10/08/2006 | rombarbar

Nature Physics 2, - p361 (2006)

In 1929, John Slater published a paper entitled "The Theory of Complex Spectra", in which he applied the emerging mathematics of quantum theory to the physics of complex multi-electron atoms. Among papers published in the Physical Review family of journals over the past century, Slater's paper now ranks 1,853rd in terms of citations — somewhat lower than might be expected, perhaps, for a landmark in the development of quantum mechanics. All physicists know Slater's name because of this paper, which introduced the 'Slater determinant' for many-body wavefunctions. Why doesn't the citation count reflect this?
We may now have an answer for Slater's paper as well as for other 'classic' papers that the citation count apparently undervalues. It seems there is a better way to use citations in judging the importance of a paper. The key idea, fittingly enough, has emerged out of the technology for searching the World Wide Web.
Google is the world's favourite search engine, largely because of the algorithm it uses to rank web sites by relevance and quality. If you search on keywords, Google ranks sites, in part, according to whether they contain those words, and how they appear on the page. But a page also scores more highly if lots of other pages have links to it, effectively 'voting' for it as being useful. Votes from more highly ranked sites count more. Google scores sites highly if they receive lots of votes from other highly ranked sites.
This circular definition expresses the self-consistency that important pages catch the interest not only of other pages, but of other important pages. Calculating this measure — its 'Google number' — turns out to be equivalent to finding an eigenvector of a matrix reflecting the links between pages.
Clearly this idea might be applied to any network, including that of research papers linked by citations, which is what Pu Chen and colleagues have now done for papers in Physical Review. The Google idea implies that important papers should not only garner lots of citations, but citations from other important papers. To calculate Google numbers, Chen and colleagues used a database of Physical Review citations for 353,268 articles published from 1893 to 2003. Their results reveal some fascinating discrepancies between crude citation rankings and the more sophisticated Google measures (http://arXiv.org/abs/physics/0604130).
For example, the famous 1944 paper of Lars Onsager, in which he reported the solution of the two-dimensional Ising model, ranks only 55th in terms of citations, but rises to 6th in the Google ranking. This change reflects the fact that papers referring to Onsager's work turned out to be exceptionally important themselves. What about Slater's paper? Ranked 1,853rd purely in terms of citations, it jumps to number 10 in the Google ranking, for similar reasons. The Slater determinant slipped into common usage and into a number of other papers that went on to become classics. Today, this paper gets few direct references, but scores points indirectly in Google terms as others papers that cited it long ago continue to accrue new citations.
So the citation count really is a blunt instrument, as we may have long suspected. There's a better way. One can hope that those who assess scientific output will begin to use it.