CN117974340A - Social media event detection method combining deep learning classification and graph clustering - Google Patents

Abstract

The invention discloses a social media event detection method combining deep learning classification and graph clustering, which is used for constructing and obtaining a message heterogram based on a message text and extracted characteristic information; constructing a plurality of shared characteristic edges in the message pairs, and constructing a multi-relation message graph; obtaining the similarity of the message pairs by using a deep learning classification model; if the similarity of the message pairs reaches a preset threshold, constructing an edge in the message pairs, and constructing to obtain a message isomorphic diagram; and clustering the message isomorphic graphs as input of a graph clustering algorithm to obtain a social media event detection result. The method avoids the process of representing the social messages as vectors, takes heterogeneous association of the message pairs and message texts as input, judges whether the message pairs belong to the same event by utilizing the deep learning classification model difference, constructs a message isomorphic graph by the prediction result of the message pairs, discovers the social message clusters with close association as social events by utilizing a graph clustering algorithm, and is used for massive social media event detection tasks.

Description

Social media event detection method combining deep learning classification and graph clustering

Technical Field

The present invention relates to the fields of natural language processing technology and text mining technology, and in particular to a social media event detection method combining deep learning classification and graph clustering.

Background technique

With the development of the Internet, social media platforms have changed people's lifestyles and become their main source of information. Social media is significantly more sensitive to the spread of information and the discovery of new events than traditional media. Therefore, it is particularly important to deeply analyze social media text information and discover social media events.

Deep learning social media event detection is the current mainstream method. This type of method uses deep neural networks to learn social message vector representations and discovers social media events through distance or density clustering algorithms. However, due to the short text of social messages and sparse word co-occurrence, it is difficult to use deep learning models to represent social messages of the same event as distance-similar vectors, and the vectors of different events are far away from each other. Event clustering is an important step in social media event detection, which aims to organize huge social media data into a collection of related content about the same event, so as to help users better understand and track the development of events. Most clustering algorithms used in event detection tasks are mainly based on word features, and external features are introduced to solve the problem of high-dimensional sparsity in short text clustering. However, excessive focus on word features will lead to excessive influence of noise and outliers, and it is impossible to find cluster centers in streaming data, which affects clustering performance. For streaming data in social media, the currently commonly used clustering algorithms have a high dependence on order and high computational overhead. They are inefficient when processing massive social media texts, which greatly affects the performance of the model.

Summary of the invention

To this end, the present invention provides a social media event detection method that combines deep learning classification and graph clustering to solve the problems that the existing deep learning social media event detection method represents the social messages of the same event as distance-similar vectors, which makes it difficult for the social message vectors of different events to be far away from each other, and the clustering algorithm focuses too much on the features of words and is inefficient when processing massive social media texts.

In order to achieve the above object, the present invention provides the following technical solutions:

According to a first aspect of an embodiment of the present invention, a social media event detection method combining deep learning classification and graph clustering is proposed, the method comprising:

Feature information is extracted from message texts in social media data streams, message texts and feature information are used as nodes, and message texts and their extracted feature information are connected to form edges to construct a message heterogeneous graph;

Based on the message heterogeneous graph, two message text nodes are randomly selected to construct a message pair, and multiple shared feature edges are constructed in each message pair to obtain a multi-relation message graph;

Based on the multi-relation message graph, using a deep learning classification model to obtain similarities between message pairs;

Determine whether the similarity of the message pair reaches a preset threshold. If the similarity of the message pair reaches the preset threshold, construct an edge between the two message text nodes constituting the message pair to obtain a message isomorphism graph.

The constructed message isomorphism graph is used as the input of the graph clustering algorithm for clustering, and the clustering results are used as the detected social media events.

Furthermore, the method further comprises:

Obtain a social media text dataset and divide it into multiple data blocks according to the publishing time to simulate the streaming properties of social media messages for continuous model training.

Furthermore, feature information is extracted from the message text in the social media data stream, the message text and feature information are used as nodes, and the message text and the extracted feature information are connected to form edges to construct a message heterogeneous graph, which specifically includes:

Given a single message text , extract the characteristic information of the message from the text, the characteristic information includes entity information, user information, topic tag information and time information, defined as ;

Defining a message heterogeneous graph ,in A collection of nodes representing the social media message text itself and various types of event-related feature information therein. Represents the set of edges between messages and corresponding features.

Furthermore, based on the message heterogeneous graph, two message text nodes are randomly selected to construct message pairs, and multiple shared feature edges are constructed in each message pair to construct a multi-relation message graph, which specifically includes:

Define the multi-relation message graph as ;

in Is a collection of message text nodes , is the number of message text nodes, each node has a different feature representation and has a representation of dimensional feature vector, representing the message node type as , , , and Represent messages, users, entities, topic tags and time information respectively. The set of all node features is expressed as , randomly sample nodes from other message texts Build into message pairs ;

When the message texts share different types of feature information, edges belonging to different shared feature relationships are established respectively. Yes, the message is The edges in can be associated with multiple shared feature relationships:

in It is a submatrix of the adjacency matrix of the message heterogeneous graph. The rows represent all feature information nodes and the columns represent the nodes belonging to the relationship. All message nodes, is the transpose of the matrix.

