acceptodds
Under review as a conference paper at ICLR 2027

REBASE: Relational Basis Learning for Heterogeneous Multi-Source Tabular Data

Abstract

Multi-source learning for tabular data is fundamentally challenged by feature heterogeneity across distributed sources, arising from both structural heterogeneity (differences in feature dimensionality and composition) and semantic inconsistency (variation in feature definitions, naming, and units). This challenge is particularly pronounced in healthcare, where clinical variables are collected under institution-specific protocols. While such surface-level discrepancies exist, the underlying disease mechanisms are largely shared across populations, manifesting through complex interactions among clinical features. To address this challenge, we propose REBASE, a framework that learns reusable relational bases that can be shared across heterogeneous tabular sources. REBASE represents each sample as a graph, where nodes and edges are grounded in large language model embeddings, mapping heterogeneous features into a unified semantic space. Over this space, our model learns a set of relational bases, each capturing a distinct relational pattern. It aligns each sample graph to these bases using Fused Gromov-Wasserstein distance, which jointly accounts for semantic similarity and structural correspondence. This alignment produces sample-specific coordinates that compose predictions from the bases. Experiments on three real-world benchmarks demonstrate that REBASE consistently outperforms existing multi-source tabular pretraining methods, in both robust generalization and few-shot adaptation to unseen, data-scarce domains, with its learned relational bases exhibiting clinically meaningful patterns across domains. Code is available at https://anonymous.4open.science/r/REBASE-E3B2/

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Healthcare AI, Heterogeneous Multi-Source tabular learning, few-shot adaptation.

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_GvbvOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes REBASE, a framework for learning from heterogeneous tabular datasets with different feature sets and semantic descriptions. It represents each sample as a graph using language-model embeddings and a shared graph attention encoder, then matches the sample graph to learned prototype graphs using Fused Gromov–Wasserstein distance. These distances determine sample-specific weights for combining the prototypes’ predictions. The shared prototypes support zero-shot prediction and few-shot adaptation to unseen domains. Experiments on cardiovascular, ICU, and breast cancer benchmarks show generally competitive or improved performance over existing tabular methods, particularly when target-domain labels are scarce.

Critical strengths

The method brings together components that fit the problem well. Semantic embeddings help compare feature descriptions, graphs handle different numbers of features, and FGW considers both feature representations and their relationships.

The model can make predictions on a new domain without target labels or training a new input layer. This is useful for hospitals with limited labeled data.

The experiments cover three benchmarks with 16 domains, evaluating both learning across multiple sources and adaptation to unseen domains.

Critical weaknesses

My main concern is whether prototype matching keeps enough patient information for prediction. Two patients may differ in an important clinical feature but receive the same prototype weights, leading to the same prediction.

Also, the same clinical pattern may have different risks across hospitals because of differences in treatment, patient populations, or outcome definitions. It is unclear how well shared prototype predictions handle these differences.

The main tables compare full models, so it is hard to tell how much semantic embeddings, graph encoding, prototypes, and FGW matching each contribute to the gains.

Tables 2–3 report averages without standard deviations or confidence intervals. Although the experiments are repeated, these tables alone do not show whether the smaller improvements are reliable.

Questions for authors

See Weaknesses

5 comments
AuthorsOpenReview ↗

Response to Reviewer Gvbv (1/3)

Thank you for your careful and constructive review. We especially appreciate your observation that predicting on a new domain without target labels or a new input layer is useful for hospitals with limited labeled data, which is the setting REBASE was designed for. Results from experiments newly conducted for this rebuttal are marked [new].

W1. We agree that this is a central question for prototype-based prediction. Our analyses indicate that the coordinates change with the clinical variables characterizing each basis. Patients who differ in such a variable therefore receive different coordinates, and since each prediction is the coordinate-weighted combination of the bases, they also receive different predictions. We examine this at the level of each basis (1) and of individual patients (2), and overall performance is consistent with it (3).

1. The variables characterizing each basis are clinically important. For each basis, we compare the samples assigned to it by their largest coordinate (high-risk) with the rest. The variables that separate the two groups are those used in established organ dysfunction and acute physiology scores; for instance, urea separates them for ϕ2\phi_2 and prothrombin time for ϕ4\phi_4 (Figure 3; Appendix D.2).

