Hypergraph Isomorphism Networks

• The paper demonstrates that HWL-HIN achieves expressivity equivalent to the Hypergraph Weisfeiler–Lehman test via injective node–hyperedge updates and sum-aggregation principles.
• It employs an alternating message passing between node and hyperedge embeddings with a jumping-knowledge style readout, enabling robust hypergraph-level regression and prediction.
• The categorical generalization lifts hypergraphs using functorial mappings to graded posets, enhancing expressivity beyond standard hypergraph neural networks.

Searching arXiv for the specified papers and closely related work on Hypergraph Isomorphism Networks. arxiv_search(query="Hypergraph Isomorphism Networks HWL-HIN (Tian et al., 26 Dec 2025) categorical Weisfeiler Lehman hypergraphs (Choi et al., 6 Feb 2026)", max_results=10) Hypergraph Isomorphism Networks (HINs) are hypergraph neural architectures designed so that their message passing and readout are aligned with Weisfeiler–Lehman-style distinguishability on higher-order structures. In the recent literature, the term primarily refers to two closely related but distinct formulations. The first is HWL-HIN, a hypergraph-level Hypergraph Isomorphism Network for hypergraph-level prediction that alternates injective node–hyperedge updates and is proven to have expressive power strictly equivalent to the Hypergraph Weisfeiler–Lehman (HWL) test under the paper’s assumptions (Tian et al., 26 Dec 2025). The second is a broader categorical formulation in which hypergraphs are lifted by a functor to graded posets and then processed by Weisfeiler–Lehman-style message passing; the resulting neural models are called HINs, with concrete instances induced by an incidence functor and a symmetric simplicial complex functor (Choi et al., 6 Feb 2026). A separate line of work uses the label “isomorphism network” for heterogeneous graphs rather than strict hypergraphs; that usage is terminologically related but structurally different (Sang et al., 11 Mar 2025).

1. Definition, object of study, and scope

In the strict hypergraph setting, the underlying object is a hypergraph

$\mathcal{H}=(V,E),$

where $V=\{v_1,\dots,v_{|V|}\}$ is the set of vertices and $E=\{e_1,\dots,e_{|E|}\}$ is the set of hyperedges, with each hyperedge $e\subseteq V$ potentially containing multiple vertices (Tian et al., 26 Dec 2025). The incidence structure is represented by

$$H \in \{0,1\}^{|V|\times |E|}, \quad H(v,e)=1 \iff v\in e.$$

Two neighborhood systems are central to hypergraph message passing: $$\mathcal{N}_e(v)=\{e\in E\mid H(v,e)=1\}, \quad d_v = |\mathcal{N}_e(v)|,$$ and

$$\mathcal{N}_v(e)=\{v\in V\mid H(v,e)=1\}, \quad c_e = |\mathcal{N}_v(e)|.$$

These encode, respectively, the incident hyperedges of a vertex and the vertices contained in a hyperedge (Tian et al., 26 Dec 2025).

The immediate motivation for HINs is that conventional Hypergraph Neural Networks typically follow a two-stage alternating aggregation between vertices and hyperedges, but usually rely on non-injective aggregators such as mean or max. The argument advanced by recent work is that such aggregation cannot attain HWL-level expressivity, whereas HINs are constructed explicitly around injective multiset processing, injective updates, and injective hypergraph-level readout (Tian et al., 26 Dec 2025). In the categorical formulation, this design principle is generalized: one first lifts a hypergraph to a graded poset by a functor and then applies graded Weisfeiler–Lehman refinement or its neural relaxation, so that the message passing topology is determined by the lifting itself (Choi et al., 6 Feb 2026).

2. Core architecture of HWL-HIN

HWL-HIN is a hypergraph neural network for hypergraph-level regression/prediction, with robustness prediction as its primary application (Tian et al., 26 Dec 2025). Its architecture maintains both node embeddings and hyperedge embeddings, and alternates updates in two directions:

1. hyperedges are updated from incident nodes;
2. nodes are updated from incident hyperedges.