Further, based on the multi-relation message graph, the similarity of the message pairs is obtained using a deep learning classification model, specifically including:

The message text is preprocessed, including word segmentation and stop word removal. Then, the message text is word embedded using the BERT pre-trained language model. Dimensional word vector Indicative words, ;

Vector message text Embed the message text The embedding vector , message text The embedding vector , concatenate the embedding vectors of the two message texts through the encoding layer to get the encoding vector of the message pair:

in, Represents the concatenation operation of vectors;

The obtained encoding vector is passed through a linear transformation layer to obtain a low-dimensional vector and through Activate the function to get the message text and Similarity ,in :

in, , , and represents the parameters in the model, Represents the activation function.

Further, it is determined whether the similarity of the message pair reaches a preset threshold. If the similarity of the message pair reaches the preset threshold, an edge is constructed between the two message text nodes constituting the message pair to construct a message isomorphism graph, which specifically includes:

The message isomorphic graph still retains all the common features of the message heterogeneous graph.

in is the adjacency matrix of the message isomorphism graph, where is the total number of message nodes in the graph, Indicates the node type, is a submatrix of the adjacency matrix in the message heterogeneous graph, containing Type of row and The type of the column, is the transpose of the matrix; if the message text node and the message text node Connect to some type At node time, will be greater than or equal to 1, then will be equal to 1;

For message text and message text The message pairs composed of Whether the preset threshold is reached , when the similarity reaches the preset threshold When an edge is constructed After judging all the message pairs, the initial message isomorphism graph is formed. ,in is a collection of message text nodes, is the edge set.

Furthermore, the constructed message isomorphism graph is used as the input of the graph clustering algorithm for clustering, and the clustering results are used as the detected social media events, specifically including:

The community is expanded according to the transitivity of feature associations between messages, and then multiple rounds of games are conducted between message nodes to select a better community, ultimately achieving a stable state of the community. The clustering results obtained by community division are used as detected social media events.

Furthermore, the community is expanded based on the transitivity of the feature associations between messages, and then multiple rounds of games are conducted between message nodes to select a better community, and finally a stable state of the community is achieved, including:

Expanding the community scale, if there are two edges between three points, a half triangle is formed. According to the ternary closure principle, two nodes are also related when they are related to a common node. It is considered that there is an implicit association between the three points, that is, the three points belong to the same community.

Multiple rounds of iterations are carried out. In each round of iteration, all nodes make better choices based on the current community division. There are three choices: do not change the current community; leave the current community and do not join any other community; leave the current community and join another community;

After multiple rounds of games, the algorithm reaches Nash equilibrium, that is, all nodes join the community that they are most satisfied with. The algorithm can be judged whether it has reached Nash equilibrium by setting a threshold.

According to a second aspect of an embodiment of the present invention, a social media event detection system combining deep learning classification and graph clustering is proposed, the system comprising:

A message heterogeneous graph construction module is used to extract feature information from message texts in social media data streams, use message texts and feature information as nodes, and connect message texts with their extracted feature information to form edges, thereby constructing a message heterogeneous graph;

A multi-relation message graph construction module is used to randomly select two message text nodes to construct message pairs based on the message heterogeneous graph, and to construct multiple shared feature edges in each message pair to obtain a multi-relation message graph;

A deep learning classification module, used to obtain the similarity of message pairs based on the multi-relation message graph using a deep learning classification model;

A message isomorphism graph construction module is used to determine whether the similarity of a message pair reaches a preset threshold. If the similarity of the message pair reaches the preset threshold, an edge is constructed between the two message text nodes constituting the message pair to construct a message isomorphism graph;

The graph clustering module is used to cluster the constructed message isomorphism graph as the input of the graph clustering algorithm, and the clustering result is used as the detected social media event.

According to a third aspect of an embodiment of the present invention, an electronic device is provided, the device comprising: a processor and a memory;

The memory is used to store one or more program instructions;

The processor is used to run one or more program instructions to execute the steps of a social media event detection method combining deep learning classification and graph clustering as described in any of the above items.

According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is proposed, on which a computer program is stored. When the computer program is executed by a processor, the steps of a social media event detection method combining deep learning classification and graph clustering as described in any one of the above items are implemented.

The present invention proposes a social media event detection method combining deep learning classification and graph clustering, which extracts feature information from message texts in a social media data stream, uses message texts and feature information as nodes, connects message texts with their extracted feature information to form edges, and constructs a message heterogeneous graph; based on the message heterogeneous graph, randomly selects two message text nodes to construct message pairs, and constructs multiple shared feature edges in each pair of message pairs to construct a multi-relation message graph; based on the multi-relation message graph, uses a deep learning classification model to obtain the similarity of the message pairs; determines whether the similarity of the message pairs reaches a preset threshold, and if the similarity of the message pairs reaches the preset threshold, constructs an edge between the two message text nodes that constitute the message pair to construct a message isomorphism graph; uses the constructed message isomorphism graph as the input of a graph clustering algorithm for clustering, and uses the clustering result as the detected social media event. The present invention combines a deep learning classification model with a graph clustering algorithm, avoids the process of representing social messages as vectors, takes the heterogeneous associations of message pairs and message texts as input, uses a deep learning classification model to differentially determine whether message pairs belong to the same event, and constructs a message isomorphism graph through the prediction results of the message pairs, and uses a graph clustering algorithm to discover closely related social message clusters as social events; it has been verified that the performance indicators of the present invention are better than those of the baseline model; the present invention ensures the computational difficulty and complexity, and can be used for massive social media event detection tasks.

