Social network event popularity prediction method
By combining bidirectional jump differential equations and variational encoders with neural differential equations, the popularity trend of social network events is explicitly modeled, which solves the problems of bidirectional dependence and insufficient overall trend capture in existing methods, and achieves more accurate and stable event popularity prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-10
- Publication Date
- 2026-03-13
AI Technical Summary
Existing methods for predicting the popularity of social network events struggle to explicitly capture overall popularity trends and bidirectional dependencies, resulting in limited predictive capabilities.
By combining bidirectional jump differential equations and variational encoders with neural differential equations, and through bidirectional attention mechanisms and multilayer perceptrons, we model the latent states and popularity trends of events, obtain prior and posterior distributions, and generate event popularity trends.
Explicitly modeling the popularity evolution process improves the robustness and stability of predictions, and enhances the accuracy and generalization ability of predicting the popularity of social network events.
Smart Images
Figure CN121660673A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of information diffusion prediction technology, and in particular to a method, apparatus, medium and device for predicting the popularity of social network events. Background Technology
[0002] With the explosive growth of social media platforms, the speed at which information is generated and spreads online has accelerated unprecedentedly. Users are not only receivers of information but also active producers and disseminators. This decentralized dissemination model enables information to cross geographical, linguistic, and community boundaries in a very short time, triggering large-scale social interaction and attention. However, not all information receives the same level of attention and spread online. Different types of information vary significantly in their speed, scope, and duration of dissemination. Therefore, understanding and predicting the popularity of information, and thus anticipating future dissemination trends, has become a fundamental issue in social networking services (SNS).
[0003] Information cascading popularity prediction focuses on modeling the overall scale and trajectory of information dissemination, rather than simply tracking individual behavior. This macro-perspective is crucial for understanding collective dynamics and emergence patterns in large-scale social systems. Furthermore, accurately estimating the future reach and impact of content is essential for various real-world social networking applications, including viral marketing, event detection, and content recommendation. Therefore, cascading popularity prediction has become a fundamental and challenging problem, attracting increasing attention.
[0004] Currently, mainstream methods for popularity prediction can be broadly categorized into feature-engineered models, generative diffusion models, and deep learning-based models. Feature-engineered models rely on manually designed features extracted from content, user profiles, and network structure, and establish a mapping between these features and popularity. For generative diffusion models, inspired by epidemiology and stochastic processes, some researchers have further improved upon specific features in social networks (such as periodicity and user heterogeneity). While these methods offer a degree of interpretability, they often struggle to capture the complex spatiotemporal patterns present in real-world cascades, thus limiting their predictive performance.
[0005] Recently, deep learning methods, particularly those utilizing sequence models such as recurrent neural networks (RNNs), Transformers, and neural ordinary differential equations (ODEs), have demonstrated superior performance in simulating continuous-time dynamics and the irregular event intervals observed in cascades. However, these methods rely solely on cascaded features to predict event popularity. CasFT is the only work attempting to generate future trends using a diffusion denoising model; however, it still lacks the ability to explicitly capture changes in overall popularity trends before the prediction time.
[0006] Most of these existing methods only consider unidirectional dependencies within a cascade, which further limits their ability to capture long-range global contextual information and fails to explicitly model the complete evolution of popularity trends, resulting in limited predictive power. Summary of the Invention
[0007] Therefore, it is necessary to provide a method, apparatus, medium, and device for predicting the popularity of social network events to address the aforementioned technical problems.
[0008] The following technical solution is adopted in this specification: This specification provides a method for predicting the popularity of social network events, including: Acquire user cascading data for target events in social networks; Based on the user concatenated data, the global and local representations of the user are obtained through a lightweight embedding method; the global and local representations are encoded through a bidirectional attention mechanism to generate the context representation of the target event; the context representation of the target event is evolved and fused through a bidirectional jump differential equation and a gated fusion module to obtain the potential state of the target event. Based on the latent state and popularity sequence of the target event, and combining multilayer perceptron and attention mechanism, an aggregated representation of the target event popularity and the event latent state is obtained; according to the aggregated representation of the event popularity and the event latent state, a prior distribution of the initial state of the event popularity trend is obtained through variational encoder; based on the prior distribution, a prior event popularity trend is generated through neural differential equation. Based on the prior event popularity trend and the potential state of the target event, predict the final popularity of the target event.
[0009] Furthermore, the user concatenation data is a collection of all users and timestamp information participating in the propagation of the target event, wherein the user concatenation data set... Represented as: ; in, The first to participate in the dissemination of informationi One user, Indicates user The timestamp of the propagation and the observation window are defined as follows: .
[0010] Furthermore, the process of obtaining the user's global and local representations includes: Construct a global graph based on all users on the social network and the interaction relationships between users. ; The global graph is described using the NetSMF algorithm. Extracting the global representation of the user ; Construct a cascade graph based on the propagation path of the target event. ; The cascade graph is described using the GraphWave algorithm. Extracting the user-local representation of the target event ; in, Represents the set of all users in a social network; A collection representing the interaction relationships between users globally; This represents the set of n users who participated in the dissemination of this information. This represents the m specific paths through which a user disseminates information.
