acceptodds
Desk rejected at ICLR 2027

scUGPT: Gene-wise Response Curves Along a Supplied Single-Cell Trajectory Axis

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

A central goal in functional genomics is to predict not only how strongly a perturbation shifts a cell, but when each gene responds. Single-cell measurements are noisy and sparse, and a response is observed at only a few experimental time points. Most perturbation models predict gene expression at a measured time point, so they do not directly describe response onset or persistence. We present \method, a trajectory-conditioned model that predicts gene-wise response curves along a supplied process axis. An analytic process kernel separates response centre, persistence, and late switching, distinguishing transient and maintained programmes in the parameterization. Perturbation-conditioned attention and a residual prediction path refine the kernel when the observed grid is too sparse to identify its parameters. On the reference GSE213069 split, anchor calibration improves full-curve reconstruction and the model reaches day-4-to-day-5 change correlation , versus without the residual path. Across four additional condition splits, change correlation averages . The results support a structured representation of response curves, with forecasting performance that depends on the condition split. Code is available at https://osf.io/xn3fh/?view_only=dc7333b3d4d74a0aa1baaf9f964306bf.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
single-cell perturbation prediction, CRISPR screen, response timing, trajectory modelling, process kernel, temporal dynamics, sequence model

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_X7UQOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper models gene-wise perturbation response patterns and magnitudes along a supplied temporal or pseudo time trajectory. It uses an analytical kernel to characterize response timing, persistence, and late reversal, as well as a transformer to learn magnitudes and refine the predicted curves. The model achieves reasonably good reconstruction accuracy, but its future prediction performance does not demonstrate robust improvements across different condition splits.

Critical strengths

Potential biological interpretability. The paper explicitly parameterizes gene-wise response timing, persistence, and late reversal using an analytical process kernel. This structured representation offers the potential for biological interpretability by distinguishing different temporal response patterns, such as transient, sustained, and biphasic responses, rather than treating gene expression changes as static endpoint predictions. However, the extent to which these parameters reflect underlying biological mechanisms remains to be established.

Critical weaknesses

1.Lack of single-cell heterogenety modeling. The proposed method represents gene responses as analytic functions along a chronological axis instead of the distribution of gene expression across cells. It can’t adequately characterize heterogeneous perturbation responses or changes in the proportions of distinct cellular subpopulations, raising concerns about whether the method fully exploits the heterogeneous information provided by single-cell data.

2.Failure to capture perturbation-induced divergence in cell trajectories. Perturbations can alter cell-state transitions and induce divergent differentiation trajectories, along which the same gene may exhibit distinct response patterns depending on the cellular state or developmental fate. However, the proposed method predicts gene-wise response curves along a given axis without explicitly modeling perturbation-dependent trajectory branching or cell-fate transitions. This limits its ability to characterize the dynamic consequences of perturbations at the cellular level.

3.Limited and insufficiently robust predictive performance. Its forecasting performance does not demonstrate robust improvements across different condition splits. In the main future-day prediction experiment, the average correlation across four additional condition splits is close to zero (0.007 ± 0.122).

4.Limited biological relevance of gene-wise response trends. Gene-expression programs are shaped by complex regulatory interactions. By focusing primarily on individual genes' overall response trends along a supplied trajectory, the proposed method captures only a limited aspect of these complex dynamics. It remains unclear whether such gene-wise predictions can reveal meaningful biological programs, illustrate coordinated responses to perturbations, or provide applicable insights into the underlying regulatory processes. Consequently, the biological relevance and downstream utility of the predicted response curves remain insufficiently demonstrated.

Questions for authors
  1. Disentanglement of response kinetics and modulation. The model explicitly parameterizes response kinetics through the analytical kernel (k_g(\tau)), while the GPUA module learns a time-dependent gate (\gamma_g(\tau)) to modulate the kernel output. Since both components vary with (\tau), how does the model distinguish their respective roles and prevent the gate from learning temporal patterns that duplicate or override the response kinetics encoded by the kernel?

  2. Selection of RBF basis functions and preservation of the explicit response prior. How is the number of RBF basis functions determined, and how sensitive are the results to this choice? Given the flexibility of the perturbation-conditioned RBF dictionary, what evidence demonstrates that it complements rather than overrides the response patterns encoded by the analytical kernel, thereby preserving the contribution and interpretability of the explicit response parameterization?