The update rules are

$$h_e^{(l+1)} = \text{MLP}_e^{(l)}\left((1+\epsilon_e^{(l)}) \cdot h_e^{(l)} + \sum_{v_i \in \mathcal{N}_v(e)} h_{v_i}^{(l)}\right),$$

and

$$h_v^{(l+1)} = \text{MLP}_v^{(l)}\left((1+\epsilon_v^{(l)}) \cdot h_v^{(l)} + \sum_{e_i \in \mathcal{N}_e(v)} h_{e_i}^{(l+1)}\right),$$

where $\epsilon_e^{(l)}$ and $V=\{v_1,\dots,v_{|V|}\}$0 are learnable scalars (Tian et al., 26 Dec 2025). The use of sum followed by an MLP is explicitly motivated by the Deep Sets / GIN-style “map-then-sum” principle, with injectivity over multisets as the key theoretical requirement.

For hypergraph-level prediction, HWL-HIN uses a layer-wise readout over both nodes and hyperedges: $V=\{v_1,\dots,v_{|V|}\}$1 This is described as a jumping-knowledge style readout. The stated purpose is to preserve information from shallow and deep propagation regimes simultaneously: earlier layers encode more local patterns, while deeper layers capture more global structure (Tian et al., 26 Dec 2025).

The model does not rely on raw topology alone. Each hyperedge receives a 1D feature given by its cardinality $V=\{v_1,\dots,v_{|V|}\}$2. Each node receives a 3D feature

$V=\{v_1,\dots,v_{|V|}\}$3

where $V=\{v_1,\dots,v_{|V|}\}$4 is normalized hyperdegree, $V=\{v_1,\dots,v_{|V|}\}$5 is normalized local cardinality, and $V=\{v_1,\dots,v_{|V|}\}$6 is normalized failure order or sequential vulnerability prior. All features are normalized to $V=\{v_1,\dots,v_{|V|}\}$7 (Tian et al., 26 Dec 2025). The paper states that these features are used to speed convergence and inject prior knowledge.

3. Expressivity and the Hypergraph Weisfeiler–Lehman test

The central theoretical claim of HWL-HIN is that it is as expressive as the Hypergraph Weisfeiler–Lehman test under explicit injectivity assumptions (Tian et al., 26 Dec 2025). The theoretical bridge is the injective multiset representation result

$V=\{v_1,\dots,v_{|V|}\}$8

which underpins the use of sum aggregation with sufficiently expressive learned maps.

The paper first states an upper-bound result: if an HGNN maps two non-isomorphic hypergraphs to different embeddings, then the HWL test must also decide them as non-isomorphic. It then gives a converse-style result for networks of the form

$V=\{v_1,\dots,v_{|V|}\}$9

provided that $E=\{e_1,\dots,e_{|E|}\}$0, $E=\{e_1,\dots,e_{|E|}\}$1, $E=\{e_1,\dots,e_{|E|}\}$2, and $E=\{e_1,\dots,e_{|E|}\}$3 are all injective and the hypergraph-level readout over node and hyperedge features is injective (Tian et al., 26 Dec 2025). Under sufficient depth, such a network can distinguish any pair of hypergraphs that HWL distinguishes.

The HWL refinement itself is represented in the paper as

$E=\{e_1,\dots,e_{|E|}\}$4

and the proof proceeds by induction to show the existence of injective maps $E=\{e_1,\dots,e_{|E|}\}$5 such that

$E=\{e_1,\dots,e_{|E|}\}$6

This establishes the claimed equivalence between the learned embeddings and the HWL color refinement, modulo the paper’s assumptions (Tian et al., 26 Dec 2025).

The assumptions are explicit. They include a countable input space, injective multiset aggregation, injective update functions, an injective hypergraph readout, and sufficient depth. The paper also notes that one-hot initialization can make the first mapping simpler, and that an MLP can be optional in the first layer (Tian et al., 26 Dec 2025). A plausible implication is that the claimed equivalence is architectural rather than automatic: it depends on the model class and not merely on training a generic hypergraph GNN.

4. Categorical generalization: CatWL, CatMPN, and functor-specific HINs

A broader formalization is given by the categorical Weisfeiler–Lehman framework, which treats lifting as a functorial mapping from a source data category to the category of graded posets (Choi et al., 6 Feb 2026). In this formulation, the target domain is a graded poset $E=\{e_1,\dots,e_{|E|}\}$7 equipped with a dimension function $E=\{e_1,\dots,e_{|E|}\}$8, and hypergraph isomorphism networks arise as neural relaxations of graded-poset WL refinement after lifting.