BRIEF DESCRIPTION OF THE DRAWINGS

In order to more clearly illustrate the implementation methods of the present invention or the technical solutions in the prior art, the drawings required for the implementation methods or the description of the prior art are briefly introduced below. Obviously, the drawings in the following description are only exemplary, and for ordinary technicians in this field, other implementation drawings can be derived from the provided drawings without creative work.

The structures, proportions, sizes, etc. illustrated in this specification are only used to match the contents disclosed in the specification so as to facilitate understanding and reading by persons familiar with the technology. They are not used to limit the conditions under which the present invention can be implemented, and therefore have no substantial technical significance. Any structural modification, change in proportion or adjustment of size shall still fall within the scope of the technical contents disclosed in the present invention without affecting the effects and purposes that can be achieved by the present invention.

FIG1 is a flow chart of a social media event detection method combining deep learning classification and graph clustering provided by an embodiment of the present invention;

FIG2 is a schematic diagram of a social media event detection method combining deep learning classification and graph clustering provided by an embodiment of the present invention;

FIG3 is a schematic diagram of the structure of a social media event detection system combining deep learning classification and graph clustering provided by an embodiment of the present invention.

Detailed ways

The following is a description of the implementation of the present invention by specific embodiments. People familiar with the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

The first embodiment of the present invention provides a social media event detection method combining deep learning classification and graph clustering, which is described below with reference to FIG. 1 and FIG. 2 .

As shown in FIG. 1 , in step S101 , feature information is extracted from message texts in a social media data stream, the message texts and feature information are used as nodes, and the message texts and the extracted feature information are connected to form edges to construct a message heterogeneous graph.

In this embodiment, social media messages are divided into data blocks according to the release time to simulate the streaming properties of social media messages for continuous training, and entity information, user information, tag information and time information in the messages are extracted as features to construct a message heterogeneous graph.

Specifically, the following four types of features are extracted from messages to maximize the use of social media data, and the extracted features are further processed in a unified way to construct a message heterogeneous graph. , extracting the entity information, user information, tag information and time information of the message from the text, defined as , taking these features and the social media message text itself as nodes, The edges are formed between the feature information extracted from it, forming a message heterogeneous graph.

Defining a message heterogeneous graph ,in A collection of nodes representing the social media message text itself and various types of event-related feature information therein. Represents the set of edges between messages and corresponding features.

As shown in FIG. 1 , in step S102 , based on the message heterogeneous graph, two message text nodes are randomly selected to construct a message pair, and a plurality of shared feature edges are constructed in each message pair to obtain a multi-relation message graph.

In this embodiment, the features in the message heterogeneous graph are mapped to message pairs to prevent the loss of heterogeneous feature information between event elements of different types. Multiple information feature edges are constructed in each pair of messages to form a multi-relationship message graph for preserving rich social media message features.

Specifically, define the multi-relation message graph as ,in is a collection of nodes in a data block , is the number of nodes, each node has a different feature representation and is represented as of dimensional feature vector, representing the node type as , , , and Represent message, user, entity, tag and time information respectively, and the set of all node features is expressed as , randomly sample from other messages to form a message pair .

When messages share different types of feature elements, edges belonging to different feature relationships are established respectively. Yes, the message is The edges between them can be associated with multiple relationships, with a total of four different types of relationships;

in It is a submatrix of the adjacency matrix of the message heterogeneous graph. The rows represent all feature information nodes and the columns represent the nodes belonging to the relationship. All message nodes, is the transpose of the matrix, taking the smaller of the two.

As shown in FIG. 1 , in step S103 , based on the multi-relation message graph, a deep learning classification model is used to obtain the similarity of message pairs.

In this embodiment, the message text and features are processed using the BERT pre-trained language model to obtain the embedded vector of the message text, and the embedded vector of the text is randomly spliced to form a message pair, which is then sent to the encoding layer to obtain the encoding vector of the message pair. The obtained encoding vector of the message pair is used as input, and a low-dimensional vector is obtained through the linear transformation layer. The activation function calculates the similarity between message pairs.

Specifically, before the text is input, it is first preprocessed to perform word segmentation and stop word removal operations, and the message text is word embedded using the BERT pre-trained language model. Dimensional word vector Indicative words, . For message text vector Embed to get the message text The embedding vector , message text The embedding vector , concatenate the embedding vectors of the two texts and pass them through the encoding layer to get the encoding vector of the message pair:

in, Represents a vector concatenation operation.

The obtained encoded vector is passed through a linear transformation layer to obtain a low-dimensional vector and through Activate the function to get the message text and Similarity ,in :

in, , , and represents the parameters in the model, Represents the activation function.

As shown in FIG. 1 , in step S104 , it is determined whether the similarity of the message pair reaches a preset threshold. If the similarity of the message pair reaches the preset threshold, an edge is constructed between the two message text nodes constituting the message pair to construct a message isomorphism graph.

