acceptodds
Under review as a conference paper at ICLR 2027

Temporal Resolution and Effect Flexibility: The Geometry of Survival Learning

Abstract

Discrete-time survival models can be highly sensitive to interval count , usually selected empirically without revealing the underlying mechanism. We trace this sensitivity to a larger model-optimisation problem over temporal resolution and effect flexibility , which governs how freely covariate effects vary across intervals. We make explicit through a unified temporal basis model family, recovering the effect structures of two existing boundaries: time-invariant models such as proportional hazards at , and interval-specific models such as Nnet-survival and DeepHit at . Under baseline and effect regularities and , our theory explains -sensitivity through separate approximation–estimation balances for temporal resolution and for effect flexibility, yielding distinct oracle scales and , where is the sample size and the representation dimension. Since these scales generally differ, tuning existing or models explores only fixed paths and cannot guarantee the joint optimum. We prove that held-out joint validation over attains the oracle rate without knowing or , making optimisation over the full geometry practical. Controlled studies confirm the predicted geometry. Across six real-world datasets and four deep-survival heads, joint selection lowers mean test loss on held-out outer folds relative to in 20 of 24 dataset–head strata and favours in 353 of 360 evaluations. Thus, -sensitivity fundamentally reflects a larger joint approximation–estimation geometry of temporal resolution and effect flexibility: existing models trace only fixed paths through it, while reaching the joint optimum requires selecting , not alone.

Primary area:
learning theory
Keywords:
survival analysis, deep survival learning, discrete-time survival models, time discretisation, temporal effect flexibility, adaptive model selection, approximation–estimation trade-off

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_Q5xfOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies the sensitivity of discrete-time survival models to the number of temporal intervals KK and proposes jointly optimizing temporal resolution KK and effect flexibility MM. The authors develop a temporal-basis parameterization with theoretical results on risk decomposition, minimax rates, and adaptive model selection. Experiments on synthetic and real-world survival datasets show improvements in prediction loss.

However, the paper has two major weaknesses: (1) limited methodological and theoretical novelty relative to established survival modeling and statistical learning literature, and (2) insufficient evidence of statistically robust and practically meaningful improvements over existing methods.

Critical strengths

The paper provides a useful perspective on distinguishing temporal discretization from the complexity of time-varying covariate effects. The unified (K ×\times M) formulation helps explain why increasing the number of time intervals can have different effects across survival model architectures.

Critical weaknesses

1. Limited methodological and theoretical novelty.

The central methodological idea is not novel. In particular, Perperoglou et al. (2006) [1] already represented time-varying covariate effects using qq temporal basis functions, yielding a pp \timesq q coefficient matrix. The parameter qq corresponds directly to the proposed MM, as both specify the number of temporal basis functions used to represent covariate effects. Their work further investigated reduced-rank constraints on the coefficient matrix to obtain models with different levels of flexibility. Similarly, Tutz and Binder (2004) [2] had already introduced regularized basis expansions for time-varying effects in discrete-time survival models.

Moreover, controlling the complexity of time-varying effects through smoothing, regularization, or model selection has been extensively investigated. Kauermann (2005) [3] studied penalized spline smoothing for varying-coefficient survival models, Yan and Huang (2012) [4] investigated model selection between time-invariant and time-varying effects, and Perperoglou (2013) [5] further developed reduced-rank hazard regression with fixed and time-varying effects. Therefore, the proposed temporal-basis parameterization and the use of a complexity parameter MM have clear methodological precedents. The paper does not convincingly explain what is fundamentally new beyond adopting a DCT basis and jointly tuning the temporal resolution and effect dimension.

The theoretical novelty is also limited. The approximation–estimation trade-offs, minimax convergence analysis, and adaptive model-selection guarantees are based on well-established statistical principles. Closely related theoretical frameworks, including adaptive estimation and minimax analysis for conditional intensity models under anisotropic regularity, have already been developed by Comte et al. (2011) [6]. Although the exact model classes and risk functions differ, the paper does not clearly demonstrate which theoretical results go beyond existing statistical estimation and model-selection theory.

Overall, the paper appears to repackage established methodological ideas and statistical learning principles into a KK \timesM M framework. The extensive theoretical analysis does not, by itself, establish a substantive methodological or theoretical advance.

References

[1] Perperoglou A, Cessie S, van Houwelingen H C. Reduced‐rank hazard regression for modelling non‐proportional hazards[J]. Statistics in Medicine, 2006, 25(16): 2831-2845.

[2] Tutz G, Binder H. Flexible modelling of discrete failure time including time‐varying smooth effects[J]. Statistics in Medicine, 2004, 23(15): 2445-2461.

[3] Kauermann G. Penalized spline smoothing in multivariable survival models with varying coefficients[J]. Computational statistics & data analysis, 2005, 49(1): 169-186.

