An Information Diffusion Prediction Method Based on Joint Modeling of Static and Dynamic Relationships

Abstract

This invention discloses an information diffusion prediction method based on joint modeling of static and dynamic relationships, belonging to the field of social network information diffusion prediction technology. The method includes: S1, constructing a static user relationship graph of the social network; S2, obtaining the propagation event sequence of the target information; S3, employing multiple sets of temporal feature extraction units with different temporal receptive fields; and S4, performing joint feature fusion of the static feature representation and the multi-scale temporal feature representation. By setting a multi-scale temporal feature extraction mechanism, this invention can extract local dynamic changes and cross-stage evolution trends during the information diffusion process, enhancing the model's ability to characterize complex propagation time patterns and overcoming the shortcomings of existing methods in multi-timescale joint modeling. By fusing static relationship features with dynamic propagation features, a unified expression of network structure information, node interaction information, and temporal evolution information is achieved.

Description

Technical Field

[0001] This invention belongs to the field of social network information diffusion prediction technology, specifically involving an information diffusion prediction method based on joint modeling of static and dynamic relationships. Background Technology

[0002] With the rapid development of internet technology, social networking platforms such as Weibo, Twitter, Douyin, and WeChat have become core carriers of information production, dissemination, and fermentation. Massive amounts of information spread rapidly through social connections between users, profoundly influencing public opinion, the trajectory of social events, and business decisions. Information diffusion prediction, which involves modeling and analyzing historical dissemination data to accurately predict the future scale, path, potential users, peak time, and evolution trend of information, is a core supporting technology for scenarios such as social network governance, public opinion emergency response, and targeted commercial advertising. It has significant theoretical research and engineering application value.

[0003] Information diffusion in social networks has four core characteristics that determine the technical challenges of prediction tasks: First, topological dependence, where information dissemination relies on long-term, stable social topological relationships between users (such as following, friends, and subscription relationships), which is the basic physical carrier of information diffusion; second, strong dynamic time-varying nature, where the propagation process exhibits significant stage characteristics, including minute-level bursts, hour-level stage fluctuations, and day-level continuous decay and eventual disappearance, with significant differences in propagation speed, participating node scale, and interaction intensity at different stages; and third, multi-scale temporal sequence, where propagation behavior simultaneously includes short-term bursts, medium-term evolution, and long-term trend changes, and a single time scale cannot fully characterize the propagation pattern.

[0004] Most existing methods focus on modeling static structural relationships or only represent the propagation process at a single time scale, resulting in insufficient extraction of multi-scale dynamic features in the diffusion process, which in turn affects the accuracy and stability of the prediction results. Summary of the Invention

[0005] The purpose of this invention is to provide an information diffusion prediction method based on joint modeling of static and dynamic relationships, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an information diffusion prediction method based on joint modeling of static and dynamic relationships, comprising the following steps: S1. Construct a static user relationship graph of the social network, and encode the static user relationship graph to obtain the static feature representation of each user node; S2. Obtain the propagation event sequence of the target information, and construct a dynamic propagation relationship representation of the corresponding information propagation process based on the temporal information of the propagation event sequence; S3. Multiple sets of temporal feature extraction units with different temporal receptive fields are used to extract features in parallel on the dynamic propagation relationship representation. The temporal features extracted from each set are fused to obtain a multi-scale temporal feature representation. S4. Perform joint feature fusion between the static feature representation and the multi-scale temporal feature representation to obtain a comprehensive feature representation, and output the diffusion prediction result of the target information based on the comprehensive feature representation; S5. Train and optimize the model based on historical propagation datasets from social networks, and deploy the trained model to real-world scenarios to predict the spread of information in unknown propagation events.

[0007] In a preferred embodiment, step S1 specifically includes: S11. Map all users in the social network to a set of nodes, and map the stable social relationships between users to a set of edges, constructing an undirected or directed static user relationship graph; the stable social relationships include friend relationships, follow relationships, subscription relationships, or historical stable interaction relationships. S12. The adjacency matrix of the user static relationship graph is normalized, and the multi-level neighborhood information of the nodes is aggregated through a multi-layer graph neural network to generate a static feature representation corresponding to each user node; the static feature representation is used to characterize the user's topological position, neighborhood structure features and the strength of association between nodes in the social network.