In this embodiment, in order to enhance the degree of association between messages, a variety of rich information features are merged into a unique feature, and the message heterogeneous graph is mapped to a message isomorphic graph to implement a graph clustering algorithm. The similarity of message pairs is used as the only basis for constructing a message isomorphic graph, which only contains message nodes, and there is only one edge between messages that shares all common elements.

Specifically, the message isomorphic graph still retains all the common features of the message heterogeneous graph.

in is the adjacency matrix of the message isomorphism graph, where is the total number of message nodes in the graph, Indicates the node type, is a submatrix of the adjacency matrix in the message heterogeneous graph, containing Type of row and The type of the column, is the transpose of the matrix, taking the smaller of the two. If the message and Message Connect to some type At node time, will be greater than or equal to 1, then will be equal to 1.

For Messages and Message , to determine the similarity between the two , when the similarity reaches the threshold When an edge is constructed After judging all the event pairs, the initial graph is formed. ,in is a collection of nodes, is the edge set.

As shown in FIG. 1 , in step S105 , the constructed message isomorphism graph is used as the input of the graph clustering algorithm for clustering, and the clustering result is used as the detected social media event.

In this embodiment, the constructed message isomorphism graph is used as the input of the graph clustering algorithm for clustering, and the community is expanded according to the transitivity of the feature associations between messages. After that, multiple rounds of game are carried out between the message nodes to select a better community, and finally a stable state of the community is achieved. The community division results are used as clustering results to realize social media event detection.

The specific process includes expanding the community size. If there are two edges between three points, a half triangle is formed. According to the ternary closure principle, two nodes are also related when they are related to a common node. It is considered that there is an implicit association between the three points, that is, the three points belong to the same community.

Multiple rounds of iterations are carried out. In each round of iteration, all nodes can make a better choice for themselves based on the current community division. There are three choices: do not change the current community; leave the current community and do not join any other community; leave the current community and join another community;

After multiple rounds of games, the algorithm reaches Nash equilibrium, that is, all nodes join the community that they are most satisfied with, and a threshold is set to judge the algorithm status.

In order to illustrate the effect of the present invention, the present invention is compared with the existing methods, and experiments are conducted on a large-scale public data set Event2012, which has a release time of about 29 days and uses the messages in the first week to form the initial block. , and the others are divided into 21 message blocks according to the release time. The evaluation index is consistent with the comparison method. Normalized mutual information (NMI), adjusted mutual information (AMI) and adjusted Rand index (ARI) are used as clustering result evaluation indicators. The threshold for the end of the clustering algorithm is 0.1%. The experimental results are shown in Tables 1 to 3 respectively:

Table 1 Incremental evaluation NMI scores

Blocks Word2vec LDA WMD BERT BiLSTM PP-GCN EventX KPGNN FinEventd ours M1 .19±.00 .11±.00 .32±.00 .36±.00 .24±.00 .23±.00 .36±.00 .39±.00 .84±.01 0.47±.00↓.36 M2 .50±.00 .27±.01 .71±.00 .78±.00 .50±.00 .57±.02 .68±.00 .79±.01 .84±.00 0.92±.01↑.08 M3 .39±.00 .28±.00 .67±.00 .75±.00 .39±.00 .55±.01 .63±.00 .76±.00 .89±.00 0.87±.01↓.01 M4 .34±.00 .25±.00 .50±.00 .60±.00 .40±.00 .46±.01 .63±.00 .67±.00 .71±.01 0.84±.00↑.13 M5 .41±.00 .26±.00 .61±.00 .72±.00 .41±.00 .48±.01 .59±.00 .73±.01 .83±.00 0.84±.00↑.01 M6 .53±.00 .32±.00 .61±.00 .78±.00 .50±.00 .57±.01 .70±.00 .82±.01 .83±.00 0.94±.00↑.11 M7 .25±.00 .18±.01 .46±.00 .54±.00 .33±.00 .37±.00 .51±.00 .55±.01 .73±.01 0.64±.02↓.08 M8 .46±.00 .37±.01 .67±.00 .79±.00 .49±.00 .55±.02 .71±.00 .80±.00 .87±.02 0.90±.00↑.03 M9 .35±.00 .34±.00 .55±.00 .70±.00 .43±.00 .51±.02 .67±.00 .74±.02 .79±.01 0.88±.01 ↑.09 M10 .51±.00 .44±.01 .61±.00 .74±.00 .50±.00 .55±.02 .68±.00 .80±.01 .82±.01 0.95±.00↑.13 M11 .37±.00 .33±.01 .50±.00 .68±.00 .49±.00 .50±.01 .65±.00 .74±.01 .75±.00 0.90±.00↑.15 M12 .30±.00 .22±.01 .60±.00 .59±.00 .39±.00 .45±.01 .61±.00 .68±.01 .67±.01 0.81±.00↑.14 M13 .37±.00 .27±.00 .54±.00 .63±.00 .46±.00 .47±.01 .58±.00 .69±.01 .79±.00 0.84±.01 ↑.05 M14 .36±.00 .21±.00 .66±.00 .64±.00 .44±.00 .44±.01 .57±.00 .69±.00 .82±.00 0.79±.01↓.03 M15 .27±.00 .21±.00 .51±.00 .54±.00 .40±.00 .39±.01 .49±.00 .58±.00 .69±.01 0.85±.01 ↑.16 M16 .49±.00 .35±.01 .60±.00 .75±.00 .53±.00 .55±.01 .62±.00 .79±.01 .90±.01 0.91±.01↑.01 M17 .33±.00 .19±.00 .55±.00 .63±.00 .45±.00 .48±.00 .58±.00 .70±.01 .83±.00 0.84±.00↑.01 M18 .29±.00 .18±.00 .63±.00 .57±.00 .44±.00 .47±.01 .59±.00 .68±.02 .74±.01 0.85±.01 ↑.11 M19 .37±.00 .29±.01 .54±.00 .66±.00 .44±.00 .51±.02 .60±.00 .73±.01 .66±.01 0.87±.00↑.21 M20 .38±.00 .35±.00 .58±.00 .68±.00 .48±.00 .51±.01 .67±.00 .72±.02 .80±.00 0.85±.02 ↑.05 M21 .31±.00 .19±.00 .58±.00 .59±.00 .41±.00 .41±.02 .53±.00 .60±.00 .74±.01 0.75±.01 ↑.01 AVG 0.3700 0.2671 0.5742 0.6533 0.4345 0.4771 0.8024 0.6876 0.7876 0.8338

