Pith. sign in

REVIEW 3 major objections 5 minor 7 references

X-VFL: A New Vertical Federated Learning Framework with Cross Completion and Decision Subspace Alignment

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read X-VFL is a vertical federated learning framework that handles partially missing features and lets each client predict locally on new samples, using cross-completion to reconstruct absent inputs and decision-subspace alignment to keep solo p

desk verdict Headline capability breaks: 'independent inference with missing features' has no input for XCom at inference time; the partial-missing setting is novel but the paper needs major changes. read the letter →

arxiv 2508.05568 v2 pith:T65DPFU2 submitted 2025-08-07 cs.LG cs.CVcs.DCmath.OC

Qinghua Yao , Xiangrui Xu , Zhize Li This is my paper
classification cs.LGcs.CVcs.DCmath.OC
keywords verticalfederatedlearningpartiallymissingfeaturescrosscompletiondecisionsubspacealignmentlocallyindependentinferenceconvergenceratesplit
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Vertical federated learning usually requires every client to hold complete features for every aligned sample, and every client to join every prediction. This paper claims both constraints can be dropped. X-VFL trains on non-aligned samples whose features are only partially present, and lets a single client predict on new data by itself. The mechanism is to reconstruct a client's missing features from the averaged embeddings of the other clients (XCom), then pull each client's local decision into the same subspace as the joint averaged decision (DS-Align). The paper reports convergence rates of $O(1/\sqrt{T})$ for SGD-type training and $O(1/T)$ for PAGE-type training, and accuracy gains such as 15% on CIFAR-10 and 43% on MIMIC-III over the tested baselines.

What carries the argument

The framework runs on three coupled pieces: average aggregation $E_{\mathrm{avg}} = \frac{1}{k}\sum_i E_i$, cross-completion $\tilde{X}_i = XCom_i(E_{-i}/(k-1))$, and decision-subspace alignment losses $L_{\mathrm{DSAlign1}} = \sum_i \ell(h(\tilde{E}_i), h(E_i))$ and $L_{\mathrm{DSAlign2}} = \sum_i \ell(h(E_i), h(E_{\mathrm{avg}}))$. Average aggregation is what makes solo inference possible: each local embedding has the same dimension as the input to the shared top model $h$, so one client can run $h$ by itself. XCom is what extends this to missing features; DS-Align is the regularizer that keeps solo predictions consistent with collaborative ones.

What would settle it

Take a trained X-VFL model, mask 90% of one client's features on a new test sample, and then cut all communication so no other client sends its embedding $E_j$ at query time. If the client can still evaluate $\tilde{X}_i = XCom_i(E_{-i}/(k-1))$ and produce a prediction, the claim holds; if it cannot produce a reconstruction without $E_{-i}$, independent inference with missing features reduces to collaborative inference or requires precomputed imputations.

Watch

Extended reading notes

Core claim

The paper's central claim is that VFL can be restructured so that non-aligned samples with partially missing features are trainable, and a client can predict locally using only its own model and the shared top model, with no cross-client communication. The carrying mechanism is average aggregation $E_{\mathrm{avg}} = \frac{1}{k}\sum_i E_i$ instead of concatenation, so each local embedding alone is a valid input to the top model $h$. XCom reconstructs missing local features as $\tilde{X}_i = XCom_i(E_{-i}/(k-1))$, and DS-Align regularizes the model so $h(E_i)$, $h(\tilde{E}_i)$, and $h(E_{\mathrm{avg}})$ agree in the decision subspace. Across six datasets the paper shows X-VFL retains accurac

Load-bearing premise

The framework's independent-inference-with-missing-features mode assumes a client can obtain the other clients' averaged embeddings it needs to reconstruct its missing features, even when that client is otherwise operating on its own.