[0008] In a preferred implementation, step S2, which involves constructing a dynamic propagation relationship representation based on higher-order interaction relationships, specifically includes: S211: The propagation event sequence is time-sliced ​​according to a preset time step to obtain multiple consecutive time slices, each time slice corresponding to a stage of information propagation; S212: Construct a high-order relation unit from the set of user nodes participating in the propagation of target information within the same time slice, wherein the high-order relation unit corresponds to a hyperedge structure; S213: The node features within each higher-order relation unit are aggregated using an attention aggregation mechanism to generate a dynamic propagation relation representation for the corresponding time slice. The aggregation calculation formula is as follows:

[0009] in Let represent the j-th superedge at time T. Indicates user In terms of the characteristics of time T, For learnable weight matrix, This represents the attention weight.

[0010] In a preferred embodiment, step S2, which involves constructing a dynamic propagation relationship representation based on time-aware relationships, specifically includes: S221: Extract the timestamps of the interaction behaviors of each node in the propagation event sequence, and calculate the time difference of the interaction behaviors between nodes; S222: Encode the time difference using a time coding function to generate a time-aware coding vector; the time coding function adopts a periodic function mapping method or a learnable embedding mapping method; S223: Combine node features with the time-aware encoding vector to calculate the time-aware association weights between nodes, and aggregate the features of neighboring nodes based on the time-aware association weights to generate a time-aware dynamic propagation relationship representation.

[0011] In a preferred embodiment, the time encoding function using the periodic function mapping method in step S222 is:

[0012] In the formula, This is the time-aware encoding vector corresponding to the time difference t. For the frequency coefficients of the k-th dimension, t represents the dimension of the time-aware encoding vector, and k is the dimension index, with k ranging from 1 to... Positive integers.

[0013] In a preferred embodiment, step S3 specifically includes: S31. Set up multiple sets of parallel one-dimensional convolutional units as temporal feature extraction units. Each set of one-dimensional convolutional units uses convolutional kernels of different sizes. Different sizes of convolutional kernels correspond to different temporal receptive fields. Among them, small-sized convolutional kernels are used to capture local fluctuations and short-term burst features in the propagation process, while large-sized convolutional kernels are used to capture cross-stage evolution and long-term trend features in the propagation process. S32. The dynamic propagation relationship is sequentially input into each group of parallel one-dimensional convolutional units, and after nonlinear activation, the single-scale temporal features corresponding to the time scale are output. S33. By using splicing, summation, gating fusion or weighted fusion methods, all single-scale temporal features are integrated to generate a unified multi-scale temporal feature representation.

[0014] In a preferred embodiment, step S4 specifically includes: S41. Perform linear transformations on the static feature representation and the multi-scale temporal feature representation respectively to unify the vector dimension and representation space of the two types of features; S42. Employ attention fusion mechanism, gating fusion mechanism, weighted fusion mechanism or splicing mapping method to jointly model the static feature representation after dimension unification and the multi-scale temporal feature representation to generate a comprehensive feature representation. S43. Input the comprehensive feature representation into the fully connected layer and the probability mapping layer, and output the activation probability and propagation participation probability of the target user node in the subsequent propagation stage, or the propagation scale and propagation path prediction results of the target information within a preset time range, to complete the information diffusion prediction.

[0015] As a preferred embodiment, step S5 specifically includes: S51: Obtain the historical propagation event dataset of the social network, preprocess the dataset, and divide it into training set, validation set and test set; S52: Using the preceding propagation sequence of historical propagation events as the model input and the subsequent real propagation results of the corresponding propagation events as the supervision label, a cross-entropy loss function is constructed as the objective function for model optimization. S53: The gradient descent algorithm is used to iteratively optimize the learnable parameters of the model. The model hyperparameters are adjusted through the validation set until the model's performance metrics on the validation set converge, thus completing the model training.

[0016] In a preferred embodiment, the graph neural network used in step S12 is a graph convolutional network or a graph attention network.