[0011] Furthermore, the context representation for generating the target event includes: Based on the user concatenation data of the target event, the user global embedding representation is retrieved, and bidirectional attention is calculated to obtain the global bidirectional context representation; Based on the user cascaded data of the target event, the user local embedding representation is retrieved. And perform bidirectional attention calculation to obtain a local bidirectional context representation; The global bidirectional context representation and the local bidirectional context representation are concatenated to form the bidirectional context representation of the target event; The user-global embedded representation: ; in, A globally embedded representation for user-cascaded connections; User The global representation; The formula for calculating bidirectional attention is: ； ； ； in, This represents the forward global context. This represents a backward representation to the global context; and These are the forward and backward mask matrices; The user-local embedding representation: ; in, For user-cascaded local embedding representations; User Local representation; For users in the cascade sequence Temporal embedding encoded at specific locations, specifically including: ; in, Indicates the first i Each user's timestamp j Indicates the first j dimension, d It is the embedded dimension; The bidirectional context representation of the target event: and ; Where [.;.] denotes vector concatenation; Forward local context representation; This is a backward local context representation; Forward context representation; This is a backward context representation.
[0012] Furthermore, the evolution and fusion of the context representation of the target event to obtain the potential state of the target event specifically includes: The initial state of the target event is initialized as a trainable vector. H ; Based on the trainable vectors H， Initial states are generated forward using a self-gating mechanism. and backward initial state ; In the continuous time domain, the forward hidden states between events are obtained through a neural differential equation network. and the state of hiding behind Perform smooth evolution modeling; according to The actions that occur at any given time are updated using the GRU's jump procedure to update the forward initial state. for Update the backward initial state for ; Based on the updated forward initial state and the updated forward initial state Obtain the forward state of the event. and event backward state And the event forward state is achieved through a channel attention fusion mechanism. and event backward state Fusion and normalization processes generate the latent states of the sequence. ; The jump process using GRU is as follows: ; ; in, for The instant after the jump forward; for The instant after the time jump; For the first i Forward context representation at each location; For the first i The context representation of each position; The forward state of the event and event backward state Specifically: ; ; Move the event forward state and event backward state Perform weighted summation to generate the latent states of the sequence. : ; ; in, a and These are learnable parameters; Indicates for status Aggregate weights.
[0013] Furthermore, the combination of multilayer perceptron and attention mechanism to obtain an aggregated representation of event heat and event latent states specifically includes: Time interval Divide into multiple intervals on average; The popularity sequence of the target event is obtained based on the event popularity of each interval. and Specifically, it includes: ; ; in, Indicates time The previously observed event popularity sequence; express At any given moment, the popularity of the target event; ; T For hyperparameters; Indicates the future T Cumulative popularity at each point in time; Popularity sequence of the target event is obtained through a multilayer perceptron. and Encode, obtain and The specific process is as follows: ; in, Indicates time Popularity combined with timestamps; The corresponding vector representation; Based on trainability As a query, it is aggregated through an attention mechanism. and The specific process is as follows: ; ; in, Represents the popularity sequence of prior events The features obtained through encoding; Represents the posterior event popularity sequence The features obtained through encoding; Indicates the prior popularity sequence features The aggregated feature vector; Represents the posterior popularity sequence features The aggregated feature vector; This represents a trainable vector for querying popularity feature aggregation; Using trainable vectors As a query, the latent state of the target event is aggregated through an attention mechanism. Specifically, it includes: ; in, Aggregate features representing the potential states of an event; This represents a query trainable vector used to aggregate the potential states of an event sequence.
[0014] Furthermore, obtaining the prior distribution of the initial state of the event popularity trend through the variational encoder includes: Aggregation features of the popularity trend of the target event and Aggregation features of the latent state of the target event By splicing the data together, we obtain spatiotemporal characteristics that integrate macroscopic and microscopic features; The spatiotemporal features are mapped to a Gaussian distribution using a multilayer perceptron. By reparameterizing the Gaussian distribution, latent variables for modeling popularity trends are obtained. z ; The specific process is as follows: ; ; in, 、 (.)、 and It can learn from a multilayer perceptron; Represents the prior latent variables; Represents latent variables in the posterior dimension; This represents the feature vector after aggregating the features of the prior popularity sequence; This represents the feature vector after aggregating the posterior popularity sequence features; Aggregate features representing the potential states of an event.
[0015] Furthermore, the generation of prior event popularity trends via neural differential equations includes: Based on the initial value The complete popularity trajectory is generated through the constant differential equation of God, and the calculation formula is as follows: ; in, Learned by a multilayer perceptron; Based on prior latent variables Latent variables in the posterior dimension And the shared divine frequent differential equations, generating prior hidden state trajectories and posterior hidden state trajectory The calculation formula is: ; ; Based on the prior hidden state trajectory and posterior hidden state trajectory By truncating the expected value of the normal distribution, the increase in popularity is... The calculation is performed, and the specific generation process is as follows: ; ; ; in, yes The popularity of an event at any given moment; Is the popularity of the event in The time derivative at time t; and Depend on What can be learned; It is an error function; The prior future popularity trend is obtained through an ordinary differential equation solver. and posterior future popularity trend The calculation formula is: ; ; in, It is the minimum value within the time interval; The maximum value within the event interval; T This is a hyperparameter.