Table 2 Incremental evaluation AMI scores

Blocks Word2vec LDA WMD BERT BiLSTM PP-GCN EventX KPGNN FinEventd ours M1 .08±.00 .08±.00 .30±.00 .34±.00 .12±.00 .21±.00 .06±.00 .37±.00 .84±.01 0.40±.01↓.43 M2 .41±.00 .20±.01 .69±.00 .76±.00 .41±.00 .55±.02 .29±.02 .78±.01 .84±.01 0.89±.00↑.05 M3 .31±.00 .22±.01 .63±.00 .73±.00 .31±.00 .52±.01 .18±.01 .74±.00 .89±.01 0.82±.00↓.06 M4 .24±.00 .17±.00 .45±.00 .55±.00 .30±.00 .42±.01 .19±.01 .64±.01 .69±.00 0.77±.00↑.08 M5 .33±.00 .21±.00 .57±.00 .71±.00 .33±.00 .46±.01 .14±.00 .71±.01 .82±.00 0.79±.02↓.03 M6 .40±.00 .20±.00 .57±.00 .74±.00 .36±.00 .52±.02 .27±.00 .79±.01 .82±.02 0.90±.00↑.08 M7 .13±.00 .12±.01 .46±.00 .50±.00 .20±.00 .34±.00 .13±.00 .51±.01 .72±.00 0.55±.03↓.14 M8 .33±.00 .24±.01 .63±.00 .75±.00 .35±.00 .49±.02 .21±.00 .76±.01 .87±.01 0.83±.00↓.03 M9 .24±.00 .24±.00 .46±.00 .66±.00 .32±.00 .46±.02 .19±.00 .71±.02 .78±.01 0.81±.01↑.03 M10 .39±.00 .36±.01 .57±.00 .70±.00 .39±.00 .51±.02 .24±.00 .78±.01 .81±.00 0.92±.00↑.11 M11 .26±.00 .25±.01 .42±.00 .65±.00 .37±.00 .46±.01 .24±.00 .71±.01 .74±.00 0.86±.00↑.12 M12 .23±.00 .16±.01 .58±.00 .56±.00 .32±.00 .42±.01 .16±.00 .66±.01 .67±.02 0.72±.01 ↑.05 M13 .23±.00 .19±.00 .50±.00 .59±.00 .31±.00 .43±.01 .16±.00 .67±.01 .79±.00 0.80±.00↑.01 M14 .26±.00 .15±.00 .64±.00 .61±.00 .34±.00 .41±.01 .14±.00 .65±.00 .82±.01 0.72±.01↓.09 M15 .15±.00 .13±.00 .47±.00 .50±.00 .26±.00 .35±.01 .07±.00 .54±.00 .67±.01 0.80±.00↑.13 M16 .36±.00 .27±.01 .59±.00 .72±.00 .41±.00 .52±.01 .19±.00 .77±.01 .90±.01 0.88±.02↓.01 M17 .24±.00 .13±.00 .57±.00 .60±.00 .35±.00 .45±.00 .18±.00 .68±.01 .82±.01 0.79±.00↓.02 M18 .21±.00 .12±.00 .60±.00 .53±.00 .35±.00 .45±.01 .16±.00 .66±.02 .74±.00 0.81±.00↑.07 M19 .28±.00 .22±.01 .49±.00 .63±.00 .35±.00 .48±.02 .16±.00 .71±.01 .66±.01 0.83±.01 ↑.17 M20 .24±.00 .23±.00 .55±.00 .62±.00 .34±.00 .45±.02 .18±.00 .68±.02 .78±.00 0.76±.00↓.02 M21 .21±.00 .13±.00 .52±.00 .57±.00 .31±.00 .38±.02 .10±.00 .57±.00 .64±.01 0.69±.01↑.05 AVG 0.2633 0.1914 0.5362 0.6200 0.3238 0.4419 0.1733 0.6710 0.7767 0.7781

