine the behaviors that are of interest. The weighting coefficient corresponding to the forwarding behavior; The weighting coefficient corresponding to the commenting behavior; The weighting coefficient corresponding to the "like" action.

[0030] Recommended weight is (Can be adjusted according to social media platforms). Normalize the outgoing edge weights of each node by dividing the original weights of all outgoing edges of that node by the sum of its outgoing edge weights. The normalized edge weights are denoted as... The sum of the outgoing edge weights of each node is 1, eliminating the impact of differences in node activity on subsequent calculations. Observation window length. Recommended to use the prediction window Four times the amount of historical behavior information is used to ensure sufficient historical behavior information for feature calculation and node embedding training.

[0031] It should be noted that directed social networks refer to social platform networks where the relationships between users are directional, including but not limited to social networks based on follow relationships (such as Weibo, Twitter), interactive networks based on retweet / comment / like behaviors, and multi-relationship directed networks composed of a mixture of the above types of directed edges. This application does not apply to social networks such as instant messaging and offline friends, which are mainly based on bidirectional symmetrical relationships.

[0032] S2. Based on normalized edge weights, perform directed community partitioning on the directed social network to obtain the community labels of each node.

[0033] In this embodiment, step S2 specifically includes: The directed social network is converted into an undirected projection, and the Louvain algorithm is run on the undirected projection to obtain the first community label of each node. Run the strongly connected component algorithm on a directed social network to obtain the strongly connected component labels of each node; For each community corresponding to the first community label, if the community contains multiple strongly connected components, the ratio of the average directed edge density between these strongly connected components to the average directed edge density within these strongly connected components is calculated. When this ratio is lower than a preset self-consistency threshold, the community is split into multiple sub-communities according to the boundaries of the strongly connected components, and the split sub-communities are used as the updated community labels of each node. For communities that do not meet the splitting conditions, their community labels remain unchanged.

[0034] In this embodiment, directed community partitioning further includes: When the number of nodes in a strongly connected component is less than the preset minimum community size threshold, the strongly connected component is not treated as an independent community. Instead, the directed edge density between the strongly connected component and other adjacent communities is calculated, and the component is merged into the adjacent community with the highest directed edge density. When the bidirectional directed edge density across strongly connected components between two different first communities exceeds a preset merging threshold, the two first communities are merged into the same community, and the merged community is used as the community label for each node within that community.

[0035] Specifically, the constructed directed social network needs to be divided into communities. This application adopts the following community division strategy for directed social networks: Firstly, it prioritizes using directed network community detection algorithms (such as DCSBM) to directly divide communities on the directed network. When the network size is large and the directed network algorithm is too slow, this application adopts a two-step strategy: firstly, it quickly obtains the initial community division on the undirected projection of the directed social network, i.e., obtains the first community label of each node; secondly, it uses the strongly connected component structure of the directed social network for correction, balancing efficiency and accuracy. The specific steps are as follows: First, the directed social network is converted into an undirected network. Then, the Louvain algorithm is used to divide the initial communities, and the first community label of each node is obtained. , i.e., node Which community does it belong to? To eliminate the randomness of the Louvain algorithm, in practice, a fixed random seed is used and the algorithm is run multiple times to obtain the result with the highest modularity (it is recommended to run it 5 times). On the original directed social network, a strongly connected component (SCC) partition is calculated using a strongly connected component algorithm (such as the Kosaraju algorithm or the Tarjan algorithm). For each Louvain community identified in the first step, its SCC composition is examined. If a community contains multiple SCCs, and the ratio of the mean directed edge density between these SCCs to the mean edge density within these SCCs is lower than a preset self-consistency threshold, then... (default This indicates that the community is inconsistent in its directed structure. Therefore, it is split into multiple sub-communities along the SCC boundary, and the first community label of each node is updated. The splitting operation must ensure that the number of nodes in each sub-community after splitting is no less than the minimum community size threshold. (default For sizes smaller than A single node in a Louvain community (including a single SCC) is not treated as an independent community, but is instead merged into the neighboring community with the highest directed edge density. Optionally, if the bidirectional cross-SCC directed edge density between two different Louvain communities exceeds a preset merging threshold... If so, they are merged into the same community to correct the over-splitting error caused by undirected projection. A preset merging threshold is set. The edge density can be adaptively set according to the overall network edge density. It is recommended to set it to twice the average density of directed edges between communities in the entire network to ensure that the merging operation is triggered only when the connections between communities are significantly dense.

[0036] S3. Identify information intermediary nodes based on community tags and enumerate their directed information intermediary triples. Each triple consists of two consecutive directed edges from the source node to the intermediary node and then to the audience node.

[0037] In this embodiment, the source node belongs to the first community, and the audience node belongs to the second community. The first community and the second community are different communities. Step S3 specifically includes: The directed edge density matrix between communities is calculated based on normalized edge weights. The elements in the directed edge density matrix between communities are the directed edge density from the first community to the second community, which is equal to the sum of all normalized edge weights from the first community to the second community divided by the product of the number of nodes in the first community and the number of nodes in the second community. Information intermediary nodes are identified based on the directed edge density matrix between communities and a preset community isolation threshold. For any node, if the node does not belong to the first community or the second community, and there are directed edges from nodes in the first community to the node and from the node to nodes in the second community, and the density of bidirectional directed edges between the first community and the second community is lower than the community isolation threshold, then the node is marked as an information intermediary node. Enumerate the directed information mediation triples for each information mediation node. For a node marked as an information mediation node, enumerate all triples that satisfy the following conditions: the source node belongs to the first community, the audience node belongs to the second community, there exists a directed edge from the source node to the mediation node and a directed edge from the mediation node to the audience node, and there is no directed edge from the source node directly to the audience node. All triples that satisfy the above conditions constitute the set of directed information mediation triples for that mediation node.

