Semi-Supervised Heterogeneous GNN

• Semi-supervised HetGNNs are models that integrate limited labeled data with graph structure to learn versatile node representations across different types.
• They employ type-specific feature projections and relation-aware aggregations, combining meta-path based and meta-path free techniques for scalability.
• Architectural innovations include graph regularization, semantic trees, and prompt-tuning methods to address challenges like label scarcity, over-squashing, and heterophily.

Semi-supervised heterogeneous graph neural networks, often discussed under the broader umbrella of HetGNN-style modeling, are node-representation and node-classification methods for graphs with multiple node types and edge types, where only a subset of target-type nodes is labeled and unlabeled nodes still contribute through graph structure and node attributes. In the cited literature, this design space spans graph-regularized deep heterogeneous models, meta-path-based semantic encoders, meta-path-free relation-aware architectures, scalable metapath-precomputation schemes, prompt-based post-training frameworks, and heterophily-aware dual-channel models (Ju et al., 2018, Wu et al., 2022, Guan et al., 2024, Ma et al., 2023, Li et al., 6 May 2026).

1. Formal setting and problem formulation

A heterogeneous graph is commonly written either as $$\mathcal{G}=(\mathcal{V},\mathcal{E},\mathcal{X},\mathcal{F},\mathcal{R})$$, with node-type mapping $\varphi:\mathcal{V}\rightarrow\mathcal{F}$ and edge-type mapping $\phi:\mathcal{E}\rightarrow\mathcal{R}$, or as $\mathcal{G}=(\mathcal{V},\mathcal{E};\phi,\omega)$, with heterogeneity condition $|\mathcal{F}|+|\mathcal{R}|\ge 2$ or $|\mathcal{A}|+|\mathcal{R}|>2$ (Huo et al., 2022, Shao et al., 2021). The recurring difficulty is that different node types may live in different feature spaces, while different edge types or meta-paths encode different semantics.

The standard semi-supervised setting focuses on a target node type $\mathcal{V}_T$, with a labeled subset $\mathcal{V}_L\subset \mathcal{V}_T$ and unlabeled nodes $\mathcal{V}_U=\mathcal{V}_T\setminus \mathcal{V}_L$. A pre-trained encoder $f_\theta$ or an end-to-end HGNN then maps nodes to embeddings used for downstream node classification, while only $\varphi:\mathcal{V}\rightarrow\mathcal{F}$0 contributes direct supervision (Ma et al., 2023). In many benchmark and industrial settings, labels exist only for one node type, whereas other node types appear only as structural context.

This regime differs from homogeneous graph SSL in two ways emphasized repeatedly in the literature. First, type-specific feature spaces require node-type-specific projections or encoders before aggregation. Second, structural information is not a single adjacency pattern but a collection of typed relations, meta-relations, or meta-paths with unequal informational value. This is why heterogeneous graph learning papers routinely distinguish node-type mappings, edge-type mappings, typed neighborhoods, and target-type-specific tasks rather than treating the graph as a single undifferentiated topology (Ju et al., 2018, Shao et al., 2021).

2. Architectural lineages

One lineage is graph-regularized deep heterogeneous semi-supervised learning. HELP, introduced for a bipartite user–domain graph, learns type-specific embeddings

$\varphi:\mathcal{V}\rightarrow\mathcal{F}$1

predicts domain labels by

$\varphi:\mathcal{V}\rightarrow\mathcal{F}$2

and scores cross-type interactions with

$\varphi:\mathcal{V}\rightarrow\mathcal{F}$3

Its training combines supervised prediction on labeled domains with a graph regularization term over user–domain pairs. The paper explicitly states that HELP is heterogeneous and semi-supervised but “does not perform neighborhood/message aggregation over heterogeneous neighborhoods,” so it is better viewed as a graph-regularized deep heterogeneous SSL model than as a canonical heterogeneous message-passing GNN (Ju et al., 2018).

