Pith. sign in

REVIEW 3 major objections 5 minor 55 references

CausalTAD: Causal Implicit Generative Model for Debiased Online Trajectory Anomaly Detection

T0 review · 3 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read Trajectory anomaly detection should score routes by the interventional probability $P(T|do(C))$, not the conditional $P(T|C)$, because a hidden road-preference confounder biases the latter on unseen source-destination pairs.

desk verdict Novel causal framing with strong OOD gains, but the derivation of P(T|do(C)) relies on an unjustified approximation and a tuned lambda; treat it as a reweighted-likelihood heuristic. read the letter →

arxiv 2412.18820 v1 pith:KPXAC7SY submitted 2024-12-25 cs.LG

Wenbin Li , Di Yao , Chang Gong , Xiaokai Chu , Quanliang Jing , Xiaolei Zhou , Yuxuan Zhang , Yunxia Fan
show 1 more author
Jingping Bi
This is my paper · ORCID
classification cs.LG
keywords trajectoryanomalydetectioncausalinferencedo-calculusbackdooradjustmentout-of-distributiongeneralizationvariationalautoencoderroadnetworkonline
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper argues that existing trajectory anomaly detectors, which score a trip by its conditional likelihood given the source-destination pair, are biased by a hidden confounder: road preference, which affects both which trips people take and which roads drivers choose. To fix this, CausalTAD replaces the conditional score with an interventional one, $P(T|do(C))$, derived through do-calculus and backdoor adjustment, and approximates it as a likelihood term times a per-segment reweighting factor. On two public ride-hailing trajectory datasets, the authors report that this reweighting improves ROC-AUC/PR-AUC by 2.1% to 5.7% on observed SD pairs and by 10.6% to 32.7% on unseen SD pairs. The central claim is that removing the spurious correlation, rather than fitting $P(T|C)$ better, is what lets a detector generalize to routes never seen in training.

What carries the argument

The load-bearing object is the backdoor-adjusted score of Eq. (10), which factorizes the debiased anomaly score into a likelihood term, $\log P(c,t)$, and a scaling term, $\lambda \sum_i \log \mathbb{E}_{e_i \sim P(E_i|t_i)} 1/P(t_i|e_i)$. The scaling term is the mechanism that removes confounding: it reweights each road segment's anomaly contribution by the inverse reconstruction probability under the Road Preference VAE, thereby compensating for the model's overestimation of popular roads and underestimation of unpopular ones. The Trajectory Generation VAE supplies the likelihood with a road-constrained decoder that predicts only neighboring road segments, preventing popular SD pairs from dominating the road-network representation. Together they turn a purely observational generative model into an approximate interventional one that can be updated in $O(1)$ per new road segment during online detection.

What would settle it

Simulate a synthetic road network with a known confounder $E$ (e.g., a popular district) that shifts the SD distribution, generate trips under a known causal model, and compare CausalTAD's score against the analytically computed $P(T|do(C))$; if the factorized score deviates systematically as the confounder's influence grows, the debiasing claim fails. Alternatively, run the OOD evaluation without tuning $\lambda$ on the test distribution; if performance collapses unless $\lambda$ is calibrated per dataset, the method's advantage depends on tuning rather than identification.

Watch

Extended reading notes

Core claim

CausalTAD estimates $P(T=t|do(C=c))$ as the anomaly criterion, obtained by backdoor adjustment over the hidden road-preference confounder $E$. The adjustment factor is decomposed into a per-road-segment expectation $\mathbb{E}_{e_i \sim P(E_i|t_i)} \, 1/P(t_i|e_i)$, computed by a road-preference VAE, while a trajectory-generation VAE supplies the likelihood $P(c,t)$. The claim is that this interventional score, not the observational likelihood $P(t|c)$, is the right risk measure when the distribution of source-destination pairs shifts, because it cuts off the spurious path $C \leftarrow E \rightarrow T$ while preserving the causal path $C \rightarrow T$.

Load-bearing premise

The whole debiasing argument rests on approximating the hidden confounder's effect with per-segment independent factors: $P(c|e)$ is replaced by $P(t|e)$ for the single observed trajectory, and both $P(e|t)$ and $P(t|e)$ are factorized across road segments; if those approximations fail, the computed score is not the causal effect but an ad hoc reweighting.

