An intelligent emergency prediction method based on a graph convolutional neural network, a computer device and a computer readable storage medium

Abstract

The application belongs to the field of artificial intelligence, and proposes a sudden event intelligent prediction method based on a graph convolutional neural network, a computer device and a computer readable storage medium. The application first introduces a hypergraph model to represent the relationship between multiple event entities, which can better capture the high-order complexity relationship that may exist between event entities. Then, by selecting different time granularity, a local hypergraph of an event is divided, and a certain hyperedge in the local hypergraph is converted into a simple graph. Then, by using the message propagation mechanism in the graph convolutional neural network, each node on the graph can aggregate the state information of the surrounding neighbor nodes, and the feature representation of the graph is read by taking the average value of all the aggregated node features. Finally, the representation of the graph is input into a classifier to obtain the final trend level. The intelligent prediction model constructed by the graph convolutional network greatly improves the prediction accuracy by fully utilizing the semantic association information of event elements.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, and specifically relates to an intelligent prediction method for sudden events based on graph convolutional neural networks, a computer device, and a computer-readable storage medium. Background Technology

[0002] Emergencies generally refer to sudden events that cause or may cause serious social harm, requiring emergency response measures, including natural disasters, accidents, public health emergencies, and social security incidents. These major emergencies, encompassing natural disasters, accidents, and public health emergencies, pose significant threats to life and property safety in affected areas, while their inherent urgency and uncertainty present enormous challenges to the stable development of countries worldwide. Furthermore, the advent of the big data era has made traditional expert-based emergency decision-making processes inadequate for handling massive amounts of data. Therefore, the key to improving emergency prediction capabilities lies in how to automatically construct emergency trend prediction models from massive amounts of unstructured text data and information using technologies such as natural language processing, deep learning, and big data processing.

[0003] Existing methods for predicting the trend of emergencies fall into two categories: The first uses quantitative analysis based on event data, transforming events identified in news datasets into quantifiable numerical values ​​using a coding system, and then analyzing these values. The second is based on feature word sets, employing feature word encoding and assignment, with experts defining themes and feature words for each theme, quantifying themes based on word frequency information in the sample, and using a Naive Bayes model to predict the trend level of emergencies. Both methods heavily rely on domain experts and suffer from limitations in feature dimensions and timeliness. While expert data filtering can improve prediction accuracy, with the advent of the big data era and the exponential growth of data volume, relying solely on expert analysis is extremely time-consuming, labor-intensive, and even impossible. Summary of the Invention

[0004] To address the aforementioned problems, this invention provides an intelligent prediction method, computer device, and computer-readable storage medium for sudden events based on graph convolutional neural networks. The intelligent prediction model for sudden events constructed in this invention fully utilizes the high-order interrelationships between entities and relationships within events, enabling it to effectively predict the trends of sudden events.

[0005] This invention proposes an intelligent prediction method for sudden events based on graph convolutional neural networks, which includes: S1. First, event information data is extracted from text data using event extraction technology. The event information data is represented in the form of quadruples to obtain structured event information data. A dataset is constructed based on the structured event information data, and the dataset is divided into training set and test set. An emergency event hypergraph is generated based on the training set. The emergency event hypergraph includes a set of supernodes and a set of superedges. The set of nodes includes the entity types of the emergency events, and each superedge represents a complete emergency event. S2, segment the local sudden event hypergraph from the sudden event hypergraph according to the time slice, and use it as the feature graph of the current stage; Extracting text features from supernodes in a local burst event hypergraph: The event description statements from different nodes in the event dataset are segmented into words, and the average value of the word vectors of the segmented text is taken. For the remaining attribute nodes that form words, word vectors are obtained. For out-of-vocabulary words, zero vectors are used instead. The text features of the nodes in the local sudden event hypergraph are vectorized to obtain the corresponding text feature representations. The entities of different nodes are set as the corresponding category feature representations. Then, a multilayer perceptron network is used to map the text feature representation and category feature representation of each node in the local burst event hypergraph to a common style space of a specified size. The text feature representation and category feature representation of each node in the common style space are concatenated as the initial features of the local burst event hypergraph node. S3. Randomly select any hyperedge in the hypergraph of a local sudden event, perform an inner product of the initial features of different nodes in the hyperedge, and obtain the two nodes with the greatest difference as key nodes. Connect the key nodes with the remaining nodes in the hyperedge to obtain a simple graph model. S4, aggregate the state information of all nodes in the hyperedge of the simple graph model to the key node. Each node in the simple graph updates its own state information by aggregating the state information of the surrounding nodes and its own state information. Take the average value of all aggregated node features as the feature representation of the simple graph. Linearly transform the feature representation of the simple graph through a fully connected layer and then input it into a softmax layer for classification to obtain different classification probabilities. S5 takes the classification probability of each simple graph model as input, uses the cross-entropy function to calculate the prediction loss of a single sample, and adjusts the learnable parameters in the entire simple graph model based on the prediction loss of all samples in a batch. S6. Repeat steps S3 to S5 until all batches of the training set have participated in the training of the simple graph model and reached the specified number of iterations to obtain the emergency event trend prediction model; input the test set into the trained emergency event trend prediction model for prediction to realize emergency event trend prediction.

