Public opinion event detection method based on graph basic model pre-training

Abstract

The invention discloses a public opinion event detection method based on graph basic model pre-training. The method comprises the following steps: S10, collecting public opinion event detection data; data cleaning, graph construction, feature extraction, graph normalization, data enhancement and data set division are carried out on the collected data; s20, constructing a structure vocabulary learning model in Riemannian geometry, and training the model by using the obtained data set; s30, analyzing the collected public opinion event detection data by using a trained structure vocabulary learning model in Riemannian geometry to obtain a monitoring result; the structure vocabulary learning model in Riemannian geometry comprises the following steps: constructing a product cluster space; vocabulary learning is carried out; performing global learning; geometric contrast learning is carried out; the output graph information node encoding includes shared structure knowledge. According to the method, the conflict between the Euclidean space in the real world and the non-Euclidean measurement of the social network space can be effectively solved, and the problem of unknown event detection in an open domain scene can be effectively solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of public opinion event detection, and in particular relates to a public opinion event detection method based on graph-based model pre-training. Background Art

[0002] In the era of artificial intelligence-based models, public opinion event detection faces new opportunities and challenges. On the one hand, the popularity of the Internet and social media has accelerated the spread of information and deepened the impact of public opinion, while the advantages of graph-structured data in describing complex relationships between entities provide a new perspective for public opinion analysis. Graph-based models (GFMs) draw on the successful experience of large language models such as GPT-4, and improve generalization capabilities through pre-training, which is expected to improve the problems of insufficient data processing, low analysis accuracy, and difficulty in capturing complex semantics in traditional methods in public opinion detection. However, at present, the definition of public opinion events is vague, the characteristics are not clear, the generalization ability of graph neural networks is insufficient, and it is necessary to cope with realistic challenges such as small samples and multiple languages. It is still a difficult problem to achieve high-precision, strong generalization, and cross-language streaming social network public opinion event detection.

[0003] In traditional graph neural networks (GNNs), graph data is usually assumed to be structured in Euclidean space. However, social network data has complex non-Euclidean structures, such as hierarchy and cyclicity. This structure is difficult to represent effectively in Euclidean space, resulting in performance degradation of the model when processing social network data. Specific problems include the representation of hierarchical structures and cyclic structures. User relationships in social networks are often hierarchical, such as the attention relationship between users and the community structure. In Euclidean space, this hierarchy is difficult to accurately represent, resulting in the model being unable to effectively capture the hierarchical relationship between users. There are a large number of cyclic structures in social networks, such as mutual attention between users and interactions within the community. Euclidean space cannot effectively handle these cyclic structures, resulting in information loss when the model handles cyclic relationships. Summary of the invention

[0004] In order to solve the above problems, the present invention proposes a public opinion event detection method based on graph-based model pre-training, which can effectively solve the conflict between the Euclidean space in the real world and the non-Euclidean measurement of the social network space, as well as the problem of unknown event detection in open domain scenarios.

[0005] To achieve the above object, the technical solution adopted by the present invention is: a method for detecting public opinion events based on graph-based model pre-training, comprising the steps of:

[0006] S10, collect public opinion event detection data; and preprocess the data: perform data cleaning, graph construction, feature extraction, graph normalization, data enhancement and data set division on the collected data;

[0007] S20. Construct a structural vocabulary learning model in Riemannian geometry, and train the model with the obtained dataset;

[0008] S30. Use the trained structural vocabulary learning model in Riemannian geometry to analyze the collected public opinion event detection data and obtain monitoring results;

[0009] The structural vocabulary learning model in Riemannian geometry includes: constructing a product bundle space; conducting vocabulary learning; conducting global learning; conducting geometric contrast learning; and outputting graph information node encoding including shared structural knowledge.

[0010] Furthermore, data preprocessing includes:

[0011] (1) Cleaning the data includes removing noise and handling missing values;

[0012] (2) Graph construction includes converting the original data into a graph structure and constructing an adjacency matrix;

[0013] (3) Based on the graph structure, generating feature representations for nodes and edges;

[0014] (4) Using symmetric normalization to normalize the graph structure;

[0015] (5) By applying graph augmentation techniques, generating diverse training samples;

[0016] (6) Randomly partitioning the dataset into a training set, a validation set, and a test set.

