- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper proposes ResoPath, combining wavelet denoising, multiscale NCDEs, and observation-gap-conditioned FiLM modulation for irregular multivariate time series forecasting, and reports overall improvements over eighteen ODE/CDE baselines across five datasets.
Critical strengths
- Using separate frequency components to drive continuous-time dynamics provides a valuable modeling approach for handling trends, local fluctuations, and interpolation artifacts.
- Comparisons across five datasets show good overall performance, with particularly noticeable gains on Human Activity supporting the method’s practical potential.
Critical weaknesses
- The relationship between the future ODE in Section 3.4 and the linear temporal projection head in Table 3 is unexplained, leaving the forecasting architecture underlying the reported results and its continuous-time prediction capabilities unclear.
- Experiments compare only ODE/CDE models and omit cited, directly relevant methods such as APN, Hi-Patch, and TFMixer, so the evidence does not establish the claimed state-of-the-art performance for irregular time series forecasting overall.
Questions for authors
- Under Equation (15), when all channels are observed synchronously at every knot, both endpoints have zero gaps and unit recency, so how does MaDF capture declining reliability within a long intervening gap after interpolation?
- Is the Haar decomposition performed over observation indices or a uniform physical-time grid, and how does it avoid interpreting irregular sampling and zero-filled missing entries as genuine high-frequency changes?
- Could the authors provide total NFEs and end-to-end latency with units for Figure 3, clarifying whether these include wavelet processing, all NCDE branches, and future prediction, to establish the practical accuracy–cost tradeoff?
9 comments
W2: Comparison with Directly Relevant Non-ODE/CDE Baselines
Our initial experimental design focused on 18 ODE/CDE baselines to directly evaluate our core methodological objective: resolving interpolation distortions, scale entanglement, and uniform-reliability assumptions within continuous-time NCDEs. However, we agree with the reviewer that omitting cited non-ODE/CDE IMTS architectures—specifically APN (Liu et al., 2026), Hi-Patch (Luo et al., 2025), and TFMixer (Qiu et al., 2026) (as well as ReIMTS (Li et al., 2026) noted by Reviewer xBzy)—left our broader SOTA claims across general IMTS forecasting insufficiently substantiated.
We have expanded our benchmark evaluation under the exact same data splits and preprocessing protocol, to include these directly relevant IMTS methods. The comparison results (MSE ) are reported below. Bold indicates the best (lowest mean) result; italic indicates the second-best.
| Algorithm | USHCN | Human Activity | PhysioNet'12 | MIMIC-III | MIMIC-IV |
|---|---|---|---|---|---|
| PrimeNet | 7.28 ± 0.10 | 15.33 ± 0.12 | 10.18 ± 0.10 | 10.94 ± 0.08 | 9.57 ± 0.29 |
| SeFT | 6.13 ± 0.32 | 8.47 ± 0.08 | 8.55 ± 0.01 | 10.35 ± 0.06 | 7.72 ± 0.38 |
| NeuralFlows | 5.60 ± 0.25 | 2.56 ± 0.05 | 5.32 ± 0.02 | 8.29 ± 0.05 | 5.86 ± 0.27 |
| GNeuralFlow | 4.47 ± 0.54 | 2.61 ± 0.06 | 4.90 ± 0.03 | 7.07 ± 0.05 | 6.13 ± 0.12 |
| tPatchGNN | 4.15 ± 0.31 | 1.04 ± 0.06 | 4.34 ± 0.02 | 6.37 ± 0.04 | 4.86 ± 0.16 |
| Hi-Patch | 4.34 ± 0.80 | 1.39 ± 0.02 | 4.27 ± 0.05 | 5.46 ± 0.03 | 4.48 ± 0.23 |
| Warpformer | 3.77 ± 0.62 | 1.14 ± 0.03 | 4.08 ± 0.01 | 5.13 ± 0.02 | 4.54 ± 0.22 |
| TFMixer | 3.24 ± 0.69 | 0.96 ± 0.05 | 3.73 ± 0.00 | 6.02 ± 0.02 | 4.06 ± 0.14 |
| APN | 3.16 ± 0.58 | 1.32 ± 0.04 | 3.69 ± 0.01 | 5.30 ± 0.03 | 4.12 ± 0.18 |
| ReIMTS | 3.06 ± 0.02 | 0.72 ± 0.01 | 3.43 ± 0.01 | 5.15 ± 0.02 | 3.75 ± 0.09 |
| ResoPath (ours) | 3.48 ± 0.74 | 0.83 ± 0.05 | 3.60 ± 0.00 | 5.36 ± 0.05 | 3.73 ± 0.20 |
Even against top-performing discrete patch- and frequency-mixing architectures, ResoPath demonstrates SOTA competitive performance: it achieves 1st place on MIMIC-IV and ranks 2nd place on both Human Activity and PhysioNet'12, consistently outperforming Hi-Patch, APN, TFMixer, and tPatchGNN.