Editorial extensions

If this is right

  • On in-distribution trips, CausalTAD improves detection by 2.1% to 5.7%, indicating that segment-level road-preference bias also exists within observed SD pairs, not only across them.
  • On never-seen SD pairs, the method improves performance by 10.6% to 32.7% in ROC-AUC/PR-AUC, suggesting the interventional score generalizes substantially better than conditional likelihood.
  • Because the per-segment scaling factors can be precomputed and stored, the anomaly score updates in $O(1)$ time per new road segment, making online detection feasible.
  • Ablation studies show both modules are necessary: the Trajectory Generation VAE alone loses most of the OOD gain, and the Road Preference VAE alone performs poorly, confirming that the scaling factor is what restores generalization.
  • The tuning constant $\lambda$ compensates for an overestimated scaling factor caused by dropped terms in the approximation; the authors find $\lambda=0.1$ works best across their test settings.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the causal story is right, the same backdoor-adjustment idea could transfer to other sequential decision tasks where a hidden preference confounds both the choice set and the outcome, such as route recommendation or driver-behavior modeling.
  • The method's identification relies on discretizing trajectories into road segments; applying it to grid-based or free-space trajectories would require a different factorization of the scaling factor, which is a testable extension.
  • The dependence on $\lambda$ suggests the estimator is not exactly the interventional distribution; a tighter approximation of $P(c|e)$ that sums over alternative trajectories (Eq. 5) could remove the need for heuristic tuning.
  • A direct falsification would be to simulate a synthetic road network with a known confounder, compute the true $P(T|do(C))$ analytically, and check whether CausalTAD's factorized score tracks it as the confounder's influence grows.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes CausalTAD, a variational-autoencoder-based method for online trajectory anomaly detection, and claims to estimate the interventional quantity P(T|do(C)) rather than the observational conditional P(T|C), in order to remove confounding bias caused by a hidden road-preference variable E. The method decomposes the anomaly score into a likelihood term estimated by a Trajectory Generation VAE and a scaling factor estimated by a per-road-segment Road Preference VAE, combined through a tuned weight λ. Experiments on two DiDi trajectory datasets report consistent improvements over baselines, particularly on out-of-distribution SD pairs, together with an O(1) online update property.

Significance. If the causal claim were established, the paper would make a valuable contribution: the OOD generalization problem for trajectory anomaly detection is real and practically important, and a debiased criterion that works for unseen SD pairs would be significant. The paper has several genuine strengths: the code is released, the online efficiency argument is clear, the problem motivation via the confounding example in Fig. 1(b) is intuitive, and the reported empirical gains on OOD data are large and consistent across two datasets. However, the central identification argument is not sound: the final score is not P(T|do(C)) but a reweighted log-likelihood heuristic with a tuned scalar. The empirical results are interesting evidence for the heuristic's value, but they do not establish elimination of confounding bias or estimation of an interventional distribution.

