Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Mobility data teaches maps what a place is and how it is used, and the resulting embeddings beat text-only models on five map-enrichment tasks.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 07:03 UTC pith:6N7LX4O7

load-bearing objection Solid architecture and broad experiments, but the main evaluation tasks leak the training signal, so the 'function' claim is not yet proven. the 4 major comments →

arxiv 2601.21149 v3 pith:6N7LX4O7 submitted 2026-01-29 cs.LG cs.AI

Mobility-Embedded POIs: Learning What A Place Is and How It Is Used from Human Movement

Maria Despoina Siampou , Shushman Choudhury , Shang-Ling Hsu , Neha Arora , Cyrus Shahabi This is my paper
classification cs.LG cs.AI
keywords point-of-interest representationshuman mobilitycontrastive learningmap enrichmentPOI functionlong-tail sparsitytext embedding alignmentvisit distribution transfer
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper sets out to prove that 'POI function'—how a place is actually used, as opposed to what its static description says—is a missing and essential signal for general point-of-interest representations. It introduces ME-POIs, a framework that learns place embeddings from human visit sequences and combines them with text-based embeddings. Across five map-enrichment tasks (opening hours, closure detection, visit intent, busyness, price level), adding ME-POIs consistently improves over both text-only and mobility-only baselines, with particularly large gains on function-focused tasks. The paper argues this demonstrates that mobility-derived function should be part of any general-purpose POI representation.

Core claim

The central claim is that place identity (what a place is) and place function (how it is used) are complementary, and that encoding both in a single embedding improves downstream map maintenance. ME-POIs learns a prototype embedding per POI by contrasting temporally contextualized visit embeddings—built from location, arrival time, and departure time—against the prototypes of other POIs in a batch. To handle rarely visited places, it transfers empirical weekly visit distributions from frequent nearby 'anchor' POIs to sparse ones using multi-scale Gaussian kernels, and supervises anchor prototypes directly with their own visit histograms. A text-alignment loss grounds the mobility embedding i

What carries the argument

The load-bearing mechanism is the contrastive alignment between visit-level embeddings and learnable POI prototypes (InfoNCE loss). Each visit is encoded by a Transformer over Space2Vec location encodings and Time2Vec arrival/departure encodings; the prototype acts as a functional centroid that aggregates usage patterns across users and times. The multi-scale distribution transfer—a Gaussian-kernel weighted average of anchor visit histograms, injected via KL-divergence supervision—is the component that tackles long-tail sparsity, and the cosine text-alignment loss is what merges mobility-derived function with static textual identity.

Load-bearing premise

The evaluation labels for busyness and visit intent are mostly derived from the same visit times and counts that the model sees as input, so the measured gains on those tasks may partly reflect the model memorizing its own inputs rather than learning an independent notion of function.

What would settle it

Train ME-POIs on the same visit sequences but replace arrival/departure times with random permutations (or drop them entirely), then probe the frozen embeddings on the five tasks. If busyness and visit intent gains mostly disappear while opening-hours and price-level gains persist, the claimed function signal is largely direct memorization of the label-generating statistics. A complementary check is to compare against a trivial baseline that uses binned visit-count histograms as features; if that baseline matches ME-POIs on busyness and visit intent, the contrastive architecture adds little be

