Pith. sign in

REVIEW 3 major objections 6 minor 46 references

Generating Hypotheses of Dynamic Causal Graphs in Neuroscience: Leveraging Generative Factor Models of Observed Time Series

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Dynamic causal graphs in time series can be recovered as behavior-weighted mixtures of a few static nonlinear factor graphs, and the REDCLIFF-S model recovers them with 22–28% higher F1 scores than baselines on synthetic benchmarks.

desk verdict A genuinely new dynamic-graph factor model with honest limitations; the synthetic wins are real but in-distribution, and the paper oversells the no-assumption framing. read the letter →

arxiv 2505.20697 v3 pith:ZALQ7YU2 submitted 2025-05-27 cs.LG cs.AIstat.APstat.ML

Zachary C. Brown , David Carlson This is my paper · ORCID
classification cs.LGcs.AIstat.APstat.ML MSC 62M1062D2068T07
keywords dynamiccausaldiscoveryGrangercausalitynonlineartimeseriesfactormodelshypothesisgenerationbehavioralstatelabelsneurosciencestate-dependentdynamics
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's aim is to show that time-varying causal structure, of the kind that switches as the brain changes behavioral state, can be recovered from observed time series by writing the unknown system as a weighted sum of a few reusable static interaction graphs, each of which may be nonlinear. The proposed model, REDCLIFF-S, learns the interaction graphs by forecasting the next observation and learns the per-moment weights from the recent history of the signal, optionally supervised by behavioral labels. This matters because current methods assume either linear relationships or a single fixed graph, leaving hypothesis generation blind to state-dependent neural wiring. On synthetic systems and a multi-state adaptation of DREAM4, the paper reports average F1-score improvements of roughly 22–28% over those baselines, and on two real local-field-potential datasets it produces behavioral-state-specific wiring hypotheses. The authors themselves prove in an appendix that such systems are generically not identifiable, and they argue that hypothesis generation remains useful despite this.

What carries the argument

The load-bearing identity is the conditionally weighted superposition $\Phi(t) = G(t) \odot^\dagger F(t) = \sum_k G(t)_k F_k(t)$, where $\odot^\dagger$ denotes broadcast multiplication of the score vector $G(t)$ against the tensor of factor outputs. The identity splits estimation into two tasks: each factor $f_{\phi_k}$ is a cMLP whose first-layer weights are time-lagged adjacency matrices $\hat{A}_{:,:,t}$, so a factor literally is an inspectable causal graph; and the state model $g_\theta$ maps a longer history window to nonnegative scores, deciding which factor dominates at each time step. A cosine-similarity penalty between factor graphs enforces the prior that factors should serve different purposes, and in REDCLIFF-S an invertible map lets a subset of scores double as predicted behavioral labels, so the factorization carries the hypothesis directly.

What would settle it

Run the method on a system whose causal edge strengths drift continuously through many values rather than switching between a few patterns; if REDCLIFF-S's fixed factors cannot track the drift at better than static-baseline accuracy, the superposition model is the reason. A sharper test is the paper's own two-node counterexample, in which ReLU and nReLU edges become observationally equivalent: Appendix A.1 argues no method can identify which state generated a sample, and verifying empirically that REDCLIFF-S assigns such samples at chance level would confirm the stated limit rather than a recoverable signal.

Watch

Extended reading notes

Core claim

REDCLIFF-S claims that a dynamic causal graph can be modeled as $\Phi(t) = \sum_{k=1}^{K} G(t)_k F_k(t)$: a small number of static factor functions $F_k$, each mapping past variable values to present values and each yielding an explicit time-lagged adjacency matrix of Granger causal weights, combined by history-conditioned scores $G(t)$ produced by a state model $g_\theta$. Because each factor is a component-wise MLP, the relationships inside a factor can be nonlinear, and because the scores vary with time, the composite graph can change direction and strength without assuming a discrete switch between states. In the supervised variant, behavioral labels are attached to particular factor scores through invertible maps, so each factor's adjacency matrix can be read as the candidate wiring hypothesis for that behavior. The paper reports that this construction outperforms static nonlinear, dynamic-linear, and regime-switching baselines in F1 and ROC-AUC on synthetic data and on D4IC, a five-fold mixture version of DREAM4, and it demonstrates factor graphs for Open Field, Home Cage, Tail Suspension, and Social Preference states that reproduce published findings such as lateral habenula involvement in stress.

Load-bearing premise

The recorded system must actually be expressible as a small set of reusable interaction patterns whose relative importance at each moment is set by recent history; if the true interactions instead morph continuously into new forms, the factor graphs and scores have no faithful target to recover.

