News event risk indicator association relationship prediction method based on balanced manifold graph
By using a graph neural network based on a balanced manifold graph, the problems of redundant indicator information and difficulty in parsing coupling relationships in multi-indicator correlation analysis are solved, achieving efficient prediction of correlation relationships of event risk indicators and improving the accuracy and robustness of the model.
Patent Information
- Application Number
- CN202511309772.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-15
- Publication Date
- 2025-12-30
AI Technical Summary
Existing risk assessment methods based on fuzzy graphs and Bayesian networks suffer from redundant indicator information and difficulties in parsing complex coupling relationships in multi-indicator correlation analysis. They are also unable to handle high-dimensional indicator systems and have high complexity in learning network structures.
We employ a graph neural network based on a balanced manifold graph. By constructing a balanced manifold module and an event mapping module, we utilize the graph neural network to extract key features of multi-level indicators, learn the correlation between events and risk indicators, and use the cross-entropy loss function and adaptive optimization algorithm to update parameters and optimize the geometric distribution structure of the feature space.
It significantly improves the accuracy of the correlation between event risk indicators and the ability to represent complex relational data, thereby enhancing the robustness of the model in classification tasks and the accuracy of evaluation results.
Smart Images
Figure CN121234988A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of risk assessment technology, and specifically relates to a method for predicting the correlation of news event risk indicators based on balanced manifold diagrams. Background Technology
[0002] In the field of risk assessment, risk is defined as the probability of harmful consequences or anticipated losses resulting from the complex interaction between disasters and vulnerabilities in cities and societies. In recent years, risk assessment, particularly multi-indicator correlation risk analysis, has become a popular topic in risk analysis due to its ability to comprehensively consider various factors.
[0003] Existing methods based on fuzzy graphs and Bayesian networks have the following problems:
[0004] (1) Redundancy of indicator information. Current multi-indicator correlation analysis methods generally face the challenge of redundant indicator information when dealing with the evolution process of thematic events. Although traditional dimensionality reduction methods can extract principal components, their linear assumptions make it difficult to handle nonlinear redundant relationships in the event evolution process.
[0005] (2) The dilemma of index coupling analysis. The complex coupling relationships between multiple indicators cause traditional correlation analysis methods to fail, such as time-varying coupling, hierarchical coupling and implicit coupling. Existing solutions such as dynamic Bayesian networks (DBN) can model time-series dependencies, but their network structure learning complexity is O(T·n²) (T is time and n is the number of indicators), which is difficult to handle high-dimensional index systems. Summary of the Invention
[0006] To address the aforementioned problems, this invention proposes a method for predicting the correlation between news event risk indicators based on balanced manifold graphs. This invention considers that the semantic information of the indicators and the importance of multi-level indicators in assessing event risk are crucial. It utilizes graph neural networks to extract key features from multi-level indicators, and learns the final prediction level through a balanced manifold and classifier, while simultaneously effectively learning the correlation between events and risk indicators.
[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0008] The method for predicting the correlation of news event risk indicators based on balanced manifold diagrams includes the following steps:
[0009] Step 1: Collect the original text of news events, label risk indicators, and construct training, validation, and test sets;
[0010] Step 2: Construct a graph neural network based on a balanced manifold graph;
[0011] Step 3: Train the graph neural network using samples from the training set, validate it using the validation set, and obtain the accuracy of risk indicator predictions on the test set.
[0012] Step 4: Collect the text of the news event to be analyzed, use the trained model to predict the event risk indicators, and obtain one or more indicators with the highest correlation.
[0013] Furthermore, the specific method of step 1 is as follows:
[0014] We collected the original records of 2,000 historical events using web crawlers and extracted the event description text.
[0015] Using one of the indicators from a pre-defined set of indicators, the event description text is manually annotated to obtain the risk indicator corresponding to the event.
