acceptodds
Under review as a conference paper at ICLR 2027

Beyond Correlation and Shared Spaces: Conditional and Adaptive Graph Imputation

Abstract

A column-centric graph provides a natural representation for tabular imputation because its topology can be shared across rows, allowing the same structural information to support both in-sample and out-of-sample imputation. However, such graph-based formulations face two fundamental challenges: pairwise associations may not reflect direct conditional dependencies, and shared message transformations may overlook the heterogeneous representation spaces of different columns.This paper proposes CAGI (Conditional and Adaptive Graph Imputation) to address these challenges. CAGI constructs a conditional dependency graph by estimating latent correlations with a rank-based estimator and deriving a sparse graph from the resulting partial correlations, and performs column-adaptive graph diffusion through layer-wise source and target transformations. A theoretical analysis establishes when conditional neighborhoods provide sufficient information for reconstruction and when graph diffusion can preserve the information required by the Bayes-optimal imputation function. An excess-risk decomposition further characterizes the effects of information loss in diffusion and approximation error. Experiments on tabular imputation tasks demonstrate the effectiveness of CAGI.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Tabular Data Imputation, Conditional Dependence Graphs

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_XLP9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CAGI, a graph-based tabular imputation method that represents columns as nodes in a shared graph. CAGI estimates latent correlations under a Gaussian copula framework, employing rank-based, polyserial, and polychoric estimators to accommodate mixed-type columns. The model then constructs a graph using the resulting partial correlations. CAGl uses distinct source and target maps for each column using a bidirectional diffusion operator based on the CSNN (Column-wise Shared Neural Network) architecture.

Critical strengths

S1. The empirical results are consistent. CAGI ranks first in all averaged cells of Table 1 across nine datasets and diverse settings.

S2. The paper is clearly written and easy to follow.

S3. Estimating the graph with a Gaussian copula and polyserial / polychoric correlations is an interesting way to handle mixed-type columns without assuming a common marginal distribution.

Critical weaknesses

W1. I am not entirely convinced of the novelty. To me, the proposed formulation resembles a standard sheaf diffusion network, specifically CSNN (Ribeiro et al., 2026), applied with columns as nodes. Approaches that interpret columns as a graph structure already exist, such as GRAPE and BCGNN. Isn’t this an extension of that concept using sheaves? If so, what is tabular-specific in the design beyond applying CSNN?

W2. It is not clear why the sufficiency of the "conditional neighborhood" in Theorem 3 applies to the imputation setting. Theorem 3 assumes all other variables are observed, whereas in practice, one can condition only on observed variables. For instance, in a chain 1–2–3, variables 1 and 3 are conditionally independent given 2, so there is no edge between 1 and 3 in the conditional graph. However, if variable 2 is missing, variable 1 becomes directly necessary for predicting variable 3. As the missingness rate increases, these scenarios become more frequent, shifting the dependency structure required for imputation from the conditional graph toward the marginal graph.

I think multi-hop diffusion may indirectly propagate information through masked nodes, yet the paper does not empirically verify this.

W3. The experimental evidence supporting the conditional graph is weak. In Table 2, the conditional graph shows an MAE only 1.16% lower than that of a marginal-correlation graph with the same number of edges. Moreover, this result is obtained under the MCAR 30% setting, which is the most favorable scenario for the conditional graph. Following the logic in W2, the margin may shrink further at 50% and 70% missingness, where the marginal graph becomes relatively more significant. Furthermore, Gaussian copula imputation (Zhao & Udell, KDD 2020)—which employs a latent model similar to the one used for graph construction—would serve as a natural baseline, yet it is absent from the comparison.

Questions for authors

Q1. Could you report the graph ablation of Table 2 (conditional vs. marginal, complete, and random graphs) at MCAR 50% and 70%? It would be nice to run it over multiple seeds. In addition, how does CAGI compare with Gaussian copula imputation (Zhao & Udell, KDD 2020)?

Q2. Section 3.2 defines SiS_i and TiT_i as parameters learned for each column. However, the ablation study in Figure 4 replaces these maps with identity mappings. This does not merely remove column specificity; it effectively eliminates the learned transport mechanism entirely. To isolate the specific effect of column-specific maps, could you provide a comparison with a case where all columns share the same map—which is still learned during training (i.e., Si=SS_i = S and Ti=TT_i = T)?

10 comments
AuthorsOpenReview ↗

Rebuttal Part 1: W1

We thank the reviewer for the constructive comments and suggestions.

W1: Clarifying the Novelty beyond Existing Graph and Sheaf Diffusion Methods

We thank the reviewer for raising this important concern. We agree that neither representing columns as graph nodes nor using sheaf-based diffusion is, by itself, a novel contribution of our work. We also acknowledge the close connection between our propagation mechanism and Cooperative Sheaf Neural Networks (CSNN; Ribeiro et al., 2026). In particular, CAGI adopts the source/target transformation principle and the associated bidirectional sheaf diffusion formulation introduced in CSNN. We will revise the manuscript to make this relationship explicit and avoid presenting the underlying sheaf diffusion mechanism as an independent methodological novelty.

Our contribution, however, is not simply to apply CSNN to a graph whose nodes correspond to table columns. Rather, CAGI develops an imputation-oriented framework in which the dependency structure, information propagation, and reconstruction objective are designed together. We clarify the distinctions below.