Editorial extensions

If this is right

  • Training can use non-aligned samples with partial missing features, so VFL no longer discards data whenever one client's features are incomplete.
  • A single client can serve new samples locally, making real-time or single-site deployment feasible and removing the need for all clients to be online during inference.
  • Because aggregation is averaging, any number of clients can be combined at collaborative inference time without changing the top model's input dimension.
  • SGD-style training reaches $\epsilon$-stationarity in $O(1/\epsilon^4)$ communication rounds; PAGE-style training improves this to $O(1/\epsilon^2)$.
  • At missing rate $R_{\mathrm{miss}}=0.9$, X-VFL reports under 0.5% accuracy loss on UTKFace relative to no missing data, and under 0.2% gap between independent and collaborative modes.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because the decision head consumes averaged embeddings, X-VFL could in principle serve any subset of clients at inference time; the paper demonstrates one-client and all-client cases, but not arbitrary partial subsets.
  • The same cross-completion recipe could apply to other distributed settings where one active client reconstructs the contribution of inactive partners instead of waiting for them.
  • A deployment detail the paper does not specify is where the partner embeddings feeding XCom come from at independent-inference time—fetched on demand, cached from training, or served by the server—so the claimed mode has overhead beyond the local model in practice.
  • The reported convergence rates are for the joint training loss; a testable diagnostic for the solo-inference claim is to track $\mathbb{E}[\ell(h(E_i), h(E_{\mathrm{avg}}))]$ on held-out samples and use it to decide when solo inference is trustworthy.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. X-VFL is a vertical federated learning framework that claims to handle non-aligned samples with partially missing features and to support locally independent inference at each client, including a mode where a client reconstructs missing features without cross-client communication. The framework combines a Cross Completion module (XCom), which reconstructs a client's missing features from another client's embeddings (Eq. 1), with a Decision Subspace Alignment module (DS-Align), which aligns local embeddings with joint averaged embeddings through the losses in Eqs. (3) and (4). The overall objective is Eq. (5). The paper provides convergence theorems for SGD-type and PAGE-type training algorithms (Theorems 1 and 2) and reports experiments on six datasets, across varying missing rates, overlap ratios, and data imbalance, plus a four-client extension in Appendix D.

Significance. The paper addresses a genuinely practical limitation of vertical federated learning: standard VFL requires sample alignment across clients and coordinated inference. The explicit formulation of partially missing features, the average-aggregation design that makes single-client inference possible in the no-missing-features case, and the DS-Align idea of aligning decision-subspace outputs are worth attention. The experiments cover multiple datasets and settings, and the convergence proofs in the appendix are standard and appear correctly stated. However, the central advertised capability --- locally independent inference with missing features, Mode 2 in Section 2.2 --- is not actually supported by the definitions in the paper. Because this claim is load-bearing for the abstract, the contributions, and the experimental evaluation, the manuscript in its current form does not establish its main contribution.

major comments (3)
  1. [Section 2.2, Eq. (1), Eq. (28), Fig. 5(b)] Mode 2 ('Independent inference with missing features') cannot be executed from the equations given. Eq. (1) defines ~X_a = XCom_a(E_b), where E_b is the other client's embedding, and the k-client extension in Eq. (28) defines ~X_i = XCom_i(E_{-i}/(k-1)), again requiring the sum/combination of all other clients' embeddings. In Mode 2, the client is described as producing predictions independently, with no cross-client communication during inference. For a new test sample, an isolated client has no access to E_b or E_{-i}. No alternative input to XCom is specified in Fig. 5(b), the text, or the loss formulation. The forward pass for this inference mode is therefore undefined. The advertised capability reduces, as defined, to collaborative inference with imputation using other clients' embeddings, not locally independent inference with missing features.
  2. [Section 3, Theorems 1 and 2] The convergence theorems are generic SGD/PAGE statements for any smooth stochastic objective satisfying Assumptions 1 and 2. They do not use any property of XCom, DS-Align, or the X-VFL loss, and they do not establish convergence to a model that supports the independent-inference-with-missing-features mode. In particular, the bounds in Eqs. (17) and (25) are the standard O(1/sqrt(T)) and O(1/T) rates and are unrelated to the feasibility of Mode 2. As stated, the section overstates the theoretical support for X-VFL; the results should be framed as standard optimization bounds that apply to Eq. (5) if the smoothness and variance assumptions hold, without implying the framework's central capability.
  3. [Section 4, Figs. 6-8] The experimental evaluation does not specify how XCom is invoked in the independent-inference-with-missing-features mode. Without a protocol that supplies the inputs required by Eq. (1) or Eq. (28), the reported X-VFL accuracy at Rmiss > 0 in the 'Independent' panels is not reproducible. In addition, the very small independent-vs-collaborative gap reported in Fig. 8 is partly a direct consequence of LDSAlign2 in Eq. (4), which trains h(E_i) to match h((E_a+E_b)/2); this should be acknowledged explicitly so the 'gap narrowing' is not interpreted as an emergent property of the framework.
minor comments (5)
  1. [Throughout] There are several typos: 'thorems' in the Section 3 heading, 'unbaised' in Assumption 2, 'estabilished' in the conclusion, 'variace' in Appendix A.2, 'actived' in Appendix D.1, and 'to to' in Appendix B.
  2. [Abstract and Section 4] The abstract claims a 15% accuracy improvement on CIFAR-10 and a 43% improvement on MIMIC-III, but Section 4.1 presents these results only through figures and does not state the exact comparison points or standard deviations. Reporting the precise settings and error bars would make the headline numbers verifiable.
  3. [Section 4.3, Fig. 10] The figure legend uses 'Standalone' while the text and related baselines refer to 'Vanilla Standalone'. The notation should be consistent.
  4. [Section 1.1] The claim of being 'the first to introduce a practical setting with partially missing features in VFL' is stronger than the related-work discussion supports, since only a small number of recent methods are surveyed. A broader comparison to imputation-based and missing-feature VFL work would be needed to substantiate this priority claim.
  5. [Appendix D.1, Eq. (27)] The notation in Eq. (27) mixes sums and averages: E' is a sum over M, while the loss term divides by k when combining E' with ~E'. Clarifying the intended averaging convention and the consistency between the two-client and k-client formulations would improve readability.

