A social network information cascade prediction method based on user cognition

Abstract

This invention relates to the field of social network analysis technology, and more particularly to a method for predicting the cascading propagation of social network information based on user cognition. The method includes: acquiring propagation data of target information before the observation time point; extracting user fusion features based on the propagation data and constructing a fusion feature sequence in chronological order; constructing a cascading propagation graph based on the propagation data; dynamically modeling the evolution of user opinion states along the propagation path to generate user propagation bias; constructing and training a cascading prediction model, which encodes the fusion feature sequence and introduces a propagation bias during the encoding process to output a predicted value for the future propagation scale; and using the trained model to predict the propagation scale of new information cascading. This invention refines the dynamics of propagation time sequence through continuous-time encoding, collaboratively captures global and local dependencies through dual-channel time-series modeling, and quantifies the impact of user cognitive states on propagation decisions through opinion evolution, thereby improving the accuracy and robustness of information cascading scale prediction.

Description

Technical Field

[0001] This invention relates to the field of social network analysis technology, and in particular to a method for predicting the cascading information of social networks based on user cognition. Background Technology

[0002] With the rapid development of the internet and social media platforms, information spreads in a cascading manner online through interactive behaviors such as forwarding, commenting, and quoting. Accurately predicting the future scale and evolution of information dissemination is of crucial value for applications such as public opinion monitoring, hot topic discovery, rumor control, content recommendation, and maximizing influence.

[0003] Existing research typically models the information propagation process as a sequence or growth graph that evolves over time, and focuses on extracting key features from historical propagation trajectories to predict its final scale. Mainstream methods generally recognize that temporal and structural features are two core dimensions affecting prediction accuracy. Regarding temporal feature modeling, early research and some existing methods employ discretization strategies, such as fixed time window division or segmented aggregation, to convert continuous propagation events into discrete sequences. These methods struggle to accurately characterize the non-uniform, sudden propagation rhythms in real-world scenarios, easily leading to the loss of fine-grained temporal information. In terms of structural feature modeling, techniques such as graph neural networks are used to capture cascading topological structures. However, many methods focus on extracting structural features from a single scale (either global or local), failing to effectively coordinate the complementary relationship between global network structure priors and local cascading evolution structures.

[0004] In modeling temporal dependencies, recurrent neural networks and their variants are widely used to capture sequential dependencies. However, these models are better at capturing local short-term patterns and have limited ability to model long-term dependencies that may exist in cascades. Although some studies have introduced attention mechanisms or Transformer architectures to enhance global awareness, they may still lack sensitivity to local dynamic changes. Therefore, how to coordinate the modeling of long-term global trends and short-term local fluctuations in the propagation process remains a challenge.

[0005] Furthermore, most existing technologies treat users as passive, homogeneous dissemination nodes, neglecting the changes in users' internal cognitive states during information dissemination. In reality, users' acceptance of information and their attitudes towards it evolve dynamically due to their social relationships, historical interactions, and the information content itself. This cognitive state directly influences their subsequent dissemination decisions (such as whether to forward and when to forward). Traditional methods lack explicit modeling of this cognitive dynamic, making it difficult to reveal and utilize the differentiated dissemination mechanisms arising from differences in user perspectives, thus limiting the interpretability and accuracy of predictive models in real-world scenarios.

[0006] In summary, existing information cascade prediction methods still have shortcomings in terms of detailed characterization of continuous time dynamics, joint modeling of global and local multi-scale temporal and structural features, and propagation dynamics modeling that integrates user cognitive evolution. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a method for predicting cascading information in social networks based on user cognition, comprising the following steps: S1: Obtain the propagation data of the target information before the observation time point, including user identifier, propagation relationship between users and propagation timestamp; extract user fusion features based on the propagation data, and arrange them in the propagation time sequence to form a fusion feature sequence. S2: Construct a cascaded propagation graph based on the propagation data, and dynamically model the evolution of the user's viewpoint state along the path of the cascaded propagation graph to generate the user's propagation bias; S3: Construct and train a cascaded prediction model, which is configured to encode the fused feature sequence and introduce the propagation bias during the encoding process to output a predicted value of the future propagation scale of the target information; S4: Using the cascade prediction model after training, predict the propagation scale of the new information cascade.

[0008] Preferably, in step S1, extracting user fusion features based on the propagation data further includes: S11: Construct a user sparse adjacency matrix based on the global social relationship network, and normalize and decompose the user sparse adjacency matrix to obtain the user global structure embedding representation; S12: Construct an information cascade subgraph based on the propagation data, and use a multi-scale analysis method based on thermal diffusion kernels to extract the cascaded local structure embedding representation of the user; S13: Normalize the timestamps of user participation in dissemination, and use a continuous time coding function to map the normalized timestamps into a time feature embedding representation containing linear trend components and periodic components. S14: The global structure embedding representation, the cascaded local structure embedding representation, and the time feature embedding representation of the same user are concatenated to obtain the user's fusion feature, and the fusion features of all users are sorted by timestamp to form the fusion feature sequence.