Suggestion for Future Work

  1. Downstream biological utility of explicitly modeled responses. The model characterizes gene-wise response timing, persistence, and late reversal, but the biological implications of these predicted response patterns remain unclear. How could these predictions be translated into applicable biological insights, such as prioritizing candidate regulators, identifying temporally coordinated gene programs, or generating experimentally testable hypotheses? Demonstrating such downstream utility would help establish the value of explicitly modeling response dynamics beyond curve reconstruction.
3 comments
AuthorsOpenReview ↗

Response to Reviewer X7UQ, Model Scope, Component Controls and Generalization

Thank you for the detailed assessment. We added component controls and matched split reruns, and clarified the model scope.

W1 and W2. Population responses, not cell trajectories.

scUGPT predicts population mean expression along a supplied time or pseudotime axis. Cells define the means and pseudotime coordinate. It does not predict cellular distributions, proportions, branching, or fate transitions. Mean shifts can reflect changes within a state or in cell composition. Lineage panels use known lineages, not inferred perturbation induced branching. We now state these boundaries explicitly.

W3. Forecasting across condition splits.

We completed 270 matched trainings across five Ishikawa hiPSC CRISPR condition splits, nine configurations, two tasks, and three seeds per task. The original architecture and 96 gene panel are retained. Forecast training and validation targets end at day 4. Below are means ± SD across five split means, after averaging seeds within each split.

MethodFuture MSEFuture change r
scUGPT Full, ours0.01290 ± 0.005220.074 ± 0.164
scUGPT No kernel0.00927 ± 0.002630.106 ± 0.056
scUGPT No anchor0.02555 ± 0.008020.092 ± 0.029
scUGPT Neither0.02534 ± 0.008350.074 ± 0.036
scUGPT Context gate, no RBF0.01255 ± 0.005150.102 ± 0.167
Fixed RBF basis0.02477 ± 0.00937-0.010 ± 0.016
GRU0.02203 ± 0.01151-0.039 ± 0.051
Transformer0.01798 ± 0.007260.045 ± 0.039
MLP time0.01897 ± 0.01053-0.061 ± 0.035
Static broadcast0.00956 ± 0.00216NA

Change r compares predicted day 5 minus predicted day 4 with observed change. Static change correlation is undefined. No kernel retains the anchor and primary loss, fixing the kernel multiplier to one and removing kernel attention bias.

Full has lower mean MSE than the four external learned comparators, but not No kernel or static broadcast. Kernel addition improves change correlation on one of five splits, not consistently. Tables S13 through S18 retain every split. These reruns do not replace the historical 0.007 ± 0.122 result. Overlapping splits are not independent biological replicates.

Q1. Kernel and gate are not uniquely disentangled.

The bounded gate attenuates the anchor and kernel product. It cannot directly reverse its sign, but can change peak heights and timing. A constant initial gate and zero dictionary and residual outputs favor the analytic response initially. Training does not enforce unique temporal roles.

We retrained fixed and context only gates on split 2026. Fixed uses the initial gate value. Context only removes time input and broadcasts a context dependent gate. Both retain the anchor, analytic kernel, attention, lag, mixing, residual, and primary loss. Table S46 reports three seed means ± SD.

GateFuture MSEFuture change r
Dynamic, Full0.00907 ± 0.000210.278 ± 0.053
Fixed0.00909 ± 0.000310.283 ± 0.046
Context only0.00906 ± 0.000300.282 ± 0.046

Similar errors do not establish equivalence or disentanglement. Initialization permutations reduce centre rank stability from 0.943 to 0.063, while prediction rank stability remains 0.895. Table S49 supports treating coefficients as response descriptors, not validated biological kinetics.

Q2. Dictionary size and analytic contribution.

Eight Gaussian bases were an implementation choice. Table S47 varies their number with the dynamic gate and matched training. Zero disables the dictionary, not the analytic kernel. Values are three seed means. RMS ratio is the root mean square magnitude of the gained dictionary divided by that of the analytic terms.

BasesFuture MSEFuture change rRMS ratio
00.008970.3040.000
40.009060.2890.103
80.009070.2780.173
160.009170.2540.277

The analytic terms have greater aggregate magnitude here, but there is no orthogonality or norm constraint preventing overlap. More bases increase dictionary contribution without improving the displayed forecast means. We do not interpret the RMS ratio as explained variance or proof that the prior is preserved.

W4. Coordinated programs and downstream utility.

