HSCL: Hypergraph Self-Supervised Contrastive Learning

• HSCL is a representation learning technique that uses hypergraphs to capture group-level and many-to-many relations beyond traditional pairwise reductions.
• It combines hyperedge-aware augmentation with multi-scale contrastive objectives to enhance tasks like node classification, recommendation, and trust evaluation.
• Empirical studies show HSCL outperforms conventional methods by preserving higher-order semantics while improving accuracy and efficiency.

Hypergraph-enabled Self-Supervised Contrastive Learning (HSCL) denotes a family of representation-learning methods that use hypergraphs to encode group-wise or higher-order relations and self-supervised objectives to learn node, hyperedge, and sometimes membership embeddings without relying exclusively on labels. Across hyperedge prediction, node classification, recommendation, trust evaluation, and text-attributed hypergraph learning, the unifying premise is that pairwise graph reductions can discard higher-order semantics, whereas native hypergraph modeling preserves group structure and makes it possible to construct self-supervised signals at multiple semantic scales (Ko et al., 2023, Roh et al., 12 Feb 2025, Xia et al., 2022, Li et al., 2024, Pan et al., 5 Aug 2025).

1. Conceptual scope and problem setting

A hypergraph is typically written as $G=(V,E)$ or $\mathcal{H}=(\mathcal{V},\mathcal{E})$, where each hyperedge may connect an arbitrary number of nodes. In the HSCL literature, this matters because the target relation is often intrinsically group-wise rather than pairwise. Hyperedge prediction in CASH, for example, asks whether a candidate set of nodes $e' \notin E$ should form a real hyperedge, which differs fundamentally from ordinary link prediction because the target is a set-valued relation rather than a binary edge (Ko et al., 2023). HyFi similarly treats common hyperedges as the defining structural signal for contrastive supervision, arguing that hypergraphs expose many-to-many and group-level relations that ordinary homogeneous graphs do not represent faithfully (Roh et al., 12 Feb 2025).

Within this family, HSCL is not restricted to a single task. It appears in unsupervised hypergraph representation learning for node classification and clustering, semi-supervised node classification, recommendation, text-attributed hypergraph learning, and trust evaluation in Social IoT. Some systems are fully self-supervised, such as HyFi, SE-HSSL, HiTeC, and the SIoT trust model; others are hybrid systems that combine task supervision with contrastive regularization, such as CASH for hyperedge prediction, HCCF for collaborative filtering, and CHGNN for semi-supervised classification (Roh et al., 12 Feb 2025, Li et al., 2024, Pan et al., 5 Aug 2025, Zhu et al., 9 Sep 2025, Ko et al., 2023, Xia et al., 2022, Song et al., 2023).

A central distinction from ordinary graph contrastive learning is that HSCL treats hyperedges, groups, or memberships as first-class objects. This produces self-supervised signals that are unavailable in pairwise graph settings: node-level consistency, hyperedge-level consistency, node–hyperedge membership consistency, cluster-level or subgraph-level consistency, and cross-view local–global consistency. This suggests that HSCL is best understood not as a single algorithmic template, but as a design space organized around three interacting choices: how higher-order relations are represented, how views are generated, and at what semantic scale contrast is applied.

2. Hypergraph-native representations and encoders

Most HSCL systems use native hypergraph operators based on incidence structure rather than clique expansion or line-graph conversion. In CASH, with node feature matrix $\mathbf{X}\in\mathbb{R}^{|V|\times F}$ and incidence matrix $\mathbf{H}\in\{0,1\}^{|V|\times |E|}$, the encoder alternates node-to-hyperedge and hyperedge-to-node propagation: $$\mathbf{Q}^{(k)} = \sigma\!\left(\mathbf{D}^{-1}_{E}\mathbf{H}^{T}\mathbf{P}^{(k-1)} \mathbf{W}^{(k)}_{E} + b^{(k)}_{E}\right), \qquad \mathbf{P}^{(k)} = \sigma\!\left(\mathbf{D}^{-1}_{V}\mathbf{H}\mathbf{Q}^{(k)} \mathbf{W}^{(k)}_{V} + b^{(k)}_{V}\right),$$ with $\sigma$ set to PReLU (Ko et al., 2023). HiTeC and SE-HSSL use closely related HGNN-style two-stage propagation, again operating directly on incidence matrices and degree normalizers rather than graphized reductions (Pan et al., 5 Aug 2025, Li et al., 2024).