[4] Yan J, Huang J. Model selection for Cox models with time-varying coefficients[J]. Biometrics, 2012, 68(2): 419-428.

[5] Perperoglou A. Reduced rank hazard regression with fixed and time‐varying effects of the covariates[J]. Biometrical journal, 2013, 55(1): 38-51.

[6] Comte F, Gaïffas S, Guilloux A. Adaptive estimation of the conditional intensity of marker-dependent counting processes[C]//Annales de l'IHP Probabilités et statistiques. 2011, 47(4): 1171-1196.

2. Insufficient evidence of statistically robust and practically meaningful improvements.

Although the paper reports improvements in continuous-time RC-NLL, the improvements in standard survival prediction metrics are generally very small. In Appendix Figure 10, 20 of the 24 dataset–head combinations exhibit absolute Harrell's C-index changes smaller than 0.005, with most improvements outside METABRIC being on the order of 10−310^{-3} or smaller. Excluding METABRIC, the average C-index improvement is only approximately 0.00075. Similar patterns can be observed for integrated dynamic AUC and IBS.

More importantly, the paper does not report confidence intervals or statistical significance tests for these standard survival prediction metrics. Although Figure 9 provides descriptive fold-level intervals for RC-NLL, comparable uncertainty analyses are absent for C-index, AUC, and IBS. Consequently, it remains unclear whether these small improvements are statistically reliable or practically meaningful.

Furthermore, the primary comparison is against the unrestricted M=KM=K baseline. Since reducing temporal effect complexity naturally acts as a form of regularization, the observed improvements may largely reflect conventional bias–variance trade-offs rather than a distinct advantage of the proposed framework. The paper does not provide sufficiently strong comparisons against existing temporal smoothing or regularization approaches in the learned-representation experiments.

Overall, the experiments demonstrate that restricting model complexity can improve RC-NLL, but do not convincingly establish that introducing MM provides statistically robust, practically significant, or methodologically distinct benefits beyond standard regularization and hyperparameter tuning.

Questions for authors

Q1. Methodological and theoretical novelty relative to prior work.

Perperoglou et al. (2006) [1] already modeled time-varying covariate effects using qq temporal basis functions, where qq plays essentially the same role as the proposed MM. Similarly, Tutz and Binder (2004) [2] studied regularized basis expansions in discrete-time survival models, while subsequent work [3–5] explored temporal smoothing, model selection, and reduced-rank representations.

What is the fundamental methodological novelty of introducing MM, beyond adopting a DCT basis and jointly tuning KK and MM?

Furthermore, approximation–estimation trade-offs, adaptive estimation, and minimax convergence analysis have been extensively studied in statistical learning and survival analysis, including related results in Comte et al. (2011) [6]. Which theoretical results in this paper are genuinely new, rather than applications or reformulations of established statistical principles? Can the authors explicitly distinguish their contributions from these prior results?

Q2. Statistical significance and practical relevance of the reported improvements.

Appendix Figure 10 shows that most improvements in standard survival prediction metrics are extremely small. Specifically, 20 of the 24 dataset–head combinations exhibit absolute Harrell's C-index changes below 0.005, with improvements outside METABRIC averaging approximately 0.00075. Moreover, confidence intervals or significance tests for C-index, time-dependent AUC, and IBS are not reported.

Can the authors provide uncertainty estimates or appropriate paired statistical analyses to demonstrate that these improvements are statistically reliable and practically meaningful?

More importantly, the primary comparison is against the unrestricted M=KM=K baseline. Since restricting MM naturally reduces model complexity, the observed gains may simply reflect conventional regularization and hyperparameter tuning. Although a spectral-smoothing comparator is included in the fixed-representation study, comparable strong baselines are missing from the learned-representation experiments.

Can the authors demonstrate that joint K,MK,M selection provides benefits beyond appropriately tuned temporal regularization, spline smoothing, or other established complexity-control techniques under comparable computational budgets?

Q3. Clarification of the theoretical derivation in Appendix D.4.

In Appendix D.4, Equation (56) derives the baseline-profiled information using a pointwise Schur complement, which appears to allow the baseline function to vary independently at each time point. However, the proposed model parameterizes the baseline using KK piecewise-constant interval coefficients.

Is the pointwise profiling operation equivalent to profiling the finite-dimensional interval-specific baseline parameters? If not, how does this distinction affect the information energy EmE_m and the validity or interpretation of the information-supported mode criterion in Proposition 4?

Reviewer_s7ExOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper studies the separate roles of temporal resolution K and effect flexibility M (which governs how freely covariate effects vary across intervals) in discrete failure time model. The authors make M explicit through a unified temporal basis model family and introduce a DCT-based restriction on interval-specific coefficients. The theoretical analysis provides an explanation of the relevant approximation–estimation tradeoffs. However, the incremental contribution over established sieve and varying-coefficient methods is a major concern. In addition, applicability to deep learning, and advantage over existing alternatives are not sufficiently established.