1. Modeling conditional dependencies rather than merely constructing a graph over columns.

Existing graph-based approaches have explored different graph representations for incomplete tabular data. For example, GRAPE represents observations and features as two types of nodes in a bipartite graph, with observed values as edge attributes, and formulates imputation as edge-level prediction. BCGNN further investigates feature interdependencies through a graph-based formulation. We agree that these works establish the relevance of graph structures for tabular missing-data problems.

CAGI focuses on a different aspect of the problem: which inter-column relationships should guide the reconstruction of a missing value? Pairwise marginal correlations can reflect indirect associations mediated by other variables, rather than direct conditional dependencies. Consequently, a graph constructed directly from marginal correlations may introduce edges that do not represent direct conditional relationships relevant to reconstruction.

To address this issue, CAGI estimates a latent correlation matrix for heterogeneous columns using a rank-based estimator, and then estimates a sparse precision matrix to construct a conditional dependency graph. Under the Gaussian latent-variable formulation, the nonzero off-diagonal entries of the precision matrix characterize conditional dependencies. Thus, the graph is intended to encode direct statistical dependencies among columns rather than simply their pairwise marginal associations. Importantly, this graph is shared across rows, while each row provides a distinct set of node features on the same column graph.

This graph construction is not a new sheaf-theoretic contribution; its role is to provide an imputation-oriented information structure that is not specified by a generic sheaf diffusion architecture.

2. Adapting information propagation to heterogeneous tabular variables and cell-wise reconstruction.

We acknowledge that the source/target conformal transformations and bidirectional diffusion in CAGI are closely related to CSNN. We do not claim to introduce a new general-purpose sheaf diffusion operator. Instead, CAGI uses this propagation principle to address a specific challenge in tabular imputation: different columns represent heterogeneous variables, whose values may have different domains, distributions, and semantics.

Accordingly, CAGI employs column-wise feature encoders to map numerical and categorical values into column-specific representation spaces. The source and target transformations then allow information to be communicated across these heterogeneous spaces, rather than imposing a single shared transformation on all columns. The propagation is performed on the shared conditional dependency graph, with row-specific node representations and missingness information. Finally, CAGI uses a masked reconstruction objective and column-specific reconstruction heads to recover missing cell values.

Therefore, the role of the sheaf-based mechanism in CAGI is not merely to improve generic graph representation learning. It provides a structured means of aligning and communicating information between heterogeneous variables within a cell-level reconstruction framework.

AuthorsOpenReview ↗

Rebuttal Part 2: W1

3. An imputation-oriented theoretical analysis.

Another distinction from directly applying a general-purpose sheaf network is that our theoretical analysis is formulated around the statistical imputation objective. We study two complementary questions: (i) when the conditional neighborhood of a target column contains sufficient information for reconstructing that column, and (ii) when the graph propagation representation preserves sufficient information for Bayes-optimal imputation. This connects the graph construction and representation propagation to the information required by the reconstruction task, rather than analyzing diffusion only in terms of generic graph-learning properties.

In summary, we agree with the reviewer that CAGI builds upon existing ideas in column-based graph modeling and CSNN-style sheaf diffusion. We will make this relationship explicit and sharpen our novelty claims accordingly. The contribution of CAGI lies in the imputation-specific integration of conditional dependency modeling, heterogeneous column representation alignment, and masked cell reconstruction, together with a theoretical analysis of the information conditions required for effective imputation. We believe this distinction is important: the proposed framework is not intended as a new general sheaf architecture, but as a principled graph-based approach tailored to the statistical structure and reconstruction objective of heterogeneous tabular data.

AuthorsOpenReview ↗

Rebuttal Part 3: W2

W2: Mask-dependent Predictive Neighborhoods and Multi-hop Information Propagation

We thank the reviewer for this insightful comment. We agree that the sufficiency of the conditional neighborhood in Theorem 3 should not be directly interpreted as sufficiency for arbitrary missingness patterns. The reviewer’s chain example precisely illustrates an important distinction between the conditional dependency structure under complete observation and the predictive dependencies induced by a particular missingness pattern. We clarify this distinction and explain how our mask-aware theoretical analysis and multi-hop graph diffusion address this issue.

1. Conditional neighborhoods versus mask-dependent predictive neighborhoods.

Theorem 3 establishes that, under the Markov assumption and complete observation of the remaining variables, the conditional distribution of a target variable (XiX_i) depends only on its conditional neighborhood (Ni\mathcal N_i). It does not imply that (Ni\mathcal N_i) remains sufficient when some of its members are missing.

Consider the chain (1-2-3). Under the Markov property, we have

X1⊥X3∣X2X_1\perp X_3\mid X_2.

When (X2X_2) is observed, it is sufficient to condition on (X2X_2) when predicting (X3X_3), and (X1X_1) provides no additional information. However, when (X2X_2) is missing, the relevant conditional distribution becomes

∫p(X3∣X2)p(X2∣X1) dX2\int p(X_3\mid X_2)p(X_2\mid X_1)\,dX_2,

which generally depends on (X1X_1). Thus, (X1X_1) can become informative for predicting (X3X_3) after marginalizing over the missing intermediate variable.

This observation does not imply that the conditional dependency graph should be replaced by a marginal correlation graph. Rather, it shows that the set of variables sufficient for prediction is mask-dependent and may extend beyond the original one-hop conditional neighborhood. In particular, marginal association between two variables does not necessarily identify a direct conditional dependency; using it as the graph structure may introduce edges that do not represent direct dependencies in the underlying distribution.

