acceptodds
Under review as a conference paper at ICLR 2027

TempoGeom: Observation Design for Learning Switching Speeds in Temporal Graphs

Abstract

Predicting whether an edge is present does not determine how rapidly it forms and disappears. We study observation design for jointly learning switching clocks and unknown, finite-dimensional time-varying affinities from discretely observed temporal graphs. Affinity perturbations can vanish at observation times while changing between-observation dynamics, confounding the small clock-dependent lag under fast switching. uses a two-scale efficient-Fisher analysis to retain these hidden nuisance effects. In an explicit shared-affinity family, reordering a fixed multiset of observation gaps changes local log-clock information from to , where scales switching rates, without changing observation count or horizon. We characterize how the observation graph determines the number of clock directions at each rate. Because both rates vanish, we introduce separated affinity anchors followed by probes on the switching timescale. Under quantitative affinity-conditioning, graph-rank, and uniform regularity assumptions, this design preserves a positive information floor over bounded relative clock uncertainty; a fixed positive minimum gap precludes such a floor for arbitrarily fast clocks. For finite clocks, we combine nuisance-aware receding-horizon design with centered graph regularization. Matched-estimator comparisons in polynomial and Fourier synthetic studies yield clock-RMSE reductions of and at the slowest scale, respectively, with no consistent advantage at faster scales. Broader pipeline gains also reflect regularization, whose benefits can reverse under increased heterogeneity. Together, these results separate the roles of observation timing, nuisance conditioning, and regularization in learning switching speeds.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Temporal graph learning, Continuous-time dynamic graphs, Observation design, Fisher information, Identifiability

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_RAPjOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The paper studies a model for temporal graphs where edges switches on and off according to a 22-state continuous time Markov process with unknown parameters. Each edge's rate to being on or off depends on parameters set by the two nodes that it connects, and an affinity parameter that is controlled by a parameterised family of time-dependent affinity functions, like polynomials or Fourier functions. The model knows the affinity functions and chooses times to observe the graph, and aims to reconstruct parameters from these observations. The authors propose an algorithm that chooses observation times and adapts to new updates. Their theoretical results mainly focus on why observation time matters and asymptotic information rates from informative observation times, along with examples where information is insufficient from badly chosen observation times. The authors empirically test the algorithm against adapted versions of existing algorithms on synthetic data with polynomial and Fourier affinity functions. I recommend a weak reject to this paper. This paper did produce enormous amounts of analytical and experimental work to the specific theoretical question, but there are no concrete theoretical sample complexity guarentees of the suggested algorithm, and the experiments doesn't necessarily support the algorithm's efficiency claims, where the proposed observation strategy improves parameter estimation at the slowest switching scale but performs marginally better against existing methods otherwise.

Critical strengths

(1)(1) The paper identifies a specific difficult graph recovery problem, identifies the difficulty in recovering switching speeds and provides sharp information rates for the updating pattern of the problem.

(2)(2) The paper provides sufficient conditions for observation times where there is non-vanishing information about the graph and parameters.

(3)(3) The paper identifies the asymptotic information rate for the parameters under the graph updating pattern, establishes where these rates hold and shows counterintuitive cases where observation times can affect the information rates directly.

(4)(4) The paper provides significant synthetic experimental evidence to compare the algorithm with adapted existing algorithms to support the performance of the algorithm, and shows some cases where variants of the algorithm does well.

Critical weaknesses

(1)(1) The theoretical analysis does not provide sufficient evidence for the significance of the proposed method. Theorem 11 characterizes local fisher information, but does not establish overall error rate. Later applications of the theorem only derives information rate regimes for certain affinity families.

An established estimator guarantee and showing the information separation result broadly would directly address this concern.

(2)(2) The problem itself is not well motivated and the presentation misses the problem's relevance to existing literature. The relationship between this model and common temporal graph models in literature and the specific difficulty that this model has needs to be well evidenced and compared thoroughly with existing literature, as well as how resolving this problem can help resolve related problems.

Providing more background or a related works section alongside comparisons to existing literature would directly address this concern. This would be nice to have for both the problem, the algorithm and the theoretical guarantee.

