Global Similarity Hypergraph Overview

• Global similarity hypergraphs are higher-order models that encode multi-way affinities beyond traditional pairwise relations, enabling robust network analysis and clustering.
• They integrate methods like spectral embedding, generalized kernel k-means, and information-theoretic frameworks to capture and compare complex hypergraph structures.
• Implementable metrics such as Hyper NetSimile and Hyperedge Portrait Divergence provide concise structural signatures that reveal nuanced similarities across diverse datasets.

Global similarity hypergraphs refer both to a family of higher-order models that encode multi-way affinities in data (functioning as higher-order analogues of similarity graphs), and to a class of global similarity and dissimilarity measures designed to compare hypergraphs at all structural levels. This conceptual landscape spans the spectral embedding approach to hypergraph clustering, information-theoretic frameworks for quantifying hypergraph overlap, and practical structural similarity metrics tailored to capture the nuanced properties of higher-order networks (Saito, 2022, &&&1&&&, Agostinelli et al., 21 Mar 2025). The unifying theme is the move beyond pairwise relations to systematically encode and compare the rich combinatorics of multi-node interactions.

1. Multi-way Similarity and Hypergraph Construction

Global similarity hypergraphs are constructed by modeling data using multi-way, rather than just pairwise, similarities. Let $X = \{x_1, \ldots, x_n\} \subset \mathbb{R}^d$ be data points. Given an even integer $m \geq 2$ and a positive-definite kernel $$\kappa: \mathbb{R}^d \times \mathbb{R}^d \to \mathbb{R}$$ with feature map $\psi$, one can define for every $m$-tuple a hyperedge with weight

$$S(i_1,\ldots,i_m) = \sum_{\gamma=1}^{m/2} \sum_{\nu=1}^{m/2} \kappa(x_{i_\gamma}, x_{i_{m/2+\nu}}).$$

This construction induces an $m$-uniform weighted hypergraph $G=(V, E, w)$ with $V=\{1,\ldots,n\}$, $E$ the set of $m$-tuples, and $w(e)=S(i_1,\ldots,i_m)$. Such structures systematically encode global affinity by aggregating all pairwise kernel similarities between two halves of each hyperedge, generalizing the similarity graph paradigm to higher orders (Saito, 2022).

2. Spectral Cut, Laplacian, and Kernel $k$-Means Connections

The global similarity hypergraph admits a star-reduction adjacency $A_s$ and a Laplacian $L_s = D_V - A_s$, where $H\in\mathbb{R}^{n \times |E|}$ is the incidence matrix, $W_e$ the edge-weight matrix, and $D_V$ the diagonal vertex degree matrix. Spectral clustering seeks clusters $\{V_j\}$ minimizing the $k$-way normalized cut

$$\mathrm{kNCut}(\{V_j\}) = \sum_{j=1}^k \frac{\mathrm{Cut}(V_j, V \setminus V_j)}{\mathrm{vol}(V_j)},$$

which admits a relaxation to an eigenproblem for $D_V^{-1/2} L_s D_V^{-1/2}$. This approach is equivalent to a generalized weighted kernel $k$-means, using a contracted biclique-Gram matrix

$$K^{(m)}_{i,j} = n^{m-2} \langle \psi_i + \frac{m-2}{2} \Psi, \psi_j + \frac{m-2}{2} \Psi \rangle, \quad \Psi = \frac{1}{n} \sum_{\ell} \psi_\ell,$$

and gives a one-to-one correspondence between hypergraph spectral clustering and kernel methods (Saito, 2022).

This equivalence provides a principled route from multi-way similarity to practical clustering tools, with the entire pipeline scaling as $O(n^3)$ (for eigen-decomposition), similar to standard spectral clustering on graphs.

3. Information-Theoretic Frameworks for Hypergraph Similarity

Rather than encoding similarity via multi-way weights alone, information-theoretic approaches explicitly quantify global similarity between (potentially heterogeneous) hypergraphs. Let $H_1, H_2$ be hypergraphs on a fixed set $V$. The similarity is formulated via a coding protocol that computes mutual information

$\mathrm{MI}_c(H_1; H_2) = H_c(H_2) - H_c(H_2|H_1)$

for an encoding $c$, with $H_c(\cdot)$ the entropy (description length) and $H_c(\cdot|\cdot)$ the conditional entropy under $c$. The normalized mutual information (NMI) is

$$\mathrm{NMI}_c(H_1, H_2) = 1 - \min\left\{ \frac{H_c(H_2|H_1)}{H_c(H_2)}, \frac{H_c(H_1|H_2)}{H_c(H_1)} \right\},$$

with $0 \leq \mathrm{NMI}_c \leq 1$ (Felippe et al., 31 Oct 2025).

Encoding schemes include:

• Bulk encoding: treats all hyperedges as a set, measuring overall edge overlap.
• Align encoding: computes NMI per hyperedge order (layerwise).
• Cross encoding: allows encoding lower-order edges in one hypergraph using the projections of higher-order edges in the other, capturing order-nested similarities.

Coarse-grained (mesoscale) similarity is obtained by mapping nodes into super-nodes by community or group, replacing edges with their projected multisets.