2. How our mask-aware formulation addresses this issue.

We explicitly distinguish the conditional neighborhood (Ni\mathcal N_i) in Theorem 3 from the mask-aware predictive neighborhood (Bi(M)\mathcal B_i(M)) introduced in Definition 10. The latter is defined with respect to the information actually observed under mask (MM):

Xi⊥{Xj:Mj=1, j∉Bi(M)}∣{Xj:Mj=1, j∈Bi(M)},G,MX_i \perp \{X_j : M_j=1,\ j\notin\mathcal B_i(M)\} \mid \{X_j : M_j=1,\ j\in\mathcal B_i(M)\},G,M.

Equivalently, the observed variables in (Bi(M)\mathcal B_i(M)), together with the graph and mask, are sufficient for the conditional distribution of the target cell.

For the chain example, when node 2 is masked and nodes 1 and 3 are observed, node 1 may belong to (B3(M)\mathcal B_3(M)), even though (1∉N31\notin\mathcal N_3). This formulation therefore accounts for the additional predictive information induced by marginalizing over missing intermediate variables, rather than assuming that the original conditional neighborhood is sufficient under every mask.

3. Why multi-hop diffusion can access information through masked intermediate nodes.

This mask-dependent predictive neighborhood also motivates our graph diffusion analysis. Definition 11 introduces the graph coverage condition, which requires the predictive neighborhood to be contained in the target node's receptive field:

Bi(M)⊆Ri(L)(G)\mathcal B_i(M)\subseteq\mathcal R_i^{(L)}(G).

As established by Lemma 12, each diffusion layer can propagate information by at most two graph hops. Consequently, in the chain example, information from node 1 can reach node 3 through node 2 even when the value of node 2 is masked. The masked cell does not remove node 2 from the graph or prevent its representation from receiving messages. Instead, node 2 can aggregate information from its observed neighbors and subsequently transmit the updated representation to node 3.

Importantly, this is an information propagation mechanism, not an assumption that the missing value at node 2 has been observed or reconstructed perfectly. Whether the relevant information is retained in the intermediate and target representations is addressed separately by the representation-sufficiency condition in Theorem 13. Theorem 14 further decomposes the excess imputation risk into representation information loss and decoder approximation error. Thus, graph coverage ensures that the required information is structurally accessible, while the representation and decoder conditions characterize whether it can be preserved and used for reconstruction.

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

CAGI addresses tabular missing-value imputation using a graph whose nodes are columns and whose structure is shared across rows. It estimates latent correlations under a Gaussian-copula formulation, obtains partial correlations from the precision matrix, and constructs a connected graph using within-block thresholds. The model then applies CSNN-style diffusion with column-specific source and target conformal maps, trained by masking and reconstructing observed cells. The paper provides analyses of neighborhood sufficiency, propagation expressiveness, and Bayes risk. Experiments on nine datasets report consistent average improvements on numerical imputation under several missingness settings, both in-sample and out-of-sample, with additional graph and map ablations.

Critical strengths
  1. The reported numerical results are consistent across settings. CAGI is best in all 18 numerical entries of Table 1, and the appendix extends the evaluation to higher MCAR rates. All methods use the same evaluation masks within each dataset and setting. This breadth makes the empirical results worth taking seriously.
  2. The paper gives a clear motivation for considering dependency structure and message transformations separately. The BILL_AMT example illustrates how marginal and partial-correlation weights emphasize different relationships. The resulting hypothesis is concrete enough to test through controlled graph and propagation ablations.
Critical weaknesses

1 Incremental contribution and theoretical scope

The propagation construction closely follows CSNN (Ribeiro et al., 2026), including separate source and target conformal maps and the composition of in/out operators. The graph estimator also draws on established mixed-data graphical modeling, as acknowledged in Appendix C. The combination may be useful, but I would like the paper to state more clearly what is new and show which benefits arise from this combination rather than from the existing components. Theorem 5 also leaves a specific gap between the analysis and the model. Its proof allows arbitrary linear maps and sets T_i = W^T for a general shared matrix W. This choice need not satisfy the conformal restriction used in the implementation. The inclusion is therefore not established for the implemented edge-map class; whether an analogous whole-layer result holds with the shared weights is not shown. Theorem 3 concerns latent variables on a correct Markov graph, and its transfer to categorical observations, missing neighbors, and an estimated graph needs further conditions. These points limit the theoretical support for the proposed explanation of CAGI. The risk analysis gives useful sufficient conditions, but it does not show when CAGI satisfies them.

2 Evidence for the components and comparison details

Table 2 reports a 1.16% MAE and 0.72% RMSE advantage over a marginal graph with the same edge count, without uncertainty for these differences. Table 12 provides a paired comparison on default, but only for one split and mask. Appendix D does not specify the number of runs, and the described variation is across partitions with the seed fixed. I cannot tell from these results whether the conditional graph provides a stable improvement over marginal dependence. Table 10 lists baseline configurations, but their individual source suites and tuning procedures remain unclear. The MissForest substitute, for example, uses 10 trees of depth 3; its capacity deserves clarification. Figure 4 shows that the maps help, but replacing all maps by identity removes several capabilities at once while retaining the shared layer weights. This does not isolate column-specific alignment or source/target separation. More focused comparisons would help establish the source of the gains.

