Public opinion field effect and heterogeneous hypergraph fused information diffusion prediction system and implementation method thereof

Abstract

The invention relates to the technical field of social network information spreading prediction, and discloses an information spreading prediction method fusing public opinion field effect and a heterogeneous hypergraph. In order to solve the problems that in the prior art, only pairwise user relations are relied on, multi-user group influences cannot be described, different information is subjected to cascade independent processing, and multi-topic competition is not considered, the invention provides a prediction scheme fusing public opinion field effects and heterogeneous hypergraph learning. A heterogeneous hypergraph is constructed to obtain user multivariate relation representation, then a public opinion field effect is utilized to quantify attraction energy of different information topics to a user, attention competition among multiple topics is modeled, and a more real user propagation tendency is obtained; and finally, realizing joint prediction of user interest features and social influence features through an interactive fusion mechanism. The method can be used for scenes of information propagation trend analysis, public opinion monitoring, marketing recommendation, false information early warning and the like.

Description

Information Diffusion Prediction System Integrating Public Opinion Field Effect and Heterogeneous Hypergraph and Its Implementation Method Technical Field

[0001] This invention belongs to the field of big data analysis technology, specifically involving data mining, artificial intelligence and graph neural network technology, and in particular an information dissemination prediction method based on the public opinion field effect. Background Technology

[0002] With the widespread adoption of online social media, the scale and speed of information dissemination among users have increased dramatically. Information diffusion prediction has become one of the core challenges in social network analysis, holding significant value in downstream applications such as viral marketing, misinformation detection, and public health monitoring. The goal of information diffusion prediction is to predict the sequence of users who will participate in information dissemination. However, existing technologies suffer from two key limitations: First, traditional graph-based methods rely on pairwise connections between users, failing to capture the high-order group interactions prevalent in real-world social systems. In actual social networks, multiple users often collectively influence an individual's information dissemination decisions, and simple pairwise graph structures struggle to express these complex group dependencies, resulting in an incomplete characterization of the social structure by the model.

[0003] Second, existing models treat each information propagation chain (cascade) as an independent process, ignoring the reality that multiple topics compete for users' limited attention simultaneously. Users' participation in a particular piece of information depends on its relative attractiveness, and competition between different topics significantly affects the information diffusion path. Modeling each cascade independently deviates from the actual propagation scenario, thus limiting the accuracy of predictions.

[0004] To address the aforementioned issues, there is an urgent need for an information diffusion prediction technology capable of simultaneously capturing high-level social structures and the dynamics of multi-topic competitive attention. Summary of the Invention

[0005] This invention aims to provide a method and system for predicting information diffusion that integrates the public opinion field effect and heterogeneous hypergraphs, in order to solve the technical defects of existing technologies that cannot effectively capture the dynamics of high-order group interactions and multi-topic competition for attention, and to improve the accuracy and reliability of information diffusion prediction.

[0006] Technical Solution: A method for predicting information diffusion by integrating public opinion field effects and heterogeneous hypergraphs, mainly including the following processes: Data Preprocessing: Acquiring social network data, which includes user sets, social relationships between users, message sets, and message diffusion sequences, wherein the message diffusion sequence includes users participating in message propagation and corresponding propagation times; Constructing a heterogeneous cascaded hypergraph based on the social network data, wherein the node set of the heterogeneous cascaded hypergraph is the union of the user set and the message set, and the edge set includes social relationship edges and propagation relationship edges; Social Influence Learning: Based on the heterogeneous cascaded hypergraph, extracting hypergraph snapshots through different meta-paths, and constructing the association matrix, node degree matrix, and hyperedge degree matrix of each snapshot; Calculating the symmetric normalized vertex propagation operator based on the association matrix, node degree matrix, and hyperedge degree matrix, and then constructing the hypergraph Laplacian matrix and performing spectral decomposition; Constructing a multi-scale hypergraph wavelet convolution operator based on the spectral decomposition results, and performing multi-layer propagation updates on user features through a hypergraph wavelet neural network, outputting a user embedding vector driven by social influence. Public opinion field learning: Each message node in the heterogeneous cascaded hypergraph is taken as a public opinion center, and the subgraph corresponding to each public opinion center constitutes a public opinion field; the initial features of different types of nodes are projected onto a shared latent space through a type-adaptive projection matrix; the energy value of each public opinion field is calculated, which is obtained by weighted summation of the popularity contribution coefficient and attention coefficient of the neighboring nodes of the public opinion center; the nodes are divided into four categories according to the type of their neighbors, and the corresponding attention mechanism is used to update the node features, outputting a user embedding vector driven by the public opinion field. and message embedding vector Fusion and Prediction: Fusion of social influence-driven user embedding vectors through a learnable gating mechanism User embedding vectors driven by public opinion field , to obtain the fused user embedding vector Construct a user-message affinity matrix and obtain the user attention distribution to messages through row normalization; aggregate message features based on the attention distribution, and obtain the interaction-aware user embedding vector through residual enhancement and layer normalization; calculate the probability of a user participating in the next round of message propagation through the softmax function, and output the prediction result after excluding users who have already participated in the propagation.