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If ME-POIs works as claimed, map platforms can maintain richer place data largely from anonymized mobility traces, predicting opening hours, closures, busyness, and price level without relying on stale or incomplete metadata.
  • The framework shows that function-focused embeddings are complementary to text-only embeddings, suggesting future foundation models for places should be trained on both static and behavioral signals.
  • The multi-scale transfer mechanism offers a recipe for learning representations of long-tail places, which is a recurring bottleneck in location-based machine learning.
  • The same mobility-to-embedding pipeline could be adapted to other geospatial objects—road segments, administrative regions, or building footprints—as the paper itself suggests for future work.
  • Because ME-POIs without text alignment already outperforms trajectory-based mobility models, it reorients mobility-based representation learning from next-location prediction toward POI-centric semantics.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A likely source of the large gains on busyness and visit intent is that these labels are computed from the same visit counts and temporal histograms that define the visit embeddings; the paper's central 'function' claim would be stronger if evaluated against labels that are not derived from the mobility input.
  • One testable extension is to ablate the temporal encodings (arrival/departure times) and retrain: if the benefits on opening-hours and busyness vanish while identity tasks hold, the model is largely memorizing label statistics rather than learning transferable function.
  • The distribution transfer idea could be pushed further: instead of transferring only weekly histograms, one could transfer learned prototype vectors or category-conditioned priors, which might help for POIs with no nearby anchors.
  • The method implies a practical deployment pattern: continuously update ME-POIs embeddings as new visit data arrive, enabling map enrichment to track real-world changes (closures, new popular places) faster than human-authored metadata updates.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes ME-POIs, a framework that augments text-based POI embeddings with signals from large-scale mobility data. Visit sequences are encoded with a location encoder, Time2Vec arrival/departure embeddings, and a Transformer; the resulting visit embeddings are aligned with learnable POI prototypes via an InfoNCE loss. A multi-scale kernel transfer mechanism propagates empirical visit distributions from frequent ('anchor') POIs to sparse POIs, and KL-divergence losses supervise both anchor and sparse prototypes to reproduce hourly visit distributions. A text-alignment loss optionally aligns the mobility prototypes with text embeddings. The method is evaluated on five map-enrichment tasks (opening hours, permanent closure, visit intent, busyness, price level) on Los Angeles and Houston mobility datasets, comparing against six text embeddings and eleven mobility baselines. The paper claims that augmenting text embeddings with ME-POIs consistently improves all tasks, and that mobility alone can sometimes beat strong text-only models.

Significance. If the claims held, ME-POIs would be a useful contribution to geospatial representation learning: it is a concrete, POI-centric alternative to trajectory-prediction objectives, with large-scale experiments, broad baselines, and ablations isolating each loss term. The multi-scale distribution-transfer mechanism is a sensible approach to the long-tail POI problem. However, the central evidence is weakened by a self-referential evaluation design: several downstream labels are summaries of the same temporal visit signal that the model consumes and is explicitly trained to reproduce. The claimed 'consistent' improvement is also contradicted by at least one reported drop, and no uncertainty quantification is provided. The significance of the paper is therefore conditional on additional control experiments that separate genuine functional knowledge from the model's ability to recite its input statistics.

