acceptodds
Under review as a conference paper at ICLR 2027

TimeArrowMob: A Controlled Benchmark for Temporal Directionality and Irreversibility in Human Mobility and Beyond

Abstract

Real-world human mobility trajectories provide rich and realistic temporal structure, but their true temporal directionality and time irreversibility are unknown, leaving existing methods without reliable ground truth for evaluation. Conversely, trajectories generated from hand-crafted rules or a single physical model can provide controlled directionality but may differ substantially from complex human mobility. We introduce TimeArrowMob, a controlled benchmark that combines real mobility structure with analytically computable temporal-direction ground truth. Taking each YJMob100K user trajectory as an independent unit, we construct a reversible baseline from its locations and symmetrized transition counts. We then impose controllable circulation constraints on transition counts and generate closed trajectories through Eulerian circuits, matching trajectory length, location frequencies, undirected transition counts, and self-transition counts across directions and strength levels. The construction generates paired Forward/Reverse trajectories with exact direction labels and additive local log-likelihood-ratio contributions relative to a fixed reference mechanism, while reversal-symmetric sampling provides zero-direction controls. The strength of the temporal arrow is explicitly controlled through the circulation scale and computed as the normalized magnitude of the summed reference contributions. Grounded in time-reversal symmetry and probabilistic inference, TimeArrowMob provides realistic trajectory morphology together with exact supervision for direction, strength, and localization. The framework enables systematic evaluation of irreversibility estimators, learning-based direction predictors, and attribution methods, and can be extended to other time-series domains with domain-valid reversal units.

Primary area:
datasets and benchmarks
Keywords:
Human Mobility, Temporal Directionality, Temporal Irreversibility, Arrow of Time, Time-Reversal Symmetry, Controlled Benchmark, Trajectory Analysis, Ground-Truth Attribution

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_7w2BOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces TimeArrowMob, a controlled benchmark for studying temporal directionality in human mobility prediction. It modifies transition counts through circulation constraints and generates trajectories using Eulerian circuits. This construction provides exact direction labels and computable reference strengths while matching selected statistics across directions and strength levels. The paper evaluates nine existing methods using prediction accuracy, directional response R, and historical-order benefit G.

Critical strengths

The main contribution is a controlled way to vary directional information while keeping several potentially confounding statistics unchanged, including location frequencies, undirected transition counts, and self-transition counts. The combination of circulation constraints and Eulerian circuits makes these controls exact. This provides a useful basis for examining temporal directionality beyond ordinary prediction accuracy.

Critical weaknesses

The evaluated methods do not show consistently positive R and G across strength levels. However, the experiments do not include a model known to use the injected directional information, so it remains unclear whether these measurements can reliably detect such use.

Questions for authors

Can the authors add an oracle given reference mechanism and a history-only transition-count baseline to demonstrate that the visible histories contain usable directional information and that the proposed metrics detect its use?

1 comment
AuthorsOpenReview ↗

Assessing Directional Evidence and Benchmark Sensitivity

Yes. Your suggestion identifies a gap in our evaluation: the last-location baseline shows what zero response looks like, but it does not establish whether the 45-location histories contain usable directional evidence or whether RR detects a predictor acting on that evidence. We have implemented the two controls you suggest.

Reference-informed control. The predictor receives the two reference transition matrices P+P^+ and P−P^-, but neither the source-direction label nor the next location. For each history, it sums log⁡(Pij+/Pij−)\log(P^+_{ij}/P^-_{ij}) over the 44 observed transitions. We first evaluate how well this score distinguishes windows drawn from forward versus reverse trajectories, using within-user ROC AUC at each strength level. The predictor then uses the transition matrix selected by the score to predict the most likely next location. This separates two questions: can the history reveal its source orientation, and does using that information change the top-1 prediction? Even a distinguishable orientation need not change the prediction if both matrices favor the same next location.

History-only control. This predictor receives no reference matrices or direction labels. It predicts the most frequent successor of the current location among transitions observed in the visible history. If the history contains no earlier departure from that location, it predicts the current location again.

We will evaluate both controls on the original paired queries, using the same RR, the same eligible histories and interventions for GG, and user-level confidence intervals. The full-cohort estimates are not yet available. Until we have them, our conclusion is limited to what we observed in the nine frozen predictors; their results alone do not establish the sensitivity of the benchmark.

Reviewer_o5JsOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This work introduce TimeArrowMob, a benchmark that combines real mobility structure with analytically computable temporal-direction ground truth. The proposed benchmark construct a reversible baseline from its locations and symmetrized transition counts, based on YJMob100K dataset. The evaluation is extended to irreversibility estimators, learning-based direction predictors, and attribution methods.

Critical strengths
  1. The proposed problem is well-motivated. This work addresses that the forecast accuracy does not reveal whether a model uses temporal order.
  2. Based on previous dataset, the construction of the proposed benchmark provides an exact direction label and a computable strength for every generated trajectory.
  3. Beyond accuracy, this work proposed two new evaluation metrics and evaluated multiple existing methods to show their performance on both accuracy and direction awareness.
Critical weaknesses
  1. From the main results in Table 2, R and G feels difficult to interpret and do not show how much R and G change across strength levels. The corresponding actual values in the appendix table appears to have very small values. So even with the new metrics, it is still unclear of how well the models respond to temporal direction signal.