[0017] In a preferred embodiment, in step S222, the learnable embedding mapping method is to use a multilayer perceptron to map the timestamp to a high-dimensional vector space to generate a time-aware encoding vector.

[0018] Compared with the prior art, the beneficial effects of the present invention are: This invention, by setting up a multi-scale temporal feature extraction mechanism, can extract local dynamic changes and cross-stage evolution trends in the information diffusion process, thereby enhancing the model's ability to characterize complex propagation time patterns and making up for the shortcomings of existing methods in multi-timescale joint modeling.

[0019] This invention integrates static relationship features with dynamic propagation features to achieve a unified expression of network structure information, node interaction information, and temporal evolution information. This helps to improve the ability to identify propagation paths and node activation behaviors in complex social networks, thereby enhancing the reliability of diffusion prediction results.

[0020] The multi-stage modeling and fusion approach adopted in this invention has good versatility and scalability. It can be adapted to different social platforms, different dissemination scenarios and different task objectives, and is suitable for application scenarios such as hot event dissemination prediction, public opinion evolution analysis, information dissemination risk identification, user activation prediction and social recommendation.

[0021] When faced with scenarios characterized by rapid changes in the propagation process, complex propagation relationships, and significant time dependence, this invention can more fully extract and utilize multidimensional feature information in the diffusion process, thereby improving the problems of insufficient dynamic modeling and weak integration of structural and temporal features in existing technologies, and has good application value. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the information diffusion prediction structure based on dynamic modeling of higher-order relations in this invention; Figure 2 This is a schematic diagram of the information diffusion prediction structure based on time-aware dynamic modeling of the present invention; Figure 3 This is a schematic diagram illustrating the steps of the information diffusion prediction method based on joint modeling of static and dynamic relationships according to the present invention. Detailed Implementation

[0023] The present invention will be further described below with reference to embodiments.

[0024] The following embodiments are used to illustrate the present invention, but should not be used to limit the scope of protection of the present invention. The conditions in the embodiments can be further adjusted according to specific conditions, and simple improvements to the method of the present invention under the premise of the concept of the present invention are all within the scope of protection claimed by the present invention.

[0025] Please see Figures 1-3 This invention provides an information diffusion prediction method based on joint modeling of static and dynamic relationships, comprising the following steps: Step S1: Construct a static user relationship graph of the social network, encode the static user relationship graph, and obtain the static feature representation corresponding to each user node; Step S2: Obtain the propagation event sequence of the target information, and construct a dynamic propagation relationship representation of the corresponding information propagation process based on the temporal information of the propagation event sequence; Step S3: Using multiple sets of temporal feature extraction units with different temporal receptive fields, parallel feature extraction is performed on the dynamic propagation relationship representation. The temporal features extracted from each set are then fused to obtain a multi-scale temporal feature representation. Step S4: Perform joint feature fusion between the static feature representation and the multi-scale temporal feature representation to obtain a comprehensive feature representation, and output the diffusion prediction result of the target information based on the comprehensive feature representation; Step S5: Train and optimize the model based on the historical propagation dataset of social networks, and deploy the trained model to the real-world scenario to predict the spread of information for unknown propagation events.

[0026] Further, step S1 specifically includes: S11: Map all users in the social network to a set of nodes V, and map the stable social relationships between users to a set of edges E, constructing an undirected or directed static user relationship graph G=(V,E); The stable social relationships include, but are not limited to, friend relationships, two-way follow relationships, one-way follow relationships, subscription relationships, and historical stable interaction relationships. The stable interaction relationship is the interaction relationship between users whose interaction frequency exceeds a preset threshold within a preset period. S12: Perform symmetric normalization on the adjacency matrix A of the user static relationship graph to obtain the normalized adjacency matrix. The calculation formula is as follows:

[0027] In the formula, It is an identity matrix, used to add self-loops to each node to avoid the loss of node features; The degree matrix of the user static relationship graph. That is, the diagonal elements of the degree matrix are the degrees of the corresponding nodes; S13: Aggregate multi-level neighborhood information of nodes through a multi-layer graph neural network to generate static feature representations for each user node. ; The static features are used to characterize the user's topological location, neighborhood structure features, and inter-node correlation strength in the social network. The graph neural network includes, but is not limited to, graph convolutional network (GCN), graph attention network (GAT), and graph SAGE. When using a 2-layer graph convolutional network, the layer propagation formula is:

[0028] In the formula, The initial feature matrix for the nodes can be formed using one-hot encoding of user IDs, user attribute features, or pre-trained node embeddings. , For different layers, there are learnable weight matrices; These are non-linear activation functions, including but not limited to ReLU, LeakyReLU, and GELU.