[0009] Preferably, in step S13, the normalized timestamp is mapped to a time feature embedding representation containing linear trend components and periodic components using a continuous-time encoding function, further including: Normalized timestamps The input continuous-time encoding function Time2Vec(·) is mapped to a (k+1)-dimensional embedding vector, as shown below: in, , , ... and , , ... These are the learnable parameters during model training.

[0010] Preferably, in step S12, the user's cascaded local structure embedding representation is extracted based on the multi-scale analysis method of thermally diffusing kernels, including: Construct a cascaded adjacency matrix based on the aforementioned information cascaded subgraph. and its degree matrix According to the cascaded adjacency matrix and the degree matrix Calculate the normalized Laplacian matrix As shown below: in, It is the identity matrix; Through thermal diffusion kernel functions at different diffusion scales The transformation result is obtained by transforming the cascaded graph structure, wherein the thermal diffusion kernel function is... As shown below: By aggregating the transformation results, the user's cascaded local structure embedding representation is extracted, as shown below: in, For the embedded dimension.

[0011] Preferably, in step S2, the dynamic evolution modeling of the user's viewpoint state along the path of the cascading propagation graph to generate the user's propagation bias includes: S21: Set the initial view state of the cascaded source node to the first value, and set the initial view state of the inactive node to the second value. S22: Based on the actual time sequence of information propagation, when a user node is activated, the view status of the activated node is updated according to the current view status of all its activated upstream neighbor nodes that are the direct source of propagation. S23: After completing the update of the viewpoint state corresponding to all propagation events before the observation point, use the final viewpoint state value of each user as its corresponding propagation bias.

[0012] Preferably, in step S22, updating the view state of the activated node further includes: Calculate the activated node viewpoint status , is the activated node All activated propagation upstream neighbor nodes viewpoint status The weighted average is shown below: ,in For users For users Influence weighting coefficient This indicates the impact on users during this communication event. The set of upstream users that have an impact.

[0013] Preferably, in step S3, constructing and training the cascaded prediction model includes: S31: Construct a parallel dual-channel temporal coding architecture, wherein the first channel of the parallel dual-channel is a global dependency modeling channel based on a Transformer encoder, and the second channel of the parallel dual-channel is a local dependency modeling channel based on a bidirectional gated recurrent unit network. S32: The propagation bias is used as a dynamic adjustment term and introduced into the gating calculation process of the second channel to update the modulation information; S33: Input the fused feature sequence into the first channel and the second channel simultaneously to obtain the global temporal representation and the local temporal representation respectively, and then adaptively fuse the global temporal representation and the local temporal representation through learnable weight parameters to obtain the fused temporal representation; S34: Perform summation and pooling on the fused temporal representation to obtain a cascaded global state vector of fixed dimensions; S35: Input the cascaded global state vector into the multilayer perceptron and output the propagation scale prediction value; S36: Use the mean square logarithmic error as the loss function to perform end-to-end training on all parameters of the model.

[0014] Preferably, in step S32, the gating calculation process for the second channel further includes: For users Its update gate in the bidirectional gated loop unit and reset door The calculation method is as follows: in, For users The fusion characteristics The state was hidden in the previous moment. For users The propagation bias, For learnable scaling factors, It is the sigmoid activation function. , , , These are the network weight parameters.

[0015] Preferably, in step S33, the adaptive fusion of the global temporal representation and the local temporal representation using learnable weight parameters further includes: The adaptive fusion uses learnable weight parameters. accomplish: in, This is the global timing representation of the first channel output. This is the local timing representation of the second channel output. To integrate temporal representations, These are learnable fusion weight parameters.

[0016] Preferably, step S4 further includes: S41: Obtain new propagation data of the new information to be predicted cascaded before the observation time point; S42: Based on the new propagation data, extract the new cascade fusion feature sequence of the new cascade; S43: Construct a new cascaded propagation graph based on the new propagation data, and generate a new propagation bias corresponding to the user; S44: The cascaded prediction model trained with the new cascaded fusion feature sequence and the new propagation bias input is directly used to obtain the predicted value of its future propagation scale.

