Event propagation prediction method based on group influence

Abstract

The invention relates to the technical field of social network information propagation prediction, in particular to an event propagation prediction method based on group influence. The method comprises the following steps: constructing a social network graph of target event participating users, calculating an activation probability matrix among the users by using historical propagation data, and performing social graph structure enhancement based on a threshold value to obtain a de-noised social graph; applying a graph neural network on the de-noised social graph and introducing information bottleneck constrained self-supervised contrast learning optimization node embedding, and reserving features useful for group division and propagation prediction; dynamically dividing user groups related to propagation according to the similarity and propagation context of de-noising node embedding to form a group-level propagation sequence; based on the group representation, utilizing an attention mechanism and a group relation graph to carry out modeling on propagation influence among the groups to obtain global group representation of the event; global group representation is combined with candidate user embedding, propagation probability distribution is calculated, and future participating users of corresponding events are predicted.

Description

Technical Field

[0001] This invention relates to the field of social network information propagation prediction technology, specifically to an event propagation prediction method based on group influence. The method combines multiple advanced technologies such as Graph Neural Networks (GNNs), attention mechanisms, group partitioning, and information bottleneck theory to achieve efficient and accurate modeling and prediction of large-scale, long-chain event propagation processes in social networks. Background Technology

[0002] With the rapid development of social media and online social platforms, the propagation behavior of trending events on social networks has become an important research subject. Predicting the future propagation path, participating users, and scale of an event is not only crucial for public safety, public opinion guidance, and marketing decisions, but also provides a scientific basis for understanding human behavior patterns. Existing technologies mostly model the interaction sequences of individual users, treating the propagation process as a time series, and employing Recurrent Neural Networks (RNNs), Long Short-Term Memory Networks (LSTMs), or Transformer-based models to model the influence relationships between users. However, these methods have the following shortcomings: when the event scale is large and the propagation chain is long, the time and space complexity increases significantly; the influence relationships between individual users are very sparse and easily affected by noise, resulting in poor robustness of the prediction results; existing models struggle to simultaneously capture both tight relationships within local groups and global long-range dependencies. Therefore, a method is urgently needed that can fully utilize the group influence relationships between users in social networks, effectively reduce computational complexity, and improve prediction accuracy and robustness to noise. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention proposes an event propagation prediction method based on group influence. This method uses groups as the modeling unit, dynamically identifying groups and modeling the propagation process at the group level, significantly reducing complexity and improving robustness. The method mainly includes the following key technological innovations:

[0004] Based on historical cascaded data, user activation probability is estimated. By combining threshold sparsity and structure enhancement, noisy edges in the social graph are effectively removed, and node embedding quality is improved.

[0005] Introducing information bottleneck constraints and self-supervised contrastive learning into node representation learning enables node embeddings to retain core features useful for group identification and propagation prediction.

[0006] Based on the denoised node embedding and propagation context, users are dynamically aggregated into groups, and propagation is modeled at the group level.

[0007] Improve prediction accuracy by capturing potential influence relationships between groups through group-level attention mechanisms and graph neural networks;

[0008] A multi-task learning framework is adopted to jointly optimize the propagation prediction task and the information bottleneck comparison learning task, so as to ensure that the model has good robustness while maintaining accuracy.

[0009] The aforementioned technological innovations maintain the accuracy of propagation prediction while reducing computational resource consumption, making them particularly suitable for large-scale, long-chain propagation prediction scenarios for hot events.

[0010] The technical solution adopted in this invention is as follows:

[0011] A method for predicting event propagation based on group influence includes the following steps:

[0012] Step 1: Construct a social network graph G = (V, E) of the target event participants, where the node set V represents the set of event participants, and the edge set... The social network graph G is represented by social or propagation relationships between users. Then, the activation probability matrix between users is calculated using historical propagation data, and edges with activation probabilities less than a preset threshold are removed. This process is then used to sparsify and denoise the social network graph G, resulting in the denoised social network graph.