Editorial extensions

If this is right

  • Each trained factor yields a fixed, inspectable adjacency matrix, so a REDCLIFF-S run outputs concrete candidate edges, with direction, lag, and strength, that can seed interventional studies such as dynamic causal models.
  • Attaching behavioral labels to factor scores means the output includes a wiring diagram per behavior rather than one averaged graph; on the TST data this produces distinct Open Field, Home Cage, and Tail Suspension hypotheses from a single training run.
  • Because factors are nonlinear, the approach beats linear dynamic factor models on systems with ReLU and nReLU edges, where the ground truth is unrepresentable by any linear model.
  • The ablation results attribute most of the gain to multiple factors, learned scores, and the supervised term, while the cosine-similarity penalty helps on sparse systems and can be dropped on dense ones.
  • The authors' non-identifiability proof implies that no method of this family can uniquely recover true factor graphs in general, so the practical claim is narrower: the recovered graphs are useful hypotheses even when not unique.

Reading between the lines

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

  • If real neural dynamics only approximate the superposition assumption, factor graphs may still serve as a low-dimensional grammar of the dominant interaction modes, with the scores acting as a behavioral readout; under misspecification the model would remain a dimension-reduction tool rather than an exact recovery method.
  • The cosine-similarity penalty is a proxy for functional dissimilarity of graphs, and a testable extension would replace it with a sparsity prior on factor usage to check whether rare behavioral states are recovered better.
  • The TST hypothesis that the anterior-lateral habenula drives several regions in the stressful Open Field condition is directly interventional: silencing the lateral habenula during tail suspension would test whether that edge carries causal weight.
  • Because the model needs no interventional data and accepts any scalar time series with global state labels, it could transfer to sleep staging, anesthesia depth monitoring, or economic regime detection, wherever states mark distinct dynamics.
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 / 6 minor

Summary. The paper proposes REDCLIFF-S, a factor-based Granger causal discovery method that models dynamic causal graphs as a history-conditioned weighted sum of static factor graphs, with each factor implemented as a cMLP and regularized for sparsity and factor dissimilarity; optional supervised labels are included through an invertible state model. The authors claim substantial improvements over static and linear baselines in detecting nonlinear, state-dependent causal interactions, based on synthetic benchmarks (Synthetic Systems and D4IC) and two real-world LFP case studies (TST and Social Preference). The paper also releases code and reports standard errors and ROC-AUC values alongside F1 scores.

Significance. If the central claims were fully supported, this would be a useful contribution to hypothesis generation in neuroscience: the method is explicitly designed for interpretability (fixed factor graphs, behavior-linked factor scores), combines nonlinear factors with dynamic weighting in a natural way, and is evaluated against multiple baselines with released code and standard errors. The design principle of reusable factor graphs with state-dependent weights is plausible, and the in-distribution synthetic results are encouraging. However, as detailed below, the external validity of the evaluation and the consistency between the paper's identifiability analysis and its benchmark design require substantial work before the abstract's claims about detecting complex, time-varying interactions can be accepted.

