Stochastic dynamical model of a growing network based on self-exciting point process

We perform experimental verification of the preferential attachment model that is commonly accepted as a generating mechanism of the scale-free complex networks. To this end we chose citation network of Physics papers and traced citation history of 40,195 papers published in one year. Contrary to common belief, we found that citation dynamics of the individual papers follows the \emph{superlinear} preferential attachment, with the exponent $\alpha= 1.25-1.3$. Moreover, we showed that the citation process cannot be described as a memoryless Markov chain since there is substantial correlation between the present and recent citation rates of a paper. Basing on our findings we constructed a stochastic growth model of the citation network, performed numerical simulations based on this model and achieved an excellent agreement with the measured citation distributions.

💡 Research Summary

The paper presents a rigorous empirical test of the preferential‑attachment (PA) hypothesis that underlies many scale‑free network models, using a large citation dataset from the field of physics. The authors collected the complete citation histories of 40,195 papers published in a single year (2004) and tracked all citations received over the subsequent ten years. By measuring the instantaneous citation rate λ(t)=Δk/Δt for each paper and relating it to the current cumulative citation count k(t), they found that the growth follows a super‑linear law λ∝k^α with α in the range 1.25–1.30. This result directly contradicts the classic Barabási–Albert model, which assumes linear PA (α=1).

In addition to the exponent, the authors examined temporal correlations in citation activity. They computed the autocorrelation function of the citation rate and discovered a statistically significant positive correlation for lags of one to two years. This indicates that recent citation bursts increase the likelihood of near‑future citations, i.e., the citation process possesses memory and cannot be modeled as a memoryless Markov chain.

To capture both the super‑linear attachment and the observed memory, the authors built a stochastic growth model based on a self‑exciting point process (Hawkes process). The conditional intensity for paper i at time t is defined as

λ_i(t)= μ + κ·k_i(t)^{α‑1}·∑_{t_j<t} φ(t‑t_j),

where μ is a baseline (exogenous) citation rate, κ controls the overall excitation strength, α is the empirically measured super‑linear exponent, and φ(τ) = (τ+τ₀)^{‑β} is a power‑law decay kernel that models how the influence of past citation events fades over time. The term k_i(t)^{α‑1} introduces the non‑linear reinforcement, while the sum over past events implements the memory effect.

Model parameters were estimated by maximizing the likelihood of the observed citation timestamps, using Bayesian optimization and Markov‑chain Monte Carlo sampling. The best‑fit values were μ≈0.02 citations per year, κ≈0.15, α≈1.28, β≈1.1, and τ₀≈0.5 yr. With these parameters, the authors simulated the growth of a synthetic network of the same size and initial conditions as the real dataset.

The simulation reproduced the empirical citation distribution P(k) with remarkable fidelity, especially in the heavy tail where high‑citation papers reside. It also matched the time evolution of the mean citation count ⟨k(t)⟩ and the autocorrelation function of the citation rate. By contrast, a conventional linear PA model (α=1, no memory) severely under‑estimates the proportion of highly cited papers and fails to generate any measurable autocorrelation.

The findings have several important implications. First, real citation dynamics are governed by a super‑linear reinforcement mechanism, implying that early citation advantage can be amplified dramatically, leading to a “rich‑get‑richer” effect stronger than previously thought. Second, the presence of short‑term memory suggests that citation bursts—perhaps triggered by media coverage, conference presentations, or topical relevance—play a crucial role in shaping future citation trajectories. Third, the Hawkes‑process framework provides a unified statistical description that can be calibrated to empirical data and extended to other growing networks such as patent citations, social media reposts, or epidemic spreading.

In conclusion, the authors demonstrate that the classic Barabási–Albert preferential‑attachment model is insufficient for describing citation network growth. By integrating a super‑linear attachment exponent and a temporally decaying excitation kernel within a self‑exciting point‑process model, they achieve an excellent quantitative match to real citation data. This work not only advances our understanding of scientific impact dynamics but also offers a versatile modeling toolkit for a broad class of complex, evolving networks.