pith. sign in

arxiv: 2606.10938 · v1 · pith:TICYREM3new · submitted 2026-06-09 · 💻 cs.LG

A Systematic Approach for Selecting Trajectories for Data Augmentation

Adam Nordling This is my paper

Pith reviewed 2026-06-27 13:45 UTC · model grok-4.3

classification 💻 cs.LG
keywords trajectory data augmentationselection strategiesoutliernessuncertaintydata densityspatio-temporal coherencemachine learningdata scarcity
0
0 comments X

The pith

Systematic selection of trajectories for augmentation yields more stable results than random choice but only when data is sparse.

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

The paper develops a framework to test which trajectories to augment using geometric perturbations, comparing five strategies including Outlierness, Diversity, Representativeness, Uncertainty, and Random across animal, maritime, and urban datasets. It shows that Outlierness and Uncertainty deliver higher stability and avoid the performance degradation that random sampling produces in dense data. The work also finds that augmentation itself only helps under conditions of data sparsity and lower velocities, as it repairs gaps in sparse sets but adds noise in dense ones and diverges in fast-moving domains.

Core claim

The evaluation reveals that systematic selection strategies, particularly Outlierness and Uncertainty, provide higher stability than random sampling and reduce the performance degradation seen with random choice in dense datasets. Visual analysis via UMAP shows that systematic augmentation repairs topological fragmentation in sparse data yet acts as corrupting noise in high-quality dense data. The value of augmentation is strictly conditional on data density and domain velocity, with standard perturbation techniques producing feature-space divergence in high-velocity regimes.

What carries the argument

A comparative evaluation framework that applies five selection strategies (Outlierness, Diversity, Representativeness, Uncertainty, Random) to choose trajectories for geometric perturbation, combined with Optuna hyperparameter optimization per dataset.

If this is right

  • Systematic strategies like Outlierness and Uncertainty maintain performance stability where random selection causes degradation in dense datasets.
  • Augmentation repairs topological fragmentation in sparse datasets but introduces corrupting noise in dense ones.
  • Standard geometric perturbations lead to feature-space divergence in high-velocity domains.
  • The benefit of augmentation depends on the density and velocity characteristics of the domain.

Where Pith is reading between the lines

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

  • A preliminary density check on any new dataset could guide whether to apply systematic selection or skip augmentation entirely.
  • Velocity-aware perturbation methods would be needed to make the framework useful in additional high-speed domains.
  • The conditional value of augmentation may extend to other forms of sequential data where coherence must be preserved.

Load-bearing premise

The geometric perturbation methods preserve enough spatio-temporal coherence for the augmented trajectories to remain valid training examples in every domain and velocity regime tested.

What would settle it

A new experiment on a dense, high-velocity trajectory dataset in which Outlierness and Uncertainty selection produces performance degradation at least as severe as random selection would falsify the stability advantage.

Figures

Figures reproduced from arXiv: 2606.10938 by Adam Nordling.

Figure 3.1
Figure 3.1. Figure 3.1: Trajectory Data Augmentation & Evaluation Pipeline [PITH_FULL_IMAGE:figures/full_fig_p013_3_1.png] view at source ↗
Figure 4.2
Figure 4.2. Figure 4.2: Visual analysis of Manifold Densification (Foxes, Seed 1415). The Outlier [PITH_FULL_IMAGE:figures/full_fig_p032_4_2.png] view at source ↗
Figure 4.3
Figure 4.3. Figure 4.3: Visual analysis of Decision Margin Corruption (Starkey, Seed 2862). The [PITH_FULL_IMAGE:figures/full_fig_p033_4_3.png] view at source ↗
Figure 4.4
Figure 4.4. Figure 4.4: Visual analysis of Feature Space Divergence (Car Traffic, Seed 3832). Due to [PITH_FULL_IMAGE:figures/full_fig_p034_4_4.png] view at source ↗
read the original abstract

