Method and system for predicting information popularity
By constructing a cascaded perception graph and a hybrid time-series module, the problems of noise filtering and dynamic characterization in information popularity prediction in social networks are solved, and more accurate information popularity prediction is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- JILIN UNIVERSITY
- Filing Date
- 2026-01-05
- Publication Date
- 2026-04-21
AI Technical Summary
Existing technologies struggle to effectively filter social noise, accurately characterize the long-range dependencies and continuous time dynamics of information diffusion, and quantify the uncertainties in the prediction process when predicting the popularity of information in social networks.
A cascaded perceptual graph is constructed, and node embedding representations are generated by fusing node degree encoding, temporal enhanced attention, and historical interaction enhanced attention using a graph Transformer encoder. Joint modeling is then performed using a hybrid temporal module that integrates gated recurrent units and neural stochastic differential equations to output predicted values of information popularity.
It effectively filters noisy connections in social networks, captures long-range dependencies, accurately characterizes the continuous dynamics of information diffusion, and quantifies uncertainties in the prediction process, thereby improving prediction accuracy.
Smart Images
Figure CN121903593A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of social network analysis technology, specifically a method and system for predicting the popularity of information. Background Technology
[0002] Information popularity is a crucial topic in social network analysis. Its core task is to predict how many users will ultimately receive or engage with a message (i.e., information popularity) based on the message's diffusion trajectory (i.e., information cascading) during the early observation period and the social attention relationships among users. This technology is essential for quantifying information influence and understanding information dissemination mechanisms, and has broad application value in fields such as academic citation prediction, online advertising, personalized recommendation systems, and infectious disease control.
[0003] Currently, the research methods in this technical field have mainly evolved through the following stages:
[0004] Feature engineering-based methods: an early mainstream approach. This method relies on domain knowledge to manually design and extract a series of statistical features (e.g., propagation tree depth, node degree distribution, early propagation speed, etc.) from information content, propagation users, or early cascading topology. These features are then used to train traditional machine learning models (such as support vector machines and random forests) for regression or classification prediction. The core of this method is that the quality of the manually designed features directly determines the upper limit of the model's performance.
[0005] Point process-based methods: To more naturally characterize the diffusion of information over time, researchers have introduced point process theories (such as Hawkes processes). These methods treat cascades as a sequence of temporal events, modeling future propagation probabilities by defining an intensity function that depends on historical events. Their key feature is the explicit consideration of temporal dependencies between events, providing a probabilistic explanation of the propagation process.
[0006] Deep neural network-based methods: With the development of deep learning, end-to-end models based on neural networks (such as recurrent neural networks and graph neural networks) have become a research hotspot. These methods can automatically learn complex mappings from raw data or low-dimensional features to the prediction target, reducing reliance on manual features and showing significant advantages in generalization ability, but they still have the following drawbacks.
[0007] Typically, the entire social attention network is used indiscriminately, but the numerous inactive or weak connections (i.e., social noise) within the network can interfere with the model's judgment of the true information flow path. Ordinary graph neural networks are prone to oversmoothing when modeling information propagation along cascading paths, making it difficult to capture long-range dependencies spanning multiple nodes. Most models operate on discrete time steps, making it difficult to accurately characterize the dynamics of information diffusion in continuous time dimensions, and also unable to quantify the inherent uncertainties in the prediction process. Summary of the Invention
[0008] The purpose of this invention is to provide a method and system for predicting the popularity of information, so as to solve the problems mentioned in the background art.
[0009] To achieve the above objectives, the present invention provides the following technical solution:
[0010] A method for predicting the popularity of information, the method comprising:
[0011] Acquire cascaded information data and social attention relationship data between users during the observation period; construct a cascaded perception map based on the cascaded information data and social attention relationship data;
[0012] The cascaded perceptual graph is input into the graph Transformer encoder, and node embedding representations containing global dependencies are generated by fusing node degree encoding, temporal enhanced attention and historical interaction enhanced attention, and graph-level embedding vectors are extracted.
[0013] The nodes are embedded into a hybrid temporal module that represents the input fusion gated recurrent unit and neural stochastic differential equations to jointly model the discrete event sequence and continuous time dynamics of information diffusion, and output the fused temporal hidden state.