The Categorical WL test lifts an object $E=\{e_1,\dots,e_{|E|}\}$9 to a graded poset $e\subseteq V$0 and then applies graded Weisfeiler–Lehman refinement. The color update for an element $e\subseteq V$1 in a graded poset is

$e\subseteq V$2

where boundary, coboundary, lower adjacency, and upper adjacency are all included (Choi et al., 6 Feb 2026). The corresponding neural relaxation is called a Categorical Message Passing Network (CatMPN), and when the source category is the category of hypergraphs, the resulting models are called Hypergraph Isomorphism Networks.

The paper introduces two functors from hypergraphs to graded posets.

The first is the incidence poset functor

$e\subseteq V$3

which maps a hypergraph $e\subseteq V$4 to the incidence poset $e\subseteq V$5, where $e\subseteq V$6 iff $e\subseteq V$7. This yields a message-passing topology that resembles a bipartite scheme, but the paper emphasizes that the full poset refinement also introduces peer interactions through lower and upper adjacencies (Choi et al., 6 Feb 2026).

The second is the symmetric simplicial complex functor $e\subseteq V$8, which expands each hyperedge into a full face lattice. It defines

$e\subseteq V$9

and for $$H \in \{0,1\}^{|V|\times |E|}, \quad H(v,e)=1 \iff v\in e.$$0,

$$H \in \{0,1\}^{|V|\times |E|}, \quad H(v,e)=1 \iff v\in e.$$1

This produces a hierarchy across many dimensions, allowing message passing over nodes, $$H \in \{0,1\}^{|V|\times |E|}, \quad H(v,e)=1 \iff v\in e.$$2-faces, $$H \in \{0,1\}^{|V|\times |E|}, \quad H(v,e)=1 \iff v\in e.$$3-faces, and higher-dimensional subrelations within each hyperedge (Choi et al., 6 Feb 2026). The intended advantage is preservation of internal subset geometry and intersection structure that a pure incidence representation may flatten.

The appendix instantiates the HIN update by computing message functions over each adjacency type with sums, MLPs, concatenation, and learnable self-preservation coefficients $$H \in \{0,1\}^{|V|\times |E|}, \quad H(v,e)=1 \iff v\in e.$$4 (Choi et al., 6 Feb 2026). The paper states that the message functions and aggregator are learnable, and proves the familiar upper-bound-and-attainability statement: $$H \in \{0,1\}^{|V|\times |E|}, \quad H(v,e)=1 \iff v\in e.$$5-CatWL is at least as powerful as $$H \in \{0,1\}^{|V|\times |E|}, \quad H(v,e)=1 \iff v\in e.$$6-CatMPN, and with injective update and aggregation functions and sufficient depth and width, $$H \in \{0,1\}^{|V|\times |E|}, \quad H(v,e)=1 \iff v\in e.$$7-CatMPN has the same expressive power as $$H \in \{0,1\}^{|V|\times |E|}, \quad H(v,e)=1 \iff v\in e.$$8-CatWL (Choi et al., 6 Feb 2026).

5. Expressivity beyond standard HWL and practical consequences of the lifting choice

The categorical paper makes a stronger relative claim than mere equivalence to standard HWL: both the incidence-based and symmetric simplicial approaches are proven to be not less powerful than HWL (Choi et al., 6 Feb 2026). Specifically, it states: $$H \in \{0,1\}^{|V|\times |E|}, \quad H(v,e)=1 \iff v\in e.$$9 and

$$\mathcal{N}_e(v)=\{e\in E\mid H(v,e)=1\}, \quad d_v = |\mathcal{N}_e(v)|,$$0

These results are established by exhibiting hypergraph pairs that are indistinguishable by standard HWL but distinguishable by both $$\mathcal{N}_e(v)=\{e\in E\mid H(v,e)=1\}, \quad d_v = |\mathcal{N}_e(v)|,$$1-CatWL and $$\mathcal{N}_e(v)=\{e\in E\mid H(v,e)=1\}, \quad d_v = |\mathcal{N}_e(v)|,$$2-CatWL (Choi et al., 6 Feb 2026). This suggests that, in that framework, HINs induced by these functors subsume the expressive power of the standard hypergraph WL benchmark rather than merely reproducing it.

