acceptodds
Withdrawn from ICLR 2027

GraphBridge: Graph-Mediated Transfer under Structural Modality Missingness

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Machine learning systems are often trained in domains with rich information but deployed in domains where an entire modality is unavailable and raw feature spaces do not align. External knowledge graphs (KGs) can provide a semantic bridge across such domains, but they can be noisy, incompletely aligned to data, and may contain relations that are useful in one population but harmful in another. We propose GraphBridge, a framework for KG-mediated transfer under structural modality missingness. GraphBridge anchors raw variables from each domain to graph node families with abstention, learns a source-target-specific transferable graph by filtering relations according to scores from literature support, source and target utility, and cross-domain stability, and constructs a low-dimensional spectral bridge for source-to-target signal transport. It then learns the transformation in the source domain from this bridge to the source-only modality and applies this transformation in the target domain. We provide guarantees for anchor and graph recovery, spectral bridge perturbation, transported-feature error, target excess risk, and a validation-protected model selection. On synthetic benchmarks, GraphBridge consistently improves over target-only prediction across various settings. On a real benchmark of PLATO pharmacogenomic transfer with BioKG, GraphBridge achieves strong improvement, reaching Pearson on GDSC cell-line transfer and nearly matching the oracle on a PDX transfer pair with compared to oracle . These results show that \method can support transfer via noisy external knowledge graphs across structurally mismatched domains.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
knowledge graph, transfer learning, missing modality

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_tuSpOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies structural modality missingness. Here the source domain has an auxiliary modality that the target domain lacks. The raw features of the two domains may be disjoint, and a noisy external knowledge graph is the only link between them. GraphBridge anchors the variables of each domain to node families of the graph and keeps the relations that are supported by the literature and useful in both domains. From this transferable graph it builds a low-dimensional spectral bridge. It then learns on the source a transport map from the bridge to a latent factor of the missing modality and applies it to the target. The paper gives finite-sample results for a simplified version and evaluates the method on synthetic data and the PLATO drug-response benchmark with BioKG.

Critical strengths
  1. The paper formulates a practical setting that, existing missing-modality and domain-adaptation methods typically do not cover. The method gives each difficulty its own component. An abstention option handles variables with weak or ambiguous evidence, and a validation safeguard chooses between the transfer predictor and a target-only baseline to avoid negative transfer.
Critical weaknesses
  1. Some GDSC results are hard to reconcile with the method as defined. By the paper's equations, the target bridge is a linear map of fewer than a hundred target genes. With LightGBM, however, the bridge arm beats the oracle arm, which uses the true source-only factor, on most pairs (Tables 7 and 8). An explanation would help.

  2. The two main steps, learning the transferable graph and transporting the source-only modality, are not shown to help. On GDSC with LightGBM, the bridge arm beats the transfer arm on every pair. App. H.6 also shows that in the analyzed model, adding the transported factor does not change the least-squares fit. Results for the naive-KG variant listed in App. E.7 would show what the learned graph adds.

Questions for authors
  1. What is a row in the PLATO cohorts, and could the masked drug columns reach the target bridge through anchoring, aggregation or standardization? Do the results hold when those columns are removed before running GraphBridge?

  2. What do the naive-KG variant, a random graph and the validation-selected predictor achieve on the PLATO pairs?

  3. How are the "C. Br." and oracle in Tables 7 to 10 built? With LightGBM, "C. Br." is the best non-oracle arm on every GDSC pair. The LightGBM oracle equals the raw arm on the four pairs with CH as the target.

  4. On the GDSC pairs in Table 2, KGAT's raw reaches 0.87 to 0.90, close to the Ridge and PLATO oracles. What inputs do the KGAT and PLATO receive?

  5. How were the default hyperparameters in App. L chosen? For example, were they tuned on the evaluated pairs?

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

This paper proposes GraphBridge for supervised transfer when the target domain lacks an auxiliary modality available in the source domain. It uses an external knowledge graph to connect different input feature spaces, constructs a spectral bridge representation, and learns a source-side mapping to estimate the missing target modality. The method is evaluated through synthetic experiments and drug-response prediction across GDSC and PDX cohorts.

Critical strengths
  1. Transferring information from a better-measured source domain to a target domain with missing measurements is a relevant problem. The framework also considers differences in input features and uncertainty in the external knowledge graph.

  2. The synthetic experiments examine several transfer conditions, including graph noise and negative transfer. The validation-based safeguard provides a practical fallback, and the real-data experiments cover multiple source–target cohort pairs.

Critical weaknesses
  1. I am not convinced that the real-data setup reflects a realistic missing-modality scenario. In the real experiments, the definitions of the observed features X, missing modality Z, and prediction target Y need a clearer explanation. The fixed 500-feature construction and masking of drug-related variables are not convincingly connected to a deployment scenario in which these variables would actually be unavailable. This makes it difficult to judge whether the experiments demonstrate useful modality transfer beyond the constructed benchmark setting.