[0007] Based on the above-described method, this invention also provides an information diffusion prediction system that integrates public opinion field effect and heterogeneous hypergraph, including a data preprocessing module, a social influence learning module, a public opinion field learning module, a fusion prediction module, and a storage module. Each module corresponds to a step of the above-described method and works together to realize the information diffusion prediction function.

[0008] Furthermore, the study of social influence constructs a simple diffusion graph using data containing friendship relationships and information propagation processes among users, extracting snapshots based on different meta-paths. Each snapshot is then transformed into a hypergraph snapshot, containing nodes and hyperedges of the same type. An association matrix is ​​constructed based on each hypergraph snapshot. ,in It is the number of nodes. It is the number of hyperedges, defining the node degree matrix. and hypermarginality matrix Based on this, the symmetric normalized vertex propagation operator is defined as:

[0009] The corresponding hypergraph Laplacian operator is After spectral decomposition, the following was obtained , at scale Construct a graph wavelet kernel: .

[0010] The final inter-layer propagation rule is:

[0011] in It is obtained through random initialization. It is the ReLU activation function. It is a learnable weight matrix. After propagation through L layers, the user representation is finally obtained. , where d is the embedding dimension.

[0012] Furthermore, in studying the public opinion field effect, a simple diffusion diagram is used to construct the public opinion field, in which multiple topics coexist and compete for user attention. Each topic has a central node, thus defining the energy of the central node. This indicates the degree to which a topic is attractive to users. Neighboring nodes Able to give the central node The contribution to popularity Representing neighboring nodes For the central node Attention scores. Their calculation formulas are as follows:

[0013] in, Represents the attention mechanism of graphs. In social networks The impact, This represents the learnable attention weight matrix.

[0014] Nodes in the network can be divided into four categories: 1) Location-sensitive nodes (POCs); 2) User nodes connected only to user neighbors; 3) User nodes connected only to POC neighbors; and 4) User nodes connected to both user neighbors and POC neighbors. The representation learning process for each of these four cases is described below. The node representation update formulas for these four cases are as follows: Category 1 and Category 2:

[0015] Category 3:

[0016] Category 4:

[0017] in, and These are the attention scores for neighboring user nodes and neighboring center nodes, respectively. After processing all nodes according to the above rules, we obtain the final user representation. and final message representation .

[0018] Furthermore, feature fusion and prediction, following the user representation and information representation obtained in the above two steps, involves designing a collaborative attention mechanism to jointly model the bidirectional interaction between the user and the information. Specific steps include: firstly, fusing the two user embedding vectors through a learnable gating mechanism. and : ;in, They are all learnable linear transformation matrices.

[0019] Then calculate a user-information affinity matrix:

[0020] in These are all learnable transformation matrices. To obtain an interaction-aware user representation, each row of matrix A is normalized to obtain the attention weights of each user for all messages:

[0021] Subsequently, message information is aggregated through an attention mechanism to optimize user representation, followed by residual augmentation and LayerNorm for further processing.

[0022] in It is a trainable transformation matrix. The probability formula for calculating the user's participation in information dissemination is:

[0023] After obtaining the participation probabilities of all users in the propagation, sorting them together will yield the sequence of users most likely to participate in the propagation of information.

[0024] Beneficial effects: This invention can accurately target potential users for viral marketing, provide early intervention evidence for detecting false information, and provide technical support for information dissemination early warning in public health events, thus having broad application prospects. Its substantial features and significant advancements include: Strong ability to capture high-order structures: Through heterogeneous hypergraphs and hypergraph wavelet neural networks, it breaks through the limitations of traditional pairwise graph structures, effectively capturing high-order group interactions in social networks and extracting richer structural semantic information.