[0013] Step 2: In the original social network graph G and the denoised social network graph The graph neural network is applied to encode the nodes to obtain the original embedding and the denoised embedding that reflect the characteristics of user propagation behavior. The optimization direction of the node embedding is constrained by the information bottleneck theory. Noise information that is not related to propagation is removed from the node embedding through self-supervised contrastive learning, while retaining the core feature information that is useful for group partitioning and propagation prediction, thus generating robust node embeddings.

[0014] Step 3: Based on the similarity of the denoised node embeddings and the propagation context features, dynamically identify user groups and use clustering algorithms to form multiple propagation-related user groups, thereby constructing a group-level propagation sequence;

[0015] Step 4: Based on the group-level propagation sequence, apply a group propagation modeling framework that combines group attention mechanism and multi-head self-attention to calculate the propagation influence relationship between groups and obtain the global propagation embedding representation of the event;

[0016] Step 5: Calculate the propagation probability distribution based on the global propagation embedding and the node embedding of the candidate users to predict the propagation status of the event at future times.

[0017] Furthermore, in step 1, the activation probability matrix between users is calculated using historical propagation data, and edges with activation probabilities less than a preset threshold are removed. This process is then used to sparsify and denoise the social network graph G, resulting in a denoised social network graph. The specific process is as follows:

[0018] For noisy edges in a social network graph G, the activation probability matrix P between users is calculated using historical propagation data, where P ij This indicates that user v is in the historical event propagation data. i Trigger user v j The probability of P is calculated, and then G is sparsified according to a preset threshold ∈, removing those that satisfy P. ij Edges <∈ are used to obtain the denoised social network graph. Where Φ(·) represents the graph structure enhancement operation based on diffusion features.

[0019] Furthermore, the specific process of step 2 is as follows:

[0020] In the original social network graph G and the denoised social network graph The above uses graph neural networks for node encoding, respectively for each user v i Generate the original embedding z i and denoising embedding z' i Then for each user v i Define a positive sample set pos(i) and a negative sample set neg(i), and introduce self-supervised contrastive learning with information bottleneck constraints. Use the Info-NCE loss function to contrastively optimize the denoised embedding z'. i :

[0021]

[0022] Where γ is the equilibrium hyperparameter, sim(a,b)=exp cos(a,b) / τ cos(·,·) is the cosine similarity function, z j and z k For users v j and user v k The original embedding, the positive sample set pos(i) includes users v whose interaction probability is higher than a set value. j The negative sample set neg(i) includes users v whose interaction probability is lower than a set value. k ;

[0023] By minimizing L IB This process removes irrelevant noise from the node representation while retaining core information useful for group identification, generating a more robust node embedding z'. i .

[0024] Furthermore, the specific process of step 3 is as follows:

[0025] Based on the preset number of groups K or a similarity threshold, a clustering algorithm is used to embed the denoised nodes obtained in step 2 into z'. i Users with similar propagation characteristics are grouped into the same group, ultimately resulting in a group-level propagation sequence {s1,s2,…,s}. K}, where each group s j This includes a group of users who meet the similarity requirements in terms of propagation characteristics.

[0026] Furthermore, step 4 is as follows:

[0027] Constructing the embedding matrix of group-level propagation sequences The multi-head self-attention model is used to compute the representation of the sequence.

[0028]

[0029] in, and The weight matrix is ​​a learnable matrix; the masking matrix is ​​a learningable matrix. Set the upper triangle to -∞; Q l K represents the query value of the l-th attention head; l V represents the key value of the l-th attention head; l d represents the value of the l-th attention head; d represents the dimension of the embedding representation.

[0030] By applying a multi-head attention mechanism and a feedforward neural network, MHA(Q) is... l ,K l V l The result is a tie, which ultimately yields the event propagation embedding representation.

[0031] Furthermore, step 5 is as follows:

[0032] Let z' x For candidate user v x Denoising social network graphs The node embedding on is calculated in the following form: x Predicted score for propagation at the next moment

[0033]

[0034] Where σ(·) is the Sigmoid function and τ is the temperature parameter, used to control the smoothness of the predicted distribution;

[0035] Ultimately, event C is determined based on the predicted score ranking or threshold of each candidate user.i The next step is to gather participating users.

