[Paper Review] Analytical Formulation of the Block-Constrained Configuration Model

Giona Casiraghi|arXiv (Cornell University)|Nov 12, 2018

Complex Network Analysis Techniques32 references4 citations

TL;DR

This paper introduces the block-constrained configuration model (BCCM), a generative random graph model that extends the configuration model by enforcing block structure constraints during edge rewiring, preserving both degree distributions and community structure. The model is analytically tractable and efficiently fitted using the generalized hypergeometric ensemble framework, enabling accurate community detection and model selection via AIC/BIC on real-world networks.

ABSTRACT

We provide a novel family of generative block-models for random graphs that naturally incorporates degree distributions: the block-constrained configuration model. Block-constrained configuration models build on the generalised hypergeometric ensemble of random graphs and extend the well-known configuration model by enforcing block-constraints on the edge generation process. The resulting models are analytically tractable and practical to fit even to large networks. These models provide a new, flexible tool for the study of community structure and for network science in general, where modelling networks with heterogeneous degree distributions is of central importance.

Motivation & Objective

• To address the limitation of standard stochastic block models (SBMs) in reproducing empirical degree sequences by integrating degree heterogeneity naturally into block structure modeling.
• To develop a generative model that preserves the degree sequence of real networks while enforcing community or block structure constraints.
• To enable fast, analytically tractable inference and model selection for large networks using the generalized hypergeometric ensemble framework (gHypEG).
• To provide a flexible, principled alternative to degree-corrected block models by embedding degree correction directly into the edge-generation process.
• To demonstrate the utility of BCCM through model fitting and comparison using AIC, BIC, and information criteria on real-world graphs.

Proposed method

• The BCCM extends the configuration model by constraining edge rewiring to occur only within predefined blocks, preserving the original degree sequence.
• It uses the generalized hypergeometric ensemble of random graphs (gHypEG) as a foundation, where edge probabilities are determined by vertex degrees and block-specific propensities.
• The model defines a block-structured propensity matrix Ω, where edge probabilities depend on both degrees and block memberships, ensuring block-constrained edge formation.
• The closed-form probability distribution of gHypEG enables efficient likelihood computation, parameter estimation, and model evaluation via information criteria (AIC/BIC).
• The framework supports directed and undirected graphs with or without self-loops, and allows for extensions to heterogeneous vertex or edge properties.
• Model fitting is performed by maximizing the likelihood under the gHypEG framework, with block structure inferred from community detection algorithms.

Experimental results

Research questions

• RQ1 Can a generative random graph model preserve both the empirical degree sequence and block structure of real networks?
• RQ2 Does the block-constrained configuration model (BCCM) provide a more accurate fit to real-world networks than standard SBMs or degree-corrected models?
• RQ3 Can the gHypEG framework enable fast, analytically tractable inference and model selection for large-scale networks with community structure?
• RQ4 What is the optimal number of blocks for modeling real-world networks, as determined by information criteria like AIC and BIC?
• RQ5 How does the BCCM compare to existing community detection algorithms in terms of model fit and statistical significance?

Key findings

• The BCCM successfully preserves the degree distribution of real-world networks while enforcing block structure, avoiding the need for ad hoc degree corrections.
• On the rfid network, the fast_greedy and louvain algorithms produced the lowest AIC (44721.18) and BIC (68161.88), indicating the best model fit among tested methods.
• For the karate network, the louvain algorithm yielded the lowest AIC (1707.768) and BIC (3669.947), suggesting optimal block partitioning with minimal information loss.
• On the USairports network, the louvain algorithm detected 21 blocks and achieved the lowest AIC (1210517) and BIC (5808828), outperforming other algorithms despite spinglass failing due to graph disconnection.
• The enron network showed that the louvain algorithm achieved the lowest AIC (328924.2) and BIC (659185.1), indicating strong statistical support for its block partitioning.
• The framework enables likelihood-ratio tests and information criteria-based model selection, providing a principled method to assess the statistical significance of block structures.