[0025] Accurate multi-topic competition modeling: By introducing the public opinion field effect and quantifying the energy value and competitive relationship of each topic, the model clearly models the dynamic competition of multiple topics for user attention, solving the shortcomings of existing models that handle cascading independently, and is more in line with actual communication scenarios.

[0026] Efficient and flexible feature fusion: By adopting a gating mechanism and a co-attention fusion strategy, it achieves adaptive fusion of social influence features and public opinion field features, while capturing two-way interaction between users and messages, thus enhancing the expressive ability of user embedding.

[0027] Excellent predictive performance: Experiments on four real datasets—Twitter, Douban, Android, and Christianity—show that the method of this invention significantly outperforms existing mainstream methods in both MAP@k and Hits@k metrics, with a maximum improvement of 3.85% in MAP@100 and a maximum improvement of 2.32% in Hits@100, especially in accurately identifying core disseminating users. Attached Figure Description

[0028] Figure 1 is an overall framework diagram of the POF-HG model in an embodiment of the present invention. Detailed Implementation

[0029] In summary, this invention discloses an information dissemination prediction method based on the public opinion field effect, belonging to the fields of data mining and graph neural networks. It addresses the shortcomings of existing methods based on independent cascades and simple graphs, and mainly has the following two functions: modeling friendship relationships and information forwarding relationships in social networks through heterogeneous hypergraphs, and learning user features through hypergraph wavelet neural networks; and constructing a public opinion field effect to model the competition for limited user attention among different information sources in scenarios where multiple information sources coexist. Through the learning and feature fusion prediction module of these two functions, the information dissemination prediction task is ultimately achieved.

[0030] A method for predicting information diffusion that integrates public opinion field effects and heterogeneous hypergraphs includes four core steps: data preprocessing, social influence learning, public opinion field learning, fusion, and prediction. Data preprocessing involves collecting user sets, social relationships between users (e.g., friend relationships), message sets (e.g., tweets, posts), and message diffusion sequences (recording participating users and diffusion times) from social networks. Based on this data, a heterogeneous cascaded hypergraph is constructed. The nodes of this hypergraph contain both user and message entities, while the edges contain social relationships between users and propagation relationships between users and messages, thus comprehensively characterizing heterogeneous interactions within social networks.

[0031] Social Influence Learning: To capture high-order group interactions, this step utilizes a heterogeneous cascaded hypergraph. Multiple hypergraph snapshots are extracted through different meta-paths (e.g., "user-social relationship-user" and "user-propagation-message-propagation-user"), each snapshot corresponding to a semantic relationship. For each snapshot, an association matrix H (representing the association between nodes and hyperedges) and a node degree matrix are constructed. and hypermarginality matrix And compute the symmetric normalized vertex propagation operator. .based on Construct the Laplacian matrix of the hypergraph and perform spectral decomposition, thereby generating a multi-scale hypergraph wavelet convolution operator. This operator can simultaneously capture both local and global structural dependencies. The hypergraph snapshot is input into a hypergraph wavelet neural network (HWNN), and user features are updated through multi-layer convolutional propagation, ultimately outputting a social influence-driven user embedding vector. .

[0032] Public Opinion Field Learning: To model the dynamics of multi-topic competitive attention, this step treats each message node as a Public Opinion Center (POC), and its corresponding subgraph constitutes a Public Opinion Field (POF). First, the initial features of both user and message nodes are projected onto a shared latent space using a type-adaptive projection matrix, eliminating feature space differences. The energy value of each public opinion field is calculated, which comprehensively considers the popularity contribution and attention level of the public opinion center's neighbor nodes, quantifying the attractiveness of the public opinion field to users. Nodes are divided into four categories based on their neighbor types (user neighbors only, public opinion center neighbors only, and both), and targeted attention update mechanisms are designed for each category to achieve refined learning of node features. Finally, the public opinion field-driven user embedding vector is output. and message embedding vector .

[0033] Fusion and Prediction: To fully leverage the complementary information from social influence and public opinion field effects, this step first dynamically fuses information through a learnable gating mechanism. and Generate fused user embedding vectors This gating mechanism adaptively adjusts the weights of the two types of features. A user-message affinity matrix is ​​constructed to measure the semantic relevance between users and messages, and row normalization is used to obtain the user's attention distribution across all messages. Message features are aggregated based on this attention distribution, and residual enhancement and layer normalization are combined to obtain the interaction-aware user embedding vector. Finally, the probability of a user participating in the next round of propagation is calculated using a softmax function, and users who have already participated in propagation are excluded using a mask, outputting the final prediction result.

