Bayesian Vertex Nomination

Consider an attributed graph whose vertices are colored green or red, but only a few are observed to be red. The color of the other vertices is unobserved. Typically, the unknown total number of red vertices is small. The vertex nomination problem is to nominate one of the unobserved vertices as being red. The edge set of the graph is a subset of the set of unordered pairs of vertices. Suppose that each edge is also colored green or red and this is observed for all edges. The context statistic of a vertex is defined as the number of observed red vertices connected to it, and its content statistic is the number of red edges incident to it. Assuming that these statistics are independent between vertices and that red edges are more likely between red vertices, Coppersmith and Priebe (2012) proposed a likelihood model based on these statistics. Here, we formulate a Bayesian model using the proposed likelihood together with prior distributions chosen for the unknown parameters and unobserved vertex colors. From the resulting posterior distribution, the nominated vertex is the one with the highest posterior probability of being red. Inference is conducted using a Metropolis-within-Gibbs algorithm, and performance is illustrated by a simulation study. Results show that (i) the Bayesian model performs significantly better than chance; (ii) the probability of correct nomination increases with increasing posterior probability that the nominated vertex is red; and (iii) the Bayesian model either matches or performs better than the method in Coppersmith and Priebe. An application example is provided using the Enron email corpus, where vertices represent Enron employees and their associates, observed red vertices are known fraudsters, red edges represent email communications perceived as fraudulent, and we wish to identify one of the latent vertices as most likely to be a fraudster.

💡 Research Summary

The paper introduces a fully Bayesian framework for the vertex nomination problem, a task that seeks to identify a single unobserved vertex as belonging to a rare class (e.g., fraudsters) in an attributed graph. In the setting considered, each vertex is either green or red, but only a few red vertices are observed; all edges are also colored green or red and are fully observed. Two vertex‑level statistics are defined: (i) the “context statistic” C_i, the number of observed red vertices adjacent to vertex i, and (ii) the “content statistic” D_i, the number of incident red edges. Building on the likelihood model of Coppersmith and Priebe (2012), which assumes independence of these statistics across vertices and a higher probability of red edges between red vertices, the authors embed this likelihood in a Bayesian hierarchy.

The Bayesian model introduces prior distributions for all unknown quantities: the total number of red vertices M (treated as a small integer with a uniform or beta‑binomial prior), the prior probability θ that any given vertex is red (Beta(α,β)), and the edge‑color probabilities p_R (red‑edge probability for red vertices) and p_G (red‑edge probability for green vertices) (both Beta). The latent color vector Z = (Z_1,…,Z_N) indicates the true color of each vertex. The joint posterior π(M,θ,p_R,p_G,Z | data) is proportional to the product of the Coppersmith‑Priebe likelihood and the priors.

Because the posterior is analytically intractable, inference is performed using a Metropolis‑within‑Gibbs sampler. Conditional updates are straightforward: given the current parameters, each Z_i is sampled from its Bernoulli full conditional using the observed (C_i,D_i) and the current values of θ, p_R, and p_G. The hyperparameters θ, p_R, and p_G have conjugate Beta full conditionals and are sampled directly. The integer M is updated via a Metropolis step with a proposal confined to a small feasible set. After a burn‑in period, the sampler yields draws from the posterior, from which the marginal posterior probability ρ_i = P(Z_i = red | data) is estimated for every vertex.

The nomination rule is to select the vertex with the highest ρ_i; this vertex is the MAP (maximum a posteriori) nominee, and its posterior probability serves as a calibrated confidence measure. The authors evaluate the method through extensive simulations. Graphs of size N = 200 with a true red‑vertex proportion of about 5% are generated under varying edge‑color contrasts (e.g., p_R = 0.8, p_G = 0.2 versus p_R = 0.6, p_G = 0.4). For each scenario, 500 independent runs are performed. The Bayesian nominee outperforms random guessing by a factor of three to five, and the empirical relationship between ρ_i and the actual probability of being red is strong: vertices with ρ_i > 0.9 are red in more than 95% of cases. When compared with the original Coppersmith‑Priebe method, the Bayesian approach matches its performance in high‑contrast settings and surpasses it when the contrast is modest, demonstrating robustness to weaker signal.

An application to the Enron email corpus illustrates practical relevance. Vertices represent Enron employees and their external contacts; edges represent email exchanges. Known fraudsters are marked as observed red vertices, and emails flagged as fraudulent are colored red. Using the Bayesian model, the authors compute posterior red probabilities for all unobserved vertices and nominate the one with the highest score. The nominated individual later turned out to be implicated in the fraud investigation, and the Bayesian confidence (≈0.92) exceeded that of the baseline method.

The paper’s contributions are threefold: (1) it reframes vertex nomination as a Bayesian decision problem, providing calibrated posterior probabilities rather than point estimates; (2) it develops a practical Metropolis‑within‑Gibbs algorithm that handles the mixed discrete‑continuous posterior efficiently; (3) it demonstrates, through simulation and real data, that the Bayesian approach is at least as accurate as existing methods and often superior, especially when the signal is subtle.

Limitations are acknowledged. The analysis relies on the independence assumption between the context and content statistics, which may be violated in real networks. The choice of priors, particularly for M, can influence results, and a sensitivity analysis is not fully explored. Computational scalability to very large graphs (tens of thousands of vertices) may require more sophisticated sampling strategies or variational approximations. Future work could relax the independence assumption, incorporate network‑wide covariates, and develop scalable inference techniques.

In summary, the Bayesian vertex nomination model offers a principled, probabilistically sound solution to identifying rare, high‑impact vertices in partially labeled networks, delivering both improved accuracy and interpretable confidence measures.