Generalized friendship paradox in complex networks: The case of scientific collaboration

The friendship paradox states that your friends have on average more friends than you have. Does the paradox “hold” for other individual characteristics like income or happiness? To address this question, we generalize the friendship paradox for arbitrary node characteristics in complex networks. By analyzing two coauthorship networks of Physical Review journals and Google Scholar profiles, we find that the generalized friendship paradox (GFP) holds at the individual and network levels for various characteristics, including the number of coauthors, the number of citations, and the number of publications. The origin of the GFP is shown to be rooted in positive correlations between degree and characteristics. As a fruitful application of the GFP, we suggest effective and efficient sampling methods for identifying high characteristic nodes in large-scale networks. Our study on the GFP can shed lights on understanding the interplay between network structure and node characteristics in complex networks.

💡 Research Summary

The paper extends the classic Friendship Paradox (FP), which states that “your friends have on average more friends than you do,” to arbitrary node attributes in complex networks, coining the term Generalized Friendship Paradox (GFP). The authors formalize GFP at two levels: (i) the individual level, where a node i experiences the paradox if its attribute x_i is smaller than the average attribute of its neighbors, and (ii) the network level, where the overall average attribute across all nodes is smaller than the average attribute of neighbors (weighted by degree). When the attribute x is set to the node degree k, GFP reduces to the traditional FP.

Two empirical collaboration networks are analyzed. The first is a co‑authorship network constructed from Physical Review (PR) journals, comprising 24,259 authors; four attributes are examined: number of co‑authors, total citations, number of publications, and average citations per publication. The second is a Google Scholar (GS) profile‑based network of 29,968 researchers, for which number of co‑authors and total citations are considered. For each node i, the degree k_i and attribute x_i are recorded, and the Pearson correlation ρ_kx between degree and attribute is computed, along with degree assortativity r_kk and attribute assortativity r_xx.

Empirical results show that the probability H that a randomly chosen node satisfies the individual‑level GFP exceeds 0.7 for every attribute, indicating that most nodes have neighbors with higher attribute values. The conditional probability h(k, x) that a node with degree k and attribute x experiences the paradox declines as x increases for fixed k, confirming the intuitive notion that higher‑valued nodes are less likely to be “under‑paradox.” The dependence of h(k, x) on degree varies across attributes: for citations and publications (ρ_kx≈0.79) h(k, x) rises with k, whereas for average citations per paper (ρ_kx≈0.07) it remains essentially flat. This behavior is explained analytically: the difference F between neighbor‑average and node‑average attributes can be expressed as F = ρ_kx σ_k σ_x ⟨k⟩, so a positive degree‑attribute correlation guarantees F > 0 and thus the GFP.

The authors then propose two sampling strategies that exploit GFP to locate high‑attribute nodes in large networks where full mapping is infeasible. (i) Friend sampling selects a random set of seed nodes and then picks one random neighbor for each seed, forming a “friend group.” (ii) Biased sampling also starts from random seeds but chooses, for each seed, the neighbor with the highest attribute value, forming a “biased group.” Both methods are compared against pure random sampling. Across both PR and GS networks, biased sampling consistently yields the highest concentration of top‑attribute nodes, while friend sampling outperforms random sampling whenever ρ_kx is appreciably positive. In the special case of average citations per paper (ρ_kx≈0.07), friend sampling offers little advantage over random sampling, yet biased sampling still provides a clear benefit, demonstrating its robustness even when degree‑attribute correlation is weak.

To further test the role of ρ_kx, the authors generate an auxiliary attribute X using a Choelesky‑type decomposition that allows controlled manipulation of the correlation between degree and X. Experiments confirm that as ρ_kX increases, the performance gap between friend sampling and random sampling widens, while biased sampling remains superior regardless of the correlation magnitude.

The discussion highlights the broader implications of GFP. Because individuals often compare themselves to their friends, the GFP offers a mechanistic explanation for systematic over‑ or under‑estimation of personal status (popularity, income, reputation, happiness). Moreover, in dynamical processes such as epidemic spreading, information diffusion, or resource allocation, identifying high‑degree or high‑activity nodes quickly is crucial; GFP‑based sampling provides a low‑cost, locally implementable method for doing so. Limitations include incomplete data in the GS network and the requirement of neighbor‑attribute information for biased sampling. Future work is suggested on networks where degree‑attribute correlation is negative (anti‑GFP) and on dynamic or multiplex networks.

In conclusion, the study demonstrates that the friendship paradox is a special case of a more general phenomenon that applies to any node attribute positively correlated with degree. The positive degree‑attribute correlation is identified as the universal driver of the GFP. By leveraging this insight, the authors develop simple yet effective sampling techniques for locating high‑attribute nodes, offering both theoretical insight into network‑driven perception biases and practical tools for network analysis and intervention.