Multi-Hypergraph Stochastic Block Model

• MHSBM is a probabilistic framework that generalizes the stochastic block model to analyze latent communities across multiple, heterogeneous hypergraphs.
• It employs Poisson-based likelihoods and variational EM inference to model both high-order intra-hypergraph and inter-hypergraph relationships.
• Empirical evaluations show that MHSBM outperforms single-layer models in community detection and link prediction across diverse real-world datasets.

The Multi-Hypergraph Stochastic Block Model (MHSBM) is a probabilistic framework for modeling and inferring latent community structure in settings where multiple hypergraphs—potentially of heterogeneous type, node set, or domain—encode higher-order relations among elements. The model generalizes the classical stochastic block model (SBM) for graphs to collective ensembles of hypergraphs, accounting both for high-order intrarelationships (hyperedges of arbitrary cardinality) and inter-hypergraph couplings, thus providing a unifying mechanism to analyze integrative, multi-relational datasets encountered in biological, social, and information systems (Alaluusua et al., 2023, Ni et al., 8 May 2025).

1. Model Specification: Structural and Generative Elements

MHSBM operates over $L$ hypergraphs $H^{1}, \dots, H^{L}$, with each $H^{l} = (V^{l}, E^{l})$ and $E^{l}$ containing hyperedges of arbitrary cardinality. Nodes $v_{i}^{l}$ possess community memberships $u_{i}^{l} \in \Delta_{K^{l}}$ (the unit simplex in $\mathbb{R}^{K^{l}}$), allowing for mixed-membership assignments.

Each hypergraph $l$ is parameterized by an affinity matrix $w^{l} \in \mathbb{R}_{\ge 0}^{K^{l} \times K^{l}}$ dictating community-level interaction strengths, while each ordered hypergraph pair $(l, l')$ is associated with an inter-hypergraph affinity $H^{1}, \dots, H^{L}$0. Realizations of intra-hypergraph hyperedges $H^{1}, \dots, H^{L}$1 and inter-hypergraph connections $H^{1}, \dots, H^{L}$2 are modeled as independent Poisson random variables with rates defined by the corresponding membership and affinity parameters.

To capture node-specific influence in hyperedge formation and model preferential attachment, MHSBM incorporates a hyperedge-internal-degree matrix $H^{1}, \dots, H^{L}$3, where $H^{1}, \dots, H^{L}$4 quantifies node $H^{1}, \dots, H^{L}$5’s weighted participation in hyperedge $H^{1}, \dots, H^{L}$6. For $H^{1}, \dots, H^{L}$7, $H^{1}, \dots, H^{L}$8 always holds.

The generative process for hyperedge weights is: $H^{1}, \dots, H^{L}$9 with $H^{l} = (V^{l}, E^{l})$0 the normalization for node pairs.

Cross-hypergraph edges are realized as: $H^{l} = (V^{l}, E^{l})$1 This structure allows MHSBM to encode arbitrary interdependence between hypergraphs and flexible within-hypergraph connectivity, subsuming both assortative and disassortative regimes.

2. Likelihood, Posterior, and Variational Inference

Given observed $H^{l} = (V^{l}, E^{l})$2, the likelihood under latent parameters $H^{l} = (V^{l}, E^{l})$3 factorizes as: $H^{l} = (V^{l}, E^{l})$4 In the maximum-likelihood setting, parameters are estimated directly. A Bayesian formulation introduces Dirichlet/Gamma priors on $H^{l} = (V^{l}, E^{l})$5, $H^{l} = (V^{l}, E^{l})$6 and yields a posterior proportional to $H^{l} = (V^{l}, E^{l})$7.

Inference employs expectation–maximization (EM) with variational approximations. Due to combinatorial explosion in possible hyperedges, negative sampling is used: for each $H^{l} = (V^{l}, E^{l})$8, only the observed ($H^{l} = (V^{l}, E^{l})$9) and sampled “negative” ($E^{l}$0) hyperedges contribute to the likelihood.

E-step updates variational weights for each hyperedge and cross-hypergraph edge; M-step computes closed-form parameter updates given current variational parameters. Multiple EM restarts are advocated to mitigate local optima (Ni et al., 8 May 2025).

3. Aggregate Similarity Matrix and SDP-Based Recovery

When restricting to multilayer, $E^{l}$1-uniform, single-node-set hypergraphs, MHSBM reduces to the multilayer hypergraph SBM analyzed via semidefinite programming (SDP) (Alaluusua et al., 2023). Each layer $E^{l}$2 possesses adjacency tensor $E^{l}$3. The aggregate similarity matrix $E^{l}$4, with $E^{l}$5 the total number of hyperedges (across all layers) containing both $E^{l}$6 and $E^{l}$7, forms the weighted adjacency matrix for clustering.

