Darya Filippova, University of Maryland, dfilippo@cs.umd.edu [PRIMARY contact]
Carl Kingsford, University of Maryland, carlk@umiacs.umd.edu [Faculty advisor]

We used Python, NLTK library (collection of tools for natural language processing), numpy (scientific computing tools), and matplotlib (python plotting) all of which are publicly available. Writing a script for parsing and filtering tweets requires some programming experience, but no specific knowledge is needed. Modifying scripts to investigate a particular peculiarity in data is easy and straightforward.

Our analysis shows that the epidemic started on May 18^th, likely near the Vastopolis Dome and Convention Center. We determined this by parsing and filtering tweets (via Python+NLTK script). We excluded tweets that did not contain words related to sickness from the task description:

We overlaid a heatmap of the remaining 42275 tweets on the Vastopolis map which highlighted areas of Downtown in red. We then divided the tweets by day and looked at daily heatmaps. Daily analysis revealed a spike in the number of 'sick' tweets on 18^th of May. Looking at hourly tweet counts on May 18^th-20^th, we observed that tweet counts increased significanly on May 18^th at 8a.m. (see Fig. 1). The heatmap analysis for May 18^th confirms that disease started spreading on that day from the Downtown area (see Fig. 2). In particular, there was a significant number of tweets from around Vastopolis Dome, Convention Center, and Vastopolis City Hospital.

While heatmap is good for identifying locations with highest tweet counts, such hotspots may oveshadow medium-count locations that are of equal importance. To give less weight to the highest counts and highlight locations with lower tweet counts, we plotted a log(c+1) where c is a tweet count in a given cell of the heatmap grid. Figure 3 shows that, apart from Downtown, part of Eastside and the banks of the Vast river get contaminated. Hospitals in each of the burrows stand out as well.

A plot of tweet log maps over 3 days shows how "sick" tweet distribution changes over time. We plotted tweet log heatmaps for every day (Fig. 4 shows three last days of the data when disease really takes off). On May 18^th, the disease started in the Downtown area and spread to the Eastside - possibly carried by the strong wind from the West (as indicated in the Weather.csv). Sick tweets started dispersing through the burrows in the evening after 5p.m. (evening commute from center to the suburbs) and were published at a steady rate throughout the night (see Fig. 1 above). On May 19^th, we observed further disease spread in the westward direction due to the strong wind. However, there was a significant number of "sick" tweets along the banks of Vast river downstream of infected Downtown. (We determined the direction of the river by the discharge on one of the dams).

These two observations made us hypothesize that the disease may spread by air as well as by water. While evidence for disease being airborne is strong, we were puzzled to see that there was no increase in "sick" tweets around other bodies of water in the city (lakes Pasta, Bread, Rice, Disco, and Twin Lakes). However, it is better to be safe than sorry, so we conclude that both ways of disease spread are highly likely.