[0006] In one embodiment, in step S1, the event information data includes event description, time, location, participants, and event type; the event information data is represented in the form of a quadruple, which includes: head entity, relation, tail entity, and event stamp.

[0007] In one embodiment, in step S1, a sudden event hypergraph is generated based on the training set. The sudden event hypergraph includes a set of supernodes and a set of superedges. The set of nodes includes the entity types of sudden events, and each superedge represents a complete sudden event. Specifically: The sudden event super graph Among them, the set of supernodes in the event hypergraph , For supernodes, the supernode set includes the entity types of the event and the subscript. N Represents the total number of supernodes; the set of superedges in the event hypergraph. , A hyperedge is a node consisting of a supernode, representing a complete burst event. The subscript... M This represents the total number of superedges.

[0008] In one embodiment, in S2, a local sudden event hypergraph is segmented from the sudden event hypergraph according to the time slice and used as the feature graph of the current stage; Feature map of the previous stage ,in, Represents the current set of supernodes. Represents the current set of superedges. t It belongs to a specific month in the dataset. n Indicates time slice granularity. T Indicates a time slice.

[0009] In one embodiment, during S2, when extracting the text features of the supernodes in the local burst event hypergraph, The text features of nodes in the local event hypergraph are vectorized to obtain the corresponding text feature representations. Entities at different nodes are represented by corresponding category features. ; Then, a multilayer perceptron network is used to map the text feature representation and category feature representation of each node in the local burst event hypergraph to a common style space of a specified size. The text feature representation and category feature representation of each node in the common style space are concatenated as the initial features of the nodes in the local burst event hypergraph. ; Where MLP() represents a multilayer perceptron network, Operators indicate concatenation.

[0010] In one embodiment, S4, the state information of all nodes in the hyperedge of the simple graph model is aggregated to the key node. Each node in the simple graph updates its own state information by aggregating the state information of surrounding nodes and its own state information. The average value of all aggregated node features is taken as the feature representation of the simple graph. The feature representation of the simple graph is linearly transformed through a fully connected layer and then input into a softmax layer for classification to obtain different classification probabilities. The specific method is as follows: S41 aggregates the state information of all nodes in the hyperedge of the simple graph model to the key node. Each node in the simple graph updates its own state information by aggregating the state information of the surrounding nodes and its own state information. Nodes in a simple graph model After aggregating the state information of surrounding nodes, the updated state information of itself is represented as follows: in, Represents a node The state representation after k node interactions; and These are the weight matrix and the bias matrix, respectively. It is an adjacency matrix that reflects the connection relationships between nodes. When a node... With nodes When there is a connection between them 1. If not connected 0; and These represent the output weight matrix and the output offset matrix, respectively. and These represent the input weight matrix and the input offset matrix, respectively. S42, take the average of all aggregated node features as the feature representation of the simple graph; The characteristics of a simple graph are represented as follows: in, For simple graph feature representation vectors, These are the aggregated node features. This represents the number of nodes.

[0011] S43, representing the characteristics of a simple graph After linear transformation through a fully connected layer, the data is then input into a softmax layer for classification, resulting in different classification probabilities. in, This indicates that the predicted category is... iThe probability, h g express f g After linear transformation by a fully connected layer, the input is then... softmax The results obtained by classifying the layers b i Indicates belonging to a category i The bias, Indicates the number of categories. This represents the shared, learnable weight vector.

[0012] The present invention also proposes a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the intelligent prediction method for sudden events based on graph convolutional neural networks.

[0013] The present invention also proposes a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the intelligent prediction method for sudden events based on graph convolutional neural networks.