[0038] Specifically, in directed social networks, the specific method for identifying information intermediary nodes is as follows: First, define the directed edge density matrix between communities. , From the first community Pointing to the second community The directed edge density, i.e., all edges from the first community Pointing to the second community The sum of the normalized edge weights, divided by the product of the number of nodes in the two communities. (That is, the maximum number of edges that can theoretically exist between two communities).

[0039] Subsequently, choose any two different first communities. Second Community If node At the same time, it does not belong to the first community. Second Community And simultaneously satisfy the connection conditions in the following two directions: (1) Node Receive from the first community The inner edge of the node (i.e., the edge that exists) Make the directed edge Existence indicates Consumption (Content) (2) Node To the second community Internal nodes emit outgoing edges (i.e., there exist outgoing edges) Make the directed edge Existence indicates Consumption (content); and and Directed edge density between and (in The community isolation threshold is set to 0.0002 by default. This threshold can be adjusted based on the platform's network sparsity. It is recommended to use the average directed edge density among communities across the entire network. times, default To ensure that isolation decisions have consistent relative semantics across networks of different sizes, at this point, the node... The node is marked as an information intermediary node, and this node is enumerated. All directed information intermediary triples ,in There are directed edges and And there are no directly directed edges. All triples that satisfy the above conditions constitute a node. The set of triples .

[0040] It is important to note that triplets It describes nodes Connecting the First Community (by (as representative) and the second community (by Structural information channels (represented by [example]): Information flows from... Inflow of the community ,through Flow after forwarding / posting The community in which the content is shared, rather than referring to a specific link in the forwarding chain of the same content.

[0041] The directed information mediator node identification conditions in this application are conceptually consistent with the structural hole theory, but they are formalized again for the information propagation characteristics of directed social networks and directly serve the subsequent triplet competition dynamics modeling. Compared with existing structural hole algorithms, they have stronger directed network adaptability and operability.

[0042] It should be noted that defining information intermediary nodes... The Directed Intermediary Score (DIS) serves as a comprehensive quantitative indicator of its propagation power. ; in, For the first community Pointing to the second community Directed edge density; edge weight All values ​​are weighted values ​​normalized by the out-degree of the nodes, and their range is [range missing]. . What is being measured is the node The average quality of each cross-community messaging channel: strongly dependent on both ends of the channel. (High marginal power), and the communities at both ends are highly isolated. (Small), this passage is the most valuable.

[0043] Using averaging instead of summation reduces the direct amplification of the indicator by the difference in the number of triples controlled by the nodes, making the indicator more focused on the structural quality of a single cross-community information channel.

[0044] S4. For each triple, extract the features used to quantify the competitive motivation of the audience node to bypass the intermediary node and directly connect to the source node, and obtain the embedding vector of each node through a graph neural network.

[0045] In this embodiment, the features include information complementarity, indirect perception, and connection strength; for each triple, features are extracted to quantify the competitive motivation of the audience node to bypass the intermediary node and directly connect to the source node, specifically including: Extract the information complementarity between the community to which the source node belongs and the community to which the audience node belongs. The information complementarity is the quotient obtained by subtracting the size of the intersection of the sets of objects of interest of the first community and the second community from the size of the union. Extract the structural dependence degree of the intermediary node on the content of the first community. The structural dependence degree is the quotient obtained by dividing the sum of the normalized edge weights from the first community in the incoming edges of the intermediary node by the sum of the normalized edge weights of all incoming edges of the intermediary node. Extract the coverage of the intermediary node to the second community. The coverage is the quotient obtained by dividing the number of users in the second community who pay attention to the intermediary node by the total number of people in the second community. The product of the degree of structural dependence and the degree of coverage is used as the indirect perception. Extract the connection strength between the intermediary node and the source node and the audience node respectively. The connection strength is the average of the normalized edge weight from the source node to the intermediary node and the normalized edge weight from the intermediary node to the audience node. Information complementarity, indirect perception, and connection strength are used as features to quantify competitive motivation.

[0046] In this embodiment, the embedding vector of each node is obtained through a graph neural network, specifically as follows: For each node, an initial embedding vector is constructed. The initial embedding vector shall contain at least the node’s in-degree, out-degree, one-hot encoding of its community, number of active days within the observation window, and number of published messages. In each layer of the graph neural network, messages are aggregated along the in-degree neighbor direction and the out-degree neighbor direction, respectively. The in-degree neighbor is other nodes that are followed by this node, and the out-degree neighbor is other nodes that follow this node. The message in each direction is calculated by the message function based on the dynamic edge features of the connection between the current embedding vector of the neighbor node and the node. The in-degree direction aggregation result, the out-degree direction aggregation result, and the current embedding vector of the node are concatenated and then transformed by the Message Passing Perceptron (MLP) to obtain the embedding vector of the next layer of the node.

[0047] Specifically, the dynamic edge features are calculated based on normalized edge weights and behavior timestamps in an exponential decay manner. The exponential decay is an exponential decay function with a base of the natural constant and an exponent of a negative learnable decay rate multiplied by the difference between the current time and the most recent behavior timestamp.

[0048] Specifically, at the node Information intermediary triple In the middle, information along Path flow: The content of the community is accessed through transit arrival The community you belong to. If... With nodes in triples Directly establish connections (i.e., form directed edges) Information from Direct flow to If so, the information flow path is "short-circuited". The value of this information intermediary is thus reduced. This node-level closure event is used to characterize... Bypass direct access Trends in the information sources within your community.