major comments (3)
  1. [§V-C, Eqs. (5)-(6)] The replacement of P(c|e) by P(t|e) is not justified. P(c|e) is by definition the sum of P(t'|e) over all trajectories sharing the SD pair c, and the fact that e is drawn from P(E|c,t) does not imply that P(t|e) dominates that sum pointwise for the e values with non-negligible posterior mass. If several plausible routes exist for the same SD pair, the sum can be several times P(t|e), and the error is exactly the confounding effect the method claims to remove. No bound, asymptotic argument, or empirical check is provided for this step, so the quantity after Eq. (6) is not the backdoor-adjusted causal effect.
  2. [§V-C, Eq. (7)] The mean-field factorization P(t|e)=∏P(ti|ei) and P(e|t)=∏P(ei|ti) is an additional modeling assumption that does not follow from the causal graph in Fig. 1(a). In that graph, E is a single network-level confounder, and road preferences are plausibly correlated across segments through congestion, connectivity, and area attractiveness. Replacing E by independent per-segment latents ei learned from the same observational trajectories changes the estimand and makes the adjustment circular: the ei are reconstruction latents of a VAE fitted on P(ti), not observations of the confounder. The paper does not show that the VAE posterior approximates the true P(Ei|ti) or that the factorization error is small.
  3. [§V-D, Eq. (10) and §VI-H] The free parameter λ is not part of the backdoor adjustment derivation and is introduced specifically to compensate for the acknowledged overestimation of the scaling factor. With λ=0.1 the score is approximately -log P(c,t) - 0.1 Σ_i log E_{ei~P(Ei|ti)} 1/P(ti|ei), which is a particular reweighted log-likelihood heuristic rather than the interventional quantity P(T|do(C)). The large OOD margins in Table II demonstrate empirical value, but they do not establish that confounding bias has been eliminated. The paper should either validate the causal estimate on synthetic data with a known confounder or substantially reframe the contribution as a debiasing-inspired heuristic.
minor comments (5)
  1. [§VI-F] In the training scalability paragraph, the name 'GM-VSAE' is duplicated in the list of compared methods.
  2. [Figure 4 and §VI-C] The text says the scaling factor part of the anomaly scores was 'centralized', but no definition of this centering operation is given.
  3. [§VI-A] The sentence 'we filter out trajectories shorter than 30' should specify the unit (number of road segments, trip duration, or distance).
  4. [§V-C, Eq. (5)] The equality P(c|t',e)=1 for t'∈T(c) presumes a deterministic map from trajectory to SD pair; this assumption should be stated explicitly.
  5. [Tables I and II] The paper reports point estimates only; confidence intervals or significance tests would strengthen the claim that the OOD improvements are reliable.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the backdoor-adjusted criterion is externally grounded, and the approximate reweighted score with tuned λ is an approximation/identification limitation, not a circular reduction.

full rationale

CausalTAD's derivation begins with a standard backdoor adjustment identity (Eqs. 1–2) whose assumptions are stated independently of the model outputs. The likelihood P(c,t) and the scaling factor are then estimated by TG-VAE and RP-VAE trained on observational trajectories; the replacements in Eqs. (5)–(7) — P(c|e) → P(t|e) and the mean-field factorization — are explicit approximations, and the paper openly concedes in Section VI-H that the scaling factor is overestimated and requires a tuned λ. These are substantive correctness and identification risks: the final score in Eq. (10) is not literally P(T|do(C)), but rather a reweighted observational likelihood. However, they are not cases where an input is redefined as the output by construction, where a fitted parameter is renamed a prediction, or where a load-bearing premise rests on a self-citation. No uniqueness theorem or ansatz is imported from the authors' prior work. The empirical gains in Tables I–II are benchmarked against external baselines and therefore provide independent evidence for the method's practical value, even if they cannot validate the causal label. No circular step is exhibited; score 0.

Assumptions & free parameters 1 free parameters · 5 assumptions · 2 invented entities

The central claim depends on one free parameter lambda and on several assumptions about the causal graph, the factorization of the hidden confounder, and the approximation in Eq. (6). The hidden confounder E and its per-segment versions ei are invented entities without independent falsifiable handles.

free parameters (1)
  • lambda = 0.1
    Balances the trajectory likelihood and the scaling factor in Eq. (10). Chosen by grid search on validation (Section VI-H). The paper states it compensates for the overestimated scaling factor, so it is a tuning parameter, not a derived constant.
assumptions (5)
  • domain assumption The causal graph E -> C, E -> T, C -> T correctly describes trajectory generation, with E an unobserved road-preference confounder.
    Introduced in Section IV and Fig. 2(a); no observational or experimental evidence is provided for the hidden confounder E.
  • standard math Backdoor adjustment over E is valid: E satisfies the backdoor criterion and positivity holds.
    Invoked in Eq. (1) following Pearl; the mathematical formula is standard, but its applicability requires the causal graph and positivity.
  • ad hoc to paper The hidden confounder factorizes over road segments: P(e|t)=prod_i P(ei|ti) and P(t|e)=prod_i P(ti|ei).
    Stated as a mean-field approximation in Section V-C before Eq. (7); no support is given for the independence of road preferences.
  • ad hoc to paper P(c|e) can be approximated by P(t|e), i.e., all trajectories t' != t with the same SD pair can be ignored.
    Equation (6) in Section V-C; the paper justifies this by the sampling distribution, but it is not a mathematical consequence.
  • domain assumption Anomaly ground truth is generated via the detour and switch strategies, and these synthetic anomalies match real anomalies.
    Section VI-A2; no real labels are available, so evaluation depends on this assumption.
invented entities (2)
  • E (road network preference)
    purpose: Hidden confounder that explains why SD distribution and trajectories share correlation; target of do-calculus adjustment.
    Introduced in Section IV as a mixture of road factors; never measured or validated, so the claimed debiasing cannot be checked against a ground-truth causal effect.
  • ei (per-road-segment road preference latent)
    purpose: Enables the mean-field factorization of E and the per-segment scaling factor in RP-VAE.
    Added in Section V-C; it is a VAE latent with no causal interpretation beyond the factorization assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CausalTAD: Causal Implicit Generative Model for Debiased Online Trajectory Anomaly Detection." pith.science (2026). https://pith.science/paper/KPXAC7SY

