Line Hypergraph Convolution Networks

• Line Hypergraph Convolution Networks are advanced neural architectures that transform hypergraphs into line-expanded forms to model complex higher-order relationships efficiently.
• They employ projection matrices and specialized convolutional operators to propagate features between vertices, hyperedges, and line-graph domains, unifying classical expansion methods.
• Empirical studies demonstrate that LHCNs improve accuracy in node and hyperedge classification as well as recommendation tasks, while maintaining scalability on large datasets.

Line Hypergraph Convolution Networks (LHCNs) represent a class of neural network architectures that enable the extension of graph convolutional methods to hypergraphs via a transformation into a line-expanded or line-graph domain. This approach unifies various hypergraph learning paradigms and supports principled, scalable, and expressively powerful learning on higher-order relational data, with applicability to node classification, hyperedge classification, variable-sized expansion problems, and recommendation systems.

1. Mathematical Foundations of Line Expansion and Hypergraph Line Graphs

Let $\mathcal{H} = (V, E)$ be a hypergraph, with vertex set $V$ and hyperedge set $E$. The classic incidence matrix $H \in \{0,1\}^{|V| \times |E|}$ encodes the hypergraph's structure, where $H(v, e) = 1$ if $v \in e$. The degree matrices $D_v$ and $D_e$ are diagonal with $D_v(v, v) = d(v)$ (vertex degree) and $D_e(e, e) = \delta(e)$ (hyperedge degree), $V$0, $V$1 (Yang et al., 2020).

Several notions of line expansion have been proposed:

• LE-GCN and ADHCN Line Expansion: The line-expanded graph $V$2 consists of one node $V$3 per $V$4 with $V$5. Nodes $V$6 in $V$7 are adjacent if $V$8 or $V$9, yielding the adjacency $E$0 if $E$1 or $E$2, 0 otherwise. This construction is bijective: from $E$3 one can recover the original hypergraph by inverting the line graph of the star expansion using Whitney's theorem (Yang et al., 2020, Liu et al., 2022).
• Classical Hypergraph Line Graph (as in LHCN, DHCN, HeIHNN, etc.): The line graph $E$4 has $E$5 as nodes. Two hyperedges $E$6 are adjacent if $E$7. Edge weights are typically $E$8 or $E$9 (Bandyopadhyay et al., 2020, Xia et al., 2020, Ye et al., 2024).

Projection matrices, such as $H \in \{0,1\}^{|V| \times |E|}$0, $H \in \{0,1\}^{|V| \times |E|}$1, and their back-projections, facilitate transfer of features between the vertex, edge, and line-expanded domains (Yang et al., 2020, Liu et al., 2022). These matrices play a key role in propagating representations between spaces and in performing aggregation/back-aggregation steps.

2. Convolutional Operators and Propagation Schemes

Line Hypergraph Convolution Networks apply variants of standard GCN propagation to the constructed line-graph representation:

• LE-GCN (Line Expansion): The main propagation rule on $H \in \{0,1\}^{|V| \times |E|}$2 is

$H \in \{0,1\}^{|V| \times |E|}$3

where initial features are projected from $H \in \{0,1\}^{|V| \times |E|}$4 to line-nodes via $H \in \{0,1\}^{|V| \times |E|}$5, and $H \in \{0,1\}^{|V| \times |E|}$6. After $H \in \{0,1\}^{|V| \times |E|}$7 layers, outputs are back-projected to vertices: $H \in \{0,1\}^{|V| \times |E|}$8 (Yang et al., 2020, Liu et al., 2022).

• Classical Line-Graph-based Convolution: Given a hypergraph line graph $H \in \{0,1\}^{|V| \times |E|}$9 with weighted adjacency $H(v, e) = 1$0, features on hyperedges (or sessions) are propagated as

$H(v, e) = 1$1

with $H(v, e) = 1$2 and $H(v, e) = 1$3 (Xia et al., 2020). No learnable weights or nonlinearities are required in each layer for this channel.

• Hyperedge Interaction-aware Convolution (HeIHNN): Extends to a three-stage scheme: (i) Node-to-Hyperedge (N2HE) with $H(v, e) = 1$4, (ii) Hyperedge-to-Hyperedge (HE2HE) using $H(v, e) = 1$5, and (iii) Hyperedge-to-Node (HE2N) using $H(v, e) = 1$6. Self-loops and normalization are included; hyperedge outlier removal further prunes weak incidences during propagation (Ye et al., 2024).
• Injective Set-Aggregator Message Passing: For higher-order permutation-invariance, embeddings of hyperedges $H(v, e) = 1$7 are updated via injective aggregation of their prior state, neighbor hyperedges (in the line-graph), and their member vertices (with their local line-neighborhood) (Srinivasan et al., 2021).

The explicit consideration of both node/hyperedge duality (LE-GCN), or hyperedge-to-hyperedge communication (HeIHNN, line-graph GCNs), improves expressive power over classical clique or star expansions.

