Pith. sign in

REVIEW 4 major objections 5 minor 34 references

Causal Sensitivity Identification using Generative Learning

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Pith's one-line read The paper claims that causal features and paths can be identified without a known causal graph by training a CVAE predictor on data with features altered by hand, and that conditioning on the identified features improves prediction.

desk verdict The causal claims don't survive contact with the equations, but the GeoLife feature-conditioning results are real and competitive; resubmit as a generative feature-importance paper. read the letter →

arxiv 2509.01352 v1 pith:U4XNPDFQ submitted 2025-09-01 cs.LG cs.AI

classification cs.LGcs.AI
keywords causalsensitivityconditionalvariationalautoencoderinterventionalanalysiscounterfactualpathidentificationnext-locationpredictionGeoLifegenerative
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 claims you can find which input features actually cause a prediction without knowing the causal graph, by comparing a CVAE predictor trained on factual data with one trained on data where a candidate feature's values are replaced. If the intervened-data model predicts the same test set more accurately, the feature is labelled causally sensitive. If predictions on counterfactually altered test data are less accurate than factual ones, the paper infers a direct causal path from the altered variable to the outcome. The practical payoff is that conditioning the predictor on these causal features improves next-location accuracy on the GeoLife dataset and recovers known causal dependencies on the Asia benchmark.

What carries the argument

The central object is the Conditional Variational Autoencoder (CVAE) used as a generative predictor. The load-bearing identity is the accuracy gap ΔAcc = Acc_interventional − Acc_factual: a positive gap marks a causally sensitive feature, and Acc_counterfactual − Acc_factual negative marks a causal path. The counterfactual latent representation Z_CF, obtained by encoding altered test inputs with the factual model, carries the counterfactual comparison.

What would settle it

On a dataset whose true causal graph is known, pick a variable Z that is not an ancestor of Y, and choose an alteration that strongly shifts its marginal distribution (e.g., replace the most frequent category with a rare one). If the intervened-data CVAE achieves higher test accuracy on Y than the factual CVAE, the paper's rule would label Z causally sensitive; observing that would falsify the claim that ΔAcc>0 is a causal signal. Conversely, on the Asia graph, intervening on a known cause via true do(X=x') should reproduce the paper's sign pattern for bronc and either; if it does not, the han

Watch

Extended reading notes

Core claim

The paper's central claim is that causal sensitivity can be read off a prediction-performance gap. Train a CVAE on factual data (GP-F) and another on data where a candidate feature's values have been replaced by hand (GP-I). If GP-I predicts the original test set more accurately, the feature is causally sensitive; the paper interprets this as blocking a backdoor path through a common cause. Then, using only GP-F, encode altered test inputs to get a counterfactual latent representation; if counterfactual predictions are less accurate than factual ones, the paper infers a direct causal path X→Y. Applied to Asia's known graph, the method flags bronc and either as causal for dysp; on GeoLife, st

Load-bearing premise

The load-bearing assumption is that hand-replacing a feature's values acts like forcing that feature to a chosen value in the real system, so any accuracy change is causal rather than just a reaction to changed input patterns.

Editorial extensions

If this is right

  • A feature whose intervened-data model beats the factual model on the same test set can be treated as a causal influencer even when no causal graph is provided.
  • A counterfactual accuracy drop under altered inputs is interpreted as evidence of a direct path from the altered variable to the outcome.
  • Conditioning the CVAE decoder on identified causally sensitive features improves factual next-location accuracy and MRR relative to no conditioning.
  • Because the procedure imposes no acyclicity or graph-structure constraints, it can be applied to any supervised prediction task with categorical features.

Reading between the lines

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

  • The hand-altered value replacement is a proxy for do(X=x'); the paper does not establish equivalence. A direct test would compare the sign of ΔAcc against true do-calculus interventions on a synthetic DAG.
  • The counterfactual step encodes altered inputs through GP-F; if altered inputs fall off the training manifold, the latent shift may reflect out-of-distribution encoding rather than a true counterfactual. A calibration test could fit GP-F on both factual and intervened data and compare latent densities.
  • The method could be extended to continuous features by learning the alteration (e.g., adversarial perturbations) instead of swapping mode categories, which would test whether the criterion generalizes beyond categorical replacements.
  • Because accuracy is the only signal, a feature with true causal influence but weak predictive signal may be missed; combining ΔAcc with latent-space distance could tighten the criterion.
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