Gene mixing links the predicted curves. Across three historical checkpoints, coupling matrices have pairwise Spearman correlations of 0.70, 0.63, and 0.70. Among 4560 gene pairs, the strongest coupling decile has mean observed coresponse 0.134, versus 0.066 for the rest. Figure S2 illustrates coordinated pairs, including RPS12 with RPL10A and HMGCS1 with MVD.

This supports descriptive covariation, not regulatory direction. Association with STRING after controlling for coexpression is weak, partial Spearman 0.013. These outputs can nominate gene sets for follow up, but do not establish validated regulators or causal mechanisms.

I appreciate the authors' detailed responses and additional experimental results. However, I believe that some of my concerns remain insufficiently addressed. Therefore, I have decided to maintain my original rating.

Thank you for your continued consideration of our work. We understand that some of your concerns may not have been fully addressed. To ensure that we respond appropriately, could you please clarify which specific concerns you believe remain insufficiently addressed? We would be very happy to provide further explanations, analyses, or experimental results where appropriate. Your detailed guidance would help us better address the issues you have identified.

Reviewer_31UGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors introduce scUGPT, a model that represents perturbational responses along time or pseudo-time axis using a gene-wise process kernel. The three empirical response curves are transient, persistent, and late-response shapes. Evaluation focuses on pseudo-time fitting, reconstruction, forecasting and one-step perturbation prediction. Model improvement is not consistent across tasks and splits, and curve fitting seems empirical rather than biologically motivated.

Critical strengths

The paper features thorough benchmark across task setups beyond time course and against baselines and negative controls, and across multiple datasets. The paper included hyper-parameter sensitivity analysis which demonstrates robustness.

Critical weaknesses
  • The curves are real but seems empirical, which could explain why generalization is inconsistent across tasks and splits.
  • Benchmark metrics are not comprehensive. Community accepted benchmarks could include more metrics for example from cell-eval.
Questions for authors
  • What are the biological motivations behind the three curve shapes chosen? What are counter-examples or outlier trajectories?
  • Beyond the kernels, attention and other layers also affect output. What are the relative contribution of each component to performance?
1 comment
AuthorsOpenReview ↗

Response to Reviewer 31UG, Biological Scope, Metrics and Component Ablations

Thank you for recognizing the breadth of the evaluation. We revised the biological interpretation and added direct component controls.

W1 and Q1. Biological motivation and limits.

Transient represents a temporary excursion followed by recovery, persistent a maintained change, and signed late a later reinforcement or reversal. Shared location and width limit free parameters. These are descriptive response motifs, not estimates of transcription, degradation, or feedback rates.

Counterexamples include oscillations, multiple waves, abrupt switches, and opposing lineage responses. Population means can also change with cell proportions. Our model does not separate these causes or infer branching.

Table S43 reports the five largest forecast errors among 480 condition and gene pairs on the fixed Ishikawa split. Predictions below average three seeds.

PerturbationGeneObserved day 4Observed day 5Predicted day 5
SOX2AC104257.1-0.402-0.561-0.152
SOX2MT1X-0.688-0.778-0.413
SOX2RARRES2-0.354-0.416-0.073
ZIC2H4C3-0.1230.220-0.106
ZIC2AC104257.10.2450.4370.118

These forecast errors do not prove that the analytic family cannot fit the curves. Independent initialization permutations reduce centre rank stability from 0.943 to 0.063, while prediction rank stability remains 0.895. Table S49 therefore supports descriptive coefficients, not recovered biological kinetics.

W2. Metrics beyond expression levels.

We report MSE and MAE for levels, Pearson correlation for levels, adjacent difference correlation for shape, and onset, peak, and turning point diagnostics where defined. Future change correlation compares predicted day 5 minus predicted day 4 with the observed change. It is distinct from perturbation effect correlation relative to controls.

Tables S19 and S20 report endpoint effect correlation and top 20 differential expression overlap on unseen single perturbations. Below, r is effect Pearson and DE is overlap. Higher values are better. The fixed split contains 22 Adamson UPR, 37 Norman CRISPRa, and 273 Replogle K562 holdouts.

MethodAdamson rAdamson DENorman rNorman DEReplogle rReplogle DE
scUGPT, ours0.7900.4860.4830.3610.4750.253
Ridge0.8230.5250.4130.2820.4860.283
scGen style MLP0.8380.5520.4990.3350.4740.250
GEARS graph adapter0.7050.4360.5060.3360.3730.143
Train mean0.7110.4410.4980.3340.3730.143

scUGPT has the highest listed overlap on Norman, but does not lead every metric. These local comparisons are endpoint diagnostics, not temporal forecasting or official cell-eval scores. We have not completed an official cell-eval run. Such a comparison requires aligned expression units, gene panels, controls, and aggregation. A mean predictor cannot establish cellular distribution fidelity by copying its mean into artificial cells.

