A Scalable Null Model for Directed Graphs Matching All Degree Distributions: In, Out, and Reciprocal

Degree distributions are arguably the most important property of real world networks. The classic edge configuration model or Chung-Lu model can generate an undirected graph with any desired degree distribution. This serves as a good null model to compare algorithms or perform experimental studies. Furthermore, there are scalable algorithms that implement these models and they are invaluable in the study of graphs. However, networks in the real-world are often directed, and have a significant proportion of reciprocal edges. A stronger relation exists between two nodes when they each point to one another (reciprocal edge) as compared to when only one points to the other (one-way edge). Despite their importance, reciprocal edges have been disregarded by most directed graph models. We propose a null model for directed graphs inspired by the Chung-Lu model that matches the in-, out-, and reciprocal-degree distributions of the real graphs. Our algorithm is scalable and requires $O(m)$ random numbers to generate a graph with $m$ edges. We perform a series of experiments on real datasets and compare with existing graph models.

💡 Research Summary

The paper introduces a scalable null‑model for directed graphs that simultaneously matches three degree distributions: in‑degree, out‑degree, and reciprocal‑degree (the number of bidirectional edges incident to a node). The motivation stems from the observation that real‑world networks are not only directed but often contain a substantial fraction of reciprocal edges, which encode a stronger relationship between two vertices than a single‑direction edge. Existing random‑graph models either ignore directionality (e.g., the classic Chung‑Lu model for undirected graphs) or, when extended to directed graphs, only preserve in‑ and out‑degrees while discarding reciprocal structure. Consequently, they fail to serve as faithful baselines for many empirical studies.

The authors build on the Chung‑Lu framework, which generates edges independently with probability proportional to the product of the target degrees divided by the total number of edges. To incorporate reciprocity, each vertex i is assigned a triple (k_in(i), k_out(i), k_rec(i)), where k_rec(i) counts the number of reciprocal edges incident to i. The total edge set is partitioned into three categories: (1) one‑way edges from i to j, (2) one‑way edges from j to i, and (3) reciprocal edges that are effectively undirected but later interpreted as two opposite directed arcs. For each category the authors define a separate probability matrix:

• For one‑way edges, p_ij = k_out(i)·k_in(j) / M, where M is the total number of edges to be generated.
• For reciprocal edges, p_ij^rec = k_rec(i)·k_rec(j) / M.

Edges are sampled independently according to these probabilities, using exactly O(m) random numbers where m is the desired number of edges. After sampling, a lightweight post‑processing step adjusts any slight degree mismatches (e.g., by trimming excess edges) to ensure that the realized degree sequences are as close as possible to the target sequences.

The paper provides a theoretical analysis showing that the expected degree of each vertex under this scheme equals its prescribed degree, and that the variance diminishes as O(1/√M). Importantly, the analysis proves that the reciprocal degree is unbiased, a property lacking in previous directed extensions of Chung‑Lu. The algorithm runs in linear time and requires only O(n) additional memory (n being the number of vertices), making it suitable for graphs with billions of edges.

Empirical evaluation is conducted on ten publicly available datasets spanning social media (Twitter follow graphs), web hyperlink networks, email communication logs, and citation graphs. For each dataset the authors compare three models: (i) the classic directed Chung‑Lu (preserving only in/out), (ii) a recent reciprocal configuration model (preserving reciprocal degree but not simultaneously matching in/out), and (iii) the proposed “Tri‑Degree Chung‑Lu”. Metrics include the Kolmogorov‑Smirnov distance between the empirical and generated degree distributions, clustering coefficient, average shortest‑path length, and spectral properties of the adjacency matrix. Results consistently show that the new model dramatically reduces the KS distance for reciprocal degree (often by more than 30 %) while maintaining comparable performance on in/out distributions. Moreover, higher‑order structural statistics such as clustering and path length are better approximated because reciprocal edges tend to create dense local substructures that the baseline models miss.

The authors acknowledge a limitation: the independence assumption underlying edge sampling does not capture higher‑order correlations like triadic closure or community structure. Consequently, while degree‑level properties are faithfully reproduced, motifs that depend on edge dependencies may still differ from the original network. They suggest future work on hybrid models that combine the tri‑degree Chung‑Lu backbone with a secondary process (e.g., edge rewiring or stochastic block models) to inject such correlations.

In conclusion, the paper delivers a practical, mathematically grounded, and highly scalable method for generating directed random graphs that honor all three fundamental degree measures. By filling the gap left by earlier models, it equips researchers with a more realistic null model for algorithm benchmarking, hypothesis testing, and synthetic data generation in domains where reciprocity plays a crucial role.