Identifying Relevant Messages in a Twitter-based Citizen Channel for Natural Disaster Situations

TL;DR

This paper develops and evaluates a machine learning classifier to identify relevant tweets during natural disasters, specifically earthquakes, to improve information access for affected citizens.

Contribution

It presents a detailed process for building and validating a tweet classifier, analyzing the impact of class imbalance and dimensionality reduction on model performance.

Findings

Class imbalance significantly affects classifier accuracy.

Dimensionality reduction improves model efficiency.

Model performance varies across different classifiers.

Abstract

During recent years the online social networks (in particular Twitter) have become an important alternative information channel to traditional media during natural disasters, but the amount and diversity of messages poses the challenge of information overload to end users. The goal of our research is to develop an automatic classifier of tweets to feed a mobile application that reduces the difficulties that citizens face to get relevant information during natural disasters. In this paper, we present in detail the process to build a classifier that filters tweets relevant and non-relevant to an earthquake. By using a dataset from the Chilean earthquake of 2010, we first build and validate a ground truth, and then we contribute by presenting in detail the effect of class imbalance and dimensionality reduction over 5 classifiers. We show how the performance of these models is affected by these variables, providing important considerations at the moment of building these systems.