Trajectory data augmentation is a promising approach to mitigate data scarcity in machine learning applications, but its utility has been limited by the complexity of preserving spatio-temporal coherence. Although prior work demonstrated the viability of geometric perturbation, it relied on naive random selection, leaving a critical gap in understanding which trajectories should be augmented for maximal benefit. This thesis addresses this gap by developing a systematic and scalable framework to evaluate five systematic selection strategies: Outlierness, Diversity, Representativeness, Uncertainty, and Random selection. These strategies were rigorously tested across four datasets covering animal behavior (Foxes and Starkey), maritime traffic (AIS), and urban traffic (Car) using a suite of linear and non-linear machine learning models. As part of this evaluation, an Optuna-based hyperparameter optimization loop was integrated to empirically identify the best-performing augmentation parameters for each dataset within the explored search space. The results indicate that, while systematic selection is not a universal solution, it offers distinct advantages over the random baseline. Systematic strategies, particularly Outlierness and Uncertainty, demonstrated higher stability and were less prone to performance degradation observed with random sampling in dense datasets. However, the findings also reveal that the value of augmentation is strictly conditional. Visual analysis via UMAP demonstrates that while systematic augmentation successfully repairs topological fragmentation in sparse datasets, it can act as a corrupting noise signal in high-quality, dense datasets. Furthermore, the study identified physical limitations in high-velocity domains, where standard perturbation techniques lead to divergence in feature space...

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript develops a systematic framework for selecting which trajectories to augment via geometric perturbations, evaluating five strategies (Outlierness, Diversity, Representativeness, Uncertainty, Random) on four trajectory datasets (Foxes, Starkey, AIS, Car) spanning animal behavior and traffic domains. Using linear and non-linear ML models with Optuna-tuned augmentation parameters, it reports that Outlierness and Uncertainty yield higher stability and avoid the degradation seen with random selection in dense data, while UMAP visualizations indicate augmentation repairs fragmentation in sparse data but introduces noise in dense data; the value of augmentation is claimed to be strictly conditional on data density and domain velocity, with noted physical limitations in high-velocity regimes.

Significance. If the empirical findings hold after addressing validity concerns, the work supplies actionable guidance on conditional use of trajectory augmentation and moves beyond naive random selection. Credit is due for the multi-domain evaluation, integration of hyperparameter search, and explicit acknowledgment that benefits are not universal. The absence of independent checks on augmented-trajectory validity, however, weakens attribution of performance differences to selection strategy alone.

major comments (2)
  1. [Abstract] Abstract: the central claim that 'the value of augmentation is strictly conditional' on data density and domain velocity rests on the premise that geometric perturbations produce valid training examples; yet the manuscript reports no independent verification (velocity bounds, continuity metrics, or physical-constraint satisfaction) that the perturbed trajectories remain semantically coherent, so observed stability differences cannot be unambiguously attributed to selection rather than invalid data.
  2. [Results] Results (Optuna-tuned runs and UMAP analysis): the statements that Outlierness and Uncertainty 'demonstrated higher stability' and are 'less prone to performance degradation' are load-bearing for the recommendation of systematic strategies, but the provided description supplies neither error bars, dataset statistics, nor statistical tests comparing strategies, leaving the magnitude and reliability of the advantage unquantified.
minor comments (1)
  1. [Abstract] The abstract refers to the work as 'this thesis'; if the manuscript is intended as a journal article, the framing should be adjusted for consistency.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback highlighting two important aspects of our work: the need for explicit validation of augmented trajectory validity and the requirement for statistical quantification of performance differences. We address each major comment below and outline the revisions we will make.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim that 'the value of augmentation is strictly conditional' on data density and domain velocity rests on the premise that geometric perturbations produce valid training examples; yet the manuscript reports no independent verification (velocity bounds, continuity metrics, or physical-constraint satisfaction) that the perturbed trajectories remain semantically coherent, so observed stability differences cannot be unambiguously attributed to selection rather than invalid data.