2. Among patients who are similar in all variables but the key variable, the coordinate changes with it [new]. To show this, we newly tested whether patients who are similar in all other variables but differ in one important variable receive different coordinates, in three steps.

  • Step 1. For each basis, the key variable is chosen as the one that most strongly separates its high-risk samples from the rest, as shown above.
  • Step 2. For each reference patient, we take the 200 patients closest to it in Gower distance [1] over all variables except the key variable and select every tenth of them along the key variable, giving a group of 20. Within these groups, the other variables vary only about half as much as in groups of 20 randomly drawn patients, while the key variable varies as widely as in the full cohort.
  • Step 3. Within each group, we compute the Spearman correlation ρ\rho of the key variable with the corresponding coordinate and with the predicted risk, and report the median across groups. The table below shows the results on the ICU benchmark.
BasisSourceKey variableρ\rho (variable vs. coordinate)ρ\rho (variable vs. predicted risk)
ϕ2\phi_2 (Cardiorenal Hypoperfusion)MIMIC-IVBUN0.580.43
HiRIDurea0.520.41
ϕ4\phi_4 (Hemostatic Consumption)MIMIC-IVPT (seconds)0.460.35
SICdbINR0.450.35

In every case, the coordinate and the predicted risk rise with the key variable, whereas an unrelated variable, such as body temperature, shows almost no correlation with the coordinate (0.03–0.06). Patients assigned to the same basis therefore do not share one coordinate; it varies with the key variable, and so does the predicted risk. We have added this analysis to Appendix B.12 of the revised manuscript.

3. Overall performance. Expressing patients from all sources through the same prototype graphs does not cost predictive performance: REBASE outperforms both Single (GAT), which trains the same encoder on each source alone, and Multi (GAT), which trains it on all sources without the prototypes, at every sampling ratio on the cardiovascular benchmark (Appendix B.2).


[1] Gower JC. A general coefficient of similarity and some of its properties. Biometrics. 1971;27(4):857–871.

AuthorsOpenReview ↗

Response to Reviewer Gvbv (2/3)

W2. We agree that the same clinical pattern may carry different risks across hospitals. For each patient, REBASE computes coordinates that reflect both the hospital’s textual metadata (the name, description, data type, and unit of each variable) and the patient’s values, and uses them to combine the shared prototype predictions (ϕk\phi_k) into that patient’s risk. Differences between hospitals are therefore reflected in the coordinates, whereas the ϕk\phi_k capture patterns that appear consistently across hospitals even when the hospitals record them differently.

First, REBASE represents each sample as a graph whose nodes are the features recorded by its source and whose edge attributes are derived from the features’ descriptors. Since the coordinates are computed from the FGW distances between this graph and each prototype graph, they reflect not only the patient’s values measured at each hospital but also how the hospital defines and records its variables. Figure 2 is consistent with this: for each source, we compute how strongly each basis is associated with positive outcomes (its coordinate-weighted positive rate) and compare these profiles across sources by cosine similarity. The profiles differ across sources; for instance, SUPPORT, which enrolled patients at advanced stages of one of nine illnesses rather than at intensive-care admission, shows the lowest similarity to every other source (at most 0.14; Section 5.2), likely reflecting its distinct patient population.

Second, within each basis, variables that record the same clinical quantity may differ across sources in name or definition, yet move together toward high-risk (Figure 3). For instance, in ϕ2\phi_2, the basis that appears to represent cardiorenal hypoperfusion, BUN in MIMIC-IV and urea in HiRID are defined differently, yet both are higher in the high-risk samples.

Thus, the bases shared across hospitals capture patterns that hold across sources, while differences between hospitals appear in the coordinates.

AuthorsOpenReview ↗

Response to Reviewer Gvbv (3/3)

W3. We have additionally conducted the suggested ablations.