@misc{pith2026241218820,
  author       = {Pith},
  title        = {Pith review of: CausalTAD: Causal Implicit Generative Model for Debiased Online Trajectory Anomaly Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KPXAC7SY}},
  note         = {Machine review of arXiv:2412.18820}
}
abstract

Trajectory anomaly detection, aiming to estimate the anomaly risk of trajectories given the Source-Destination (SD) pairs, has become a critical problem for many real-world applications. Existing solutions directly train a generative model for observed trajectories and calculate the conditional generative probability $P({T}|{C})$ as the anomaly risk, where ${T}$ and ${C}$ represent the trajectory and SD pair respectively. However, we argue that the observed trajectories are confounded by road network preference which is a common cause of both SD distribution and trajectories. Existing methods ignore this issue limiting their generalization ability on out-of-distribution trajectories. In this paper, we define the debiased trajectory anomaly detection problem and propose a causal implicit generative model, namely CausalTAD, to solve it. CausalTAD adopts do-calculus to eliminate the confounding bias of road network preference and estimates $P({T}|do({C}))$ as the anomaly criterion. Extensive experiments show that CausalTAD can not only achieve superior performance on trained trajectories but also generally improve the performance of out-of-distribution data, with improvements of $2.1\% \sim 5.7\%$ and $10.6\% \sim 32.7\%$ respectively.

Figures

Figures reproduced from arXiv: 2412.18820 by the authors.

Figure 1
Figure 1. (a) The causal graph of trajectory generation. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. (a) An example of the causal graph. (b) The causal graph after an [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Overview of the CausalTAD. Upper-left: TG-VAE models the patterns [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: The anomaly scores of a normal trajectory with an unseen SD pair [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Performance under different ratios of distribution shift. (a) The ROC [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 7
Figure 7. Figure 7: (a) Comparison of training scalability. (b) The average runtime per [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 8
Figure 8. Figure 8: Performance of CausalTAD under different values of [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

55 extracted references · 50 canonical work pages

  1. [1]

    Trajectory data mining: An overview,

    Y . Zheng, “Trajectory data mining: An overview,” ACM Trans. Intell. Syst. Technol., vol. 6, no. 3, pp. 29:1–29:41, 2015

    work page 2015

  2. [2]

    Spatio-temporal trajectory similarity measures: A comprehensive survey and quantitative study,

    D. Hu, L. Chen, H. Fang, Z. Fang, T. Li, and Y . Gao, “Spatio-temporal trajectory similarity measures: A comprehensive survey and quantitative study,” CoRR, vol. abs/2303.05012, 2023

    arXiv 2023

  3. [3]

    Modeling spatial trajectories with attribute representation learning,

    M. Chen, Y . Zhao, Y . Liu, X. Yu, and K. Zheng, “Modeling spatial trajectories with attribute representation learning,” IEEE Trans. Knowl. Data Eng., pp. 1902–1914, 2022

    work page 1902

  4. [4]

    STL: online detection of taxi trajectory anomaly based on spatial- temporal laws,

    B. Cheng, S. Qian, J. Cao, G. Xue, J. Yu, Y . Zhu, M. Li, and T. Zhang, “STL: online detection of taxi trajectory anomaly based on spatial- temporal laws,” in Database Systems for Advanced Applications - 24th International Conference, DASFAA 2019, Chiang Mai, Thailand, April 22-25, 2019, Proceedings, Part II , ser. Lecture Notes in Computer Science, vol. ...