A second lineage is meta-path-free heterogeneous GNN design. “Meta-path Free Semi-supervised Learning for Heterogeneous Networks” proposes simple and effective graph neural networks for heterogeneous graph, excluding the use of meta-paths, and states that the models focus on relaxing the heterogeneity stress for model parameters by expanding model capacity of general GNNs (Park et al., 2020). “Heterogeneous Graph Tree Networks” makes a related claim from a depth perspective: most existing HGNNs employ shallow Conv-GNNs such as GCN and GAT, whereas HetGTCN and HetGTAN use node-type-specific transformation, edge-type-specific neighbor aggregation, and target-specific aggregation across edge types, and are reported to go deep without compromising performance on semi-supervised node classification (Wu et al., 2022).

A third lineage keeps semantic decomposition explicit. MV-HetGNN treats each meta-path as a semantic view, learns a view-specific representation from a meta-path-based ego graph, and fuses the resulting embeddings through hierarchical autoencoders with orthogonal regularization. The paper is explicit that this architecture supports semi-supervised node classification through a labeled node set $\varphi:\mathcal{V}\rightarrow\mathcal{F}$4 and a downstream classification loss on labeled nodes only, even though embeddings are computed using graph structure and features that involve unlabeled nodes (Shao et al., 2021).

3. Semi-supervised objectives and label utilization

Semi-supervision in heterogeneous graphs is implemented through several distinct objective families. HELP uses supervised binary cross-entropy on labeled domains together with a graph regularization term over linked and unlinked user–domain pairs: $\varphi:\mathcal{V}\rightarrow\mathcal{F}$5 This makes labels sparse but graph context dense, and the model is inductive because embeddings are generated from node features rather than stored in a per-node table (Ju et al., 2018).

MV-HetGNN writes the semi-supervised downstream loss over labeled nodes as

$\varphi:\mathcal{V}\rightarrow\mathcal{F}$6

and combines it with intra-view reconstruction, inter-view reconstruction, and orthogonal regularization: $\varphi:\mathcal{V}\rightarrow\mathcal{F}$7 Here labels enter only through $\varphi:\mathcal{V}\rightarrow\mathcal{F}$8, but unlabeled nodes still affect view-specific ego-graph encoding and multi-view fusion (Shao et al., 2021).

Other models use labels not only as targets but also as input-side information. HetTree masks labels so that only training labels are visible,

$\varphi:\mathcal{V}\rightarrow\mathcal{F}$9

then aggregates these masked labels over metapath neighborhoods and fuses them with metapath features in a semantic tree. HG-MDA combines cross-entropy on labeled data with consistency regularization on unlabeled data,

$\phi:\mathcal{E}\rightarrow\mathcal{R}$0

and sharpens fused predictions from multiple heterogeneous augmentations before consistency matching (Guan et al., 2024, Chen et al., 2022).

Prompt-based post-training introduces a further variant. HetGPT freezes a contrastively pre-trained HGNN and contrastive head, injects a heterogeneous feature prompt into the input space, introduces virtual class prompts, and reformulates semi-supervised node classification as node–class contrastive matching. Its prompt-tuning loss is an InfoNCE objective over labeled target nodes plus orthogonality regularization on class tokens, while the encoder itself remains frozen (Ma et al., 2023).

4. Semantic structure beyond one-hop aggregation

A central divide in semi-supervised HetGNN research is how heterogeneous semantics are represented. Meta-path-based methods treat a typed composite path as a semantic unit. MV-HetGNN makes this explicit by defining a meta-path as

$\phi:\mathcal{E}\rightarrow\mathcal{R}$1

constructing a meta-path-based ego graph $\phi:\mathcal{E}\rightarrow\mathcal{R}$2, and encoding each view bottom-up before multi-view fusion. The fused embedding is intended to be “versatile” in the sense that it should be at least as informative as each single-view representation (Shao et al., 2021).

HetTree argues that flat semantic attention over metapaths ignores a natural prefix hierarchy among metapaths. It therefore constructs a semantic tree over all metapaths up to hop $\phi:\mathcal{E}\rightarrow\mathcal{R}$3, aggregates offline features

$\phi:\mathcal{E}\rightarrow\mathcal{R}$4

matches target-ending metapath features with propagated labels, and encodes the resulting tree bottom-up with subtree attention. The parent representation is updated by

$\phi:\mathcal{E}\rightarrow\mathcal{R}$5

so shorter metapaths absorb information from their extensions (Guan et al., 2024).

