Degree-Corrected Stochastic Block Models

• Degree-Corrected SBMs are statistical models that extend traditional stochastic block models by incorporating node-specific degree parameters to capture heterogeneous network connectivity.
• They improve community detection in networks with heavy-tailed degree distributions by addressing the limitations of classic SBMs that assume uniform node behavior.
• Inference methods such as likelihood maximization, spectral clustering, and Bayesian approaches enable robust, consistent recovery of community structures in complex networks.

The degree-corrected stochastic block model (DC-SBM) is a broad generalization of the traditional stochastic block model (SBM) designed to capture community structure in networks exhibiting within-group degree heterogeneity. DC-SBMs introduce node-specific degree parameters while retaining the latent community structure of classic block models, enabling accurate modeling and detection of communities in networks where empirical degree distributions are heavy-tailed, heterogeneous, or systematically distinct within blocks (Karrer et al., 2010).

1. Model Specification and Identifiability

Formally, a DC-SBM on $n$ nodes with $K$ communities consists of the following generative process:

• Each node $i$ is assigned a block $z_i \in \{1, \ldots, K\}$.
• Each node is endowed with a latent positive degree parameter $\theta_i > 0$.
• The block-connectivity matrix $B \in \mathbb{R}_+^{K \times K}$ is symmetric and full rank.
• For $i < j$, edges $A_{ij}$ are independent with

$$\mathbb{E}[A_{ij}] = \theta_i \theta_j B_{z_i, z_j}$$

or $$A_{ij} \sim \mathrm{Poisson}(\theta_i \theta_j B_{z_i,z_j})$$ in the multi-edge regime.

Identifiability of $K$0 is ensured up to permutation and community-wise scaling if every community contains at least three nodes, under the constraint $K$1 for each $K$2 (Park et al., 2024). The only remaining freedom is a scaling factor per community and a relabeling of the communities; for most purposes, community labels are considered equivalent under permutation.

2. Model Rationale and Statistical Properties

Classical SBMs assume all nodes in a community are stochastically equivalent, which leads to poor fit and misleading inference on empirical networks with broad degree distributions or hubs. DC-SBM explicitly models degree variation within blocks via node-specific parameters. This modification corrects the tendency of uncorrected SBMs to partition nodes by degree rather than true community (Karrer et al., 2010, Zhao et al., 2011).

Key theoretical properties include:

• Likelihood Decomposition and Profile Likelihood: The log-likelihood for a given partition can be written as

$K$3

where $K$4 is the summed edge count between blocks $K$5, and $K$6 is the total degree of nodes in block $K$7 (Karrer et al., 2010).

• Consistency of Community Recovery: Likelihood-based community assignments under DC-SBM are consistent under very mild conditions—specifically, when the expected degree diverges ($K$8) and without further constraints on the block parameters or node degrees (Zhao et al., 2011).
• Consistency for Number of Communities: Penalized likelihood criteria targeting the number of communities are strongly consistent under semi-sparse to dense regimes, with penalties that account both for block parameters and the node-specific degree sequence (Cerqueira et al., 2023, Ma et al., 2018).

3. Algorithmic Paradigms

Methods for inference in DC-SBMs include likelihood maximization, spectral clustering, convex relaxation, and Bayesian inference.

Likelihood Maximization

• The classical heuristic for maximizing the profile likelihood is a local, iterative algorithm akin to the Kernighan–Lin method: nodes are reassigned to new communities to maximize the objective, with best-prefix rollback to escape poor local maxima. This procedure achieves robust performance and directly optimizes the theoretically justified criterion (Karrer et al., 2010).

Spectral Methods

• Normalized or Regularized Laplacian: Spectral clustering methods for DC-SBM use the normalized Laplacian $K$9 or its regularized variant. The top $i$0 eigenvectors are embedded and input to $i$1-means; row normalization is essential to remove the "star-shape" caused by degree heterogeneity (Wan et al., 2021, Qin et al., 2013).
• Consistent Recovery and Min-Degree: Recovery is strongly consistent provided the minimum expected degree grows at least logarithmically (with regularization, no strict lower bound is required) (Su et al., 2017, Qin et al., 2013).

Convex and SDP Relaxations

• Convexified Modularity Maximization: The modularity objective is adapted to the degree-corrected setting and relaxed to an SDP, followed by $i$2-median or $i$3-means rounding. Theoretical guarantees cover both exact and approximate recovery, and the method achieves optimal rates even for bounded-degree regimes (Chen et al., 2015).
• Outlier-Robust Extensions: SDP relaxations with additional node-wise penalties can exactly recover communities even in the presence of adversarial outliers and heavy-tailed degree distributions (Qian et al., 2019).

