As its name suggests, Big Data involves enormous quantities of data. At first glance, these massive datasets are passive. Without an algorithm to process them, they provide nothing. So how does it work in practice?
Tracking a flu epidemic =============================
In 2009, using the number of searches for flu symptoms ("fever," "aches," "flu," etc.) in a given location and how they changed over time and across regions, Google thought it could predict the course of a flu epidemic and launched Google Flu Trends. It withdrew the tool a few years later because Google's forecasts proved wrong in 90% of cases! There were several reasons. To begin with, many people confuse colds with flu, distorting their searches. Furthermore, what matters is not geographical distance but travel time. Paris is therefore closer to Nice than Digne is, because Nice and Paris are connected by air, whereas Nice and Digne are linked only by the Train des Pignes and by road. In short, Google inadvertently demonstrated that, however massive they may be, data are worthless unless they are selected and processed by an appropriate algorithm. At the very least, it seems necessary to correlate search-engine query data with actual statistics on cases recorded at a given time.
Ranking web pages ========================
Although Google got it wrong when predicting flu outbreaks, it ranks web pages effectively. The US giant uses the mathematical concept of a Markov chain together with data on clicks from one web page to another. The best way to understand how this works is to picture the Web as a gigantic graph whose nodes are pages, connected whenever one links to another—that is, whenever the second can be reached from the first with a single mouse click. This graph is immense: it contains several billion pages. Most of these pages (90%) are connected to one another in one direction, the other, or both. Google uses the graph of the Web to return pages about a given subject, ranking them by relevance. This ranking is determined by internet users themselves, who choose whether or not to link their pages to particular others. Each page is therefore assigned a score based on the number of pages linking to it, and it distributes that score among the pages to which it links. To make the calculations more precise, imagine that the part of the Web dealing with the subject of a search is limited to four pages, labelled A, B, C and D.