A stacked convolutional neural network for detecting the resource tweets during a disaster.

Area of Science:

• Natural Language Processing
• Machine Learning
• Disaster Informatics

Background:

• Social media platforms like Twitter are crucial for real-time information dissemination during disasters.
• Detecting tweets related to the Need and Availability of Resources (NAR) is vital for effective disaster response.
• Existing methods for NAR tweet detection have limitations in performance and focus.

Purpose of the Study:

• To develop a reliable methodology for detecting NAR tweets during disasters.
• To address the limitations of existing NAR tweet detection techniques.
• To improve the accuracy and efficiency of identifying critical resource-related information from social media during crises.

Main Methods:

• Proposed a stacked model combining Convolutional Neural Networks (CNN) with traditional feature-based classifiers.
• Engineered informative features such as 'aid,' 'need,' 'food,' and 'earthquake' for classification.
• Utilized a meta-classifier (SVM) with learned features from base classifiers (KNN, Decision Tree, Naive Bayes) and CNN.

Main Results:

• The proposed stacked model, particularly using KNN as a base classifier and SVM as a meta-classifier with CNN, outperformed other algorithms.
• Achieved the best accuracy compared to baseline methods in detecting NAR tweets.
• Demonstrated the effectiveness of the model on 2015 and 2016 Nepal and Italy earthquake datasets.

Conclusions:

• The developed stacked model offers a robust solution for identifying NAR tweets during disasters.
• Combining CNN with traditional classifiers and informative features significantly enhances detection accuracy.
• This approach provides a valuable tool for humanitarian organizations and disaster management agencies to leverage social media data effectively.