Questions for authors
  1. What differs from CSNN in the propagation and map generation? Please specify the inputs and architecture of the S_i/T_i generator, including dependence on column identity, row features, missingness, and layer.
  2. Can Theorem 5 be established for the actual conformal parameterization, accounting for the shared layer weights where needed? Otherwise, could the claim be limited to the class that is proved?
  3. What conditions connect Theorem 3 to observed mixed-type data with missing entries? For nominal columns, how is category order chosen, and are the graph and results stable under relabeling? A representative check would help.
  4. How many runs support each result? Please provide per-dataset paired conditional-versus-marginal differences with uncertainty, especially for Table 2, and clarify which sources of randomness are varied.
  5. Which implementation, source suite, and tuning procedure produced each baseline configuration? In particular, do the conclusions hold with a higher-capacity MissForest configuration selected using the same permitted validation information?
  6. Would a focused comparison with shared learnable maps, or with source and target maps tied, support the claimed benefit of column-specific alignment? This would help distinguish the explanation from the broader effect measured by the identity ablation.
11 comments
AuthorsOpenReview ↗

Rebuttal Part 1: W1 & Q1

We thank the reviewer for the constructive comments and suggestions.

W1: Incremental contribution and theoretical scope

Q1: Difference from CSNN and the S/T generator

We thank the reviewer for raising this important concern. We agree that the source-target conformal maps and the bidirectional propagation construction in CAGI are closely related to Cooperative Sheaf Neural Networks (CSNN; Ribeiro et al., 2026). In particular, we do not claim the source-target factorization or the underlying in/out diffusion construction as new contributions. Our contribution is instead to develop a task-specific instantiation of this propagation principle for heterogeneous tabular imputation, where the graph structure, node features, transformation generators, and learning objective have distinct statistical meanings.

1. What is different from CSNN?

The key difference lies in how the graph and its node-wise transformations are constructed and conditioned for tabular imputation.

First, the graph in CAGI is not an arbitrary input graph. Its nodes correspond to table columns, and its topology is estimated from mixed-type data using latent correlation estimation followed by sparse precision estimation. The resulting graph is intended to encode direct conditional dependencies among columns, rather than generic pairwise associations. Importantly, this graph is shared across rows, while the node features are instantiated separately for each row. This provides a common dependency structure through which information can be propagated for both in-sample and out-of-sample imputation.

Second, although CSNN also employs node-dependent source and target maps, CAGI specifies their conditioning information according to the semantics of tabular data and the missing-value reconstruction task. In particular, the map generators receive (i) a column identity embedding, (ii) a row-conditioned cell embedding that represents either the observed cell value or a dedicated missing-value embedding, and (iii) the current-layer node representation. Thus, the generated maps are conditioned jointly on the identity of the variable, the value and observation status of the corresponding cell, and the information accumulated during graph propagation. This allows the transformations to vary not only across columns, but also across sample rows and propagation layers.

Third, CAGI connects this design to the imputation objective. The conditional dependency graph determines the candidate information pathways among columns, whereas the row-conditioned adaptive transformations determine how information from heterogeneous and potentially missing cells is aligned and communicated. Our theoretical analysis further studies conditional neighborhoods, mask-aware predictive information, and the excess imputation risk induced by information loss during diffusion. These questions are specific to the information requirements of missing-value reconstruction and are different from the cooperative communication and over-squashing analyses in CSNN.

We have revised the Introduction and Method sections to make these distinctions explicit. We also clarify that the rank-based mixed-data dependence estimation and the source-target diffusion construction are established components that we adopt, rather than claiming them as standalone methodological novelties.

AuthorsOpenReview ↗

Rebuttal Part 2: W1 & Q1

2. Inputs and architecture of the (Si/TiS_i/T_i) generators

We have expanded the description of the transformation generators. For a row (rr), column (ii), and diffusion layer (ℓ\ell), the generator input is

[eicol  ∥  er,icell  ∥  Hr,i(ℓ)][e_i^{\mathrm{col}} \;\Vert\; e_{r,i}^{\mathrm{cell}} \;\Vert\; H_{r,i}^{(\ell)}]

where (eicole_i^{\mathrm{col}}) is the column identity embedding, (er,icelle_{r,i}^{\mathrm{cell}}) is the row-conditioned cell embedding, and (Hr,i(ℓ)H_{r,i}^{(\ell)}) is the current-layer representation. The cell embedding is obtained from the column-specific value encoder for an observed cell, while a dedicated missing embedding is used when the cell is missing or masked during training.

The source and target transformations are generated by two separate GraphSAGE-based networks, with independent parameters and output heads:

Zr,iS,(ℓ)=GraphSAGE(G,Ur(ℓ);θS(ℓ))Z_{r,i}^{S,(\ell)} = \mathrm{GraphSAGE}(G,U_r^{(\ell)};\theta_S^{(\ell)}), Zr,iT,(ℓ)=GraphSAGE(G,Ur(ℓ);θT(ℓ))Z_{r,i}^{T,(\ell)} = \mathrm{GraphSAGE}(G,U_r^{(\ell)};\theta_T^{(\ell)}).

Sr,i(ℓ)=gS(ℓ)(Zr,iS,(ℓ))S_{r,i}^{(\ell)} = g_S^{(\ell)}(Z_{r,i}^{S,(\ell)}), Tr,i(ℓ)=gT(ℓ)(Zr,iT,(ℓ))T_{r,i}^{(\ell)} = g_T^{(\ell)}(Z_{r,i}^{T,(\ell)}).