major comments (4)
  1. [§4.2, Appendix A.1.2, Eq. (2), §3.4] Opening-hours and busyness are not independent tests of the claim that ME-POIs captures a missing 'function' signal. The model input already contains arrival and departure times (Eq. 2), and the pretraining objective directly supervises anchor prototypes to reproduce empirical visit distributions r_pa over hourly bins (§3.4). Appendix A.1.2 defines opening hours as a 168-dim open/closed vector and busyness as a weekly average of hourly activity levels — essentially summaries of the same histograms. A model that simply memorized or smoothed per-POI visit-time histograms would be expected to do well on these tasks, so the large relative gains (e.g., +16.2% F1 for opening hours, -24.7% MAE for busyness) do not establish generalizable functional knowledge. I request control experiments: report the performance of simple baselines built directly from the empirical visit distributions (e.g., a
  2. [§4.2, Table 1] The abstract and §4.2 claim that augmenting text embeddings 'consistently' outperforms text-only baselines. Table 1 contradicts this: GEMINI permanent closure F1 decreases from 0.756 to 0.753 (-0.4%), and OPENAI-SMALL improves by only +0.1% F1 on the same task. The paper reports means over 5 runs but gives no standard deviations or significance tests, so it is unclear which of the many 1–3% relative improvements are reliable. Given that the headline claim is 'consistent' improvement, the paper should either hedge the claim or provide error bars and significance testing over the five runs.
  3. [§4.2, Tables 3–4, Table 7] Price-level evidence is fragile. The text-only and mobility-only baselines have very low absolute F1 values (e.g., 0.177–0.329 in Los Angeles, 0.177–0.294 in Houston), often near the performance of a trivial classifier on an imbalanced task. Table 7 shows the highest price class has only 36 samples in Los Angeles and 27 in Houston, making F1 unstable and relative improvements (e.g., +75.1%) potentially driven by a handful of examples. The paper should report per-class precision/recall, macro-averaged metrics with confidence intervals, and consider a balanced evaluation protocol before using price-level gains as evidence for the method.
  4. [§4.2, Appendix A.1.2] Visit intent is defined from Google Maps direction queries, which are likely correlated with overall visit volume. ME-POIs is trained with visit counts and arrival/departure times, so strong performance on visit-intent may reflect popularity rather than a distinct functional signal. To support the functional interpretation, I ask for an analysis that controls for total visit frequency (e.g., stratify by visit-count quartiles, or include total visits as an input feature to the baselines). This is important because visit-intent is one of the tasks used to argue that ME-POIs captures 'how a place is used' beyond movement regularities.
minor comments (5)
  1. [Throughout] The method name is inconsistently typeset as 'ME-POIs' and 'ME-POIS'; please standardize. Also fix typos: 'Sine mobility-based models' should be 'Since'; 'intristic' should be 'intrinsic'; 'the these representations' in the introduction; 'do not longer exist' in Appendix A.1.2.
  2. [§4.3, Table 5] The ablation table caption says 'for open hours in Houston' but the table contains both Los Angeles and Houston columns; clarify.
  3. [§3.3] Eq. (10) uses a normalized Gaussian kernel, but the normalization is over anchors at each scale; this means the multi-scale average in Eq. (11) is equally weighted across scales. It would be helpful to state explicitly whether equal scale weights are intentional or whether the scales should be learned/validated.
  4. [Figure 4] The figure compares models on anchor vs sparse POIs, but the caption does not state how 'anchor' and 'sparse' are defined in the evaluation split. Please add this information.
  5. [Appendix A.1.4] The hyperparameter section says 'λ_α = λ_s = λ_t = 1', but the loss in Eq. (17) uses λ_a, λ_s, λ_t; the subscript α in the appendix is a typo.

Circularity Check

1 steps flagged

Opening-hours and busyness 'predictions' read out the same weekly hourly visit histogram the embeddings are explicitly trained to reproduce; part of the headline function gains re-measure the input.

specific steps
  1. fitted input called prediction [§3.3–3.4 (Eqs. 11, 15); §A.1.2; §4.1]
    "For each anchor POI pa ∈ Panchor, we construct an empirical visit distribution rpa ∈∆T by aggregating visits into T fixed temporal bins (e.g., hourly slots over a week) and normalizing the resulting histogram. ... We represent this as a 168-dimensional binary vector, where each dimension corresponds to one hour of the week ... prediction of busyness, as a weekly average of hourly activity levels"

    LKL-anchor (Eq. 15) explicitly trains the anchor embedding to output qθ(pa) ≈ rpa, the normalized histogram of visits over weekly hourly bins. The 'Weekly Opening Hours' label is the open/closed version of the same 168 hourly bins, and 'Busyness Estimation' is 'a weekly average of hourly activity levels'—the same histogram. So on these two headline tasks (+16.2% F1; −24.7% MAE) the downstream head decodes the visit-time distribution the embedding was built to encode; gains over text-only models largely re-measure the input mobility signal rather than an independent 'function' signal. Pretraining on the full dataset before the 60/20/20 split (A.1.4) means test POI histograms were directly seen, compounding the overlap.

full rationale

The derivation of the ME-POIs embeddings themselves is not circular: pretraining is self-supervised on visit times and never fits downstream labels, and the contrastive/text-alignment components are standard. The circularity is confined to the evaluation design. Three of the five 'function' tasks—opening hours, busyness, and to a lesser degree visit intent—are summaries of the same temporal visit signal that Eq. (2) feeds into the encoder and that Eqs. (13)–(15) explicitly supervise the embeddings to reproduce. Busyness in particular is, up to normalization, the exact weekly hourly histogram rpa that anchor embeddings are trained to predict, so the reported improvements on busyness (up to −24.7% MAE) and opening hours (up to +16.2% F1) are expected from reading out the fitted histogram, not independent evidence of a novel generalizable 'function' representation. The permanent-closure and price-level tasks give some independent grounding, but the price-level baselines are very low (F1 ≈ 0.18–0.33), making relative gains fragile, and closure gains are modest (≤6.5% F1). The paper's self-citations (Siampou et al. 2025a,b; Hsu et al. 2024; Chu & Shahabi 2025) are used for context, encoder alternatives, or baselines, not as load-bearing mathematical or empirical support, so no self-citation circularity is present. Overall: partial circularity, score 6.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

