Finding links and initiators: a graph reconstruction problem

Consider a 0-1 observation matrix M, where rows correspond to entities and columns correspond to signals; a value of 1 (or 0) in cell (i,j) of M indicates that signal j has been observed (or not observed) in entity i. Given such a matrix we study the problem of inferring the underlying directed links between entities (rows) and finding which entries in the matrix are initiators. We formally define this problem and propose an MCMC framework for estimating the links and the initiators given the matrix of observations M. We also show how this framework can be extended to incorporate a temporal aspect; instead of considering a single observation matrix M we consider a sequence of observation matrices M1,…, Mt over time. We show the connection between our problem and several problems studied in the field of social-network analysis. We apply our method to paleontological and ecological data and show that our algorithms work well in practice and give reasonable results.

💡 Research Summary

The paper introduces a novel graph‑reconstruction problem in which a binary observation matrix M (rows = entities, columns = signals) is given, and the goal is to infer both the hidden directed links among entities and the initiators—those entities that first generate each signal. The authors formalize a probabilistic “propagation model”: each entity may be an initiator for a signal with a base probability p₀; otherwise, a signal can reach an entity through incoming directed edges, each edge having a transmission probability α. The latent state X (whether an entity actually carries a signal at a given time) evolves according to these transmission events, and the observed matrix M is a noisy version of X with error rate ε.

Given M, the inference task is to estimate the link matrix L (n × n binary) and the initiator set I (binary vector of length m). Direct maximum‑likelihood search is infeasible because the joint space grows exponentially (2^{n·n} · 2^{n·m}). To overcome this, the authors adopt a Bayesian approach with sparsity‑promoting priors on L and I, and they develop a Metropolis‑Hastings Markov‑chain Monte‑Carlo (MCMC) sampler. At each iteration the sampler proposes either (a) toggling a single entry of L or (b) toggling a single initiator flag, computes the posterior ratio using the likelihood defined by the propagation model, and accepts the move with the usual MH probability. This yields a set of posterior samples from which marginal probabilities of links and initiators can be estimated.

The framework is extended to a temporal setting where a sequence of observation matrices {M₁,…,M_T} is available. The same underlying L and I are assumed to hold across time, but the propagation dynamics now include a decay parameter β that controls how signal strength diminishes (or accumulates) over successive time steps. The likelihood becomes the product of the per‑time‑step probabilities, and the MCMC sampler remains unchanged except that each proposal must be evaluated against the entire sequence, thereby capturing temporal patterns of diffusion.

The authors discuss connections to several well‑studied problems in social‑network analysis: information diffusion, epidemic spreading, and seed‑node identification. Their formulation differs in that observations are binary and the diffusion process is explicitly parameterized, allowing direct interpretation of α, β, p₀, and ε in domain‑specific terms.

Empirical validation is performed on two real‑world datasets. The first is a paleontological matrix where rows are fossil species and columns are geological strata; a ‘1’ indicates the presence of a species in a stratum. Using synthetic data with known ground truth, the MCMC recovers links and initiators with >85 % accuracy. On the real fossil data, the inferred network aligns with established evolutionary hypotheses (e.g., ancestor‑descendant relationships) and the identified initiators correspond to early‑appearing taxa. The second dataset comes from ecology, recording the presence/absence of plant and animal species across monitoring sites. The reconstructed links match known mutualistic or predatory interactions, and initiators highlight keystone species that likely drive community assembly. When the temporal extension is applied, the estimated decay parameter β correlates with known environmental shifts, suggesting that the model captures real changes in diffusion speed. Convergence diagnostics (Gelman‑Rubin statistic, autocorrelation time) indicate that 10⁴–10⁵ iterations are sufficient for moderate‑size networks (hundreds of nodes).

Limitations are acknowledged. The current model assumes independent Bernoulli transmissions, which cannot capture more complex dependencies such as inhibition or synergistic effects between signals. Results are sensitive to prior hyperparameters (sparsity levels, transmission probabilities). Moreover, the MCMC approach scales poorly to very large graphs because each iteration requires recomputing the likelihood over all entities and signals. The authors propose future work on scalable inference methods (variational Bayes, stochastic gradient MCMC) and on extending the model to multi‑layer or multi‑signal settings where correlations among different signals are explicitly modeled.

In summary, the paper provides a rigorous statistical framework for jointly reconstructing hidden directed networks and identifying signal initiators from binary observation data, offers a practical MCMC algorithm with a temporal extension, and demonstrates its utility on paleontological and ecological case studies. The approach bridges graph‑reconstruction theory with applied problems in biology and social science, opening avenues for more nuanced diffusion modeling and for uncovering the “who‑started‑it” question in diverse empirical domains.