Study On Traffic Incident Identification Based On The Social Network Data

In recent years,with the rapid development of the Internet,the netizens are growing continually.The social networks,as a medium for the netizens to communicate with each other,is gradually becoming an indispensable part of our daily life.From the 13 th five-year plan to the 14 th five-year plan,the Chinese government has been vigorously advocated making full use of social network information,opening an access to them.In the field of transportation,the massive real-time information shared by netizens provides a potential rich resource,which contributes to mining traffic accidents,traffic congestion and so on.The Sina Weibo,as one of the largest social networking platforms in China,provides rich text resources.Thus,taking Sina Weibo as the research object,this paper explored to mine and identify the non-recurrent freeway traffic incidents,as well as make the analysis of public opinions.This paper used the web crawler to crawl two kinds of the sample sets.After that,this paper preprocessed the sample sets by text denoising,text deduplication,text marking,word segmentation and stop-words filtering.In addition,a location library was established to make the collected information be classified by region.After that,the machine-learning-based methods were applied to identify traffic accidents.The paper employed a method to select feature words based on feature weights.The feature weight is calculated by normalizing,weighting,and combining the word frequency and the ratio of the text containing that word.Accordingly,the feature weight of each unique word in the training set of the traffic incident text could be achieved,used as a criterion for selecting feature words,and as the inputs of classifiers.A test was conducted to compare different classifiers and methods to select feature words.The results showed that the proposed method to select feature words combined with the XGBoost classifier has the optimal performance.Next,this paper constructed the Word2Vec-Conv BILSTM model,which integrates a Text CNN-layer with a BILSTM-layer,to identify traffic congestion.In addition,the word vectors were constructed by the Word2Vec-based tool,which were embedded into the Conv BILSTM model.Finally,some experiments were conducted.The results validated the effectiveness of the proposed model,and it could be an alternative data source to detect the nonrecurrent freeway traffic congestion.Finally,this paper also studied public opinions.Through a word-frequency analysis,a keyword-based analysis,an Apriori-Association rule and a co-occurrence network-based analysis,this paper established the relevance between word semantics.Subsequently,the paper constructed two models,including an index weighting-based model and a K-Means clusteringbased model.The paper also built a visualization system based on Tkinter.The system contributed to showing a holistic and systematic research idea.In short,using the machine-learning-based methods and the deep-learning-based methods,this paper identified the non-recurrent traffic incidents effectively.Furtherly,they have the advantages,e.g.,the real-time and the wide coverage.Therefore,Sina Weibo could be used as a potential information source to assist the traffic departments partly.Additionally,making an analysis of public opinions is helpful to understand the evolution of them and build a management and control system.