major comments (3)
  1. [Section 4.1 and Eq. (4)] The synthetic evaluation is in-distribution for the model class. Section 4.1 generates each recording as a time-varying weighted superposition of VAR models with ReLU/nReLU edges, exactly the form of Eq. (4), and Section 4.3 constructs D4IC by the same mixing mechanism over DREAM4 folds. The Section 1 statement that 'we make no assumptions regarding the underlying generative processes' is therefore contradicted by both the model form and the evaluation design. Because the real-data case studies in Section 5 lack ground truth, the paper provides no evidence for performance when the true dynamics are not a small weighted superposition of reused static graphs. Please add experiments where the target system is outside the model class (e.g., continuously morphing edge weights, or edge functions that change form over time) or substantially soften the no-assumptions claim and the abstract's characterization of what is supported.
  2. [Appendix A.1 and Section 4.1] The paper's own identifiability result directly weakens the benchmark's validity. Appendix A.1 proves that a two-node system with ReLU and nReLU edges and symmetric innovations has two observationally equivalent state-dependent graphs, and Section 4.1 explicitly uses ReLU and nReLU activations in the synthetic systems. The same sign/flip symmetry may therefore be present in the synthetic ground truth, which would mean that edge-level F1 and ROC-AUC scores in Section 4.2 and Tables 2, 5, and 6 penalize algorithms for recovering observationally equivalent alternatives. Please either establish that the synthetic data generation avoids the A.1 symmetry (e.g., via asymmetric innovation distributions or initial conditions), or evaluate against equivalence classes rather than a single ground-truth graph; at minimum, the relationship between A.1 and the benchmark's ground truth should be discussed explicitly.
  3. [Section 3.5 and Section 5] The behavioral link is partly imposed by the supervised construction and should not be described as an emergent discovery. In Eq. (8), the first B factor weights are assigned to behavioral labels, and Eq. (9) trains the state model to predict those labels through the lambda*MSE term. Consequently, the finding that specific factors 'correspond' to Open Field or Home Cage states in the TST case study reflects the imposed supervised structure as much as an independent property of the data. The real-data analysis should be framed as hypothesis generation with the caveat that the behavior-factor association is built into the loss, and the authors should provide a corresponding unsupervised or held-out validation to support the claim of uncovering behavior-linked relationships.
minor comments (6)
  1. [Abstract and Appendix C.1] The phrase 'in some of our experiments' makes the headline improvement of 22-28% (and 'well over 60%') untestable; please cite the specific systems and figure numbers (e.g., Figure 6 and Supplementary Figure 9) in the abstract or clarify the qualifier in the main text.
  2. [Section 4.2] The system naming convention '6-2-2', '6-4-2', '12-11-2', '12-11-5' is not defined; please state at first use that the triple denotes (number of nodes, number of inter-variable edges, number of factors).
  3. [Table 3 and Appendix C.6] The up/down arrows in Table 3 are ambiguous because the baseline for comparison is not specified; please state that arrows indicate change relative to the full REDCLIFF-S model, and note that for the two systems where the rho=0 ablation improves performance, this should be acknowledged in the main text discussion of the cosine-similarity penalty.
  4. [Figure 5] The color scale for 'difference in mean normalized causal estimates' is not defined; please describe how the normalization and difference are computed, since the figure is the primary evidence for the TST case study.
  5. [Appendix A.2.2] The proof sketch of Lemma 1 is difficult to follow: the notation p(f_i, X) is used in a way that conflates a joint density and a conditional likelihood, and the step from Eq. (13) to Eq. (14) relies on the uniform-state prior in a way that should be stated explicitly; please rewrite the derivation for clarity.
  6. [General] There are minor typographical and formatting issues, including 'it's use' in Section 4.3, inconsistent 'Open Field (OP)' vs 'OF' in Appendix D.1 and Section 5, and the unexpanded 'V AR' spacing in several places.

Circularity Check

2 steps flagged · score 4.0 of 10

All synthetic benchmarks generate the true system inside the model's own Eq. 4 superposition class, so the paper's central structural assumption is never tested; the behavior-linked factors in the TST case study are partly imposed by the supervised loss of Eq. 9.

  1. self definitional [Section 3.1 (Problem Statement); Section 3.2/Eq. 4; Section 4.1 (Synthetic Data Generation); Section 4.3 (D4IC)]
    "Indeed, in Sections 4.1, 4.2, and 4.3 we implement Φ as a factor model itself (see Section 3.2 for formalization). ... Recordings obtained from each VAR model were then weighted over time with linearly interpolated weights (randomly selected between 0 and 1). These weighted recordings were added together along with a level of Gaussian noise."

    The benchmarks define the target system by the model's own hypothesis. Eq. 4 sets REDCLIFF's form as Φ(t) = Σ_k G(t)_k F_k(t), and Section 3.1 states that in all experiments 'we implement Φ as a factor model itself': each synthetic system is a set of fixed VAR factors combined by time-varying scalar weights ('weighted over time with linearly interpolated weights... added together'), and D4IC likewise superposes DREAM4 folds by adding down-weighted recordings. Hence every ground-truth graph the method is tested on is, by construction, a small weighted sum of reused static factor graphs — exactly the class the model assumes.

  2. fitted input called prediction [Section 3.5 (REDCLIFF-S, Eqs. 8-9, λ=100 in Tables 19-20); Section 5 (TST case study, Figure 5)]
    "Thus, we follow prior literature and assign behaviors to the first B elements of α (Mague et al., 2022) ... we add a supervised component to L in Eq. 7 to obtain L(ϕ, θ,D) = Lf (ϕ, D) + Lg(θ, D) + λMSE(Y, Ŷ) (9) ... In Figure 5 we show the difference between mean normalized causal estimates for the REDCLIFF-S factors assigned to the Open Field (OF) and Home Cage (HC) behavioral paradigms."

