HGNet: Hypergraph Isomorphism Network

• HGNet is a hypergraph-level isomorphism network that leverages injective node–hyperedge updates to capture higher-order interactions.
• It represents both nodes and hyperedges with learnable embeddings and processes them using a two-stage message passing scheme.
• HWL-HIN achieves expressive power equivalent to the Hypergraph Weisfeiler-Lehman test, ensuring superior robustness predictions compared to standard HGNNs.

Searching arXiv for the target paper and closely related context papers. arXiv search query: (Tian et al., 26 Dec 2025) HWL-HIN Hypergraph Weisfeiler-Lehman Hypergraph Isomorphism Network HGNet, in the sense used by "HWL-HIN: A Hypergraph-Level Hypergraph Isomorphism Network as Powerful as the Hypergraph Weisfeiler-Lehman Test with Application to Higher-Order Network Robustness," denotes HWL-HIN: a hypergraph-level Hypergraph Isomorphism Network for predicting hypergraph robustness directly from hypergraph structure (Tian et al., 26 Dec 2025). The model represents both nodes and hyperedges with learnable embeddings, updates them through an injective two-stage node–hyperedge message-passing scheme, and produces a single embedding for the entire hypergraph through an injective readout over all nodes and hyperedges across layers. Its defining claim is theoretical as well as algorithmic: the architecture is constructed to be strictly equivalent in expressive power to the Hypergraph Weisfeiler-Lehman (H-WL) isomorphism test, and it is used as a surrogate for higher-order network robustness prediction (Tian et al., 26 Dec 2025).

1. Definition and problem setting

Within this paper, HGNet is not a generic label for hypergraph learning; it is the specific model HWL-HIN, explicitly introduced as a hypergraph-level architecture. “Hypergraph-level” means that the network outputs a single representation for an entire hypergraph, suitable for graph- or hypergraph-level regression tasks such as robustness prediction, rather than node classification (Tian et al., 26 Dec 2025).

The motivation is tied to the mismatch between conventional robustness surrogates and higher-order systems. CNN- and GNN-based predictors have been used for rapid robustness estimation, but the paper argues that these approaches neglect higher-order correlations that are naturally represented by hypergraphs. Standard hypergraph neural networks are described as insufficient for topological tasks because they commonly rely on non-injective pooling operators such as mean or max, which weaken their ability to distinguish subtle structural differences (Tian et al., 26 Dec 2025).

This positioning is consistent with the broader hypergraph-learning literature, where hypergraph neural networks are used to model higher-order interactions that cannot be faithfully reduced to ordinary pairwise edges (Kim et al., 2024). In that broader context, HWL-HIN is a specific design choice within the HNN family: one that prioritizes topological discriminability through injective aggregation rather than only scalable message passing.

2. Representation, input features, and message passing

HWL-HIN uses both node features and hyperedge features. Each hyperedge $e$ receives a 1D feature based on its cardinality $c_e$. Each node $v_i$ receives a 3D feature vector

$$x_i = [\tilde{k}_i, \tilde{c}_i, \tilde{o}_i]^\top,$$

where $\tilde{k}_i$ is the normalized hyperdegree, $\tilde{c}_i$ is the normalized local cardinality, and $\tilde{o}_i$ is a normalized feature encoding the node’s failure order / sequential vulnerability under the attack setting. All features are scaled to $[0,1]$ (Tian et al., 26 Dec 2025).

The core computation alternates between node-to-hyperedge and hyperedge-to-node aggregation. For layer $l$, hyperedges are updated first: $$h_e^{(l+1)} = \mathrm{MLP}_e^{(l)} \left( (1+\epsilon_e^{(l)})\,h_e^{(l)} + \sum_{v_i \in \mathcal{N}_v(e)} h_{v_i}^{(l)} \right),$$ where $c_e$0 is a learnable scalar self-term. Nodes are then updated from their incident hyperedges: $c_e$1 The design purpose is explicit: the model retains transformed self-information, sums incident representations, and applies an MLP so that the update is injective on bounded multisets (Tian et al., 26 Dec 2025).

The final hypergraph representation concatenates pooled summaries from all layers, using both node and hyperedge states: $c_e$2 The paper describes this as a hypergraph analogue of Jumping Knowledge or multi-layer readout. Because the node and hyperedge representations have already passed through MLPs, the per-layer sums remain injective, and concatenation across layers preserves injectivity while exposing patterns from local to global scales (Tian et al., 26 Dec 2025).

3. Injectivity and equivalence to the Hypergraph Weisfeiler-Lehman test

The central theoretical result is that HWL-HIN is as powerful as the Hypergraph Weisfeiler-Lehman test. The paper relies on the standard Deep Sets characterization

$c_e$3

emphasizing that injectivity over multisets requires a nonlinear map $c_e$4 before summation; raw summation alone is not sufficient because different multisets can collide (Tian et al., 26 Dec 2025).

The theorem stated in the paper considers iterative node and hyperedge updates of the form

$c_e$5

with injective $c_e$6, $c_e$7, $c_e$8, and $c_e$9, together with an injective hypergraph-level readout. Under these conditions, the model can distinguish any pair of hypergraphs that H-WL declares non-isomorphic (Tian et al., 26 Dec 2025).