[0049] This application will Will the community be with The establishment of direct information channels within a community is modeled as a directed competitive dynamic, with the active party being the node. The decision is whether to match the nodes in the triplet. Establish direct connections (add corresponding directed edges in the directed network) ,Right now focus on Information from Direct flow to ),thereby Bypass Get directly The content of the community you belong to.

[0050] Specifically, we define a directed competitive payoff function. ,measure How much motivation is there to directly pay attention? The node of the community, instead of through... To acquire content. The stronger the motivation, The more easily an information intermediary position is lost, the more likely it is to decline.

[0051] Given information mediator node , Active direct connection Net revenue of the community node: in, For information complementarity Learnable weight parameters; For indirect perception Learnable weight parameters; For connection strength The learnable weight parameters. The above three features All were constructed as The normalized indices within the intervals have comparable numerical ranges; however, since the three originate from different statistical mechanisms, their empirical distributions may differ. Therefore, learnable weights are used to differentiate them. Adaptively adjust various contributions.

[0052] Information complementarity, a measure The content of the community The degree of complementarity between two communities is important; the less overlap the accounts followed by two communities, the greater the difference in their content ecosystems. From The stronger the motivation for direct engagement, the more new content a community gains access to that it wouldn't normally have. Conversely, if two communities follow the same groups of accounts, There's no need to specifically go to the link. The calculation is as follows: ; in, Represents a node The community Represents a node The community For nodes The union of the sets of objects followed by all users within the community; It is a community-level feature; all triples within the same community share this value, which can be pre-computed and cached.

[0053] Indirect perceptuality, used for approximate representation pass transit The degree of indirect perception of community content. The higher the degree of indirect perception, right The more familiar the community content, the stronger the motivation for direct engagement. Indirect perception is estimated from two dimensions: To what extent does the content originate from Community (content source) exist The proportion of users covered in the community (reach) and the breadth of reach are jointly determined. pass Indirect perception The level of community content.

[0054]

[0055] in, express From the edge The sum of the normalized edge weights of the community accounts for The proportion of the sum of all normalized weights of incoming edges measures right The degree of structural dependence of community content; when the denominator is 0, let , indicating the current window There are no valid incoming edges that can be used to characterize the content source structure; Indicates the source node The community to which it belongs; community Any node in; Represents a node The neighbor set, that is, to The set of nodes from which the output information originates; Represents a set Any incoming neighbor node in the network; and These represent the edge weights of the corresponding directed edges after out-degree normalization. Represents a node Received from the community The sum of the normalized weights of the borders. Represents a node It receives the sum of normalized incoming edge weights from all incoming neighbors; the ratio of these two weights is used to measure the node's performance. For the community The degree of structural dependence on content sources.

[0056]

[0057] Right now Followed in the community users The proportion of the total number of members in the community, a measure right Community reach; statistical community There exists a middle node Pointing to the community The number of users on the directed edges of the internal node, divided by the community. The total number of users.

[0058] Connection strength, a measure As an information intermediary The higher the value provided, the more significant the difference. The more difficult it is for communities to be replaced by direct connections Its information intermediary value.

[0059]

[0060] Right now The average normalized edge weight connected to the two endpoints reflects The strength of the connection and the closeness of the relationship on this channel. As the source node, As an information intermediary node, For the audience node; Indicates the source node Pointing to the intermediary node Normalized border weights Indicates a middleman node Pointing to the audience node The normalized edge weights are calculated using the average of the normalized edge weights of two connecting edges, which is used to measure the weights of intermediate nodes. The connection strength and mediation value within this information transmission channel. A higher value indicates bypassing [the intermediary role]. The higher the cost of substitution.

[0061] S5. Based on the features and embedding vectors of each triplet, the triplet is processed by the first multilayer perceptron and the second multilayer perceptron respectively to determine the closure probability of each triplet within the prediction window.

[0062] In this embodiment, step S5 specifically includes: The embedding vectors of the audience node, the source node, the intermediary node, the information complementarity, the indirect perception, and the connection strength are concatenated and input into the first multilayer perceptron to obtain the approximate competitive payoff value of each triple. The approximate competitive payoff value is used to characterize the payoff of the audience node to bypass the intermediary node and directly connect to the source node. The embedding vector of the intermediary node is input into the second multilayer perceptron and mapped to obtain the adaptive closure threshold of the intermediary node. The adaptive closure threshold is used to characterize the different benefit thresholds that different intermediary nodes need to reach to transform competitive benefits into actual closure behavior. The difference between the approximate competitive payoff and the adaptive closure threshold is mapped using the sigmoid function to obtain the closure probability of the triple within the prediction window.

[0063] Specifically, triplet Closure occurs within the prediction window (i.e.) direct connection Bypass The probability of ) is:

[0064] in, For the sigmoid function, For nodes The adaptive closure threshold is determined by the node. The node embedding vectors are predicted by a second multilayer perceptron. The ontology of the revenue function is theoretically defined as follows: However, in the process of engineering implementation and training inference, its approximate estimate is used. Participating in the closure probability calculation. Different information intermediary nodes. The social network environments they operate in are different. Some nodes... It connects highly active communities, some nodes What is being connected is a relatively silent community. The same level of benefit has different probabilities of translating into actual closed-loop behavior in different environments. Its function is to provide each node with Set a personalized threshold; closure can only actually occur when the approximate return exceeds this threshold.

[0065] For large-scale networks, if each triple is directly computed in analytical form... The overall computational cost is relatively high. Therefore, graph neural networks do not replace these three explicit features, but rather utilize node embedding vectors. Implicitly encode the global topological information of each node beyond the local triplet, and combine it with the above. , Three explicit features, processed by the first multilayer perceptron. Approximate calculations are performed to reduce end-to-end computing costs.