[0029] Furthermore, step S2 provides two dynamic propagation relationship representation construction methods that can be adapted to different application scenarios: a dynamic modeling method based on high-order interaction relationships and a dynamic modeling method based on time-aware relationships. The appropriate method can be flexibly selected based on the characteristics of the propagation scenario and the size of the dataset.

[0030] Among them, the dynamic modeling method based on high-order interaction relationships is suitable for scenarios involving the spread of hot topics, public opinion, and large-scale public events with significant group propagation characteristics, specifically including: S211: The propagation event sequence is sliced ​​according to a preset time step to obtain multiple consecutive time slices, each time slice corresponding to a stage of information propagation; the time step can be a fixed step (such as 5 minutes, 10 minutes, 1 hour, 1 day) or an adaptive step; the adaptive step is dynamically adjusted according to the propagation speed, with a smaller step in the propagation burst stage (when the rate of increase in forwarding volume per unit time exceeds a preset threshold) and a larger step in the propagation steady stage, so as to achieve precise division of propagation stages; S212: Construct a high-order relation unit from the set of user nodes participating in the propagation of target information within the same time slice. The high-order relation unit corresponds to the hyperedge in the hypergraph structure. All high-order relation units within the same time slice constitute the dynamic hypergraph of that time slice, thereby realizing the representation of multi-node group propagation behavior. S213: The node features within each higher-order relation unit are aggregated using an attention aggregation mechanism to generate a dynamic propagation relation representation for the corresponding time slice. The aggregation calculation formula is as follows:

[0031] In the formula, For the first Within the time frame, the first Dynamic representation of a high-order relational unit It is a non-linear activation function. Higher-order relational unit The first One user node, For user nodes In higher-order relation units Internal attention weights The weight matrix is ​​a learnable matrix. For user nodes In the The initial node features of each time slice can be initialized using the static feature representation obtained in step S1. Among them, attention weight Calculated using the attention mechanism, the formula is as follows:

[0032] In the formula, For learnable attention vectors, For feature concatenation operations, LeakyReLU is a linear rectified activation function with leakage, used to solve the neuron death problem of the ReLU function; S214: Concatenate the dynamic propagation relationships of all time slices in chronological order to form a dimension of... The dynamic propagation sequence matrix of ×d, where d represents the number of time slices, and d represents the feature dimension, which serves as the input for subsequent multi-scale temporal feature extraction.

[0033] Among them, the dynamic modeling method based on time-aware relationships is suitable for scenarios with strong time-series dependencies and complex propagation paths, such as rumor dissemination, precision marketing, and prediction of niche information diffusion. Specifically, it includes: S221: Extract the timestamps of the interaction behaviors of each node in the propagation event sequence; S222: Encode the time difference \(\Deltat\) using a time coding function to generate a time-aware coding vector; the time coding function adopts a periodic function mapping method or a learnable embedding mapping method; The time coding function using the periodic function mapping method is:

[0034] In the formula, This is the time-aware encoding vector corresponding to the time difference t. For the frequency coefficients of the k-th dimension, t represents the dimension of the time-aware encoding vector, and k is the dimension index, with k ranging from 1 to... Positive integers.