4 major / 5 minor

Summary. This paper proposes Causal Sensitivity Identification using a Conditional VAE. It trains a factual predictor GP-F on original data and an interventional predictor GP-I on data with one feature's values replaced, and declares a feature causally sensitive if GP-I's test accuracy exceeds GP-F's (Eq. 9). It then claims that encoding altered test inputs with GP-F's encoder and decoding produces counterfactuals, and that a drop in accuracy relative to factual predictions identifies a causal path X→Y (Eq. 11). The method is evaluated on Asia and GeoLife, and conditioning on the identified features is reported to improve next-location prediction.

Significance. If the proposed equivalence were valid, the method would be a practically attractive way to identify causal features and paths without a known graph, with a generative predictor. The paper also tackles a real application and includes a nontrivial spatiotemporal evaluation. However, the central identification step is not justified: retraining on feature-altered data estimates an observational conditional distribution, not a do-interventional one, and no identifiability argument is supplied. The counterfactual construction is not an SCM counterfactual. Consequently the main causal claims, and the performance gains attributed to causal conditioning, are not supported. The Asia experiment cannot rescue the premise because it uses the same operation. The contribution at present is a heuristic sensitivity analysis, not causal sensitivity identification.

major comments (4)
  1. [§3.2(b), Eq. (6)] The intervention step is not a do-intervention. Replacing a feature's values in the training set and retraining GP-I estimates p(Y | X_altered) under the original data-generating process; it does not estimate p(Y | do(X=x')) because Y was generated with the confounders intact and no adjustment set or graph is given. Equation (6) writes a backdoor-style sum over FCS, but FCS is exactly the quantity being inferred from ΔAcc, so the definition is circular. Any distribution shift—class-balance change, changed marginal frequencies, or model sensitivity—can move test accuracy, so ΔAcc>0 in Eq. (9) is not evidence of a confounder or a causal effect.
  2. [§3.2, Eq. (11)] The counterfactual construction does not implement the Pearl SCM counterfactual. Eq. (11) writes P(Y_{X=x'} | X=x, Y=y) as an integral over qφ(z|X=x,Y=y), but this requires the latent z to invert the true structural equations (abduction) and the decoder to be the structural outcome mechanism; neither condition is shown or derivable from CVAE training. Using ZCF = GP-F.encoder(Xtest altered) is not abduction-action-prediction. Therefore ΔAcc<0 and the inferred causal path X→Y in Algorithm 1 are unsupported.
  3. [§4.1, Tables 1–3] The Asia validation cannot establish the premise. The same flawed alteration operation is applied to the known-graph data, and 'agreement' of a few accuracy deltas with known edges is not an identifiability result. The tables also report differences of 0.005–0.01 as meaningful (e.g., [either, smoke, bronc] 0.850 vs 0.855) without confidence intervals or significance tests, while some entries go in the opposite direction (e.g., [either] factual 0.660 vs interventional 0.580). The CausalVAE comparison in Table 3 reports exactly 0.000 delta for every variable, which suggests a degenerate or ill-specified comparison rather than a substantive contrast.
  4. [§4.2, Table 4 and Algorithm 2] The reported performance improvement is selected on the same test set used for evaluation. Algorithm 1 uses Xtest accuracy to choose FCS, and Algorithm 2 then reports accuracy on the same Xtest after conditioning on that FCS; no held-out selection split or nested procedure is described. Additionally, the comparison in Table 4 is not apples-to-apples: the proposed GCSP generates n=20 samples per instance and reports the best-performing prediction, while the LSTM/Attention/DeepMove/MHSA baselines are evaluated as deterministic single predictions. These issues account for the Acc@1/MRR gains independently of any causal mechanism.
minor comments (5)
  1. [Eq. (5)] Equation (5) is not a valid probability statement: the RHS sums over FCS but FCS does not appear in the summand, and the equality is dimensionally meaningless.
  2. [§3.2 heading] The first sub-problem is labeled 'Causally sensitive feature (ZCF)', but ZCF is later used for the counterfactual latent; the intended symbol is FCS.
  3. [§4.2 Altered data] The two altered versions of the location sequence are both named LS1; the second should presumably be LS2.
  4. [References] The reference list contains duplicate entries for CausalVAE: [Yang et al., 2021a] and [Yang et al., 2021b] refer to the same paper.
  5. [Algorithm 1] Algorithm 1 lists 'Causally sensitive features' as an output, but the pseudocode only outputs ΔAcc and predictions; the feature set is never explicitly returned.