No new physical entities are introduced. The framework is assembled from standard components plus hand-set hyperparameters; the main domain risk is that temporal visit patterns and the chosen tasks are both author-defined proxies for 'function,' which makes the benchmark partly self-referential.

free parameters (5)
  • Loss weights λ_a, λ_s, λ_t = 1.0 each (λ_a=λ_s=λ_t=1)
    Weights in Eq. (17) balancing contrastive, anchor-KL, sparse-KL, and text-alignment losses; set to 1 without sensitivity analysis.
  • Anchor visit threshold M = 100 (Los Angeles), 50 (Houston)
    POIs with at least M visits are anchors (A.1.1), determining which POIs receive direct supervision vs. transfer; chosen per dataset.
  • Spatial bandwidths σ_m = 0.3, 1.0, 3.0 km
    Multi-scale kernel bandwidths in Eq. (10) for distribution transfer; chosen by hand, no search reported.
  • InfoNCE temperature τ = not reported
    Temperature in Eq. (9) controls contrastive sharpness; value is omitted from hyperparameters.
  • Architecture sizes = w=32, d_h=512, N=4, heads=8, MLP hidden 256
    Transformer and MLP capacities in A.1.4; conventional, but no evidence they were externally benchmarked.
axioms (5)
  • domain assumption Visit timestamps (arrival/departure) and staypoint attribution faithfully encode how a POI is used
    Entered in Eq. (2) and A.1.1; if GPS staypoints/POI attribution introduce noise or duplicates, the function signal is corrupted.
  • domain assumption Nearby anchor POIs have similar temporal visit patterns, so visit-distribution transfer is valid
    Section 3.3 states POIs of the same urban environment tend to exhibit similar activity patterns; this motivates Eqs. (10)-(13).
  • domain assumption GeoLLM-style prompts produce text embeddings that encode POI identity
    Section 3.5 uses prompt design from Manvi et al. 2024 without validating the prompt separately.
  • ad hoc to paper Map-enrichment tasks and label sources (SafeGraph, Google Maps) measure POI identity and function
    Section 4.1 introduces five tasks; no external benchmark or consensus task set is used, so the yardstick is defined by the authors.
  • standard math Standard building blocks (Space2Vec, Time2Vec, Transformer, InfoNCE) behave as claimed
    Component choices in Sections 3.1-3.2 rely on prior methods; no proofs are reproduced.

pith-pipeline@v1.3.0-alltime-deepseek · 18449 in / 15676 out tokens · 163233 ms · 2026-08-03T07:03:49.056412+00:00 · methodology

0 comments
read the original abstract

Recent progress in geospatial foundation models highlights the importance of learning general-purpose representations for real-world locations, particularly points-of-interest (POIs) where human activity concentrates. Existing approaches, however, focus primarily on place identity derived from static textual metadata, or learn representations tied to trajectory context, which capture movement regularities rather than how places are actually used (i.e., POI's function). We argue that POI function is a missing but essential signal for general POI representations. We introduce Mobility-Embedded POIs (ME-POIs), a framework that augments POI embeddings derived, from language models with large-scale human mobility data to learn POI-centric, context-independent representations grounded in real-world usage. ME-POIs encodes individual visits as temporally contextualized embeddings and aligns them with learnable POI representations via contrastive learning to capture usage patterns across users and time. To address long-tail sparsity, we propose a novel mechanism that propagates temporal visit patterns from nearby, frequently visited POIs across multiple spatial scales. We evaluate ME-POIs on five newly proposed map enrichment tasks, testing its ability to capture both the identity and function of POIs. Across all tasks, augmenting text-based embeddings with ME-POIs consistently outperforms both text-only and mobility-only baselines. Notably, ME-POIs trained on mobility data alone can surpass text-only models on certain tasks, highlighting that POI function is a critical component of accurate and generalizable POI representations.