[0036] Compared with the prior art, the present invention has the following advantages:

[0037] 1. Innovatively shifting propagation modeling from the user level to the group level significantly reduces computational complexity;

[0038] 2. Improve node embedding quality and enhance group partitioning and prediction accuracy by optimizing graph sparsity and information bottlenecks;

[0039] 3. The group-level attention mechanism takes into account both local and global propagation features, improving the model's performance in long-chain propagation scenarios;

[0040] 4. The multi-task learning framework enhances model robustness and alleviates overfitting;

[0041] 5. The algorithm is suitable for predicting the spread of hot events over large scale and long time spans, and has good scalability. Attached Figure Description

[0042] Figure 1 This is an overall framework diagram of the event propagation prediction method according to an embodiment of the present invention. Detailed Implementation

[0043] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0044] To enable those skilled in the art to more clearly understand the present invention, the event propagation prediction method based on group influence of the present invention will be described in detail below with reference to embodiments. The present invention can be further adjusted and optimized according to specific needs, and the steps are not limited to the following order.

[0045] like Figure 1 This embodiment discloses a method for predicting the propagation of social network events based on the circle effect, including the following steps:

[0046] Step 1: First, define the relevant notation and construct the social network model. Define a directed graph G = (V, E) to represent the social network, where the set of nodes V represents the set of users participating in the event, and the set of edges... Represents social or propagation relationships between users; cascade C of the propagation of any event i. i Represented as a sequence of user participation sorted by timestamp:

[0047] C i =((v1,t1),(v2,t2),...(v M ,t M ))

[0048] Where M is the number of users, and each tuple (v j ,tj ) represents user v j At time t j They participated in the dissemination of this event. To introduce the concept of groups, user groups are defined. This represents a set of users with similar propagation characteristics;

[0049] Then, the social graph is sparsified and enhanced. For noisy edges in the social network graph G, the activation probability matrix P between users is first calculated using historical propagation data, where P... ij This indicates that user v is in the historical event propagation data. i Trigger user v j The probability of P is calculated, and then G is sparsified according to a preset threshold ∈, removing those that satisfy P. ij Edges <∈ are used to obtain the denoised social network graph. Where Φ(·) represents the graph structure enhancement operation based on diffusion features, which effectively reduces the interference of noisy connections on subsequent representation learning.

[0050] Step 2: Node Embedding Learning and Information Bottleneck Optimization. This involves learning the original social network graph G and optimizing the denoised social network graph. Graph Neural Networks (GNNs) are applied to encode nodes in both the original social network graph G and the denoised social network graph G. Specifically, graph convolutional networks (GCNs) and graph attention networks (GATs) can be used to encode the nodes. Perform graph learning separately and generate the original embedding z for each node. i and denoising embedding z' i Then, to optimize node embedding and remove irrelevant information, a self-supervised contrastive learning approach with an information bottleneck constraint is introduced: for each user v i Define a positive sample set pos(i) (containing a high interaction probability P) ij User v j ) and negative sample set neg(i) (users with low interaction probability v) k The Info-NCE loss function is used to compare and optimize the denoising embedding z'. i To make the user embeddings of the same group (with similar interaction features) as similar as possible, and to be similar to the original embedding z i And the differences in user embeddings (with different interaction characteristics) among different groups should be as large as possible:

[0051]

[0052] Where γ is the equilibrium hyperparameter, sim(a,b)=exp cos(a,b) / τ cos(·,·) is the cosine similarity function, z j and zk For users v j and user v k The original embedding, the positive sample set pos(i) includes users v whose interaction probability is higher than a set value. j The negative sample set neg(i) includes users v whose interaction probability is lower than a set value. k ;

[0053] By minimizing L IB This process removes irrelevant noise from node representations while retaining core information useful for group identification, thereby guiding the model to generate more robust node embeddings z'. i .