[0016] Using a pre-trained BERT model, semantic knowledge is learned in an unsupervised manner from event description text, transforming the event description text into vectors;
[0017] Graph structured data is constructed from vectors generated from event description texts. Each event yields one graph structured data point, which is defined as follows: , among which, nodes This represents the current event and each indicator in the set of evaluation metrics, with edges... Adjacency matrices are used to represent the relationships between nodes. To record Given a topological structure A, the elements A[i][j] in A satisfy the following equation:
[0018]
[0019] in, express The line connecting them is an edge in E;
[0020] All graph-structured data constitute the final sample set, which is then divided proportionally into training, validation, and test sets.
[0021] Furthermore, the graph neural network described in step 2 includes a balanced manifold module and an event mapping module; wherein, the balanced manifold module includes a KNN layer and an MDS layer, the KNN layer finds the k nearest neighbors of each graph node including the event, and constructs an adjacency graph between these nodes, the edges in the adjacency graph represent the connection between the event and the feature, and the shortest path algorithm is used to calculate the shortest path distance between the event and the indicator, and these shortest path distances are used as the geodesic distance between the features; the MDS layer constructs a distance matrix based on the calculated geodesic distance, where each element represents the geodesic distance between two features, thereby representing the geometric relationship of the symptom indicator;
[0022] The event mapping module maps the event node representation to the MDS vector space to obtain the MDS representation of the event, and at the same time obtains the correlation weight between the syndrome index and the event.
[0023] Furthermore, in step 3, cross-entropy is used as the loss function, the gradient of the loss function is calculated through backpropagation, and the parameters are updated using an adaptive optimization algorithm. At the end of each training cycle, the performance of the model is evaluated based on the validation set. The training process continues until the maximum number of iterations is reached, and then the learning process is terminated. The best-performing model parameters in the entire training process are automatically identified and stored through a global comparison mechanism. Then, the saved model parameters are loaded, and the risk indicators corresponding to the events are predicted on the test set to obtain the accuracy on the test set.
[0024] Furthermore, step 4 is specifically implemented as follows:
[0025] Collect event text data to be analyzed through web crawlers or other means;
[0026] The event text data is converted into word embeddings through a pre-trained BERT model to construct the corresponding graph structure data, where nodes represent events and various indicators in the evaluation indicator set, and edges represent the relationships between nodes.
[0027] The processed graph structure data is input into the trained model to obtain the three evaluation metrics that are most relevant to the current event.
[0028] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:
[0029] 1. This invention considers that the semantic information of indicators and the crucial role of multi-level indicators in assessing event risk are essential. It utilizes graph neural networks to learn the feature representations of multi-level indicators, fuses key features using graph neural networks, strengthens the discriminative features in the feature space using balanced manifolds, and updates network parameters based on the true risk category label loss. The trained model is then used to predict the correlation between event risk indicators. This invention can fully explore the correlation between events and symptom indicators, improving accuracy.
[0030] 2. This invention utilizes graph neural networks to aggregate node neighborhood information and adaptively extracts key feature representations from graph structures, significantly improving the representation ability of complex relational data and the robustness of classification tasks.
[0031] 3. Based on balanced manifolds, this invention significantly enhances the intra-class compactness and inter-class separability of discriminative features by optimizing the geometric distribution structure of the feature space, thereby improving the accuracy of the final evaluation results. Attached Figure Description
[0032] Figure 1 This is a schematic diagram of a graph neural network based on a balanced manifold.
[0033] Figure 2 This is a flowchart of a method for predicting the correlation of news event risk indicators based on balanced manifold diagrams. Detailed Implementation
[0034] To make the objectives, technical methods, and advantages of this invention clearer, the invention will be further described below with reference to the accompanying drawings and specific embodiments. However, the following embodiments are by no means intended to limit the invention in any way.
[0035] A method for predicting the correlation of news event risk indicators based on balanced manifold diagrams, such as Figure 2 As shown, it includes the following steps:
[0036] Step 1: Collect the original text of news events, label risk indicators, and construct training, validation, and test sets; the specific method is as follows:
[0037] We collected the original records of 2,000 historical events using web crawlers and extracted the event description text.