[0066] It should be noted that, The three explicit features still need to be calculated, but among them... Community-level features can be pre-computed and cached. , and These are node-level features that can be computed incidentally during message passing in a Graph Neural Network (GNN). The core contribution of GNNs lies in their use of node embeddings to... The approximate calculation supplements global structural information beyond the local triplet, rather than redefining the payoff function itself.

[0067] Initial embedding vector of nodes Define as a node The result of concatenating the basic feature vectors, i.e. .in, and Representing nodes respectively In-degree and out-degree at the end of the observation window; Represents a node The one-hot encoded vector of the community to which it belongs, if the total number of communities is ,but And only in The corresponding dimension value for the community is 1; Represents a node The number of active days with content posting, forwarding, or commenting within the observation window; Represents a node The number of information items published or forwarded within the observation window; all the above scalar features are uniformly normalized before being input into the graph neural network.

[0068] In directed social networks, message passing requires distinguishing between in-degree and out-degree neighbors to capture the information receiving and output capabilities of nodes respectively (Note: According to the edge direction convention in this application—directed edges point from content producers to content consumers, and in-degree neighbors...). yes Followed users The source of the output content, out-degree neighbor. It is a concern Users from (Node receiving information)

[0069] in For nodes Self-first The embedding vector of the wheel, For nodes in-degree neighbor set ( (The users you follow, i.e., the content sources) For nodes Out-degree neighbor set (focus) (the users, i.e., the content audience). and The dynamic edge characteristics at the current moment are represented by the normalized edge weights combined with the time decay term. As a message function, combining neighbor embeddings with edge features can yield... (Weighted) or (Transformation after splicing) and other forms, the specific implementation can be selected according to the platform requirements. and These represent the aggregation operations on incoming neighbor messages and outgoing neighbor messages, respectively. The preferred method is mean aggregation, which means averaging the message vectors in the corresponding message sets element by element to reduce the direct impact of differences in the number of neighbors on the node representation scale. This represents a vector concatenation operation. Bidirectional message aggregation enables... Simultaneous encoding Attention behavior (through) ) and the relationship of being followed (through It is recommended that GNNs have 2 to 3 layers to strike a balance between capturing sufficient neighborhood information and avoiding oversmoothing.

[0070] The temporality of directed edges is encoded with exponentially decaying weights, with recently formed attention / interaction relationships contributing more to node embedding: ; in, To normalize edge weights, for and The timestamp of the most recent occurrence among all directed actions (follow / share / comment / like). The decay rate (a learnable parameter).

[0071] in, Represents the return function This is an approximate estimate, not a redefinition of the payoff function. The above process will calculate... Three explicit interpretable features are directly input. Then embedded with GNN structure The splicing provides complementary information; explicit features preserve the interpretability of the payoff function, while node embeddings supplement global topological information beyond the local context of the triples.

[0072] The adaptive threshold prediction formula is: ; in, For a threshold prediction network that outputs a single scalar, with nodes Embedded vector As input, the output is the personalized closure threshold of the node; its parameters are learned through supervised training of triplet closure events within the historical window.

[0073] In engineering implementation, the main computational overhead of GNN message passing is concentrated in edge-level neighborhood information aggregation; among which, community-level features... Pre-compute and cache node-level features , and It can be combined with synchronous statistics during message transmission, and the revenue is similar to that of a network. Batch forward inference can be used for execution. Therefore, this application is more advantageous in controlling the overall computational overhead in large-scale network scenarios and supports real-time or near real-time prediction.

[0074] S6. Aggregate the closure probabilities of all triples of each intermediary node to obtain the propagation decay probability of each intermediary node.

[0075] In this embodiment, step S6 specifically includes: For each second community, the closure probability of all triples in the second community is compared, and the maximum value is taken as the bypass probability of the second community. The bypass probability is used to characterize the trend strength of the second community as a whole bypassing the intermediary node and directly accessing the community information source to which the source node belongs. Calculate the complement of the bypass probability of each second community, multiply the complements of all bypass probabilities of the second community together, and then subtract the result of the multiplication by one to obtain the propagation power decay probability of the intermediary node. The propagation power decay probability is used to characterize the overall probability that the intermediary node has at least one cross-community information channel bypassed within the prediction window, thus causing its propagation power to decay.

[0076] Specifically, after obtaining the node embedding, this application further predicts the information mediator node in the prediction window. There are two types of evolutionary outcomes: one is the probability of decreased propagation power due to the bypassing of the information intermediary node's position. Secondly, the probability of increased propagation power of the information intermediary due to the formation of a new information intermediary triplet. In this process, the intermediary role of information is bypassed, specifically by the direct replacement of its original cross-community communication channels, thereby weakening its bridging role in the information dissemination chain and resulting in a decline in its dissemination power. The predefined prediction window length represents the time span for predicting the evolution state of node propagation force.

[0077] Specifically, the propagation decay model addresses the risk that existing information intermediary triples will be bypassed within a future window. This application constructs a neural probabilistic hybrid model, first outputting triple-level intermediate quantities from a neural network, then aggregating them layer by layer through a deterministic probability mapping link to finally obtain the overall decay probability at the node level.

[0078] The specific forward computation process is as follows: For any triple First, based on the output node embeddings and the defined explicit features, an approximate estimate of the profit function is calculated:

[0079] In its theoretical definition, the profit function ontology is still denoted as... In the process of engineering implementation and training inference, the following methods are adopted: Approximate estimate of the output It participates in subsequent closure probability calculations to reduce computational overhead on large-scale networks while preserving the interpretability of explicit features.

