acceptodds
Under review as a conference paper at ICLR 2027

CG-DIFF: CAUSAL-GRAPH-GUIDED DIFFUSION MODELS FOR TABULAR IMPUTATION UNDER SELF-MASKING MNAR

Abstract

Missing-value imputation under Self-Masking Missing Not At Random (MNAR) is challenging because observed and missing values may follow different distributions, making it difficult to recover missing values from observed data alone. We propose a Causal-Graph-guided DIFFusion model (CG-DIFF) for tabular missing-value imputation under Self-Masking MNAR. CG-DIFF introduces an observation mask into the forward diffusion process to enable training on observed values only, and develops a Causal-Graph-guided Noise Estimator (CG-NE) to estimate diffusion noise from incomplete data. CG-NE combines Kolmogorov–Arnold network projections with graph-masked attention, using either parents plus self or the Markov blanket plus self as the graph support. At each reverse step, CG-DIFF replaces the values at observed positions with their forward-noised counterparts at the corresponding timestep, thereby preserving observed information throughout sampling. We establish sufficient conditions under which the population predictor based on the Markov blanket has lower missing-case squared-error risk than the predictor based only on the parent set, providing theoretical motivation for the Markov-blanket-plus-self design. We evaluate CG-DIFF against nine competing imputation methods on simulated datasets and a real-world dataset. Experimental results show that CG-DIFF with the Markov-blanket-plus-self graph support achieves better imputation performance than the competing methods in most settings.

Primary area:
generative models
Keywords:
Causal Graph, Diffusion Models, Diffusion Noise, Kolmogorov–Arnold Network, Self-Masking MNAR

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_1pXROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CG-DIFF, a diffusion-based method for continuous tabular imputation under the Self-Masking MNAR case. The proposed method combines a masked forward-diffusion objective trained only at observed positions, a Causal-Graph-guided Noise Estimator using KAN projections and graph-masked attention, and reverse sampling that repeatedly restores observed coordinates to their time-matched noisy values. Two graph supports are considered: parents plus self and Markov blanket plus self. Evaluation uses four synthetic SCM families with artificially generated Self-Masking MNAR and one complete ADNI subset. CG-DIFF is compared with nine imputation methods. Its Markov-blanket variant has the lowest reported mean RMSE in 9/12 synthetic settings and all three ADNI masking-rate settings.

Critical strengths

A coherent method for incorporating structural side information. CG-DIFF combines masked diffusion training, graph-restricted information exchange, and repeated restoration of observed values in a technically coherent framework. The model is motivated by the difficulty of learning missing-case behavior from observed values alone under MNAR.

Promising results with unusually explicit graph provenance. CG-DIFF achieves the lowest reported mean RMSE across all synthetic conditions between its two graph variants and obtains the lowest mean in all three controlled ADNI conditions. The authors also appropriately label the ADNI graph as a literature-informed working prior rather than ground truth and disclose the evidence behind each edge

Critical weaknesses

The evaluation does not substantiate real-world MNAR effectiveness or robustness to the required graph. ADNI supplies real covariates but not real missingness: the paper selects a complete, small subset and imposes the same artificial mechanism used in simulation. Only one monotone self-masking family is evaluated, with no lower-tail, two-sided, nonlinear, variable-specific, multivariate, site-dependent, or naturally occurring missingness. Moreover, synthetic CG-DIFF receives the generator DAG, while ADNI uses a single hand-curated DAG. There is no edge-deletion/addition/reversal study, estimated-graph comparison, alternative plausible ADNI graph, or random/dense/self-only matched-capacity control. So the evidence does not show whether the method works under the uncertain graphs and heterogeneous mechanisms encountered in practice.

The theoretical and empirical evidence is incomplete and inconsistent. Theorem 3.1 motivates an oracle Markov-blanket conditional-mean predictor under an unverified covariance condition; it does not establish a result for CG-DIFF’s noised inputs, mask dropout, multi-hop attention, or reverse sampler. Empirically, the Markov-blanket advantage over parents is small on average in Table 6, reverses in several Table 1 settings, and is unsupported by paired confidence intervals or tests. Baseline tuning spaces, budgets, stopping rules, selected configurations, and run seeds are also not sufficiently documented.