The two generators use the same conditional dependency graph but do not share their GraphSAGE parameters. Their output heads produce the positive scale and orthogonal factors of the conformal maps, with the orthogonal factors parameterized using Householder reflections. The generators are evaluated at each diffusion layer using the corresponding current-layer representations. We also clarify that no global pooling is applied in these generators: each row is treated as a graph instance with row-specific node features over the shared column graph.

Consequently, the induced transformation from column (jj) to column (ii) is row- and layer-dependent:

(Tr,i(ℓ))⊤Sr,j(ℓ)(T_{r,i}^{(\ell)})^\top S_{r,j}^{(\ell)}.

This explicit conditioning is important for tabular imputation: two rows can share the same column graph and column identities while having different observed values and missingness patterns, and therefore can induce different source-target transformations.

3. How do the benefits arise from the combination?

The two components address complementary aspects of the imputation problem. The conditional dependency graph specifies which columns are structurally relevant to one another, reducing reliance on marginal associations that may include indirect relationships. The adaptive transformations then provide a mechanism for communicating information across heterogeneous column representations while accounting for row-specific values and missingness.

Our existing identity-map ablation provides evidence that the learned conformal transformations contribute to performance: replacing the learned maps with identity maps increases both in-sample and out-of-sample reconstruction errors across all reported dataset-metric comparisons. We have clarified the interpretation of this ablation and avoid treating it as a direct comparison against every possible shared-map parameterization.

To make the contribution of the two components more transparent, we have also revised the discussion of their complementary roles and the scope of our empirical claims. The revised manuscript distinguishes the benefit of learning conformal transformations from the benefit of constructing a conditional dependency graph, rather than attributing all improvements to either component in isolation.

Overall, CAGI is not intended as a new general-purpose sheaf diffusion architecture. Its contribution is an imputation-oriented framework that combines a mixed-data conditional dependency structure with row-conditioned, missingness-aware, column-adaptive propagation, together with theoretical analysis tailored to the information requirements of tabular imputation.

AuthorsOpenReview ↗

Rebuttal Part 3: W1 & Q2

Q2: Expressiveness under the Actual Conformal Parameterization

We thank the reviewer for identifying an important gap in our original theoretical analysis. We agree that the original proof of Theorem 6 was insufficient. Specifically, it established the inclusion by setting (Sj=IqS_j=I_q) and (Ti=W⊤T_i=W^\top) for an arbitrary shared matrix (WW). However, (W⊤W^\top) is not necessarily conformal, and thus this construction does not respect the parameterization used in our implementation. We have therefore revised the theoretical analysis to distinguish the expressiveness of the induced edge maps from that of the complete bidirectional propagation layer.

1. We characterize the conformal transformation class (new Lemma 4).

We introduce an explicit conformal transformation class and establish its relevant closure properties. In particular, the transpose of a conformal transformation remains conformal, and the composition of two conformal transformations is also conformal. This result provides the algebraic foundation for analyzing the source-target factorization under the actual conformal constraint.

2. We revise Theorem 6 to establish expressiveness under the conformal constraint.

We redefine the shared and column-adaptive edge-map classes using conformal transformations, rather than arbitrary linear maps. We then establish

Fsharedconf⊆Fadaptconf\mathcal F_{\mathrm{shared}}^{\mathrm{conf}} \subseteq \mathcal F_{\mathrm{adapt}}^{\mathrm{conf}}.

The inclusion follows by choosing (Sj=IqS_j=I_q) and (Ti=C⊤T_i=C^\top) for every node, where (CC) is any shared conformal edge map. This construction satisfies the conformal constraint and therefore resolves the gap in the original proof. We further establish strict inclusion under a mild graph condition, provided that the conformal class contains a realizable non-identity transformation.

Importantly, the revised theorem concerns the algebraic edge-map classes and does not rely on arbitrary unconstrained linear transformations.

3. We extend the analysis to the complete bidirectional propagation layer (new Theorem 8).

We agree that matching the off-diagonal edge maps alone is insufficient to establish equality of the complete propagation operators. In our actual propagation mechanism, the out- and in-oriented operators also contain the self-interaction terms (Si⊤SiS_i^\top S_i) and (Ti⊤TiT_i^\top T_i), respectively. Their normalized forms further depend on the corresponding degree operators.

To address this issue, we introduce a separate layer-level result. We define the shared conformal propagation layer by assigning the same source and target transformations to every node, and compare it with the column-adaptive parameterization. By assigning these same transformations to every node in the adaptive parameterization, we match both the off-diagonal edge interactions and the diagonal self-interaction terms. Consequently, the corresponding unnormalized operators, degree normalizations, normalized operators, and composed bidirectional diffusion operators are identical.

Throughout this construction, the graph, edge weights, shared feature transformations (W1(ℓ)W_1^{(\ell)}) and (W2(ℓ)W_2^{(\ell)}), and the remaining layer components are kept unchanged. Thus, Theorem 8 establishes that every shared conformal propagation layer can be reproduced by the column-adaptive parameterization under matched layer components.

4. We retain and clarify the geometric interpretation of the conformal factorization (Proposition 9).

We retain the geometric characterization of the source-target factorization and renumber the corresponding proposition as Proposition 9. Specifically, the induced edge transformation can be decomposed as

ciTcjS(QiT)⊤QjSc_i^T c_j^S (Q_i^T)^\top Q_j^S.