[0014] The advantages of this invention compared to the prior art are: (1) This invention uses a hypergraph model to represent the relationships between different event entities, which can better capture the higher-order correlations between entities and relationships in events; (2) This invention uses message propagation of graph neural convolutional networks to perform information aggregation and representation learning on the node representation of simple graphs, which can better aggregate the information of multiple event entities, thereby better representing the features of event entities; (3) This invention improves prediction accuracy by using fully connected layer linear transformation and softmax layer to classify graph feature representations for trend. Attached Figure Description

[0015] Figure 1 This is a flowchart of the intelligent prediction method for sudden events based on graph convolutional neural networks according to the present invention. Detailed Implementation

[0016] This invention proposes an intelligent prediction method for sudden events based on graph convolutional neural networks, drawing on deep learning and big data processing technologies. First, a hypergraph model is introduced to represent the relationships between multiple event entities, enabling better capture of potentially high-order complex relationships. Then, a local hypergraph of a given event is segmented by selecting different temporal granularities, and a hyperedge within this local hypergraph is transformed into a simpler graph. Next, the message propagation mechanism in the graph convolutional neural network is utilized, allowing each node in the graph to aggregate the state information of its neighbors. The feature representation of the graph is obtained by averaging the features of all aggregated nodes. Finally, the graph representation is input into a classifier to obtain the final trend level.

[0017] This invention transforms the prediction of emergency trends into a trend level classification problem, and proposes an intelligent prediction method for emergencies based on graph convolutional neural networks. The overall flowchart is as follows: Figure 1 As shown, the model's input is the event hypergraph, and its output is the trend level.

[0018] First, by fully exploring the spatiotemporal relationships and global information of event data, a hypergraph of sudden events with high-order correlation is constructed. Then, by selecting the time granularity, the local hypergraph of sudden events is segmented. At the same time, the hyperedges in the event hypergraph are transformed into simple graphs to enhance the coupling ability of the model. Furthermore, a graph convolutional neural network is used to iteratively update the node features. By inputting the node features into a classifier to classify the trend level, the trend of sudden events is finally predicted.

[0019] The steps of the intelligent prediction method for sudden events based on graph convolutional neural networks in this invention are as follows: Step 1: 1-1) First, event extraction techniques are used to extract structured event information data, including event description, time, location, participants, and event type, represented as quadruples, including head entity, relation, tail entity, and event stamp. To ensure the accuracy of event extraction, within a defined domain, pattern matching-based event extraction techniques are generally used to extract events from the original unstructured text data. Domain experts determine the event types and formulate event description templates. Based on the patterns defined by the experts, structured event information data is extracted, and a dataset is constructed. The dataset is then divided into training, validation, and test sets.

[0020] 1-2) Generate an event hypergraph based on the constructed training set The generated event hypergraph is then stored in a graph database. Let be the set of supernodes in the hypergraph. For supernodes, subscripts NIt represents the total number of supernodes, including seven event entity types: time, location, event subject, event object, event type, event source, and event sentence. Let be the set of hyperedges in the hypergraph. For each hyperedge, a sub-edge represents a complete event. M This represents the total number of hyperedges. For example, a nuclear test by a country that triggered an artificial earthquake event is represented as a hyperedge. By supernode composition.

[0021] Step Two: 2-1) The sudden event trends studied in this invention are generally predicted on a monthly basis. The constructed sudden event hypergraph contains complete time periods, while the prediction of sudden events in a certain stage is only related to events that occurred in the previous month or several months. Therefore, it is necessary to segment a local hypergraph from the sudden event hypergraph according to time slices as the feature map of the current stage. .in, Represents the current set of supernodes. Represents the current set of superedges. t It belongs to a specific month in the dataset. n Indicates time slice granularity. T This represents a time slice. Taking the prediction of the trend level for April 2022 as an example, let's assume the time slice granularity... n The value is 3, which means retrieving all supernodes and superedges whose event time attribute is between January 2022 and March 2022 from the event hypergraph and generating a local event hypergraph.

[0022] 2-2) In the hypergraph of local emergencies, all hypernodes are described in Chinese. Therefore, the bag-of-words model is used to extract the text features of the nodes. Based on the event descriptions of different nodes in the event dataset, word segmentation is performed, and the average of the segmented word vectors is taken. For attribute nodes that directly form words, word vectors are directly obtained. Out-of-vocabulary words are replaced with zero vectors. The vectorized representation of the node text features in the hypergraph is as follows: Furthermore, considering the differences in entity categories across different nodes, entities in different nodes are represented using different category features. Then, a multilayer perceptron (MLP) network is used to map the text feature representation and category feature representation of each node in the hypergraph to a common style space of size d. The text feature representation and visual feature representation of each node in this style space are concatenated as the initial feature representation of the hypergraph node. The initial state information of each node in the common space is represented as follows: Where MLP() represents a multilayer perceptron network, "Operators represent concatenation. The Multilayer Perceptron (MLP) network is a three-layer architecture consisting of one LeakyReLU nonlinear layer and two tanh nonlinear layers."