Q2. Contributions beyond the kernel.

We added anchor and kernel factorial controls and direct gate and dictionary interventions. All rows below use Ishikawa hiPSC CRISPR, 96 genes, split 2026, and three seeds per task. Reconstruction uses the observed day 5 anchor. Forecast fitting and checkpoint selection use days 2 through 4. Values are seed means, with SDs in Tables S3, S4, and S46.

VariantRecon MSEShapeFuture MSEFuture change r
Full, ours0.007670.2120.009070.278
No kernel, anchor retained0.005850.3580.007240.042
No anchor, kernel retained0.013990.3950.016470.082
Neither anchor nor kernel0.013920.4060.016230.053
Fixed gate0.007700.2280.009090.283
Context only gate0.007680.2270.009060.282
No dictionary0.007330.2860.008970.304

No kernel fixes the multiplier to one and removes kernel attention bias while retaining the anchor and other paths. All rows retain the primary loss. The anchor has a large effect on level error. Kernel addition improves change correlation on this split, but not MSE. Fixed and context only gates give similar errors, and dictionary removal improves the displayed means. Small differences do not establish equivalence, and parameter counts differ.

The causal attention mask study changes ordering, not the presence of attention. Separate attention, lag, and mixing removal studies remain unavailable. We do not present these controls as a complete attribution of all layers.

Across five matched condition splits, Full gives future change correlation 0.074 ± 0.164, versus 0.106 ± 0.056 without the kernel. These are SDs over split means after averaging three seeds. Tables S13 through S18 therefore support an explicit response representation, not a stable forecasting advantage.

Reviewer_pe4aOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces a parameterized “process kernel” that can describe changes in pseudobulk expressions of genes across trajectories in single-cell datasets. The paper claims this kernel can express a wide variety of gene trajectories, such as genes that peak early and then decay slowly or genes that have a sustained response throughout. A neural network is then wrapped out this process kernel to enable a variety of prediction tasks, including forecasting gene expression in held-out future timepoints, backcasting gene expression in held-out earlier timepoints, and predicting trajectories for entirely held-out conditions. The paper tests this framework on six different single-cell datasets.

Critical strengths
  • The problem the paper is trying to solve is a reasonable one, and the method they have proposed is quite flexible given that it can (in theory) perform forecasting, backcasting, and trajectory prediction on held-out conditions.
  • The paper is very candid about its shortcomings. For instance in the “Discussion and Limitations” section the author(s) suggest that the results “do not establish a split-independent forecasting advantage” for their method.
Critical weaknesses

There are many small issues with this paper, but the most fundamental issue is that there are several seemingly arbitrary choices that go into this framework:

  • Why use this specific combination of seven parameters in the construction of the kernel?
  • Why the Huber loss function, compared to MSE or L1? Where do the 1 / 0.05 / 0.25 weights comes from?

These choices would not be problematic on their own were there sufficiently robust ablation studies included in the paper, but the key ablation studies are missing and the ones that are present are severely lacking. For instance, the “No kernel” ablation isn’t cleanly assessing the impact of removing the kernel, since it sets h = 0, which removes contributions from the anchor in addition to the kernel. (And elsewhere in the paper we see that the anchor is really important, since ablating the anchor significantly degrades performance.) There are also ablations for various kernel parameters in Table 9, but many of them show that performance actually improves under ablation: dictionary gain fixed at r = 0 is better on every metric, and late amplitude fixed at s = 0 gives comparable performance as allowing s to be freely optimized.

Questions for authors
  • Can you report Tables 3 and 5 averaged over all condition splits, rather than using the fixed seed of 2026?
  • How does GEARS score exactly 0.000 +/- 0.000 on Hagai Delta-Pearson? Is this a true readout or a bad implementation?
  • What happens if you replace the kernel by a constant or by a simpler kernel?
3 comments
AuthorsOpenReview ↗

Response to Reviewer pe4a, Clean Ablations and Cross Split Evaluation

Thank you for identifying these gaps. We added the missing controls. Tables below use Ishikawa hiPSC CRISPR, 96 genes.

W2. Separating kernel and anchor.