Circularity Check

3 steps flagged · score 6.0 of 10

The causal sensitivity labels are defined by the accuracy comparison used to report them, the backdoor formula in Eq. 6 presupposes the FCS set that the procedure outputs, and the counterfactual test feeds altered inputs into the factual model, making the causal-path conclusion a construction rather than an inference.

  1. self definitional [Section 3.2(e), Eq. (9), Algorithm 1]
    "If the prediction error is higher (i.e., accuracy is lower) in the factual scenario (c) than in the interventional scenario (d), this indicates, that the feature is causally sensitive and acts as a common influencer... ∆Acc = Accinterventional − Accfactual; ∆Acc > 0 ... if ∆Acc > 0 then Feature is causally sensitive"

    A feature is declared 'causally sensitive' exactly when the accuracy difference is positive. No independent causal quantity (e.g., a do-probability estimated from an identified adjustment set) is computed. The output label is therefore the input criterion by construction: 'causal influence' is defined as 'positive accuracy gap under altered training.' The later claim that conditioning on these features gives causally guided prediction is a restatement of the selection rule, not an independent confirmation.

  2. self definitional [Section 3.2(b), Eq. (6); Introduction, Intervention bullet]
    "P(Yt+1| do(X=Xaltered)) = Σ_FCS P(Yt+1| Xaltered, FCS)P(FCS) (6) ... To detect FCS, we compare prediction performance on identical test data using two models, one trained on the original (factual) train data and the other trained on intervened data where candidate features are altered."

    Equation 6 is presented as a backdoor-style adjustment over a set FCS, but FCS is not a known adjustment set: it is the output of the accuracy comparison in Eq. 9. The causal effect is written in terms of the very features whose causal status the procedure is supposed to establish. The formula therefore cannot justify the interventional accuracy gap; it assumes the conclusion (which features are confounders) as an input.

1 more flagged steps
  1. self definitional [Section 3.2, step 2(b)-(c), Eqs. (12)-(14)]
    "Obtain the counterfactual latent representation from the Xtest altered. ZCF = GP-F.encoder(Xtest altered) (12) ... Y(t+1)counterfactual = GP-F.decoder(ZCF, Xtest altered) (14) ... The difference in accuracy between counterfactual and factual scenarios ∆Acc < 0 ... signifies causal path X → Y."

    The counterfactual prediction is produced by passing the deliberately altered input through both the encoder and decoder of the factual model. Since the input was changed by construction (e.g., most frequent location replaced), the factual model's accuracy on this out-of-distribution input will drop by design. That drop is then labeled 'causal path X → Y.' The measurement is equivalent to the input perturbation: no abduction on the factual example is performed, so the counterfactual gap is a statement about distribution shift, not an identified causal effect.

full rationale

The paper's central estimator is definitionally circular: causal sensitivity is defined in Section 3.2(e) as a positive accuracy difference (Eq. 9), and the counterfactual causal path is defined as a negative accuracy difference after feeding altered inputs through the factual model (Eqs. 12-14). Equation 6 compounds this by writing the interventional distribution as an adjustment over FCS, the very set that the accuracy comparison is supposed to output. These are not merely unproven causal assumptions; the claimed 'predictions' of causal impact reduce to the accuracy comparisons used to define them. The Asia benchmark provides an external check on the labels and partly offsets the circularity in terms of empirical evaluation, and the self-citation to Bandyopadhyay and Sarkar (2023) is not load-bearing. However, the derivation chain itself does not identify causal effects: it renames a sensitivity-to-input-replacement measure as causal sensitivity and counterfactual evidence. This is partial, not total, circularity, so score 6 rather than 8 or 10; a full 8-10 would require the absence of any external benchmark or a derivation that is logically identical to its input throughout.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method's causal claims rest on several unverified assumptions: that data perturbation equals do-intervention, that the CVAE learns correct conditionals, that the identified features satisfy the backdoor criterion, and that the latent variable captures confounders. The free parameters are primarily the hand-chosen alteration rules and the candidate feature set. No new physical entities are introduced.