[0034] Step 1: Data Preprocessing

[0035] (11) Data collection: Obtain raw data from social networking platforms, including: user sets ,in For the number of users; for the set of social relationships Such as friend relationships, follow relationships; message collections Such as tweets, Q&A posts; diffusion sequences in Indicates user In time Participation message The spread of.

[0036] (12) Construction of heterogeneous cascaded hypergraphs: Constructing hypergraphs , where the node set It includes user nodes and message nodes; edge It includes social hyperedges (connecting user groups with social relationships) and propagation hyperedges (connecting users and message nodes participating in the same message propagation). For node types (user / message). Edge type (social / propagation). The edge weight is initially set to 1 to avoid scale bias.

[0037] Step Two: Learning Social Influence

[0038] (21) Extract hypergraph snapshots based on two meta-paths: Meta-path 1: User-social relationship-user, corresponding to social hypergraph snapshot, which describes the social relationship between users; Meta-path 2: User-propagation-message-propagation-user, corresponding to propagation hypergraph snapshot, which describes the propagation relationship between users and messages and users.

[0039] (22) Matrix construction: For each snapshot, construct the correlation matrix. ( This represents the number of snapshot nodes. (where the number of superedges is 1). Represents a node Belongs to superedge Calculate the node degree matrix (Diagonal) elements are the sum of the weights of the hyperedges associated with the nodes and the hyperedge degree matrix. (The diagonal element is the number of nodes contained in the superedge).

[0040] (23) Construction of Hypergraph Wavelet Convolution Operator: Computation of Symmetric Normalized Vertex Propagation Operator: ,in The hyperedge weight matrix (identity matrix); the hypergraph Laplacian matrix: ,right Perform spectral decomposition Multi-scale hypergraph wavelet kernel: Set the scale =0.5.

[0041] (24) HWNN Propagation: Initial Embedding (Embedding dimension); Propagation formula ,in It is the ReLU activation function. The weight matrix is ​​learnable; after passing through a 2-layer HWNN, the output is a user embedding driven by social influence. .

[0042] Step 3: Learning about the public opinion field

[0043] (3.1) Feature projection: Initial features of user nodes and message nodes Projected into the shared latent space using a type-adaptive projection matrix A: The A-matrix of user nodes and message nodes is learned independently.

[0044] (3.2) Calculation of public opinion field energy: Popularity contribution coefficient: ,in User influence weight (calculated based on historical dissemination frequency). Attention coefficient Public opinion field energy: After normalization, we get .

[0045] (3.3) Node feature update: Case 1 (central node of public opinion) and Case 2 (user nodes containing only user neighbors):

[0046] Case 3 (Including only user nodes of the neighborhood in the center of public opinion):

[0047] In the formula The updated public opinion center features for Case 1; Case 4 (containing user nodes with two types of neighbors):

[0048] Output (Public opinion field drives user embedding) and (Message embedding).

[0049] Step 4: Fusion and Prediction

[0050] (4.1) Gated fusion: Gating coefficient: Integrating user embeddings: .

[0051] (4.2) Co-attention fusion: User-message affinity matrix: Attention distribution: Interaction-aware user embedding: , .

[0052] (4.3) Prediction and Training: Prediction Probability: , The mask matrix is ​​set to -∞ for users already propagated; the loss function is: ,in Indicates user In the Step-by-step propagation; Training parameters: learning rate 0.001, training epochs 100, batch size 32, using Adam optimizer.

[0053] System Implementation: The information diffusion prediction system of this invention can be implemented using the Python language. The hardware environment is an NVIDIA GeForce RTX 4060 GPU (8GB), and the software relies on the PyTorch deep learning framework, NumPy, and SciPy numerical computing libraries. Each module of the system is implemented through modular programming, and data storage uses a MySQL database. It supports both batch processing and real-time prediction modes: batch processing is suitable for analyzing the diffusion path of historical data, while real-time prediction is suitable for predicting the propagation trend of new messages, with a response time of less than 1 second.

[0054] Case Study

[0055] This invention relates to an information diffusion prediction system based on public opinion field effect and heterogeneous hypergraph fusion (POF-HG system for short). This system is particularly suitable for predicting information diffusion behavior on online social networking platforms and can be effectively applied to practical business scenarios such as false information warning, hot topic tracking, and precision marketing promotion. The technical effects of this invention are illustrated below with specific experimental data.