[0014] Based on the graph-level embedding vector and the temporal hidden state, the predicted value of the popularity of the network output information is obtained.
[0015] As a further aspect of the present invention, the specific methods for constructing the cascaded sensing graph include:
[0016] For each cascaded node, select the K neighbor nodes with the highest historical forwarding counts from its global social relationships and establish directed edges; retain the actual participation timestamp for the cascaded node, and set the timestamp to 0 for newly added social neighbor nodes; calculate the interaction strength of each edge according to the following formula:
[0017] ;
[0018] Where n is the historical number of forwards on the edge, and M is the maximum number of forwards between all user pairs in the dataset.
[0019] As a further embodiment of the present invention, the calculation formula for the attention mechanism used in the graph Transformer encoder is as follows:
[0020] ;
[0021] Among them, h i For user characteristics, W Q W K Let be a learnable parameter matrix, and d be the dimension of the features. Let be the impulse function, α be a learnable parameter, and t be... i Represents a timestamp, S ij This represents the strength of the edge.
[0022] As a further embodiment of the present invention, the method of integrating the gated recurrent unit and the neural stochastic differential equation in the hybrid temporal module is as follows: the event sequence is discretely modeled by the gated recurrent unit to obtain the first hidden state; the hidden state is evolved in a continuous time interval by the neural stochastic differential equation to obtain the second hidden state; and the two are weighted and fused using a temporal gating mechanism to obtain the final hidden state.
[0023] As a further embodiment of the present invention, the prediction network is a multilayer perceptron, and its input is the concatenation vector of the graph-level embedding vector and the temporal hidden state.
[0024] As a further aspect of the present invention, the mean square logarithmic error is used as the loss function for end-to-end training.
[0025] The present invention also provides a system for predicting the popularity of information, the system comprising:
[0026] The perception graph construction module is used to acquire information cascade data and social attention relationship data between users during the observation period; and to construct a cascaded perception graph based on the information cascade data and social attention relationship data.
[0027] The encoding module is used to input the cascaded sensing graph into the graph Transformer encoder, and generate a node embedding representation containing global dependencies by fusing node degree encoding, temporal enhanced attention and historical interaction enhanced attention, and extract graph-level embedding vectors.
[0028] The fusion module is used to embed the node into a hybrid temporal module that represents the input fusion gated recurrent unit and the neural stochastic differential equation, jointly model the discrete event sequence and continuous time dynamics of information diffusion, and output the fused temporal hidden state.
[0029] The prediction output module is used to output a predicted value of the information popularity through the prediction network based on the graph-level embedding vector and the temporal hidden state.
[0030] Compared with existing technologies, the beneficial effects of this invention are: it proposes a selective graph construction strategy. Based on the real-time diffusion state of information cascading, it selectively integrates the most relevant social connections into the graph structure, thereby effectively filtering noise and focusing on truly active propagation paths.
[0031] A graph Transformer is introduced as the encoder. The self-attention mechanism can directly model the global interaction between any two nodes in the cascade, overcoming the limitations of traditional graph neural networks in deep propagation, thereby generating more expressive node feature representations.
[0032] A hybrid architecture integrating gated recurrent units (GRUs) and neural stochastic differential equations was designed. This architecture leverages the GRU's ability to process discrete sequence events while simultaneously using neural stochastic differential equations to smoothly model the diffusion process in the continuous time domain. Together, they achieve an accurate description of the continuous dynamics of information diffusion and quantify prediction uncertainty. Attached Figure Description
[0033] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention.
[0034] Figure 1 A flowchart illustrating a method for predicting the popularity of information provided in an embodiment of the present invention. Detailed Implementation
[0035] To make the technical problems, solutions, and beneficial effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0036] Figure 1 This is a flowchart illustrating a method for predicting the popularity of information. In this embodiment of the invention, a method for predicting the popularity of information includes:
[0037] Acquire cascaded information data and social attention relationship data between users during the observation period; construct a cascaded perception map based on the cascaded information data and social attention relationship data;
[0038] The cascaded perceptual graph is input into the graph Transformer encoder, and node embedding representations containing global dependencies are generated by fusing node degree encoding, temporal enhanced attention and historical interaction enhanced attention, and graph-level embedding vectors are extracted.