[0017] Furthermore, input the graph structure data after data preprocessing into the structural vocabulary learning model in Riemannian geometry, sample trees and cycles in the graph, and model the structural vocabulary; use the hyperbolic space and the hypersphere space to model the substructures of trees and cycles.

[0018] Furthermore, construct a product bundle, and stack general Riemannian layers on the product bundle; in Riemannian geometry, the tangent bundle consists of the hyperbolic space and the hypersphere space that highlight local geometry and the tangent space that describes supplementary information; each node i in the bundle is associated with a node coordinate and a node encoding;

[0019] Derive Riemannian operations, derive a closed-form Riemannian linear operation, and introduce a geometric midpoint as a mathematical preparation, with the arithmetic mean in the hyperbolic space and the hypersphere space as the geometric midpoint;

[0020] In Riemannian geometry, the operation output remains on the manifold; use left matrix multiplication to formulate the linear operation.

[0021] Furthermore, during the vocabulary learning process, cross-geometric attention is performed to embed structural vocabulary into hyperbolic space and hyperspherical space without considering a specific graph, thereby providing shared structural knowledge with cross-domain transferability.

[0022] Furthermore, cross-geometric attention is utilized to learn the node coordinates in substructures, represented by a hyperbolic factor tree; in the hyperbolic manifold, the tree is updated in a bottom-up manner to induce node coordinates from its descendant nodes; the node coordinates are given by attention aggregation and attention weights.

[0023] A transformer network is introduced, where each substructure has cross-geometric keys in the corresponding Riemannian manifold, and the transformer network places the input globally in hyperbolic space.

[0024] Furthermore, during the global learning process, different substructures are aligned with the global view, and node encodings are generated through bundle convolution; multiple substructures are sampled from the graph to examine the entire graph and learn node encodings from a global perspective.

[0025] Furthermore, the global learning process is achieved through the following two stages:

[0026] First, determine the node encodings at the substructure level; the node encodings are located in the tangent bundle around the manifold, and messages on the tangent bundle are passed through bundle convolution, and a unified form is derived for any curvature.

[0027] Second, obtain the encodings of the output nodes at the graph level.

[0028] Furthermore, during the geometric contrast learning process, geometric contrast on the product bundle is performed for the self-supervised learning of the model.

[0029] Furthermore, during the geometric contrast learning process, the node encodings in the tangent space serve as the geometric views of the corresponding manifold; a shared tangent space at the north pole of the Lorentz / sphere model is adopted.

[0030] Beneficial effects of adopting this technical solution:

[0031] The present invention explores the structural geometry of graphs, especially within the framework of Riemannian geometry, and uses hyperbolic space and hyperspherical space to more effectively represent and process graph data, capturing the hierarchical and cyclic structures of graphs, and solving the limitations of traditional graph neural networks in cross-domain applications and generalization capabilities. In the task of public opinion event detection, the model can effectively solve the conflict between the real-world Euclidean space and the non-Euclidean metric of the social network space, as well as the problem of detecting unknown events in open-domain scenarios, and has important practical application value.

[0032] The present invention proposes a new graph-based model that utilizes Riemannian geometry to learn the structural knowledge of graphs in order to achieve cross-domain transferability. Specifically, Riemannian geometry pre-trains a general model by discovering a simple yet effective structural vocabulary (composed of trees and cycles) and exploring its inherent connection with Riemannian geometry, so as to learn the structural knowledge of any graph; it can effectively solve the conflict between the Euclidean space in the real world and the non-Euclidean metrics in the social network space, as well as the problem of detecting unknown events in open-domain scenarios, and has important practical application value.

[0033] The present invention can improve the generalization ability: current graph neural networks are usually designed for specific tasks and lack cross-task generalization ability. In public opinion event detection, the data sources are extensive and diverse, and traditional models are difficult to adapt to public opinion data in different fields. By introducing structural vocabulary and learning the basic structural features of graphs, the present invention enables the model to adapt to various types of graph data and perform well even in graphs lacking rich text attributes, thereby enhancing its generalization ability in different public opinion scenarios.