[0056] To verify the effectiveness of this invention, we selected four real-world datasets from different social platforms for testing, including Twitter, Douban, Android, and Christianity community data. Under a unified experimental environment, we compared this invention with eight existing mainstream prediction methods, including feature-based methods (such as DeepDiffuse and TopoLSTM), graph-based methods (such as FOREST and DyHGCN), and fusion methods (such as Inf-VAE). The evaluation metrics used were the industry-standard mean precision (MAP@k) and top k hits (Hits@k), where k was set to 10, 50, and 100.

[0057] (1) Overall prediction performance comparison

[0058] This invention demonstrates optimal performance across all datasets and evaluation metrics, as shown in the table below: Table 1 Performance Comparison of Hits@k (Unit: %)

[0059] Table 2 Performance Comparison of MAP@k (Unit: %)

[0060] Experimental results show that the present invention significantly outperforms the existing technology in all indicators, especially on the Twitter dataset, where Hits@100 is improved by more than 3 percentage points and MAP@100 is improved by more than 1.5 percentage points, demonstrating the robust prediction ability of the system in real and complex social environments.

[0061] (2) Verification of contributions from key technology modules

[0062] To verify the necessity of each core module in this invention, we conducted a decomposition test on the system through ablation experiments. The specific results are as follows: Table 3 Ablation Experiment Performance Analysis (Taking Twitter Dataset as an Example)

[0063] Ablation experiments show that: (1) the public opinion field module contributes the most, and its removal leads to a significant drop in performance, indicating that modeling the multi-topic competition effect plays a key role in the accuracy of prediction; (2) the hypergraph wavelet network module effectively captures high-order group interaction relationships, and its absence also leads to significant performance loss; (3) the collaborative attention fusion mechanism further optimizes the interaction representation between users and information, and improves the system's refined prediction capability.

[0064] The experimental data above demonstrates that the POF-HG system provided by this invention effectively solves two major problems in existing technologies: insufficient modeling of high-order relationships and neglect of multi-topic competition, by integrating public opinion field effects and heterogeneous hypergraph structures. Systematic validation on multiple public datasets shows that this system significantly outperforms existing solutions in terms of prediction accuracy, ranking quality, and robustness, possessing outstanding practical value and promising prospects for widespread application.

[0065] In terms of application, this invention is applicable to recommendation systems targeting social networks. First, it constructs a dynamic interest profile based on user information and their social relationship network (including follow lists, fan interactions, group relationships, etc.), analyzing and generating a series of candidate push notifications. Then, it introduces a public opinion field effect mechanism to simulate the competitive relationship between different pieces of information in the user's current attention field. By quantifying the real-time dissemination energy and social popularity of each piece of content, it dynamically adjusts the push ranking. Finally, the system not only considers users' long-term interests but also combines the real-time competitive situation between information to achieve intelligent push notifications based on competition awareness, preventing high-quality content from being buried due to a current "low-energy" competitive environment.

[0066] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting information diffusion that integrates public opinion field effects and heterogeneous hypergraphs, characterized in that, The process includes the following steps: Data preprocessing: Acquiring social network data, which includes user sets, social relationships between users, message sets, and message propagation sequences, wherein the message propagation sequence includes users participating in message propagation and their corresponding propagation times; Constructing a heterogeneous cascaded hypergraph based on the social network data, wherein the node set of the heterogeneous cascaded hypergraph is the union of the user set and the message set, and the edge set includes social relationship edges and propagation relationship edges; Social influence learning: Based on the heterogeneous cascaded hypergraph, extracting hypergraph snapshots through different meta-paths, and constructing the association matrix, node degree matrix, and hyperedge degree matrix of each snapshot; Based on the aforementioned correlation matrix, node degree matrix, and hyperedge degree matrix, a symmetric normalized vertex propagation operator is calculated, and then the hypergraph Laplacian matrix is ​​constructed and spectral decomposition is performed. A multi-scale hypergraph wavelet convolution operator is constructed based on the spectral decomposition results. User features are updated through multi-layer propagation using a hypergraph wavelet neural network, and the output user embedding vector driven by social influence is generated. Public opinion field learning: Each message node in the heterogeneous cascaded hypergraph is taken as the public opinion center, and the subgraph corresponding to each public opinion center constitutes the public opinion field; the initial features of different types of nodes are projected to the shared latent space through a type adaptive projection matrix; The energy value of each public opinion field is calculated, which is obtained by weighted summation of the popularity contribution coefficient and attention coefficient of the neighboring nodes of the public opinion center; Nodes are categorized into four classes based on their neighbor types, and corresponding attention mechanisms are used to update node features for each category, outputting user embedding vectors driven by the public opinion field. and message embedding vector Fusion and Prediction: Fusion of social influence-driven user embedding vectors through a learnable gating mechanism User embedding vectors driven by public opinion field , to obtain the fused user embedding vector Construct a user-message affinity matrix and obtain the user attention distribution to messages through row normalization; aggregate message features based on the attention distribution, and obtain the interaction-aware user embedding vector through residual enhancement and layer normalization; calculate the probability of a user participating in the next round of message propagation through the softmax function, and output the prediction result after excluding users who have already participated in the propagation.