[0080] Based on node-embedded computing nodes Personalized closing threshold: ; The triplet-level closure probability is obtained from the approximate return and the personalized threshold: ; in, This indicates the audience node within the prediction window. Bypassing intermediary nodes And with the source node The probability of establishing a direct connection; For the competitive payoff function Approximate estimate; The sigmoid function is used to map input values ​​to... Interval.

[0081] Aggregating the triplet-level closure probabilities by audience community yields the community-level bypass probabilities: ; That is, take the second community The maximum probability of closure of all triples within the community is taken as the overall bypass probability of the community. Probability estimation. Using... The rationale for this action is that within the same community, as long as there is a sufficiently strong direct connection motivation, it may trigger the transfer of information channels at the community level.

[0082] At the node level, the bypass probabilities of different audience community channels are further aggregated to obtain the node. Overall transmission capacity decline probability:

[0083] The formula calculates the nodes using a probability product method. In the prediction window The overall probability of decline within, of which Indicates at node The set of triples As the audience node The set of all audience communities appearing in the given community assumes that bypass decisions between different audience community channels are independent. Specifically, This indicates the joint probability that no bypass event occurred across all audience community channels. Subtracting this joint probability yields the result that at least one channel is bypassed, leading to... The overall probability of diffusion decline. This modeling approach defines node-level decline as the result of multi-channel competition, which better reflects the isolated characteristics of user behavior in different communities within a social network than the independence assumption at the single triple level.

[0084] S7. Obtain the expansion behavior characteristics of each intermediary node, and combine them with the embedding vector of each intermediary node to predict the probability of enhanced propagation power of each intermediary node. The expansion behavior characteristics are used to quantify the trend of each intermediary node establishing new cross-community connections within the observation window.

[0085] In this embodiment, step S7 specifically includes: Obtain the expansion behavior feature vector of each intermediary node. The expansion behavior feature vector includes the intensity of expansion to new audiences, the breadth of community coverage, the expansion speed of acquiring new content sources, and historical expansion records. The strength of expansion to new audiences is the quotient of the number of new outgoing edges of the intermediary node within the observation window divided by the length of the observation window; the breadth of community coverage is the percentage of all outgoing edges of the intermediary node pointing to new community nodes that have not been connected in the past; the expansion speed of acquiring new content sources is the quotient of the number of incoming edges of the intermediary node from new communities that have not been contacted in the past within the observation window divided by the length of the observation window; the historical expansion record is a binary indicator of whether the current set of directed information intermediary triples has added at least one directed information intermediary triple compared to the previous observation window; The embedding vectors of each intermediary node are concatenated with the feature vectors of the expansion behavior. The concatenated vectors are then mapped through a third multilayer perceptron and activated by the sigmoid function to obtain the propagation enhancement probability of each intermediary node within the prediction window.

[0086] Specifically, within the framework of triplet competition dynamics, enhanced propagation power can be understood as nodes... The tendency to add new information intermediary triples, i.e., nodes Discovered by new community users and forming new cross-community connections, enabling new... Triads form within the window period. (And) (Based on modeling the mechanism of existing triplet destruction) different, The modeling focuses on the potential emergence of triples that do not yet exist. Since newly added information intermediary triples do not explicitly exist at the prediction time, their potential connection objects and community combinations are unknown, making direct modeling based on existing competitive relationships difficult, unlike existing triples. Therefore, this application employs a combination of node embedding and behavioral statistical features for prediction, with the forward computation as follows: ; in, The node embedding vector obtained from GNN message passing is used to encode the nodes. Local network structure information; The extended feature vector is calculated directly from behavioral data statistics; the two are concatenated and then input. Make predictions.

[0087] ; in, for In the observation window The number of newly added outgoing edges divided by the window length Corresponding node The rate at which a node gains attention or connection from new content consumers is used to measure its performance. The intensity of expansion to new audiences; For nodes The percentage of outgoing edges pointing to new community nodes that have never been connected in the past is used to measure the number of nodes. The breadth of community coverage (existing penetration rate) in social networks. For nodes Within the observation window, the number of incoming edges from historically uncontacted new communities divided by the window length. Used to measure nodes The expansion speed of acquiring new content sources (incremental new user acquisition rate); For the present Compared to the previous observation window (strictly limited to) , and prediction window Whether at least one new information mediator triple (0 / 1) has been added (no overlap), encoding The recent expansion momentum. and Based on different mechanisms: the former is based on the competitive dynamics of triples, while the latter is based on behavioral statistics and network structure characteristics. Both can be high or low at the same time.

[0088] S8. Output the probability of propagation decay and the probability of propagation enhancement as independent bi-objective prediction results.

[0089] This application employs a multi-task joint learning paradigm, using a unified loss function to perform end-to-end parameter optimization on each neural network module. Module (Message Passing Perceptron) First multilayer perceptron Second multilayer perceptron Third multilayer perceptron Functionally decoupled, each task undertakes structural representation learning, benefit approximation, threshold decision-making, and propagation enhancement prediction. Each subtask is driven by an independent supervision signal, maintaining decoupling and complementarity in parameter optimization.

[0090] The training set is constructed based on historical time-series slices of a directed social network. For each historical moment... Construct an observation window and the corresponding prediction window And targeting information intermediary nodes Construct the following supervision labels: Tag construction: Positive samples ( ): In the prediction window Inside, node The set of triples The triplet is closed (i.e., a direct directed edge appears). ); negative samples ( No triplet closure occurs within the window; Because of the imbalance between positive and negative samples (the number of triplets that actually close is far less than the number that do not), negative samples are downsampled to make the ratio of positive to negative samples 1:4 (i.e., retaining 4 times the number of negative samples as positive samples). Tag construction: Positive samples ( ): In the prediction window Inside, At least one new directed information intermediary triple has been added; negative samples ( No new information intermediary triples have been added to the window; Negative samples are downsampled to make the ratio of positive to negative samples 1:3; To reduce the impact of community segmentation drift Due to the influence of labels, this application adopts a fixed strategy for dividing the community into baseline periods: using the starting time of the observation window for each time slice in the training set. Based on this, fix the community segmentation results at this moment. In the judgment prediction window When adding a new triplet, always use The community labels are used instead of being redefined. The baseline community partitioning is only updated when moving to the next non-overlapping training slice.