Community recovery is formulated as the NP-hard quadratic program: $E^{l}$8 relaxed to an SDP over $E^{l}$9, with constraints $v_{i}^{l}$0, $v_{i}^{l}$1, and $v_{i}^{l}$2. Disassortative cases replace $v_{i}^{l}$3. Exact recovery conditions are characterized in terms of an information rate parameter $v_{i}^{l}$4; if $v_{i}^{l}$5, exact recovery of the true labeling (up to global sign) occurs with high probability as $v_{i}^{l}$6 (Alaluusua et al., 2023).

Proofs leverage dual-certificate construction, matrix concentration, and diagonal lower bounds to establish uniqueness of the SDP optimum.

4. Performance Evaluation and Empirical Results

Empirical assessment of MHSBM targets three axes: community detection, hyperedge reconstruction, and inter-hypergraph edge inference. Experimental datasets span single-type multi-hypergraphs (e.g., face-to-face contact, high school, hospital ward), multi-domain instances (gene–disease, arXiv co-authorship, Author–Citation), and legislative cosponsorship.

Performance is measured via F1-score, normalized mutual information (NMI), cosine similarity (CS), and area under curve (AUC) for link and edge prediction. MHSBM consistently outperforms single-layer baselines (Hy-MMSBM [Contisciani et al., 2022], Hypergraph-MT [De Bacco et al., 2022]) by 10–20% (CS), achieves AUC $v_{i}^{l}$7 for hyperedge prediction, and attains AUC in [0.77, 0.97] on cross-hypergraph edge prediction tasks. Removal of inter-hypergraph couplings leads to smooth degradation to the single-hypergraph baseline, evidencing the benefit of integrative modeling (Ni et al., 8 May 2025).

5. Extensions, Scalability, and Modeling Flexibility

MHSBM readily accommodates extensions: multi-view settings (where $v_{i}^{l}$8 are identical but $v_{i}^{l}$9 differ), multi-domain set-ups (heterogeneous node and edge types), dynamic or temporal multi-hypergraphs, and fully Bayesian versions with parameter priors. For more than two communities, unbalanced partitions, or non-uniform hyperedge sizes, analogous aggregate or tensor-based relaxations and inference mechanisms apply (Alaluusua et al., 2023, Ni et al., 8 May 2025).

When $u_{i}^{l} \in \Delta_{K^{l}}$0 (number of hypergraphs/layers) is fixed or grows slowly ($u_{i}^{l} \in \Delta_{K^{l}}$1), current theoretical results on exact recovery and SDP concentration continue to hold. Computationally, SDP solvers require $u_{i}^{l} \in \Delta_{K^{l}}$2 (interior-point), though practical methods exploit low-rank or first-order algorithms for $u_{i}^{l} \in \Delta_{K^{l}}$3 cost.

MHSBM’s hyperedge-internal-degree $u_{i}^{l} \in \Delta_{K^{l}}$4 parameter allows nuanced modeling of node importance and preferential attachment, increasing fidelity to empirical phenomena such as “star” nodes or organizers in social/event hypergraphs. The framework handles both assortative and disassortative structures: nonzero off-diagonal $u_{i}^{l} \in \Delta_{K^{l}}$5 or $u_{i}^{l} \in \Delta_{K^{l}}$6 influence whether communities prefer internal or cross-type connections.

6. Applications and Practical Implications

MHSBM enables integrative discovery of latent communities, robust prediction of missing hyperedges, and accurate inference of cross-graph edges in complex systems. Integrating multiple hypergraphs (via $u_{i}^{l} \in \Delta_{K^{l}}$7) substantially improves interpretability and predictive power over single-layer approaches, particularly in heterogeneous or multi-modal situations such as gene–protein, author–citation, or hybrid communication networks. In practice, MHSBM can serve as a foundation for high-order clustering, multi-relational link prediction, and structural analysis across domains (Alaluusua et al., 2023, Ni et al., 8 May 2025).

7. Summary Table: Key Structural Objects in MHSBM

Object	Definition/Role	Domain
$u_{i}^{l} \in \Delta_{K^{l}}$8	Mixed-membership vector for $u_{i}^{l} \in \Delta_{K^{l}}$9	$\mathbb{R}^{K^{l}}$0
$\mathbb{R}^{K^{l}}$1, $\mathbb{R}^{K^{l}}$2	Intra- and inter-hypergraph community affinities	$\mathbb{R}^{K^{l}}$3 matrices
$\mathbb{R}^{K^{l}}$4	Internal degree of node $\mathbb{R}^{K^{l}}$5 in hyperedge $\mathbb{R}^{K^{l}}$6 (preferential weight)	$\mathbb{R}^{K^{l}}$7
$\mathbb{R}^{K^{l}}$8	Hyperedge / cross-hypergraph edge random variables	$\mathbb{R}^{K^{l}}$9
$l$0	Observed / negative-sampled hyperedges for computational tractability	Hyperedge sets

The MHSBM unifies high-order community detection and link prediction with rigorous likelihood-based and spectral inference, scales to complex multi-relational settings, and is empirically validated for real-world, multigraph data sources (Alaluusua et al., 2023, Ni et al., 8 May 2025).