What did Larry and Sergey knew, and when did they know it?

August 1, 2011

By now, it is widely understood that Google’s PageRank algorithm builds on the insight that the eigenvector of the largest eigenvalues for the adjacency matrix representing the set of all web pages and the links between them, yields a pretty good measure of a page’s influence. Probably less well-known is that the use of eigenvector centrality or its variants to measure the influence of particular nodes was well-known to network sociologists ever since Phil Bonacich‘s pioneering work in the early 1970s.

From my admittedly cursory search of the citation record, it does not seem that Brin and Page credited this earlier literature in their early efforts to rank web pages’ influence. How could this be? I see three possibilities.

  1. They may have been genuinely ignorant of Bonacich’s contributions, and simply reinvented the wheel with a 25 year lag. As an economist, I am genuinely distraught by this possibility, since it implies that the social return to network sociology research was zero (or at least a very small fraction of what it could have been).
  2. The use of eigenvector centrality to measure influence in networks might have been well known to contemporaries of Bonacich in computer science, but it circulated within their field in the form of examples rather than a codified body of knowledge, maybe because it was not clear what kind of problems it could be applied to.
  3. Brin and Page might have been aware of the network sociology literature, but chose to credit computer scientists (including their advisers) for parochial reasons.

Which of these possibilities is the correct one? Writing this particular piece of scientific history could tell us much about how the market for scientific credit works ­­­— or sometimes fail to work.

NB1: of course, none of this is meant to belittle Page and Brin’s enormous contribution. First, they recognized the usefulness of couching the search challenge in network terms; this was a key recombinative step. Second, the complementary algorithmic innovations that made it feasible to compute the centrality scores for an extremely large (and constantly growing) network were essential to realize their vision.

NB2: Steven Levy’s rather excellent history of Google does not enable one to adjudicate between these interpretations.

NB3: The citation behaviors of physicists who study networks also exhibit pronounced parochialism. They generally ignore the contributions of Linton Freeman and other network sociologists. This tend to annoy my sociologist colleagues to no end, probably for good reasons…