[0039] The nodes are embedded into a hybrid temporal module that represents the input fusion gated recurrent unit and neural stochastic differential equations to jointly model the discrete event sequence and continuous time dynamics of information diffusion, and output the fused temporal hidden state.
[0040] Based on the graph-level embedding vector and the temporal hidden state, the predicted value of the popularity of the network output information is obtained.
[0041] In this embodiment, the information cascade and the attention relationship between users during the observation period are used as model inputs. First, the most relevant social connections and cascades are integrated to form a cascade perception graph. Then, the global interaction between any two nodes in the graph is learned through the graph Transformer. Next, the discrete events and continuous time dynamics are combined through the fusion module of GRU and neural probabilistic stochastic equations. At the same time, the inherent uncertainty in the information diffusion process is modeled. Finally, the predicted cascade popularity is obtained through the multilayer perceptron.
[0042] First, a denoised cascaded perceptual graph is dynamically constructed based on historical user interaction data. This graph effectively filters noisy connections in the social network by preserving key propagation paths. Then, a graph Transformer is used to encode the cascaded perceptual graph. By incorporating information such as node degree, timestamp, and edge weights into the attention mechanism, this module can capture long-range dependencies in the cascade and generate high-quality node embedding representations.
[0043] Next, these node representations are input into a hybrid temporal model, which integrates a gated recurrent unit (GRU) and a neural stochastic differential equation (SDE). The GRU is responsible for learning the dependencies between discrete event sequences, while the neural SDE models the dynamic process of information diffusion in the continuous time domain, thus accurately characterizing its randomness and uncertainty. Finally, the temporally modeled feature representations are input into a multilayer perceptron (MLP) predictor to obtain the final popularity prediction. The entire network is trained end-to-end using mean squared log error as the loss function.
[0044] As a preferred embodiment of the present invention, the specific method for constructing a cascaded sensing graph includes:
[0045] For each cascaded node, select the K neighbor nodes with the highest historical forwarding counts from its global social relationships and establish directed edges; retain the actual participation timestamp for the cascaded node, and set the timestamp to 0 for newly added social neighbor nodes; calculate the interaction strength of each edge according to the following formula:
[0046] ;
[0047] Where n is the historical number of forwards on the edge, and M is the maximum number of forwards between all user pairs in the dataset.
[0048] In this embodiment, to overcome the problems of low computational efficiency, noise interference, and insufficient utilization of fine-grained signals in global social graphs, this invention designs a cascaded sensing graph construction method. This method first collects the historical forwarding counts between all user pairs in the data set and records the global maximum value M for normalization calculation. Based on this, for each cascaded node, the K neighbor nodes with the highest historical forwarding counts are selected from its global social relationships, and directed edges are established from this node to its neighbors (if a neighbor already exists in the cascade, it is not added again). In the node timestamp settings, the cascaded node retains the actual participation time, while the timestamp of newly added social neighbor nodes is set to 0 to distinguish between dynamic propagation and static social relationships. Finally, the interaction strength of each edge is calculated using the normalized logarithm formula.
[0049] The interaction strength calculation formula guarantees interaction strength. The value range is [0,1], forming a unified and comparable measure of connection strength.
[0050] In a preferred embodiment of the present invention, the calculation formula for the attention mechanism used in the graph Transformer encoder is as follows:
[0051] ;
[0052] ;
[0053] Among them, h i For user characteristics, W Q W K Let be a learnable parameter matrix, and d be the dimension of the features. Let be the impulse function, α be a learnable parameter, and t be... i Represents a timestamp, S ij This represents the strength of the edge.
[0054] In this embodiment, to fully utilize the structural information in the cascaded sensing graph, this module introduces three key enhancement mechanisms on the graph Transformer architecture:
[0055] Log-centrality encoding: Addressing the wide distribution and long-tail characteristics of node degree values in social networks, a log-binding strategy is designed to map the original in-degree and out-degree values to a controllable range. By adding the compressed degree embeddings to the original node features, an initial hidden representation of the node is constructed, enabling the model to maintain expressiveness while avoiding scale bias introduced by highly connected nodes.
[0056] Temporally Enhanced Attention Mechanism: To address the insensitivity of traditional Transformer structures to the temporal participation of nodes, an absolute temporal bias distributed along the diagonal is introduced into the self-attention computation. This mechanism limits the range of temporal influence through an impulse function, enabling the model to distinguish the differences in node participation at different times while maintaining computational efficiency.
