Promise and Pitfalls of Extending Googles PageRank Algorithm to Citation Networks

We review our recent work on applying the Google PageRank algorithm to find scientific “gems” among all Physical Review publications, and its extension to CiteRank, to find currently popular research directions. These metrics provide a meaningful extension to traditionally-used importance measures, such as the number of citations and journal impact factor. We also point out some pitfalls of over-relying on quantitative metrics to evaluate scientific quality.

💡 Research Summary

The paper investigates how Google’s PageRank algorithm, originally designed for ranking web pages, can be adapted to scientific citation networks and extended to a time‑biased version called CiteRank. Using the complete internal citation dataset of the American Physical Society (APS) journals—353,268 papers published from 1893 to June 2003 with a total of 31,108,839 citations—the authors treat each paper as a node and each citation as a directed edge. In the standard PageRank formulation a random “surfer” follows an outgoing citation with probability (1‑d) and, with probability d, jumps to a randomly chosen node. The authors argue that for citation networks the appropriate damping factor is d = 0.5, reflecting the empirical observation that researchers typically follow about two citation steps before starting a new search. After iterating until convergence, each paper receives a “Google number” (its PageRank score).

The results show that while highly cited papers generally obtain high PageRank, a number of older, modestly cited works achieve surprisingly high rankings. This occurs because PageRank weights citations from influential papers more heavily and also gives extra credit when the citing paper has a short reference list. Consequently, papers such as Onsager’s 1944 exact solution of the two‑dimensional Ising model, Anderson’s 1977 work on the absence of diffusion, and Slater’s 1929 theory of complex spectra emerge as “scientific gems” – papers that are not obvious from raw citation counts but have had a deep, lasting impact. The authors present a table of the top‑ranked non‑review APS articles, illustrating how PageRank can surface seminal contributions that would otherwise be overlooked.

Recognizing that citation networks are inherently temporal—papers can only cite earlier work—the authors introduce CiteRank to capture current research trends. CiteRank modifies the initial distribution of surfers by biasing them toward recent publications, using an exponential decay with time constant τ = 2.6 years. The same damping factor d = 0.5 is retained, meaning a surfer typically follows two citation steps before restarting. This model predicts the rate at which papers acquire recent citations and therefore highlights papers that are currently “hot” in the community. The authors demonstrate that CiteRank successfully identifies contemporary focal points such as quantum information, high‑temperature superconductivity, and complex materials, complementing the lifetime‑oriented perspective of PageRank.

The paper also discusses several caveats. First, citation counts (and any derived metric) do not always reflect intrinsic intellectual value; classic works may receive few citations after becoming textbook material. Second, disciplinary differences in citation practices and collaboration sizes introduce systematic biases. Third, both PageRank and CiteRank are sensitive to network topology and parameter choices, so applying the same settings to other fields or databases would require careful recalibration. Fourth, the authors warn against over‑reliance on such quantitative indices for evaluating individual researchers, departments, or institutions, noting that metrics like the h‑index have already become easy to compute but can be misused. They argue that while PageRank and CiteRank provide computationally inexpensive, context‑aware measures that improve upon raw citation counts, they cannot replace expert judgment.

Finally, the authors make all computed PageRank and CiteRank values publicly available (http://www.cmth.bnl.gov/~maslov/citerank), encouraging further exploration and adaptation of these methods. They conclude that, when used judiciously, PageRank and its time‑biased extension offer powerful tools for uncovering both enduring scientific contributions and emerging research directions, but they must be complemented by qualitative assessment to avoid the pitfalls of metric‑driven evaluation.