    Authors: We agree that the manuscript does not report independent verification metrics (e.g., velocity bounds or continuity checks) for the semantic coherence of geometrically perturbed trajectories. Validity is assessed indirectly via downstream task performance and UMAP topology preservation. We will revise the abstract, methods, and discussion to explicitly acknowledge this as a limitation, clarify that the conditional findings rest on observed empirical patterns rather than direct validity proofs, and reference prior geometric perturbation literature for the underlying assumption. This does not change the reported results but strengthens the framing of the claims. revision: yes

  2. Referee: [Results] Results (Optuna-tuned runs and UMAP analysis): the statements that Outlierness and Uncertainty 'demonstrated higher stability' and are 'less prone to performance degradation' are load-bearing for the recommendation of systematic strategies, but the provided description supplies neither error bars, dataset statistics, nor statistical tests comparing strategies, leaving the magnitude and reliability of the advantage unquantified.

    Authors: We acknowledge that the manuscript text does not include error bars, dataset statistics, or formal statistical tests for the stability comparisons. The Optuna-tuned results and UMAP analysis are described qualitatively in the provided sections. We will revise the results section to add error bars to all performance plots, report key dataset statistics (e.g., trajectory counts and density measures), and include statistical comparisons (e.g., Wilcoxon signed-rank tests) between selection strategies to quantify the observed advantages. revision: yes

Circularity Check

0 steps flagged

Empirical evaluation only; no derivation chain present

full rationale

The manuscript is a purely experimental study that evaluates five selection strategies on four real-world trajectory datasets using linear and non-linear models, with Optuna performing external hyperparameter search. No equations, uniqueness theorems, or first-principles derivations are advanced; performance differences are reported directly from cross-validation results. The abstract and described methodology contain no self-definitional steps, fitted inputs renamed as predictions, or load-bearing self-citations that would collapse the central claim to its own inputs. The reader's assessment of score 2.0 is therefore conservative; the correct circularity score is 0.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The central claim rests on the assumption that geometric perturbation produces valid examples and that the four chosen datasets adequately sample the space of trajectory problems. No new entities are postulated.

free parameters (1)
  • augmentation parameters per dataset
    Optuna search identifies best-performing perturbation settings for each dataset; these are fitted values that directly affect reported performance.
axioms (2)
  • domain assumption Geometric perturbation preserves sufficient spatio-temporal coherence to yield useful training examples
    Invoked in the opening sentence as the prior limitation that the selection framework is meant to overcome.
  • domain assumption The four datasets (Foxes, Starkey, AIS, Car) are representative of the domains where trajectory augmentation is applied
    The evaluation scope is defined by these datasets without further justification of coverage.

pith-pipeline@v0.9.1-grok · 5788 in / 1491 out tokens · 19548 ms · 2026-06-27T13:45:31.214348+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

42 extracted references · 17 canonical work pages

  1. [1]

    A survey on trajectory generation methods with reinforcement learning,

    X. Zhao, Z. Li, X. Zhang, S. Guo, and Y . Wu, “A survey on trajectory generation methods with reinforcement learning,”Artificial Intelligence Review,

  2. [2]

    Available: https://doi.org/10.1007/s10462-023-10598-x

    [Online]. Available: https://doi.org/10.1007/s10462-023-10598-x

    work page doi:10.1007/s10462-023-10598-x

  3. [3]

    Road passenger load probability prediction based on trajectory big data,

    W. Liu, X. Lu, M. Duan, and Z. Wang, “Road passenger load probability prediction based on trajectory big data,”Applied Sciences, vol. 14, no. 17, p. 7756, 2024. [Online]. Available: https://www.mdpi.com/2076-3417/14/17/7756

    2024

  4. [4]

    Multi-constellation, multi-frequency gnss collars improve precision and accuracy in wildlife tracking,