free parameters (3)
  • Intervention alteration rule = LS1: replace most frequent location with third most frequent; LS2: replace with location ID 0
    Hand-chosen perturbation of the training data; the choice directly affects DeltaAcc and which features are deemed causal.
  • Candidate features tested as FCS = W, Smin, DS
    Only three features are considered on GeoLife; the causal sensitivity verdict depends on this pre-selected set.
  • Number of samples n for prediction = 20
    Each test instance is sampled 20 times and the best-performing prediction is reported, a test-time selection that can inflate accuracy.
assumptions (4)
  • ad hoc to paper Retraining on a dataset with one feature's values replaced implements the structural intervention do(X=x')
    No argument is given that training on altered data matches the causal do-operator; it is a distribution-shift heuristic. Invoked in Section 3.2 step (b) and Equation 6.
  • domain assumption The CVAE decoder samples from the true conditional distribution P(Y|X) after training
    The accuracy-based causal inference assumes the generative predictor is well-calibrated; this is not verified anywhere.
  • domain assumption The features identified as causally sensitive form a sufficient adjustment set (backdoor criterion) for the effect of X on Y
    The paper states that conditioning on FCS blocks the backdoor path but never shows identifiability; Section 3.2.
  • domain assumption Unobserved confounders are absent or captured by the latent variable z
    The validity of the counterfactual approximation (Equation 11) relies on the latent variable representing all confounders.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Causal Sensitivity Identification using Generative Learning." pith.science (2026). https://pith.science/paper/U4XNPDFQ

@misc{pith2026250901352,
  author       = {Pith},
  title        = {Pith review of: Causal Sensitivity Identification using Generative Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/U4XNPDFQ}},
  note         = {Machine review of arXiv:2509.01352}
}
read the original abstract

In this work, we propose a novel generative method to identify the causal impact and apply it to prediction tasks. We conduct causal impact analysis using interventional and counterfactual perspectives. First, applying interventions, we identify features that have a causal influence on the predicted outcome, which we refer to as causally sensitive features, and second, applying counterfactuals, we evaluate how changes in the cause affect the effect. Our method exploits the Conditional Variational Autoencoder (CVAE) to identify the causal impact and serve as a generative predictor. We are able to reduce confounding bias by identifying causally sensitive features. We demonstrate the effectiveness of our method by recommending the most likely locations a user will visit next in their spatiotemporal trajectory influenced by the causal relationships among various features. Experiments on the large-scale GeoLife [Zheng et al., 2010] dataset and the benchmark Asia Bayesian network validate the ability of our method to identify causal impact and improve predictive performance.

Figures

Figures reproduced from arXiv: 2509.01352 by the authors.

Figure 1
Figure 1. Functional components of the proposed causal sensitivity identification framework with the factual (blue), interventional (red), and [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Causal graph depicting backdoor path (c) Factual prediction: Test the GP-F using Xtest. Y(t+1)factual = GP-F.decoder(ZF C , Xtest) (7) (d) Interventional prediction: Test the GP-I using Xtest. Y(t+1)Interventional = GP-I.decoder(ZIF , Xtest) (8) (e) Features having causal influence: If the predic￾tion error is higher (i.e., accuracy is lower) in the factual scenario (c) than in the interventional sce￾nario (d), this… view at source ↗
Figure 3
Figure 3. Causal graph - Asia [Ke et al., 2020] [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Causal Sensitivity Identification: a: Causally sensitive features, b: Performance across factual, intervention, and counterfactual [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Comparison of latent spaces ZF C and ZCF from the GP-F model: (a) shows the t-SNE projection; (b) shows the Jensen-Shannon divergence between their distributions. with softmax activation to predict the next location. The re￾construction loss uses sparse categorical cro…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 32 canonical work pages

  1. [1]

    Bandyopadhyay and S

    [Bandyopadhyay and Sarkar, 2023] S. Bandyopadhyay and S. Sarkar. Exploring causality aware data synthesis. In Proc. ACM AIMLSystems,

  2. [9]

    Learning neural causal models from unknown interventions

    [Ke et al., 2020] Nan Rosemary Ke, Olexa Bilaniuk, Anirudh Goyal, Stephan Bauer, Hugo Larochelle, Chris Pal, and Yoshua Bengio. Learning neural causal models from unknown interventions. In International Conference on Learning Representations (ICLR),