The old intervention also removed the anchor contribution. No kernel now fixes the multiplier to one and removes kernel attention bias, retaining the refined anchor, gate, lag and gene mixing, residual, and primary loss. No anchor zeroes the static effect before normalization. All four variants are retrained.

Below are split 2026 means ± SD over three seeds. Reconstruction uses the day 5 anchor. Forecasting observes days 2 through 4. Future Δr correlates predicted change from day 4 to day 5 with observed change. One means enabled.

AnchorKernelWeightsRecon MSEFuture MSEFuture Δr
119510140.00767 ± 0.000220.00907 ± 0.000210.278 ± 0.053
107593970.00585 ± 0.000200.00724 ± 0.000410.042 ± 0.067
019510140.01399 ± 0.000070.01647 ± 0.000050.082 ± 0.015
007593970.01392 ± 0.000070.01623 ± 0.000210.053 ± 0.031

The kernel improves change correlation here, not MSE. Parameter counts are not matched. See Table 5 and Tables S3 and S4.

Q1. Averaging condition splits.

We completed 270 runs, five splits, nine configurations, two tasks, and three seeds. All use a 200 epoch limit and validation early stopping. Forecast training and validation targets end at day 4. Below are means ± SD across five split means, after averaging seeds within each split.

MethodRecon MSEFuture MSEFuture Δr
scUGPT Full, ours0.00730 ± 0.001340.01290 ± 0.005220.074 ± 0.164
scUGPT No kernel0.00595 ± 0.001140.00927 ± 0.002630.106 ± 0.056
scUGPT No anchor0.01725 ± 0.002540.02555 ± 0.008020.092 ± 0.029
scUGPT Neither0.01708 ± 0.002560.02534 ± 0.008350.074 ± 0.036
scUGPT Context gate, no RBF0.00728 ± 0.001430.01255 ± 0.005150.102 ± 0.167
Fixed RBF basis0.01122 ± 0.004320.02477 ± 0.00937-0.010 ± 0.016
GRU0.01095 ± 0.004560.02203 ± 0.01151-0.039 ± 0.051
Transformer0.01150 ± 0.002310.01798 ± 0.007260.045 ± 0.039
MLP time0.01190 ± 0.004010.01897 ± 0.01053-0.061 ± 0.035
Static broadcast0.01149 ± 0.003860.00956 ± 0.00216NA

Full has lower mean future MSE than the four external learned comparators, but not No kernel or static broadcast. Kernel addition improves change correlation on only one of five splits, so we do not claim a stable kernel advantage. See Tables S13 through S18. Overlapping splits are not independent biological replicates.

W1 and Q3. Parameterization and simpler kernels.

The seven analytic fit quantities are amplitude, location, width, persistence, late gap, late amplitude, and offset. Transient, maintained, and signed late terms describe response patterns, not biological rates. The neural model has additional weights.

Below are split 2026 means for scUGPT variants. SDs and other datasets are in Tables S35 through S42.

KernelRecon MSEFuture MSEFuture Δr
Full0.007670.009070.278
Constant0.005700.007140.036
Linear0.006030.007230.149
Transient0.008530.014240.356
Persistence0.007260.008550.356
Late, no dictionary0.007330.008970.304
Contextual RBF0.011160.016270.060

Simpler variants can improve performance. Constant and Linear retain the temporal gate and residual. Persistence adds the maintained term, and Late adds the late term. These are optional components, not universally necessary terms.

W3. Loss and weights.

Huber reduces sensitivity to large errors. The implemented point, shape, and amplitude weights are 1, 0.5, and 0.25. We correct the reported shape weight from 0.05 to 0.5. The separate primary weight is 0.45. Loss replacement covers all active terms, retaining normalization and validation MSE selection.

LossShape weightAmplitude weightRecon MSEFuture MSEFuture Δr
Huber0.50.250.007670.009070.278
MSE0.50.250.007720.009060.275
L10.50.250.007380.009150.266
Huber1.01.00.007720.009490.254
Huber0.250.050.007530.008840.262

These are split 2026 means, with SDs in Table S48. No setting improves every metric. Weights are empirical, not established optima.

Q2. GEARS zero correlations.

The local static adapter broadcasts one profile and returns zero for undefined zero variance correlations. The plotted export uses a different adapter. Historical checkpoint mapping remains unresolved. We removed the GEARS row from the Hagai comparison and retained this caveat. These zeros are not valid evidence of GEARS performance.