    work page 2019

  5. [5]

    ibat: detect- ing anomalous taxi trajectories from GPS traces,

    D. Zhang, N. Li, Z. Zhou, C. Chen, L. Sun, and S. Li, “ibat: detect- ing anomalous taxi trajectories from GPS traces,” in UbiComp 2011: Ubiquitous Computing, 13th International Conference, UbiComp 2011, Beijing, China, September 17-21, 2011, Proceedings , 2011, pp. 99–108

    work page 2011

  6. [6]

    Coupled IGMM-GANs for deep multimodal anomaly detection in human mobility data

    K. Gray, D. Smolyak, S. Badirli, and G. O. Mohler, “Coupled igmm- gans for deep multimodal anomaly detection in human mobility data,” CoRR, vol. abs/1809.02728, 2018

    work page Pith review arXiv 2018

  7. [7]

    Some practical guidance for the imple- mentation of propensity score matching,

    M. Caliendo and S. Kopeinig, “Some practical guidance for the imple- mentation of propensity score matching,” Journal of economic surveys , vol. 22, no. 1, pp. 31–72, 2008

    work page 2008

  8. [8]

    Inverse probability weighting,

    M. A. Mansournia and D. G. Altman, “Inverse probability weighting,” Bmj, vol. 352, 2016

    work page 2016

Show all 55 references
  1. [9]

    Evolutionary clustering of moving objects,

    T. Li, L. Chen, C. S. Jensen, T. B. Pedersen, Y . Gao, and J. Hu, “Evolutionary clustering of moving objects,” in 38th IEEE International Conference on Data Engineering, ICDE 2022, Kuala Lumpur, Malaysia, May 9-12, 2022 , 2022, pp. 2399–2411

    2022

  2. [10]

    DBSCAN- MS: distributed density-based clustering in metric spaces,

    K. Yang, Y . Gao, R. Ma, L. Chen, S. Wu, and G. Chen, “DBSCAN- MS: distributed density-based clustering in metric spaces,” in 35th IEEE International Conference on Data Engineering, ICDE 2019, Macao, China, April 8-11, 2019 , 2019, pp. 1346–1357

    2019

  3. [11]

    Trajectory clustering via deep representation learning,

    D. Yao, C. Zhang, Z. Zhu, J. Huang, and J. Bi, “Trajectory clustering via deep representation learning,” in IJCNN, 2017, pp. 3880–3887

    2017

  4. [12]

    Computing trajectory similarity in linear time: A generic seed-guided neural metric learning approach,

    D. Yao, G. Cong, C. Zhang, and J. Bi, “Computing trajectory similarity in linear time: A generic seed-guided neural metric learning approach,” in 35th IEEE International Conference on Data Engineering, ICDE 2019, Macao, China, April 8-11, 2019 , 2019, pp. 1358–1369

    2019

  5. [13]

    Deep representation learning for trajectory similarity computation,

    X. Li, K. Zhao, G. Cong, C. S. Jensen, and W. Wei, “Deep representation learning for trajectory similarity computation,” in 34th IEEE Interna- tional Conference on Data Engineering, ICDE 2018, Paris, France, April 16-19, 2018, 2018, pp. 617–628

    2018

  6. [14]

    Efficient trajec- tory similarity computation with contrastive learning,

    L. Deng, Y . Zhao, Z. Fu, H. Sun, S. Liu, and K. Zheng, “Efficient trajec- tory similarity computation with contrastive learning,” in Proceedings of the 31st ACM International Conference on Information & Knowledge Management, Atlanta, GA, USA, October 17-21, 2022 , 2022, pp. 365– 374

    2022

  7. [15]

    Trajgat: A graph- based long-term dependency modeling approach for trajectory similarity computation,

    D. Yao, H. Hu, L. Du, G. Cong, S. Han, and J. Bi, “Trajgat: A graph- based long-term dependency modeling approach for trajectory similarity computation,” in KDD ’22: The 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Washington, DC, USA, August 14 - 18, 202...