Semantic embeddings [new]. Semantic embeddings place variables recorded under different names in a shared space where they can be compared. Since REBASE distinguishes variables only by these embeddings, without them two variables could correspond only if their names were identical. However, because each source follows its own naming conventions for the same underlying concept, most variables would then lose their counterparts in other sources: on the ICU benchmark, only 97 (24%) of the 398 cross-source pairs that record the same clinical feature share an identical name. This would break REBASE’s premise of learning jointly from sources with different schemas, so an ablation that removes only the semantic embeddings is not possible. Instead, we newly ran an ablation that progressively enriches the textual metadata in the descriptor given to the language model in three levels, thereby varying how much information about each variable the embeddings carry:

  • L0: the column name alone
  • L1: the column name with its clinical name
  • L2: the full descriptor used by REBASE, with name, description, data type, and unit

As the metadata becomes richer, variables recorded under different names are matched more accurately (see our response to Reviewer r1Wt, Q2) and zero-shot AUROC on the ICU benchmark improves on average, with L0 the lowest in every source.

Zero-shot AUROC (ICU)MIMIC-IVeICUHiRIDSUPPORTSICdbZigongMean
REBASE + L071.0±2.6965.3±3.5173.9±4.1363.5±3.4174.9±3.0259.9±3.3568.1
REBASE + L172.5±4.5167.9±3.2480.5±5.6565.0±5.9976.3±4.9960.6±5.9170.5
REBASE + L273.6±5.3868.3±4.4580.4±5.9865.6±6.4276.7±5.1460.8±6.3970.9

We have added these results to Appendix B.10 of the revised manuscript.

Graph encoding [new]. As with the semantic embeddings, an ablation that removes only graph encoding is not possible. Each sample is represented as an instance graph so that a single encoder can process sources with different variables; these graphs are aligned with the prototype graphs via FGW, and the prototype graphs are initialized from the encoder’s [CLS] embeddings, so removing graph encoding would also remove the FGW alignment and the prototype graphs. Meanwhile, self-attention over variables effectively operates on a fully connected graph, so replacing the encoder with self-attention would not remove the graph either. What graph encoding adds over self-attention is the edge attribute ej,j′\mathbf{e}_{j,j'}, which encodes the semantic relation between two variables, so we newly ablated only these attributes. As shown in the table below, removing them lowers the mean zero-shot AUROC on the ICU benchmark, so graph encoding not only underlies the FGW alignment and the prototype graphs but also improves zero-shot transfer through the semantic relations it encodes.

VariantEdge attributesZero-shot AUROC (ICU mean)
REBASE✓70.9
w/o edge attributes✗67.5

We have added this ablation to Appendix B.6 of the revised manuscript.

Prototypes. Figures 5–6 (Appendix B.2) compare GAT trained on a single source, GAT trained on multiple sources without prototypes, and REBASE, showing that basis decomposition through prototypes improves performance. Basis exclusion (Appendix B.8) shows that removing a single basis lowers performance mainly in the sources that rely on it, indicating that each basis plays a distinct predictive role.

FGW matching. Figure 10 (Appendix B.6) shows that zero-shot AUROC on HD peaks at 78.0 with FGW (β=0.5\beta = 0.5), compared with 64.7 when aligning on structure alone (GW, β=1\beta = 1). Tables 12–13 (Appendix B.7) show that replacing FGW with pooled [CLS] cosine or cross-attention alignment lowers zero-shot AUROC, averaged over the cardiovascular targets, from 76.8 to 61.2 and 55.9, and 64-shot AUROC from 89.7 to 81.1 and 80.0, at a comparable per-epoch cost (16.08 s vs. 13.94 s and 14.22 s).

W4. Tables 2–3 summarize few-shot results averaged over all target domains, so we do not report a single standard deviation there. Per-domain results with standard deviations are reported in Appendix E.4 (Cardiovascular), E.5 (ICU), and E.6 (Breast Cancer), and the statistical significance of the rank differences across methods is assessed with critical difference diagrams (Nemenyi test, α = 0.05) in Appendix E.2.

We hope this addresses your concerns.

Reviewer_VpUTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

REBASE looks at learning/transfer across clinical tabular sources where feature schemas differ. each sample becomes a graph w/ attributes grounded in LM embeddings, a shared graph encoder gives feature reps + relational structure, and then the graph is matched to a set of learned prototype graphs via FGW distances. those distances give sample-specific coordinates which combine the prototype-derived representations for prediction. training also has source specific auxiliary losses and a coordinate regularizer.