Figures

Figures reproduced from arXiv: 2601.21149 by Cyrus Shahabi, Maria Despoina Siampou, Neha Arora, Shang-Ling Hsu, Shushman Choudhury.

Figure 1
Figure 1. Figure 1: Illustration of ME-POIS. ME-POIS augment static text-based POI representations with mobility-derived signals, to learn POI embeddings that capture their identity and function. 2024). While effective at modeling movement regularities, these learned POI embeddings are optimized for trajectory prediction and therefore do not capture POI function. For instance, a gym and a bar near the same office may both be … view at source ↗
Figure 2
Figure 2. Figure 2: Overview of ME-POIS pretraining. The framework includes: (i) a transformer-based visit sequence encoder, (ii) contrastive alignment of contextualized visits (h) with global POI embeddings (z ME p ) to capture usage patterns, (iii) multi-scale distribution transfer to propagate temporal visit information to under-visited POIs (ps), (iv) direct supervision on anchor POIs (pa) to regularize embeddings via vis… view at source ↗
Figure 3
Figure 3. Figure 3: Impact of LKL-anchor and LKL-sparse on sparse and anchor POIs on open hours in Houston. Anchors Sparse 0.2 0.4 0.6 0.8 F1 Score OpenAI-large Gemini CTLE TrajGPT ME-POIs (Ours) [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Example prompt for Taco Man POI in Los Angeles. A.3. Computational Efficiency The pre-training cost of ME-POIs is dominated by running the visit encoder on sequences of visits. For a sequence length of L and an embedding dimension d, the overall computation complexity is O(L 2 · d + L · d 2 ) for. The contrastive module 13 [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. TRAJGANR: Trajectory-Centric Urban Multimodal Learning via Geospatially Aligned Neural Representations

    cs.CV 2026-05 unverdicted novelty 7.0

    TrajGANR learns continuous neural representations of trajectories to enable fine-grained alignment with street-view images and locations in a joint multimodal self-supervised objective, outperforming prior geospatial ...

Reference graph

Works this paper leans on

15 extracted references · 8 linked inside Pith · cited by 1 Pith paper

  1. [3]

    Deepmove: Predicting human mobility with attentional recurrent networks

    Feng, J., Li, Y ., Zhang, C., Sun, F., Meng, F., Guo, A., and Jin, D. Deepmove: Predicting human mobility with attentional recurrent networks. InProceedings of the 2018 world wide web conference, pp. 1459–1468,

    2018

  2. [6]

    Due to its year-long temporal coverage, the Los Angeles dataset contains approximately an order of magnitude more visits than the Houston dataset, which spans 20 days

    The number of POIs in Los Angeles and Houston is comparable, although Los Angeles covers a larger geographic region and therefore contains more POIs. Due to its year-long temporal coverage, the Los Angeles dataset contains approximately an order of magnitude more visits than the Houston dataset, which spans 20 days. Table 6.Summary of Datasets Statistics....

    2019

  3. [7]

    Stan: Spatio-temporal at- tention network for next location recommendation

    Luo, Y ., Liu, Q., and Liu, Z. Stan: Spatio-temporal at- tention network for next location recommendation. In Proceedings of the web conference 2021, pp. 2177–2185,

    2021

  4. [8]

    9 Mobility-Embedded POIs: Learning What A Place Is and How It Is Used from Human Movement Mikolov, T., Chen, K., Corrado, G., and Dean, J

    URL https://openreview.net/forum? id=TqL2xBwXP3. 9 Mobility-Embedded POIs: Learning What A Place Is and How It Is Used from Human Movement Mikolov, T., Chen, K., Corrado, G., and Dean, J. Efficient estimation of word representations in vector space.arXiv preprint arXiv:1301.3781,

    Pith/arXiv arXiv

  5. [10]