[0054] Step 3: User Group Identification. Use the denoised node embedding z' obtained in Step 2. i Grouping can be performed using clustering algorithms such as K-means or spectral clustering. Specifically, based on a preset number of groups K or a similarity threshold, users who are similar in the embedding space are grouped into the same group, ultimately obtaining a group set {s1, s2, ..., s...}. K}, where each group s j This includes a group of users who meet the similarity requirement in terms of propagation characteristics; this process effectively cascades the user set C involved in the original large-scale event propagation. i Divided into a relatively small number of group-based event propagation cascades Subsequent propagation modeling was conducted at the group level.

[0055] Step 4: Group-level propagation modeling. An attention-based model is used on the group sequences to extract global propagation features; specifically, a group sequence matrix can be constructed. The multi-head self-attention model is used to compute the representation of the sequence.

[0056]

[0057] in, It is a learnable weight matrix; masking matrix Set the upper triangle to -∞ to prevent information leakage, Q l K represents the query value of the l-th attention head; l V represents the key value of the l-th attention head; l d represents the value of the l-th attention head; d represents the dimension of the embedding representation.

[0058] By applying a multi-head attention mechanism and a feedforward neural network, MHA(Q) is... l ,K l V l The result is a tie, which ultimately yields the event propagation embedding representation. Used for the analysis and prediction of event propagation.

[0059] Step 5: Propagation Prediction and Inference. Based on the output representation of the group-level propagation model in Step 4, predict future propagation; specifically, let... For cascaded C i The final group indicates that z' x For candidate user v x Denoising social network graphs The node embedding on the node can be computed in the following form: x Predicted score for propagation at the next moment

[0060]

[0061] Where σ(·) is the Sigmoid function, and τ is a temperature parameter used to control the smoothness of the prediction distribution; finally, based on the ranking of the prediction scores of each candidate user or the threshold, event C can be obtained. i The next step involves participating user sets; through the representation of a group-level model, both the influence between local groups and the global propagation information are considered, thus having a larger processable sequence length and higher efficiency.

[0062] During the parameter training process in the above steps, a multi-task learning strategy is adopted: the main task is supervised learning for propagation prediction, using a prediction loss L based on cross-entropy. pre The auxiliary task is the self-supervised contrastive loss L for the aforementioned information bottleneck. IB Specifically, the prediction loss L based on cross-entropy pre as follows:

[0063]

[0064] Wherein, if user v x If y participates in event propagation in step i, then x,i =1, otherwise y x,i =0, C k V represents the set of all participants in a specific event, and V represents all users of the social network. Let L represent the prediction result of the participant in step i. The total loss is L = L pre +L IB Through joint optimization, the model can accurately predict future propagation while effectively removing noise components from user embeddings, thus achieving a highly efficient and effective prediction model.

[0065] Each step in the above method can be implemented using corresponding algorithm modules, including: a Graph Neural Network module for node encoding of social and group graphs, a clustering algorithm for user group identification, an attention mechanism-based sequence modeling module for group-level propagation modeling, and an information bottleneck optimization module for embedding denoising and representation enhancement. These modules work together to realize the event propagation prediction function based on group influence described in this invention.

[0066] The preferred embodiments of the present invention have been described in detail above with reference to the accompanying drawings. These embodiments are merely descriptions of preferred embodiments and are not intended to limit the scope or concept of the invention. The specific technical features described in the above embodiments can be combined in any suitable manner without contradiction. Such combinations, as long as they do not violate the spirit of the present invention, should also be considered as part of this disclosure. To avoid unnecessary repetition, the present invention will not further describe the various possible combinations.