Eval is on 3 clinical benchmarks (16 cohorts), multi-source integration plus leave-one-domain-out zero/few-shot. there's also kidney function regression, ablations, compute cost and some clinically interpretable patterns of the learned bases. Results look promising and mostly favor REBASE over the multi-source baselines. but the probability-valued basis decomposition in the paper vs the latent-mixture predictor in the released code needs to be sorted out.

Critical strengths
  1. Solid empirical eval on a practically relevant problem Combining clinical datasets with different feature sets/inconsistent descriptions is a real obstacle, so the problem is well motivated. 16 cohorts over 3 benchmarks, and the paper separates integrating seen sources from transfering to a fully held out one, which is nice. In Table 1 REBASE gets best mean AUROC on 14/16 cohorts. LODO results are encouraging too esp. at low label budgets. repeated support-set sampling, multiple seeds and source-validation based HP selection all help. Overall decent evidence the approach is useful under heterogeneous schemas.
  2. Component analyses are actually informative Appendices go beyond aggregate numbers. single vs multi source GAT comparison helps separate the prototype module effect from just having more data. W-only / GW-only / alt. coordinate experiments support using both attributes and structure for matching. basis removal adds some functional evidence on top of the clinical visualizations, and runtime/memory numbers make the tradeoff clear. I appreciated these.
Critical weaknesses
  1. Prediction formulation / adaptation in paper doesn't quite match the released code Eq (1) and (4) define the predictor as a convex combination of class-probability basis outputs (each basis on the simplex). But in the anon code linked in App. C.1, models/TabularFLM_S_.py, predict() mixes prototype-derived latent vectors and then applies ghead (LayerNorm + nonlinear MLP). So thats a nonlinear classifier on a latent mixture, not a convex mix of fixed class prob vectors — in general these aren't equivalent.

Also set_freeze_target() updates the instance encoder and the existing global head while freezing prototype graphs and prototype GNNs, and Sec 4.4 doesn't really describe this. Since the basis decomposition is central to the framing and interpretation, I think equations, code and exact configs behind the reported numbers should be reconciled. Probably fixable, but right now its hard for me to tell which model the experiments are validating.

  1. Increment over related FGW-template classifiers not isolated enough Vincent-Cuaz et al. NeurIPS 2022 (“Template based Graph Neural Network with Optimal Transport Distances”) already learns graph templates and classifies via FGW distances to them. this connection isn't discussed.

The heterogeneous tabular representation + multi-source training could well be valuable contributions. However the current GAT-only and cosine/cross-attn ablations don't isolate the benefit of the specific basis readout/training design vs a classifier on the same FGW-template representation — and this matters more given the latent-mixture implementation above. A controlled comparison (same graph construction, data, encoder, model selection) would help clarify the contribution a lot.

Ref: Vincent-Cuaz et al. (2022), Template based Graph Neural Network with Optimal Transport Distances

Questions for authors
  1. Which predictor produced the reported tables, the convex mixture in Eq (1)/(4) or the nonlinear head after latent mixing in the code? Could you give the exact forward equation, dims/meaning of each basis output, and all params+losses used in target adaptation? also which code version/configs. This affects my soundness score and how I read the basis interpretation.

  2. How does REBASE differ from a TFGW style classifier on the same semantic graphs and FGW-template rep? Is a controlled comparison possible (same data, encoder, selection protocol, comparable readout capacity)? the existing ablations show FGW matching helps but don't fully answer this.

  3. FGW stability: GraphQuantizer.compute_fgw() reuses the previous call's transport plans when shapes match, uses 5 outer iters, and the plan cache gets updated during eval too. Are predictions/metrics stable if you reset plans, change eval order or batch size, or use more solver iterations? Also what feature cost + solver config was used for main results? README examples and App. A seem to describe different feature costs.

  4. Preprocessing: what normalization/imputation is used, and which samples are used to fit the stats in zero-shot and few-shot? The embedding entry point calls maker.fit(X, y) before the downstream split, and the embedding implementation itself isn't released yet. Does fitting only use the right training partition, and is an unlabeled target pool allowed? Releasing that preprocessing code would clear this up.

3 comments
AuthorsOpenReview ↗

Response to Reviewer VpUT (1/2)

Thank you for your careful review and for checking our code in detail. We especially appreciate your recognition that combining clinical datasets with different feature sets and inconsistent descriptions is a real obstacle, and that our component analyses are informative. Results from experiments newly conducted for this rebuttal are marked [new].