[0035] When using a learnable embedding mapping method, the timestamp t is mapped to a high-dimensional vector space through a 1-3 layer multilayer perceptron (MLP) to generate a time-aware encoded vector; This approach can learn time coding patterns that adapt to different scenarios through data-driven learning, resulting in stronger representation capabilities and suitability for large-scale datasets. S223: Combining node features with the time-aware encoding vector, calculate the time-aware association weights between nodes, and aggregate the features of neighboring nodes based on the time-aware association weights to generate a time-aware dynamic propagation relationship representation; the time-aware association weights are dynamically adjusted with the interaction time difference, and the closer the interaction time, the higher the weight, which conforms to the time decay law of information propagation, can distinguish the degree of influence of neighboring nodes on the current node under different time intervals, and improve the ability to represent the propagation time-series dependency relationship; S224: Concatenate the dynamic node representations of different timestamps in chronological order to form a dynamic propagation sequence matrix, which serves as the input for subsequent multi-scale temporal feature extraction.

[0036] Furthermore, step S3 specifically includes: S31: Multiple sets of parallel one-dimensional convolutional units are set as temporal feature extraction units. Each set of one-dimensional convolutional units uses convolutional kernels of different sizes, and different sizes of convolutional kernels correspond to different temporal receptive fields. Among them, small-sized convolutional kernels (such as 2, 3, 5) are used to capture local fluctuations and short-term burst features in the propagation process, medium-sized convolutional kernels (such as 7, 10, 15) are used to capture the stage evolution features in the propagation process, and large-sized convolutional kernels (such as 20, 30, 50) are used to capture the long-term trend features across stages in the propagation process. S32: The dynamic propagation sequence matrix is ​​input into each group of parallel one-dimensional convolutional units. The output of each group of convolutional units is processed by a nonlinear activation function to output the single-scale temporal features corresponding to the time scale. All convolutional units use the same padding method to ensure that the length of the output sequence is consistent with the length of the input sequence and to avoid loss of temporal information. S33: Employing splicing, summation, gated fusion, or weighted fusion methods, all single-scale temporal features are integrated to generate a unified multi-scale temporal feature representation. ; When using a gated fusion approach, learnable gate weights adaptively adjust the contribution of features at different time scales. During the propagation burst phase, the weight of small-scale features is automatically increased, while during the propagation steady phase, the weight of large-scale features is automatically increased, further enhancing the model's adaptability to different propagation stages.

[0037] Further, step S4 specifically includes: S41: representing the static features respectively. With the multi-scale temporal feature representation A linear transformation is performed, and the vector dimensions and representation spaces of the two types of features are unified to the same dimension through a fully connected layer, so as to avoid fusion bias caused by feature dimension mismatch. S42: Employ attention fusion mechanism, gating fusion mechanism, weighted fusion mechanism or splicing mapping method to jointly model the static feature representation after dimension unification and the multi-scale temporal feature representation to generate a comprehensive feature representation Z; S43: Input the comprehensive feature representation Z into the fully connected layer and the output layer of the corresponding task, and output the diffusion prediction result of the target information; For the node activation prediction task, the output layer adopts the Softmax activation function to output the activation probability and propagation participation probability of the target user node within a subsequent preset time range, thereby realizing the prediction of potential propagation users. For the task of predicting the scale of dissemination, the output layer uses a linear activation function to output dissemination scale indicators such as the number of forwards, the number of participating users, and the number of reads of the target information within a preset time range, thereby predicting the scale of dissemination. For the task of predicting the dissemination path, the output layer uses a ranking layer to output the ranking score of the dissemination path between nodes, thereby predicting the future dissemination path of the information. For the task of warning of rumors, the output layer uses a sigmoid activation function to output the probability of rumors and the level of dissemination risk of the target information, thereby enabling early warning of rumors.

[0038] Furthermore, when the gating fusion mechanism is used in step S42, the following steps are specifically included: S421: Input the static feature representation after dimension unification and the multi-scale temporal feature representation into the gating unit, and calculate the dynamic gating weight coefficient through learnable mapping parameters. The value range of the dynamic gating weight coefficient is 0 to 1. S422: Based on the dynamic gating weight coefficients, the static feature representation and the multi-scale temporal feature representation are weighted and allocated respectively. The weighted sum is then used to generate a comprehensive feature representation. The corresponding calculation formula is as follows:

[0039] In the formula, This is the comprehensive feature representation of the output. For dynamic gating weight coefficients, This is a static feature representation after dimensional unification. This represents the multi-scale temporal feature representation after dimensional unification. This is the learnable weight matrix corresponding to the static features. This is the learnable weight matrix corresponding to multi-scale temporal features. The bias term is learnable, and the sigmoid is a sigmoid activation function used to map the input value to the 0-1 range.

