acceptodds
Under review as a conference paper at ICLR 2027

Subgraph Representation Learning on Heterogeneous Graphs

Abstract

Subgraph representation learning on heterogeneous graphs requires accounting jointly for subgraph membership and the types of nodes and relations in a shared base graph. Existing approaches primarily address node- or graph-level heterogeneity, leaving subgraph-level challenges less explored. We propose Cross-Type Anchor Projection (CTAP), which decomposes the target subgraphs and the base graph by node type, representing each type-specific node set within its surrounding graph. Cross-type projection maps each other-type node (an anchor) to the subgraph of its neighbors, yielding views that relate these sets to anchors and capture long-range dependencies. Semantic fusion combines representations learned on the views, while membership pooling incorporates cross-type neighbor and constituent-node information before the final fusion across node types. Among the compared methods, CTAP achieves the highest mean test micro-F1 on five of six new datasets drawn from team discovery, rare-disease diagnosis, and gene/protein set analysis.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
subgraph representation learning, heterogeneous graphs, graph neural networks, graph pooling

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_iJ1UOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces CTAP, which extends subgraph representation learning to heterogeneous graphs, providing a more diverse and flexible setting.

CTAP first decomposes a shared base heterogeneous graph into node-overlap and edge-connectivity views for each node type, followed by GCN-like message passing, SeHGNN-like semantic fusion, and finally membership pooling.

Experiments show promising gains over existing homogeneous subgraph-based GNNs as well as heterogeneous GNN baselines on five datasets.

Critical strengths
  1. Empirical validation is conducted thoroughly.
  2. The illustration figures are helpful.
  3. Strong gains over existing baselines.
  4. One of the first extensions of subgraph representation learning to heterogeneous graphs.
Critical weaknesses
  1. [Presentation] The challenges of extending existing methods to heterogeneous graphs are not clearly discussed in the main text. Explaining why such an extension is difficult and non-trivial should be one of the most important elements in establishing the significance of this work.

  2. [Presentation] Some experiments lack informative discussion. For example, Section 6.2, "Performance with Reduced Training Data", plainly shows the superior performance of the proposed method. However, the discussion does not explain why or how CTAP is particularly effective in limited-data scenarios, nor does it justify why reducing the training data is particularly meaningful in this setting. A similar concern applies to Figure 6.

Questions for authors
  1. How are homogeneous subgraph baselines (e.g., SubGNN) adapted to heterogeneous graphs, given that they typically expect homogeneous graphs as input and may not support heterogeneous features across different node types? Are the heterogeneous graphs converted into homogeneous graphs for these baselines?

  2. How are node features constructed in the datasets? Heterogeneous graphs such as DBLP and Movie (IMDB) often lack intrinsic features and typically use bias features (e.g., node-type IDs or zero features), identity features (e.g., node IDs), propagated features (e.g., propagated term IDs as bag-of-words features) [1], or PLM-based term embeddings [2].

  3. As discussed in [1], the choice of node features (e.g., bias features, identity features) can have a large impact on downstream performance. This may be particularly relevant to the proposed approach due to its decomposition and projection. Could the authors briefly discuss how feature construction affects the proposed method?

[1]: Are we really making much progress? Revisiting, benchmarking, and refining heterogeneous graph neural networks. KDD 2021 [2]: Open Graph Benchmark: Datasets for Machine Learning on Graphs. NIPS 2020

Reviewer_NHrJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

This paper proposes CTAP for subgraph representation learning and classification within a shared heterogeneous graph. CTAP first decomposes the base graph and target subgraphs by node type, then projects nodes of other types into subgraphs induced by their neighbors. It constructs node-overlap, edge-connectivity, and same-type views, combining graph neural networks, attention-based fusion, and membership pooling to obtain target subgraph representations. The authors construct six datasets covering publications, movies, rare diseases, and biological functions. CTAP achieves the highest mean micro-F1 on five of them. The evaluation is primarily limited to subgraph classification within a given shared base graph.

Critical strengths

CTAP effectively integrates subgraph membership and heterogeneous graph context, achieving the highest mean micro-F1 on five of six datasets.

Critical weaknesses
  1. Compared with existing approaches, CTAP demonstrates limited methodological novelty and no clear advantage in its design, primarily combining multi-view graph construction with standard GNN propagation and feature fusion.
  2. Summarizing inter-set relations through membership overlap counts and aggregated edge weights may lose fine-grained structural information. Whether the resulting views can distinguish different structural patterns requires further evaluation.
  3. The target subgraphs of DBLP, Movie, and Protein Set Function all contain only a single node type, which can verify the utilization of heterogeneous background information, but is insufficient to verify the contribution of cross-type fusion within the target subgraphs.
  4. Existing ablations do not isolate the contribution of cross-type projection, since its removal also disables dependent pooling operations. Consequently, it remains unclear which relations projection captures more effectively than standard message passing.
Questions for authors

See weakness

Reviewer_ArKMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CTAP for subgraph representation learning on heterogeneous graphs. The core idea is to first decompose both the base graph and the target subgraphs by node type into type-specific subproblems; within each subproblem, neighbors of another type are projected as induced subgraphs on the type-τ base graph, which are then translated together with the target subgraph into nodes via Subgraph-To-Node (S2N) translation to construct three views, a node-overlap view, an edge-connectivity view, and a same-type view. Each view passes through an MPNN, outputs are combined via Transformer-style semantic fusion, then membership pooling incorporates cross-type neighbor features into node representations and constituent-node features into subgraph representations, iterating layer by layer, and finally fusing across node types to obtain the subgraph representation.

Critical strengths
  1. The problem decomposition reflects a genuine structural insight.
  2. The combination of view design and S2N is natural and implementable.
  3. The ablation conclusions are conducted well.
Critical weaknesses
  1. The projection granularity is fixed at one hop, and the sensitivity of this choice is not discussed.
  2. The 10.9 pp gain on Movie is anomalously large, and the paper does not rule out information asymmetry in view construction.
  3. The complexity analysis gives two different bounds, but the main text cites only the looser one, and the implementation-version bound relies on the assumption of at most one relation per pair of types.
  4. The edge-weight definitions of the node-overlap and edge-connectivity views are conceptually redundant, and the paper does not justify when the two provide complementary information.
Questions for authors
  1. Jumping Knowledge concatenation is adopted by all six configurations, but Section 5.4 mentions JK only in passing as implementation also considers, and Table 2 has no JK ablation.
  2. The analysis of failure cases stays at the level of task saturation / long-range dependencies already provide signal and lacks diagnosis of CTAP's own failure modes.

Comments on the paper

1 comment

Hi Reviewers,

We are now near the end of the rebuttal period. If you haven't done so yet, please review the authors' responses and post a reply addressing whether their clarifications or new results resolve your concerns. Even if your score or recommendation remains unchanged, a brief note confirming what did or didn't shift your assessment—and why—is essential for the discussion and final decision. Thank you for your time and engagement.

Best regards,

Area Chair

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.