[0034] The present invention can utilize geometric knowledge: public opinion events in social networks have complex hierarchical and cyclic structures, and traditional graph neural networks are difficult to effectively capture these characteristics in Euclidean space. By combining Riemannian geometry and using hyperbolic space and hypersphere space to embed the structural vocabulary, the present invention can better represent the geometric characteristics of graphs, capture local and global relationships, and thus improve the detection accuracy of public opinion events.

[0035] The present invention can achieve cross-domain transfer: public opinion event detection involves multiple fields such as social media and news websites, and the data characteristics in different fields vary greatly. By constructing a shared structural knowledge base, the present invention allows the model to perform effective knowledge transfer between different fields, reduces the need for re-training, quickly adapts to new tasks and new data sets, and improves the overall efficiency.

[0036] The present invention can promote the development of graph-based models: graph-based models still face challenges in processing diverse graph structures, especially in graphs lacking text attributes. By introducing structural vocabulary and geometric knowledge, the present invention provides new perspectives and methods for the research of graph-based models, and promotes the innovation and development of graph processing technology in public opinion event detection.

[0037] By combining the concepts of Riemannian geometry and structural vocabulary, the present invention develops a new type of graph-based model aimed at enhancing the generalization ability, self-adaptability, and flexibility of graph data analysis, so as to achieve effective graph knowledge transfer and application in a wide range of application scenarios such as public opinion event detection. Description of the Drawings

[0038] Figure 1Schematic flow chart of a public opinion event detection method based on pre-training of a graph-based model according to the present invention;

[0039] Figure 2 Schematic framework diagram of the vocabulary learning module in an embodiment of the present invention;

[0040] Figure 3 Schematic framework diagram of the global learning module in an embodiment of the present invention. Detailed implementation manners

[0041] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described below with reference to the accompanying drawings.

[0042] In this embodiment, as shown in Figure 1 the present invention proposes a public opinion event detection method based on pre-training of a graph-based model, including the steps of:

[0043] S10, collecting public opinion event detection data;

[0044] S20, constructing a structural vocabulary learning model in Riemannian geometry, and training the model with the obtained data set;

[0045] S30, using the trained structural vocabulary learning model in Riemannian geometry to analyze the collected public opinion event detection data to obtain a monitoring result;

[0046] The structural vocabulary learning model in Riemannian geometry includes:

[0047] Constructing a product bundle space;

[0048] Performing vocabulary learning;

[0049] Performing global learning;

[0050] Performing geometric contrast learning;

[0051] Outputting graph information node encoding including shared structural knowledge;

[0052] Among them, collect data for detecting public opinion events; social networks such as Twitter, Facebook, and LinkedIn can also provide valuable data for research, recording users and their relationships. Such data usually forms a directed graph, effectively representing interactions such as user attention, likes, and comments, and is suitable for social network analysis and recommendation systems. By leveraging these public graph databases and social network data, diverse and high-quality graph-structured data can be obtained, enhancing the performance of the model. By carefully selecting data sources, reasonably determining data types and attribute information, ensuring an adequate amount of data, a solid foundation for subsequent model pre-training is laid. And preprocess the data: perform data cleaning, graph construction, feature extraction, graph normalization, data augmentation, and dataset division on the collected data;

[0053] Preferably, the data preprocessing includes:

[0054] (1) Cleaning the data includes removing noise and handling missing values; when removing noise, irrelevant or duplicate nodes and edges, such as isolated nodes and redundant edges, should be deleted to improve the clarity and effectiveness of the graph. When handling missing values, methods such as mean imputation (using the mean of similar nodes) or interpolation (estimating based on adjacent node attributes) can be used. These measures ensure the integrity and consistency of the data, providing high-quality input for subsequent model training.