A further technical contribution concerns redundancy among adjacency channels. Under the condition

$$\mathcal{N}_e(v)=\{e\in E\mid H(v,e)=1\}, \quad d_v = |\mathcal{N}_e(v)|,$$3

the paper proves that the full four-adjacency refinement has the same expressive power as a restricted version using only boundary and upper adjacencies, omitting coboundary and lower adjacencies (Choi et al., 6 Feb 2026). This result is used to justify efficient experimental variants such as $$\mathcal{N}_e(v)=\{e\in E\mid H(v,e)=1\}, \quad d_v = |\mathcal{N}_e(v)|,$$4, even though the full formalism includes four message types.

The lifting choice is therefore not merely a representational convenience. The paper’s own interpretation is that the incidence functor is best thought of as a bipartite hypergraph lift suited to global hyperedge connectivity and dense incidence patterns, whereas the symmetric simplicial functor preserves fine internal subrelations within hyperedges (Choi et al., 6 Feb 2026). This is not presented as a universal ranking. Instead, the theory and experiments are organized around the claim that no single lifting is uniformly best.

6. Empirical behavior, applications, and terminological boundaries

The application emphasized by HWL-HIN is robustness prediction for higher-order networks under static attacks and dynamic cascading attacks (Tian et al., 26 Dec 2025). The target robustness label is defined through the percolation curve $$\mathcal{N}_e(v)=\{e\in E\mid H(v,e)=1\}, \quad d_v = |\mathcal{N}_e(v)|,$$5: $$\mathcal{N}_e(v)=\{e\in E\mid H(v,e)=1\}, \quad d_v = |\mathcal{N}_e(v)|,$$6 For dynamic cascading attacks, the paper extends a load redistribution model to hypergraphs, using

$$\mathcal{N}_e(v)=\{e\in E\mid H(v,e)=1\}, \quad d_v = |\mathcal{N}_e(v)|,$$7

together with hyperedge-mediated redistribution rules (Tian et al., 26 Dec 2025). Because full simulation is expensive, labels are generated with adaptive Simpson integration using a strict tolerance, specifically $$\mathcal{N}_e(v)=\{e\in E\mid H(v,e)=1\}, \quad d_v = |\mathcal{N}_e(v)|,$$8 and maximum recursion depth $$\mathcal{N}_e(v)=\{e\in E\mid H(v,e)=1\}, \quad d_v = |\mathcal{N}_e(v)|,$$9 (Tian et al., 26 Dec 2025).

The experiments use synthetic hypergraphs with $$\mathcal{N}_v(e)=\{v\in V\mid H(v,e)=1\}, \quad c_e = |\mathcal{N}_v(e)|.$$0 nodes. Five homogeneous datasets are considered—ER, WS, SF, SBM, and UF—plus a Mix dataset formed from 500 samples of each generator. Each homogeneous dataset contains 1200 samples with 1000 train and 200 test, while Mix contains 2500 training samples and 200 test samples (Tian et al., 26 Dec 2025). Baselines include HGNNs, GIN-MAS, NRL-GT, SPP-CNN, ATTRP, KNN, and DT. The paper reports mean error $$\mathcal{N}_v(e)=\{v\in V\mid H(v,e)=1\}, \quad c_e = |\mathcal{N}_v(e)|.$$1 standard deviation, average rank in parentheses, paired $$\mathcal{N}_v(e)=\{v\in V\mid H(v,e)=1\}, \quad c_e = |\mathcal{N}_v(e)|.$$2-tests at significance level $$\mathcal{N}_v(e)=\{v\in V\mid H(v,e)=1\}, \quad c_e = |\mathcal{N}_v(e)|.$$3, and computational efficiency. Its summary finding is that HWL-HIN is the best or near-best method across datasets and attack types, that it outperforms standard HGNNs especially in static attack settings, and that graph-based baselines degrade in dynamic cases because bipartite conversion loses higher-order structure (Tian et al., 26 Dec 2025). It also reports that HWL-HIN and HGNNs have comparable inference time, while matrix-based methods are much slower and surrogate ML methods are hundreds of times faster than adaptive integration label generation (Tian et al., 26 Dec 2025).