[0017] Compared with the prior art, the beneficial effects of the present invention are: This invention acquires the propagation data of target information before the observation time point, and extracts a fusion feature sequence that integrates the user's global network structure, cascaded local propagation structure, and continuous time dynamics, thereby achieving a multi-dimensional and fine-grained representation of the information propagation process. This overcomes the information loss problem caused by traditional methods due to single or discretized feature processing, laying a solid foundation for subsequent accurate prediction and significantly improving the model's ability to characterize complex propagation patterns. This invention constructs a cascading propagation graph and introduces an opinion evolution model along its path to dynamically model user states, generating propagation biases. This achieves an explicit characterization of users' internal cognitive dynamics during information dissemination. This allows the model to quantify and utilize the impact of user opinion similarity on propagation decisions, overcoming the limitations of traditional methods that treat users as homogeneous nodes. Consequently, in scenarios such as public opinion analysis and rumor warning, it can more accurately simulate and predict the diffusion path and scale of information. This invention constructs a dual-channel temporal coding prediction model with fused propagation bias. It utilizes a Transformer encoder to capture global long-term dependencies while employing a bias-introduced bidirectional gated recurrent unit network to capture local short-term dynamics. An adaptive mechanism fuses the outputs of both models, ultimately achieving end-to-end prediction of future propagation scale. This collaboratively addresses the shortcomings of existing methods in simultaneously handling long-term and short-term dependencies, ensuring the stability and accuracy of the prediction model across different propagation rhythms and lifecycle cascades. Therefore, it provides reliable technical support for content management and risk control on social media platforms. Attached Figure Description

[0018] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0019] Figure 1 The flowchart of the social network information cascading prediction method based on user cognition of the present invention is shown below; Figure 2 This is a schematic diagram of the cascaded prediction model of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages 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 of the invention and are not intended to limit the invention. Obviously, the described embodiments are only some, not all, of the embodiments described in this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without creative effort are within the scope of protection of this application.

[0021] Those skilled in the art will understand that, unless otherwise stated, the singular forms “a” and “an” used herein, and “the”, may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0022] Please see Figure 1 As shown in the figure, this embodiment provides a social network information cascading prediction method based on user cognition. By introducing the Degroot viewpoint evolution model and representing non-uniform time series in a continuous time manner, and combining global network structure and local cascading structure features, it addresses the shortcomings of traditional methods in time and structure modeling, and improves the accuracy and robustness of information cascading prediction. The method includes the following steps: S1: Obtain the propagation data of the target information before the observation time point, including user identifiers, propagation relationships between users, and propagation timestamps. Extract user fusion features based on the propagation data and arrange them in the propagation time sequence to form a fusion feature sequence.

[0023] Preferably, in step S1, extracting user fusion features based on propagation data further includes: S11: Construct a user sparse adjacency matrix based on the global social relationship network to represent the structural connection relationships between user nodes. Normalize and perform low-rank decomposition on the user sparse adjacency matrix to obtain the user global structural embedding representation. Specifically, in this embodiment, when a user... With users When there is a focus relationship or interaction relationship between the elements in the adjacency matrix, ,otherwise Due to the large scale of the social network, the adjacency matrix is ​​stored and computed in sparse matrix form.

[0024] More preferably, to reduce the impact of different node degree distributions on the structural representation results, the sparse adjacency matrix is ​​normalized to construct a normalized adjacency matrix, as shown below: in, Let be a degree matrix, whose diagonal elements satisfy... .

[0025] Based on this, the normalized adjacency matrix A low-rank approximate decomposition is performed to extract the latent structural features of the user in the global network. The decomposition form is expressed as: in, and It is a low-dimensional feature mapping matrix. For a singular value diagonal matrix, by selecting the first... Each principal component can be used to obtain the global structural embedding representation for each user.

[0026] Therefore, users The global structure embedding is represented as:

[0027] S12: Construct an information cascade subgraph based on propagation data, and use a multi-scale analysis method based on thermal diffusion kernels to extract the user's cascaded local structure embedding representation. Specifically, in this embodiment, for a given observation time point... In a previously established information propagation process, user nodes participating in the information propagation are selected, and a corresponding information cascade network structure is constructed. The cascade network consists of user nodes participating in the propagation and the actual propagation path of information between nodes. Based on the wavelet analysis method of heat diffusion, the cascade graph is modeled in a multi-scale structure, and the role characteristics of nodes in the local propagation structure are extracted to form an embedded representation of the cascade structure.

[0028] More preferably, a cascaded adjacency matrix is ​​constructed based on the information cascaded network. When information is provided by the user Distributed to users season ,otherwise And construct the corresponding degree matrix accordingly. Its diagonal elements satisfy:

[0029] Preferably, in step S12, the user's cascaded local structure embedding representation is extracted based on the multi-scale analysis method of thermally diffusing kernels, including: To characterize the local structural position and dissemination role of users in the specific information dissemination process, a cascaded adjacency matrix is ​​constructed based on the information cascaded subgraph. and its degree matrix According to the cascaded adjacency matrix Sum-degree matrix Calculate the normalized Laplacian matrix As shown below: ,in, It is the identity matrix; Through thermal diffusion kernel functions at different diffusion scales The following transformation is performed on the cascade diagram structure to obtain the transformation result, the thermal diffusion kernel function. As shown below: Specifically, in this embodiment, the diffusion scale parameter This is used to control the propagation range of structural information in a cascaded network. By calculating the thermal diffusion response matrix at different diffusion scales and extracting multi-scale features from the results, the local structural feature representation of the user in the cascaded propagation structure is obtained. The aggregation transformation result is used to extract the user's information. The cascaded local structure embedding representation is shown below: ,in, As an embedding dimension, specifically in this embodiment, the cascaded structure embedding vector is used to characterize the local propagation position and structural role features of the user in the specific information dissemination process, and serves as the input for subsequent continuous time feature encoding and multi-source feature fusion steps.