[0055] (2) Graph construction includes converting the original data into a graph structure and constructing an adjacency matrix; first, the definitions of nodes and edges need to be clarified. For example, in a social network, users can be used as nodes, and the relationships between users can be used as edges. In this way, the relationships in the original data are mapped into a graph structure. Next, an adjacency matrix is constructed based on the connection relationships between nodes. The adjacency matrix is a square matrix used to represent the connection situation of each node in the graph, preparing for the input of the subsequent graph neural network. The constructed adjacency matrix provides the necessary structural information for the model, ensuring effective learning and reasoning.

[0056] (3) Based on the graph structure, generate feature representations for nodes and edges; first, node feature generation is achieved by converting the attributes of nodes into feature vectors, and these vectors may need to be standardized or normalized to ensure that different features are on the same scale, thereby improving the training effect of the model. Second, if the edges have attributes (such as weights, types, etc.), corresponding edge features need to be generated based on these attributes. These edge features help capture the relationships and interactions between nodes, further enhancing the model's understanding and learning ability of the graph structure. Through effective feature extraction, the model can better adapt to complex graph data.

[0057] (4) Normalize the graph structure using symmetric normalization; in graph neural networks, the difference in node connectivity may cause the model to be unbalanced during training. By using symmetric normalization techniques, the adjacency matrix of each node can be normalized to ensure that the influence of each node is on the same scale. This normalization method can reduce the dominant role of high-degree nodes in the model learning process, making the model more stable during training and thus accelerating convergence. This effective normalization process provides a more balanced input for subsequent graph neural networks, promoting the improvement of model performance.

[0058] (5) Generate diverse training samples by applying graph augmentation techniques such as random walk and graph pruning; the random walk method creates new subgraphs by randomly selecting paths in the graph, thus increasing the diversity of samples. Graph pruning, on the other hand, randomly removes some nodes or edges while maintaining the basic structure of the graph to construct different graph morphologies. These augmentation techniques can effectively expand the training dataset, help the model better adapt to unknown data, reduce the risk of overfitting, and thus improve its performance in practical applications.

[0059] (6) Randomly divide the dataset into a training set, a validation set, and a test set. This division helps to avoid overfitting of the model and provides a reliable benchmark for model evaluation. The training set is used for model training, the validation set is used to adjust model hyperparameters, and the test set is used for final performance evaluation.

[0060] As an optimized solution to the above embodiment, input the graph structure data after data preprocessing into the structure vocabulary learning model in Riemannian geometry, sample trees and cycles in the graph, and model the structure vocabulary; use the hyperbolic space and hypersphere space to model the substructures of trees and cycles.

[0061] Construct a product bundle and stack general Riemannian layers on the product bundle; in Riemannian geometry, the tangent bundle consists of the hyperbolic space and hypersphere space that highlight local geometry and the tangent space that describes supplementary information; each node i in the bundle is associated with node coordinates and node encoding;

[0062] Specifically, the coordinates in the manifold contain the relative positions in the substructures (i.e., structure vocabulary), while the encoding in the tangent space carries the information of the global structure.

[0063] To combine the substructures of different geometries, a product bundle is constructed as follows:

[0064]

[0065] where represents the Cartesian product, and d(·) and κ(·) are the dimension and curvature respectively.

[0066] For each node of this product, there is

[0067]

[0068] where || is the vector concatenation,

[0069] The Riemannian metric of the product bundle is generated as:

[0070]

[0071] where is the identity matrix in d H +1 dimensions, and denotes the direct sum between matrices.

[0072] The hyperbolic bundle and the hypersphere bundle are responsible for trees and cycles respectively. Represent the hyperbolic space and the hypersphere space in a unified form as.

[0073] Derive the Riemannian operations, derive a closed-form Riemannian linear operation, and introduce a geometric midpoint as a mathematical preparation, with the arithmetic mean in the hyperbolic space and the hypersphere space as the geometric midpoint;

[0074] Take the arithmetic mean as the geometric midpoint, and the arithmetic mean in the hyperbolic space and the hypersphere space is in the form of the formula:

[0075]

[0076] The arithmetic mean of this formula lies on the manifold and is the geometric midpoint with respect to the squared distance d.

[0077] In Riemannian geometry, the operation output remains on the manifold. However, the lack of isometry and possible mapping errors motivate a fully Riemannian formula to formulate the linear operation using left matrix multiplication. The operation strategy parameterized by W is as follows:

[0078]

[0079] where the redefined factor α is ||·|| denotes the L2 norm.