HyFi uses the HGNN model of Feng et al. as encoder $f_\theta(\cdot)$, with

$$Q^{(k)} = \sigma\left( D_E^{-1} H^T D_V^{-1/2} P^{(k-1)} \Theta_E^{(k)} + b_E^{(k)} \right), \qquad P^{(k)} = \sigma\left( D_V^{-1/2} H W Q^{(k)} \Theta_V^{(k)} + b_V^{(k)} \right),$$

and $W$ set to the identity matrix (Roh et al., 12 Feb 2025). The SIoT trust model uses the same structural pattern, expressed as device-to-hyperedge and hyperedge-to-device aggregation in a parameter-sharing HGNN, because devices and social groups are both explicit entities in the learned representation (Zhu et al., 9 Sep 2025).

Several methods depart from the standard HGNN operator in task-specific ways. HCCF does not use a fixed incidence-matrix hypergraph convolution; instead it learns separate user-side and item-side hypergraph dependency matrices,

$\mathcal{H}=(\mathcal{V},\mathcal{E})$0

and propagates through the induced operator $\mathcal{H}=(\mathcal{V},\mathcal{E})$1 with an additional hierarchical hypergraph mapping $\mathcal{H}=(\mathcal{V},\mathcal{E})$2 (Xia et al., 2022). CHGNN likewise emphasizes hypergraph-specific encoding, but its distinctive choice is to weight incident hyperedges by a learned notion of hyperedge homogeneity during node updates, rather than presenting the model in the usual incidence-Laplacian form (Song et al., 2023).

The common pattern is nevertheless clear: HSCL methods generally preserve the node–hyperedge duality in the encoder itself. A plausible implication is that native node–hyperedge message passing has become the default representational substrate for HSCL, while latent hypergraphs, homogeneity-aware propagation, and text-aware variants arise when the downstream task imposes additional structural biases.

3. View generation and augmentation

View construction is one of the main axes along which HSCL methods differ. CASH uses hyperedge-aware augmentation: for each hyperedge $\mathcal{H}=(\mathcal{V},\mathcal{E})$3, it masks a random fraction $\mathcal{H}=(\mathcal{V},\mathcal{E})$4 of that hyperedge’s members independently, rather than masking a global fraction of incidence entries, and also applies node feature masking by zeroing a random fraction $\mathcal{H}=(\mathcal{V},\mathcal{E})$5 of feature dimensions (Ko et al., 2023). The explicit goal is to preserve the distribution and existence of group-wise relations better than random global masking.

HyFi takes the opposite position and avoids topology perturbation altogether. It keeps the hypergraph structure fixed and perturbs only node features,

$\mathcal{H}=(\mathcal{V},\mathcal{E})$6

arguing that random node, hyperedge, or incidence dropping corrupts the very topology that carries higher-order signal. Among the tested augmentation types, Gaussian noise achieved the best average rank (Roh et al., 12 Feb 2025).

CHGNN learns its augmentations adaptively. Its view generator chooses among preserve, remove, and mask operations for each hyperedge, samples these decisions via Gumbel-softmax, and regularizes the two augmentation vectors with a similarity loss $\mathcal{H}=(\mathcal{V},\mathcal{E})$7 motivated by a minimal-sufficient-view perspective (Song et al., 2023). SE-HSSL uses a simpler paired augmentation scheme—node feature masking and membership masking—but its contribution lies less in the view generator itself than in reducing dependence on negative-sample-heavy contrast at later stages (Li et al., 2024).

HiTeC extends augmentation into the text-attributed regime. Its two semantic-aware strategies are prompt-enhanced text augmentation and semantic-aware hyperedge drop. The prompt-enhanced view concatenates original text with domain knowledge, global topology, and neighbor context, while semantic-aware hyperedge drop modulates incidence removal by a semantic cohesiveness score computed from average pairwise cosine similarity among member node features (Pan et al., 5 Aug 2025). In the SIoT trust model, augmentation acts directly on devices, hyperedges, and device–hyperedge incidences through independently sampled Bernoulli masks, yielding two structurally perturbed social hypergraphs (Zhu et al., 9 Sep 2025).

A more specialized variant appears in motif-based recommendation. Motifs-Res does not rely on stochastic perturbation as its primary source of views; instead, it treats social, joint, and purchase motifs as distinct semantic channels and constructs matched cross-channel representations before applying inter-channel contrast. The paper explicitly reports that directly adding Triplet or InfoNCE between raw motif channels degrades performance, which identifies a failure mode of naive cross-view alignment when the views are semantically heterogeneous rather than augmentation-equivalent (Sun et al., 2021).