[0016] Furthermore, it also includes an event popularity prediction model composed of a bidirectional attention mechanism, a bidirectional jump differential equation, a gated fusion module, a multilayer perceptron, an attention mechanism, a variational encoder, and a neural constant differential equation. The parameter training process of the event popularity prediction model includes: Based on the event popularity representation and the aggregated representation of the latent state of the target event, the posterior distribution of the initial state of the event popularity trend is obtained through a variational encoder. Based on the posterior distribution, a posterior event popularity trend is generated using a neural differential equation. The final spread scale is predicted based on the posterior event popularity trend and event potential state, and the loss function is calculated to train the parameters of the popularity prediction model. Specifically, the calculation of the loss function to train the parameters of the popularity prediction model includes: By based on KL By applying knowledge distillation loss and regularization constraints to the divergence, we obtain consistent prior and posterior latent variable evolution trajectories, calculated as follows: ; in, The knowledge distillation loss represents the prior and subsequent hidden states. Indicates in The prior hidden state at a given moment; Indicates in The posterior hidden state at time t; The main loss function is calculated using the mean square logarithmic error, and the formula is as follows: ; Through regression loss The generated trend and the true trend are obtained within the range. Inside T The mean square logarithmic error at each time point; Indicates the true from arrive The amount of popularity growth at any given moment; This indicates the use of posterior hidden states and event latent states. and The predicted increase in popularity, Represents a multilayer perceptron; The overall loss function is: ; ; in, and This is a hyperparameter used to balance the contributions of various losses.
[0017] Furthermore, the final popularity of the predicted event includes: Trends are generated from prior latent variables. Then, based on the potential state of the event and To predict the incremental popularity of an event, the specific process is as follows: ; in, Represents a multilayer perceptron; It represents the future trend generated by prior latent variables.
[0018] The above-mentioned technical solutions adopted in this specification can achieve the following beneficial effects: The social network event popularity prediction method provided by this invention models the potential state evolution of events through bidirectional jump ordinary differential equations, extracts features from cascade sequences and popularity trajectories, obtains the prior and posterior distributions of popularity trends using a variational encoder, and then uses a neural ordinary differential equation to generate prior and posterior popularity trends. This not only explicitly models the complete popularity evolution process, but also improves the robustness of event popularity prediction by modeling the uncertainty of prior and posterior distributions. Furthermore, the proposed method for predicting the popularity of social network events based on situational evolution awareness, which addresses user cascading sequences in social network events, can characterize the bidirectional dependence of user propagation sequences and the uncertainty of popularity evolution, thereby enhancing the stability and generalization ability of the prediction. Attached Figure Description
[0019] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:
[0020] Figure 1 This document provides a flowchart illustrating a method for predicting the popularity of information on social networks. Figure 2 This specification provides a schematic diagram of a social network information popularity prediction method; wherein, Figure 2 (a) represents the input data; Figure 2 (b) represents a two-way jump differential equation; Figure 2 (c) represents the variational neural differential equation; Figure 3 A comparison chart of the performance and training time of the different methods provided in this manual; Figure 3 (a) represents the Twitter dataset; Figure 3 (b) represents the APS dataset; Figure 3 (c) represents the Weibo dataset. Detailed Implementation
[0021] To make the objectives, technical solutions, and advantages of this specification clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments in this specification without creative effort are within the scope of protection of this application.
[0022] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.
[0023] Figure 1 This is a schematic diagram of the system method for predicting the popularity of social network events based on situational evolution awareness, as described in this specification. The method specifically includes the following steps: S101: Obtain user cascading data for the target event in the social network.
[0024] Input the collected event user propagation cascade sequence data, and its user set It records several events that make up the sequence. The first to participate in the dissemination of information i One user; Indicates user The timestamps of the propagation are used to define the observation time window. .
[0025] S102: Based on the user cascaded data, obtain the user's global and local representations through a lightweight embedding method.
[0026] Using the NetSMF algorithm in the global graph Extracting the global representation of the user The GraphWave algorithm is used in cascaded graphs. Extracting local representations of users .
[0027] S103: Encode the global and local representations using a bidirectional attention mechanism to generate a contextual representation of the target event.
[0028] (1) For the global representation of the user: retrieve the user embedding representation based on the user sequence involved in the information cascade. ,in User The global representation is then used for bidirectional attention computation, which specifically includes: ; ; ; in, and This indicates a bidirectional context representation. and It is the forward and backward mask matrix.
[0029] (2) For user local representation: retrieve user local embedding representation based on the user sequence involved in the information concatenation. ,in User Local representation, For users in the cascade sequence Temporal embedding encoded at specific locations, specifically including: ; in, Indicates the first i Each user's timestamp j Indicates the first j dimension, d This is the embedding dimension. Then, the bidirectional representation of the cascaded graph can be obtained using the bidirectional attention encoding method described above for the global representation. and .
[0030] (3) The global and local representations after bidirectional processing are concatenated to form the final context representation. and ,in[ .;. ] indicates a vector concatenation.