    2022

  8. [16]

    Effective and efficient trajectory outlier detection based on time-dependent popular route,

    J. Zhu, W. Jiang, A. Liu, G. Liu, and L. Zhao, “Effective and efficient trajectory outlier detection based on time-dependent popular route,” World Wide Web, vol. 20, no. 1, pp. 111–134, 2017

    2017

  9. [17]

    Few-shot learning for trajectory outlier de- tection with only normal trajectories,

    Y . Su, D. Yao, and J. Bi, “Few-shot learning for trajectory outlier de- tection with only normal trajectories,” in International Joint Conference on Neural Networks, IJCNN 2023, Gold Coast, Australia, June 18-23, 2023, 2023, pp. 1–8

    2023

  10. [18]

    Sub-trajectory- and trajectory-neighbor-based outlier detection over trajectory streams,

    Z. Zhu, D. Yao, J. Huang, H. Li, and J. Bi, “Sub-trajectory- and trajectory-neighbor-based outlier detection over trajectory streams,” in PAKDD, 2018, pp. 551–563

    2018

  11. [19]

    Online learning and sequential anomaly detection in trajectories,

    R. Laxhammar and G. Falkman, “Online learning and sequential anomaly detection in trajectories,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 36, no. 6, pp. 1158–1173, 2014

    2014

  12. [20]

    Continuous trajectory similarity search for online outlier detection,

    D. Zhang, Z. Chang, S. Wu, Y . Yuan, K. Tan, and G. Chen, “Continuous trajectory similarity search for online outlier detection,” IEEE Trans. Knowl. Data Eng. , vol. 34, no. 10, pp. 4690–4704, 2022

    2022

  13. [21]

    A fast trajectory outlier detection approach via driving behavior modeling,

    H. Wu, W. Sun, and B. Zheng, “A fast trajectory outlier detection approach via driving behavior modeling,” in Proceedings of the 2017 ACM on Conference on Information and Knowledge Management, CIKM 2017, Singapore, November 06 - 10, 2017 , 2017, pp. 837–846

    2017

  14. [22]

    Online anomalous trajectory detection with deep generative sequence modeling,

    Y . Liu, K. Zhao, G. Cong, and Z. Bao, “Online anomalous trajectory detection with deep generative sequence modeling,” in 36th IEEE International Conference on Data Engineering, ICDE 2020, Dallas, TX, USA, April 20-24, 2020 , 2020, pp. 949–960

    2020

  15. [23]

    Auto-encoding variational bayes,

    D. P. Kingma and M. Welling, “Auto-encoding variational bayes,” in 2nd International Conference on Learning Representations, ICLR 2014, Banff, AB, Canada, April 14-16, 2014, Conference Track Proceedings , 2014

    2014

  16. [24]

    Deeptea: Effective and efficient online time-dependent trajectory outlier detection,

    X. Han, R. Cheng, C. Ma, and T. Grubenmann, “Deeptea: Effective and efficient online time-dependent trajectory outlier detection,” Proc. VLDB Endow., vol. 15, no. 7, pp. 1493–1505, 2022

    2022

  17. [25]

    Trajectory outlier detection: A partition-and- detect framework,

    J. Lee, J. Han, and X. Li, “Trajectory outlier detection: A partition-and- detect framework,” in Proceedings of the 24th International Conference on Data Engineering, ICDE 2008, April 7-12, 2008, Canc ´un, Mexico, 2008, pp. 140–149

    2008

  18. [26]

    iboat: Isolation-based online anomalous trajectory detection,

    C. Chen, D. Zhang, P. S. Castro, N. Li, L. Sun, S. Li, and Z. Wang, “iboat: Isolation-based online anomalous trajectory detection,” IEEE Trans. Intell. Transp. Syst. , vol. 14, no. 2, pp. 806–818, 2013

    2013

  19. [27]

    Online anomalous subtrajectory detection on road networks with deep reinforcement learning,

    Q. Zhang, Z. Wang, C. Long, C. Huang, S. Yiu, Y . Liu, G. Cong, and J. Shi, “Online anomalous subtrajectory detection on road networks with deep reinforcement learning,” in 39th IEEE International Conference on Data Engineering, ICDE 2023, Anaheim, CA, USA, April 3-7, 2023 , 2...

