acceptodds
Withdrawn from ICLR 2027

A Unified Framework for Vertical Federated Learning with Partially Missing Features and Locally Independent Inference

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Vertical Federated Learning enables collaborative training across clients holding disjoint feature subsets of data samples. Existing methods often treat a client's local features as either fully available or entirely missing, and thus fail to make full use of partially observed features. Moreover, standard VFL typically relies on joint inference involving all participating clients, limiting the ability of an individual client to make predictions independently. To address these limitations, we propose XVFL, a unified framework that uses non-aligned data with partially missing features and supports locally independent inference for each client. Specifically, XVFL introduces two modules: Cross Completion (XCom), which reconstructs missing local features for non-aligned samples, and Decision Subspace Alignment (DSAlign), which aligns local, completed, and joint representations in the decision subspace to improve independent inference. Furthermore, we establish convergence guarantees for XVFL, proving an rate for SGD-type and for PAGE-type algorithms, where denotes the number of training steps. Extensive experiments on real-world datasets demonstrate that XVFL significantly outperforms existing methods across different feature missing rates, e.g., achieving an improvement of around 43 percentage points on the medical MIMIC-III dataset, validating XVFL's practical effectiveness and robustness.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Vertical Federated Learning, Partially Missing Features, Locally Independent Inference

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_YQnDOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This work proposes a new VFL model called X-VFL which handles non-aligned data with missing features and supports locally independent inference.

Critical strengths

This work proposes a new VFL architecture to handle non-aligned data with missing features in VFL. Theoretical analysis on convergence and experimental results on performance are provided to validate the model.

Critical weaknesses

The demonstrated scope is narrow and a comparison study with the SOTA models is not included. There are some concerns on on the proposed model, for instance, X~a\tilde{X}_a and E~a\tilde{E}_a as explained in questions.

Questions for authors
  1. There are three missing mechanisms in general [1], missing completely at random, missing at random, and missing not at random for the training and test datasets. Not all of these scenarios are considered in this work, so the model and its evaluation are limited. Moreover, the proposed model is applied to labeled datasets and no explanation is provided on how to extend it to unlabeled data, while recent works on VFL consider unlabeled as well as labeled datasets. [1] Roderick JA Little and Donald B Rubin. Statistical analysis with missing data, volume 793. John Wiley & Sons, 2019.

  2. The comparison study is limited. For instance, it is necessary to compare the proposed model with the following two recent models: [2] Yuanqin He, Yan Kang, Xinyuan Zhao, Jiahuan Luo, Lixin Fan, Yuxing Han, and Qiang Yang. A hybrid self-supervised learning framework for vertical federated learning. IEEE Transactions on Big Data, 2024. [3] K. Hong et. al, Deep Latent Variable Model Based Vertical Federated Learning with Flexible Alignment and Labeling Scenarios, In International Conference on Learning Representation 2026.

  3. In the two client setting, X~a\tilde{X}_a is computed from client B and the components of X~a\tilde{X}_a for missing features are adopted to obtain E~a\tilde{E}_a. So, the resulting E~a\tilde{E}_a seems to be more dependent on EbE_b from the other client B than the client A's features. Moreover, the average (E~a+Eb)/2(\tilde{E}_a + E_b)/2 is the input for the function hh, meaning that the output from hh is more dependent on client B and the features of client A seem to affect the output of hh relatively less. Since the existing features of client A are important information, does the proposed method make sense from this perspective? Do we need to consider a weighted average instead of the equal-weight average? The features of some clients may be more important than those of the other clients.

  4. In (3) both EaE_a and E~a\tilde{E}_a are used, but E~a\tilde{E}_a is generated with the completed features X~a\tilde{X}_a. Recall that X~a\tilde{X}_a is obtained by adopting from the output of XCom_a for the missing features in the data. If Ea=fa(x)E_a = f_a(x) is possible or all features of client A exist, then E~a\tilde{E}_a has no adoptation from XCom and both are the same. In this case, is it meaningful to consider both EaE_a and E~a\tilde{E}_a? If the reviewer misunderstands this, a detailed explanation would be helpful.

  5. In (3) h(E~a)h(\tilde{E}_a) and h(Ea)h(E_a) are used as the inputs of the similarity loss function, but it is also possible to consider E~a\tilde{E}_a and EaE_a as the inputs of the same similarity loss function. Obviously, we can train XCom in the latter case.

Reviewer_dws1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies vertical federated learning with partially observed local features and locally independent inference. XCom completes missing raw features from other clients' embeddings while retaining observed entries, and DS-Align encourages output consistency between original/completed and local/joint predictions. The framework uses averaged embeddings and reports results on six datasets under multiple missingness levels and two- and four-client settings, together with SGD and PAGE convergence analyses.