    S. Garrido-Carretero, G. Escribano-Avila, M. Peichl, J. Morales-González, and J. M. Fedriani, “Multi-constellation, multi-frequency gnss collars improve precision and accuracy in wildlife tracking,”The Journal of Wildlife Management, vol. 87, no. 3, p. e22378, 2023. [Online]. Available: https://wildlife.onlinelibrary.wiley.com /doi/10.1002/jwmg.22378

    work page doi:10.1002/jwmg.22378 2023

  5. [5]

    A survey on big data for trajectory analytics,

    D. Ribeiro de Almeida, C. de Souza Baptista, F. Gomes de Andrade, and A. Soares, “A survey on big data for trajectory analytics,”ISPRS International Journal of Geo- Information, vol. 9, no. 2, p. 88, 2020

    2020

  6. [6]

    A trajectory scoring tool for local anomaly detection in maritime traffic using visual analytics,

    F. H. Abreu, A. Soares, F. V . Paulovich, and S. Matwin, “A trajectory scoring tool for local anomaly detection in maritime traffic using visual analytics,”ISPRS Inter- national Journal of Geo-Information, vol. 10, no. 6, p. 412, 2021

    2021

  7. [7]

    Enhancing global mar- itime traffic network forecasting with gravity-inspired deep learning models,

    R. Song, G. Spadon, R. Pelot, S. Matwin, and A. Soares, “Enhancing global mar- itime traffic network forecasting with gravity-inspired deep learning models,”Sci- entific reports, vol. 14, no. 1, p. 16665, 2024

    2024

  8. [8]

    Un- derstanding evolution of maritime networks from automatic identification system data,

    E. Carlini, V . M. de Lira, A. Soares, M. Etemad, B. Brandoli, and S. Matwin, “Un- derstanding evolution of maritime networks from automatic identification system data,”GeoInformatica, pp. 1–25, 2022

    2022

  9. [9]

    Crisis: Integrating ais and ocean data streams using semantic web standards for event detection,

    A. Soares, R. Dividino, F. Abreu, M. Brousseau, A. W. Isenor, S. Webb, and S. Matwin, “Crisis: Integrating ais and ocean data streams using semantic web standards for event detection,” in2019 International conference on military com- munications and information systems (ICMCIS). IEEE, 2019, pp. 1–7

    2019

  10. [10]

    Multi-path long-term vessel trajectories forecasting with probabilistic feature fusion for problem shifting,

    G. Spadon, J. Kumar, D. Eden, J. van Berkel, T. Foster, A. Soares, R. Fablet, S. Matwin, and R. Pelot, “Multi-path long-term vessel trajectories forecasting with probabilistic feature fusion for problem shifting,”Ocean Engineering, vol. 312, p. 119138, 2024

    2024

  11. [11]

    A semi-supervised method- ology for fishing activity detection using the geometry behind the trajectory of mul- tiple vessels,

    M. D. Ferreira, G. Spadon, A. Soares, and S. Matwin, “A semi-supervised method- ology for fishing activity detection using the geometry behind the trajectory of mul- tiple vessels,”Sensors, vol. 22, no. 16, p. 6063, 2022

    2022

  12. [12]

    Predicting fishing effort and catch using semantic trajectories and ma- chine learning,

    P. Adibi, F. Pranovi, A. Raffaetà, E. Russo, C. Silvestri, M. Simeoni, A. Soares, and S. Matwin, “Predicting fishing effort and catch using semantic trajectories and ma- chine learning,” inInternational Workshop on Multiple-Aspect Analysis of Semantic Trajectories. Springer International Publishing Cham, 2019, pp. 83–99. 36

    2019

  13. [13]

    A dash- board tool for mobility data mining preprocessing tasks,

    Y . J. Haranwala, S. Haidri, T. S. Tricco, V . P. da Fonseca, and A. Soares, “A dash- board tool for mobility data mining preprocessing tasks,” in2022 23rd IEEE In- ternational Conference on Mobile Data Management (MDM). IEEE, 2022, pp. 278–281

    2022

  14. [14]