  2. The results are aggregated by level names (weak, medium, strong, strongest). It would be helpful to show quantitatively how the metrics change as the directional signal increases.

  3. Currently no evaluated models shows a positive R, it in unclear that whether the benchmark can detect direction usage when it exists.

Questions for authors

Same as in weaknesses.

2 comments

Thank you. We agree that the signs in Table 2 conceal the magnitudes of the effects. MoVER illustrates why this matters. The following are its existing results from Table D.3; brackets give pointwise 95% user-bootstrap intervals.

Within-user levelDirectional response RR, multiplied by 1,000Historical-order benefit GG, percentage points
Weak−0.396 [−0.814, −0.00084]+0.307 [+0.099, +0.515]
Medium−2.743 [−3.213, −2.288]+0.176 [−0.056, +0.408]
Strong−4.329 [−4.872, −3.803]+0.444 [+0.235, +0.649]
Strongest−6.576 [−7.248, −5.946]+0.347 [+0.083, +0.650]

At the strongest level, MoVER was 0.347 percentage points more accurate with the original history than after reversal of the selected closed segment, on the same eligible queries. Yet its R=−0.006576R=-0.006576: its predicted transitions shifted, on average, against the injected reference direction. Thus a predictor can benefit from historical order without favoring the particular directional signal we injected. RR is a normalized edge-weight contrast, not an accuracy difference.

We also agree that the four level names do not give the numerical strength: “strong” for one user need not equal “strong” for another. We will report the realized strength in nats per transition (median and interquartile range across users), together with the within-user change in RR and GG from weakest to strongest and its confidence interval.

Finally, none of the 36 model–level intervals for RR was wholly above zero. That finding alone does not tell us whether the benchmark detects reference-aligned behavior when it is present. To address this, we have implemented a reference-informed predictor and a history-only transition-count baseline. We will report their results on the original queries alongside the frozen predictors.

Thank authors for the response. But main concern remains. The current results do not sufficiently demonstrate that the benchmark can detect direction use when it exists. Also the results are still reported only by level names, without showing quantitatively how the metrics change as the directional signal increases. I will keep my rating.

Reviewer_9dzAOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The author(s) introduce a controlled benchmark for analysing human mobility trajectories. The data augments a real-world mobility dataset, i.e. YJMob100K, with analytically generated information. For the augmented information, the authors consider transition counts and generated closed trajectories. The new semi-synthetic dataset provides paired Forward/Reverse trajectories with exact direction labels.

Critical strengths

The temporal aspect of the human mobility prediction is discussed.

Critical weaknesses

Several synthetic operations are proposed without sufficient justification.

Questions for authors

No

1 comment
AuthorsOpenReview ↗

Motivation and Design Rationale for TimeArrowMob

Thank you for pointing out that the motivation for the synthetic operations was difficult to follow. Our goal is to answer a specific question: if we change first-order directional evidence while keeping location frequencies and undirected transition counts fixed, do a predictor’s outputs change in the corresponding direction?

Our work is motivated by a validation challenge in studying temporal directionality in real human mobility trajectories. Although the chronological order of observations is known, it does not directly provide ground truth for statistical temporal asymmetry, its strength, or its local contributions. Consequently, results on real data alone are insufficient to fully assess either the effectiveness of directionality detection methods or whether the local segments identified by an algorithm provide directional evidence. Intuitions about temporal direction in human behavior can motivate research questions, but they are insufficient as an evaluation criterion.

With this consideration, we aim to construct a controlled dataset that balances three requirements:

  1. Preserving selected statistical properties of real trajectories, so that the test data retain a connection to human mobility.
  2. Providing direction labels and computable reference strengths under an explicitly defined reference mechanism, thereby establishing a verifiable basis for evaluation.
  3. Allowing reference directional evidence to be decomposed additively into contributions from individual transitions, supporting the analysis and evaluation of local directional attribution.

The synthetic operations in TimeArrowMob are designed around these objectives: they construct data with known reference directional signals and controlled strengths while preserving selected statistical properties.

We acknowledge that this semi-synthetic construction simplifies the complex temporal structure of real human mobility, and that its conclusions are limited by the reference mechanism and evaluation protocol. We therefore position TimeArrowMob as a diagnostic tool with explicit controls, intended to assess whether algorithms can detect and use the constructed directional information. Generalizing its findings to real human mobility trajectories requires further validation.

A three-location example illustrates the construction. We first symmetrize the observed transition counts so that each edge has the same count in both directions. We can then add one traversal to each edge of the cycle A→B→C→AA\to B\to C\to A and subtract one from each reverse edge, provided those counts remain feasible. This changes the preferred orientation without changing how often each location is visited, how often each undirected edge is traversed, or how often the trajectory stays in place. The Eulerian circuit turns these prescribed counts into an actual sequence that a next-location predictor can consume, using each directed edge exactly as many times as specified.

The resulting direction is known because we imposed it. We do not claim that it recovers the original person’s direction of travel or preserves their daily routine. We will use this example in Section 3 to explain what each operation contributes and what the construction preserves.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.