4. Contrastive objectives and supervisory granularity

HSCL is defined as much by what it contrasts as by how it encodes. CASH combines supervised prediction with a dual contrastive regularizer over nodes and hyperedges: $\mathcal{H}=(\mathcal{V},\mathcal{E})$8 Its contrastive branch is therefore node-level and group-level, but the paper does not provide a full InfoNCE-style formula with temperature, explicit positive and negative sets, or denominator terms over negatives. The most faithful characterization is that CASH maximizes cosine similarity across two augmented views at node and hyperedge levels rather than specifying a temperature-scaled instance-discrimination objective (Ko et al., 2023).

HyFi makes the contrastive relation itself hypergraph-native. It introduces the notion of “commonality,” defined through shared hyperedges, and adds a third sample type beyond positives and negatives: the weak positive pair. With

$\mathcal{H}=(\mathcal{V},\mathcal{E})$9

the number of shared hyperedges between nodes determines a weak-positive weight,

$e' \notin E$0

Strict positives are the same node across original and noisy views, weak positives are different nodes sharing at least one hyperedge, and negatives are nodes sharing no hyperedge. The same logic is mirrored at the hyperedge level through shared nodes (Roh et al., 12 Feb 2025). This is one of the clearest examples of HSCL redefining “positive pair” in a way that only makes sense in a hypergraph.

Other methods broaden the semantic scale further. HCCF aligns graph-local embeddings $e' \notin E$1 with hypergraph-global embeddings $e' \notin E$2 node by node and layer by layer, thereby using cross-view contrast between explicit local collaborative signals and latent global dependencies (Xia et al., 2022). CHGNN adds a multi-part objective containing cluster-level and hyperedge-level basic contrast, plus cross-validation losses for node–cluster, node–hyperedge, and hyperedge–cluster pairs, and an enhanced training process that adaptively changes temperatures for harder negatives (Song et al., 2023). SE-HSSL replaces node-level and group-level instance discrimination with sampling-free CCA-style invariance and decorrelation objectives, while retaining a hierarchical membership-level contrastive loss over adjacent membership shells defined by overlap-induced hop structure (Li et al., 2024).

HiTeC adds a hierarchical multi-scale objective for text-attributed hypergraphs: $e' \notin E$3 The distinctive term is $e' \notin E$4, a subgraph-level InfoNCE loss built from $e' \notin E$5-walk sampled subgraphs, where consecutive hyperedges in a walk must overlap in at least $e' \notin E$6 nodes (Pan et al., 5 Aug 2025). The SIoT trust model instead uses a three-part InfoNCE decomposition over devices, hyperedges, and device–hyperedge memberships, with separate temperatures $e' \notin E$7, $e' \notin E$8, and $e' \notin E$9 (Zhu et al., 9 Sep 2025).

Taken together, these formulations show that HSCL has moved beyond a single node-only contrastive recipe. Node-, hyperedge-, membership-, cluster-, local–global-, motif-, and subgraph-level supervision all appear, and the choice among them is strongly task dependent.

5. Representative systems and empirical record

The empirical record of HSCL is spread across multiple tasks and evaluation protocols, but a consistent pattern is that hypergraph-native self-supervision improves either effectiveness, efficiency, or both.

System	Task domain	Headline empirical finding
CASH (Ko et al., 2023)	Hyperedge prediction	Best averaged AUROC and AP on six hypergraphs; average AUROC improvements over AHP reach $\mathbf{X}\in\mathbb{R}^{|V|\times F}$0 on $\mathbf{X}\in\mathbb{R}^{|V|\times F}$1
HyFi (Roh et al., 12 Feb 2025)	Unsupervised node classification	Average rank $\mathbf{X}\in\mathbb{R}^{|V|\times F}$2 across 10 datasets; lower training time and GPU memory cost than TriCL
HCCF (Xia et al., 2022)	Collaborative filtering	Average relative gains in NDCG@20 of $\mathbf{X}\in\mathbb{R}^{|V|\times F}$3 over LightGCN, $\mathbf{X}\in\mathbb{R}^{|V|\times F}$4 over DHCF, $\mathbf{X}\in\mathbb{R}^{|V|\times F}$5 over MHCN, and $\mathbf{X}\in\mathbb{R}^{|V|\times F}$6 over SGL
CHGNN (Song et al., 2023)	Semi-supervised node classification	Outperforms 13 competitors on nine datasets; on Cora with $\mathbf{X}\in\mathbb{R}^{|V|\times F}$7 labels, $\mathbf{X}\in\mathbb{R}^{|V|\times F}$8 versus TriCL $\mathbf{X}\in\mathbb{R}^{|V|\times F}$9
SE-HSSL (Li et al., 2024)	Unsupervised node classification and clustering	Average gains over TriCL of $\mathbf{H}\in\{0,1\}^{|V|\times |E|}$0 NMI and $\mathbf{H}\in\{0,1\}^{|V|\times |E|}$1 ARI; at least $\mathbf{H}\in\{0,1\}^{|V|\times |E|}$2 faster on PubMed and ModelNet40
HiTeC (Pan et al., 5 Aug 2025)	Text-attributed hypergraphs	Best node classification on all six datasets and best hyperedge prediction on 5 of 6; exceeds second-best baseline by $\mathbf{H}\in\{0,1\}^{|V|\times |E|}$3 on Cora and $\mathbf{H}\in\{0,1\}^{|V|\times |E|}$4 on Fitness
HSCL for SIoT (Zhu et al., 9 Sep 2025)	Trust evaluation	Better separation of trusted and untrusted nodes than TSTCM and ATM; identifies higher-trust collaborators on Sigcomm-2009