[0057] Historical interaction-enhanced attention: The intensity of historical interactions, represented by edge weights, is used as a direct bias term in the attention score, enabling the model to explicitly focus on user connections with frequent historical interactions. This design incorporates normalized interaction frequencies between users into the attention calculation, strengthening the weight of key propagation paths in information aggregation.
[0058] A pre-normalization layer structure is employed to improve training stability, and a trainable virtual node is introduced as a graph-level representation. This virtual node interacts with all nodes in the graph through a self-attention mechanism, and its final layer output serves as an embedding vector representing the entire cascaded structure for use by the downstream prediction module.
[0059] In a preferred embodiment of the present invention, the hybrid time-series module integrates the gated recurrent unit and the neural stochastic differential equation in the following manner: the event sequence is discretely modeled by the gated recurrent unit to obtain the first hidden state; the hidden state is evolved in a continuous time interval by the neural stochastic differential equation to obtain the second hidden state; and the two are weighted and fused using a time-series gating mechanism to obtain the final hidden state.
[0060] In this embodiment, to address the temporal irregularities in information diffusion (such as forwarding intervals ranging from seconds to days), this invention proposes a hybrid module integrating a gated recurrent unit (GRU) and a neural stochastic differential equation (SDE). This module first uses the GRU unit to discretely model the observed event sequence, generating a semantically rich latent representation, i.e., the first hidden state:
[0061] ;
[0062] Where, s i-1 It is the hidden state of the previous time step, and θ is a learnable parameter of the GRUCell. For user embedding.
[0063] The GRU hidden states are then input into the neural SDE. A drift term neural network characterizes the deterministic evolution trend, a diffusion term neural network describes the uncertainty, and a Wiener process is used to introduce stochasticity. The Euler-Maruyama numerical method is used to solve the SDE, achieving modeling of the continuous evolution of the hidden states within irregular time intervals. The second hidden state solution process is as follows:
[0064] ;
[0065] in, For SDE from time t i-1 At time t i The solution, For drift terms, used to capture deterministic trends. This is a diffusion term used to model uncertainty.
[0066] To further integrate discrete observations with continuous dynamics, a time-series gating mechanism is introduced. This mechanism generates a fused hidden state based on the contribution of the GRU output and the SDE solution, adaptively weighted according to the time interval between events. The module's final output is the ultimate hidden state containing both the history of discrete events and the continuous time-series dynamics.
[0067] The formula for fusion is:
[0068] ;
[0069] ;
[0070] in, This represents element-wise multiplication. λ represents the time difference between two activities, and λ represents the scaling factor.
[0071] In a preferred embodiment of the present invention, the prediction network is a multilayer perceptron, and its input is a concatenated vector of the graph-level embedding vector and the temporal hidden state.
[0072] In this embodiment, the graph-level embedding h is performed. g The hidden state s after fusion with the sequence n The data is then concatenated and fed into a multilayer perceptron (MLP) to obtain the predicted popularity.
[0073] ;
[0074] in, This represents the predicted popularity.
[0075] In a preferred embodiment of the present invention, the mean square logarithmic error is used as the loss function for end-to-end training.
[0076] In this embodiment, the loss function is expressed as:
[0077] .
[0078] This invention also provides an information popularity prediction system, the system comprising:
[0079] The perception graph construction module is used to acquire information cascade data and social attention relationship data between users during the observation period; and to construct a cascaded perception graph based on the information cascade data and social attention relationship data.
[0080] The encoding module is used to input the cascaded sensing graph into the graph Transformer encoder, and generate a node embedding representation containing global dependencies by fusing node degree encoding, temporal enhanced attention and historical interaction enhanced attention, and extract graph-level embedding vectors.
[0081] The fusion module is used to embed the node into a hybrid temporal module that represents the input fusion gated recurrent unit and the neural stochastic differential equation, jointly model the discrete event sequence and continuous time dynamics of information diffusion, and output the fused temporal hidden state.
[0082] The prediction output module is used to output a predicted value of the information popularity through the prediction network based on the graph-level embedding vector and the temporal hidden state.