The relative transformation ((QiT)⊤QjS(Q_i^T)^\top Q_j^S) is orthogonal, while (ciTcjSc_i^T c_j^S) controls the scale. This clarifies how the source-target factorization implements relative coordinate alignment between heterogeneous column representations.

5. Scope of the revised theoretical claims.

The revised analysis explicitly distinguishes edge-map expressiveness from complete propagation-layer expressiveness. Theorem 6 establishes strict inclusion between the conformal edge-map classes. Theorem 8 establishes inclusion between the shared and column-adaptive conformal propagation layers under matched graph, edge-weight, self-interaction, normalization, and shared feature-transformation components. We do not claim unrestricted function-class inclusion between the complete neural map generators, nor do we infer strict inclusion of the complete normalized propagation-operator classes solely from the strict edge-map inclusion.

We have revised the corresponding definitions, theorem statements, and proofs in the main text and appendices to make the theoretical claims consistent with the conformal source-target parameterization and the actual bidirectional diffusion mechanism.

Reviewer_bUHNOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CAGI, a graph-based method for tabular imputation. Columns are represented as nodes in a graph constructed from estimated latent partial correlations. The model then uses column-specific source and target transformations to propagate information between heterogeneous features and reconstruct missing values. The paper also provides theoretical results on information preservation and reconstruction risk. Experiments on nine datasets cover different missingness mechanisms, in-sample and out-of-sample imputation, and downstream prediction.

Critical strengths
  1. The imputation results are consistently competitive across the evaluated datasets and missingness settings. CAGI achieves the best average numerical imputation metrics under MCAR, MAR, and MNAR at 30% missingness, both in-sample and out-of-sample.

  2. The paper examines both graph construction and message transformation rather than treating the graph as a fixed preprocessing choice. The comparisons with marginal, random, and complete graphs, together with the transformation ablation, provide some evidence that both components contribute to the results.

Critical weaknesses
  1. Scalability with respect to the number of columns is not adequately evaluated. All nine datasets have at most 47 columns. However, graph construction involves estimating a full correlation matrix, performing an eigen decomposition, and computing a matrix inverse. The default threshold also retains the top 50% of partial correlations within each block, so the number of edges can still grow quadratically with the number of columns. Table 11 reports training costs but excludes graph construction time. As a result, the current experiments do not establish whether the method remains practical for wider tables, where these costs could become substantial.

  2. The evidence for the necessity of column-specific transformations is incomplete. Figure 4 replaces all source and target maps with identity matrices, removing learned alignment altogether. This does not distinguish the benefit of column-specific transformations from the benefit of simply having learnable transformations. A comparison with learned maps shared across columns, ideally with a parameter-matched control, would more directly test the paper's claim that shared transformations are insufficient for heterogeneous features. Similarly, the conditional graph improves average MAE and RMSE over the marginal graph by only 1.16% and 0.72% in Table 2, and uncertainty is not reported for these differences. These results leave some uncertainty about how much of the overall gain comes from the two central design choices.

Questions for authors
  1. How do graph construction time, training time, and peak memory change as the number of columns increases? Can you provide results on a wider dataset or a controlled scaling experiment? Please include the graph construction cost omitted from Table 11.

  2. How does the model perform when the source and target transformations are learned but shared across columns, rather than fixed to the identity? This comparison would help establish whether column-specific alignment is responsible for the improvement.

  3. Are the improvements of the conditional graph over the matched-edge-count marginal graph consistent across random seeds and missingness masks? Reporting variation for this comparison would help assess the importance of conditional graph construction.

5 comments
AuthorsOpenReview ↗

Rebuttal Part 1: W1 & Q1

We thank the reviewer for the constructive comments and suggestions.

W1 & Q1: Scalability with respect to the number of columns

We have measured graph-construction cost and peak GPU memory and have added both to Table 11.

Real datasets. Across the nine datasets, one-off construction takes 0.02–29.4 s. Relative to the training time in Table 11, this is 0.0016%–1.53% (median ≈ 0.45%). The variation reflects the sample size, the column-type combination and the categorical cardinality; the categorical block is the dominant cost in mixed tables, where polychoric/polyserial pairs require a one-dimensional optimization per pair. For example, adult (n = 22,792, 8 categorical columns) takes 29.4 s while obesity (n = 1,477, 8 categorical columns) takes 2.63 s, and news (d = 47, 2 categorical columns) takes 2.85 s.

Controlled scaling. With n = 10,000 fixed and all-numeric columns, construction time follows a log-log slope of 1.98 against d (the number of columns; d = 20, 50, 100, 200, 400, 800; Table R1-Q1.2, d-scan rows), i.e. approximately quadratic empirical scaling over the tested range, reaching 110.1 s at d = 800. The eigendecomposition and inversion retain their cubic worst case, but they are not the dominant term in this range. With d = 100 fixed, time is linear in n over n = 1,000–50,000 (t = 1.08×10⁻⁴·n + 0.61 s, R² = 0.999; Table R1-Q1.2, n scan).

Categorical columns. The dominant cost on real tables is the categorical block. With the categorical cardinality fixed at the median of the 39 categorical columns in our datasets (7) and a 50:50 numeric/categorical split, construction takes 24.7 s at d = 20 versus 0.075 s for the all-numeric table of the same size—a 329× difference (Table R1-Q1.2, mixed rows), caused by the one-dimensional optimization in the polychoric/polyserial estimators. The mixed graph also shows approximately quadratic empirical scaling over the tested range: from d = 20 to 100, time grows by 28.4× (log-log slope 2.08).