Ablation studies are especially informative because they isolate which HSCL ingredients matter. In CASH, self-supervised contrastive learning improves prediction accuracy consistently over the no-contrastive variant, hyperedge-aware augmentation outperforms simple random augmentation, and context-aware aggregation adds further gains; the full model improves averaged AUROC over the baseline by $\mathbf{H}\in\{0,1\}^{|V|\times |E|}$5 on $\mathbf{H}\in\{0,1\}^{|V|\times |E|}$6, $\mathbf{H}\in\{0,1\}^{|V|\times |E|}$7 on $\mathbf{H}\in\{0,1\}^{|V|\times |E|}$8, $\mathbf{H}\in\{0,1\}^{|V|\times |E|}$9 on $$\mathbf{Q}^{(k)} = \sigma\!\left(\mathbf{D}^{-1}_{E}\mathbf{H}^{T}\mathbf{P}^{(k-1)} \mathbf{W}^{(k)}_{E} + b^{(k)}_{E}\right), \qquad \mathbf{P}^{(k)} = \sigma\!\left(\mathbf{D}^{-1}_{V}\mathbf{H}\mathbf{Q}^{(k)} \mathbf{W}^{(k)}_{V} + b^{(k)}_{V}\right),$$0, and $$\mathbf{Q}^{(k)} = \sigma\!\left(\mathbf{D}^{-1}_{E}\mathbf{H}^{T}\mathbf{P}^{(k-1)} \mathbf{W}^{(k)}_{E} + b^{(k)}_{E}\right), \qquad \mathbf{P}^{(k)} = \sigma\!\left(\mathbf{D}^{-1}_{V}\mathbf{H}\mathbf{Q}^{(k)} \mathbf{W}^{(k)}_{V} + b^{(k)}_{V}\right),$$1 on $$\mathbf{Q}^{(k)} = \sigma\!\left(\mathbf{D}^{-1}_{E}\mathbf{H}^{T}\mathbf{P}^{(k-1)} \mathbf{W}^{(k)}_{E} + b^{(k)}_{E}\right), \qquad \mathbf{P}^{(k)} = \sigma\!\left(\mathbf{D}^{-1}_{V}\mathbf{H}\mathbf{Q}^{(k)} \mathbf{W}^{(k)}_{V} + b^{(k)}_{V}\right),$$2 (Ko et al., 2023). In HyFi, removing weak positive pairs causes the largest degradation, changing average rank from $$\mathbf{Q}^{(k)} = \sigma\!\left(\mathbf{D}^{-1}_{E}\mathbf{H}^{T}\mathbf{P}^{(k-1)} \mathbf{W}^{(k)}_{E} + b^{(k)}_{E}\right), \qquad \mathbf{P}^{(k)} = \sigma\!\left(\mathbf{D}^{-1}_{V}\mathbf{H}\mathbf{Q}^{(k)} \mathbf{W}^{(k)}_{V} + b^{(k)}_{V}\right),$$3 to $$\mathbf{Q}^{(k)} = \sigma\!\left(\mathbf{D}^{-1}_{E}\mathbf{H}^{T}\mathbf{P}^{(k-1)} \mathbf{W}^{(k)}_{E} + b^{(k)}_{E}\right), \qquad \mathbf{P}^{(k)} = \sigma\!\left(\mathbf{D}^{-1}_{V}\mathbf{H}\mathbf{Q}^{(k)} \mathbf{W}^{(k)}_{V} + b^{(k)}_{V}\right),$$4, which indicates that shared-group structure contributes more than ordinary strict positive pairs in many cases (Roh et al., 12 Feb 2025). In SE-HSSL, removing any of node-level CCA, group-level CCA, or hierarchical membership loss degrades performance, supporting the claim that efficient HSCL benefits from complementary signals rather than a single contrastive term (Li et al., 2024).