Circularity Check

1 steps flagged · score 6.0 of 10

Mode 2 'independent inference with missing features' is self-definitional: the only defined XCom input is other clients' embeddings, so the independence claim reduces to collaboration.

  1. self definitional [Section 2.2 (Mode 2), Section 2.1 Eq. (1), Appendix D Eq. (28)]
    "2. Independent inference with missing features (Fig. 5b): Each client leverages its XCom module to reconstruct missing features and produce predictions independently. ... ˜Xa = XComa(Eb), ˜Xb = XComb(Ea), (1) ... ˜Xi = XComi( E−i/k−1 ), 1 ≤ i ≤ k, (28)"

    Mode 2 promises locally independent inference with no cross-client communication at inference time. But the only definition of XCom takes as input an embedding produced by another client: Eq. (1) uses E_b = f_b(X_b), and Eq. (28) uses E_{-i}/(k-1), the sum of embeddings from all other clients. For a new test sample, an isolated client with missing local features has no access to E_b or E_{-i}, so its XCom has no defined input. The advertised capability is therefore not a derived, independent result; executing Eq. (1) at inference requires exactly the cross-client information the mode claims to remove. The mode reduces by construction to collaborative inference with imputation, making the 'independent inference with missing features' claim circular rather than merely underspecified.

full rationale

Aside from this central self-definitional gap, the paper does not exhibit other circular derivations. Theorems 1 and 2 are standard SGD/PAGE convergence analyses applied to the smooth nonconvex objective in Eq. (6); no fitted constants or X-VFL-specific predictions are fed back into the convergence rates. The empirical gains are measured against external datasets and baselines, not produced by an equation that assumes the reported accuracies. The PAGE lemmas cited in Appendix A.2 come from prior published work and are general-purpose, so the author-overlap citation is not load-bearing in a circular sense. The DS-Align2 term in Eq. (4) explicitly trains individual predictions toward the averaged prediction, so the reported small gap between independent and collaborative accuracy is a trained property rather than a surprising empirical discovery; I treat this as a minor secondary issue. The central Mode-2 claim is self-definitional because the mechanism offered to enable independence is defined in terms of the very cross-client information independence excludes, so the score is 6 rather than 0-2.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

Two hyperparameters are tuned per dataset, and the reconstruction premise is the primary extra assumption. No new physical entities are introduced; XCom and DS-Align are algorithmic modules. The convergence theorems rely on generic smoothness and variance assumptions rather than any X-VFL-specific structure.

free parameters (2)
  • lambda_1 (DS-Align1 weight) = Per-dataset grid search over Table 1 ranges, e.g., CIFAR-10: 0.01 to 0.5; MIMIC-III: 5e-4 to 1e-2
    Hyperparameter in Eq. (5) controlling the reconstruction-alignment loss; selected separately for each dataset and affects all reported accuracies.
  • lambda_2 (DS-Align2 weight) = Per-dataset grid search over Table 1 ranges, e.g., CIFAR-10: 1e-5 to 5e-4; MIMIC-III: 0.02 to 0.1
    Hyperparameter in Eq. (5) controlling the independent-inference alignment loss; tuned per dataset, with final values not reported.
assumptions (4)
  • standard math Stochastic gradients of L in Eq. (5) satisfy average smoothness and bounded variance (Assumptions 1 and 2).
    Invoked in Theorems 1 and 2; these are standard nonconvex finite-sum assumptions, not verified for the X-VFL loss in Section 3.1.
  • domain assumption XCom modules can reconstruct missing features from other clients' embeddings.
    Eq. (1) and Appendix D Eq. (28) define completion this way; this is the premise that conflicts with locally independent inference in Section 2.2.
  • domain assumption The server holds labels and can average client embeddings during training.
    Standard VFL setup used throughout Section 2.1 and Eq. (2), with labels on the server and a shared top model h.
  • domain assumption Missingness in experiments is simulated by masking random feature proportions.
    Appendix B applies masks with rates Rmiss to non-aligned samples; no structural missingness model or validation on naturally missing data is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of X-VFL: A New Vertical Federated Learning Framework with Cross Completion and Decision Subspace Alignment." pith.science (2026). https://pith.science/paper/T65DPFU2