[0030] S13: Normalize the timestamps of user participation in dissemination, and use a continuous time coding function to map the normalized timestamps into a time feature embedding representation containing linear trend components and periodic components. Specifically, in this embodiment, a continuous time feature coding method is introduced to perform high-dimensional mapping on the normalized timestamps. By mapping the time information into a vector representation that simultaneously contains linear change components and periodic change components, the dynamic characteristics of information dissemination evolving over time can be characterized.

[0031] More ideally, for each forwarding or interaction during the cascading propagation of information, obtain the corresponding timestamp information, and set the user... The timestamp for participating in information dissemination is The observation deadline is To eliminate differences in time scales across different cascades and to unify the time reference standard, timestamps are normalized to obtain a normalized time representation: ,in, This represents the maximum timestamp of the cascade within the observation window, used to ensure the numerical stability of the time values.

[0032] Preferably, in step S13, the normalized timestamp is mapped to a time feature embedding representation containing linear trend components and periodic components using a continuous-time encoding function, further including: Normalized timestamps The input continuous-time encoding function Time2Vec(·) is mapped to a (k+1)-dimensional embedding vector, as shown below: ,in, , , ... and , , ... These are the learnable parameters during model training.

[0033] More preferably, the first dimension is used to represent the linear trend characteristics over time, while the remaining dimensions are used to represent the periodic changes over time. This allows users to... The temporal feature embedding in the information dissemination process is represented as follows: Continuous temporal feature embedding vectors are used to characterize the temporal dynamics of user participation in information dissemination. They can reflect non-uniform and sudden temporal changes during the dissemination process and serve as an important input for the time dimension in subsequent multi-source feature fusion steps.

[0034] S14: Concatenate the global structure embedding representation, cascaded local structure embedding representation, and time feature embedding representation of the same user to obtain the user's fused features, and sort the fused features of all users by timestamp to form a fused feature sequence.

[0035] Even better, for any user participating in information dissemination Its corresponding global network structure embedding representation is as follows The local structure embedding of a cascaded network is represented as Continuous-time feature embedding is represented as The three types of embedding vectors are concatenated in a preset order to construct the user's fused feature representation vector, which is represented as follows: in," "" indicates a vector concatenation operation, through which a fused feature representation is constructed. It simultaneously includes the user's structural position information in the global social network, their local propagation role information in specific information cascades, and the temporal dynamic characteristics of the user's participation in propagation, thus forming a comprehensive description of the user's propagation behavior. Furthermore, according to the temporal order of information propagation, the fused feature representations of users participating in cascade propagation are sorted to construct a propagation sequence representation: in, It represents the temporal order of user participation in information dissemination. The dissemination sequence is a unified sequence of user dissemination behavior and is input into subsequent time series modeling and prediction steps to capture the temporal dependencies and evolutionary patterns in the process of information cascading dissemination.

[0036] S2: Construct a cascaded propagation graph based on the propagation data, and dynamically model the evolution of users' viewpoint states along the path of the cascaded propagation graph to generate users' propagation biases. Specifically, in this embodiment, the Degroot viewpoint evolution model is introduced to dynamically model the changes in users' viewpoints during the information propagation process, obtain the user's viewpoint value in the cascade, and further construct propagation biases based on the differences in viewpoints among users to characterize the impact of cognitive similarity on propagation decisions.

[0037] Preferably, in step S2, the user's viewpoint state is dynamically evolved along the path of the cascading propagation graph to generate the user's propagation bias, including: S21: Set the initial viewpoint state of the cascade source node to the first value, and set the initial viewpoint state of the inactive node to the second value. Specifically, in this embodiment, for a certain information cascade, before a given observation time point, construct a corresponding cascade propagation directed graph according to the actual propagation order of the information. The node represents the user participating in the propagation, and the edge represents the propagation relationship from the source user to the user being propagated. At the start of cascade propagation, set the initial viewpoint state of the users. For the source node of the information cascade, its initial viewpoint state is set to 1; for users who have not yet participated in the propagation or have not yet been affected by the information, their initial viewpoint state is set to 0, thereby depicting the situation where the information is only fully accepted by the source node in the early stage of propagation, and the other users have not yet formed a clear cognitive state.