[0038] Using one of the indicators from a pre-defined set of indicators, the event description text is manually annotated to obtain the risk indicator corresponding to the event.
[0039] Using a pre-trained BERT model, semantic knowledge is learned in an unsupervised manner from event description text, transforming the event description text into vectors;
[0040] Graph structured data is constructed from vectors generated from event description texts. Each event yields one graph structured data point, which is defined as follows: , among which, nodes This represents the current event and each indicator in the set of evaluation metrics, with edges... Adjacency matrices are used to represent the relationships between nodes. To record Given a topological structure A, the elements A[i][j] in A satisfy the following equation:
[0041]
[0042] in, express The line connecting them is an edge in E;
[0043] All graph-structured data constitute the final sample set, which is then divided proportionally into training, validation, and test sets.
[0044] Step 2: Construct a graph neural network based on a balanced manifold;
[0045] like Figure 1 As shown, the balanced manifold network includes a balanced manifold module and an event mapping module. The balanced manifold module includes a KNN (K-Nearest Neighbors) layer and an MDS (Multidimensional Scaling) layer. The KNN layer finds the k nearest neighbors for each graph node, including events, and constructs an adjacency graph between these nodes. The edges in the adjacency graph represent the connections between events and features. The shortest path algorithm is used to calculate the shortest path distance between events and indicators. These shortest path distances are used as geodesic distances between features. The MDS layer constructs a distance matrix based on the calculated geodesic distances, where each element represents the geodesic distance between two features, thereby representing the geometric relationship of the symptom indicators.
[0046] The event mapping module maps the event node representation to the MDS vector space to obtain the MDS representation of the event, and at the same time obtains the correlation weight between the syndrome index and the event.
[0047] The k-Nearest Neighbors (kNN) algorithm is a classic data mining technique used to find the k nearest neighbors to a given data point in a dataset. In this example, the KNN algorithm is applied to an event dataset to find the k nearest neighbor events for each event.
[0048] First, it's necessary to define a similarity metric between events. This typically involves calculating some kind of distance or similarity index between events, such as Euclidean distance, Manhattan distance, or cosine similarity. The choice of metric depends on the characteristics of the event data and the purpose of the analysis; in this example, Euclidean distance is used as the metric.
[0049] Next, the event dataset is traversed, and the KNN algorithm is applied to each event. Specifically, for each event in the dataset, a similarity metric between its features and those of all other events is calculated, and the events are sorted according to these values. Then, the top k events are selected as the nearest neighbors of that event.
[0050] After determining the k nearest neighbor events for each event, an adjacency graph is further constructed to represent the connection relationships between these events.
[0051] An adjacency graph is a graph structure that uses nodes to represent data points (in this case, events) and edges to represent the connections between data points. In this example, each event is represented as a node, and the connections between an event and its k nearest neighbor events are represented as edges.
[0052] Constructing an adjacency graph is relatively simple. First, create a node for each event and add it to the graph. Then, for each event, find its k nearest neighbor events and add edges between them. This results in an adjacency graph containing all events and their k nearest neighbor connections.
[0053] The edges in this adjacency graph represent connections between events and their k nearest neighbors. The length of the edges can be the Euclidean distance in the original space, but subsequent calculations will be based on these connections to estimate the geodesic distance.
[0054] In mathematics and geometry, geodesic distance is commonly used in distance calculations for graph networks and mesh surfaces. Classic algorithms for calculating geodesic distance include Dijkstra's algorithm and the Floyd-Warshall algorithm, which calculate the geodesic distance by finding the shortest path between two points.
[0055] This example uses Dijkstra's algorithm to calculate the shortest path distance between all pairs of points, and these shortest path distances are used as approximations of the geodesic distance between data points.
[0056] Dijkstra's algorithm calculates the shortest path from one vertex to all other vertices, solving the shortest path problem in a weighted graph. Its main characteristic is that it starts from the starting point and employs a greedy algorithm strategy, iterating through the nearest unvisited vertex's adjacent nodes each time until it reaches the destination.