Scope and mitigations. At batch size 128, training fits at d = 150 (15.3 GB) and exceeds our GPU memory at d = 200 (19,900 directed edges). The peak is about 1.4 MB per directed edge, so a smaller batch size reduces it. For wider tables, the natural extensions are (i) vectorizing the pairwise estimator, (ii) parallelizing the three type blocks, and (iii) sparsifying edges with top-k or an absolute threshold.

Benchmark scope. The nine benchmark datasets span 8 to 47 modeled columns, the same range used by the DiffPuter and CACTI suites we compare against. Wider tables also occur outside this benchmark, so we measure the scaling curve up to d = 800 instead of extrapolating from d ≤ 47: construction shows approximately quadratic empirical scaling with a per-pair cost of 0.33–0.40 ms for numeric columns and 130–141 ms when half the columns are categorical, and the training peak is about 1.4 MB per directed edge, with batch size as the parameter that extends d.

Table R1-Q1.1 Real datasets (MCAR 30%). n is the number of training rows; num and cat are the numeric and categorical column counts (d = num + cat). Construction is the one-off build_feature_graph call on a fixed MCAR 30% mask. Epochs run, Training and Per epoch are from Table 11 (RTX 3090, early stopping, including the periodic validation pass and excluding construction; Per epoch = Training / Epochs run). Params is the model parameter count. Peak GPU is a separate measurement on the same fixed mask at batch size 128.

DatasetnnumcatConstruction (s)Epochs runTraining (s)Per epoch (s)ParamsPeak GPU (MB)
adult22,7926829.371663,404.220.51112,580182.6
bean9,5271600.04884712,640.25.6199,496226.2
beijing29,2296521.462563,586.414.01108,666125.0
california14,447800.02072011,164.55.7997,82481.3
default21,00014917.022163,785.017.52110,675433.4
magic13,3141000.02911911,805.89.4598,242110.3
news27,7504522.8541117,844.543.42107,5381,607.0
obesity1,477882.63336414.11.23101,543225.6
shoppers8,63110713.13171860.05.03108,003252.1
AuthorsOpenReview ↗

Rebuttal Part 2: W1 & Q1

Table R1-Q1.2 Construction-side scaling (synthetic tables). All rows use the one-off build_feature_graph; time is the median of three repetitions, except the mixed d = 100 row, which is one measured repetition. The d scans fix n = 10,000; the n scan fixes d = 100. Mixed rows use a 50:50 numeric/categorical split with categorical cardinality 7 (the median over the 39 categorical columns of the nine datasets). Peak memory is Python-level tracemalloc.

ScandnnumcatDirected edgesConstruction (s)Per pairPeak Python (MB)
d (all-numeric)2010,0002001900.075395 μs2.40
d (all-numeric)5010,0005001,2240.440360 μs5.85
d (all-numeric)10010,00010004,9501.651334 μs11.63
d (all-numeric)20010,000200019,9006.782341 μs23.31
d (all-numeric)40010,000400079,80026.600333 μs47.11
d (all-numeric)80010,0008000319,600110.142345 μs96.55
n (d = 100)1001,00010004,9500.785158.6 μs1.35
n (d = 100)1005,00010004,9501.142230.7 μs5.86
n (d = 100)10010,00010004,9501.601323.4 μs11.64
n (d = 100)10050,00010004,9506.0241,216.9 μs57.7
d (50:50 mixed)2010,000101018824.685129.9 ms1.39
d (50:50 mixed)4010,0002020780107.711138.1 ms2.41
d (50:50 mixed)6010,00030301,768240.073135.6 ms3.57
d (50:50 mixed)8010,00040403,160432.015136.7 ms4.74
d (50:50 mixed)10010,00050504,948700.063141.4 ms5.91

Rows are grouped by scan: first the all-numeric d scan (vary d at fixed n = 10,000), then the all-numeric n scan (vary n at fixed d = 100), then the 50:50 mixed d scan.

d, all-numeric: log-log slope 1.98, i.e. approximately quadratic over the tested range; per pair stays at 333–395 μs. This is an empirical trend, not the asymptotic complexity; factorization and inversion retain their cubic worst case.

n, all-numeric: linear, t = 1.08×10⁻⁴·n + 0.605 s, R² = 0.9993.

d, 50:50 mixed: log-log slope 2.08 over d = 20–100, i.e. approximately quadratic over the tested range; time grows 28.4× from d = 20 to 100; at d = 20 it is 329× slower than the all-numeric table (24.7 s vs 0.075 s); per pair 130–141 ms.

Table R1-Q1.3 Training-side scaling (synthetic all-numeric tables, n = 10,000, batch size 128, steps = 5). The build+setup column is measured inside the training harness, so it includes surrounding setup and is not the isolated construction time of Table R1-Q1.2 (e.g., 2.682 s vs 1.651 s at d = 100).

dDirected edgesBuild + setup (s)Step (ms)Epoch est. (s)Peak GPU (MB)Peak / edge (MB)
501,2240.748247.119.521,795.71.47
1004,9502.682720.356.906,914.31.40
15011,1746.2621,388.4109.6815,337.01.37
20019,90011.347OOMOOMOOMOOM

Peak memory is linear in the directed edge count: ≈ 1.36 MB per edge + 0.13 GB (the three non-OOM rows give 1.47/1.40/1.37 MB per edge). Because the edge count is approximately C(d,2) in these synthetic all-numeric graphs, peak memory is O(d²): log-log slope 1.95 over d = 50–150. Step time grows more slowly (slope 1.57 against d, 0.78 against edges) because of a fixed per-step overhead; the epoch estimate grows 19.5 → 109.7 s.

