Detecting and responding to real-time emergencies with Machine Learning

Published 17 May 2021 | by Gary Devenay

Detecting and responding to real-time emergencies with Machine Learning

Since the launch of Safe & the City i3 in March 2021, we have been focused on delivering a packed roadmap of features and improvements to both the i3 Risk and i3 React products. Our improvement objectives are simple: Deliver more accurate information about safety emergencies and risks to our customers faster.

A key part of delivering information quickly to our i3 React API customers is incident discovery. We have to be able to discover unfolding, real-world events across the country as they are happening. The typical method of learning of these events, be it a protest, riot or terrorist threat, is from national news coverage. This can take 30 minutes or substantially longer for a news organisation to gather enough information to publish a story. This takes too much time for us to provide meaningful information to those on the ground and potentially in harm's way.

In order for us to identify events as they happen, we instead turn to Twitter. Twitter’s API allows us to ingest a real-time, filtered stream of Tweets using in-built rules and operations (more information on Twitter APIs). Of course— thousands of Tweets are posted each second so this would create an unwieldy task for any one person to read and categorise in a timely fashion.

Over the last few months, we have been able to dramatically improve the speed and accuracy of our detection system with Machine Learning. With guidance provided by Dr. Marek Rei of Imperial Consultants at Imperial College London, we have been implementing and training Natural Language Processing (NLP) models using PyTorch and Hugging Face Transformers with the objective of detecting and flagging new public safety incidents in real-time as they unfold.

In preliminary training and testing, we were able to detect protests and riots unfolding with very high accuracy. Our initial results show an F1 score of 0.92 out of a maximum score of 1.0. We measure our success in F1-score because it takes into account false positives, as well as false-negative results (read more about F-score).

Each new incident detected is then flagged to a member of the Safe & the City team to review for accuracy and measured against a number of metrics for information reliability. Upon acceptance, positive results are forwarded to our i3 React customers via their registered webhook endpoints.

Once live, customers can now expect to receive i3 React notifications in as little as a few minutes. This i3 React upgrade will come as standard, with no change of existing implementations required.

This is just one example of how Safe & the City is utilising state-of-the-art Machine Learning models to increase public and personal safety. If you are interested in learning more about how you can integrate our i3 Intelligence APIs for consumer safety, business intelligence or risk management get in touch