Research On Spatio-temporal Event Detection In Social Media

Social media is a new type of online media with many users participating.In recent years,social media sites represented by Twitter have attracted millions of users.Events that occur in the real world will be quickly received in social media.Social media contains rich information about events.The event detection method in social media is a method that uses machine learning and natural language processing and other technologies to find events that occur in real life from massive social media data.Effective event detection allows people to understand the social events in a timely manner.Information on the hot events can help the government and other agencies respond to social events in a timely manner and take measures to deal with these problems.For influential events,event-related content is usually widely discussed in social media.So a basic idea of existing social media event detection is to find text related to the event through text clustering,and then to analyze the text cluster to determine whether it is an emergency.However,the existing research has two problems: 1)the text clustering effect on the short text stream of social media is not good enough;2)the feature extraction and determination of events are not accurate enough.In view of the above problems,this thesis studies social media event detection methods from the perspective of obtaining more accurate event characteristics through spatio-temporal information.The main contributions are summarized as follows:(1)An online density clustering method based on fusion similarity is proposed.In terms of similarity calculation,considering the characteristics of short text length in social media,this thesis uses short text stream to construct a dynamic word association space,and based on this space,constructs text similarity with multiple indicators.In terms of text clustering,for the problem of single-pass incremental clustering method is not effective,this paper uses online density clustering method to cluster online short text streams.Experiments show that the proposed method has achieved good results in text similarity measurement and text online clustering.(2)Propose an emergency decision method based on geographic entities.Since most of the events are related to geographic information,this article filters the events to obtain real emergencies based on the detection of textual geographic entity bursts.In the judgment of the suddenness of geographic entities,the frequency distribution of geographic entities has the characteristics of long-tail effect.In this thesis,it is converted into normal distribution through logarithmic standardization,and the suddenness level of geographic entities is measured by Z scores.For candidate event clusters associated with the detected sudden geographic entity,this thesis uses a multi-level filter to filter it.And a tweet is extracted as a description of the event.Through the test in the actual data,it is found that this method can timely and accurately detect the relevant information of emergencies,and achieve a better event detection effect.