[0057] Based on geodesic distance, a distance matrix is constructed, where each element represents the geodesic distance between two data points. Then, Medium-Density Sequence (MDS) is used to transform the distance matrix into coordinates of points in a low-dimensional space. The goal is to find a set of coordinates such that the distances between these points are as close as possible to the distances in the distance matrix.
[0058] First, any N×N real symmetric matrix has N linearly independent eigenvectors and can be orthogonally normalized. A real symmetric matrix A can be decomposed into:
[0059]
[0060] Where Q is an orthogonal matrix, Λ is a real diagonal matrix, and the superscript T denotes the transpose of the matrix.
[0061] Given m samples X = {x1, x2, ..., xn} m}∈R d×m The distance matrix D∈R m×m The element dist in the i-th row and j-th column ij For sample x i To x j The distance. The goal is to obtain a low-dimensional representation:
[0062] Z = {z1, z2, …, z m} ∈ R d’×m
[0063] where d’ < d, and the Euclidean distance between any two samples in the d’-dimensional space remains unchanged.
[0064] From the known conditions, the only relationship that can be obtained between the samples after dimensionality reduction and those before dimensionality reduction is as follows:
[0065]
[0066] To more clearly observe the relationship between z i , z j and dist ij , square the left side of the equation to get:
[0067]
[0068] It can be seen that disti j2 is related to the norms and inner products of z i , z j respectively. To be able to uniformly represent this relationship, the inner product matrix B = Z T Z ∈ R m × m is introduced here, where b ij = z i T z j , and we can get:
[0069]
[0078]
[0079]
[0080] Combining the above formulas, we get:
[0081]
[0082] At this point, we have obtained all the relationships between B and D. The next step is to derive the final goal Z from B.
[0083] By the properties of real symmetric matrices, B = ,in =diag( , ) is a diagonal matrix composed of eigenvalues. Let V be the eigenvector matrix. Then we have:
[0084]
[0085]
[0086] In general dimensionality reduction tasks, d′≪d is always taken as the target dimension, so d' largest eigenvalues are selected to form a diagonal matrix. The corresponding eigenvector matrix is ,but
[0087]
[0088] The input to the entire algorithm is a distance matrix. Let its element distij be the distance from sample xi to xj, and the dimension of the low-dimensional space be d'. First calculate... ; Calculate the inner product matrix B; Perform eigenvalue decomposition on B; Take Let d' be a diagonal matrix composed of the d' largest eigenvalues. The corresponding eigenvector matrix; final output .
[0089] The geometric relationship characterization of symptom indicators based on geodesic distance can be obtained through the above.
[0090] Step 3: Train the balanced manifold network using samples from the training set, validate it using the validation set, and obtain the accuracy of risk indicator predictions on the test set; specifically, use cross-entropy as the loss function L:
[0091]
[0092] Where M represents the number of risk levels; The sign function is set to 1 if the true class of sample i is c, and 0 otherwise. Let i be the predicted probability of observation sample i being risk level c.
[0093] By analyzing geometric relationships, the risk level of an event can be predicted more accurately. The correlation between the indicators and the event and their weights can be obtained through the prediction weights automatically learned by the model.
[0094] The gradient of the loss function is calculated through backpropagation, and the parameters are updated using an adaptive optimization algorithm. At the end of each training cycle, the performance of the model is evaluated based on the validation set. The training process continues until the maximum number of iterations is reached, and then the learning process is terminated. The optimal model parameters are automatically identified and stored through a global comparison mechanism. The saved model parameters are then loaded, and the risk indicators corresponding to the events are predicted on the test set to obtain the accuracy on the test set.
[0095] Step 4: Collect the text of the news events to be analyzed, use the trained model to predict the event risk indicators, and obtain one or more indicators with the highest correlation; the specific method is as follows:
[0096] Collect event text data to be analyzed through web crawlers or other means;
[0097] The event text data is converted into word embeddings through a pre-trained BERT model to construct the corresponding graph structure data, where nodes represent events and various indicators in the evaluation indicator set, and edges represent the relationships between nodes.