@misc{pith2026250805568,
  author       = {Pith},
  title        = {Pith review of: X-VFL: A New Vertical Federated Learning Framework with Cross Completion and Decision Subspace Alignment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/T65DPFU2}},
  note         = {Machine review of arXiv:2508.05568}
}
abstract

Vertical Federated Learning (VFL) enables collaborative learning by integrating disjoint feature subsets from multiple clients/parties. However, VFL typically faces two key challenges: i) the requirement for perfectly aligned data samples across all clients (missing features are not allowed); ii) the requirement for joint collaborative inference/prediction involving all clients (it does not support locally independent inference on a single client). To address these challenges, we propose X-VFL, a new VFL framework designed to deal with the non-aligned data samples with (partially) missing features and to support locally independent inference of new data samples for each client. In particular, we design two novel modules in X-VFL: Cross Completion (XCom) and Decision Subspace Alignment (DS-Align). XCom can complete/reconstruct missing features for non-aligned data samples by leveraging information from other clients. DS-Align aligns local features with completed and global features across all clients within the decision subspace, thus enabling locally independent inference at each client. Moreover, we provide convergence theorems for different algorithms used in training X-VFL, showing an $O(1/\sqrt{T})$ convergence rate for SGD-type algorithms and an $O(1/T)$ rate for PAGE-type algorithms, where $T$ denotes the number of training update steps. Extensive experiments on real-world datasets demonstrate that X-VFL significantly outperforms existing methods, e.g., achieving a 15% improvement in accuracy on the image CIFAR-10 dataset and a 43% improvement on the medical MIMIC-III dataset. These results validate the practical effectiveness and superiority of X-VFL, particularly in scenarios involving partially missing features and locally independent inference.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

7 extracted references · 5 canonical work pages

  1. [1]

    Self-supervised vertical federated learning

    Timothy Castiglia, Shiqiang Wang, and Stacy Patterson. Self-supervised vertical federated learning. In Workshop on Federated Learning: Recent Advances and New Challenges, in Conjunction with 12 NeurIPS 2022 , 2022a. Timothy J Castiglia, Anirban Das, Shiqiang Wang, and Stacy Patterson. Compressed-VFL: Communication-efficient learning with vertically partit...

    work page 2022

  2. [2009]

    Vertical semi-federated learning for efficient online advertising

    Wenjie Li, Qiaolin Xia, Hao Cheng, Kouyin Xue, and Shu-Tao Xia. Vertical semi-federated learning for efficient online advertising. arXiv preprint arXiv:2209.15635 , 2022a. 13 Wenjie Li, Qiaolin Xia, Junfeng Deng, Hao Cheng, Jiangming Liu, Kouying Xue, Yong Cheng, and Shu-Tao Xia. VFed-SSD: Towards practical vertical federated advertising. arXiv preprint a...

    work page arXiv 2006

  3. [2016]

    Vertical federated knowledge transfer via repre- sentation distillation for healthcare collaboration networks

    Chung-ju Huang, Leye Wang, and Xiao Han. Vertical federated knowledge transfer via repre- sentation distillation for healthcare collaboration networks. In Proceedings of the ACM Web Conference 2023, pages 4188–4199,

    work page 2023

  4. [2018]

    Personalized federated learning with feature alignment and classifier collaboration

    Jian Xu, Xinyi Tong, and Shao-Lun Huang. Personalized federated learning with feature alignment and classifier collaboration. arXiv preprint arXiv:2306.11867 ,

    arXiv

  5. [2021]

    According to simplified PAGE update step Eq

    effectively reduces the variance and thus largely decreases the total number of communication rounds (Theorem 2). According to simplified PAGE update step Eq. (10) and Lemma 4 of PAGE (Li et al., 2021), we have the following variance reduction lemma: Lemma 1 Suppose that Assumptions 1 and 2 hold. For the probabilistic gradient estimator gt defined in Eq. ...

    work page 2021

  6. [2022]

    and FedCVT (Kang et al., 2022), that are relevant for comparison, as discussed in Section 1.2. Concretely, we compare our X-VFL with these four baselines for both independent Inference and collaborative inference modes, using test classification accuracy as the performance metric. All results are averaged over five random seeds to ensure robustness. 17 C ...

    work page 2022

  7. [2025]

    Split learning for health: Distributed deep learning without sharing raw patient data

    Praneeth Vepakomma, Otkrist Gupta, Tristan Swedish, and Ramesh Raskar. Split learning for health: Distributed deep learning without sharing raw patient data. arXiv preprint arXiv:1812.00564,

    arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.