    PTRAIL — A python package for parallel trajectory data preprocessing,

    S. Haidri, Y . J. Haranwala, V . Bogorny, C. Renso, V . P. da Fonseca, and A. Soares, “PTRAIL — A python package for parallel trajectory data preprocessing,”Soft- wareX, vol. 19, p. 101176, 2022

    2022

  15. [15]

    A data augmentation al- gorithm for trajectory data,

    Y . J. Haranwala, G. Spadon, C. Renso, and A. Soares, “A data augmentation al- gorithm for trajectory data,” in1st ACM SIGSPATIAL International Workshop on Methods for Enriched Mobility Data: Emerging issues and Ethical perspectives 2023 (EMODE ’23). Hamburg, Germany: ACM, November 2023, pp. 1–5, [On- line]. Available at: doi:10.1145/3615885.3628008

    work page doi:10.1145/3615885.3628008 2023

  16. [16]

    Assessing com- pression algorithms to improve the efficiency of clustering analysis on ais vessel trajectories,

    M. D. Ferreira, J. Campbell, E. Purney, A. Soares, and S. Matwin, “Assessing com- pression algorithms to improve the efficiency of clustering analysis on ais vessel trajectories,”International Journal of Geographical Information Science, vol. 37, no. 3, pp. 660–683, 2023

    2023

  17. [17]

    Deep learning for trajectory classification: A survey,

    H. Lee, S. Kim, and J. Park, “Deep learning for trajectory classification: A survey,”Pattern Recognition, vol. 135, p. 109084, 2023. [Online]. Available: https://doi.org/10.1016/j.patcog.2022.109084

    work page doi:10.1016/j.patcog.2022.109084 2023

  18. [18]

    Analytic: An active learning system for trajectory classification,

    A. S. Júnior, C. Renso, and S. Matwin, “Analytic: An active learning system for trajectory classification,”IEEE computer graphics and applications, vol. 37, no. 5, pp. 28–39, 2017

    2017

  19. [19]

    A study on the geometric and kine- matic descriptors of trajectories in the classification of ship types,

    Y . Tavakoli, L. Peña-Castillo, and A. Soares, “A study on the geometric and kine- matic descriptors of trajectories in the classification of ship types,”Sensors, vol. 22, no. 15, p. 5588, 2022

    2022

  20. [20]

    Trajectory clustering analysis,

    Y . Wang and Y . Y . Tang, “Trajectory clustering analysis,” inMachine Learning for Data Science Handbook, L. Rokach, O. Maimon, and E. Shmueli, Eds. Springer, 2023, pp. 171–189. [Online]. Available: https://doi.org/10.1007/978-3-031-24628 -9_10

    work page doi:10.1007/978-3-031-24628 2023

  21. [21]

    A three-dimensional hail trajectory clustering technique,

    M. T. Johnson and E. L. Smith, “A three-dimensional hail trajectory clustering technique,”Monthly Weather Review, vol. 151, no. 9, pp. 3001–3020, 2023. [Online]. Available: https://doi.org/10.1175/MWR-D-22-0345.1

    work page doi:10.1175/mwr-d-22-0345.1 2023

  22. [22]

    Trajectory clustering-based anomaly detection in indoor human movement,

    D. T. Lan and S. Yoon, “Trajectory clustering-based anomaly detection in indoor human movement,”Sensors, vol. 23, no. 6, p. 3318, 2023. [Online]. Available: https://doi.org/10.3390/s23063318

    work page doi:10.3390/s23063318 2023

  23. [23]

    Motion segmentation of pedestrian trajectories using angular gaussian mixture model,

    W. Tan, M. Liu, and L. Zhao, “Motion segmentation of pedestrian trajectories using angular gaussian mixture model,” inProceedings of the 2023 5th World Symposium on Software Engineering (WSSE ’23). ACM, 2023, pp. 102–110. [Online]. Available: https://doi.org/10.1145/3631991.3632040

    work page doi:10.1145/3631991.3632040 2023

  24. [24]

