Laura (right) and two other teammates from the Fire Prevention team do the demo of their project, an app they created during the Women in Tech machine learning product hackathon.
They created a predictive model which will take in factors about natural disaster such as location, date, storm type (hail, high wind, wildfire, heavy snow, etc.) and predict what other fire call types are prone to occur. The intended outcome of their project is to use the expected fire call predictions to create action recommendations both on the personal level (watch for candles, stay inside) and at the government administration level (scan for down power lines).
As the product of the hackathon, they exposed the predictive model via a "shiny web app" that the fire marshall teams can use to understand potential fire calls. In the future they would like to work with subject matter experts to turn predictions into recommendations and deliver them real time -- for example, how phone alerts are delivered to warn people about flooding.
They built this model based on of fragmented datasets available to them:
-- general fire department calls (all calls to fire departments in Kansas in 2016 and 2017 that occurred only during storm events)
-- detailed fire events (all confirmed fires and extra details)
-- public storm data (public storm events and associated meta data)
Their best decision tree model had accuracy of 29.53%.