We will incorporate these non-ODE/CDE baselines into a new Appendix G (with full implementation details in Appendix E), and refine the claims in the Abstract, Introduction, and Conclusion to precisely highlight our architectural advances within the continuous-time family.
Q1: Missingness Descriptor Construction under Synchronous Observations (Eq. 15)
We agree that Eq. (15) in Appendix B was overly compressed and omitted key details of our interval-wise descriptor construction. Our actual implementation avoids the degenerate zero-gap / unit-recency collapse through FD-spaced auxiliary knots combined with one-sided pre/post-reset limits:
-
Auxiliary Descriptor Knots within Inter-Observation Intervals: Consider two consecutive synchronous observations across all channels at and with no intervening measurements in . Whenever , the Missingness Generator (Figure 2) inserts temporary auxiliary evaluation knots spaced according to the channel-specific Freedman–Diaconis (FD) timescale defined in Eq. (14). At each auxiliary knot , descriptors are computed from the latest actual observation at :
Crucially, these auxiliary knots serve solely as descriptor-evaluation points (), not new measurements, and therefore do not reset elapsed time.
-
Pre-Reset Left Limits vs. Post-Reset Right Limits: Because elapsed time resets discontinuously upon observation, interpolation over is anchored at the post-reset right limit at () and terminates at the pre-reset left limit at immediately prior to assimilating the measurement at :
The post-reset values and apply only to the subsequent interval after the observation at is incorporated.
-
Interval-Wise Piecewise-Linear Interpolation: Piecewise-linear interpolation is constructed separately over the augmented knot sequence within each inter-observation interval , rather than connecting directly to . Consequently, strictly increases and exponentially decays across , faithfully capturing long unobserved gaps even under synchronous sampling. We will rewrite Eq. (15) and Appendix B to explicitly reflect the one-sided limits and auxiliary knot set.
We sincerely thank you for the constructive feedback of our formulation and experimental evaluation. Below we address each weakness and question in detail.
W1: Forecast-Head Specification & Reconciliation of Section 3.4 with Table 3
We clarify that all reported results in Tables 1–2 were produced using the autonomous latent ODE rollout followed by a point-wise output readout, as described in Section 3.4. The phrase "Shared linear temporal projection from the input length to the forecasting horizon" in Table 3 (Line 959) was an editing remnant from an early discrete-grid baseline configuration table and does not reflect the ResoPath forecast head.
Specifically, our forecasting stage operates as follows:
Cross-Scale Terminal State Fusion (Eq. 5): At the historical boundary T, the terminal states of the J+1 scale-specific NCDE branches, ZT=[z(0)(T),…,z(J)(T)]∈R(J+1)dh, are fused via a learnable 1×1 scale aggregation into the initial forecasting state zˉ(T)∈Rdh.
Future Autonomous Latent ODE Rollout (Reconciling Line 279 with Eq. 8): Because no control observations arrive during (T,T+H], the matrix-valued historical field in Eq. (8) (∈Rdh×(D+1)) cannot be applied unchanged. Our statement in Line 279 ("where Fθ(⋅) is obtained from Eq. (8)") meant that the future drift field gψ:Rdh×Rdm→Rdh retains the same FiLM-modulated residual backbone as Eq. (8), but replaces the matrix reshaping projection with a vector-valued output projection into Rdh. To account for growing uncertainty as the forecast horizon extends without new measurements, we track the elapsed observation age a(t)∈RD, yielding an autonomous system in the augmented state (u(t),a(t)) for t∈[T,T+H]:
\qquad \dot{a}(t) = \mathbf{1}_D, \qquad u(T) = \bar{z}(T), \quad a(T) = \delta(T).$$ Since no new observations are assimilated after $T$, $a(t) = \delta(T) + (t - T)\mathbf{1}_D$ increases linearly in closed form without resetting, and $\eta_{\mathrm{fut}}(a(t)) = E_m(M(a(t)))$ continuously informs the future vector field of the accumulating forecast horizon.Arbitrary-Horizon Querying & Point-wise Readout: Given any sorted set of future query timestamps tk=T+τkk=1K⊂(T,T+H], the ODE solver evaluates u(tk)∈Rdh directly along the continuous trajectory, and a shared readout πη maps each queried latent state independently to the D observable channels:
\qquad W_{\mathrm{out}} \in \mathbb{R}^{D \times d_h}.$$The identical rollout and readout mechanism is used during both training and inference. We will revise Section 3.4 (fixing Line 279) and update the
Prediction headentries in Table 3 so that all specifications and dimensions are completely consistent.