[0031] S104: The context representation of the target event is evolved and fused through bidirectional jump differential equations and gating fusion module to obtain the potential state of the target event.
[0032] (1) The initial state of the event is first initialized as a trainable vector. H Subsequently, forward initializations are generated through a self-gating mechanism. shape state and backward initial state ， The specific process is as follows:
[0033] ; ; in, This represents element-wise multiplication. W and b These are learnable parameters.
[0034] Hidden state and In the interval and The continuous-time evolution is modeled by the neural ODE, and the specific process is as follows: ; in, and It is a multilayer perceptron (MLP).
[0035] When the action is in time When it occurs, the state of the cascade sequence and Pass It is updated through a jump process using GRU, the specific process is as follows: ; ; in ， and The first one obtained from the attention layer i A location context representation is used to enhance sequence modeling.
[0036] (2) Obtain the event status and Then, a channel attention fusion mechanism is applied to generate the final sequence state. The specific process is as follows: ; in, a and These are learnable parameters. At each time step, the forward and backward hidden states are weighted and summed to obtain the final sequence representation. To further improve training stability, layer normalization is applied to the output results to obtain the final latent state representation of the events. .
[0037] S105: Based on the latent state of the target event and the target event heat sequence, and combining the multilayer perceptron and attention mechanism, obtain the aggregated representation of the event heat representation and the event latent state.
[0038] (1) The interval Divided into average T The popularity of each interval is used to construct a popularity trajectory during training. and Specifically, it includes: ; in Indicates time The previously observed temporal popularity can be used as a basis for obtaining... Specifically, it includes: ; The above formula includes the future. T The cumulative popularity at each point in time, among which , T This is a hyperparameter.
[0039] (2) For the popularity sequence and Encoding is obtained through MLPs. and The specific process is as follows: ; in, Indicates time Popularity combined with timestamps This is the corresponding vector representation.
[0040] Then use a trainable As a query, it is aggregated through an attention mechanism. and The specific process is as follows: ； ； Similarly, in a similar manner, trainable vectors are utilized. As a query, the latent state of events is aggregated through an attention mechanism. ， Specifically, it includes: ; S106: Use a variational encoder to obtain the prior and posterior distributions of the initial state of the event popularity trend.
[0041] By aggregating features from popularity trends and Aggregated features from the event's latent state By concatenating these features, both macroscopic and microscopic spatiotemporal characteristics are integrated. A multilayer perceptron is used to map the features to a Gaussian distribution, and a reparameterization technique is employed to obtain latent variables for modeling popularity trends. z The specific process is as follows:
[0042] ； ； in, 、 (.)、 and It can learn from a multilayer perceptron. and These represent the prior and posterior latent variables, respectively. Meanwhile, to ensure... >0 Parameterize the standard deviation as .
[0043] S107: Use the constant differential equation to generate the prior event popularity trend and the posterior event popularity trend respectively.
[0044] (1) Obtaining initial values Subsequently, a complete popularity trajectory was generated using a regular differential equation. Simultaneously, to model the uncertainties in the generation process and ensure that the instantaneous increment is non-negative, the popularity increment was... The expected value of a truncated normal distribution is used for calculation, and the specific generation process is as follows:
[0045] ; ; ; ; in, 、 and Learned from MLPs It is an error function.
[0046] (2) Given the initial prior and posterior It uses a shared, frequently differential equation to sequentially generate prior and posterior future popularity trends via an ODE solver. and At the same time, the trajectory of the hidden state is obtained. and .
[0047] S108: During the training phase, posterior-generated trend and event sequence features are used to predict the final diffusion scale, the loss function is calculated, and the model parameters are trained.
[0048] (1) Although the prior and posterior latent variables share the same neural frequent differential equation for evolution, the small differences in the initial state will accumulate over time, eventually leading to significant bias. Therefore, a knowledge distillation loss based on KL divergence is introduced at the final time step to maintain the consistency of the evolution trajectories of the two variables through regularization constraints. The calculation formula is as follows:
[0049] ; This symmetry penalizes differences in both directions, thus enabling time-varying calculations. To more effectively align prior and posterior latent representations.
[0050] (2) During the training process, the final prediction result is obtained through the decoder, and the specific process is as follows: ; in, This represents a multilayer perceptron. It represents the future trend generated by the posterior latent variables.
[0051] The main loss function is calculated using the mean square logarithmic error, and the formula is as follows: ; Meanwhile, to make the generated trend more closely resemble the real data, regression loss is introduced. Defined as the difference between the generated trend and the true trend within a certain range. Inside TThe mean squared logarithmic error at each time point. The final overall loss function is as follows:
[0052] ; ; in, and This is a hyperparameter used to balance the contributions of various losses.
[0053] S109: During the testing phase, predict the final popularity of an event based on prior generation trends and event sequence features.
[0054] The process of predicting the popularity of an event first generates trends through prior latent variables. , then use the event's potential state and To predict the incremental popularity of an event, the specific process is as follows: ; Ultimately, the predicted event popularity increment is obtained. And output it.