Reviewer_pe4aOpenReview ↗

Response to Rebuttal

I thank the authors for the work they put into addressing my concerns in their rebuttal. However, the ablation studies I requested in my review suggest that this method is even less robust than I initially thought:

  • Table 5: The kernel hurts on most metrics: adding the kernel worsens reconstruction MSE (0.00585 → 0.00767) and future MSE (0.00724 → 0.00907). It also worsens reconstruction shape correlation (0.358 → 0.212). Its only gain is future change correlation (0.042 → 0.278).
  • Table 6: scUGPT (Full) is never the outright best kernel form on either metric in any dataset, and on the main split it is matched or beaten on all metrics by two of its own simplified variants.

In other words, my concerns about the arbitrariness of the design have now been vindicated by ablation studies. So while I thank the authors for their candor, my overall rating will remain unchanged.

Dear Reviewer,

Thank you for reading the revised manuscript and ablation results closely. The main point we would like to clarify is that Transient, Persistence, and Late in Table 6 are nested ablation variants of our proposed analytic kernel, rather than external competing methods. Their strong performance supports the analytic kernel design, while showing that the additional RBF component in Full does not improve performance on this split.

Table 5 addresses a related distinction between predicting expression levels and recovering temporal changes. Ishikawa contains 4 timepoints, and predictions that remain close to the static anchor can achieve low MSE when temporal measurements are sparse and noisy. Tables 3 and 6 illustrate this, with Static broadcast and Constant achieving Future MSE values of 0.00902 and 0.00714, respectively. This pattern is consistent with the strong performance of static baselines on level metrics reported by Ahlmann Eltze et al., 2025.

With Anchor On and Kernel Off, Future change rr is 0.042±0.0670.042 \pm 0.067. This near zero correlation indicates poor recovery of expression changes from day 4 to day 5, despite the lower MSE. Kernel Off also achieves higher Recon Shape, suggesting that fitting observed differences does not necessarily translate into better extrapolation. Without the temporal constraint imposed by the kernel, the residual and gate paths have more freedom to fit those differences.

Adding the kernel raises Future change rr from 0.0420.042 to 0.278±0.0530.278 \pm 0.053, more than 6 times the mean correlation, while increasing MSE by approximately 0.0018. The lower MSE of Kernel Off therefore needs to be considered alongside its poor change recovery. In this setting, the anchor provides a useful estimate of expression levels, while the kernel improves forecasts of temporal changes.

Table 6 then examines which parts of our kernel contribute to this improvement. Constant, Linear, and RBF control are alternative parameterizations. Transient contains term 1 of Equation 1, Persistence contains terms 1 and 2, and Late contains terms 1 to 3. These variants retain progressively more of our analytic formulation. Full adds the generic RBF dictionary to the same analytic terms.

On the main Ishikawa split, Constant achieves Future change rr of 0.036. RBF control uses 8 generic basis functions without our analytic onset, persistence, and late switching terms. It achieves Future change rr of 0.0600.060 and Future MSE of 0.016270.01627, the highest MSE in this comparison.

By comparison, Future change rr is 0.356 for Transient, 0.356 for Persistence, 0.304 for Late, and 0.278 for Full, with Future MSE values ranging from 0.00855 to 0.00907. All 4 variants contain our analytic terms. The stronger results of Transient and Persistence therefore show that a simpler version of our proposed kernel works well on this split. They also show that including every component is not necessary to obtain the benefit of the analytic formulation.

The difference between Late and Full is the additional RBF term, rg∑j=1Jαgjϕj(τ)r_g \sum_{j=1}^{J} \alpha_{gj}\phi_j(\tau), with JJ equal to 8. Late achieves rr of 0.304 and MSE of 0.00897, compared with 0.278 and 0.00907 for Full. One plausible explanation is that the extra flexibility of the RBF dictionary is difficult to estimate from only 4 timepoints. These results favor the compact analytic variants on this grid, while providing no evidence that the additional RBF dictionary improves forecasting here.

As stated in line 400, we retained Full as the architecture specified before examining these test results, rather than selecting Persistence or Late afterward. The ablations support the contribution of our analytic kernel, but do not establish Full as the best configuration for this split. Table 5 shows the improvement in change recovery when the kernel is included, and Table 6 locates that benefit in the analytic terms shared by our proposed variants.

Thank you for prompting these comparisons. They clarify why the kernel helps recover temporal changes and why a simpler variant can be preferable when only a few timepoints are available.

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

Related papers

Discussion (0)

Sign in to comment.