    Sws: an unsupervised trajectory segmentation algorithm based on change detection with in- terpolation kernels,

    M. Etemad, A. Soares, E. Etemad, J. Rose, L. Torgo, and S. Matwin, “Sws: an unsupervised trajectory segmentation algorithm based on change detection with in- terpolation kernels,”GeoInformatica, vol. 25, no. 2, pp. 269–289, 2021. 37

    2021

  25. [25]

    A semi- supervised approach for the semantic segmentation of trajectories,

    A. S. Junior, V . C. Times, C. Renso, S. Matwin, and L. A. Cabral, “A semi- supervised approach for the semantic segmentation of trajectories,” in2018 19th IEEE international conference on mobile data management (MDM). IEEE, 2018, pp. 145–154

    2018

  26. [26]

    Grasp-uts: an algorithm for unsupervised trajectory segmentation,

    A. Soares Júnior, B. N. Moreno, V . C. Times, S. Matwin, and L. d. A. F. Cabral, “Grasp-uts: an algorithm for unsupervised trajectory segmentation,”International Journal of Geographical Information Science, vol. 29, no. 1, pp. 46–68, 2015

    2015

  27. [27]

    Unfolding ais transmission behavior for vessel movement modeling on noisy data leveraging machine learning,

    G. Spadon, M. D. Ferreira, A. Soares, and S. Matwin, “Unfolding ais transmission behavior for vessel movement modeling on noisy data leveraging machine learning,” IEEE Access, vol. 11, pp. 18 821–18 837, 2022

    2022

  28. [28]

    Challenges in vessel behavior and anomaly detection: From classical machine learning to deep learning,

    L. May Petry, A. Soares, V . Bogorny, B. Brandoli, and S. Matwin, “Challenges in vessel behavior and anomaly detection: From classical machine learning to deep learning,” inAdvances in Artificial Intelligence: 33rd Canadian Conference on Ar- tificial Intelligence, Canadian AI 2020, Ottawa, ON, Canada, May 13–15, 2020, Proceedings 33. Springer, 2020, pp. 401–407

    2020

  29. [29]

    Uncovering vessel movement patterns from ais data with graph evolution analysis,

    E. Carlini, V . M. de Lira, A. Soares, M. Etemad, B. B. Machado, and S. Matwin, “Uncovering vessel movement patterns from ais data with graph evolution analysis,” inEDBT/ICDT Workshops, 2020

    2020

  30. [30]

    Trajectory Data Augmentation Framework,

    A. Nordling, “Trajectory Data Augmentation Framework,” 2026. [Online]. Available: https://doi.org/10.5281/zenodo.19323839

    work page doi:10.5281/zenodo.19323839 2026

  31. [31]

    An empirical survey of data augmentation for limited data learning in nlp,

    J. Chen, D. Tam, C. Raffel, M. Bansal, and D. Yang, “An empirical survey of data augmentation for limited data learning in nlp,”Transactions of the Association for Computational Linguistics, vol. 11, pp. 191–211, 2023. [Online]. Available: https://aclanthology.org/2023.tacl-1.12/

    2023

  32. [32]

    A survey of mix-based data augmentation: Taxonomy, methods, applications, and explainability,

    C. Cao, F. Zhou, Y . Dai, J. Wang, and K. Zhang, “A survey of mix-based data augmentation: Taxonomy, methods, applications, and explainability,”ACM Computing Surveys, vol. 57, no. 2, pp. 1–38, 2024. [Online]. Available: https://dl.acm.org/doi/10.1145/3696206

    work page doi:10.1145/3696206 2024

  33. [33]

    A comprehensive survey of image augmentation techniques for deep learning,

    M. Xu, S. Yoon, A. Fuentes, and D. S. Park, “A comprehensive survey of image augmentation techniques for deep learning,”Pattern Recognition, vol. 137, p. 109347, 2023. [Online]. Available: https://doi.org/10.1016/j.patcog.2023.109347

    work page doi:10.1016/j.patcog.2023.109347 2023

  34. [34]