Critical strengths
  1. Retaining observed feature entries while completing the missing components is a sensible extension of block-missing VFL. Combining this with local prediction objectives addresses a useful availability problem.
  2. The evaluation covers six datasets, varying missingness, overlap and data imbalance, and multiple clients. The five-seed results and module-level ablations provide evidence of gains, particularly for independent inference under substantial missingness.
Critical weaknesses
  1. The algorithmic contribution is insufficiently established. Averaging, cross-view completion, and joint-to-local transfer have precedents, while the ablations do not isolate the benefit of raw-feature completion or the two alignment losses. Moreover, Eq. (1) requires another client's embedding, leaving independent inference with missing features unclear when other clients are unavailable.

  2. The stated conditions do not justify the convergence rates. Theorem 1 only upper-bounds η\eta, leaving the 1/(ηT)1/(\eta T) term uncontrolled. In Theorem 2, η≤1/(2β)\eta \leq 1/(2\beta) does not justify the descent step for all b′≤bb' \leq \sqrt{b}. Equation (29) also retains the noise floor σ2/b\sigma^2/b, so the claimed rate requires explicit batch-size dependence on the target accuracy.

Questions for authors
  1. In Fig. 5b, where does the embedding supplied to XCom originate for a new test sample? Can inference with missing local features proceed when all other clients are unavailable? Please specify any communication or cached information and the corresponding information available to each baseline.
  2. Can matched comparisons isolate raw-feature versus representation completion and the two DS-Align losses separately? Please keep observed entries, supervision and training budgets comparable, and report the effects on local and joint accuracy separately.
Reviewer_KYs1OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes X-VFL, a vertical federated learning framework designed to handle non-aligned samples with partially missing features while supporting locally independent inference. The framework introduces two modules: Cross Completion (XCom), which reconstructs missing local features using embeddings from other clients, and Decision Subspace Alignment (DS-Align), which aligns predictions from local, reconstructed, and aggregated representations. Experiments on six real-world datasets evaluate the proposed method under varying feature missing rates, sample overlap ratios, and client configurations.

Critical strengths

Practical and relevant problem setting. The paper addresses partially missing features in VFL, moving beyond conventional block-wise missingness. The proposed framework combines feature completion and decision-space alignment to better utilize incomplete data.

Critical weaknesses

1. The distributed training protocol and its communication overhead are insufficiently specified.

  • Forward communication: XCom requires embeddings from other clients to reconstruct missing features. For example, if client B has partially missing features, client A must first upload its embedding to the server, which then forwards it to B for reconstruction. B must subsequently upload the completed embedding to the server for loss computation. These additional communication steps are not clearly described or accounted for.
  • Backward communication: Cross-client feature completion introduces additional gradient dependencies. For example, since B's XCom takes A's embedding as input, backpropagation through B's XCom produces gradients that must be transmitted back to A through the server. The paper does not clarify how these cross-client gradients are propagated, raising concerns about the claimed communication efficiency and convergence analysis.

2. The claimed locally independent inference is not fully supported by the proposed method or experiments.

When a client encounters missing features during inference, its XCom module still requires embeddings from other clients. For example, client B cannot reconstruct its missing features without obtaining client A's embedding for the same test sample. Therefore, the proposed method cannot perform the claimed independent inference when other clients are unavailable. This issue also affects the experimental evaluation, where X-VFL appears to leverage cross-client information during supposedly independent inference, while competing methods may be restricted to local information. Such potentially unequal information access undermines the fairness of the comparisons and the validity of the claimed improvements in independent inference.

3. The implementation is not publicly available.

The paper provides no code repository or implementation link. Given the unspecified distributed communication protocol and the complexity of cross-client gradient propagation, the absence of code further limits the reproducibility and verifiability of the reported results.

Questions for authors

1. Distributed Training Protocol: Could the authors clarify the complete distributed training protocol, particularly the additional embedding exchanges and cross-client gradient propagation introduced by XCom? Beyond the aggregate communication costs reported in Table 3, could the authors provide a quantitative breakdown of the communication rounds, sequential communication stages, and transmitted data volumes, explicitly accounting for both forward and backward propagation?

2. Independent Inference: How can X-VFL perform independent inference with missing features when XCom requires embeddings from other clients? Could the authors provide separate experimental results for strictly independent inference without cross-client communication and XCom-assisted inference with cross-client information, along with comparisons against baselines under equivalent test-time information access? Reporting the additional inference communication costs would also help clarify the practical applicability of the proposed method.

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

Related papers

Discussion (0)

Sign in to comment.