  2. The comparisons mainly test different predictors, with few competing modality-transfer methods. KGAT already performs strongly without the proposed transfer, and some bridge-only results outperform the full transfer arm. I therefore find it difficult to attribute the gains to modality transport itself. The real-data evaluation is also limited to one pharmacogenomic benchmark collection and one KG.

  3. The connection between the theory and the evaluated method is incomplete. The analysis uses hard anchoring and gating, simplified spectral construction, and linear estimation, while the implemented pipeline uses soft assignments, different graph scores, normalization, and regularization. The paper acknowledges these simplifications, but does not clearly establish which guarantees carry over.

  4. The dependence on BioKG and the computational cost need more evaluation. Appendix reports about 36 hours per cohort, but does not show how runtime and memory scale. The exact KG version and feature-to-entity mapping also need clearer documentation.

Questions for authors
  1. What exactly do X, Z, and Y represent in the real experiments, and why would Z be unavailable at deployment?

  2. Does modality transport consistently improve over bridge-only prediction and strong transfer baselines with the same information access?

Reviewer_qNGHOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes to perform transfer learning under structural modality missingness where an additional modality is available in the source domain but is absent in the target domain and the observed feature spaces also do not overlap. For this, the authors' main proposal is a multi-stage process using an external knowledge graph (KG) to bridge the two domains that works as follows: 1/ KG node families are used to anchor the raw features from source and target domains. The KG is filtered based on literature support, source/target applicability and cross-domain stability. 2/ The filtered KG is used to create a spectral representation and train a mapping from this representation to source-only modality. 3/ The mapping is then applied to predict the missing modality in the target domain. This and the spectral mapping of the target input is then used for final target prediction.

The authors also present theoretical formulations for spectral stability and modality transport since their main contributions are around 1/ graph filtering and 2/ missing modality generation. In my view, neither of these contributions hold well in the paper. Graph filtering via abstention and literature support for (source, target) pair is not rigorously compared to a static unfiltered graph usage baseline or even a continuously weighted graph without heuristic pruning. Furthermore, in experiments (and possibly theoretical formulation too) the value of the missed modality prediction is minimal since simpler baselines like Bridge give similar performance and are formulated similarly to infer without the missing modality.

Critical strengths
  1. The authors make a good effort to address the challenges of aligning source and target features to an external KG via the effort in designing Stage 1. The combination of anchoring, abstention for uncertain mappings, literature support and (source, target) utility provides a comprehensive approach to making a (source, target) pair specific graph. They make the explicit assumption here that KG relations are not all equally useful across domains.
  2. Over the Raw baseline (using target features only) GraphBridge improves significantly on PLATO evaluations.

Finally, the paper also presents a theoretical analysis section which analyzes the individual contributions of filtering and transfer for the missing modality.

Critical weaknesses

Both the key contributions of the paper have substantive weakness in their claims:

  1. The necessity of the multi-pronged graph-filtering process is unclear. In the paper, Stage 1 combines literature support, semantic anchoring with abstention and cross-domain discrepancy. While these components are well motivated, it is not clear how much each one contributes to the final improvements. It would be useful to see ablations for incremental benefits over fixed KG, anchoring with and without abstention, literature based support. A comparison of a filtered KG with continuously weighted KG could also be very useful to conclusively explore whether Stage 1 has to be designed exactly this way.
  2. Next, the missing modality reconstruction process is not backed by experiments. While GraphBridge wins over the Raw baselines easily, Raw is a weak baseline that uses just target features without KG enrichment. This only proves the already known benefits of KG derived transfer from source to target. Per the results in Appendix L, Bridge, without the missing modality reconstruction, frequently wins or is at par with Transfer of missing modality (Table 7-10). This shows that the largest improvements may just be coming from adding KG to the weak Raw baseline without KG rather than from modality reconstruction.

Critically, the experiments in Table 7-10 should have been summarized but still kept in the main paper to avoid Appendix hunting for critical components.

Questions for authors
  1. What is the incremental value of the missing modality reconstruction in target domain? Doesn't the theorem 4.2 also imply that the Bridge setting (without modality reconstruction) already captures the same linear function as the transfer?
  2. If feasible, please add the ablation for the various Stage 1 heuristics for graph filtering.

Suggestion:

  1. In my view, the missing modality reconstruction would be better motivated as a mechanism for removing the KG information at final inference time. even if not in predictive performance, this can have inference efficiency advantages. I recommend exploring this path for modality reconstruction value.

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

Related papers

Discussion (0)

Sign in to comment.