    AugmenTRAJ: A framework for point-based trajectory data augmentation,

    Y . J. Haranwala, “AugmenTRAJ: A framework for point-based trajectory data augmentation,” Master’s thesis, Memorial University of Newfoundland and Labrador, Department of Computer Science, 2023, [Online]. Available at: doi:10.48550/arXiv.2311.15097. Accessed: 2024-03-15

    work page doi:10.48550/arxiv.2311.15097 2023

  35. [35]

    Trajec- tory augmentation for robust neural locomotion controllers,

    D. Agrawal, M. König, J. Buhmann, R. Sumner, and M. Guay, “Trajec- tory augmentation for robust neural locomotion controllers,” inMIG ’23: Mo- tion, Interaction and Games. ACM, 2023, pp. 1–11, [Online]. Available at: doi:10.1145/3623267.3623274

    work page doi:10.1145/3623267.3623274 2023

  36. [36]

    ISBN 978-1- 72819-077-8

    D. Antotsiou, C. Ciliberto, and T.-K. Kim, “Adversarial imitation learning with tra- jectorial augmentation and correction,” in2021 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2021, pp. 3342–3348, [Online]. Available at: doi:10.1109/ICRA48506.2021.9561915. 38

    work page doi:10.1109/icra48506.2021.9561915 2021

  37. [37]

    Augmenting safety-critical driving scenarios while preserving similarity to expert trajectories,

    H. Mirkhani, B. Khamidehi, and K. Rezaee, “Augmenting safety-critical driving scenarios while preserving similarity to expert trajectories,” in2024 IEEE Intel- ligent Vehicles Symposium (IV). IEEE, 2024, pp. 1–6, [Online]. Available at: doi:10.1109/IV55156.2024.10588830

    work page doi:10.1109/iv55156.2024.10588830 2024

  38. [38]

    SelectAugment: Hierarchical deterministic sample selection for data augmentation,

    S. Lin, Z. Zhang, X. Li, and Z. Chen, “SelectAugment: Hierarchical deterministic sample selection for data augmentation,” inThe Thirty-Seventh AAAI Conference on Artificial Intelligence (AAAI-23). AAAI Press, 2023, pp. 1892–1900, [Online]. Available at: doi:10.1609/aaai.v37i2.25247

    work page doi:10.1609/aaai.v37i2.25247 2023

  39. [39]

    A survey and comparison of trajectory classification methods,

    C. L. da Silva, L. M. Petry, and V . Bogorny, “A survey and comparison of trajectory classification methods,” in2019 8th Brazilian Conference on Intelli- gent Systems (BRACIS). IEEE, 2019, pp. 479–484, [Online]. Available at: doi:10.1109/BRACIS.2019.00089

    work page doi:10.1109/bracis.2019.00089 2019

  40. [40]

    A unified approach for mining outliers,

    E. M. Knorr and R. T. Ng, “A unified approach for mining outliers,” inProceed- ings of the 1997 conference of the Centre for Advanced Studies on Collaborative research, 1997, p. 11

    1997

  41. [41]

    Movement tactics of a mobile predator in a meta- ecosystem with fluctuating resources: the arctic fox in the high arctic,

    S. Lai, J. Bêty, and D. Berteaux, “Movement tactics of a mobile predator in a meta- ecosystem with fluctuating resources: the arctic fox in the high arctic,”Oikos, vol. 126, no. 7, pp. 937–947, 2017

    2017

  42. [42]

    M. J. Wisdom,The Starkey Project: a synthesis of long-term studies of elk and mule deer. Lawrence, Kansas: Alliance Communications Group, 2005, rep. PNW-GTR- 634. 39 Appendix 1: Source Code and Reproducibility To ensure full transparency and reproducibility of the results discussed in this thesis, the complete Python framework has been open-sourced. The r...

    2005