[0055] The social network information popularity prediction system provided in this embodiment of the invention includes: Data input module: Used to input user cascade data collected from social networks; Graph embedding module: Uses lightweight graph embedding to obtain user global and local representations; Attention Encoding Module: Encodes user representations using a bidirectional attention mechanism; Event latent state evolution module: Based on bidirectional jump ordinary differential equations and gated fusion module, it represents the latent state of each event in the social network; Event popularity encoding module: Encodes event popularity using a multilayer perceptron and attention mechanism; Prior and posterior distribution acquisition module: Uses variational encoder to acquire the prior and posterior distributions of the initial state of the event popularity trend; Event popularity trend generation module: Generates prior event popularity trends and posterior popularity trends based on the neural network constant differential equation; Model parameter training module: During the training phase, posterior generation of trend and event sequence features is used to predict the final diffusion scale, calculate the loss function, and train the model parameters; Event popularity prediction module: Predicts the popularity of an event based on its potential state and generation trend.
[0056] To demonstrate the inventiveness and technical value of the technical solution of this invention, this section provides specific product or related technology application examples of the technical solution claimed.
[0057] Figure 2 This is a complete illustration of the social network information popularity prediction method provided in this embodiment of the invention. Specifically, the execution process of the social network information popularity prediction method of this invention can be divided into the following steps:
[0058] (a) Collecting user cascade data in social networks; (b) Lightweight graph embedding is used to obtain the user's global and local representations; (c) Encode the contextual representation of the event sequence through bidirectional attention; (d) The latent states of the social network at different time series are characterized using bidirectional jump differential equations and gated fusion modules. For example... Figure 2 (a) and Figure 2 (b) and Figure 2 As shown in (c), with user sequence For example, neural differential equations are used to bidirectionally model the latent states of evolutionary events, with the forward pass coming from... arrive , from behind arrive At any point in the timeline (e.g.) When real users participate, and At any given moment, the potential state of an event will change.
[0059] (e) Encoding event popularity using multilayer perceptron and attention mechanisms; (f) Use a variational encoder to obtain the prior and posterior distributions of the initial state of the event popularity trend; (g) Generate the prior event popularity trend and the posterior event popularity trend using the constant differential equation of God, respectively; (h) During the training phase, the final diffusion scale is predicted using posterior generated trend and event sequence features, the loss function is calculated, and the model parameters are trained. (i) During the testing phase, the final popularity of an event is predicted based on prior generation trends and event sequence features.
[0060] Figure 3 Provided for embodiments of the present invention, this paper compares the training efficiency and prediction performance of the proposed method (a social network event popularity prediction method based on situational evolution awareness, VNOIP) with baseline methods (CTCP, CasDo, CasFT) on three datasets, using two metrics, "seconds / cycle" and "mean squared logarithmic error," for comprehensive evaluation. (a) such as Figure 3 As shown in (a), the Twitter-1d dataset demonstrates that VNOIP significantly outperforms other methods in training time while maintaining a low MSLE; (b) such as Figure 3 As shown in (b), the training times of the models in the APS-3y dataset are similar, but VNOIP's MSLE metric is more competitive. (c) such as Figure 3 As shown in (c), VNOIP achieves both the fastest training speed and the highest prediction accuracy in the Weibo-0.5h scenario.
[0061] Overall results show that VNOIP achieves an optimal balance between computational efficiency and prediction accuracy through the synergistic design of bidirectional skipping ODE and variational neural ODE.
[0062] Table 1 Event sequence data used in the test Table 1 shows the real-world datasets and their user cascade sequence data used in the testing of this invention. Twitter is a social media platform tweet cascade dataset, containing tweet propagation chains collected based on tags from March 1st to April 15th, 2022. It reflects the dynamic characteristics of information spreading over time in social networks and can be used to test the model's performance in modeling rapid information propagation sequences within a short period. Weibo is a retweet cascade dataset from a major Chinese microblogging platform, containing records of user retweet behavior on the platform, presenting diverse information propagation patterns. It can be used to verify the model's effectiveness in modeling different types of social interaction behavior sequences. These datasets provide rich and representative test scenarios for VNOIP performance evaluation. Experiments on these datasets can comprehensively test the model's effectiveness and generalization ability in the information popularity prediction task.
[0063] Table 2 shows the performance of existing methods and the method of this invention (VNOIP) on the test dataset. Table 2 shows the performance comparison between the baseline model and VNOIP on the three datasets at different observation times, with the metrics being mean squared logarithmic error (MSLE) and mean absolute percentage error (MAPE) (lower values are better).
[0064] As shown in Table 2, the method (VNOIP) provided by this invention outperforms other existing methods across different time ranges on two datasets, particularly achieving the lowest performance in MSLE and MAPE metrics, demonstrating its effectiveness and superiority in related tasks. This shows that the method proposed in this invention is superior to other methods in terms of event prediction accuracy on social networks. Furthermore, the method of this invention also exhibits robust pattern recognition capabilities in high-noise data environments (such as user-generated content and real-time sensor monitoring data), effectively filtering out interference and extracting core features. It demonstrates significant practical value in fields such as intelligent recommendation and industrial fault diagnosis, possessing broad application potential and widespread adoption prospects.