[0091] Triple-level auxiliary label construction (for auxiliary loss): For nodes Each triple If in the prediction window Directed edges appear inside. Then the triplet label ,otherwise ; Joint training of three objectives: ; ; ; ; in, and These are the main losses for the node-level propagation decay prediction task and the propagation enhancement prediction task, respectively. Auxiliary loss for triplet-level closure prediction; for Task weight (default) ), Auxiliary loss weights for triples (default) Due to node-level decay probability The closure probability is obtained by community aggregation from triplet-level probabilities. Relying solely on node-level supervision may lead to the dilution of supervision of the intermediate closure mechanism during backpropagation. Therefore, further... Direct constraints are applied to the triplet-level closure probabilities to enhance the learnability of local closure patterns and improve overall training stability. In the formula, Represents the set of training node samples. This represents the training triplet sample set. All three losses are calculated using the mean of the samples to balance the magnitudes across different tasks. and Using independent prediction heads but sharing the node embeddings of the underlying GNN, efficient parameter utilization is achieved. Message-passing MLPs do not require a pre-defined independent loss function; their parameters are obtained through the above... Backpropagation obtains gradients from downstream tasks for updates, ensuring optimal alignment between node embedding representations and propagation evolution prediction targets.

[0092] It should be noted that the training set is constructed using historical time-series slices, with each slice corresponding to an observation window. and prediction window Adjacent slices are separated by a step size. Slide (Recommended) (i.e., non-overlapping), the training set and the test set are strictly divided in chronological order to ensure that the time window of the test set is later than all time windows of the training set, thus avoiding time leakage.

[0093] During forward computation, message passing is first performed based on the directed dynamic graph within the observation window to obtain node embeddings; then... and Calculate the probability of triplet closure Furthermore, it aggregates data at both the community and node levels to obtain the node-level decay probability. At the same time by Output node-level enhancement probability Calculate joint loss .

[0094] Backpropagation updates all learnable parameters (including message passing). Weight, Edge weight behavior type weight To fix the hyperparameters, they do not participate in gradient updates, and the iteration is repeated until convergence.

[0095] After training, during online inference, the above forward computation process is repeated using the data in the current window, outputting only the node level. and There is no need to calculate the loss function or update the model parameters; the results can be used by the subsequent resource allocation module.

[0096] After completing model training and obtaining node-level prediction results, this application further... and Three metrics are provided, offering two resource allocation interfaces for both platform and brand clients, directly outputting allocation coefficients that can be called by the system. The candidate node set is preferably the identified set of information intermediary nodes; in a platform-side scenario, a creator candidate set can be further selected from this set, while in a brand-side scenario, a KOL candidate set can be further selected.

[0097] Specifically, Interface 1 – Platform: Adjustment of traffic distribution weight; Perform log normalization on the DIS values ​​of the candidate node set: ; make ,and The dimensions are consistent. Among them, This represents any node in the current set of candidate nodes participating in traffic distribution or budget allocation. Therefore, the denominator represents the maximum DIS value in the current candidate set after logarithmic compression.

[0098] ; in, For the current time node The normalized DIS value is used to compress the huge gap between head nodes and ordinary nodes. and The coefficients for the decay penalty and the enhancement reward (both are 0.5 by default, which are fixed hyperparameters). It can be directly used as an additional node weight when distributing traffic in an advertising system. The output range is (when When, the maximum value is (Obtained at the time), it can be directly used as the relative weight coefficient of the advertising system, and it is comparable across nodes.

[0099] Interface 2 – Brand: KOL advertising budget allocation coefficient.

[0100] ; in, To perform minimum-maximum normalization on the candidate KOL set, the output is... The allocation coefficient of the range; For nodes The original directed mediation strength. The difference between Interface 2 and Interface 1 is: (1) Interface 2 directly uses the original Instead of log normalization Because it will eventually be normalized to a unified dimension through the minimum-maximum value; (2) Interface 2 is not used Attenuation coefficient, for and The response is more sensitive – brand advertising decisions are typically made quarterly, requiring a more timely response to changes in communication trends, thus employing a more aggressive decline penalty (when...). hour (Reset to zero, it is recommended to completely stop the distribution of this node) and enhance rewards. Among them, This represents the relative budget coefficient within the candidate KOL set. Brands can use this coefficient based on each candidate node. The proportion of each factor is used to normalize the allocation of the total budget; for example, for the candidate set. Middle node The suggested budget can be expressed as: ; in, This refers to the total budget that the brand plans to allocate to the current pool of potential KOLs. Represents a node The relative budget allocation coefficient; This represents the set of candidate nodes to participate in budget allocation; Represents the candidate set Any candidate node in; This represents the sum of the relative budget coefficients of all candidate nodes.

[0101] This application further provides a basis and The explanatory budget decision rules for combination relationships are used to assist in manual review and operational strategy formulation, but do not replace the calculation results of the above formulas.

[0102] Budget allocation decision-making rules: High (>0.6) Low (<0.3) Reduce the budget and find alternative nodes; Low (<0.3) High (>0.6) Increase your holdings in advance to lock in undervalued assets; high, high Maintain observation for testing new audiences; Low, Low Maintain the current budget and keep the status quo stable.

