UniSA researchers have built a forecasting system which analyses millions of social media posts each day to predict the likelihood of civil unrest events in Australia and the Asia-Pacific.
Sifting through open source data from websites, Twitter and Facebook streams, the system – called Carbon – filters vast amounts of information which provides clues about possible protests and strikes.
Dr Wei Kang, from UniSA’s School of Information Technology and Mathematical Sciences, says the system, built in collaboration with big data company D2D CRC, is designed to alert relevant authorities and the public.
Researchers generate a dictionary of keywords that are commonly used in relation to protests and strikes and the system uses these to filter several hundred million tweets and Facebook posts across Australia and Asia each day.
Carbon demonstrated its accuracy in 2017 when it forecast a planned bus strike for Adelaide after analysing records and tweets in the weeks before. On 1 April, commuters were warned of a planned bus strike on 4 April, six days after the UniSA system issued the prediction. The strike would have thrown 50,000 commuters into chaos but, fortunately, was averted after the bus operators and drivers reached an agreement.
The system uses clustering techniques which group together common words. Once these reach a threshold, it signals a pattern which alerts researchers to a possible event.
UniSA PhD candidate Jeff Ansah, working alongside Dr Kang, says the research shows that humans are creatures of habit, repetition and patterns.
“On social media the word ‘protest’ is often used in a very loose way, but our system is so advanced that it can filter the noise from useful information and then predict behaviour,” he says.
Carbon’s use could potentially be extended to predicting health outbreaks, especially pandemics, Jeff says, as people are prone to discussing their health using hashtags such as #flu.
The same analytics have transferable applications in the stock market, with the system able to mine public sentiment about products, thereby influencing company stocks.
The model could also be adapted to predict the next season’s fashions as well as forecasting traffic gridlocks for motorists on their daily commute.
“Its uses are endless,” Jeff says. “The main challenges are in storing and processing such vast amounts of data to extract the right information.
“They specialise in telling us when, why and where an event is going to take place and who is likely to be involved.”
The project is being funded by the Data to Decisions Cooperative Research Centre.