[0080] Given and κ ≠ 0, for any remains on the manifold, and for any holds. For a set of points and their weights {x i , v i}i∈Ω , where

[0081] As an optimized solution to the above embodiments, during the vocabulary learning process, cross-geometric attention is performed to embed structural vocabulary into hyperbolic space and hyperspherical space without considering a specific graph, thereby providing shared structural knowledge with cross-domain transferability.

[0082] Use cross-geometric attention to learn the node coordinates in the substructure, represented by a hyperbolic factor tree; in the hyperbolic manifold, update the tree in a bottom-up manner, inducing the node coordinates from its descendant nodes; the node coordinates are given by attention aggregation and attention weights;

[0083] As follows:

[0084]

[0085] where j is a descendant node of i, and j contains the coordinate information of i itself. In cross-geometric attention, the key, query, and value are respectively derived by Riemannian linear operations and derived. φ can be any function that returns a scalar. Therefore, the node coordinate is updated to v i . The query value is given by to utilize the compensatory information of another geometry. In addition, the proposed aggregation is unidirectional, which is different from traditional bidirectional aggregation in graph models. In traditional aggregation, each node considers its information in the neighborhood and vice versa. However, as shown in the formula, each node receives the coordinates of its descendant nodes to position itself on the manifold, and there is no reverse information path, that is, the coordinates of the node are not affected by the ancestor nodes in the bottom-up construction. Similarly, refine the loop on the hyperspherical manifold, where the node coordinates are updated by the two nodes connecting it, and the unidirectional path is from the adjacent node to the center.

[0086] Although there are differences in generalization ability, the proposed architecture is fundamentally different from that of graph transformers, which usually perform bidirectional attention on all nodes in Euclidean space. Instead, the proposed attention is unidirectional and is performed on graph substructures considering their Riemannian geometry. Introduce a transformer network, where each substructure has cross-geometric keys in the corresponding Riemannian manifold, and the transformer network places the input globally in hyperbolic space.

[0087] As an optimized solution to the above embodiments, as Figure 2 shown, align different substructures with the global view during the global learning process and generate node encodings through bundle convolution; sample multiple substructures from the graph, examine the entire graph, and learn node encodings from a global perspective.

[0088] The global learning process is achieved through the following two stages:

[0089] First, determine the node encoding at the sub-structure level; the node encoding is located in the tangent bundle around the manifold, where the tangent space of one point is not compatible with that of another point. Due to the spatial incompatibility, existing message passing formulas cannot be used. To bridge this gap, it is proposed to pass messages on the tangent bundle through bundle convolution, and a unified form is derived for arbitrary curvature;

[0090] The unified form derivation for arbitrary curvature is as follows:

[0091]

[0092] where Λ is the set of nodes of the sub-structure, and α it is the attention weight. The basic principle for solving spatial incompatibility lies in parallel transport, which is a standard way to connect different tangent spaces.

[0093] In Riemannian geometry, parallel transport (PT x→y ) with respect to the Levi-Civita connection transports the vector v in the tangent space at point x on the manifold M to another tangent space along the geodesic between x and y by a linear isometry. The unified form derivation for arbitrary curvature will encode and be parallel transported to the tangent space of the target point, where message passing is subsequently performed. The advantage of bundle convolution is that it takes into account the encoding of the global structure while encapsulating the local geometry of the manifold.

[0094] Second, obtain the encoding of the output nodes at the graph level.

[0095] For K samples of a, they are given by the geometric midpoint of the coordinates on the manifold and aligned, where the aggregation weight is set to 1. Then, the node encoding of each sample is parallel transported to the tangent space of the midpoint. Thus, the node encoding at the graph level is derived as

[0096] Regarding stacking multiple layers, the main advantage is to expand the receptive field. For example, a node in a tree is updated in one layer by its first-level descendant nodes and further influenced by second-level descendant nodes in another layer. By stacking multiple layers, a node can perceive a larger area in the sub-structure while expanding its global view by calculating the consistency of the entire graph.