Training cost and memory. On the nine real datasets, peak GPU memory is 81 MB–1.61 GB (maximum on news), and all nine train within 1.7 GB (Table R1-Q1.1); Table R1-Q1.3 gives the synthetic d-scan.

AuthorsOpenReview ↗

Rebuttal Part 3: W2 & Q2

W2: Evidence for column-specific transformations and for the marginal-graph margin

Q2: Shared source and target transformations

We ran the requested control. In the shared-map variant, the generator's per-node Householder parameters and log-scale are averaged over the columns of each row before S and T are formed, so all columns in a row share one learned transformation while different rows can differ. The averaged parameters still define a valid conformal map (positive scale and orthogonal factor); averaging the final matrices would not, since the average of orthogonal matrices need not be orthogonal. It is therefore a learned, data-dependent map, not an identity or fixed-map ablation. The two variants have the same parameter count, graph, data, masks, seed and training protocol, so the comparison isolates the sharing of the forward function while keeping the full parameter budget. Each dataset contributes three runs, one per mask, and the comparison is made per dataset on paired runs rather than between the overall column means.

At MCAR 30%, column-specific maps are better on 9/9 datasets in MAE and RMSE, 8/9 in R², and 5/6 in categorical accuracy. The equal-weight relative MAE gap is 3.33% (RMSE 1.90%, accuracy 0.49%), largest on bean (15.6%); the dataset-level paired t-test gives p = 0.0042 (Wilcoxon 0.0039), p = 0.0098 for RMSE, and p = 0.055 for accuracy. Tables R1-Q2.1 and R1-Q2.2 give the per-dataset values and tests. The identity control increases MAE by 5.4%; together with the shared-map result, the two controls show that learnable maps help and that the tested shared-map variant performs worse. Within this control, column-specific alignment remains better.

Table R1-Q2.1. In-sample (column-specific / shared-learned; mean ± s.d. over the three benchmark runs, one per missingness mask). Lower is better for MAE and RMSE, higher for R² and accuracy; the better mean in each cell is bold. The ± value is each variant's own mask-to-mask spread. Because the comparison is paired (same data, same mask, same seed), the relevant uncertainty is the paired difference reported in Table R1-Q2.2.

DatasetMAERMSER²cat_acc
adult0.4485±0.0054 / 0.4502±0.00940.8858±0.0132 / 0.8879±0.01640.2428±0.0017 / 0.2392±0.00230.7439±0.0025 / 0.7436±0.0017
bean0.1047±0.0033 / 0.1210±0.00090.2886±0.0031 / 0.3114±0.00710.9134±0.0018 / 0.8992±0.0036—
beijing0.2780±0.0016 / 0.2826±0.00100.6562±0.0054 / 0.6639±0.00540.5276±0.0039 / 0.5164±0.00260.2920±0.0040 / 0.2931±0.0072
california0.3523±0.0056 / 0.3530±0.00480.6248±0.0286 / 0.6253±0.03600.4458±0.0119 / 0.4502±0.0139—
default0.2155±0.0001 / 0.2201±0.00040.5866±0.0068 / 0.6054±0.01680.6444±0.0042 / 0.6213±0.01590.7720±0.0035 / 0.7666±0.0045
magic0.4080±0.0010 / 0.4141±0.00400.6573±0.0010 / 0.6621±0.00440.5569±0.0017 / 0.5505±0.0047—
news0.2555±0.0004 / 0.2677±0.00160.4857±0.0033 / 0.4983±0.00820.7577±0.0009 / 0.7451±0.00490.4753±0.0045 / 0.4679±0.0041
obesity0.5957±0.0103 / 0.6086±0.00860.8260±0.0216 / 0.8322±0.00880.3200±0.0297 / 0.3094±0.00730.8672±0.0065 / 0.8635±0.0023
shoppers0.3140±0.0036 / 0.3189±0.00350.7132±0.0152 / 0.7162±0.01470.4865±0.0098 / 0.4836±0.01110.6070±0.0011 / 0.6035±0.0009

Table R1-Q2.2. In-sample dataset-level summary over the nine datasets (cat_acc on six datasets). Δ = shared − column, equal-weight mean of the per-dataset differences; for MAE/RMSE a positive Δ means the shared-map variant is worse, for R²/cat_acc a negative Δ means it is worse. The 95% CI, t p, Wilcoxon p and Cohen’s dz are computed over the nine per-dataset gaps (six for cat_acc); the three masks are nested within datasets, so the dataset is the unit of inference.

MetriccolumnsharedΔ95% CI of Δt pWilcoxon pdz
MAE0.33020.3374+0.0071[+0.0030, +0.0113]0.00420.00391.32
RMSE0.63600.6447+0.0087[+0.0028, +0.0147]0.00980.00391.12
R²0.54390.5350−0.0089[−0.0150, −0.0028]0.00960.0195−1.13
cat_acc0.62620.6230−0.0032[−0.0065, +0.0001]0.0550.094−1.01

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.