[0023] Step 3: To better capture the higher-order relationships between event nodes, any hyperedge in the hypergraph is randomly selected. By performing an inner product of the initial states of different nodes in the hyperedge, the two nodes with the greatest differences are obtained as key nodes, and then connected to the remaining nodes in the hyperedge to obtain a simple graph model.

[0024] Step Four: 4-1) After obtaining the simple graph model, the message propagation mechanism in graph convolutional neural networks is used to aggregate the state information of all nodes in the hyperedge to the key node. Each node in the simple graph updates its own state information by aggregating the state information of its surrounding nodes and its own state information. After each node interaction, the nodes in the simple graph model... Represented as: in, Supernode The state representation after k node interactions. and These are the weight matrix and the bias matrix, respectively, which can help the model extract useful propagation information more effectively. It is an adjacency matrix that reflects the connection relationships between nodes. When the supernode... With supernode When there is a connection between them =1, if not connected =0. Clearly, the weight matrix... and bias matrix This determines the interaction between nodes. And by utilizing the same... and Simulating interactions between different nodes is insufficient for modeling flexible and complex interactions. Therefore, a unique weight matrix must be provided for each interaction. However, simply assigning a unique weight and bias matrix to each interaction consumes too much parameter space and runtime. To reduce time and space complexity, this invention assigns an input weight and bias matrix, and an output weight and bias matrix, to each node. When a node... Send its status information to the node Then, this status information is first passed through The output matrix and offset matrix are transformed, and then passed through the nodes. The input matrix is ​​transformed, and then the nodes... Accept this status information. Node The new state is represented as: in, and These represent the output weight matrix and the output offset matrix, respectively. and These represent the input weight matrix and the input offset matrix, respectively.

[0025] 4-2) The graph convolutional network yields the feature representation of each node, while the final task is to classify the graph, thus requiring a representation of the graph itself. The feature representation of this simple graph is obtained by averaging the features of all aggregated node features: in, For simple graph feature representation vectors, These are the aggregated node features. This represents the number of nodes.

[0026] 4-3) The obtained graph feature representation is linearly transformed through a fully connected layer before being input. softmax The layers are classified to obtain different classification trend probabilities, as shown in the following formula: in, This indicates that the predicted category is... i The probability, h g express f g After linear transformation by a fully connected layer, the input is then... softmax The results obtained by classifying the layers b i Indicates belonging to a category i The bias, Indicates the number of categories. This represents the shared, learnable weight vector.

[0027] Step 5: Using the classification probability of each simple graph as input, the prediction loss of a single sample is calculated using the cross-entropy function, and the learnable parameters in the entire model are adjusted based on the prediction loss of all samples in a batch.

[0028] Step Six: Repeat steps three through five until all batches of the training set have participated in model training and the specified number of iterations has been reached. Then, input the processed emergency event data test set into the trained emergency event trend prediction model for prediction, thereby achieving emergency event trend prediction.

[0029] The present invention also proposes a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the intelligent prediction method for sudden events based on graph convolutional neural networks.

[0030] The present invention also proposes a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the intelligent prediction method for sudden events based on graph convolutional neural networks.

[0031] This invention utilizes deep learning and big data processing technologies, introducing graph convolutional neural networks to predict sudden events. The intelligent prediction model built using graph convolutional networks significantly improves prediction accuracy by fully leveraging the semantic relationships between event elements.

Claims

1. A method for intelligent prediction of sudden events based on graph convolutional neural networks, characterized in that, It includes: S1. First, event information data is extracted from text data using event extraction technology. The event information data is represented in the form of quadruples to obtain structured event information data. A dataset is constructed based on the structured event information data, and the dataset is divided into training set and test set. An emergency event hypergraph is generated based on the training set. The emergency event hypergraph includes a set of supernodes and a set of superedges. The set of nodes includes the entity types of the emergency events, and each superedge represents a complete emergency event. S2, segment the local sudden event hypergraph from the sudden event hypergraph according to the time slice, and use it as the feature graph of the current stage; Extracting text features from supernodes in a local burst event hypergraph: The event description statements from different nodes in the event dataset are segmented into words, and the average value of the word vectors of the segmented text is taken. For the remaining attribute nodes that form words, word vectors are obtained. For out-of-vocabulary words, zero vectors are used instead. The text features of the nodes in the local sudden event hypergraph are vectorized to obtain the corresponding text feature representations. The entities of different nodes are set as the corresponding category feature representations. Then, a multilayer perceptron network is used to map the text feature representation and category feature representation of each node in the local burst event hypergraph to a common style space of a specified size. The text feature representation and category feature representation of each node in the common style space are concatenated as the initial features of the local burst event hypergraph node. S3. Randomly select any hyperedge in the hypergraph of a local sudden event, perform an inner product of the initial features of different nodes in the hyperedge, and obtain the two nodes with the greatest difference as key nodes. Connect the key nodes with the remaining nodes in the hyperedge to obtain a simple graph model. S4, aggregate the state information of all nodes in the hyperedge of the simple graph model to the key node. Each node in the simple graph updates its own state information by aggregating the state information of the surrounding nodes and its own state information. Take the average value of all aggregated node features as the feature representation of the simple graph. Linearly transform the feature representation of the simple graph through a fully connected layer and then input it into a softmax layer for classification to obtain different classification probabilities. S5 takes the classification probability of each simple graph model as input, uses the cross-entropy function to calculate the prediction loss of a single sample, and adjusts the learnable parameters in the entire simple graph model based on the prediction loss of all samples in a batch. S6. Repeat steps S3 to S5 until all batches of the training set have participated in the training of the simple graph model and reached the specified number of iterations to obtain the emergency event trend prediction model; input the test set into the trained emergency event trend prediction model for prediction to realize emergency event trend prediction.