3. Network Architectures and Training Protocols

Common architectural and training patterns across LHCN models include:

• Layer Setup: Most models utilize two GCN layers on the line (or line-expanded) graph, with ReLU or LeakyReLU activations. Hidden layers typically use 32–128 units (Yang et al., 2020, Bandyopadhyay et al., 2020, Liu et al., 2022).
• Regularization: Dropout (0.5) and weight decay ($H(v, e) = 1$8 or $H(v, e) = 1$9) are routine, with Adam as the optimizer and learning rates around $v \in e$0 (Yang et al., 2020).
• Sampling and Scalability: Neighbor sampling (e.g., up to 30 neighbors per node in LE-GCN) addresses potential quadratic blow-up in line-expanded graph edge counts for large hyperedges (Yang et al., 2020).
• Dual-Channel Models and Fusion: Several models (e.g., DHCN, ADHCN) process both the line-graph channel and a standard hypergraph-convolution channel in parallel, fusing outputs via averaging or attention mechanisms before task-specific heads (Xia et al., 2020, Liu et al., 2022).
• Projection and Back-projection: Line-graph and vertex features are bridged by projection matrices ($v \in e$1, $v \in e$2), ensuring that learned representations preserve the original hypergraph semantics (Liu et al., 2022).

4. Theoretical Properties and Unification

The line hypergraph expansion framework enjoys several strong theoretical guarantees:

• Bijectivity: The line-expansion $v \in e$3 is a bijection, enabling exact inversion under mild combinatorial preconditions (Whitney’s theorem) (Yang et al., 2020).
• Unification of Expansions: The line-expansion subsumes classical star and clique expansions:
  • Setting edge weights $v \in e$4 retrieves star-expansion GCN [Zhou et al. '06].
  • Collapsing all self-loops reduces LE-GCN to the (standard) clique-expansion GCN.
  • On 2-regular (ordinary) graphs, LE-GCN yields standard Kipf–Welling GCN up to scaling (Yang et al., 2020).
• Spectral Identities: The projection identities $v \in e$5 and $v \in e$6 relate hypergraph and line graph spectra, supporting spectral analysis and providing rigorous insight into representational fidelity (Yang et al., 2020).
• Permutation Invariance: By design, models that aggregate over sets (as in DeepSets) remain invariant to permutations of vertex or hyperedge indices at every layer (Srinivasan et al., 2021).

5. Applications and Empirical Performance

Line Hypergraph Convolution methods have demonstrated strong experimental performance on classical learning tasks involving higher-order relations.

• Node and Hyperedge Classification: LE-GCN outperforms baselines on five diverse hypergraph datasets by approximately 2% accuracy margin, with similar runtime due to neighbor sampling (Yang et al., 2020). LHCN delivers $v \in e$7 gain over best baseline on Cora, and consistent improvements on Citeseer and Pubmed (Bandyopadhyay et al., 2020).
• Recommendation and Self-Supervision: DHCN demonstrates superior performance on session-based recommendation over HGNN, HyperGCN, and others, benefiting from auxiliary self-supervised alignment between hypergraph and line-graph embeddings (mutual information maximization) (Xia et al., 2020).
• Hyperedge Classification, Expansion: Line-graph-based architectures surpass GCN/GraphSAGE on clique-expansions, RGCN, and direct set-aggregation baselines for variable-sized hyperedge expansion (with a small GAN and permutation-invariant heads) (Srinivasan et al., 2021).
• Dual-Channel and Attention Fusion: ADHCN shows that attention-based fusion of hypergraph and line-graph channels yields further accuracy improvements in technological patent data (Liu et al., 2022).

Representative results are illustrated below:

Dataset	Model	Accuracy (%)
20News	LE_GCN	75.6
	HGNN	74.3
	HyperGCN	73.6
Cora	LHCN	73.3
	HyperGCN	67.6
	HGNN	67.6

6. Limitations, Practical Considerations, and Extensions

While LHCNs provide a principled route for lifting GCNs to hypergraphs, several practical and theoretical considerations remain:

• Line Graph Density: For hypergraphs with high-order edges (large $v \in e$8), the line-graph (or line-expanded graph) may become dense, increasing computational burden ($v \in e$9 complexity, $D_v$0 average hyperedge size) (Bandyopadhyay et al., 2020).
• Attribute and Edge-weight Design: The weighting rules for line-graph edges often remain hand-designed; end-to-end learnable edge attentions or dynamic weighting schemes are conceivable extensions (Bandyopadhyay et al., 2020).
• Modular Extensions: Explicit modeling of hyperedge-to-hyperedge interactions (as in HeIHNN) introduces richer, denser mixing, which can be further enhanced via outlier pruning and higher-order spectral kernels (Ye et al., 2024).
• Inductive and Dynamic Scenarios: Extensions to time-evolving or inductive hypergraphs, as well as learned end-to-end projection/back-projection, have not been fully explored (Bandyopadhyay et al., 2020).
• Expressivity: Injective set aggregation and dual message-passing channels guarantee maximal representational invariance to input order, but the balance of computational cost and expressivity remains a subject of active research (Srinivasan et al., 2021).

7. Summary and Outlook

Line Hypergraph Convolution Networks enable the principled application of GCN-style deep learning to general hypergraphs, bridging the gap between pairwise and higher-order relation modeling. By transforming hypergraphs through line expansion or line-graph constructions, these methods deliver strong empirical accuracy, strong theoretical guarantees (bijectivity, unification, permutation-invariance), and modular extensibility to dual-channel and interaction-aware architectures.

Future work may focus on efficient sparsification/sampling for very large hyperedges, adaptive weighting and attention schemes, full end-to-end differentiability, and dynamic or inductive hypergraph settings. LHCNs constitute a foundational tool for scalable and expressive learning on higher-order relational data (Yang et al., 2020, Bandyopadhyay et al., 2020, Liu et al., 2022, Xia et al., 2020, Ye et al., 2024, Srinivasan et al., 2021).