[0097] As an optimized solution to the above embodiment, such as Figure 3As shown, the base model requires self-supervised learning to obtain shared knowledge that does not depend on specific annotations. Contrastive learning has become an effective method for self-supervised learning, but it is not easy for graphs; for example, graph augmentations that generate contrastive views are not as easy as cropping / rotating images. Due to their different designed structural geometries, they provide different views for graph contrastive learning (i.e., hyperbolic views and hyperspherical views). In the present invention, during the geometric contrastive learning process, geometric contrast on the product bundle is performed for the self-supervised learning of the model without graph augmentation;

[0098] During the geometric contrastive learning process, the node encodings in the tangent space serve as geometric views of the corresponding manifolds; thus, the remaining component is a scoring function for contrasting positive and negative samples, and the challenge lies in the incompatibility between different geometries. To bridge this gap, a shared tangent space using the Lorentz / spherical model's north pole is considered.

[0099] Therefore, the geometric contrast is as follows:

[0100]

[0101] The overall objective is formulated as where N is the number of nodes. Although the geometric contrast is performed on the node encodings in the tangent space, the parameters of the factor manifolds are encapsulated in the parallel transport between the tangent spaces. Its computational complexity is O(|v| 2 +|e|, where v and e are the node set and edge set respectively, and the proposed model supports mini-batch training.

[0102] Finally, the model can generate informative node encodings for any graph, which contain the shared structural knowledge of the graph domain learned in Riemannian geometry.

[0103] In the present invention, by combining the concepts of Riemannian geometry and structural vocabulary, a general model capable of understanding and processing various graph structures is pre-trained to promote the development of graph data analysis and machine learning techniques in public opinion event detection. The key technologies of the present invention mainly focus on the following aspects:

[0104] 1. Innovative application of the graph base model: The present invention proposes a graph base model aiming to achieve transferability across different datasets and domains. This model is specifically targeted at graph-structured data and can handle various non-Euclidean structures from recommendation systems to public opinion dissemination networks. The core advantage of the graph base model is its ability to transcend domain-specific limitations and provide a general graph data analysis solution, especially suitable for the complex and changing network structures in public opinion event detection.

[0105] 2. Combination of Structural Vocabulary and Riemannian Geometry: The present invention combines structural vocabulary (such as trees and cycles) with Riemannian geometry. These structural vocabulary are the basic elements that make up any graph, and the characteristics of Riemannian geometry can better represent and process graph data, especially in capturing the hierarchical and cyclic structures of graphs. This technology can effectively handle the complex network topologies in public opinion event detection and enhance the model's ability to capture the propagation paths and dynamic changes of public opinion.

[0106] 3. Geometric Contrastive Learning: Pre-training is carried out through geometric contrastive learning to learn the structural knowledge of graphs. The specific method is to construct a novel product bundle to integrate different geometric structures, and then stack Riemannian layers on this constructed space, enabling the structural vocabulary to learn in the Riemannian manifold. This method allows the model to learn shared structural knowledge without considering the specific graph structure and generate node encodings for any input graph, thereby enhancing the adaptability and accuracy in public opinion event detection.

[0107] 4. Shared Structural Knowledge for Cross-Domain Transfer Ability: The present invention provides the shared structural knowledge required for cross-domain transfer and explores graph-based models for a wider range of practical graphs (not limited to graphs with text attributes). This technology studies graph-based models from the perspective of structural geometry for the first time and can effectively address cross-domain issues in public opinion event detection, such as the differences between public opinion data from different platforms or different topics. Through shared structural knowledge, the model can quickly adapt to new tasks and new data sets, reducing the need for re-training.

[0108] By combining the concepts of Riemannian geometry and structural vocabulary, the present invention develops a novel graph-based model aimed at enhancing the generalization ability, self-adaptability, and flexibility of graph data analysis, thereby achieving effective graph knowledge transfer and application in a wide range of application scenarios such as public opinion event detection.

