S1 Fig displays in each country the best transformation of each input variable and suspected Zika case counts. This app is simply one of the best possible for the iPhone. The best model performance (lowest relative RMSE) in each time series by country is shown as a bolded line. As a consequence, we relied on the relative RMSE (rRMSE) to establish the quality of model prediction given the short time span of the outbreaks. The rRMSE provides an estimate of the prediction error relative to the number of true cases observed in each week over the evaluation period, and, from our perspective, allows for better comparisons across models and time horizons. Data normalization. An increase in the volume of a Web search query increases its own average over time and thus its denominator for future comparisons. Note that while some model predictions showed high correlation values with official case counts, their predictions showed large discrepancies with the data. These transformations were observed to sometimes lead to better correlation values than the original raw variables for different time periods. A futuristic step taken already, towards ubiquitous computing, based upon the idea that the Internet and computers will be accessible everywhere at any point of time without even needing to use one’s hands.
Plots comparing model predictions with the official Zika case count, by time horizon and country, are shown in Figs 1-3. Table 2 summarizes the out-of-sample predictive performance of the four models for each of the three week-ahead time horizons and for all countries, as captured by the three evaluation metrics. For each country, we produced out-of-sample predictions for the one, two, and three-week ahead time-horizons with the four models introduced in the previous section. From the multiple panels for each country, it can be seen that at least a subset of these (transformed) variables showed potential to be useful to track Zika. Devlin (2017, online) contends that those working in the big data and analytics industries are perhaps the least likely to be surprised that political figures or parties would try to use algorithms to influence public behaviour in their favour, saying that “the application – both overt and covert – of technology to affect election outcomes was arguably inevitable” (Devlin 2017, online). In addition, we evaluated models with and without the inclusion of Twitter data. Thus, the inclusion of local content by Google News had mixed effects on local outlets: it increased their traffic, especially in the short run, but it also increased the reliance of users on Google News for their choices of news, and increased the dispersion of user attention across outlets.
When born-digital content, for example online news, is written with search engine visibility in mind it is, in effect, automatically tailored towards advertising; advertisers and content creators both want to strengthen their association to the kinds of words and phrases used in search engine queries. If no one is using a language there is no incentive for advertisers to pay Google for specific words and phrases, thus Google accelerates the process of online language death. The user-centered design process included a design workshop, the sketching of a lo-fi prototype, a first think-aloud evaluation of the lo-fi prototype, adjustments and a second think-aloud evaluation of an implemented prototype app.The implementation part investigated what challenges are presented when implementing the interface in a real product (the official Firstonsite app). Location-based functionality was investigated and different solutions are described. Using a dataset of user browsing behavior, we compare users who adopt the localization feature to a sample of control users who are similar to the treatment users in terms of recent internet news consumption.