[0040] Through the above-mentioned gating fusion mechanism, the model can adaptively adjust the contribution of static features and dynamic temporal features according to the dynamic changes in the propagation stage, which solves the defect that the fixed fusion method cannot adapt to the changes in the propagation stage and greatly improves the robustness of the model.

[0041] Further, step S5, model training and optimization, specifically includes: S51: Obtaining a historical propagation event dataset from the social network, preprocessing the dataset including invalid propagation cascade filtering, data deduplication, abnormal timestamp cleaning, and user relationship extraction; after preprocessing, dividing the dataset into training, validation, and test sets according to a preset ratio; S52: Using the preceding propagation sequence of historical propagation events as model input and the subsequent real propagation results of the corresponding propagation events as supervision labels, constructing corresponding target loss functions for different prediction tasks; wherein, the cross-entropy loss function is used for node activation prediction and rumor warning tasks, the mean squared error (MSE) loss function is used for propagation scale prediction tasks, and the mean squared error (MSE) loss function is used for propagation path prediction tasks. Pairwise sorting loss function; S53: Iteratively optimize the learnable parameters of the model using gradient descent algorithm, including but not limited to Adam, AdamW, and SGD optimizers; During training, early stopping strategy, L2 regularization, and Dropout mechanism are used to avoid model overfitting. When the loss value of the validation set does not decrease for a preset number of consecutive rounds, training is stopped and the optimal model parameters are saved; S54: The performance of the trained model is validated using a test set. The model performance is evaluated using metrics such as accuracy, F1 score, mean absolute error (MAE), and root mean square error (RMSE). After successful validation, the model is deployed to a real-world application scenario to predict the spread of information about unknown propagation events.

[0042] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An information diffusion prediction method based on joint modeling of static and dynamic relationships, characterized in that, Includes the following steps: S1. Construct a static user relationship graph of the social network, and encode the static user relationship graph to obtain the static feature representation of each user node; S2. Obtain the propagation event sequence of the target information, and construct a dynamic propagation relationship representation of the corresponding information propagation process based on the temporal information of the propagation event sequence; S3. Multiple sets of temporal feature extraction units with different temporal receptive fields are used to extract features in parallel on the dynamic propagation relationship representation. The temporal features extracted from each set are fused to obtain a multi-scale temporal feature representation. S4. Perform joint feature fusion between the static feature representation and the multi-scale temporal feature representation to obtain a comprehensive feature representation, and output the diffusion prediction result of the target information based on the comprehensive feature representation; S5. Train and optimize the model based on historical propagation datasets from social networks, and deploy the trained model to real-world scenarios to predict the spread of information in unknown propagation events.

2. The information diffusion prediction method based on joint modeling of static and dynamic relationships according to claim 1, characterized in that: Step S1 specifically includes: S11. Map all users in the social network to a set of nodes, and map the stable social relationships between users to a set of edges, constructing an undirected or directed static user relationship graph; the stable social relationships include friend relationships, follow relationships, subscription relationships, or historical stable interaction relationships. S12. The adjacency matrix of the user static relationship graph is normalized, and the multi-level neighborhood information of the nodes is aggregated through a multi-layer graph neural network to generate a static feature representation corresponding to each user node; the static feature representation is used to characterize the user's topological position, neighborhood structure features and the strength of association between nodes in the social network.

3. The information diffusion prediction method based on joint modeling of static and dynamic relationships according to claim 2, characterized in that: In step S2, constructing a dynamic propagation relationship representation based on higher-order interaction relationships specifically includes: S211: The propagation event sequence is time-sliced ​​according to a preset time step to obtain multiple consecutive time slices, each time slice corresponding to a stage of information propagation; S212: Construct a high-order relation unit from the set of user nodes participating in the propagation of target information within the same time slice, wherein the high-order relation unit corresponds to a hyperedge structure; S213: The node features within each higher-order relation unit are aggregated using an attention aggregation mechanism to generate a dynamic propagation relation representation for the corresponding time slice. The aggregation calculation formula is as follows: ； in Let represent the j-th superedge at time T. Indicates user In terms of the characteristics of time T, For learnable weight matrix, This represents the attention weight.

