acceptodds
Withdrawn from ICLR 2027

Hard Latent-Table Adaptation From Link Prediction to Subgraph Tasks

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Link-prediction pretraining supplies node representations for subgraph classification. Adapting these representations, however, requires subgraph-level labels to update a node table shared by overlapping examples. These labels constrain aggregated predictions, leaving the individual node updates only indirectly supervised. We propose HardLT, a latent-table adaptation objective that couples complete-view task fitting with supervised training on globally coherent partial views. It constructs these views by masking fixed PCA directions jointly across all nodes before aggregation, then selects one common mask using the minibatch’s mean classification loss. The complete view fits the task, while the selected view supplies an additional constraint on adaptation. Dataset-specific configurations also align the two predictions through a consistency loss. With the GPEN encoder, HardLT achieves the highest mean scores on HPO-Metab, EM-User, and HPO-Neuro among the compared configurations, including a tie with GPEN on HPO-Neuro. In matched component studies, HardLT raises mean micro-F1 from 0.634 to 0.647 on HPO-Metab and from 0.916 to 0.922 on EM-User. Global sharing gives higher means than nodewise perturbations in both studies. Fixed-view evaluation on the same HPO-Metab checkpoints shows smaller performance decreases under shared latent-direction perturbations. HardLT adds no parameters or computation at inference.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
subgraph classification, graph representation learning, link prediction pretraining, transfer learning, representation regularization, graph neural networks

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_FQUrOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies the adaptation of a node table that is pretrained by link prediction and later fine-tuned for subgraph classification. A subgraph label supervises only the aggregated prediction and does not specify which individual node representations should be updated. Because overlapping subgraphs share the same table rows, they can send conflicting task signals to the model, so training on complete views alone leaves the adaptation of the shared table underdetermined. The authors propose HardLT, which first constructs a fixed reference system from a principal component analysis of the initial table, including a protected set of directions. Within each batch, the method samples several candidate views, each of which applies the same direction mask to every node, and then selects the hardest view according to the mean classification loss of the batch. The model is trained jointly on the complete view and on the selected difficult view to improve classification performance. Experiments adopt the GPEN encoder and report micro F1 on four subgraph classification datasets. Relative to GPEN, HardLT yields absolute micro‑F1 improvements of 0.004 on PPI‑BP, 0.008 on HPO‑Metab and 0.010 on EM‑User, whereas no performance gain is observed on HPO‑Neuro.

Critical strengths
  1. The problem setting is distinctive. Starting from the shift in supervision granularity, from edge level pretraining to subgraph level labels, the paper argues that a shared node table becomes underdetermined under aggregated labels.
  2. The globally shared mask is an attractive technical idea. The design aligns the scope of the perturbation, where all nodes share the same direction set, with the shared scope of the adapted parameters. Proposition A.2 rigorously shows that this is not equivalent to independent per node dropout. A global mask preserves the cross row covariance terms, whereas independent row masks remove the interactions between different rows.
Critical weaknesses
  1. According to the results in Table 1, HardLT does not achieve a significant improvement in micro F1 over the baselines.

  2. The writing quality is poor. The entire paper reads in a very stiff and awkward manner. The research motivation together with the problem introduction are not clearly conveyed. The related work section is far too brief. For example, the sentence "WLKS and ISNN develop further structural representations" does not explain how these methods work or why they are able to develop further structural representations. The section also fails to discuss other works on pretraining adaptation. In addition, the comparison table that appears in the related work section has no table caption. The figures need more annotations to be self explanatory. Figure 1 contains a large number of formulas without any explanation of what each formula does. The text inside Figure 1 should also use a uniform font size. The problem formulation in Section 3 should be a concise, symbol based formalization.