    2023

  20. [28]

    Deeper, broader and artier domain generalization,

    D. Li, Y . Yang, Y . Song, and T. M. Hospedales, “Deeper, broader and artier domain generalization,” in IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017 , 2017

    2017

  21. [29]

    Generalizing to unseen domains: A survey on domain generalization,

    J. Wang, C. Lan, C. Liu, Y . Ouyang, T. Qin, W. Lu, Y . Chen, W. Zeng, and P. S. Yu, “Generalizing to unseen domains: A survey on domain generalization,” IEEE Trans. Knowl. Data Eng. , vol. 35, no. 8, 2023

    2023

  22. [30]

    Metanorm: Learning to normalize few-shot batches across domains,

    Y . Du, X. Zhen, L. Shao, and C. G. M. Snoek, “Metanorm: Learning to normalize few-shot batches across domains,” in 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021 , 2021

    2021

  23. [31]

    Stable predic- tion across unknown environments,

    K. Kuang, P. Cui, S. Athey, R. Xiong, and B. Li, “Stable predic- tion across unknown environments,” in Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, KDD 2018, London, UK, August 19-23, 2018 , 2018, pp. 1617– 1626

    2018

  24. [32]

    Deep stable learning for out-of-distribution generalization,

    X. Zhang, P. Cui, R. Xu, L. Zhou, Y . He, and Z. Shen, “Deep stable learning for out-of-distribution generalization,” in IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, 2021, pp. 5372–5382

    2021

  25. [33]

    Environment inference for invariant learning,

    E. Creager, J. Jacobsen, and R. S. Zemel, “Environment inference for invariant learning,” in Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event , ser. Proceedings of Machine Learning Research, vol. 139, 2021, pp. 2189– 2200

    2021

  26. [34]

    Domain-adversarial training of neural networks,

    Y . Ganin, E. Ustinova, H. Ajakan, P. Germain, H. Larochelle, F. Lavio- lette, M. Marchand, and V . S. Lempitsky, “Domain-adversarial training of neural networks,” J. Mach. Learn. Res., vol. 17, pp. 59:1–59:35, 2016

    2016

  27. [35]

    Causal inference by using invariant prediction: identification and confidence intervals,

    J. Peters, P. B ¨uhlmann, and N. Meinshausen, “Causal inference by using invariant prediction: identification and confidence intervals,” Journal of the Royal Statistical Society Series B: Statistical Methodology , vol. 78, no. 5, pp. 947–1012, 2016

    2016

  28. [36]

    Learning to generalize: Meta-learning for domain generalization,

    D. Li, Y . Yang, Y . Song, and T. M. Hospedales, “Learning to generalize: Meta-learning for domain generalization,” in Proceedings of the Thirty- Second AAAI Conference on Artificial Intelligence, (AAAI-18), the 30th innovative Applications of Artificial Intelligence (IAAI-18)...

    2018

  29. [37]

    Invariant risk minimization,

    M. Arjovsky, L. Bottou, I. Gulrajani, and D. Lopez-Paz, “Invariant risk minimization,” CoRR, vol. abs/1907.02893, 2019

    1907 arXiv

  30. [38]

    Causal intervention for weakly-supervised semantic segmentation,

    D. Zhang, H. Zhang, J. Tang, X. Hua, and Q. Sun, “Causal intervention for weakly-supervised semantic segmentation,” in Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virt...

    2020

  31. [39]

    Visual commonsense R- CNN,

    T. Wang, J. Huang, H. Zhang, and Q. Sun, “Visual commonsense R- CNN,” in 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020 , 2020, pp. 10 757–10 767

    2020

  32. [40]

    On the estimation of treatment effect with text covariates,

    L. Yao, S. Li, Y . Li, H. Xue, J. Gao, and A. Zhang, “On the estimation of treatment effect with text covariates,” in Proceedings of the Twenty- Eighth International Joint Conference on Artificial Intelligence, IJCAI 2019, Macao, China, August 10-16, 2019 , 2019, pp. 4106–4113

    2019

  33. [41]

    Text and causal inference: A review of using text to remove confounding from causal estimates,

    K. A. Keith, D. Jensen, and B. O’Connor, “Text and causal inference: A review of using text to remove confounding from causal estimates,” in Proceedings of the 58th Annual Meeting of the Association for Compu- tational Linguistics, ACL 2020, Online, July 5-10, 2020 . Associati...

