Identification of medical resource tweets using Majority Voting-based Ensemble during disaster

• Published in: Social network analysis and mining Vol. 10; no. 1; p. 66
• Main Authors: Madichetty, Sreenivasulu, M, Sridevi
• Format: Journal Article
• Language: English
• Published: Vienna Springer Vienna 01.12.2020; Springer Nature B.V
• Subjects: Applications of Graph Theory and Complex Networks; Bagging; Classifiers; Computer Science; Computing time; Data Mining and Knowledge Discovery; Datasets; Decision trees; Disasters; Earthquake damage; Earthquakes; Economics; Floods; Game Theory; Humanities; Information retrieval; Infrastructure; Law; Methodology of the Social Sciences; Methods; Neural networks; Organizations; Original Article; Social and Behav. Sciences; Social networks; Statistics for Social Sciences; Support vector machines; Tsunamis; Victims; Voting; Italy; Nepal; Twitter; Disaster; Earthquake; Ensemble
• ISSN: 1869-5450; 1869-5469
• DOI: 10.1007/s13278-020-00679-y

During disaster, detecting tweets related to the target event is a challenging task. Earthquake, floods, tsunami, etc., are the examples for target event. Prior to several studies have been made on earthquake detection. The event contains many categories (classes) of information such as resources, infrastructure damage and helping requests. Different organizations need different categories (classes) of information. There have been only a few studies on the detection of a certain kind of classes and how they are interrelated during the disaster. It is difficult to design features for discriminating and detecting specific classes. Hence, this paper focuses on detection of medical resource (requirement and availability) tweets class during disaster to help medical organizations and victims. For this purpose, the Majority Voting-based Ensemble method is proposed for the detection of medical resource tweets during a disaster. It uses informative features and is fed to various classifiers such as bagging, AdaBoost, gradient boost, random forest and SVM classifiers. The output of different classifiers is combined by majority voting to detect medical resource tweets during the disaster. The proposed informative features are tested on different classifiers such as bagging, AdaBoost, gradient boosting, random forest and SVM classifiers by using the real-time Nepal earthquake dataset. And the results are compared with standard baseline BOW model. The classifiers considered in this paper with the proposed informative features outperform BOW model. The dimensionality, sparsity and computational time for features are less in case of the proposed informative features as compared with BOW model. The proposed method outperforms the state -of the art for Nepal and Italy Earthquake datasets on different parameters. It detects 82.4% of tweets that are correctly related to medical resources during a disaster.