Modelling citation networks

Abstract

The distribution of the number of academic publications against citation count for papers published in the same year is remarkably similar from year to year. We characterise the shape of such distributions by a ‘width’, $\sigma ^2$, associated with fitting a log-normal to each distribution, and find the width to be approximately constant for publications published in different years. This similarity is not surprising, after all, why would papers in a given year be cited more than another year? Nevertheless, we show that simple citation models fail to capture this behaviour. We then provide a simple three parameter citation network model which can reproduce the correct width over time. We use the citation network of papers from the hep-th section of arXiv to test our model. Our final model reproduces the data’s observed ‘width’ when around 20 % of the citations in the model are made to recently published papers in the entire network (‘global information’). The remaining 80 % of citations are made using the references from these papers’ bibliographies (‘local searches’). We note that this is consistent with other studies, though our motivation to achieve the above distribution with time is very different. Finally, we find that, in the citation network model, varying the number of papers referenced by a new publication is important as it alters the parameters in the model which are fitted to the data. This is not addressed in current models and needs further work.

Appendix

Fitting procedure

We follow the procedure used in Evans et al. (2012) and Goldberg and Evans (2012). We use logarithmic binning so that the citations in bin b $c_b \in {\mathbb {Z}}$ with $c_{b+1}$ equal to $R c_b$ rounded to the nearest integer or to $(c_{b}+1)$, whichever is the highest, where R is some fixed bin scale chosen to ensure there are no empty bins below the bin containing the highest citation values. The edge of the first bin is chosen to be the lowest integer above the value $0.1 \langle c \rangle$. In order to make the fit we compare the total value in the bth bin, $n_b= \sum _{c=c_{b}}^{c_{b+1}} n(c)$, against the expected value

$$\begin{aligned} n_b^{({\text {expect}})} = (1+A) N \int _{c_{b}-0.5}^{c_{b+1}+0.5} {\text{d}}c \,\, \frac{1}{\sqrt{2\pi } \sigma c } \exp \left\{ -\frac{(\ln (c/\langle c \rangle )+(\sigma ^2/2)-B)^2}{2\sigma ^2 } \right\} . \end{aligned}$$

(9)

This gives us a sequence of data and model values which are compared using a non-linear least squares algorithm to give us values for $\sigma ^2, A$ and B.

Lognormal out-degree distribution

In all above models the citation networks were created by determining the number of references a new node would create (via a normal distribution mean 12.0, standard deviation 3.0 references) and then having a method of deciding which nodes to reference. However, we found that a lognormal fits the out-degree distribution of the hep-th data better than a normal distribution, Figs. 4 and 21, respectively.

Fig. 21

This is a plot of the out-degree distribution of 27,000 publications from the hep-th arXiv data, in blue, on a log–log plot. Superimposed, in green, is a plot of 27,000 numbers generated by the lognormal distribution fitted to the out-degree of the hep-th arXiv data. We observe that the lognormal is a better fit to the data than the normal in Fig. 4. (Color figure online)

As further work we inputted this fitted lognormal to determine the number of references created by a new node into the model C and ran it for our final parameters $(p,q)=(0.55,0.20)$ and $\tau = 200$papers. The ratio of the $\sigma ^2$ values associated with the in-degree distribution of papers published in the same year for the data is divided by the corresponding year’s $\sigma ^2$ for this modified model C and plotted against year in Fig. 22. We find that the ratio is close to 1, however, it is not as close as the original model C, Fig. 20. Therefore the $\sigma ^2$ plot does depend on in-degree, contradicting (Ren et al. 2012), who say it is ‘innocuous’ to the in-degree distribution of the citation network. Although this out-degree distribution has been observed by the literature (Vázquez 2001) its use in a citation network model is novel and original.

Fig. 22

This is the ratio of $\sigma ^2$ associated with the in-degree distribution of papers published in the same year for the data (years from 1992, 1993 etc. relabelled to 0, 1 etc.) divided by that of the modified model C (where the out-degree is determined by a lognormal distribution, above). The plot and error bars lie within 1.0 therefore the modified model C is consistent with the data. Therefore the modified model C is promising, a significant improvement on model A and B, Figs. 7 and 14, respectively. However, the data is always lower than the modified model C; the points are always above the 1.0 line and not as close to 1.0 as the model C which implies the need for modification of the parameters in model C. So modifying the out-degree does change the in-degree, which contradicts (Vázquez 2001). We conjecture that by changing the attention span parameter this model’s $\sigma ^2$ plot could increase to match the data

Although the $\sigma ^2$ plot is lower than that of the original model C we conjecture that by varying the $\tau$ of the model C the $\sigma ^2$ plot could match that of the data’s, this may also increase the attention span to something closer to a year as expected by Simkin and Roychowdhury (2007).