    Oord, A. v. d., Li, Y ., and Vinyals, O. Representation learn- ing with contrastive predictive coding.arXiv preprint arXiv:1807.03748,

    Pith/arXiv arXiv

  6. [11]

    Geographic location encoding with spherical harmonics and sinusoidal representation networks.arXiv preprint arXiv:2310.06743,

    Rußwurm, M., Klemmer, K., Rolf, E., Zbinden, R., and Tuia, D. Geographic location encoding with spherical harmonics and sinusoidal representation networks.arXiv preprint arXiv:2310.06743,

    Pith/arXiv arXiv

  7. [12]

    Text embeddings by weakly-supervised contrastive pre-training.arXiv preprint arXiv:2212.03533,

    Wang, L., Yang, N., Huang, X., Jiao, B., Yang, L., Jiang, D., Majumder, R., and Wei, F. Text embeddings by weakly-supervised contrastive pre-training.arXiv preprint arXiv:2212.03533,

    Pith/arXiv arXiv

  8. [15]

    DETAILS ONBASELINES We evaluate ME-POISagainst both text- and mobility-based baselines

    Table 7.Visit Intent and Price Level Counts Los Angeles Houston Class Visit Intent Price Level Visit Intent Price Level 0 12840 2563 7158 2270 1 1376 2311 979 1675 2 5654 181 4841 133 3 2499 36 2654 27 A.1.3. DETAILS ONBASELINES We evaluate ME-POISagainst both text- and mobility-based baselines. For text-based comparisons, we select competitive and widely...

    2013

  9. [2008]

    Spabert: a pretrained language model from geographic data for geo- entity representation.arXiv preprint arXiv:2210.12213,

    Li, Z., Kim, J., Chiang, Y .-Y ., and Chen, M. Spabert: a pretrained language model from geographic data for geo- entity representation.arXiv preprint arXiv:2210.12213,

    Pith/arXiv arXiv

  10. [2013]

    H., Ma, J., Zhao, V ., Luan, Y ., Hall, K., Chang, M.-W., et al

    Ni, J., Qu, C., Lu, J., Dai, Z., Abrego, G. H., Ma, J., Zhao, V ., Luan, Y ., Hall, K., Chang, M.-W., et al. Large dual encoders are generalizable retrievers. InProceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pp. 9844–9855,

    2022

  11. [2017]

    Appendix A.1

    10 Mobility-Embedded POIs: Learning What A Place Is and How It Is Used from Human Movement A. Appendix A.1. Additional Details on Experimental Setup A.1.1. DATASETSTATISTICS Table 6 summarizes key statistics of the two mobility datasets used in our experiments. Both datasets consist of anonymized raw GPS trajectories, containing timestamped geographic coo...

    2008

  12. [2021]

    Gatgpt: A pre-trained large language model with graph attention network for spa- tiotemporal imputation.arXiv preprint arXiv:2311.14332,

    Chen, Y ., Wang, X., and Xu, G. Gatgpt: A pre-trained large language model with graph attention network for spa- tiotemporal imputation.arXiv preprint arXiv:2311.14332,

    Pith/arXiv arXiv

  13. [2022]

    Geolm: Empowering language models for geospatially grounded language understanding

    Li, Z., Zhou, W., Chiang, Y .-Y ., and Chen, M. Geolm: Empowering language models for geospatially grounded language understanding. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pp. 5227–5240,

    2023

  14. [2024]

    M., Goel, R., Eghbali, S., Ramanan, J., Sa- hota, J., Thakur, S., Wu, S., Smyth, C., Poupart, P., and Brubaker, M

    Kazemi, S. M., Goel, R., Eghbali, S., Ramanan, J., Sa- hota, J., Thakur, S., Wu, S., Smyth, C., Poupart, P., and Brubaker, M. Time2vec: Learning a vector representation of time.arXiv preprint arXiv:1907.05321,

    Pith/arXiv arXiv 1907

  15. [2025]

    and Shahabi, C

    Chu, C. and Shahabi, C. Geo2vec: Shape-and distance- aware neural representation of geospatial entities.arXiv preprint arXiv:2508.19305,

    Pith/arXiv arXiv