[0065] Based on the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by this invention are as follows: First, addressing the technical problems existing in the prior art and the difficulty of solving them, this paper closely analyzes, in conjunction with the technical solution to be protected by this invention and the results and data obtained during the research and development process, how the technical solution of this invention solves the technical problems, and the inventive technical effects brought about by solving these problems. The specific description is as follows:
[0066] The social network event popularity prediction method provided by this invention models the latent state evolution of events using bidirectional jump ordinary differential equations. It extracts features from cascade sequences and popularity trajectories, utilizes a variational encoder to obtain the prior and posterior distributions of popularity trends, and then uses neural ordinary differential equations to generate prior and posterior popularity trends. This not only explicitly models the complete popularity evolution process but also improves the robustness of event popularity prediction by modeling the uncertainty of prior and posterior distributions. This invention proposes a situational evolution-aware social network event popularity prediction method for user cascade sequences in social network events. This method can characterize the bidirectional dependence of user propagation sequences and the uncertainty of the popularity evolution process, thereby enhancing the stability and generalization ability of prediction.
[0067] Second, considering the technical solution as a whole or from a product perspective, the technical effects and advantages of the technical solution to be protected by this invention are specifically described as follows: When preparing to analyze the evolutionary trends of event popularity on social networks, network regulators can collect user propagation cascade data (recording event occurrence and time) of different events as input data, and utilize this invention to effectively predict the popularity of social network events. The specific technical effects and advantages are as follows:
[0068] (1) This invention introduces bidirectional jump ordinary differential equations to model the potential state of events, which can simultaneously consider the past and future directions of information dissemination, effectively solving the bidirectional dependence in information dissemination, thereby improving the accuracy of event dissemination modeling and the stability of prediction, making this invention applicable to scenarios such as monitoring the trend of event popularity in social networks, predicting future trends, and judging public opinion trends. (2) The present invention designs a variational encoder module to sample the initial state of popularity from the prior and posterior distributions respectively, which is used for uncertainty modeling of event popularity trends. This effectively solves the problem of prediction bias caused by the lack of uncertainty modeling in existing methods, and further improves the accuracy and stability of the present invention in predicting the popularity of social network events. (3) This invention models the global popularity trend by introducing the method of the God constant differential equation. By extracting features from the event potential state and popularity trajectory, and obtaining the initial state through the variational encoder, the God constant differential equation is used to generate the prior popularity trend and the posterior popularity trend, thereby explicitly capturing the overall popularity trend and effectively improving the model's ability to characterize the continuous evolution of popularity over time.
[0069] Third, as supplementary evidence of the inventive step of the claims of this invention, it is also reflected in the following important aspects: (1) The expected benefits and commercial value of the technical solution of this invention after transformation are as follows: The event popularity prediction method involved in this invention can play a key role in applications such as social network situational awareness and public opinion management. Specifically, in terms of situational awareness, for high-heat events predicted by the model and their spread trends, it can track the spread dynamics, influence boundaries and evolution direction of various events in social networks in real time, accurately identify potential public opinion focus or risk signs, provide platform operators and management departments with comprehensive situational judgment basis, and greatly improve the control and response efficiency of the network environment. In terms of public opinion management, by predicting potential public opinion hotspots or risk events, it helps management departments to discover public opinion focus in a timely manner, formulate scientific response measures and public opinion guidance strategies, effectively reduce the negative impact of emergencies on the network environment and public opinion, and improve the accuracy and effectiveness of public opinion prevention and control. In situational awareness and public opinion management, this invention helps relevant entities to achieve more scientific decision-making and more efficient public opinion control, and has significant commercial application value and broad promotion prospects.
[0070] (2) The technical solution of the present invention fills the technical gap in the industry at home and abroad: In view of the limitation that existing event popularity prediction methods generally cannot explicitly capture the overall popularity evolution trend of events, the present invention adopts the neural differential equation method for the first time to explicitly characterize the whole trend of event popularity, filling this key technical gap.
[0071] (3) The technical solution of this invention solves a technical problem that people have long desired to solve but have never been able to achieve: This invention effectively solves the problem that existing methods only model unidirectional propagation and ignore bidirectional information dependence and limit global context capture by introducing bidirectional jump differential equations to model the hierarchical relationship. At the same time, the popularity trend generation method based on variational encoder designed in this invention can characterize the potential uncertainty in the event popularity trend generation process and solve the problem that existing methods are difficult to accurately reflect the continuous evolution characteristics of popularity over time.
[0072] (4) The technical solution of the present invention overcomes technical bias: The present invention improves the accuracy of event popularity prediction by modeling the event-level connection sequence through bidirectional jump differential equations, overcoming the technical bias of previous technical solutions that only modeled unidirectional propagation and ignored bidirectional information dependence. Furthermore, the present invention uses variational encoders and neural differential equations to model the uncertainty of the complete popularity trend evolution process, overcoming the technical bias of previous technical solutions that only used observation cascade structures to predict popularity.