  3. [10]

    [Kingma and Welling, 2013] D. P. Kingma and M. Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114,

  4. [12]

    Mehta, and Shubham Choudhary

    [Krishna et al., 2018] Kalpit Krishna, Devendra Jain, Shob- hit V . Mehta, and Shubham Choudhary. An lstm-based system for prediction of human activities with durations. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, 1(4):147:1–147:31,

  5. [13]

    Lauritzen and David J

    [Lauritzen and Spiegelhalter, 1988] Steffen L. Lauritzen and David J. Spiegelhalter. Local computations with probabil- ities on graphical structures and their application to expert systems. Journal of the Royal Statistical Society: Series B (Methodological), 50(2):157–194,

  6. [16]

    Liu et al

    [Liu and others, 2021] X. Liu et al. Trajgans: Geo-privacy protection of trajectory data. In GIScience,

  7. [17]

    Causal effect inference with deep latent-variable models

    [Louizos et al., 2017] Christos Louizos, Uri Shalit, Joris Mooij, David Sontag, Rich Zemel, and Max Welling. Causal effect inference with deep latent-variable models. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS),

  8. [22]

    [Pearl, 2009] J. Pearl. Causality: Models, reasoning, and inference. Cambridge University Press, New York, 2nd edition,

Show all 34 references
  1. [24]

    Kitani, Dariu M

    [Rudenko et al., 2020] Andrey Rudenko, Luigi Palmieri, Michael Herman, Kris M. Kitani, Dariu M. Gavrila, and Kai O. Arras. Human motion trajectory prediction: A survey. International Journal of Robotics Research , 39(8):895–935,

  2. [25]

    Salzmann, B

    [Salzmann et al., 2020] T. Salzmann, B. Ivanovic, P. Chakravarty, and M. Pavone. Trajectron++: Multi-agent generative trajectory forecasting with heterogeneous data for control. CoRR, abs/2001.03093,

  3. [26]

    The bnlearn dataset reposi- tory

    [Scutari, 2009] Marco Scutari. The bnlearn dataset reposi- tory. https://www.bnlearn.com/bnrepository/,

  4. [27]

    [van der Maaten and Hinton, 2008] Laurens van der Maaten and Geoffrey Hinton

    Ac- cessed: 2025-05-25. [van der Maaten and Hinton, 2008] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne.Journal of Machine Learning Research, 9:2579–2605,

  5. [29]

    [Wang et al., 2022] S. Wang, J. Cao, and P. S. Yu. Deep learning for spatio-temporal data mining: A survey. IEEE Trans. Knowl. Data Eng., 34(8):3681–3700,

  6. [30]

    Causal models for counterfactual identification and esti- mation

    [Xia et al., 2023] Kun Xia, Yang Pan, and Elias Bareinboim. Causal models for counterfactual identification and esti- mation. In Proceedings of the International Conference on Learning Representations (ICLR),

  7. [31]

    Characterizing and learning equivalence classes of causal dags under interventions

    [Yang et al., 2018] Karren Yang, Abigail Katcoff, and Car- oline Uhler. Characterizing and learning equivalence classes of causal dags under interventions. In Proceedings of the 35th International Conference on Machine Learning (ICML), pages 5541–5550. PMLR,

  8. [32]

    Zheng, X

    [Zheng et al., 2010] Y . Zheng, X. Xie, and W.-Y . Ma. Ge- olife: A collaborative social networking service among user, location and trajectory. IEEE Data Eng. Bull. , 33(2):32–39,

  9. [34]

    [Zuo et al., 2022] A. Zuo, S. Wei, T. Liu, B. Han, K. Zhang, and M. Gong. Counterfactual fairness with partially known causal graph. In Proc. AAAI Conference on Arti- ficial Intelligence, 2022

  10. [1988]

    Cyclical annealing schedule: A simple approach to mitigating kl vanishing

    [Li et al., 2019] Chunyuan Li, Xiujun Liu, Jianfeng Gao, Asli Celikyilmaz, and Lawrence Carin. Cyclical annealing schedule: A simple approach to mitigating kl vanishing. In Proceedings of the 2019 Conference of the North Amer- ican Chapter of the Association for Computational ...