[0098] The processed graph structure data is input into the trained model to obtain the three evaluation metrics that are most relevant to the current event.
[0099] This invention utilizes graph neural networks to establish a hierarchical indicator system. By optimizing the data distribution structure of the feature space through balanced manifolds, it significantly enhances the expression of discriminative features, thereby improving the model's classification performance and generalization ability. Furthermore, by combining graph neural networks and balanced manifolds, it is possible to objectively assess event risk by integrating the importance of indicators, and quantitatively calculate indicator weights. This invention can fully explore the correlation between events and symptom indicators, improving accuracy.
Claims
1. A method for predicting the correlation of news event risk indicators based on balanced manifold graph, characterized in that, The method comprises the following steps: Step 1, collect the original text of news events, label risk indicators, and construct training set, validation set and test set; Step 2, construct a graph neural network based on balanced manifold graph; Step 3, use the samples in the training set to train the graph neural network, verify it through the validation set, and get the accuracy rate of risk indicator prediction on the test set; Step 4, collect the news event text to be analyzed, use the trained model to predict the event risk indicators, and get the one or more indicators with the highest correlation. 2.The balanced manifold graph based news event risk indicator correlation relationship prediction method according to claim 1, characterized in that, The specific way of step 1 is: Collect 2000 historical event original records through network crawler, and extract event description text; Use one indicator in the pre-defined indicator set to manually label the event description text to get the corresponding risk indicator of the event; Use the pre-trained bert model to learn semantic knowledge from the event description text in an unsupervised manner, and convert the event description text into a vector; According to the vector converted from the event description text, a graph structure data is constructed, each event obtains a graph structure data, and the graph structure data is defined as wherein, nodes represent the current event and each index in the evaluation index set, edges represent the association relationship between the nodes, and an adjacency matrix is used to record the topological structure, and elements A[i][j] in A satisfy the following formula: wherein represents the line connecting is an edge in E; All graph structure data constitutes the final sample set, which is divided into training set, validation set and test set in proportion. 3.The balanced manifold graph based news event risk indicator correlation relationship prediction method according to claim 2, characterized in that, The graph neural network in step 2 includes a balanced manifold module and an event mapping module; wherein the balanced manifold module includes a KNN layer and an MDS layer, the KNN layer finds k nearest neighbors for each graph node including events, and constructs an adjacency graph between these points, the edges in the adjacency graph represent the connection between events and features, and the shortest path algorithm is used to calculate the shortest path distance between events and indicators, which is used as the geodesic distance between features; the MDS layer constructs a distance matrix based on the calculated geodesic distance, wherein each element represents the geodesic distance between two features, and the geometric relationship between the indicators is represented; The event mapping module maps the event node representation to the MDS vector space to get the MDS representation of the event, and gets the correlation weight between the syndrome indicators and the event.
4. The balanced manifold graph-based news event risk indicator association relationship prediction method according to claim 3, characterized in that, In step 3, cross entropy is used as the loss function, the loss function gradient is calculated by back propagation, and the adaptive optimization algorithm is used to update the parameters; at the end of each training period, the performance of the model is evaluated based on the validation set, the training process continues until the maximum iteration number is reached, then the learning process is terminated, and the model parameters with the best performance in the whole training process are automatically identified and stored through the global contrast mechanism; Then load the saved model parameters to predict the risk indicators corresponding to the events on the test set to get the accuracy rate on the test set.
5. The balanced manifold graph based news event risk indicator correlation relationship prediction method according to claim 4, characterized in that, The specific way of step 4 is: Collect event text data to be analyzed through network crawler or other ways; Convert the event text data into word embedding through the pre-trained bert model to construct corresponding graph structure data, wherein the nodes represent events and each indicator in the indicator set, and the edges represent the relationship between the nodes; Input the processed graph structure data into the trained model to get the three evaluation indicators with the highest correlation to the current event.