4. The information diffusion prediction method based on joint modeling of static and dynamic relationships according to claim 1, characterized in that: In step S2, constructing a dynamic propagation relationship representation based on time-aware relationships specifically includes: S221: Extract the timestamps of the interaction behaviors of each node in the propagation event sequence, and calculate the time difference of the interaction behaviors between nodes; S222: Encode the time difference using a time coding function to generate a time-aware coding vector; the time coding function adopts a periodic function mapping method or a learnable embedding mapping method; S223: Combine node features with the time-aware encoding vector to calculate the time-aware association weights between nodes, and aggregate the features of neighboring nodes based on the time-aware association weights to generate a time-aware dynamic propagation relationship representation.

5. The information diffusion prediction method based on joint modeling of static and dynamic relationships according to claim 4, characterized in that: In step S222, the time encoding function using the periodic function mapping method is: ； In the formula, This is the time-aware encoding vector corresponding to the time difference t. For the frequency coefficients of the k-th dimension, t represents the dimension of the time-aware encoding vector, and k is the dimension index, with k ranging from 1 to... Positive integers.

6. The information diffusion prediction method based on joint modeling of static and dynamic relationships according to claim 1, characterized in that: Step S3 specifically includes: S31. Set up multiple sets of parallel one-dimensional convolutional units as temporal feature extraction units. Each set of one-dimensional convolutional units uses convolutional kernels of different sizes. Different sizes of convolutional kernels correspond to different temporal receptive fields. Among them, small-sized convolutional kernels are used to capture local fluctuations and short-term burst features in the propagation process, while large-sized convolutional kernels are used to capture cross-stage evolution and long-term trend features in the propagation process. S32. The dynamic propagation relationship is sequentially input into each group of parallel one-dimensional convolutional units, and after nonlinear activation, the single-scale temporal features corresponding to the time scale are output. S33. By using splicing, summation, gating fusion or weighted fusion methods, all single-scale temporal features are integrated to generate a unified multi-scale temporal feature representation.

7. The information diffusion prediction method based on joint modeling of static and dynamic relationships according to claim 1, characterized in that: Step S4 specifically includes: S41. Perform linear transformations on the static feature representation and the multi-scale temporal feature representation respectively to unify the vector dimension and representation space of the two types of features; S42. Employ attention fusion mechanism, gating fusion mechanism, weighted fusion mechanism or splicing mapping method to jointly model the static feature representation after dimension unification and the multi-scale temporal feature representation to generate a comprehensive feature representation. S43. Input the comprehensive feature representation into the fully connected layer and the probability mapping layer, and output the activation probability and propagation participation probability of the target user node in the subsequent propagation stage, or the propagation scale and propagation path prediction results of the target information within a preset time range, to complete the information diffusion prediction.

8. The information diffusion prediction method based on joint modeling of static and dynamic relationships according to claim 1, characterized in that: Step S5 specifically includes: S51: Obtain the historical propagation event dataset of the social network, preprocess the dataset, and divide it into training set, validation set and test set; S52: Using the preceding propagation sequence of historical propagation events as the model input and the subsequent real propagation results of the corresponding propagation events as the supervision label, a cross-entropy loss function is constructed as the objective function for model optimization. S53: The gradient descent algorithm is used to iteratively optimize the learnable parameters of the model. The model hyperparameters are adjusted through the validation set until the model's performance metrics on the validation set converge, thus completing the model training.

9. The information diffusion prediction method based on joint modeling of static and dynamic relationships according to claim 2, characterized in that: The graph neural network used in step S12 is either a graph convolutional network or a graph attention network.

10. The information diffusion prediction method based on joint modeling of static and dynamic relationships according to claim 4, characterized in that: In step S222, the learnable embedding mapping method is to use a multilayer perceptron to map the timestamp to a high-dimensional vector space to generate a time-aware encoding vector.

Images