  11. [1997]

    Martin, Y

    [Martin et al., 2022] H. Martin, Y . Hong, N. Wiedemann, D. Bucher, and R. Martin. Trackintel: An open-source python library for human mobility analysis.arXiv preprint arXiv:2206.03593,

  12. [2007]

    [Hong et al., 2023] Y . Hong, Y . Zhang, K. Schindler, and M. Raubal. Context-aware multi-head self-attentional neu- ral network model for next location prediction. Transp. Res. Part C: Emerg. Technol., 156,

  13. [2008]

    Vaswani, N

    [Vaswani et al., 2017] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polo- sukhin. Attention is all you need. In Proc. 31st NeurIPS, pages 5998–6008. Curran Associates,

  14. [2009]

    [Pearl, 2019] J. Pearl. Causal and counterfactual inference. Springer, New York,

  15. [2010]

    [Zheng et al., 2018] Xun Zheng, Bryon Aragam, Pradeep Ravikumar, and Eric P. Xing. Dags with no tears: Con- tinuous optimization for structure learning. In Advances in Neural Information Processing Systems (NeurIPS), vol- ume 31,

  16. [2013]

    Counterfactuals and causal reasoning

    [Kment, 2020] Boris Kment. Counterfactuals and causal reasoning. In Perspectives on Causation: Selected Pa- pers from the Jerusalem 2017 Workshop, pages 463–482. Springer,

  17. [2014]

    Hershey and Peder A

    [Hershey and Olsen, 2007] John R. Hershey and Peder A. Olsen. Approximating the kullback–leibler divergence be- tween gaussian mixture models. In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , volume 4, pages IV–317. IEEE,

  18. [2017]

    MacDonald and Walter Zucchini

    [MacDonald and Zucchini, 1997] Iain L. MacDonald and Walter Zucchini. Hidden Markov and Other Models for Discrete-Valued Time Series, volume 110 of Monographs on Statistics and Applied Probability . CRC Press, Boca Raton, FL,

  19. [2018]

    [Feng et al., 2020] J. Feng, Z. Yang, F. Xu, H. Yu, M. Wang, and Y . Li. Learning to simulate human mobility. In Pro- ceedings of the 26th ACM SIGKDD Conference on Knowl- edge Discovery and Data Mining, pages 3426–3433,

  20. [2019]

    A hierarchical temporal attention-based lstm encoder-decoder model for individual mobility pre- diction

    [Li et al., 2020] Fei Li, Zhen Gui, Zhili Zhang, Dawei Peng, Shuang Tian, Kai Yuan, Yafei Sun, Huayi Wu, Jing Gong, and Yinjie Lei. A hierarchical temporal attention-based lstm encoder-decoder model for individual mobility pre- diction. Neurocomputing, 403:153–166,

  21. [2020]

    [Goodfellow et al., 2014] I. J. Goodfellow, J. Pouget- Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y . Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems ,

  22. [2021]

    [Feng et al., 2018] J. Feng, L. Yong, C. Zhang, F. Sun, F. Meng, A. Guo, and D. Jin. Deepmove: Predicting hu- man mobility with attentional recurrent networks. InProc. WWW Conf., pages 1459–1468,

  23. [2022]

    [Nezhadettehad et al., 2024] Amin Nezhadettehad, Arkady Zaslavsky, Rafiq Abdur, S. A. Shaikh, Seng W. Loke, Guang-Li Huang, and Ali Hassani. Predicting next use- ful location with context-awareness: The state-of-the-art. arXiv preprint arXiv:2401.08081,

  24. [2023]

    When selection meets intervention: Additional complexities in causal discovery

    [Dai et al., 2025] Haoyue Dai, Yaqi Xue, Krzysztof Chalupka, and Elias Bareinboim. When selection meets intervention: Additional complexities in causal discovery. In International Conference on Learning Representations (ICLR),

  25. [2024]

    The Book of Why: The New Science of Cause and Effect

    [Pearl and Mackenzie, 2018] Judea Pearl and Dana Macken- zie. The Book of Why: The New Science of Cause and Effect. Basic Books,

  26. [2025]

    [Doersch, 2021] C. Doersch. Tutorial on variational autoen- coders. arXiv preprint arXiv:2111.10846,

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.