Q1. The repository now contains the version behind all reported results. We thank the reviewer for pointing out the mismatch between the code and the paper. We inadvertently uploaded our ablation version to the anonymous repository; it attaches an MLP head (ghead) to a latent mixture, and its README describes a cosine feature cost. We have replaced both with the version that produced all reported results, which predicts by the convex combination of Eqs. (1) and (4), a coordinate-weighted average of the bases’ class probabilities, and uses the l2 feature cost of Appendix A. The MLP-head variant gives lower AUROCs than Table 1 (see Q2).

Target adaptation (set_freeze_target()). The relational bases (latent_graph, gnn_experts, basis_out) are kept fixed, and a target head (target_head) is attached to the convex-combination prediction. The instance encoder (basis_layers, basis_layer_norms) is updated, since the target coordinates are computed from graphs it produces; with a fixed encoder, they could not adapt. Only the loss on the convex-combination prediction (global_loss) is used; local_loss and reg_loss are not. We have added this description to Appendix C.1 of the revised manuscript.

Q2. REBASE differs from TFGW [1] in its readout, which performs better under identical settings. TFGW feeds the FGW distances to learned templates into an MLP, so each template’s contribution to a prediction cannot be identified. REBASE computes a class probability from each template with a basis GNN and combines them by the coordinates, so each prediction decomposes into basis contributions. Under the same graphs, encoder, and templates, we newly compared FGW distances fed into an MLP (TFGW-style), the latent mixture with an MLP head (Q1), and REBASE.

Readout [new]MIMIC-IVeICUHiRIDSUPPORTSICdbZigongZero-shot (mean)
TFGW-style86.1±0.382.9±0.486.9±0.587.9±4.288.0±0.581.9±1.965.9
MLP head89.9±0.485.8±0.389.9±0.390.3±1.489.3±0.582.5±2.068.3
REBASE91.1±0.386.0±0.290.7±0.890.8±1.289.4±0.783.7±2.770.9

REBASE performs the best in every source and in zero-shot. We have added this comparison and a discussion of TFGW to Appendix B.11 of the revised manuscript.


[1] Vincent-Cuaz C, Flamary R, Corneli M, Vayer T, Courty N. Template based Graph Neural Network with Optimal Transport Distances. Advances in Neural Information Processing Systems 35 (NeurIPS), 2022.

AuthorsOpenReview ↗

Response to Reviewer VpUT (2/2)

Q3. The warm start has a negligible effect. Warm-starting each solve from the previous transport plan (T_init) only speeds up computation. We newly reran REBASE, changing one condition at a time, and AUROC changed by less than the seed standard deviation in every case.

ΔAUROC (points) [new]MIMIC-IVeICUHiRID
No warm start (T_init=None)−0.03+0.02−0.05
Shuffled evaluation order+0.01−0.01+0.02
Batch size 32+0.05+0.03−0.06
50 outer iterations+0.01+0.01+0.02

Feature cost and solver. The main results use the l2 feature cost of Appendix A (feat_distance), β∈[0.5,0.7]\beta \in [0.5, 0.7] (alpha in POT) selected on the source validation split, entropic regularization ϵ=10−2\epsilon = 10^{-2} (entropy_reg), and 5 outer and 20 inner (Sinkhorn) iterations.

Q4. No preprocessing statistic is computed from test rows in any setting. Missing values are imputed with the mean (numerical) or the mode (categorical); within each source, numerical features are then standardized and Yeo–Johnson transformed. y is used only for the stratified split.

  • Source domains: Statistics are computed from the non-test rows (training and validation) and applied to the test rows. maker.fit(X, y) reproduces the loader’s test split.
  • Target domains (zero-shot and few-shot): Zero-shot means that no target label is used, not that no target data is available. The statistics need only feature values, so they are computed from the target’s unlabeled non-test rows (the 80% outside the fixed test split) and applied to the test rows. This is allowed because a new hospital holds its patients’ records before outcomes are labeled, and neither target labels nor test rows enter the statistics. Few-shot reuses them across all 25 support-set resamplings.

