Public safety is a big concern, especially at college campuses.
At UNC, a common complaint is that the university alert system, Alert Carolina, often sends out alerts a long time after the danger is gone, or even doesn't alert the community at all. It's said that social media, including Twitter, Facebook and YikYak, are much more efficient at conveying real-time information.
We set out to create a system that could analyze data from social media and infer whether a dangerous situation is currently occurring.
We achieved this by training a machine learning algorithm with a dataset of 1000 tweets classified as conveying danger or not. After the initial training, the model could then guess with a certain accuracy whether a new message conveys immediate danger or not.
Users can use the web/mobile interface to subscribe to alert messages via SMS or email. Alerts are tailored to specific campuses and are triggered after a certain percentage of new tweets in the area are classified as dangerous.
We use Twitter, Sendgrid and Twilio APIs, as well as a Machine Learning algorithm in NodeJS.