[0038] S22: Based on the actual time sequence of information propagation, when a user node is activated, the viewpoint state of the activated node is updated according to the current viewpoint state of all its activated upstream neighbor nodes that are the direct source of propagation. Specifically, in this embodiment, at the start of cascading propagation, the initial viewpoint state of the user is set. For the source node of the information cascading, its initial viewpoint state is set to 1; for users who have not yet participated in the propagation or have not yet been affected by the information, their initial viewpoint state is set to 0, thereby depicting the situation where the information is only fully accepted by the source node in the early stage of propagation, and the other users have not yet formed a clear cognitive state.

[0039] Preferably, in step S22, updating the view state of the activated node further includes: Calculate the activated node viewpoint status , is the activated node All activated propagation upstream neighbor nodes viewpoint status The weighted average is shown below: ,in For users For users The influence weighting coefficient is used to characterize the relative influence of different sources of information during the process of updating viewpoints. For users In the The state of opinion following the secondary dissemination event. This indicates the impact on users during this communication event. The set of upstream users that have an impact, and .

[0040] Through the above update method, the user's opinion state is continuously influenced by its propagation source as cascading propagation events occur, ensuring that the opinion evolution process remains consistent with the actual propagation path and time sequence, thus forming a user opinion representation that reflects the dynamic characteristics of information propagation. After completing the opinion updates for all propagation events before a given observation time point, the final opinion state of each user in the cascade can be obtained.

[0041] S23: After completing the update of the viewpoint state corresponding to all propagation events before the observation point, use the final viewpoint state value of each user as its corresponding propagation bias.

[0042] More preferably, each user's final opinion state value serves as their corresponding propagation bias, including: A propagation bias is constructed based on the viewpoint states among users to characterize the impact of cognitive similarity on information propagation behavior. For any two users in the cascade... and Their differences in opinion are defined as follows: Furthermore, the aforementioned differences in viewpoints are mapped to propagation bias terms: Among them, propagation bias This is used to reflect the moderating effect of the similarity of opinions among users on communication behavior. The smaller the difference in opinions, the greater the corresponding communication bias, indicating that information is more likely to spread continuously among users with similar cognitive states.

[0043] In the subsequent time series modeling and prediction process, the aforementioned propagation bias is used as an adjustment factor for the intensity of information interaction between nodes and introduced into the time series modeling module of the cascaded propagation sequence. This enables the model to comprehensively consider the propagation structure, time evolution, and changes in user cognitive state when learning the dynamic characteristics of propagation, thereby improving the accuracy and stability of the information cascade prediction results.

[0044] S3: Construct and train a cascaded prediction model. The cascaded prediction model is configured to encode the fused feature sequence and introduce a propagation bias during the encoding process to output a predicted value for the future propagation scale of the target information. Specifically, in this embodiment, the user propagation behavior representation sequence is used. The constructed propagation bias term is simultaneously input into the parallel-configured Transformer coding layer and the bidirectional gated recurrent unit network (BiGRU) for processing.

[0045] Preferably, in step S3, constructing and training the cascaded prediction model includes: S31: Construct a parallel dual-channel temporal coding architecture. The first channel of the parallel dual-channel is a global dependency modeling channel based on a Transformer encoder, and the second channel of the parallel dual-channel is a local dependency modeling channel based on a bidirectional gated recurrent unit network (BiGRU) to capture global long-term dependencies and local short-term dependencies in cascading propagation.

[0046] More preferably, the Transformer encoding layer models each user representation in the propagation sequence through a self-attention mechanism to capture the global dependencies between different propagation nodes, and its output representation is as follows: in Presentation layer normalization operation, , and Representing query, key, and value respectively, in this embodiment, the self-attention mechanism uses a user embedding matrix. As input, generate corresponding values ​​through different linear transformations. , and and output Transformer representation t represents the number of encoder layers. BiGRU consists of two GRUs, forward and backward, and can simultaneously capture bidirectional dependencies in a sequence.

[0047] S32: The propagation bias is introduced as a dynamic adjustment term into the gating calculation process of the second channel to modulate information updates. Specifically, in this embodiment, considering that the degree of information acceptance by nodes during propagation is different, their information update behavior should be adjusted by the viewpoint state. In this embodiment, the viewpoint bias is introduced into the gating calculation of BiGRU as a gating bias. Through this design, the viewpoint bias only acts on the gating calculation process to adjust the information update ratio between the historical state and the current input, without changing the basic state update structure of GRU.

[0048] Preferably, in step S32, the gating calculation process for the second channel further includes: For users Its update gate in the bidirectional gated loop unit and reset door The calculation method is as follows: in, For users The fusion characteristics The state was hidden in the previous moment. For users The propagation bias, A learnable scaling factor used to control the degree to which viewpoint bias affects gating behavior. It is the sigmoid activation function. , , , These are the network weight parameters.

