- Rating
- 3(+1)
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes QiYao-I, a foundation model for irregular multivariate time series forecasting. It introduces sampling-conditioned temporal manifold attention to capture irregular temporal dependencies and learnable spectral representations to guide asynchronous cross-variable interactions. Experiments on 8 datasets demonstrate competitive zero-shot and few-shot forecasting performance compared with existing foundation models and irregular forecasting methods.
Critical strengths
-
The paper addresses an important and practical problem by jointly modeling irregular temporal dependencies and asynchronous cross-variable interactions.
-
Incorporating local sampling structures into temporal attention is an interesting design that enables the model to distinguish different sampling patterns.
-
The experiments are relatively comprehensive, covering 8 datasets with comparisons against foundation models and irregular forecasting methods, along with ablation and efficiency analyses.
Critical weaknesses
-
The necessity of the temporal manifold formulation is not fully justified. Although the ablation studies demonstrate its effectiveness, comparisons with simpler timestamp-distance-based attention biases using the same sampling-conditioned bandwidth would better isolate the contribution of the learned manifold formulation.
-
The spectral representations may be sensitive to irregular sampling density. In Section 3.3, the global spectral coefficients are computed by averaging point-level projections without explicitly accounting for non-uniform sampling intervals. Thus, the representations may vary with the sampling distribution even for the same underlying signal. Further analysis would help establish their robustness to different sampling patterns.
-
The robustness to varying levels of sampling irregularity could be further evaluated. Although the paper considers different sampling patterns, it remains unclear whether QiYao-I maintains its advantages over existing methods as the degree of sampling irregularity increases or decreases. Controlled experiments under different irregularity levels would better demonstrate its robustness.
Questions for authors
-
The necessity of the temporal manifold formulation is not fully justified. Although the ablation studies demonstrate its effectiveness, comparisons with simpler timestamp-distance-based attention biases using the same sampling-conditioned bandwidth would better isolate the contribution of the learned manifold formulation.
-
The spectral representations may be sensitive to irregular sampling density. In Section 3.3, the global spectral coefficients are computed by averaging point-level projections without explicitly accounting for non-uniform sampling intervals. As a result, the representations may vary with the sampling distribution even for the same underlying signal. Further analysis would help establish their robustness to different sampling patterns.
-
The robustness to varying levels of sampling irregularity could be further evaluated. Although the paper considers different sampling patterns, it remains unclear whether QiYao-I maintains its advantages over existing methods as the degree of sampling irregularity increases or decreases. Controlled experiments under different irregularity levels would better demonstrate its robustness.
6 comments
Response to Q2 & W2 — Sensitivity of spectral representations to irregular sampling density
Thanks for your valuable suggestion. We pretrain two variants from scratch on the same corpus under a common budget of same optimizer updates. Due to time constraints, we adopt a limited pretraining budget for these ablation experiments. Our downstream datasets have different sampling patterns, which is detailed in Table 7. We compare time-weighted aggregation with our equal-weight aggregation. As shown in Table R1, equal-weight aggregation achieves lower errors on 10 out of 16 dataset. These results indicate that equal-weight aggregation already provides an effective representation of spectral information under irregular sampling, and explicitly incorporating time-weighted aggregation does not necessarily improve forecasting performance.
Table R1.
| Dataset | Equal-weight aggregation | Time-weighted aggregation |
|---|---|---|
| CESNET (10⁻³) | 22.17 / 90.37 | 22.19 / 90.52 |
| APTC (10⁻⁵) | 392.50 / 4640.84 | 392.38 / 4639.72 |
| ClusterTrace (10⁻³) | 38.55 / 144.92 | 38.86 / 144.86 |
| Seabirds (10⁻⁶) | 162.83 / 7458.76 | 169.70 / 7618.67 |
| Human Activity (10⁻⁴) | 41.20 / 391.18 | 40.97 / 390.14 |
| FNSPID (10⁻⁵) | 324.98 / 3331.04 | 326.56 / 3334.38 |
| GDELT (10⁻⁴) | 201.29 / 856.52 | 202.36 / 854.18 |
| USHCN (10⁻³) | 768.79 / 404.17 | 774.37 / 406.61 |
Response to Q3 & W3 — Robustness to varying degrees of sampling irregularity 1
Thanks for your valuable suggestion. Following your suggestions, we add two experiments.
Reducing the number of historical observations. For each input window, we varied the retention rate from 100% to 10%, retaining the corresponding percentage of historical observations while preserving the first and last observations. We compare QiYao-I with Chronos-2 on identical retained observations and future targets. Chronos-2 uses the original linear-interpolation adaptation.