(3)(3) The empirical experiment does not establish a clear advantage of the proposed algorithm. While the experiments show advantage in some settings, it seems much of the gains come from regularization, and the algorithm only shows clear improvement in the slowest tested switching scales.

clarifying the conclusions supported by existing evidence and provide evidence that isolate the scheduling contribution and show its significance would directly address this concern

Minor concerns that did not affect my score:

(1)(1) The setting is underexplained at section 22, for example theta is not defined in section 22 when the paper uses it immediately in section 33. Or certain notation are not explained, XeX_e being one row of BB was never explained.

(2)(2) Theorem 11's results are not well explained, it is unclear what "information in a direction" means, or introducing "nuisance" without defining it.

(3)(3) Section 3.23.2 begins by introducing the construction without explaining what it's meant to show.

(4)(4) More unexplained terms like "weighted scenarios" and "candidate pool" makes the paper difficult to follow.

(5)(5) Algorithm 22 would be easier to follow if steps 5 and 7, updating FsF_s and heh_e, wew_e were explicitly referenced to the formulas they are following for the update. I was intuitively unclear of how we are updating it.

Questions for authors

1, What motivated this problem? Is there a concrete application or existing learning problem that leads naturally to these assumptions?

2, How exactly are Step-DAD, Inside-Out SMC2SMC^2, and PASOA adapted to this setting? Which parts of the original methods are retained, and what objectives do the adapted methods optimize?

3, Why were polynomial and Fourier affinity families chosen?

4, What explains why the scheduling improvements are concentrated at the slowest tested switching scale? When should we expect adaptation to improve over a fixed anchor–probe schedule?

Reviewer_PpvSOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes TEMPOGEOM, an observation design framework to jointly learn switching clocks and unknown, finite-dimensional time-varying affinities from observed temporal graphs. This paper uses a two-scale efficient-Fisher analysis to retain hidden nuisance effects where intermediate affinity perturbations mimic clock-dependent lags and combines separated affinity anchors with probes on the switching timescale, avoiding the fixed-resolution obstructions of standard observation gaps. Experiments show that matched-estimator comparisons in polynomial and Fourier synthetic studies yield clock-RMSE reductions of 8.3% and 11.5% at the slowest scale, respectively, with no consistent advantage at faster scales.

Critical strengths

The paper studies observation design for learning switching speeds in temporal graphs with unknown time-varying affinities, which is a meaningful problem.

This paper provides detailed theoretical analysis.

Critical weaknesses

The motivation for endpoint-invisible affinity variation mainly relies on a constructed shared-affinity family, and it is still unclear how it affects clock information in more general settings. The connection between these theoretical guarantees and the proposed finite-clock method is unconvincing.

There is limited experimental evidence for the benefits of the proposed observation design. The main experiments are limited to polynomial and Fourier synthetic families on small complete graphs. The matched-estimator comparisons show that much of the full-pipeline improvement comes from regularization, while acquisition gains are mainly observed at scale 8, with no consistent advantage at faster scales. Thus, these results do not sufficiently demonstrate the advantages and generality of TempoGeom-R.

Questions for authors

For W1, how common is endpoint-invisible affinity variation beyond the explicitly constructed shared-affinity family? Can the authors provide more direct evidence of its impact on clock information or clock recovery?

Do Theorem 5's information guarantees still apply to TempoGeom-R's finite-clock schedules when there is a fixed positive minimum gap?

For W2, given that the matched-estimator gains mainly appear at scale 8, what is the additional benefit of TempoGeom-R over fixed anchor-and-probe design or other simpler timing strategies?

Can the authors clarify how the finite-support beliefs and the prestructured timing grid affect comparisons with recent acquisition methods? Would the conclusions hold under more general observation graphs or continuous parameter uncertainty?

Reviewer_Wt9NOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies the problem of estimating switching speeds in temporal graphs from discrete observations, where the actual edge formation and dissolution events between observations are unobserved. The objective is to estimate how quickly edges change between present and absent states, when the affinity trajectories are unknown.