[0073] In summary, this invention provides a method for predicting the popularity of social network events based on situational evolution awareness. This method collects cascaded user data from the social network, uses lightweight graph embedding to obtain global and local user representations, encodes the contextual representation of event sequences through bidirectional attention, and utilizes bidirectional jump differential equations and a gating fusion module to characterize the latent states of social network events. It combines a multilayer perceptron and attention mechanism to encode event popularity, effectively capturing the deep correlation between user interaction and event evolution, thus improving the accuracy of event popularity prediction. Furthermore, this invention uses a variational encoder to obtain the prior and posterior distributions of the initial state of the event popularity trend, and uses ordinary differential equations to generate prior and posterior event popularity trends. During the training phase, the posterior generated trend and event sequence features are used to predict the final diffusion scale, calculate the loss function, and train the model parameters. During the testing phase, the final event popularity is predicted based on the prior generated trend and event sequence features. This invention's method models both the micro-level evolution of the event's own state through jump differential equations and the macro-level overall evolution of the event popularity trend, enhancing the representational ability and accuracy of the event popularity prediction model. This gives VNOIP's event popularity prediction method a solid foundation for practical applications and provides valuable guidance.
[0074] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented using hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or using software executed by various types of processors, or using a combination of the above-described hardware circuitry and software, such as firmware.
[0075] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
Claims
1. A method for predicting the popularity of information on social networks, characterized in that, include: Acquire user cascading data for target events in social networks; Based on the aforementioned user concatenated data, the global and local representations of the user are obtained through a lightweight embedding method; The global and local representations are encoded using a bidirectional attention mechanism to generate the context representation of the target event; the context representation of the target event is then evolved and fused using a bidirectional jump differential equation and a gated fusion module to obtain the latent state of the target event. Based on the latent state and popularity sequence of the target event, and combining multilayer perceptron and attention mechanism, an aggregated representation of the target event popularity and the event latent state is obtained; according to the aggregated representation of the event popularity and the event latent state, a prior distribution of the initial state of the event popularity trend is obtained through variational encoder; based on the prior distribution, a prior event popularity trend is generated through neural differential equation. Based on the prior event popularity trend and the potential state of the target event, predict the final popularity of the target event.
2. The method for predicting the popularity of social network information as described in claim 1, characterized in that, The user concatenation data is a collection of all users and timestamp information participating in the propagation of the target event, wherein the user concatenation data set Represented as: ； in, The first to participate in the dissemination of information i One user, Indicates user The timestamp of the propagation and the observation window are defined as follows: .
3. The method for predicting the popularity of social network information as described in claim 1, characterized in that, The process of obtaining the user's global and local representations includes: Construct a global graph based on all users on the social network and the interaction relationships between users. ; The global graph is described using the NetSMF algorithm. Extracting the global representation of the user ; Construct a cascade graph based on the propagation path of the target event. ; The cascade graph is described using the GraphWave algorithm. Extracting the user-local representation of the target event ; in, Represents the set of all users in a social network; A collection representing the interaction relationships between users globally; This represents the set of n users who participated in the dissemination of this information. This represents the m specific paths through which a user disseminates information.
4. The method for predicting the popularity of social network information as described in claim 1, characterized in that, The context representation for generating the target event includes: Based on the user concatenation data of the target event, the user global embedding representation is retrieved, and bidirectional attention is calculated to obtain the global bidirectional context representation; Based on the user cascaded data of the target event, the user local embedding representation is retrieved. And perform bidirectional attention calculation to obtain a local bidirectional context representation; The global bidirectional context representation and the local bidirectional context representation are concatenated to form the bidirectional context representation of the target event; The user-global embedded representation: ； in, A globally embedded representation for user-cascaded connections; User The global representation; The formula for calculating bidirectional attention is: ； ； ； in, This represents the forward global context. This represents a backward representation to the global context; and These are the forward and backward mask matrices; The user-local embedding representation: ； in, For user-cascaded local embedding representations; User Local representation; For users in the cascade sequence Temporal embedding encoded at specific locations, specifically including: ； in, Indicates the first i Each user's timestamp j Indicates the first j dimension, d It is the embedded dimension; The bidirectional context representation of the target event: and ; Where [.;.] denotes vector concatenation; Forward local context representation; This is a backward local context representation; Forward context representation; This is a backward context representation.
5. The method for predicting the popularity of social network information as described in claim 1, characterized in that, The process of evolving and fusing the context representation of the target event to obtain the potential state of the target event specifically includes: The initial state of the target event is initialized as a trainable vector. H ; Based on the trainable vectors H， Initial states are generated forward using a self-gating mechanism. and backward initial state ; In the continuous time domain, the forward hidden states between events are obtained through a neural differential equation network. and the state of hiding behind Perform smooth evolution modeling; according to The actions that occur at any given time are updated using the GRU's jump procedure to update the forward initial state. for Update the backward initial state for ; Based on the updated forward initial state and the updated forward initial state Obtain the forward state of the event. and event backward state And the event forward state is achieved through a channel attention fusion mechanism. and event backward state Fusion and normalization processes generate the latent states of the sequence. ; The jump process using GRU is as follows: ； ； in, for The instant after the jump forward; for The instant after the time jump; For the first i Forward context representation at each location; For the first i The context representation of each position; The forward state of the event and event backward state Specifically: ； ； Move the event forward state and event backward state Perform weighted summation to generate the latent states of the sequence. : ； ； in, a and These are learnable parameters; Indicates for status Aggregate weights.