The lines below select the rows used for fitting; the fitted transform is then applied without refitting. The full code will be released upon acceptance.

# Table2EmbeddingSlimTransformer.fit(X, y):
train_idx, test_idx = train_test_split(
    list(range(len(X))), test_size=0.2,
    stratify=list(y), random_state=self.args.random_seed)
X_tr = X.iloc[train_idx]
self.num_fill = X_tr[self.num_col_names].mean()
self.cat_fill = X_tr[self.cat_col_names].mode().iloc[0]
self.num_transformer = make_pipeline(
    StandardScaler(), PowerTransformer(standardize=False)).set_output(transform="pandas")
self.num_transformer.fit(X_tr[self.num_col_names].fillna(self.num_fill))

# data loader (released)
train_idx, test_idx = train_test_split(
    indices, test_size=0.2, stratify=labels, random_state=args.random_seed)
AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer VpUT,

Thank you again for your careful review and for examining our code in detail. As the discussion period ends soon, we would like to kindly ask whether our response has successfully addressed your concerns. If you have any remaining concerns, we would be willing to address them.

Best regards, Submission 23832 Authors

Reviewer_r1WtOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes REBASE, a framework for learning transferable relational representations from heterogeneous tabular datasets. It represents each sample as a feature graph using language model embeddings and learns shared relational bases through Fused Gromov–Wasserstein alignment. These bases support prediction across different feature schemas without requiring direct feature correspondence. Experiments on cardiovascular, ICU, and breast cancer datasets demonstrate competitive performance in multi-source learning, zero-shot generalization, and few-shot adaptation.

Critical strengths
  1. The paper addresses an important practical challenge: transferring knowledge across heterogeneous tabular datasets with inconsistent feature definitions and structures.
  2. The experimental evaluation is relatively comprehensive, covering three clinical benchmarks, multiple baselines, zero-shot and few-shot settings, and interpretability analyses.
Critical weaknesses
  1. REBASE assumes that transferable relational patterns can be learned across heterogeneous sources. However, the cardiovascular benchmark combines datasets targeting different clinical outcomes, including CAD, CVD, CHD, and heart failure. Although these outcomes are related, they differ in clinical definitions, underlying mechanisms, patient populations, and associated risk factors. Consequently, improved cross-source performance does not necessarily demonstrate that the learned relational bases capture stable mechanisms underlying the same clinical outcome. An additional experiment separating sources with consistent clinical endpoints from those with heterogeneous endpoints would help clarify how outcome heterogeneity affects transfer performance.

  2. REBASE relies on language model embeddings to place heterogeneous clinical features into a shared semantic space. However, the experiments do not sufficiently isolate whether this semantic grounding remains reliable under realistic schema inconsistencies. Clinical features may have different names but equivalent meanings, similar names but different definitions, or inconsistent units and reference ranges. Semantic similarity in an embedding space does not necessarily imply clinical equivalence. Systematic evaluations involving feature-name paraphrasing, abbreviations, unit perturbations, and semantic mismatches, potentially across different embedding models, would strengthen the evidence that REBASE effectively addresses semantic inconsistency.

  3. REBASE treats the attention-derived affinity matrix from the GAT encoder as the relational structure of each sample and aligns it with learned prototype graphs using FGW. However, since these affinities are learned through the prediction objective, they do not necessarily represent meaningful feature interactions or clinical relationships. More importantly, standard self-attention can also model sample-dependent interactions among features without explicitly constructing and aligning prototype graphs. Although the paper provides component ablations, it remains unclear whether the FGW-based relational formulation offers benefits beyond a sufficiently expressive self-attention architecture under comparable model capacity and computational budgets. A controlled comparison against a matched self-attention baseline would better establish the necessity of the proposed relational alignment mechanism.

Questions for authors

See weakness.

6 comments
AuthorsOpenReview ↗

Response to Reviewer r1Wt (1/3)

Thank you for your careful and constructive review. We especially appreciate your recognition that transferring knowledge across tabular datasets with inconsistent feature definitions and structures is an important practical challenge, which is what motivated REBASE's reusable relational bases shared across domains. Results from experiments newly conducted for this rebuttal are marked [new].