MGA-HHN generalizes the semantic view notion from graphs to hypergraphs. It constructs meta-path-based heterogeneous hypergraphs in which each symmetric meta-path defines a semantic view and the central node type of the symmetric meta-path becomes the hyperedge type. Node-level attention then operates within a view, while hyperedge-level attention fuses across views: $\phi:\mathcal{E}\rightarrow\mathcal{R}$6 The paper argues that pairwise meta-path message passing fails to capture higher-order relations among nodes and that hyperedges preserve the group structure implicit in shared semantic contexts (Jin et al., 7 May 2025).

The existence of strong meta-path-free work complicates any simple identification of semi-supervised HetGNN with meta-path-based design. The abstract of (Park et al., 2020) explicitly calls meta-path design a suboptimal remedy because “the features from the first-order neighbors are not fully leveraged for training and inference,” while (Wu et al., 2022) argues that meta-path construction is either manual or expensive. This suggests that contemporary semi-supervised HetGNN design now spans both explicit semantic decomposition and meta-path-free typed message passing.

5. Scarcity, robustness, and failure modes

Label scarcity is a foundational theme. HeCo states that most HGNNs follow a semi-supervised learning manner and proposes a self-supervised alternative based on co-contrastive learning between a network-schema view and a meta-path view. HGCL makes a related argument, claiming that most existing HGNN-based approaches are supervised or semi-supervised and that heterogeneous contrastive learning must account for noise in both node attributes and graph topologies. Neither method is semi-supervised during representation learning, but both are directly relevant to low-label HetGNN pipelines because their embeddings are evaluated by downstream node classification with limited labels (Wang et al., 2021, Huo et al., 2022).

Over-squashing and graph perturbation are treated explicitly in HG-MDA. The paper argues that heterogeneous graph augmentation must respect heterogeneous relations and the information contribution of different types of neighbors and edges, then proposes node augmentation, topology augmentation, and triangle-based edge adding and removing. Its semi-supervised loss combines cross-entropy for labeled data and consistency regularization for unlabeled data, while sharpening fuses predictions from various heterogeneous augmentations (Chen et al., 2022).

Topological bias has been extended from homogeneous GNNs to heterogeneous node learning. HTAD introduces a meta-weighted adjacency matrix,

$\phi:\mathcal{E}\rightarrow\mathcal{R}$7

constructs a projection called HLID from a PageRank-like impact matrix and a label-availability vector,

$\phi:\mathcal{E}\rightarrow\mathcal{R}$8

and then uses HLID differences to sample an augmented heterogeneous graph for contrastive debiasing. The paper argues that degree is an inadequate proxy for structural bias in heterogeneous graphs and that topological bias is universal across datasets with and without intra-type connections (Zhang, 4 Dec 2025).

Heterophily has become a distinct axis of heterogeneous graph learning. Hetero$\phi:\mathcal{E}\rightarrow\mathcal{R}$9Net defines metapath-based label homophily and metapath-based Dirichlet energy,

$\mathcal{G}=(\mathcal{V},\mathcal{E};\phi,\omega)$0

and argues that existing HGNNs fail to generalize to heterogeneous graphs with heterophily or low level of homophily. HeterSEED pushes this further by decoupling a heterogeneous semantic channel from a structure-aware heterophily channel, partitioning metapath-induced neighbors into homophilic and heterophilic subsets via pseudo-labels, weighting them by metapath counts, and fusing the two channels adaptively per node (Li et al., 2023, Li et al., 6 May 2026).