[0049] Based on the update rules of GRU units, BiGRU models the cascaded sequence in both forward and backward directions to capture the temporal dependencies of nodes during propagation. The hidden states obtained in both directions are updated at time steps. The nodes are concatenated to form the final BiGRU representation, and its output is as follows:

[0050] S33: Input the fused feature sequence into the first and second channels simultaneously to obtain the global temporal representation and the local temporal representation respectively. Then, adaptively fuse the global temporal representation and the local temporal representation through learnable weight parameters to obtain the fused temporal representation.

[0051] Preferably, in step S33, the adaptive fusion of the global temporal representation and the local temporal representation using learnable weight parameters further includes: Adaptive fusion uses learnable weight parameters accomplish: in, This is the global timing representation of the first channel output. This is the local timing representation of the second channel output. To integrate temporal representations, The fusion weights are learnable parameters used to adjust the relative contribution ratio of global and local features in the final representation. The fusion weights can be automatically determined by a learnable function.

[0052] Obtaining the fused temporal feature representation Subsequently, in order to map the variable-length propagation sequence to a fixed-dimensional concatenated representation, a summation pooling operation is performed on the temporal features along the time dimension.

[0053] S34: Perform summation pooling on the fused temporal representation to obtain a fixed-dimensional cascaded global state vector. Specifically, in this embodiment, the feature vectors corresponding to each propagation node are accumulated to obtain the cascaded overall representation vector, which is represented as follows: ,in, Indicates the first The fusion temporal feature representation corresponding to each propagation node. The number of nodes participating in the cascading propagation of information.

[0054] S35: Input the cascaded global state vector Z into the multilayer perceptron (MLP) and output the propagation scale prediction value. Specifically, in this embodiment, the MLP prediction model estimates the propagation scale of information within the prediction time window, and the prediction process is expressed as follows: in, This is a prediction of the scale of information cascading over a future time period.

[0055] S36: Using the mean squared logarithmic error as the loss function, perform end-to-end training on all parameters of the model. Specifically, in this embodiment, the loss function is defined as: ,in, Indicates the number of training samples. For the first The predicted propagation scale corresponding to each sample The '+1' operation, representing the true propagation scale of the sample, serves as a smoothing term to avoid the undefined logarithmic operation when the predicted or true value is 0. By introducing a logarithmic transformation between the predicted and true values, the loss function effectively mitigates the numerical instability caused by the long-tailed distribution of information cascade propagation scale, making the training process of the model more stable on cascade samples of different scales. By minimizing the loss function, the model parameters are iteratively updated, thereby completing the training of the information cascade prediction model. Through the above parallel temporal modeling, feature fusion, and prediction process, this embodiment can achieve accurate prediction of the future propagation scale of information cascade based on a full characterization of global propagation trends and local dynamic changes.

[0056] S4: Using the trained cascade prediction model, predict the propagation scale of new information cascades.

[0057] Preferably, step S4 further includes: S41: Obtain new propagation data of the new information to be predicted cascaded before the observation time point; S42: Based on the new propagation data, extract the new cascade fusion feature sequence of this new cascade; S43: Construct a new cascading propagation graph based on the new propagation data, and generate a new propagation bias corresponding to the user; S44: The cascaded prediction model trained with the new cascaded fusion feature sequence and the new propagation bias input directly obtains the predicted value of its future propagation scale.

[0058] This embodiment compares the performance of the above algorithm using two real-world network datasets widely used in the field of community detection. Table 1 shows the specific information of the two real-world datasets.

[0059] (1) Twitter: This dataset comes from the X platform and contains publicly available English tweets and their dissemination information published between March 24 and April 25, 2012. In the experiment, the observation time... The predicted time is set to 1 day and 2 days respectively. A uniform 32-day period is set to predict the final propagation scale of information cascading.

[0060] (2) Weibo: This dataset comes from Sina Weibo, one of the most representative microblogging social platforms in China. It includes the original Weibo post published on June 1, 2016, and all repost records within the following 24 hours. In the experiment, the observation time... Set to 0.5 hours and 1 hour, predict time The timeframe was set to 24 hours to evaluate the model's performance in short-term information diffusion prediction tasks.

[0061] Table 1: Detailed information of the two real datasets For each dataset, it was randomly divided into training, validation, and test sets at a ratio of 70%, 15%, and 15%, respectively. Detailed statistical information for the three datasets is summarized in Table 2.

[0062] Table 2: Statistical information of datasets at different observation times The model's performance is evaluated using two commonly used metrics: mean squared logarithmic error and mean absolute percentage error, calculated as follows: in, The total number of cascades, For the first The actual scale of cascading propagation, This is the corresponding predicted value. The "" in the formula The operation is used to avoid logarithmic operations or zero values ​​in the denominator. For both of these metrics, the lower the value, the better the model's predictive performance.