2. The method according to claim 1, characterized in that, In step S1, the event information data includes event description, time, location, participants, and event type; the event information data is represented in the form of a quadruple, which includes: head entity, relation, tail entity, and event stamp.

3. The method according to claim 1, characterized in that, In step S1, a sudden event hypergraph is generated based on the training set. The sudden event hypergraph includes a set of supernodes and a set of superedges. The set of nodes includes the entity types of sudden events, and each superedge represents a complete sudden event. Specifically: The sudden event super graph Among them, the set of supernodes in the event hypergraph , For supernodes, the supernode set includes the entity types of the event and the subscript. N Represents the total number of supernodes; the set of superedges in the event hypergraph. , A hyperedge is a node consisting of a supernode, representing a complete burst event. The subscript... M This represents the total number of superedges.

4. The method according to claim 1, characterized in that, In S2, a local sudden event hypergraph is segmented from the sudden event hypergraph according to the time slice, and used as the feature graph of the current stage; Feature map of the previous stage ,in, Represents the current set of supernodes. Represents the current set of superedges. t It belongs to a specific month in the dataset. n Indicates time slice granularity. T Indicates a time slice.

5. The method according to claim 1, characterized in that, In S2, when extracting the text features of supernodes in the local burst event hypergraph, The text features of nodes in the local event hypergraph are vectorized to obtain the corresponding text feature representations. Entities at different nodes are represented by corresponding category features. ; Then, a multilayer perceptron network is used to map the text feature representation and category feature representation of each node in the local burst event hypergraph to a common style space of a specified size. The text feature representation and category feature representation of each node in the common style space are concatenated as the initial features of the nodes in the local burst event hypergraph. ; Where MLP() represents a multilayer perceptron network, Operators indicate concatenation.

6. The method according to claim 1, characterized in that, S4 aggregates the state information of all nodes in the hyperedge of the simple graph model to the key node. Each node in the simple graph updates its own state information by aggregating the state information of its surrounding nodes and its own state information. The average value of all aggregated node features is taken as the feature representation of the simple graph. The feature representation of the simple graph is linearly transformed through a fully connected layer and then input into a softmax layer for classification to obtain different classification probabilities. The specific method is as follows: S41 aggregates the state information of all nodes in the hyperedge of the simple graph model to the key node. Each node in the simple graph updates its own state information by aggregating the state information of the surrounding nodes and its own state information. Nodes in a simple graph model After aggregating the state information of surrounding nodes, the updated state information of itself is represented as follows: in, Represents a node The state representation after k node interactions; and These are the weight matrix and the bias matrix, respectively. It is an adjacency matrix that reflects the connection relationships between nodes. When a node... With nodes When there is a connection between them 1. If not connected 0; and These represent the output weight matrix and the output offset matrix, respectively. and These represent the input weight matrix and the input offset matrix, respectively. S42, take the average of all aggregated node features as the feature representation of the simple graph; The characteristics of a simple graph are represented as follows: in, For simple graph feature representation vectors, These are the aggregated node features. This represents the number of nodes. S43, representing the characteristics of a simple graph After linear transformation through a fully connected layer, the data is then input into a softmax layer for classification, resulting in different classification probabilities. in, This indicates that the predicted category is... i The probability, h g express f g The result is obtained by linear transformation through a fully connected layer and then inputting it into a softmax layer for classification. b i Indicates belonging to a category i The bias, Indicates the number of categories. This represents the shared, learnable weight vector.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method according to any one of claims 1 to 6.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 6.

Images