Bayesian Approaches

• Nonparametric Bayesian DC-SBM: Infinite (Dirichlet process) versions of the DC-SBM automatically infer the number of clusters and the degree of correction, with analytical posterior updates for the conjugate construction (Herlau et al., 2013).
• Logistic Link/Polya-Gamma Augmentation: Bayesian logistic DC-SBM with node-wise additive parameters and Polya–Gamma augmentation allows for fully sampled inference and principled centroid estimators that address label-switching (Peng et al., 2013).

4. Statistical Theory: Hierarchies, Generalizations, and Limits

DC-SBM sits within a hierarchy of random graph models:

• SBM $i$4 DC-SBM $i$5 PABM (Popularity Adjusted Block Model): Degree correction is a single scalar per node (PLD-SBM introduces a prior to approximate power-laws), while PABM allows for connection probabilities to depend on both sender and receiver clusters (requiring $i$6 parameters) (Noroozi et al., 2020, Qiao et al., 2019).
• Nested Block Models: Intermediate forms between DC-SBM and PABM allow hierarchical mixtures, controlling the parameter complexity (Noroozi et al., 2020).

For multi-layer networks, DC-SBM extensions enable inference and Gaussian asymptotics for layer-specific connectivity matrices, supporting hypothesis testing and confidence intervals for link intensities across layers (Su et al., 2024).

5. Information-Theoretic and Probabilistic Limits

There exist fundamental detectability limitations in sparse DC-SBMs:

• Non-Reconstruction Threshold: Reconstruction of communities is information-theoretically impossible when

$i$7

where $i$8 and $i$9 parameterize the within- and between-block connection probabilities, $z_i \in \{1, \ldots, K\}$0 is the number of communities, and $z_i \in \{1, \ldots, K\}$1 is the second moment of the degree distribution. In this regime, even the mutual information between any algorithmic estimate and the planted assignment vanishes asymptotically (Gulikers et al., 2015).

• Local Weak Limits and Branching Process Coupling: Sparse DC-SBM local neighborhoods converge to multi-type Galton–Watson trees, with component structures rigorously characterized via breadth-first walks and excursion representations (Gulikers et al., 2015, Jr et al., 2024).
• Reduction to Wigner-Type Ensembles: Random matrix theory can be leveraged for community detection and testing by centering and scaling the adjacency matrix to a standardized Wigner form, allowing for explicit significance control and spectral hypothesis testing in block models with degree correction (Malinas et al., 2023).

6. Empirical Performance and Applications

Extensive empirical evidence demonstrates the necessity of DC-SBM when degree variation is present:

• On synthetic and real networks (e.g., political blogs, social networks), DC-SBMs substantially outperform non-degree-corrected models, which otherwise cluster by degree instead of community labels (Karrer et al., 2010, Chen et al., 2015, Qiao et al., 2019).
• Degree correction is only statistically necessary when substantial within-community heterogeneity exists; otherwise, standard SBMs perform similarly but involve estimating fewer parameters (Zhao et al., 2011, Herlau et al., 2013).

DC-SBM is used for unsupervised community detection, model selection (number of communities), link prediction, hypothesis testing concerning network structure, and for robust recovery in contaminated or adversarial settings.

7. Model Extensions, Open Problems, and Hierarchies

DC-SBM forms the basis for a range of advanced models:

• Power-law and Heavy-tailed Extensions: PLD-SBM and Bayesian DC-SBM variants use explicit priors or parameterizations to ensure scale-free structure in the degree sequence (Qiao et al., 2019, Herlau et al., 2013).
• Cascading and Multi-layer Extensions: Statistical inference for layered networks and time-varying settings generalize naturally via scalable spectral and variance-adjusted procedures (Su et al., 2024).
• Open Directions: Challenges include optimal algorithms for ultra-sparse regimes, sharper non-reconstructibility phases, robust scalable inference for massive graphs, and information-theoretic limits in the presence of adversarial or dynamic contamination (Gulikers et al., 2015, Qian et al., 2019).

Practical recommendations emphasize using DC-SBM when empirical degrees are heterogeneous, leveraging model selection criteria tailored for degree-corrected settings, and employing spectral or convexified clustering methods with regularization and row normalization. For model-based inference, either direct likelihood maximization or conjugate Bayesian methods can be employed, with the choice guided by application needs and network scale.