[0063] Table 3: Performance Comparison of Different Algorithms As shown in Table 3, the experimental results presented in this paper demonstrate that the proposed method achieves the lowest MSLE and MAPE values ​​across both datasets and observation settings, exhibiting a stable and consistent advantage. This result clearly demonstrates that the proposed method can more effectively characterize the complex interaction between temporal dynamics and structural features in information cascading, thereby significantly improving prediction accuracy. On the Twitter dataset, the significant temporal heterogeneity and burstiness of the information diffusion process place higher demands on the model's temporal modeling capabilities. Under the most challenging 1-day observation window, the proposed method achieves a performance improvement of up to 14.84% in MSLE compared to Casformer, showing a particularly significant advantage. This result indicates that the continuous-time encoding mechanism (Time2Vec) has a significant advantage in capturing fine-grained temporal signals and can effectively compensate for the information loss in the early diffusion stage of methods based on discrete time windows or methods with limited temporal modeling capabilities (such as CasDO). Although the diffusion pattern on the Weibo platform is constrained to some extent by platform mechanisms and user behavior patterns, the proposed method still maintains a significant performance advantage. Under a short observation setting of 0.5 hours, the proposed method achieved varying degrees of reduction in MSLE and MAPE compared to the optimal baseline model. This indicates that even in scenarios with relatively limited behavioral patterns and more regular diffusion rhythms, the proposed method can still effectively integrate local temporal burst features with the overall network propagation structure, thereby improving the robustness and accuracy of predictions.

[0064] Table 4: Ablation Experiment Structure Table To further verify the impact of various technical features on the prediction performance of information cascading, an ablation experiment was designed, and the results are shown in Table 4. When both Transformer and BiGRU (w / o TB) were removed simultaneously, the model performance significantly decreased on all three datasets, especially in the short-term prediction tasks of Twitter and Weibo, where the error increased significantly. This indicates that relying solely on structural features and cognitive factors is insufficient to effectively capture the complex relationships between nodes during information propagation. Replacing Time2Vec with positional encoding (w / o T2V) also led to a significant degradation in prediction performance, particularly on the Weibo and APS datasets. This phenomenon suggests that, compared to positional encoding which only reflects sequential information, the continuous-time representation of Time2vec is more conducive to capturing the non-uniform temporal characteristics in cascading propagation, especially in scenarios with significant differences in propagation rhythm or long time spans. After removing the Degroot viewpoint evolution model suitable for information cascading (w / o CDg), the model performance decreased across all datasets and prediction windows. This result demonstrates that relying solely on propagation structure and temporal information is insufficient to fully characterize the intrinsic evolutionary dynamics of cascading events. Introducing a cognitive evolutionary mechanism helps the model depict the differentiated behavioral patterns of nodes during propagation, thereby improving the model's ability to predict the scale of subsequent information diffusion. When the original form of the DeGroot mechanism (Dg) is used instead of the improved version designed in this paper, the model performance significantly decreases. This indicates that directly using the traditional DeGroot update rule is difficult to adapt to complex information cascading scenarios. This paper, by combining it with the cascading propagation process, introduces viewpoint evolution as a bias into temporal modeling, enabling it to more effectively serve the prediction objective. The results show that by incorporating a cognitive evolutionary mechanism, the model exhibits stronger robustness and prediction accuracy across different propagation environments and time scales. Ablation experiments fully validate the rationality of the module design and further illustrate that incorporating a cognitive evolutionary mechanism provides an effective modeling approach for cascading prediction.

[0065] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for predicting cascading information in social networks based on user cognition, characterized in that, Includes the following steps: S1: Obtain the propagation data of the target information before the observation time point, including user identifier, propagation relationship between users and propagation timestamp; extract user fusion features based on the propagation data, and arrange them in the propagation time sequence to form a fusion feature sequence. S2: Construct a cascaded propagation graph based on the propagation data, and dynamically model the evolution of the user's viewpoint state along the path of the cascaded propagation graph to generate the user's propagation bias; S3: Construct and train a cascaded prediction model, which is configured to encode the fused feature sequence and introduce the propagation bias during the encoding process to output a predicted value of the future propagation scale of the target information; S4: Using the cascade prediction model after training, predict the propagation scale of the new information cascade.

2. The social network information cascading prediction method based on user cognition according to claim 1, characterized in that, In step S1, extracting user fusion features based on the propagation data further includes: S11: Construct a user sparse adjacency matrix based on the global social relationship network, and normalize and decompose the user sparse adjacency matrix to obtain the user global structure embedding representation; S12: Construct an information cascade subgraph based on the propagation data, and use a multi-scale analysis method based on thermal diffusion kernels to extract the cascaded local structure embedding representation of the user; S13: Normalize the timestamps of user participation in dissemination, and use a continuous time coding function to map the normalized timestamps into a time feature embedding representation containing linear trend components and periodic components. S14: The global structure embedding representation, the cascaded local structure embedding representation, and the time feature embedding representation of the same user are concatenated to obtain the user's fusion feature, and the fusion features of all users are sorted by timestamp to form the fusion feature sequence.