[0103] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting the propagation power of information mediators based on the competitive dynamics of triplet groups, characterized in that, Includes the following steps: We acquire user social behavior data collected from social platforms and construct a directed social network. The nodes of the directed social network are users, and the directed edges point from content producers to content consumers. We perform weighted aggregation and out-degree normalization on the directed edges to obtain normalized edge weights. Directed community partitioning is performed on the directed social network based on normalized edge weights to obtain the community labels of each node; Identify information intermediary nodes based on community tags and enumerate their directed information intermediary triples. Each triple consists of two consecutive directed edges from the source node to the intermediary node and then to the audience node. For each triple, features are extracted to quantify the competitive motivation of the audience node to bypass the intermediary node and directly connect to the source node, and the embedding vector of each node is obtained through a graph neural network. Based on the features and embedding vectors of each triplet, the triplet is processed by the first multilayer perceptron and the second multilayer perceptron respectively to determine the closure probability of each triplet within the prediction window; Aggregate the closure probabilities of all triples of each intermediary node to obtain the propagation decay probability of each intermediary node; The expansion behavior features of each intermediary node are obtained, and the propagation power enhancement probability of each intermediary node is predicted by combining the embedding vector of each intermediary node. The expansion behavior features are used to quantify the trend of each intermediary node establishing new cross-community connections within the observation window. The probability of transmission decay and the probability of transmission enhancement are output independently as bi-objective prediction results.

2. The method for predicting the propagation power of information mediators based on triplet competition dynamics according to claim 1, characterized in that, The user social behavior data includes users associated with the behavior, behavior types, and behavior timestamps. Users associated with the behavior include content consumers and content producers, and behavior types include following, forwarding, commenting, and liking. The construction of the directed social network specifically includes: With users as nodes and social behaviors between users as directed edges, the directed edges point from content producers to content consumers. When a content consumer follows a content producer, a directed edge points from the content producer to the content consumer; when a content consumer shares content produced by a content producer, a directed edge points from the content producer to the content consumer; when a content consumer comments on content produced by a content producer, a directed edge points from the content producer to the content consumer; when a content consumer likes content produced by a content producer, a directed edge points from the content producer to the content consumer. For behavior data related to following relationships, if the data collected by the social platform cannot obtain the specific timestamp of establishing the following relationship, then the time when the following relationship was first collected shall be used as the behavior timestamp of the directed edge.

3. The method for predicting the propagation power of information mediators based on triplet competition dynamics according to claim 2, characterized in that, The weighted aggregation specifically involves setting weight coefficients for each of the following, forwarding, commenting, and liking behaviors, and using the sum of the products of the occurrence frequency of each behavior type and its corresponding weight coefficient on the same edge as the aggregation weight of that directed edge; the out-degree normalization specifically involves dividing the aggregation weight of each out-edge of each node by the sum of the aggregation weights of all out-edges of that node, so that the sum of the normalized edge weights of each out-edge of that node is 1.

4. The method for predicting the propagation power of information mediators based on triplet competition dynamics according to claim 1, characterized in that, The method of performing directed community partitioning on a directed social network based on normalized edge weights to obtain community labels for each node specifically includes: The directed social network is converted into an undirected projection, and the Louvain algorithm is run on the undirected projection to obtain the first community label of each node. Run the strongly connected component algorithm on a directed social network to obtain the strongly connected component labels of each node; For each community corresponding to the first community label, if the community contains multiple strongly connected components, the ratio of the average directed edge density between these strongly connected components to the average directed edge density within these strongly connected components is calculated. When this ratio is lower than a preset self-consistency threshold, the community is split into multiple sub-communities according to the boundaries of the strongly connected components, and the split sub-communities are used as the updated community labels of each node. For communities that do not meet the splitting conditions, their community labels remain unchanged.

5. The method for predicting the propagation power of information mediators based on triplet competition dynamics according to claim 4, characterized in that, The directed community segmentation also includes: When the number of nodes in a strongly connected component is less than the preset minimum community size threshold, the strongly connected component is not treated as an independent community. Instead, the directed edge density between the strongly connected component and other adjacent communities is calculated, and the component is merged into the adjacent community with the highest directed edge density. When the bidirectional directed edge density across strongly connected components between two different first communities exceeds a preset merging threshold, the two first communities are merged into the same community, and the merged community is used as the community label for each node within that community.

6. The method for predicting the propagation power of information mediators based on triplet competition dynamics according to claim 1, characterized in that, The source node belongs to the first community, and the audience node belongs to the second community. The first community and the second community are different communities. The step of identifying information intermediary nodes based on community tags and enumerating their directed information intermediary triples specifically includes: The directed edge density matrix between communities is calculated based on normalized edge weights. The elements in the directed edge density matrix between communities are the directed edge density from the first community to the second community, which is equal to the sum of all normalized edge weights from the first community to the second community divided by the product of the number of nodes in the first community and the number of nodes in the second community. Information intermediary nodes are identified based on the directed edge density matrix between communities and a preset community isolation threshold. For any node, if the node does not belong to the first community or the second community, and there are directed edges from nodes in the first community to the node and from the node to nodes in the second community, and the density of bidirectional directed edges between the first community and the second community is lower than the community isolation threshold, then the node is marked as an information intermediary node. Enumerate the directed information mediation triples for each information mediation node. For a node marked as an information mediation node, enumerate all triples that satisfy the following conditions: the source node belongs to the first community, the audience node belongs to the second community, there exists a directed edge from the source node to the mediation node and a directed edge from the mediation node to the audience node, and there is no directed edge from the source node directly to the audience node. All triples that satisfy the above conditions constitute the set of directed information mediation triples for that mediation node.