4. Structural and Statistical Metrics for Hypergraph Comparison

Complementing the information-theoretic perspective, recent advances provide implementable metrics:

• Hyper NetSimile (HNS): Each hypergraph is summarized by a 45-dimensional signature vector of nine structural node features (degree, hyperdegree, hyper-clustering, incident edge-size statistics, neighbor aggregates, 2-hop ego size) with five summary statistics apiece. The normalized Canberra distance between these vectors becomes the dissimilarity $d(H_1, H_2)$, with similarity $s(H_1, H_2)=1-d(H_1,H_2)$ (Agostinelli et al., 21 Mar 2025).
• Hyperedge Portrait Divergence (HPD): The "hyperedge portrait" $\Gamma_{m,n,l,k}$ records for each hyperedge size $m$, the count of hyperedges of size $n$ at path distance $l$ with $k$ such neighbors, normalized to a probability tensor $P$. The Jensen–Shannon divergence $\mathrm{JS}(P_1, P_2)$ measures global structure, again yielding similarity by $1-\mathrm{JS}$.

Both methods are size-invariant, relabeling invariant, and sensitive to higher-order structural nuances. HNS is computationally lighter; HPD requires all-pairs shortest paths on the hyperedge adjacency and scales as $O(E^2)$.

5. Algorithmic Steps and Computational Complexity

Spectral Embedding Pipeline for Clustering (global similarity hypergraph):

1. Compute $n \times n$ Gram matrix $K_0$ via the base kernel.
2. Compute contracted biclique-Gram $K^{(m)}$ via a closed-form update.
3. Build vertex degrees $D_V$ and normalized adjacency $M = D_V^{-1/2} K^{(m)} D_V^{-1/2}$.
4. Perform eigen-decomposition of $M$, selecting top $k$ eigenvectors.
5. (Optional) Row-normalize and run $k$-means in the reduced space.

Total complexity is cubic in $n$ for dense inputs (Saito, 2022).

NMI-based Cross-Order Similarity Computation:

For each order pair $(k, \ell)$, the projection overlap $E_1^{(k \rightarrow \ell)}$ and $E_{1 \rightarrow 2}^{(k \rightarrow \ell)}$ are computed recursively using maps and hashing; overall complexity is $O(E^2 L^2)$ for $E$ edges and maximum order $L$ (Felippe et al., 31 Oct 2025).

HNS and HPD Computation:

• HNS: Main bottleneck is the hyper-clustering coefficient per node; total cost scales as $O(NK^2M^2)$ where $K$ is average hyperdegree, $M$ is max edge size.
• HPD: Main cost is all-pairs shortest paths among $E$ hyperedges, $O(E^2)$; for large $E$, sampling or $\ell$-truncation (max path length) yields approximations (Agostinelli et al., 21 Mar 2025).

6. Empirical Validation and Use Cases

Global similarity hypergraph models and metrics have been validated on synthetic generative models (Erdős–Rényi, configuration, Watts–Strogatz) and diverse empirical datasets:

• Information-theoretic NMI distinguishes block-nested and fully random hypergraphs, detects multiplex cross-order similarity, and tracks mesoscale (community) structure under coarse-graining (Felippe et al., 31 Oct 2025).
• HNS and HPD accurately cluster both generative and real networks (face-to-face proximity, co-authorship, online community, legislative committee networks), outperforming pairwise methods and revealing data-type-driven clustering (Agostinelli et al., 21 Mar 2025).
• HPD is uniquely sensitive to changes in maximum hyperedge size and null-model reshufflings, confirming its global structure sensitivity.

A plausible implication is that genuine higher-order patterns—in collaborations, social gatherings, biological complexes, etc.—are not well-captured by pairwise-only metrics, and necessitate global similarity hypergraph tools for robust detection and analysis.

7. Limitations and Practical Considerations

• For large-scale hypergraphs ($E \gg 10^4$), HPD and information-theoretic NMI become computationally demanding; sampling or truncation yields scalable approximations.
• HNS is sensitive to feature selection and may miss certain structural motifs; feature augmentation (e.g., with centralities or core indices) may be needed for targeted applications.
• Existing global similarity measures are invariant to node labeling and ignore explicit node alignments; alignment-sensitive tasks require graph/hypergraph matching frameworks.
• When comparing hypergraphs with non-overlapping hyperedge size support, HPD and NMI measures may report maximal dissimilarity; preprocessing or layered restriction may be necessary (Agostinelli et al., 21 Mar 2025, Felippe et al., 31 Oct 2025).

Table: Summary of Major Global Hypergraph Similarity Methods

Method/Class	Core Principle	Computational Cost
Spectral Biclique Hypergraph	Multi-way kernel; spectral cut; $k$-means	$O(n^3)$
NMI (Information-Theoretic)	Coding overlap; intra/cross-order; mesoscale	$O(E^2 L^2)$
HNS	Feature vector (node stats); Canberra distance	$O(E^2 M^2/N)$
HPD	Hyperedge-path tensor; Jensen–Shannon div.	$O(E^2)$

Global similarity hypergraph frameworks are foundational to higher-order data mining, clustering, and network comparison, enabling robust, scalable, and order-sensitive analyses that transcend the limitations of pairwise models.

Key references: (Saito, 2022, Felippe et al., 31 Oct 2025, Agostinelli et al., 21 Mar 2025)