3. The social network information cascading prediction method based on user cognition according to claim 2, characterized in that, In step S13, the normalized timestamp is mapped to a time feature embedding representation containing linear trend components and periodic components using a continuous-time encoding function, further including: Normalized timestamps The input continuous-time encoding function Time2Vec(·) is mapped to a (k+1)-dimensional embedding vector, as shown below: in, , , ... and , , ... These are the learnable parameters during model training.

4. The social network information cascading prediction method based on user cognition according to claim 2, characterized in that, In step S12, based on the multi-scale analysis method of thermally diffusing kernels, the user's cascaded local structure embedding representation is extracted, including: Construct a cascaded adjacency matrix based on the aforementioned information cascaded subgraph. and its degree matrix According to the cascaded adjacency matrix and the degree matrix Calculate the normalized Laplacian matrix As shown below: in, It is the identity matrix; Through thermal diffusion kernel functions at different diffusion scales The transformation result is obtained by transforming the cascaded graph structure, wherein the thermal diffusion kernel function is... As shown below: By aggregating the transformation results, the user's cascaded local structure embedding representation is extracted, as shown below: in, For the embedded dimension.

5. The social network information cascading prediction method based on user cognition according to claim 1, characterized in that, In step S2, the dynamic evolution modeling of the user's viewpoint state along the path of the cascading propagation graph to generate the user's propagation bias includes: S21: Set the initial view state of the cascaded source node to the first value, and set the initial view state of the inactive node to the second value. S22: Based on the actual time sequence of information propagation, when a user node is activated, the view status of the activated node is updated according to the current view status of all its activated upstream neighbor nodes that are the direct source of propagation. S23: After completing the update of the viewpoint state corresponding to all propagation events before the observation point, use the final viewpoint state value of each user as its corresponding propagation bias.

6. The social network information cascading prediction method based on user cognition according to claim 5, characterized in that, In step S22, updating the view state of the activated node further includes: Calculate the activated node viewpoint status , is the activated node All activated propagation upstream neighbor nodes viewpoint status The weighted average is shown below: in For users For users Influence weighting coefficient This indicates the impact on users during this communication event. The set of upstream users that have an impact.

7. The social network information cascading prediction method based on user cognition according to claim 1, characterized in that, In step S3, the cascaded prediction model is constructed and trained, including: S31: Construct a parallel dual-channel temporal coding architecture, wherein the first channel of the parallel dual-channel is a global dependency modeling channel based on a Transformer encoder, and the second channel of the parallel dual-channel is a local dependency modeling channel based on a bidirectional gated recurrent unit network. S32: The propagation bias is used as a dynamic adjustment term and introduced into the gating calculation process of the second channel to update the modulation information; S33: Input the fused feature sequence into the first channel and the second channel simultaneously to obtain the global temporal representation and the local temporal representation respectively, and then adaptively fuse the global temporal representation and the local temporal representation through learnable weight parameters to obtain the fused temporal representation; S34: Perform summation and pooling on the fused temporal representation to obtain a cascaded global state vector of fixed dimensions; S35: Input the cascaded global state vector into the multilayer perceptron and output the propagation scale prediction value; S36: Use the mean square logarithmic error as the loss function to perform end-to-end training on all parameters of the model.

8. The social network information cascading prediction method based on user cognition according to claim 7, characterized in that, In step S32, the gating calculation process for the second channel further includes: For users Its update gate in the bidirectional gated loop unit and reset door The calculation method is as follows: in, For users The fusion characteristics The state was hidden in the previous moment. For users The propagation bias, For learnable scaling factors, It is the sigmoid activation function. , , , These are the network weight parameters.

9. The social network information cascading prediction method based on user cognition according to claim 7, characterized in that, In step S33, the global temporal representation and the local temporal representation are adaptively fused using learnable weight parameters, further including: The adaptive fusion uses learnable weight parameters. accomplish: in, This is the global timing representation of the first channel output. This is the local timing representation of the second channel output. To integrate temporal representations, These are learnable fusion weight parameters.

10. The social network information cascading prediction method based on user cognition according to claim 1, characterized in that, Step S4 further includes: S41: Obtain new propagation data of the new information to be predicted cascaded before the observation time point; S42: Based on the new propagation data, extract the new cascade fusion feature sequence of the new cascade; S43: Construct a new cascaded propagation graph based on the new propagation data, and generate a new propagation bias corresponding to the user; S44: The cascaded prediction model trained with the new cascaded fusion feature sequence and the new propagation bias input is directly used to obtain the predicted value of its future propagation scale.

Images