    2020

  34. [42]

    Fairness and explanation in clustering and outlier detec- tion,

    I. Davidson, “Fairness and explanation in clustering and outlier detec- tion,” in KDD ’21: The 27th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Virtual Event, Singapore, August 14-18, 2021, 2021, p. 4037

    2021

  35. [43]

    Improving ad click prediction by considering non-displayed events,

    B. Yuan, J. Hsia, M. Yang, H. Zhu, C. Chang, Z. Dong, and C. Lin, “Improving ad click prediction by considering non-displayed events,” in Proceedings of the 28th ACM International Conference on Information and Knowledge Management, CIKM 2019, Beijing, China, November 3-7, 2019...

    2019

  36. [44]

    Uncovering causal effects of online short videos on consumer behaviors,

    Z. Tan, S. Zhang, N. Hong, K. Kuang, Y . Yu, J. Yu, Z. Zhao, H. Yang, S. Pan, J. Zhou, and F. Wu, “Uncovering causal effects of online short videos on consumer behaviors,” in WSDM ’22: The Fifteenth ACM International Conference on Web Search and Data Mining, Virtual Event / Te...

    2022

  37. [45]

    Deconfounded recommendation for alleviating bias amplification,

    W. Wang, F. Feng, X. He, X. Wang, and T. Chua, “Deconfounded recommendation for alleviating bias amplification,” in KDD ’21: The 27th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Virtual Event, Singapore, August 14-18, 2021 , 2021, pp. 1717– 1725

    2021

  38. [46]

    Pearl, Causality

    J. Pearl, Causality. Cambridge university press, 2009

    2009

  39. [47]

    Glymour, J

    M. Glymour, J. Pearl, and N. P. Jewell, Causal inference in statistics: A primer. John Wiley & Sons, 2016

    2016

  40. [48]

    A theory of inferred causation,

    J. Pearl and T. Verma, “A theory of inferred causation,” in Proceedings of the 2nd International Conference on Principles of Knowledge Repre- sentation and Reasoning (KR’91). Cambridge, MA, USA, April 22-25, 1991, 1991, pp. 441–452

    1991

  41. [49]

    A general identification condition for causal effects,

    J. Tian and J. Pearl, “A general identification condition for causal effects,” in Proceedings of the Eighteenth National Conference on Artifi- cial Intelligence and Fourteenth Conference on Innovative Applications of Artificial Intelligence, July 28 - August 1, 2002, Edmonton,...

    2002

  42. [50]

    Complete identification methods for the causal hierarchy,

    I. Shpitser and J. Pearl, “Complete identification methods for the causal hierarchy,” J. Mach. Learn. Res. , vol. 9, pp. 1941–1979, 2008

    1941

  43. [51]

    Adam: A method for stochastic optimiza- tion,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimiza- tion,” in 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015

    2015

  44. [52]

    A note on two problems in connexion with graphs,

    E. W. Dijkstra, “A note on two problems in connexion with graphs,” in Edsger Wybe Dijkstra: His Life, Work, and Legacy , K. R. Apt and T. Hoare, Eds., 2022, pp. 287–290

    2022

  45. [53]

    Lstm-based encoder-decoder for multi-sensor anomaly detection,

    P. Malhotra, A. Ramakrishnan, G. Anand, L. Vig, P. Agarwal, and G. Shroff, “Lstm-based encoder-decoder for multi-sensor anomaly detection,” CoRR, vol. abs/1607.00148, 2016. [Online]. Available: http://arxiv.org/abs/1607.00148

    2016 arXiv

  46. [54]

    beta-vae: Learning basic visual concepts with a constrained variational framework,

    I. Higgins, L. Matthey, A. Pal, C. P. Burgess, X. Glorot, M. M. Botvinick, S. Mohamed, and A. Lerchner, “beta-vae: Learning basic visual concepts with a constrained variational framework,” in 5th Inter- national Conference on Learning Representations, ICLR 2017, Toulon, France...

    2017

  47. [55]

    Disentangling by factorising,

    H. Kim and A. Mnih, “Disentangling by factorising,” in Proceedings of the 35th International Conference on Machine Learning, ICML 2018, Stockholmsm¨assan, Stockholm, Sweden, July 10-15, 2018, vol. 80, 2018, pp. 2654–2663

    2018

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.