Active online Multiple Prototype Classifier to Support Crisis Management

Abstract

In times of crisis, social networks have emerged as crucial channels for open communication. The traditional methods employed for analysing social media data in crisis situations have faced criticism due to their mixed results and limited applicability beyond the scope of the initial study. To address these issues, a novel online active multi-prototype classifier, known as AOMPC, has been proposed. AOMPC operates with data streams and incorporates active learning mechanisms to actively request labels for unlabeled and ambiguous data points, managing the number of requests through a fixed budget strategy. Typically, AOMPC is designed to handle partially labelled data streams. To assess its effectiveness, AOMPC was evaluated using two types of data: synthetic data and Twitter data related to two specific crises—the Colorado floods and the Australian wildfires. During the evaluation, established parameters were utilized to gauge the quality of results, and a sensitivity analysis was conducted to understand how AOMPC's parameters impacted result accuracy. Furthermore, a comparative study was carried out to contrast AOMPC with other available e-learning algorithms. The experiments demonstrated AOMPC's capability to perform exceptionally well in processing partially labelled scalable data streams, potentially offering valuable insights during crises.