Table 3 Incremental evaluation ARI scores

Blocks Word2vec LDA WMD BERT BiLSTM PP-GCN EventX KPGNN FinEventd ours M1 .01±.00 .01±.00 .04±.00 .03±.00 .03±.00 .05±.00 .01±.00 .07±.01 .90±.00 0.11±.01↓.78 M2 .49±.00 .08±.00 .48±.00 .64±.00 .49±.00 .67±.03 .45±.02 .76±.02 .90±.01 0.91±.00↑.01 M3 .16±.00 .02±.01 .28±.00 .43±.00 .17±.00 .47±.01 .09±.01 .58±.00 .89±.01 0.78±.01↓.10 M4 .07±.00 .07±.00 .11±.00 .19±.00 .11±.00 .24±.01 .07±.01 .29±.01 .27±.01 0.56±.01 ↑.29 M5 .17±.00 .06±.00 .26±.00 .44±.00 .19±.00 .34±.00 .04±.00 .47±.03 .63±.02 0.76±.01 ↑.13 M6 .25±.00 .07±.01 .16±.00 .44±.00 .18±.00 .55±.03 .14±.00 .72±.03 .74±.00 0.90±.00↑.16 M7 .02±.00 .01±.00 .08±.00 .07±.00 .08±.00 .11±.02 .02±.00 .12±.00 .45±.01 0.15±.02↓.28 M8 .17±.00 .03±.00 .22±.00 .50±.00 .08±.00 .43±.04 .09±.00 .60±.01 .72±.01 0.79±.00↑.07 M9 .08±.00 .03±.01 .12±.00 .33±.00 .27±.00 .31±.02 .07±.00 .46±.02 .68±.00 0.63±.02↓.03 M10 .23±.00 .09±.02 .20±.00 .44±.00 .22±.00 .50±.07 .13±.00 .70±.06 .74±.01 0.88±.00↑.14 M11 .09±.00 .03±.01 .12±.00 .27±.00 .17±.00 .38±.02 .16±.00 .49±.03 .60±.01 0.93±.00↑.33 M12 .09±.00 .02±.01 .27±.00 .31±.00 .13±.00 .34±.03 .07±.00 .48±.01 .26±.00 0.60±.01↑.14 M13 .06±.00 .01±.00 .13±.00 .14±.00 .13±.00 .19±.01 .04±.00 .29±.03 .75±.02 0.73±.01↓.01 M14 .10±.00 .02±.00 .33±.00 .30±.00 .16±.00 .29±.01 .10±.00 .42±.02 .81±.01 0.58±.00↓.22 M15 .09±.00 .01±.00 .16±.00 .10±.00 .14±.00 .15±.00 .01±.00 .17±.00 .46±.00 0.86±.00 ↑.40 M16 .10±.00 .11±.01 .32±.00 .41±.00 .10±.00 .51±.03 .08±.00 .66±.05 .88±.01 0.85±.01↓.02 M17 .06±.00 .02±.00 .26±.00 .24±.00 .17±.00 .35±.03 .12±.00 .43±.05 .81±.01 0.70±.03↓.08 M18 .21±.00 .02±.00 .35±.00 .24±.00 .19±.00 .39±.03 .08±.00 .47±.04 .52±.01 0.76±.01 ↑.24 M19 .28±.00 .03±.00 .12±.00 .32±.00 .16±.00 .41±.02 .07±.00 .51±.03 .35±.01 0.75±.02 ↑.40 M20 .24±.00 .02±.01 .19±.00 .33±.00 .20±.00 .41±.01 .11±.00 .51±.04 .71±.01 0.67±.01↓.03 M21 .21±.00 .01±.01 .19±.00 .18±.00 .16±.00 .20±.03 .01±.00 .20±.01 .48±.00 0.51±.00↑.03 AVG 0.1514 0.0367 0.2090 0.3024 0.1681 0.3471 0.0933 0.4476 0.6452 0.6861

The experimental results show that compared with the baseline model FinEvent, the NMI value of the experimental group model increased by 4.62% on average, the AMI value increased by 0.14% on average, and the ARI value increased by 4.09% on average. It can be seen that the present invention pays attention to the graph structure and uses a graph clustering algorithm to mine the connection relationship in social media messages while ensuring the difficulty and complexity of calculation, which can better improve the performance of social media event detection.

Corresponding to the above-disclosed social media event detection method combining deep learning classification and graph clustering, an embodiment of the present invention further discloses a social media event detection system combining deep learning classification and graph clustering, as shown in FIG3 , which specifically includes:

A message heterogeneous graph construction module is used to extract feature information from message texts in social media data streams, use message texts and feature information as nodes, and connect message texts with their extracted feature information to form edges, thereby constructing a message heterogeneous graph;

A multi-relation message graph construction module is used to randomly select two message text nodes to construct message pairs based on the message heterogeneous graph, and to construct multiple shared feature edges in each message pair to obtain a multi-relation message graph;

A deep learning classification module, used to obtain the similarity of message pairs based on the multi-relation message graph using a deep learning classification model;

A message isomorphism graph construction module is used to determine whether the similarity of a message pair reaches a preset threshold. If the similarity of the message pair reaches the preset threshold, an edge is constructed between the two message text nodes constituting the message pair to construct a message isomorphism graph;

The graph clustering module is used to cluster the constructed message isomorphism graph as the input of the graph clustering algorithm, and the clustering results are used as the detected social media events.

