[Paper Review] Community Detection in the Labelled Stochastic Block Model

Simon Heimlicher, Marc Lelarge+1 more|arXiv (Cornell University)|Sep 13, 2012

Complex Network Analysis Techniques10 references73 citations

TL;DR

This paper proposes a generalized detectability threshold for community detection in labelled stochastic block models, where interactions carry multiple types (labels). Using belief propagation and tree reconstruction analysis, it establishes that a threshold in the signal-to-noise ratio of labels and connection probabilities determines whether communities can be reconstructed with correlation to the true partition. The key contribution is a unified condition that predicts the onset of detectability across multiple inference models.

ABSTRACT

We consider the problem of community detection from observed interactions between individuals, in the context where multiple types of interaction are possible. We use labelled stochastic block models to represent the observed data, where labels correspond to interaction types. Focusing on a two-community scenario, we conjecture a threshold for the problem of reconstructing the hidden communities in a way that is correlated with the true partition. To substantiate the conjecture, we prove that the given threshold correctly identifies a transition on the behaviour of belief propagation from insensitive to sensitive. We further prove that the same threshold corresponds to the transition in a related inference problem on a tree model from infeasible to feasible. Finally, numerical results using belief propagation for community detection give further support to the conjecture.

Motivation & Objective

• To extend the detectability threshold from unlabelled to labelled stochastic block models, where interaction types are observed.
• To determine under what conditions community structure can be reconstructed from labelled network data.
• To validate the threshold through belief propagation, tree reconstruction, and numerical experiments.
• To unify theoretical transitions in belief propagation sensitivity and tree reconstruction feasibility under a single condition.

Proposed method

• Formulates a generalized detectability condition based on the ratio of label-dependent edge probabilities and connection rates.
• Applies belief propagation to infer node communities from labelled graphs, analyzing its sensitivity to initial conditions.
• Uses a tree-based model to study reconstruction feasibility, linking it to the detectability threshold.
• Employs Rayleigh's monotonicity law and effective resistance in random graphs to analyze convergence and stability.
• Derives a critical threshold τ = 1 as the boundary between infeasible and feasible inference.
• Uses the overlap metric Q to numerically evaluate reconstruction performance across varying ε and a=b parameters.

Experimental results

Research questions

• RQ1 Under what conditions can community structure be detected in a labelled stochastic block model?
• RQ2 How does the inclusion of interaction labels affect the detectability threshold compared to unlabelled models?
• RQ3 Does the same threshold govern the sensitivity of belief propagation and the feasibility of tree reconstruction?
• RQ4 Can numerical belief propagation results validate the theoretical threshold for detectability?

Key findings

• The proposed threshold τ > 1 correctly identifies the transition from insensitivity to sensitivity in belief propagation.
• The same threshold τ = 1 corresponds to the onset of feasible reconstruction in the associated labelled tree model.
• Numerical results show that belief propagation achieves non-zero overlap Q only when ε > 1/(2√a), confirming the theoretical threshold.
• For a = b, reconstruction is only possible when ε > 1/(2√a), demonstrating that labels alone enable detection in symmetric cases.
• When a < b, the threshold shifts leftward, showing that both edge density and label informativeness jointly improve detectability.
• Below the threshold, overlap Q remains near zero across multiple seeds, indicating consistent failure; above it, Q increases steadily, showing reliable detection.