Theoretically, the work also provdides a two-scale efficient Fisher information expansion that accounts for nuisance-affinity directions invisible at observation endpoints. and demonstrates that two observation schedules with identical observation counts, horizons.

Experiments on synthetic temporal graphs with polynomial and Fourier affinity trajectories suggest that the proposed acquisition method can improve clock estimation at relatively slow switching scales, while the benefits disappear at faster scales.

Critical strengths

The major strength of the work is its theoretical contritution.

It is shown that affinity parameters invisible at observation endpoints cannot necessarily be discarded from the Fisher information analysis. Their derivatives can contribute at the same asymptotic order as the switching-clock signal. Two-scale nuisance projection provides a principled way to characterize this effect. The result shows why conventional analyses based only on endpoint-identifiable affinity directions can overestimate information about switching rates.

By keeping the observation budget, horizon, and gap multiset fixed, the authors isolate the effect of temporal ordering. The theoretical analysis provides a concrete example of how observation scheduling changes the statistical difficulty of estimating temporal dynamics. The paper derives matching asymptotic rates instead of elying exclusively on numerical evidence.

Critical weaknesses

1.The theoretical analysis is rigorous, However, its applicability to realistic temporal graphs is unclear. The main results assume a finite-dimensional, correctly specified affinity family with sufficient smoothness and regularity. The guarantees requires a uniformly well-conditioned anchor Jacobian and a bounded range of relative clock uncertainty. However, in real temporal networks, affinity trajectories may be irregular, piecewise stationary, affected by abrupt events, or difficult to represent using a small number of smooth basis functions. Moreover, the affinity family is generally unknown.

  1. The experimental results do not consistently demonstrate that the proposed adaptive acquisition method significantly outperforms simpler observation strategies. Table 1 shows that TempoGeom-R/J with regularization the advantages largely disappear at higher switching scales. The results suggest that the benefit of optimizing acquisition schedules depends strongly on the switching regime.
Questions for authors
  1. How sensitive are the theoretical conclusions to misspecification of the affinity basis?

  2. Can the method accommodate piecewise-smooth or nonparametric affinity trajectories?

  3. Why does TempoGeom-R fail to provide consistent improvements at faster switching scales?

  4. How robust is the algorithm across substantially different degrees of clock and affinity heterogeneity?

  5. What is the computational complexity of TempoGeom-R in terms of the number of dyads, observation candidates, scenarios, and future observation pairs?

  6. How does runtime scale as the number of observation points increases?

Reviewer_iZq2OpenReview ↗
Rating
1
Confidence
1
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes an observation design based on Fisher analysis methods to learn the clock switching speeds in temporal graphs. Every pair of nodes at each time of a temporal graph can be connected or not, and the probability of connection is based on the affinity. The clock controls the switching between different affinity states. Given that the observed data is incomplete with only partial time and the affinities are unobserved, it makes learning the switching rates a challenge. This paper addresses this challenge and provides theoretical guarantees on the lower bound of the Fisher information.

Critical strengths

This paper seems to be addressing a challenging problem, and provides theoretical guarantees as well as empirical validations. I cannot verify the soundness and significance of the results since I do not fully understand the paper.

Critical weaknesses

The paper is very notation-heavy and difficult to follow. There are not enough plain-text explanations on the notations, definitions, assumptions, and theorems. Many notations and terms are used without defining or explaining. Even the crucial concepts including nodes, edges, dyads, affinity, clock, switching rates, etc. are never properly defined. The theoretical results starting from section 3 are full of abstract math expressions without explanations on the idea or the intuition. It is thus difficult to justify the soundness of the results, and also it is not clear how the theorems connects to the proposed method. It can be helpful to instantiate those with meaningful concrete concepts or practical applications, and include examples.

The framework also seems overly complicated. As also mentioned above, it is hard to gauge how each theorem or other theoretical result connects to the main method. I think the paper can benefit from simplifying the framework and making the high-level picture clear. As for the current paper, the problem and solution strategy are both not specified clearly enough.

The model defined in section 2 is parameterized without justification. It is unclear if the model definition is standard or a specific choice. There is no explanation on how the functions and parameters are chosen and how general they are.

Questions for authors

None.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.