- Rating
- 1
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper proposes GapFlow, a flow-matching framework for multivariate time-series imputation, motivated by the limitations of per-variable interpolation priors under contiguous missingness. GapFlow combines a cross-variable conditional prior based on ridge-regularized channel covariance, a probability path adapted to the two-sided distance from observations, and iterative self-conditioning. The paper evaluates the method across six datasets, four missingness mechanisms, two missing ratios, and multiple window lengths, reporting the lowest MAE among the external baselines in all 72 settings, with larger gains over the interpolation-prior counterpart under block missingness than under MCAR.
Critical strengths
The paper motivates using observed peer channels to improve imputation within contiguous gaps, where a variable’s own observations may provide limited information. This focus addresses a meaningful challenge in multivariate time-series imputation and makes the proposed direction worth investigating.
Critical weaknesses
W1. The paper does not sufficiently establish the validity and scope of its central problem formulation. FM-Interp performing worse than a Transformer does not demonstrate that interpolation priors themselves cause “active damage.” The comparison does not isolate the prior’s effect, and the conditions under which this claimed failure occurs remain insufficiently characterized. This weakens the foundation motivating the proposed method.
W2. Section 3.5 describes recomputing the cross-variable conditional prior using pseudo-observations. However, the pseudo-observation branch in sample_selfcond() passes an all-observed mask, disabling cross-variable mixing and overwriting the mean with pseudo-observations. The resulting operation blends predictions with a new prior sample, undermining the stated mechanism and interpretation of its ablation gains.
W3. Setting recovers the interpolation mean, but adaptive variance, path scheduling, and self-conditioning can remain active. The paper should clarify whether the exact-reduction claim concerns only the prior mean or the complete method, and specify the conditions required for the latter.
Questions for authors
Q1. In sample_selfcond(), the pseudo-observation branch passes an all-observed mask to prior.sample(). This makes the distance and cross-variable mixing weight zero, while the mean is overwritten by pseudo-observations. Thus, this branch returns pseudo-observations plus noise rather than the cross-variable conditional update described in Section 3.5. How do the authors reconcile this discrepancy? Did the reported self-conditioning results use this released implementation or a different version?
Q2. Is there a controlled comparison between Gaussian and interpolation priors with the network, path, training protocol, and sampling budget fixed? What evidence isolates the interpolation prior as the cause of “active damage” under block missingness?
Q3. Does the independence-based reduction claim concern only the mean in Equation (4), or the complete method? If the latter, under what conditions do the adaptive variance, probability path, and self-conditioning also reduce to FM-Interp?
If these concerns, particularly the implementation discrepancy, are satisfactorily addressed, I would consider increasing my rating.