6. The method for predicting the popularity of social network information as described in claim 1, characterized in that, The combination of multilayer perceptron and attention mechanism to obtain an aggregated representation of event heat and event latent state specifically includes: Time interval Divide into multiple intervals on average; The popularity sequence of the target event is obtained based on the event popularity of each interval. and Specifically, it includes: ； ； in, Indicates time The previously observed event popularity sequence; express At any given moment, the popularity of the target event; ; T For hyperparameters; Indicates the future T Cumulative popularity at each point in time; Popularity sequence of the target event is obtained through a multilayer perceptron. and Encode, obtain and The specific process is as follows: ； in, Indicates time Popularity combined with timestamps; The corresponding vector representation; Based on trainability As a query, it is aggregated through an attention mechanism. and The specific process is as follows: ； ； in, Represents the popularity sequence of prior events The features obtained through encoding; Represents the posterior event popularity sequence The features obtained through encoding; Indicates the prior popularity sequence features The aggregated feature vector; Represents the posterior popularity sequence features The aggregated feature vector; This represents a trainable vector for querying popularity feature aggregation; Using trainable vectors As a query, the latent state of the target event is aggregated through an attention mechanism. Specifically, it includes: ； in, Aggregate features representing the potential states of an event; This represents a query trainable vector used to aggregate the potential states of an event sequence.
7. The method for predicting the popularity of social network information as described in claim 1, characterized in that, The prior distribution of the initial state of the event popularity trend obtained through the variational encoder includes: Aggregation features of the popularity trend of the target event and Aggregation features of the latent state of the target event By splicing the data together, we obtain spatiotemporal characteristics that integrate macroscopic and microscopic features; The spatiotemporal features are mapped to a Gaussian distribution using a multilayer perceptron. By reparameterizing the Gaussian distribution, latent variables for modeling popularity trends are obtained. z ; The specific process is as follows: ； ； in, 、 (.)、 and It can learn from a multilayer perceptron; Represents the prior latent variables; Represents latent variables in the posterior dimension; This represents the feature vector after aggregating the features of the prior popularity sequence; This represents the feature vector after aggregating the posterior popularity sequence features; Aggregate features representing the potential states of an event.
8. The method for predicting the popularity of social network information as described in claim 1, characterized in that, The generation of prior event popularity trends through neural differential equations includes: Based on the initial value The complete popularity trajectory is generated through the constant differential equation of God, and the calculation formula is as follows: ； in, Learned by a multilayer perceptron; Based on prior latent variables Latent variables in the posterior dimension And the shared divine frequent differential equations, generating prior hidden state trajectories and posterior hidden state trajectory The calculation formula is: ； ； Based on the prior hidden state trajectory and posterior hidden state trajectory By truncating the expected value of the normal distribution, the increase in popularity is... The calculation is performed, and the specific generation process is as follows: ； ； ； in, yes The popularity of an event at any given moment; Is the popularity of the event in The time derivative at time t; and Depend on What can be learned; It is an error function; The prior future popularity trend is obtained through an ordinary differential equation solver. and posterior future popularity trend The calculation formula is: ； ； in, It is the minimum value within the time interval; The maximum value within the event interval; T This is a hyperparameter.
9. The method for predicting the popularity of social network information as described in claim 1, characterized in that, It also includes an event popularity prediction model composed of a bidirectional attention mechanism, bidirectional jump differential equations, gated fusion module, multilayer perceptron, attention mechanism, variational encoder and neural constant differential equations; The parameter training process of the event popularity prediction model includes: Based on the event popularity representation and the aggregated representation of the latent state of the target event, the posterior distribution of the initial state of the event popularity trend is obtained through a variational encoder. Based on the posterior distribution, a posterior event popularity trend is generated using a neural differential equation. The final spread scale is predicted based on the posterior event popularity trend and event potential state, and the loss function is calculated to train the parameters of the popularity prediction model. Specifically, the calculation of the loss function to train the parameters of the popularity prediction model includes: By based on KL By applying knowledge distillation loss and regularization constraints to the divergence, we obtain consistent prior and posterior latent variable evolution trajectories, calculated as follows: ； in, The knowledge distillation loss represents the prior and subsequent hidden states. Indicates in The prior hidden state at a given moment; Indicates in The posterior hidden state at time t; The main loss function is calculated using the mean square logarithmic error, and the formula is as follows: ； Through regression loss The generated trend and the true trend are obtained within the range. Inside T The mean square logarithmic error at each time point; Indicates the true from arrive The amount of popularity growth at any given moment; This indicates the use of posterior hidden states and event latent states. and The predicted increase in popularity, Represents a multilayer perceptron; The overall loss function is: ； ； in, and This is a hyperparameter used to balance the contributions of various losses.
10. The method for predicting the popularity of social network information as described in claim 1, characterized in that, The predicted final popularity of the event includes: Trends are generated from prior latent variables. Then, based on the potential state of the event and To predict the incremental popularity of an event, the specific process is as follows: ； in, Represents a multilayer perceptron; It represents the future trend generated by prior latent variables.