[Q1. Shared mechanisms across different endpoints]

We agree that performance gains alone do not show that the bases capture stable mechanisms. However, as defined in Section 3, the sources in REBASE share a label space for a common clinical outcome, and what may differ across sources is the specific clinical endpoint that defines this label. The six cardiovascular sources likewise all predict the same label, whose positive class denotes a clinically adverse cardiovascular outcome, and CAD, CVD, CHD, and HF, raised by the reviewer, are all endpoints of this label (Appendix D.1).

REBASE aims to capture the mechanisms shared across sources as bases, while patterns that apply only to some sources, such as those that depend on the endpoint, are separated into other bases and do not interfere with this learning. To verify this, we newly ran the basis exclusion experiment on the cardiovascular benchmark.

Basis exclusion [new]. We removed each basis from the combination in Eq. (4) at inference and renormalized the remaining coordinates. Removing ϕ1\phi_1 (Ischemic Cascade) lowers the three CAD sources and ERB (CVD), whose endpoint differs from theirs, and removing ϕ2\phi_2 (Behavioral Risk Profile) lowers ERB (CVD) and SAH (CHD), which have different endpoints (see the table below). Sources with different endpoints thus share the same basis, and sources that do not share it are barely affected.

Source (endpoint)All basesRemove ϕ1\phi_1 (Δ)Remove ϕ2\phi_2 (Δ)
MED (CAD)97.5−1.4−0.2
CDD (CAD)98.7−1.20.0
HD (CAD)96.7−1.6−0.3
ERB (CVD)97.7−0.9−1.2
SAH (CHD)76.2−0.3−2.4
HF (mortality)92.2−0.3−0.5

We have included these results as a new figure in Appendix B.8 of the revised manuscript.

On the ICU benchmark, the endpoint likewise differs slightly across sources under the same label, mortality (in-hospital death vs. death at ICU discharge), and the same ablation already reported in the original manuscript shows a similar tendency (Appendix B.8, Figure 11).

Source exclusion (Appendix B.3). Excluding MED from pretraining lowers CDD and HD, which share ϕ1\phi_1 with it, by 2.7 and 3.3 points, and excluding SAH (CHD) lowers ERB (CVD), which shares ϕ2\phi_2 despite its different endpoint, by 2.0 points. In both cases, no other source drops, and even excluding HF, whose endpoint differs the most, changes the other sources by at most 0.7 (Figure 9).

Same vs. mixed endpoints [new]. As the reviewer suggested, we newly ran an experiment that separates the sources with the same endpoint from those with different endpoints. Under the same setup as Section 5.1, we pretrained REBASE only on the three CAD sources (MED, CDD, HD) and compared it with REBASE pretrained on all six sources (Table 1), which additionally include ERB (CVD), SAH (CHD), and HF (mortality). If the sources with other endpoints interfered with the bases learned for the CAD sources, adding them would lower the AUROC of the CAD sources. However, the AUROC of the CAD sources under six-source pretraining (97.5, 98.7, 96.7) remains within one standard deviation of that under CAD-only pretraining (97.3, 98.7, 96.5). This is likely because patterns present only in sources with other endpoints are separated into other bases, such as ϕ2\phi_2. We have added these results to Appendix B.9 of the revised manuscript.

We will state the scope of this explanation, namely that the endpoints differ within a common clinical outcome, more explicitly in Section 3. We will also correct the wording in Appendix D.1 ("relaxes the shared-label assumption"), which conflates the label with the endpoint and may have suggested that the sources predict different outcomes.

AuthorsOpenReview ↗

Response to Reviewer r1Wt (2/3)

[Q2. Semantic grounding across names, definitions, and units]

We agree that similar embeddings do not by themselves make two variables clinically equivalent. This is most likely when variables are embedded by their names alone. REBASE therefore semantically grounds each feature through its descriptor, the textual metadata of its name, description, data type, and unit, so variables recorded under abbreviations or other expressions are grounded by what they measure.

