Stochastic block model and exploratory analysis in signed networks

We propose a generalized stochastic block model to explore the mesoscopic structures in signed networks by grouping vertices that exhibit similar positive and negative connection profiles into the same cluster. In this model, the group memberships are viewed as hidden or unobserved quantities, and the connection patterns between groups are explicitly characterized by two block matrices, one for positive links and the other for negative links. By fitting the model to the observed network, we can not only extract various structural patterns existing in the network without prior knowledge, but also recognize what specific structures we obtained. Furthermore, the model parameters provide vital clues about the probabilities that each vertex belongs to different groups and the centrality of each vertex in its corresponding group. This information sheds light on the discovery of the networks’ overlapping structures and the identification of two types of important vertices, which serve as the cores of each group and the bridges between different groups, respectively. Experiments on a series of synthetic and real-life networks show the effectiveness as well as the superiority of our model.

💡 Research Summary

The paper introduces a generalized stochastic block model (SBM) designed specifically for signed networks, which contain both positive and negative edges. Traditional SBMs and most community‑detection methods assume only positive links, limiting their ability to uncover the rich mesoscopic structures that can arise when antagonistic relationships are present. To address this, the authors propose the Signed Stochastic Block Model (SSBM), which augments the classic SBM with two separate block matrices: ω⁺ governing the probability of a positive edge from group r to group s, and ω⁻ governing the probability of a negative edge for the same ordered pair of groups.

Vertices are assigned soft group memberships: θᵢʳ denotes the probability that vertex i belongs to group r as a source, while φⱼˢ denotes the probability that vertex j belongs to group s as a target. These parameters capture the “centrality” of a vertex within each group from the perspective of outgoing and incoming edges, respectively. All parameters are normalized (∑ᵣθᵢʳ = 1, ∑ₛφⱼˢ = 1, ∑ᵣ,ₛ ω⁺ᵣₛ = 1, ∑ᵣ,ₛ ω⁻ᵣₛ = 1).

Model fitting is performed via the Expectation–Maximization (EM) algorithm. In the E‑step, the posterior probabilities q⁺ᵢⱼʳˢ and q⁻ᵢⱼʳˢ that a given edge (i, j) originates from group r and terminates in group s are computed using the current estimates of ω, θ, and φ. In the M‑step, these posteriors are used to update the parameters analytically: ω⁺ᵣₛ becomes the weighted fraction of positive edges assigned to (r, s), ω⁻ᵣₛ analogously for negative edges, while θ and φ are updated by aggregating the contributions of all incident edges (both signs) weighted by the corresponding q values. The updates automatically satisfy the normalization constraints.

The computational cost per EM iteration is O(m c²), where m is the number of edges and c the number of groups; the algorithm typically converges in a modest number of iterations, and multiple random initializations are used to mitigate the risk of local optima.

A key advantage of SSBM is that it yields a soft partition of the network. Each vertex receives a probability distribution over groups, enabling the detection of overlapping (or “fuzzy”) community structures that are common in real‑world signed networks. Moreover, the magnitude of θ and φ provides a quantitative measure of a vertex’s centrality within its group, allowing the identification of two special classes of vertices: (1) group cores—vertices with high intra‑group centrality, and (2) bridges—vertices that have substantial membership probabilities in multiple groups and thus mediate inter‑group connectivity.

The authors validate the model on synthetic benchmarks designed to exhibit pure assortative (community), disassortative (bipartite/multipartite), core‑periphery, and mixed patterns. In each case, the learned ω⁺ and ω⁻ matrices clearly reflect the underlying structure, and the soft assignments correctly recover overlapping nodes. Comparisons with existing signed‑network methods—such as signed modularity maximization, signed Potts models, and standard SBM extensions—show that SSBM achieves higher accuracy, especially when negative edges dominate the structural signal.

Real‑world experiments on several signed social and reputation networks further demonstrate the model’s practical utility. The method uncovers meaningful mesoscopic patterns without any prior specification of the number or type of structures, and the identified core and bridge vertices correspond to influential users or opinion leaders in the datasets.

Limitations are acknowledged: EM converges only to local maxima, the number of groups c must be supplied (though the authors discuss model‑selection heuristics), and the O(c²) term may become prohibitive for very large c. Future work is suggested in the direction of Bayesian non‑parametric extensions to infer c automatically, scalable sparse‑matrix implementations, and dynamic extensions for temporally evolving signed networks.

Overall, the paper delivers a comprehensive, mathematically grounded framework for exploratory analysis of signed networks, unifying community, bipartite, core‑periphery, and overlapping structures under a single probabilistic model and providing interpretable parameters for both global patterns and vertex‑level roles.