Table R1.
| History retained | Model | APTC (10⁻⁵) | ClusterTrace (10⁻³) | Seabirds (10⁻⁶) | FNSPID (10⁻⁵) | CESNET (10⁻³) |
|---|---|---|---|---|---|---|
| 100% | QiYao-I | 385.30 / 4597.09 | 37.22 / 142.67 | 165.67 / 7411.77 | 249.80 / 3016.12 | 22.01 / 90.24 |
| Chronos-2 | 401.85 / 4681.75 | 45.99 / 157.43 | 295.50 / 9384.00 | 311.00 / 3253.60 | 22.00 / 90.94 | |
| 90% | QiYao-I | 386.39 / 4603.18 | 37.25 / 142.81 | 167.51 / 7447.23 | 250.62 / 3021.41 | 22.04 / 90.24 |
| Chronos-2 | 399.31 / 4666.21 | 45.46 / 156.56 | 292.31 / 9358.46 | 306.44 / 3233.96 | 21.99 / 90.81 | |
| 80% | QiYao-I | 387.63 / 4610.06 | 37.27 / 142.88 | 170.08 / 7478.49 | 253.88 / 3032.98 | 22.03 / 90.19 |
| Chronos-2 | 399.09 / 4664.77 | 44.76 / 155.83 | 302.36 / 9570.26 | 302.94 / 3227.62 | 22.00 / 90.82 | |
| 70% | QiYao-I | 389.10 / 4618.69 | 37.23 / 142.89 | 170.65 / 7471.20 | 256.95 / 3052.17 | 22.01 / 90.27 |
| Chronos-2 | 397.82 / 4658.42 | 44.75 / 155.60 | 325.17 / 9667.50 | 302.31 / 3240.92 | 22.05 / 90.93 | |
| 60% | QiYao-I | 391.28 / 4631.50 | 37.59 / 143.31 | 169.80 / 7433.18 | 260.72 / 3071.55 | 22.15 / 90.39 |
| Chronos-2 | 399.44 / 4666.99 | 44.57 / 155.36 | 319.17 / 9058.79 | 298.28 / 3227.25 | 22.10 / 90.99 | |
| 50% | QiYao-I | 394.39 / 4649.12 | 37.71 / 143.73 | 168.06 / 7388.24 | 266.03 / 3100.93 | 22.28 / 90.60 |
| Chronos-2 | 402.28 / 4682.21 | 45.00 / 155.94 | 300.70 / 9130.18 | 297.12 / 3245.13 | 22.13 / 91.17 | |
| 40% | QiYao-I | 398.41 / 4672.02 | 38.21 / 144.92 | 166.40 / 7390.86 | 271.61 / 3118.03 | 22.50 / 90.87 |
| Chronos-2 | 406.75 / 4707.31 | 44.13 / 154.89 | 337.75 / 9944.06 | 287.61 / 3238.47 | 22.25 / 91.41 | |
| 30% | QiYao-I | 404.88 / 4708.15 | 37.96 / 144.40 | 170.65 / 7534.77 | 277.56 / 3148.20 | 22.82 / 91.20 |
| Chronos-2 | 412.92 / 4737.82 | 43.25 / 153.25 | 206.21 / 7504.69 | 320.62 / 3372.11 | 22.55 / 91.80 | |
| 20% | QiYao-I | 415.29 / 4766.49 | 38.27 / 145.00 | 164.76 / 7400.93 | 286.48 / 3181.00 | 22.99 / 91.66 |
| Chronos-2 | 425.98 / 4812.52 | 43.03 / 153.23 | 214.68 / 8351.13 | 302.13 / 3322.02 | 23.00 / 92.86 | |
| 10% | QiYao-I | 438.26 / 4892.17 | 40.83 / 147.93 | 162.10 / 7321.07 | 310.15 / 3287.56 | 23.37 / 92.48 |
| Chronos-2 | 459.45 / 5007.46 | 44.10 / 154.80 | 250.22 / 9705.68 | 369.79 / 3904.63 | 24.31 / 94.72 |
Progressively removing historical observations while preserving the endpoints creates sparser inputs with more gaps. QiYao-I remains robust to moderate thinning, with MSE increasing by at most 2.66% at 80% retention and 6.50% at 50% retention across the five datasets. Compared with Chronos-2, QiYao-I achieves lower MSE on four of the five datasets at both retention levels, with average relative reductions of 15.88% and 14.41%, respectively.
Response to Q1 & W1 — Learnable manifold distance under sampling-conditioned bandwidth
Thanks for your valuable suggestion. To demonstrate the contribution of the learnable temporal manifold distance, we compare it with a simpler timestamp-distance-based attention bias, while retaining the same anchor-based sampling-conditioned bandwidth. Both variants are pretrained from scratch on the same corpus with a matched number of optimizer updates, using a reduced pretraining budget due to time constraints.
Table R1.
T1 achieves lower MSE on seven of eight datasets. These results support the additional benefit of the learned temporal manifold formulation over a simpler timestamp-distance-based attention bias.