The categorical HIN paper evaluates hypergraph classification on six real-world benchmarks: IMDB_dir_form, IMDB_dir_genre, IMDB_wri_form, IMDB_wri_genre, steam_player, and twitter_friend (Choi et al., 6 Feb 2026). Using 10-fold cross-validation and reporting best mean validation accuracy and standard deviation, it compares against MLP, HCHA, HNHN, HyperGCN, UniGCNII, ED-HNN, AllSetTransformer, HIC, and 2-FHNN. The reported summary is that CatMPN variants achieve the best performance on five of six benchmarks. The paper highlights that on IMDB_dir_genre, $$\mathcal{N}_v(e)=\{v\in V\mid H(v,e)=1\}, \quad c_e = |\mathcal{N}_v(e)|.$$4-HIN exceeds HIC by 1.92%; on twitter_friend, the gain over HIC is 6.49%; and on IMDB_wri_genre, $$\mathcal{N}_v(e)=\{v\in V\mid H(v,e)=1\}, \quad c_e = |\mathcal{N}_v(e)|.$$5-HIN attains 61.02%, outperforming HNHN by 6.85% (Choi et al., 6 Feb 2026). The exception is steam_player, where 2-FHNN remains best; the paper interprets this as a case in which highly uniform hyperedge cardinality reduces the benefit of the extra structure in HINs (Choi et al., 6 Feb 2026).

The practical limitations are explicit. For the incidence lift $$\mathcal{N}_v(e)=\{v\in V\mid H(v,e)=1\}, \quad c_e = |\mathcal{N}_v(e)|.$$6, construction cost is roughly

$$\mathcal{N}_v(e)=\{v\in V\mid H(v,e)=1\}, \quad c_e = |\mathcal{N}_v(e)|.$$7

For the symmetric simplicial lift $$\mathcal{N}_v(e)=\{v\in V\mid H(v,e)=1\}, \quad c_e = |\mathcal{N}_v(e)|.$$8, generating all simplices is combinatorial; for $$\mathcal{N}_v(e)=\{v\in V\mid H(v,e)=1\}, \quad c_e = |\mathcal{N}_v(e)|.$$9-simplices the cost scales like

$$h_e^{(l+1)} = \text{MLP}_e^{(l)}\left((1+\epsilon_e^{(l)}) \cdot h_e^{(l)} + \sum_{v_i \in \mathcal{N}_v(e)} h_{v_i}^{(l)}\right),$$0

To control this, the paper uses a threshold $$h_e^{(l+1)} = \text{MLP}_e^{(l)}\left((1+\epsilon_e^{(l)}) \cdot h_e^{(l)} + \sum_{v_i \in \mathcal{N}_v(e)} h_{v_i}^{(l)}\right),$$1 and excludes hyperedges with cardinality $$h_e^{(l+1)} = \text{MLP}_e^{(l)}\left((1+\epsilon_e^{(l)}) \cdot h_e^{(l)} + \sum_{v_i \in \mathcal{N}_v(e)} h_{v_i}^{(l)}\right),$$2 from lifting for $$h_e^{(l+1)} = \text{MLP}_e^{(l)}\left((1+\epsilon_e^{(l)}) \cdot h_e^{(l)} + \sum_{v_i \in \mathcal{N}_v(e)} h_{v_i}^{(l)}\right),$$3-HIN (Choi et al., 6 Feb 2026). This is an implementation-level restriction rather than a change in the underlying theory.

A recurrent source of confusion is the overlap between hypergraph isomorphism networks and the similarly named heterogeneous graph isomorphism network in HeGMN (Sang et al., 11 Mar 2025). The latter is not about hypergraphs in the strict sense. It models heterogeneous graphs with node types $$h_e^{(l+1)} = \text{MLP}_e^{(l)}\left((1+\epsilon_e^{(l)}) \cdot h_e^{(l)} + \sum_{v_i \in \mathcal{N}_v(e)} h_{v_i}^{(l)}\right),$$4 and edge types $$h_e^{(l+1)} = \text{MLP}_e^{(l)}\left((1+\epsilon_e^{(l)}) \cdot h_e^{(l)} + \sum_{v_i \in \mathcal{N}_v(e)} h_{v_i}^{(l)}\right),$$5, but every edge still connects two nodes, and its encoder is a relation-aware variant of GIN rather than a hypergraph model over hyperedges of arbitrary cardinality (Sang et al., 11 Mar 2025). That paper explicitly states that its model is best understood as a type- and relation-aware GIN variant. The distinction matters because HWL-HIN and categorical HINs are defined around higher-order incidence structure, injective hyperedge-aware updates, and hypergraph-level expressive power, whereas HGIN in HeGMN is a heterogeneous-graph encoder for graph similarity learning rather than a strict hypergraph isomorphism network (Sang et al., 11 Mar 2025).