Here we provide greater technical detail concerning our historical significance ranking methods. We will discuss (1) exactly what data we used, (2) how the data was normalized, and (3) the statistical methodology we used to develop our rankings. Interested readers should begin with the high-level overview of our methods that we presented in Chapter 2.

Feature Set

We used the October 11, 2010 Wikipedia release (http://dumps.wikimedia.org/enwiki/20101011/) as the basis for our analysis. We extracted six input variables from this distribution and other available datasets:

1. • Page Rank (NPR and PPR) – Wikipedia is a hyperlinked document, defining a network with vertices corresponding to articles and directed edges (x, y), meaning that article x references article y. The well-established PageRank measure of network centrality [Brin and Page, 1998] computes a significance score based on the number and strength of the in-links to each node. We compute two forms of PageRank, based on two different graphs derived from Wikipedia. The first contains all Wikipedia pages, while the second consists only of the pages corresponding to people. Page annotations were determined using Freebase (http://www.freebase.com/), a collaborative knowledge database. We employed the Cloud9 map-reduce library (https://github.com/lintool/Cloud9) to compute Page-Rank.

2. • Page Hits (PH) – Web logging data reveal how often eachWikipedia page is viewed. More famous/significant entities should have their pages read more frequently. We analyzed six months of log data, collected immediately prior to the date of our Wikipedia release. In order to reduce the large monthly variance in readership owing to news events, we used the median monthly frequency as our measure of page hits.