The equivalence is bidirectional. If two hypergraphs are distinguished by H-WL, HWL-HIN can map them to different embeddings; conversely, if HWL-HIN distinguishes two hypergraphs, then H-WL can do so as well. This yields strict equivalence in expressive power. In the paper’s framing, this is the hypergraph analogue of the role played by GIN in ordinary graph learning, but it operates natively on node–hyperedge incidence structure rather than on a reduced graph representation (Tian et al., 26 Dec 2025).

4. Relation to standard HGNNs and to graph isomorphism networks

The paper defines HWL-HIN partly by contrast with standard HGNNs. Conventional hypergraph neural networks are described as typically using mean/max pooling, often omitting explicit hyperedge features, and generally not enforcing injectivity. As a result, they are weaker at distinguishing fine-grained hypergraph topology. HWL-HIN instead uses strictly injective aggregation, updates both node and hyperedge embeddings symmetrically, includes hyperedge-level information explicitly, and employs an injective hypergraph-level readout (Tian et al., 26 Dec 2025).

The comparison to GIN is precise rather than rhetorical. GIN is a graph model whose expressive power matches 1-WL on ordinary graphs; HWL-HIN is presented as the corresponding hypergraph construction. The analogy lies in the use of a sum + MLP design to realize injective multiset aggregation, but the domain of aggregation is different: GIN aggregates over node neighborhoods in a graph, whereas HWL-HIN alternates aggregation between nodes and hyperedges in a hypergraph incidence structure (Tian et al., 26 Dec 2025).

A common simplification is to convert a hypergraph into a bipartite graph and then apply an ordinary graph network. The paper explicitly rejects this as sufficient for the target setting, arguing that such conversion loses higher-order structure, whereas HWL-HIN operates natively on the hypergraph and preserves the full incidence-based topology (Tian et al., 26 Dec 2025). This suggests that its advantage is not merely architectural depth or feature engineering, but the preservation of higher-order relational structure under an injective update rule.

5. Robustness prediction in higher-order networks

The model’s application domain is connectivity robustness prediction for hypergraphs under node-removal attacks. The target scalar is the robustness score

$v_i$0

where $v_i$1 is the largest connected component fraction after removing a fraction $v_i$2 of nodes. The paper also notes the conventional discrete form

$v_i$3

but uses the integral form to obtain more accurate labels (Tian et al., 26 Dec 2025).

The hypergraph semantics are direct: vertices represent system components and hyperedges represent multi-way relations or group interactions. Two attack regimes are considered. The first is static targeted attacks, where failure order is predetermined by structural metrics such as hyperdegree. The second is dynamic cascading attacks, where failure propagates through a higher-order load redistribution process (Tian et al., 26 Dec 2025).

For the dynamic regime, the paper extends a Motter–Lai-style load model to hypergraphs: $v_i$4 When node $v_i$5 fails, its load is redistributed through incident hyperedges and then among surviving nodes in those hyperedges. Additional failures occur if node load exceeds capacity, and hyperedges are also considered failed when too many nodes inside them fail. These simulations generate the robustness labels used for training (Tian et al., 26 Dec 2025).

Ground-truth labels are computed by adaptive integration using Adaptive Simpson’s method, with tolerance chosen to be much tighter than the surrogate model’s expected error. The stated purpose is to reduce simulation cost while keeping label noise small (Tian et al., 26 Dec 2025).

6. Empirical behavior, optimization, and significance

Training uses AdamW with cosine annealing learning-rate scheduling. The empirical study is conducted on synthetic hypergraph datasets: ER, WS, SF, SBM, UF, and a mixed dataset (Tian et al., 26 Dec 2025).

The reported results are structurally consistent across experiments. HWL-HIN achieves the best overall prediction accuracy across most settings; it substantially outperforms standard HGNNs, especially in static scenarios where topological expressiveness is most consequential; and it also outperforms graph-based baselines after hypergraph-to-bipartite conversion, reinforcing the claim that preserving higher-order structure matters (Tian et al., 26 Dec 2025). In dynamic cascading settings, however, the performance gap between HWL-HIN and simpler HGNNs narrows because the failure-sequence feature becomes highly informative. This is an important caveat: the model’s structural advantage is most visible when topology, rather than exogenous sequence information, dominates the task.

The ablation study supports the same interpretation. Removing engineered features hurts standard HGNNs much more than HWL-HIN; HWL-HIN remains substantially stronger when learning from adjacency or topology alone; and removing the cosine annealing scheduler hurts convergence for both the proposed model and the baselines (Tian et al., 26 Dec 2025). The paper therefore presents the model as not merely feature-driven, but as having stronger structural learning capacity.

Efficiency claims are also central. HWL-HIN and other graph or hypergraph neural surrogates have comparable inference latency; standard HGNNs are slightly faster because they use fewer parameters; matrix-based methods such as SPP-CNN and ATTRP are much slower and can encounter GPU OOM as resolution increases; and all machine-learning surrogates are hundreds of times faster than the adaptive integration procedure used to generate labels (Tian et al., 26 Dec 2025). In this sense, HWL-HIN is positioned as a model that couples a maximal topological expressiveness guarantee, at the H-WL level, with practical training and inference efficiency.

A recurring ambiguity in the literature is that “HGNet” can denote a broad hypergraph-neural family rather than a single canonical architecture (Kim et al., 2024). In this paper, by contrast, the term identifies a specific construction: HWL-HIN, a native hypergraph-level isomorphism network whose distinguishing features are injective node-to-hyperedge and hyperedge-to-node updates, explicit hyperedge features, and an injective multiscale readout, all directed toward robustness prediction in higher-order networks (Tian et al., 26 Dec 2025).