23 September 2016 - Last week, third year Business Analytics students M. Thirukkumaran, Andre Tan, Eddy Pang participated in finals of the 2016 Teradata University Network Analytics Challenge (TUNAC 2016) and won the Best Analytics and Visualisation Award.
There were a total of 77 entries for the annual business analytics competition and 10 finalists were selected to present their projects to experts and professionals in the field, at the 2016 Teradata PARTNERS Conference in Atlanta, Georgia, USA. The NUS Computing team was the only non-American team in the finals.
Thiru, Andre and Eddy’s project, ‘Terror-Eyes’ is an immediate response system that empowers authorities to react intelligently and more quickly to terror threats. Speaking on behalf of the team, Thiru explained, “The immediate problem following a terrorist attack is the lack of information that authorities face. Who did it? Will subsequent attacks happen? Where? It would be immensely helpful for authorities to get these answers immediately, and these answers may even help even stop subsequent attacks. We came up with this project because we were deeply concerned about the increasing frequency and seriousness of terror attacks across the world. We are more vigilant than ever, yet the attacks seem to be getting worse. Although it’s hard to predict the first attack, we think that we could narrow down the problem, and instead, use the information from the first attack to quickly identify the perpetrator, and prevent subsequent attacks (if any). Thus, minimising the total lives lost. Anti-Terrorism is a communal effort and we hope that Terror-Eyes can play an important role against the fight on terror.”
According to the team, many of the other teams focused on healthcare and most only presented visualisations of data. The NUS Computing team, however, tackled a more unusual current issue, and had a working prototype that visitors could test at their booth and received raves about their solution. “It was far from a smooth journey – we encountered many obstacles along the way with getting the ensemble learning to work, and phrasing a complex machine learning problem in an easily comprehensible fashion. The months leading up to the finals in Atlanta were particularly trying. We really wanted to submit the best project we could build, so countless nights were spent racking our brains on how to improve the web platform, and practising the way we would pitch the ideas to the judges. This effort ultimately paid off and it was immensely satisfying watching our idea come to fruition, and that other experts found our work meaningful,” said Thiru.
Mr. Benedict The, the team’s faculty advisor, observed that the competition was very close, with no clear front runner until the winners were announced as all the finalists presented very impressive ideas. Further describing the experience, Thiru said, “The feeling of competing on an international arena was both exhilarating and an eye-opener. It was great to be given the opportunity to represent NUS on the global stage, and we learnt a lot from interacting with some of the best in the field, and other bright students as well! It was very insightful learning how larger enterprises were leveraging data and technology to change the way they do business, yet humbling at the same time when we saw that many industry professionals found our application useful.”
How it works:
Within seconds after an attack, Terror-Eyes identifies the perpetrator responsible, and predicts other critical information. Some of which include the total number killed, the probability of subsequent attacks and the location of those attacks, which are then reflected on a heat map, allowing efficient allocation of manpower.
Dashboard in action
To accomplish this, we utilized the Global Terrorism Dataset, an open source dataset with information on terror attacks from 1970s till 2014 and utilized ensemble learning, in which we combined 5 different classifiers to improve accuracy of our predictions. This is further augmented with a twitter bot, aimed at crowdsourcing antiterrorism efforts. Upon an attack, users can tweet suspicious activity to this bot with their location and photo of the activity. The bot aggregates this information and displays the locations on another heat map for comparison, while the photos are downloaded into a folder, which an image classifier (Inception V3 on Tensorflow) analyses to determine what the picture is about. This saves the authorities having to comb through the tweets themselves to glean information. The accuracy of the ensemble of classifier's to predict the given perpetrator behind an attack was 75%.