Critical strengths
  1. The separation between temporal resolution and effect complexity is useful.
  2. The theoretical analysis provides an explanation of the relevant approximation–estimation tradeoffs.
Critical weaknesses
  1. The methodological novelty compared with established sieve and smoothing methods is insufficiently demonstrated. The proposed estimator seems a sieve likelihood estimator with separate baseline and effect dimensions. Existing varying-coefficient spline models and piecewise-exponential likelihoods with smooth baseline and nonlinear or time-varying covariate effects already develop separate complexity control. In addition, adaptive conditional-intensity estimation has established related anisotropic risk theory. Please clarify which results are new and compare against existing literature.
  2. The theoretical model differs from the neural models evaluated. In addition, M=1 does not imply PH for every output transformation. The manuscript acknowledges these limitations, but its main claims should be correspondingly narrower.
  3. The oracle scales involve unknown smoothness and constants, and the real-data procedure still uses extensive validation search. Thus, the practical benefit of the theory remains unclear.
Questions for authors
  1. In addition to the comments mentioned above, please include comparisons with existing continuous time survival methods, e.g. DeepSurv.
  2. Although mean RC-NLL improves in 20 of 24 combinations, C-index and IBS changes aresmall outside METABRIC. Please report absolute performance; predictive improvements should be assessed by magnitude, not only frequency.
1 comment

Dear Reviewer,

Thank you for your thoughtful review and contribution to the discussion.

As the discussion period approaches its conclusion, I would like to kindly remind you to review the authors' rebuttal and engage in the discussion if you have not already done so.

In particular, it would be helpful if you could indicate whether the authors have adequately addressed your concerns and, if appropriate, reconsider your assessment and score.

Thank you again for your time and effort!

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

The paper studies two choices in survival models. K sets the number of time intervals. M controls how much feature effects can change over time. The authors explain the costs of increasing K and M and give a method to choose both. Their theory assumes fixed features. Tests on six datasets and four model heads show lower mean test loss in 20 of 24 settings. The method chooses M < K in 353 of 360 runs.

Critical strengths

1.The model can use many time intervals without giving each interval a separate set of feature effects. This helps explain why a finer time grid can sometimes hurt prediction.

2.The authors test their main claims with both simulated and real data. They use separate test folds to measure performance. They also test several model heads.

Critical weaknesses

1.The neural tests use weight decay and early stopping. However, they do not report a direct comparison with a full head using tuned time smoothing. Such a model could also limit changes in feature effects. It is unclear how much choosing M adds over this option.

2.The theory assumes fixed features, but the neural tests learn them during training. Appendix F.3 also appears to use the same validation set for early stopping and model selection. The proof in Appendix D.8 does not directly cover this setup. The paper should state this limit more clearly.

Questions for authors

1.How does choosing K and M compare with choosing K and a time-smoothing strength for a full head? Please use similar tuning and compute budgets.

2.Do early stopping and model selection use the same validation set? If so, does the proof cover this use of the data?

Comments on the paper

1 comment
AuthorsOpenReview ↗

Thank you and summary of revisions

We sincerely thank the PC members and all reviewers for their time and thoughtful engagement. This discussion has made the paper clearer and stronger, and we are grateful for it.

Reviewer veit asked how choosing MM compares with tuned time smoothing, and whether our selection guarantee covers early stopping. These questions led us to state clearly that truncation and smoothing are two ways to control the same quantity, and to extend the selection guarantee in App. D.8 to early stopping.

Reviewer s7Ex pushed us to position the work precisely against the sieve and varying-coefficient literature, to narrow our claims to the scope the theory supports, and to compare with continuous-time models. The DeepSurv and Cox-Time comparison now shows a practical consequence of the theory.

Reviewer Q5xf pointed us to the classical literature on time-varying effects and found a genuine imprecision in Eq. (56). We read each work closely: they control effect flexibility on a given time scale, and none contains our main results (Thm 2, Cor 5, Prop 6). Some were already cited in the submission, and we will add the rest and correct the display. With no further concern raised, we understand the concerns on novelty, empirical benefit and Eq. (56) to be resolved.

In summary, the revision will (i) state the contribution as a mechanism, the hidden M=KM = K coupling in interval-specific models, (ii) narrow the claims in the abstract and Sec. 4 to fixed representations and time-invariant effect structure, (iii) add the paired tests, tuned-regulariser comparisons and continuous-time baselines, and (iv) correct Eq. (56), which affects no result.

Thank you all again. It has been a genuinely constructive discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.