[0109] The above has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. What is described in the above embodiments and the specification only illustrates the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for detecting public opinion events based on graph-based model pre-training, characterized in that: Includes steps: S10, collect public opinion event detection data; and preprocess the data: perform data cleaning, graph construction, feature extraction, graph normalization, data enhancement and data set division on the collected data; S20, constructing a structural vocabulary learning model in Riemannian geometry, and training the model with the acquired data set; S30 uses the trained structural vocabulary learning model in Riemannian geometry to analyze the collected public opinion event detection data to obtain monitoring results; The structural vocabulary learning model in Riemannian geometry includes: constructing a product bundle space; performing vocabulary learning; performing global learning; performing geometric contrast learning; and outputting graph information node encoding including shared structural knowledge.

2. According to claim 1, a method for detecting public opinion events based on graph-based model pre-training is characterized in that: Data preprocessing includes: (1) Cleaning data includes removing noise and processing missing values; (2) Graph construction includes converting raw data into a graph structure and constructing an adjacency matrix; (3) Generate feature representations for nodes and edges based on the graph structure; (4) Use symmetric normalization to normalize the graph structure; (5) Generate diverse training samples by applying graph enhancement techniques; (6) Randomly divide the dataset into training set, validation set and test set.

3. According to claim 1, a method for detecting public opinion events based on graph-based model pre-training is characterized in that: The graph structure data after data preprocessing is input into the structural vocabulary learning model in Riemannian geometry, and trees and cycles are sampled in the graph to model the structural vocabulary; The substructures of trees and cycles are modeled using hyperbolic and hyperspherical spaces.

4. According to claim 3, a method for detecting public opinion events based on graph-based model pre-training is characterized in that: Construct a product bundle and stack a universal Riemannian layer on it. In Riemannian geometry, a tangent bundle consists of a hyperbolic space and a hyperspherical space that highlight the local geometry, and a tangent space that describes the supplementary information. Each node i in the bundle is associated with a node coordinate and a node code. Derive the Riemann operation, derive a closed form Riemann linear operation, and introduce a geometric midpoint as a mathematical preparation, the arithmetic mean in hyperbolic space and hyperspherical space as the geometric midpoint; In Riemannian geometry, the output of an operation remains on a manifold; linear operations are formulated using matrix left multiplication.

5. According to claim 1, a method for detecting public opinion events based on graph-based model pre-training is characterized in that: During vocabulary learning, a cross-geometric attention is performed to embed the structural vocabulary into hyperbolic and hyperspherical spaces regardless of the specific graph, thus providing shared structural knowledge for cross-domain transferability.

6. According to claim 5, a method for detecting public opinion events based on graph-based model pre-training is characterized in that: We use cross-geometric attention to learn node coordinates in a substructure, represented by a hyperbolic factor tree. We update the tree in a bottom-up manner in a hyperbolic manifold, inducing node coordinates from their descendant nodes. Node coordinates are given by attention aggregation and attention weights. A transformer network is introduced where each substructure has a cross-geometry bond in the corresponding Riemannian manifold, while the transformer network places the input as a whole in a hyperbolic space.

7. According to claim 1, a method for detecting public opinion events based on graph-based model pre-training is characterized in that: In the global learning process, different substructures are aligned with the global view and node encodings are generated through clustered convolutions; multiple substructures are sampled from the graph, the entire graph is examined, and node encodings are learned from a global perspective.

8. According to claim 7, a method for detecting public opinion events based on graph-based model pre-training is characterized in that: The global learning process is achieved through the following two stages: First, the node encoding of the substructure level is determined; the node encoding is located in the tangent bundle around the manifold, and the message on the tangent bundle is transmitted through the bundle convolution, and the arbitrary curvature is derived in a unified form; Secondly, the encoding of the output nodes is obtained at the graph level.

9. According to claim 1, a method for detecting public opinion events based on graph-based model pre-training is characterized in that: In the process of geometric contrast learning, geometric contrast is performed on the product bundle for self-supervised learning of the model.

10. A method for detecting public opinion events based on graph-based model pre-training according to claim 9, characterized in that: During geometric contrastive learning, nodes in the tangent space are encoded as geometric views of the corresponding manifold; a shared tangent space of the North Pole of the Lorentz / sphere model is adopted.

Images