It should be noted that for the detailed description of a social media event detection system combining deep learning classification and graph clustering provided in an embodiment of the present invention, reference can be made to the relevant description of a social media event detection method combining deep learning classification and graph clustering provided in an embodiment of the present application, which will not be repeated here.

In addition, an embodiment of the present invention also provides an electronic device, comprising: a processor and a memory; the memory is used to store one or more program instructions; the processor is used to run one or more program instructions to execute the steps of a social media event detection method combining deep learning classification and graph clustering as described above.

It should be noted that, for a detailed description of an electronic device provided in an embodiment of the present invention, reference can be made to the relevant description of a social media event detection method combining deep learning classification and graph clustering provided in an embodiment of the present application, which will not be repeated here.

In addition, an embodiment of the present invention further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the social media event detection method combining deep learning classification and graph clustering are implemented as described above.

It should be noted that, for a detailed description of a computer-readable storage medium provided in an embodiment of the present invention, reference can be made to the relevant description of a social media event detection method combining deep learning classification and graph clustering provided in an embodiment of the present application, which will not be repeated here.

In the embodiment of the present invention, the processor may be an integrated circuit chip having the signal processing capability. The processor may be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

The methods, steps and logic block diagrams disclosed in the embodiments of the present invention can be implemented or executed. The general processor can be a microprocessor or the processor can also be any conventional processor, etc. The steps of the method disclosed in the embodiments of the present invention can be directly embodied as a hardware decoding processor for execution, or can be executed by a combination of hardware and software modules in the decoding processor. The software module can be located in a mature storage medium in the field such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, etc. The processor reads the information in the storage medium and completes the steps of the above method in combination with its hardware.

The storage medium may be a memory, which may be, for example, a volatile memory or a nonvolatile memory, or may include both volatile and nonvolatile memory.

Among them, the non-volatile memory can be a read-only memory (ROM), a programmable ROM (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), or a flash memory.

The volatile memory may be a random access memory (RAM) which is used as an external cache. By way of example and not limitation, many forms of RAM are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM), and direct RAM bus random access memory (DRRAM).

The storage media described in the embodiments of the present invention are intended to include, but are not limited to, these and any other suitable types of memory.

Those skilled in the art will appreciate that in one or more of the above examples, the functions described in the present invention may be implemented using a combination of hardware and software. When software is used, the corresponding functions may be stored in a computer-readable medium or transmitted as one or more instructions or codes on a computer-readable medium. Computer-readable media include computer storage media and communication media, wherein communication media include any media that facilitates the transmission of computer programs from one place to another. Storage media may be any available media that can be accessed by a general or special purpose computer.

Although the present invention has been described in detail above by general description and specific embodiments, it is obvious to those skilled in the art that some modifications or improvements can be made to the present invention. Therefore, these modifications or improvements made without departing from the spirit of the present invention all belong to the scope of protection claimed by the present invention.

Claims (10)