6. Empirical landscape, scope, and common misconceptions

The empirical record is not limited to small benchmarks. MV-HetGNN reports DBLP Macro-F1 around $\mathcal{G}=(\mathcal{V},\mathcal{E};\phi,\omega)$1–$\mathcal{G}=(\mathcal{V},\mathcal{E};\phi,\omega)$2 and Micro-F1 around $\mathcal{G}=(\mathcal{V},\mathcal{E};\phi,\omega)$3–$\mathcal{G}=(\mathcal{V},\mathcal{E};\phi,\omega)$4, while IMDb is around $\mathcal{G}=(\mathcal{V},\mathcal{E};\phi,\omega)$5–$\mathcal{G}=(\mathcal{V},\mathcal{E};\phi,\omega)$6 Macro/Micro-F1. HetGTCN and HetGTAN report $\mathcal{G}=(\mathcal{V},\mathcal{E};\phi,\omega)$7 Macro-F1 on ACM, $\mathcal{G}=(\mathcal{V},\mathcal{E};\phi,\omega)$8 Macro/Micro-F1 on IMDB, and $\mathcal{G}=(\mathcal{V},\mathcal{E};\phi,\omega)$9 on DBLP. HetTree reaches $|\mathcal{F}|+|\mathcal{R}|\ge 2$0 test accuracy on ogbn-mag and $|\mathcal{F}|+|\mathcal{R}|\ge 2$1 test accuracy on a real-world email graph, and HeterSEED reports $|\mathcal{F}|+|\mathcal{R}|\ge 2$2 on IMDB and $|\mathcal{F}|+|\mathcal{R}|\ge 2$3 on MAG, with strongest gains in strongly heterophilic regimes (Shao et al., 2021, Wu et al., 2022, Guan et al., 2024, Li et al., 6 May 2026).

A recurrent misconception is that every model with “HetGNN” in its name belongs to semi-supervised heterogeneous node classification. The literature explicitly rejects this. SR-HetGNN is a session-based recommendation model trained with cross-entropy over the next clicked item; it is “not a semi-supervised heterogeneous node-classification paper.” SAGE-Het is a supervised, application-specific heterogeneous GNN for railway delay regression and is “not framed as semi-supervised, self-supervised, or contrastive learning” (Chen et al., 2021, Li et al., 2023).

A second misconception is that all relevant work must itself be semi-supervised at the representation-learning stage. HeCo and HGCL are self-supervised; HetGPT is a post-training prompting framework on top of contrastively pre-trained HGNNs. Yet all three are relevant because they target the low-label regime that motivates semi-supervised HetGNNs in the first place (Wang et al., 2021, Huo et al., 2022, Ma et al., 2023).

A third misconception is that every semi-supervised heterogeneous model is a message-passing HGNN in the strict sense. HELP is directly relevant to semi-supervised learning on heterogeneous graphs and is inductive, but it “does not perform neighborhood/message aggregation over heterogeneous neighborhoods.” This makes it a precursor or adjacent model rather than a canonical heterogeneous GNN (Ju et al., 2018).

Work	Learning regime	Relation to semi-supervised HetGNN
HELP (Ju et al., 2018)	Semi-supervised	Heterogeneous, inductive, graph-regularized; not a canonical message-passing HGNN
HeCo (Wang et al., 2021)	Self-supervised	Label-free pretraining alternative for low-label heterogeneous node classification
HGCL (Huo et al., 2022)	Self-supervised	Robust contrastive pretraining usable in downstream semi-supervised settings
HetGPT (Ma et al., 2023)	Semi-supervised post-training	Prompt-tuning framework for pre-trained HGNNs on semi-supervised node classification
SR-HetGNN (Chen et al., 2021)	Supervised	Session recommendation, not semi-supervised node classification
SAGE-Het (Li et al., 2023)	Supervised	Heterogeneous node regression, not semi-supervised node classification

Taken together, these works show that “Semi-Supervised Heterogeneous Graph Neural Network (HetGNN)” is not a single settled architecture but a research field organized around a stable problem—typed graph learning with limited labels—and an evolving set of solutions. The main axes of variation are whether heterogeneity is handled through relation-specific message passing or semantic views, whether higher-order structure is represented by meta-paths, trees, or hypergraphs, how labels enter the model, and how the method reacts to label scarcity, topological bias, over-squashing, and heterophily.