A Cascade of Learners for Firemen' Emergency Events Classification

The main objective of firefighters is to optimise readiness in response to hazardous events and to minimise their collateral effects. In this context, few but growing research is investigating machine learning algorithms to support firefighters' work. Hence, this paper presents a decision support system to promptly identify relevant interventions, which are those events for which the national control room needs to alert the competent authorities because they could be dangerous for the community. The aim is to provide firefighters useful information for the management of such interventions and of the available resources. We define a set of new hand-crafted features specifically designed for the task, which catch both static and dynamic characteristics of the events. Furthermore, we design a learning approach based on a cascade of binary classifiers, which exploits the ability of most of the available classification algorithms to learn binary functions and it takes advantage of some characteristic of the dataset. The experiments were performed in leave-one-out on a real-world data set provided by the Italian National Fire Corps, analysing how the system works to distinguish among relevant and not-relevant interventions, both at the time of the call and during the events updates. The results show that machine learning-based decision support system significantly outperform the human operator.