Questions for authors
  1. I was wondering if you tested with more diverse MNAR mechanisms or stronger real-world scenarios? If not, the real-world MNAR claim should be softened/narrowed.
  2. How robust is CG-DIFF to missing, incorrect, or estimated graph edges, especially for the ADNI graph?
  3. How does Theorem 3.1 connect to the actual CG-DIFF model, and does its covariance condition hold empirically?
  4. Please report paired uncertainty estimates and clarify whether baseline tuning was comparable.
  5. Would you please explain why relevant diffusion baselines like CSDI, ForestDiffusion are not included in the experiments?
Reviewer_3KGkOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces CG-DIFF, a diffusion framework for imputation that adopts a predictor for diffusion noise from incomplete data. This predictor combines Kolmogorov–Arnold network projections and graph-masked attention to enhance its capability.

Critical strengths
  • The paper is well written and clearly organized. I am comfortable following this paper.
Critical weaknesses
  • I didn't get the core idea of why this paper adopts Kolmogorov–Arnold network projections as the noise predictor, which theoretically can be any network. I think there is a missing part of the ablation study on the network choice.
  • This is no convergence analysis for the diffusion initial noise predictor; will the final imputed tabular output still follow the original data distribution?
  • The paper provides experimental validation that shows the superiority of the proposed method; however, the experimental results is rather limited in size. More experiment dataset, such as in [1], is encouraged.

[1] Zhichao Chen, Haoxuan Li, Fangyikang Wang, Odin Zhang, Hu Xu, Xiaoyu Jiang, Zhihuan Song, and Hao Wang. Rethinking the diffusion models for missing data imputation: A gradient flow perspective. In Advances in Neural Information Processing Systems, pp. 112050–112103, 2024.

Questions for authors

please see weakness

Reviewer_UA67OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a model called Causal-Graph-Guided Diffusion (CG-DIFF) for tabular missing-value imputation under Self-Masking MNAR. It trains a diffusion model using only observed values and uses a given causal graph to guide noise estimation through KAN projections and graph-masked attention. The model considers either parent variables or the Markov blanket and preserves observed information during the reverse diffusion process. Experiments on four types of synthetic data and the ADNI dataset show that CG-DIFF achieves better imputation performance than nine baseline methods in most settings.

Critical strengths
  1. Strong experimental results.

CG-DIFF achieves better imputation performance than existing methods in most experimental settings. In particular, it shows clear improvements on both synthetic datasets and the ADNI dataset. These results suggest that the proposed method can be effective for missing-value imputation under Self-Masking MNAR.

  1. Interesting use of causal graph information.

The paper combines diffusion models with causal graph information for missing-value imputation. Using the Markov blanket to guide information sharing between variables is an interesting idea. The paper also provides theoretical analysis to support this design choice.

Critical weaknesses
  1. Limited explanation of how the method handles MNAR.

The model is trained using only observed values. However, under Self-Masking MNAR, observed and missing values can follow different distributions. It is unclear how the model learns to predict missing values accurately without directly modeling the missingness mechanism. The theoretical analysis also does not fully explain why the proposed diffusion model can overcome this problem. This raises concerns about the main claim of the paper.

  1. Insufficient evidence for the benefit of causal graph information.

The experiments do not clearly show how much the causal graph contributes to the performance improvement. In the synthetic experiments, the proposed method uses the ground-truth causal graph, while the baseline methods do not use the same information. Also, comparisons with random graphs or fully connected attention are missing. Therefore, it is difficult to determine whether the improvements come from the proposed method or from the additional graph information.

Questions for authors
  1. [Clarification] How does the proposed method learn to predict missing values under Self-Masking MNAR when the training loss uses only observed values? Could the authors explain why this training approach is expected to work when observed and missing values follow different distributions?

  2. [Additional experimental evidence] Could the authors compare the proposed method using the ground-truth causal graph with versions using random graphs, perturbed graphs, or fully connected attention? These comparisons would help clarify whether the causal graph structure is important for the reported improvements.

  3. [Clarification] Theorem 3.1 requires an additional covariance condition to show that the Markov blanket is better for missing cases. Is this condition expected to hold in the experimental settings? How does this theoretical result relate to the performance of the proposed diffusion model?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.