1. A social media event detection method combining deep learning classification and graph clustering, characterized in that the method comprises: Feature information is extracted from message texts in social media data streams, message texts and feature information are used as nodes, and message texts and their extracted feature information are connected to form edges to construct a message heterogeneous graph; Based on the message heterogeneous graph, two message text nodes are randomly selected to construct a message pair, and multiple shared feature edges are constructed in each message pair to obtain a multi-relation message graph; Based on the multi-relation message graph, using a deep learning classification model to obtain similarities between message pairs; Determine whether the similarity of the message pair reaches a preset threshold. If the similarity of the message pair reaches the preset threshold, construct an edge between the two message text nodes constituting the message pair to obtain a message isomorphism graph. The constructed message isomorphism graph is used as the input of the graph clustering algorithm for clustering, and the clustering results are used as the detected social media events. 2. According to claim 1, a social media event detection method combining deep learning classification and graph clustering is characterized in that the method further comprises: Obtain a social media text dataset and divide it into multiple data blocks according to the publishing time to simulate the streaming properties of social media messages for continuous model training. 3. According to claim 1, a social media event detection method combining deep learning classification and graph clustering is characterized in that feature information is extracted from message texts in social media data streams, message texts and feature information are used as nodes, and message texts and their extracted feature information are connected to form edges to construct a message heterogeneous graph, specifically including: Given a single message text , extract the characteristic information of the message from the text, the characteristic information includes entity information, user information, topic tag information and time information, defined as/> ; Defining a message heterogeneous graph , where/> A node set representing the social media message text itself and various types of event-related feature information therein, /> Represents the set of edges between messages and corresponding features. 4. According to claim 3, a social media event detection method combining deep learning classification and graph clustering is characterized in that based on the message heterogeneous graph, two message text nodes are randomly selected to construct message pairs, and multiple shared feature edges are constructed in each pair of message pairs to construct a multi-relation message graph, which specifically includes: Define the multi-relation message graph as ; in Is a message text node collection/> ,/> is the number of message text nodes, each node has a different feature representation and is represented as/> /> dimension feature vector, representing the message node type as/> ,/> ,/> ,/> and/> Represent messages, users, entities, topic tags and time information respectively. The set of all node features is expressed as , randomly sample nodes from other message texts/> Constructed into message pairs/> ; When the message texts share different types of feature information, edges belonging to different shared feature relationships are established respectively. It is a message to/> The edges in can be associated with multiple shared feature relationships: Where/> It is a submatrix of the adjacency matrix of the message heterogeneous graph. The rows represent all feature information nodes and the columns represent the relationships./> All message nodes, /> is the transpose of the matrix. 5. According to claim 1, a social media event detection method combining deep learning classification and graph clustering is characterized in that based on the multi-relation message graph, a deep learning classification model is used to obtain the similarity of message pairs, specifically including: The message text is preprocessed, including word segmentation and stop word removal. Then, the message text is word embedded using the BERT pre-trained language model. Dimensional word vector/> Indicates word, /> ; Vector message text Embed the message text/> Embedding vector of , message text/> Embedding vector of , the embedding vectors of the two message texts are concatenated through the encoding layer to obtain the encoding vector of the message pair: Among them,/> Represents the concatenation operation of vectors; The obtained encoding vector is passed through a linear transformation layer to obtain a low-dimensional vector , and by /> Activate the function to get the message text/> and/> Similarity of/> , where/> : ,/> , where /> ,/> ,/> and/> Represents the parameters in the model, /> Represents the activation function. 6. According to claim 3, a social media event detection method combining deep learning classification and graph clustering is characterized in that it is determined whether the similarity of the message pair reaches a preset threshold. If the similarity of the message pair reaches the preset threshold, an edge is constructed between the two message text nodes constituting the message pair to construct a message isomorphism graph, which specifically includes: The message isomorphic graph still retains all the common features of the message heterogeneous graph. , where/> is the adjacency matrix of the message isomorphism graph, where/> is the total number of message nodes in the graph, /> Indicates the node type, /> is a submatrix of the adjacency matrix in the message heterogeneous graph, containing Type of line and /> Type of column, /> is the transpose of the matrix; if the message text node /> and the message text node /> Connect to some type of /> Node, /> will be greater than or equal to 1, then/> will be equal to 1; For message text and the message text/> The message pairs are composed of the following: Whether the preset threshold is reached/> , when the similarity reaches the preset threshold/> When an edge is constructed/> , after judging all the message pairs, the initial message isomorphism graph is formed/> , where/> Is a collection of message text nodes, /> is the edge set. 7. According to claim 1, a social media event detection method combining deep learning classification and graph clustering is characterized in that the constructed message isomorphism graph is used as the input of the graph clustering algorithm for clustering, and the clustering result is used as the detected social media event, specifically including: The community is expanded according to the transitivity of feature associations between messages, and then multiple rounds of games are conducted between message nodes to select a better community, ultimately achieving a stable state of the community. The clustering results obtained by community division are used as detected social media events. 8. According to claim 7, a social media event detection method combining deep learning classification and graph clustering is characterized in that community expansion is performed according to the transitivity of feature associations between messages, and then multiple rounds of game selection are performed between message nodes to select a better community, and finally a stable state of the community is achieved, which specifically includes: Expanding the community scale, if there are two edges between three points, a half triangle is formed. According to the ternary closure principle, two nodes are also related when they are related to a common node. It is considered that there is an implicit association between the three points, that is, the three points belong to the same community. Multiple rounds of iterations are carried out. In each round of iteration, all nodes make better choices based on the current community division. There are three choices: do not change the current community; leave the current community and do not join any other community; leave the current community and join another community; After multiple rounds of games, the algorithm reaches Nash equilibrium, that is, all nodes join the community that they are most satisfied with. The algorithm can be judged whether it has reached Nash equilibrium by setting a threshold. 9. A social media event detection system combining deep learning classification and graph clustering, characterized in that the system includes: A message heterogeneous graph construction module is used to extract feature information from message texts in social media data streams, use message texts and feature information as nodes, and connect message texts with their extracted feature information to form edges, thereby constructing a message heterogeneous graph; A multi-relation message graph construction module is used to randomly select two message text nodes to construct message pairs based on the message heterogeneous graph, and to construct multiple shared feature edges in each message pair to obtain a multi-relation message graph; A deep learning classification module, used to obtain the similarity of message pairs based on the multi-relation message graph using a deep learning classification model; A message isomorphism graph construction module is used to determine whether the similarity of a message pair reaches a preset threshold. If the similarity of the message pair reaches the preset threshold, an edge is constructed between the two message text nodes constituting the message pair to construct a message isomorphism graph; The graph clustering module is used to cluster the constructed message isomorphism graph as the input of the graph clustering algorithm, and the clustering result is used as the detected social media event. 10. An electronic device, characterized in that the device comprises: a processor and a memory; The memory is used to store one or more program instructions; The processor is used to run one or more program instructions to execute the steps of a social media event detection method combining deep learning classification and graph clustering as described in any one of claims 1 to 8.