The recommendation literature exhibits a related pattern. HCCF attributes its robustness on sparse data to the combination of local graph encoding, global hypergraph encoding, and cross-view contrast; the gains are especially large on Amazon-book, the sparsest of its datasets (Xia et al., 2022). Motifs-Res reaches smaller but consistent improvements over $$\mathbf{Q}^{(k)} = \sigma\!\left(\mathbf{D}^{-1}_{E}\mathbf{H}^{T}\mathbf{P}^{(k-1)} \mathbf{W}^{(k)}_{E} + b^{(k)}_{E}\right), \qquad \mathbf{P}^{(k)} = \sigma\!\left(\mathbf{D}^{-1}_{V}\mathbf{H}\mathbf{Q}^{(k)} \mathbf{W}^{(k)}_{V} + b^{(k)}_{V}\right),$$5-MHCN in both general and cold-start settings, and its most distinctive empirical result is negative: direct cross-channel Triplet or InfoNCE on raw motif embeddings degrades performance, which is evidence that HSCL must distinguish between compatible and incompatible positive-pair definitions (Sun et al., 2021).

6. Limitations, misconceptions, and open directions

Several misconceptions recur in the HSCL literature. One is that hypergraph contrastive learning is simply graph contrastive learning after clique expansion. The surveyed systems contradict this directly: they use incidence-based encoders, hyperedge-aware augmentation, group-level or membership-level objectives, and in some cases hyperedge or subgraph entities that have no direct analogue in ordinary graphs (Ko et al., 2023, Roh et al., 12 Feb 2025, Li et al., 2024, Pan et al., 5 Aug 2025). A related misconception is that any two available views should be aligned directly. Motifs-Res shows the opposite for semantically heterogeneous motif channels: direct cross-channel contrast causes representation homogenization, and a matching representation is required before inter-view contrast becomes useful (Sun et al., 2021).

There are also important methodological limits. CASH does not fully formalize its contrastive loss in the standard InfoNCE sense, so temperature scaling, explicit positive/negative pair indexing, and negative-set construction are absent (Ko et al., 2023). CHGNN is not a purely self-supervised system; its classifier and cluster projection are tied to the downstream label space, and its evaluation is transductive node classification rather than generic pretraining (Song et al., 2023). HCCF’s latent hypergraphs improve flexibility but reduce interpretability relative to explicit hyperedges (Xia et al., 2022). HyFi assumes that nodes sharing more hyperedges are more semantically similar; if shared membership is weakly informative, weak positives may become noisy (Roh et al., 12 Feb 2025). SE-HSSL defines membership hierarchy through clique-expansion distances rather than a native hypergraph geodesic, which is an effective engineering compromise but not a fully intrinsic hypergraph metric (Li et al., 2024). HiTeC depends on text quality, prompt design, and semantic similarity estimates, and its benchmarks are partly reconstructed because no public TAHG benchmark exists in the required form (Pan et al., 5 Aug 2025). The SIoT trust model is evaluated on a relatively small dataset with 76 nodes, uses assigned 2D coordinates because the dataset lacks geographic coordinates, and does not provide a detailed internal ablation isolating each design choice (Zhu et al., 9 Sep 2025).

An adjacent design space is illustrated by "Fusion Self-supervised Learning for Recommendation" (Zhang et al., 2024). That work is not hypergraph-enabled—it operates on a user–item bipartite graph with no hyperedge construction—but it shows that augmentation-free high-order view construction and fusion-style contrastive denominators can be effective. This suggests a neighboring line of inquiry for HSCL rather than a direct instance of it.

Natural directions already visible in the literature include richer hypergraph-specific negative sampling inside contrastive losses, membership-level contrast beyond node and group consistency, temperature-scaled formulations where they are presently absent, time-aware hypergraph embedding for recommendation, and broader validation of trust-oriented HSCL beyond small SIoT settings (Ko et al., 2023, Xia et al., 2022, Zhu et al., 9 Sep 2025). A plausible implication is that future HSCL systems will be judged less by whether they use contrastive learning at all than by whether their notions of view, positive pair, and semantic scale are genuinely native to higher-order structure.