Questions for authors
  1. Why not use node representations generated by pretraining on node classification or other tasks for subgraph prediction? What advantage do node representations derived from link prediction specifically offer? Are there any relevant applications for subgraph classification models pre-trained via link prediction?
  2. Why should the partial view discard the PCA residual entirely? If downstream adaptation produces new directions outside the initial PCA subspace, does existing method hinder that adaptation?
  3. Does the method generalize to subgraph encoders other than GPEN?
Reviewer_axV9OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper focuses on the problem of subgraph classification. The authors that fitting a model on the complete set of nodes only is limiting, as there might be conflicting information. Instead, the authors propose to also consider training on partial views. The partial views are created by projecting the nodes via PCA and masking a portion of the dimensions before computing the loss. The full loss includes the full view, partials views, in addition to a consistency loss. The authors also propose to pre-train via link prediction. The authors detail their results on 4 datasets.

Critical strengths
  1. Subgraph-level tasks are an interesting problem that have seen far less attention that the other main graph tasks (node, link, graph). So it's nice to see more work on this topic.
Critical weaknesses
  1. The paper is very poorly written. Many things are not explained. I give just a few examples below:
  • Complete-view (line 45) is not defined. I had to infer it from the rest of the paper
  • The use of the word "Table" is odd and doesn't appear in prior literature. Why not just use "node embeddings" or something like that?
  • The authors define an "ambient graph" and "table" without defining what is meant by ambient.
  • The paragraph between lines 175-180 is incredibly difficult to understand. The authors mention a Jacobian but don't provide any formal explanation or intuition in the main text. Instead, they simple point to the appendix which contains points of proofs. As such, it's unclear to me what the authors are trying to communicate.
  1. The motivation to me is unclear. Specifically, I don't understand what the problem with complete-view fitting is. I read the explanation on lines 166-180 and am still confused. The authors briefly mention a theoretical explanation on lines 175-180 but don't properly explain it in the main text.

  2. It's unclear to me why the authors use PCA at all. Why not just do the masking directly on E0E_0? It's not explained clearly.

  3. The ablation (Table 2) seems to indicate that none of the components are actually necessary. This is because the ablated results are always within variance of the full model.

  4. The final results (Table 1) aren't very encouraging. There's no statistically significant difference between the proposed method and ISNN and GPEN.

Questions for authors
  1. Can you better explain the motivation of the method (see weakness 2)?
  2. Can you explain why PCA is used (weakness 3)?
Reviewer_vUt6OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The author's propose HardLT which adapts LP node tables on subgraph views for the complete table, providing a difficult partial view which they optimize with minibatch loss selection. The problem formulation treats this as regularization on tunable node-table which translates to an implementation combining PCA-projection and a globally-consistent mask to produce views for ranking hardness. This is then encoded into adaptive steps that incurs no additional computational overhead during inference to boost GPEN Encoder performance for PPI-BP, HPO-Metab and EM-User.

Critical strengths

The HardLT framework is clean and relatively-simple with direct benefits versus full-scale adaptation for the LP task. The hyperparameter analysis is particularly promising, clearly indicating HardLT retaining direction at evaluation versus full-scale adaptation. This shows a clear gain between GPEN and HardLT within Table 1, further component comparisons, component analysis, and hyperparameter analysis indicate promise for the HardLT's proposed target.

Critical weaknesses

My critical concern is around the problem formulation for HardLT and it's subsequent practical utility. Given that HardLT is proposed as a framework that functions on top of existing subgraph encoders, it then appears limiting to show that HardLT works only for the GPEN encoder. Although HardLT is distinct in it's application of PCA and adaptation, there is significant literature on downstream adaptation for the link prediction task which precedes HardLT and largely follows the same framework principles [1] . Although direct testing is not entirely necessary, this raises concerns about the novelty of HardLT and requires positioning on what makes HardLT's contribution unique to existing adaptation methods.

[1] Song, Yu, et al. "A Scalable Pretraining Framework for Link Prediction with Efficient Adaptation." Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. 2025.

Questions for authors
  1. Would it be feasible to test HardLT with other encoder mechanisms? In particular, does HardLT provide a benefit to improve other strong subgraph encoders, such as ISNN and WLKS?

  2. How is HardLT distinct from existing adaptation methods, such as [1] from the critical weakness? Currently, it appears that HardLT follows the same abstract workflow as existing methods while combining PCA Reconstruction. It appears that the masking strategy could indicate new utility but the current manuscript leaves me uncertain.

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

Related papers

Discussion (0)

Sign in to comment.