Descriptor ablation [new]. To test whether the full descriptor, rather than the name alone, aligns variables with their true counterparts, we newly ran an ablation on the descriptor. We collected 301 cross-source pairs of variables that record the same clinical feature under different names (excluding pairs defined differently, such as PT/INR) and varied the descriptor at three levels: the column name alone (L0, e.g., chem_bun_max), the column name with its clinical name (L1, e.g., "Blood urea nitrogen, maximum"), and REBASE's full descriptor (L2). The proportion of variables whose nearest embedding neighbor is their true counterpart rises from 79.8% (L0) to 93.2% (L1) and 94.5% (L2), with the same order under a general-purpose encoder, Qwen3-Embedding (74.1%, 89.5%, 91.3%), and the mean zero-shot AUROC on the ICU benchmark rises accordingly (68.1, 70.5, 70.9). Most of the gain in alignment comes from adding the clinical name (L0 → L1). We have added this analysis to Appendix B.10 of the revised manuscript.

Beyond the embeddings, variables that differ in name, definition, or unit move in the same direction within the bases (ICU benchmark; Figure 3; Appendix E.1). We compare the samples assigned to each basis by their largest coordinate (high-risk) with the rest.

  • Different names, equivalent meanings. Heart rate is recorded as vs_heart_rate_max, hrt, and heart_rate_max, yet is higher in the high-risk samples of Oxygen Delivery Failure (ϕ11\phi_{11}) in all three sources.
  • Different definitions. In Hemostatic Consumption (ϕ4\phi_4), coag_pt (seconds) and inr (a ratio) are higher in the high-risk samples, whereas SICdb's prothrombin_time, a percentage of normal activity, is lower; all three indicate slower clotting.
  • Different units and reference ranges. Kidney function is recorded as BUN in mg/dL in MIMIC-IV and as urea in mmol/L in HiRID, yet both are higher in the high-risk samples of Cardiorenal Hypoperfusion (ϕ2\phi_2), as within-source standardization removes linear unit conversions. Consciousness, recorded on the GCS and on the oppositely scored scoma, is impaired in the high-risk samples of Neurologic Depression (ϕ9\phi_9) in every source.

Text encoder [new]. We also newly evaluated REBASE with a general-purpose text encoder, Qwen3-Embedding, in place of the medical one (EmbeddingGemma-300M-medical). The per-source AUROC of Table 1 changes by at most 0.8 points, showing that REBASE works robustly even with a general-purpose encoder. We have added these results to Appendix B.6 of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer r1Wt (3/3)

[Q3. FGW alignment vs. self-attention]

FGW alignment compares the instance graph of every sample, including the relations among its features, with prototype graphs shared across the heterogeneous sources. Self-attention has no such shared structure.

REBASE learns the relations among features across sources, whereas self-attention uses them only within each sample. Our GAT encoder, like self-attention, learns sample-dependent relations among features, and we do not claim this as a contribution. In self-attention, the attention weights only determine how each feature token aggregates the others within the same sample, and sources are connected only through the shared weights and the labels. Samples with the same label therefore need only fall on the same side of the decision boundary, and nothing in the objective requires samples with different relations to be distinguished. In REBASE, by contrast, the affinity AA of each instance graph is aligned with the shared prototype graphs via FGW, and the encoder is trained through this alignment (Eq. (6)). The affinities are thus trained to be comparable across sources rather than only to fit the labels, so that samples with corresponding relations can be represented by the same bases even when their feature sets differ.

As a result, REBASE determines where a sample from an unseen source lies. It is aligned with the same prototype graphs and thus represented by the same bases as training samples with corresponding relations. Self-attention has no such component, so we expect the gap to be largest in zero-shot transfer.

Controlled comparison with self-attention [new]. To test this, we newly compared REBASE with three self-attention models that use the same language model embeddings. For a fair comparison, two of them match REBASE in parameter count (L=2 and L=4) and one matches it in time per step (L=6), so that any gap cannot be attributed to model size or computation.

Model#Paramss/stepAUROC (6-source mean)Zero-shot AUROC (mean)
Self-attention (L=2, d=540)3.65M0.07183.161.9
Self-attention (L=4, d=384)3.66M0.07584.062.6
Self-attention (L=6, d=384)5.43M0.10984.563.1
REBASE3.66M0.11088.670.9

REBASE outperforms all three under either matching condition, with the largest margin in zero-shot transfer, as expected. REBASE uses more memory than the time-matched model (3.17 vs. 1.85 GB), mainly because it aligns each sample with all prototype graphs via FGW and computes the output of every basis. We have added the per-source results to Appendix B.7 of the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.