Claims

1. A method for predicting event propagation based on group influence, characterized in that, Includes the following steps: Step 1: Construct a social network graph G = (V, E) of the target event participants, where the node set V represents the set of event participants, and the edge set... The social network graph G is represented by social or propagation relationships between users. Then, the activation probability matrix between users is calculated using historical propagation data, and edges with activation probabilities less than a preset threshold are removed. This process is then used to sparsify and denoise the social network graph G, resulting in the denoised social network graph. Step 2: In the original social network graph G and the denoised social network graph The graph neural network is applied to encode the nodes to obtain the original embedding and the denoised embedding that reflect the characteristics of user propagation behavior. The optimization direction of the node embedding is constrained by the information bottleneck theory. Noise information that is not related to propagation is removed from the node embedding through self-supervised contrastive learning, while retaining the core feature information that is useful for group partitioning and propagation prediction, thus generating robust node embeddings. Step 3: Based on the similarity of the denoised node embeddings and the propagation context features, dynamically identify user groups and use clustering algorithms to form multiple propagation-related user groups, thereby constructing a group-level propagation sequence; Step 4: Based on the group-level propagation sequence, apply a group propagation modeling framework that combines group attention mechanism and multi-head self-attention to calculate the propagation influence relationship between groups and obtain the global propagation embedding representation of the event; Step 5: Calculate the propagation probability distribution based on the global propagation embedding and the node embedding of the candidate users to predict the propagation status of the event at future times.

2. The event propagation prediction method based on group influence according to claim 1, characterized in that, In step 1, the activation probability matrix between users is calculated using historical propagation data, and edges with activation probabilities less than a preset threshold are removed. This process is then used to sparsify and denoise the social network graph G, resulting in the denoised social network graph. The specific process is as follows: For noisy edges in a social network graph G, the activation probability matrix P between users is calculated using historical propagation data, where P ij This indicates that user v is in the historical event propagation data. i Trigger user v j The probability of P is calculated, and then G is sparsified according to a preset threshold ∈, removing those that satisfy P. ij Edges <∈ are used to obtain the denoised social network graph. Where Φ(·) represents the graph structure enhancement operation based on diffusion features.

3. The event propagation prediction method based on group influence according to claim 1, characterized in that, Step 2 is as follows: In the original social network graph G and the denoised social network graph The above uses graph neural networks for node encoding, respectively for each user v i Generate the original embedding z i and denoising embedding z' i Then for each user v i Define a positive sample set pos(i) and a negative sample set neg(i), and introduce self-supervised contrastive learning with information bottleneck constraints. Use the Info-NCE loss function to contrastively optimize the denoised embedding z'. i : Where γ is the equilibrium hyperparameter, sim(a,b)=exp cos(a,b) / τ cos(·,·) is the cosine similarity function, z j and z k For users v j and user v k The original embedding, the positive sample set pos(i) includes users v whose interaction probability is higher than a set value. j The negative sample set neg(i) includes users v whose interaction probability is lower than a set value. k ; By minimizing L IB This process removes irrelevant noise from the node representation while retaining core information useful for group identification, generating a more robust node embedding z'. i .

4. The event propagation prediction method based on group influence according to claim 3, characterized in that, Step 3 is as follows: Based on the preset number of groups K or a similarity threshold, a clustering algorithm is used to embed the denoised nodes obtained in step 2 into z'. i Users with similar propagation characteristics are grouped into the same group, ultimately resulting in a group-level propagation sequence {s1,s2,…,s}. K }, where each group s j This includes a group of users who meet the similarity requirements in terms of propagation characteristics.

5. The event propagation prediction method based on group influence according to claim 1, characterized in that, Step 4 is as follows: Constructing the embedding matrix of group-level propagation sequences The multi-head self-attention model is used to compute the representation of the sequence. in, and The weight matrix is ​​a learnable matrix; the masking matrix is ​​a learningable matrix. Set the upper triangle to -∞; Q l K represents the query value of the l-th attention head; l V represents the key value of the l-th attention head; l d represents the value of the l-th attention head; d represents the dimension of the embedding representation. By applying a multi-head attention mechanism and a feedforward neural network, MHA(Q) is... l ,K l V l The result is a tie, which ultimately yields the event propagation embedding representation.

6. The event propagation prediction method based on group influence according to claim 5, characterized in that, Step 5 is as follows: Let z' x For candidate user v x Denoising social network graphs The node embedding on is calculated in the following form: x Predicted score for propagation at the next moment Where σ(·) is the Sigmoid function and τ is the temperature parameter, used to control the smoothness of the predicted distribution; Finally, based on the predicted scores or thresholds of each candidate user, the next set of participating users for event Ci is obtained.

Images