2. The information diffusion prediction method integrating public opinion field effect and heterogeneous hypergraph as described in claim 1, characterized in that, In the aforementioned social influence learning, the construction process of the hypergraph wavelet convolution operator is as follows: the calculation method based on the symmetric normalized vertex propagation operator is as follows: ,in Let the node degree matrix be... It is an incidence matrix. This is the hyperedge weight matrix. Here is the hyperedge degree matrix; the Laplacian matrix of the hypergraph is calculated as follows: Spectral decomposition yields ,in It is the identity matrix. The eigenvector matrix, It is an eigenvalue diagonal matrix; Construction scale is The hypergraph wavelet kernel, used as a multi-scale hypergraph convolution operator, is calculated as follows: 。 3. The information diffusion prediction method integrating public opinion field effect and heterogeneous hypergraph as described in claim 1, characterized in that, During the learning process of the public opinion field, the energy value of the public opinion field is calculated as follows: in, For the first A public opinion arena, As the center of public opinion The neighborhood group, Neighboring nodes right The popularity contribution coefficient, Neighboring nodes right The attention coefficient and popularity contribution coefficient of the represented message are calculated as follows: in, Neighboring nodes Influence weight in social networks For learnable attention weight vectors, To share the linear transformation matrix, For splicing operations, It is the ReLU activation function. 、 They are the center of public opinion and neighboring nodes Feature vectors in a shared latent space; attention coefficients The calculation method is as follows: in, For feature dimension, 、 They are respectively 、 The 3D eigenvalues 、 They are respectively 、 The characteristic mean.

4. The information diffusion prediction method integrating public opinion field effect and heterogeneous hypergraph as described in claim 1 or 3, characterized in that, In the process of learning the public opinion field, the feature update methods for four types of nodes include: the public opinion center node and user nodes containing only user neighbors are updated using a self-attention mechanism, and the update formula is: in, This is the attention coefficient; user nodes that are only neighbors of the opinion center are updated based on the fusion of self-attention and opinion field energy, and the update formula is: in, For the energy of the normalized public opinion field, The updated features are for nodes that are the center of public opinion; user nodes that contain both the center of public opinion neighbors and user neighbors are updated by calculating the attention coefficients of the two types of neighbors using a dual attention weight matrix, and then weighting and aggregating them.

5. The information diffusion prediction method integrating public opinion field effect and heterogeneous hypergraph as described in claim 1, characterized in that, The fusion expression for the gating mechanism in the fusion and prediction process is as follows: in, 、 For learnable linear transformation matrices, This is the gating coefficient vector. Element-wise multiplication; the user-message affinity matrix is ​​represented as: In the formula, and It is a learnable transformation matrix.

6. The information diffusion prediction method integrating public opinion field effect and heterogeneous hypergraph as described in claim 1 or 5, characterized in that, The expression for the attention distribution aggregating message features during the fusion and prediction process is: In the formula, It is a trainable transformation matrix; the probability formula for calculating the user's participation in information dissemination is: In the formula, This indicates a general forecast. It is a mask matrix.

7. An information diffusion prediction system integrating public opinion field effects and heterogeneous hypergraphs, characterized in that, include: Data preprocessing module: used to acquire social network data and construct heterogeneous cascaded hypergraphs; The social influence learning module is used to extract user embedding vectors driven by social influence through a hypergraph wavelet neural network; the public opinion field learning module is used to model the dynamics of multi-topic competitive attention and output user embedding vectors and message embedding vectors driven by the public opinion field; the fusion prediction module is used to predict the probability of user participation in dissemination through gating mechanism and co-attention fusion; the storage module is used to store social network data, hypergraph structure, embedding vectors and prediction results.

Images