[0083] This invention constructs an end-to-end information popularity prediction model using deep learning methods, which has high application value. In terms of efficiency, this invention is more efficient than traditional manual feature extraction methods. In practical applications, this invention directly uses information cascades and user attention relationships as input, requiring no additional processing and making it easier to apply directly. In terms of versatility, this invention has not only verified its effectiveness on social network platforms but also demonstrated good performance on academic citation networks. Furthermore, in publicly available APS, Twitter, and Weibo datasets, the prediction accuracy of this invention outperforms other existing deep learning-based information popularity prediction models.
[0084] This invention was tested on three publicly available datasets (APS, Twitter, and Weibo), with 70% of the datasets used as the training set, 15% as the validation set, and 15% as the test set. For each dataset, two observation windows, one short and one long, were set. The specific attributes of the datasets are shown in the table below:
[0085]
[0086] The model was trained until the MSLE on the test set did not decrease after 50 epochs or for 10 consecutive epochs. The final prediction metrics of the model on the test set are shown in the table below:
[0087]
[0088] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for predicting the popularity of information, characterized in that, The method includes: Acquire cascaded information data and social attention relationship data between users during the observation period; construct a cascaded perception map based on the cascaded information data and social attention relationship data; The cascaded perceptual graph is input into the graph Transformer encoder, and node embedding representations containing global dependencies are generated by fusing node degree encoding, temporal enhanced attention and historical interaction enhanced attention, and graph-level embedding vectors are extracted. The nodes are embedded into a hybrid temporal module that represents the input fusion gated recurrent unit and neural stochastic differential equations to jointly model the discrete event sequence and continuous time dynamics of information diffusion, and output the fused temporal hidden state. Based on the graph-level embedding vector and the temporal hidden state, the predicted value of the popularity of the network output information is obtained.
2. The method for predicting the popularity of information according to claim 1, characterized in that, Specific methods for constructing cascaded sensing graphs include: For each cascaded node, select the K neighbor nodes with the highest historical forwarding counts from its global social relationships and establish directed edges; retain the actual participation timestamp for the cascaded node, and set the timestamp to 0 for newly added social neighbor nodes; calculate the interaction strength of each edge according to the following formula: ； Where n is the historical number of forwards on the edge, and M is the maximum number of forwards between all user pairs in the dataset.
3. The method for predicting the popularity of information according to claim 1, characterized in that, The calculation formula for the attention mechanism used in the graph Transformer encoder is as follows: ； Among them, h i For user characteristics, W Q W K Let be a learnable parameter matrix, and d be the dimension of the features. Let be the impulse function, α be a learnable parameter, and t be... i Represents a timestamp, S ij This represents the strength of the edge.
4. The method for predicting the popularity of information according to claim 1, characterized in that, The hybrid temporal module integrates gated recurrent units and neural stochastic differential equations as follows: the event sequence is discretely modeled by the gated recurrent unit to obtain the first hidden state; the hidden state is evolved in a continuous time interval by the neural stochastic differential equation to obtain the second hidden state; and the two are weighted and fused using a temporal gating mechanism to obtain the final hidden state.
5. The method for predicting the popularity of information according to claim 1, characterized in that, The prediction network is a multilayer perceptron, and its input is the concatenation vector of the graph-level embedding vector and the temporal hidden state.
6. The method for predicting the popularity of information according to claim 1, characterized in that, End-to-end training is performed using mean squared logarithmic error as the loss function.
7. A system for predicting the popularity of information, used to implement the method for predicting the popularity of information as described in any one of claims 1-6, characterized in that, The system includes: The perception graph construction module is used to acquire information cascade data and social attention relationship data between users during the observation period; and to construct a cascaded perception graph based on the information cascade data and social attention relationship data. The encoding module is used to input the cascaded sensing graph into the graph Transformer encoder, and generate a node embedding representation containing global dependencies by fusing node degree encoding, temporal enhanced attention and historical interaction enhanced attention, and extract graph-level embedding vectors. The fusion module is used to embed the node into a hybrid temporal module that represents the input fusion gated recurrent unit and the neural stochastic differential equation, jointly model the discrete event sequence and continuous time dynamics of information diffusion, and output the fused temporal hidden state. The prediction output module is used to output a predicted value of the information popularity through the prediction network based on the graph-level embedding vector and the temporal hidden state.