7. The method for predicting the propagation power of information mediators based on triplet competition dynamics according to claim 6, characterized in that, The features include information complementarity, indirect perception, and connection strength; For each triple, features are extracted to quantify the competitive motivation of the audience node to bypass the intermediary node and directly connect to the source node, specifically including: Extract the information complementarity between the community to which the source node belongs and the community to which the audience node belongs. The information complementarity is the quotient obtained by subtracting the size of the intersection of the sets of objects of interest of the first community and the second community from the size of the union. Extract the structural dependence degree of the intermediary node on the content of the first community. The structural dependence degree is the quotient obtained by dividing the sum of the normalized edge weights from the first community in the incoming edges of the intermediary node by the sum of the normalized edge weights of all incoming edges of the intermediary node. Extract the coverage of the intermediary node to the second community. The coverage is the quotient obtained by dividing the number of users in the second community who pay attention to the intermediary node by the total number of people in the second community. The product of the degree of structural dependence and the degree of coverage is used as the indirect perception. Extract the connection strength between the intermediary node and the source node and the audience node respectively. The connection strength is the average of the normalized edge weight from the source node to the intermediary node and the normalized edge weight from the intermediary node to the audience node. Information complementarity, indirect perception, and connection strength are used as features to quantify competitive motivation.

8. The method for predicting the propagation power of information mediators based on triplet competition dynamics according to claim 1, characterized in that, The embedding vectors of each node are obtained through a graph neural network, specifically as follows: For each node, an initial embedding vector is constructed. The initial embedding vector shall contain at least the node’s in-degree, out-degree, one-hot encoding of its community, number of active days within the observation window, and number of published messages. In each layer of the graph neural network, messages are aggregated along the in-degree neighbor direction and the out-degree neighbor direction, respectively. The in-degree neighbor is other nodes that are followed by this node, and the out-degree neighbor is other nodes that follow this node. The message in each direction is calculated by the message function based on the dynamic edge features of the connection between the current embedding vector of the neighbor node and the node. The in-degree direction aggregation result, the out-degree direction aggregation result, and the current embedding vector of the node are concatenated and then transformed by the Message Passing Perceptron (MLP) to obtain the embedding vector of the next layer of the node. The dynamic edge features are calculated based on normalized edge weights and behavior timestamps in an exponential decay manner. The exponential decay is an exponential decay function with a base of the natural constant and an exponent of a negative learnable decay rate multiplied by the difference between the current time and the most recent behavior timestamp.

9. The method for predicting the propagation power of information mediators based on triplet competition dynamics according to claim 1, characterized in that, The process of determining the closure probability of each triplet within the prediction window, based on the features and embedding vectors of each triplet and processed by a first multilayer perceptron and a second multilayer perceptron respectively, specifically includes: The embedding vectors of the audience node, the source node, the intermediary node, the information complementarity, the indirect perception, and the connection strength are concatenated and input into the first multilayer perceptron to obtain the approximate competitive payoff value of each triple. The approximate competitive payoff value is used to characterize the payoff of the audience node to bypass the intermediary node and directly connect to the source node. The embedding vector of the intermediary node is input into the second multilayer perceptron and mapped to obtain the adaptive closure threshold of the intermediary node. The adaptive closure threshold is used to characterize the different benefit thresholds that different intermediary nodes need to reach to transform competitive benefits into actual closure behavior. The difference between the approximate competitive payoff and the adaptive closure threshold is mapped using the sigmoid function to obtain the closure probability of the triple within the prediction window. The aggregation of the closure probabilities of all triples of each intermediary node to obtain the propagation decay probability of each intermediary node specifically includes: For each second community, the closure probability of all triples in the second community is compared, and the maximum value is taken as the bypass probability of the second community. The bypass probability is used to characterize the trend strength of the second community as a whole bypassing the intermediary node and directly accessing the community information source to which the source node belongs. Calculate the complement of the bypass probability of each second community, multiply the complements of all bypass probabilities of the second community together, and then subtract the result of the multiplication by one to obtain the propagation power decay probability of the intermediary node. The propagation power decay probability is used to characterize the overall probability that the intermediary node has at least one cross-community information channel bypassed within the prediction window, thus causing its propagation power to decay.

10. The method for predicting the propagation power of information mediators based on triplet competition dynamics according to claim 1, characterized in that, The process involves acquiring the expansion behavior features of each intermediary node, combining them with the embedding vectors of each intermediary node to predict the probability of increased propagation power. These expansion behavior features are used to quantify the trend of each intermediary node establishing new cross-community connections within the observation window, specifically including: Obtain the expansion behavior feature vector of each intermediary node. The expansion behavior feature vector includes the intensity of expansion to new audiences, the breadth of community coverage, the expansion speed of acquiring new content sources, and historical expansion records. The strength of expansion to new audiences is the quotient of the number of new outgoing edges of the intermediary node within the observation window divided by the length of the observation window; the breadth of community coverage is the percentage of all outgoing edges of the intermediary node pointing to new community nodes that have not been connected in the past; the expansion speed of acquiring new content sources is the quotient of the number of incoming edges of the intermediary node from new communities that have not been contacted in the past within the observation window divided by the length of the observation window; the historical expansion record is a binary indicator of whether the current set of directed information intermediary triples has added at least one directed information intermediary triple compared to the previous observation window; The embedding vectors of each intermediary node are concatenated with the feature vectors of the expansion behavior. The concatenated vectors are then mapped through a third multilayer perceptron and activated by the sigmoid function to obtain the propagation enhancement probability of each intermediary node within the prediction window.