    The factor-behavior coupling is imposed by the loss, then reported as a discovery. Eq. 9 supervises the first B factor scores to reproduce the behavioral labels (λ = 100), and Eq. 8 defines the predicted label as an invertible function of exactly those scores, ŷ = gy(gα^{-1}(α:B)). The TST case study then presents the difference between 'REDCLIFF-S factors assigned to the Open Field (OF) and Home Cage (HC) behavioral paradigms' as evidence of behavior-linked causal structure. Under Eq. 9, factors are forced to track the behavioral states during training, so the finding that the OF factor differs from the HC factor is partly guaranteed by construction; only the edge weights inside each factor are genuinely learned from the forecasting objective.

full rationale

The paper's principal empirical claims — roughly 22-28% mean optimal-f1 improvements over static/linear baselines on Synthetic Systems and D4IC — are anchored to externally defined ground-truth graphs and external baselines, so they are not fabricated by fitting. The model genuinely learns nontrivial quantities: recovered accuracy varies with complexity and SNR (ROC-AUC 0.63-0.80; ablations degrade performance), and the baselines are independent implementations. However, the method's central structural premise — that real dynamics are a small weighted superposition of reused static factor graphs (Eq. 4) — is never stress-tested: Section 3.1 states that in all experiments 'we implement Φ as a factor model itself,' with VAR factors and linearly interpolated weights, so the Synthetic Systems and D4IC ground truths are generated inside the model's own hypothesis class. This is compounded by the supervised construction of Eqs. 8-9, where behaviors are assigned to the first B factor scores and used as training targets, so the behavior-linked factor separation in the TST case study is in part imposed by the objective rather than discovered. Real-data corroboration also leans partly on same-group prior work (Talbot et al., 2023; Mague et al., 2022), but the synthetic comparisons do not depend on those citations, so self-citation is not load-bearing for the main result. Net: partial circularity concentrated in the self-definitional benchmarks and the supervised factor-behavior coupling; score 4.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

No new physical entities are postulated. The factors and state model are internal model components whose outputs, factor scores and adjacency matrices, are testable hypotheses rather than new ontology. The main ledger entries are the modeling assumptions that the system is a fixed finite superposition of reusable nonlinear graphs and that behavior labels align with factor identities.

free parameters (5)
  • Loss coefficients eta, omega, rho, gamma, lambda = omega=10.0; rho=1/sum(i); gamma=0.001; lambda=100.0; eta=0.1/(nk*sqrt(nc^2-1))
    Grid-searched on a single synthetic repeat and transferred to other experiments; these coefficients control sparsity, forecasting, factor dissimilarity, and supervision, and directly shape the learned adjacency matrices.
  • Number of factors nk = Set equal to dataset B for synthetic and D4IC; nk=9 for TST; nk in [2,4,6,9,18,36] for SP
    Chosen by model-selection criteria, not derived from theory. The entire superposition dimension is a free modeling choice.
  • State-model context length tau_cl (embedded lag tau_in + tau_cl) = 16 for synthetic and real datasets; 20 for D4IC
    A design choice justified by Lemma 1 but still selected per dataset and affecting which history the state model sees.
  • Behavioral presence threshold c_b = Not specified in the paper
    In Equation 8, predicted behavioral label presence is decided by thresholding y_hat at c_b, but the chosen value is not reported.
  • Input lag order tau_in for generative factors = 4 for REDCLIFF-S experiments; 2 for synthetic system generation
    Fixed by hand; the maximum lag over which causal relationships are estimated, so it directly bounds which edges can be found.
assumptions (5)
  • domain assumption The observed time series are generated by a system Phi(t) that is exactly a finite weighted superposition of K* static factor functions F_k with weights G(t), as in Section 3.2 and Equation 4.
    This defines the model class. If the real system has continuously evolving or non-factorized causal functions, the estimated factor graphs and weights have no faithful target.
  • domain assumption The first-layer weights of each cMLP factor faithfully represent Granger causal relationships between variables.
    Borrowed from Tank et al. (2021) and used throughout the paper to extract adjacency matrices from the learned factors.
  • domain assumption Global state or behavioral labels are available and correspond to one dominant generative factor at each time step.
    Section 3.5 and the synthetic label generation procedure assume that a one-hot behavioral label identifies which factor weight is largest, which is what makes the supervised variant meaningful.
  • domain assumption Cosine similarity between estimated factor graphs is a useful proxy for graph dissimilarity and for model selection and stopping.
    Appendix B.3 reports empirical correlation with ROC-AUC, but the cosine penalty is not proven to equal any true graph-distance measure, and ablations show it can hurt on dense systems.
  • domain assumption A dynamic causal graph can be fairly summarized by lag-summed adjacency matrices and binary thresholded edges for evaluation.
    Appendix B.2 compresses all baseline graph representations to R^{nc x nc} by summing over lags or features, which discards timing and sign information that may matter in neuroscience.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Generating Hypotheses of Dynamic Causal Graphs in Neuroscience: Leveraging Generative Factor Models of Observed Time Series." pith.science (2026). https://pith.science/paper/ZALQ7YU2

@misc{pith2026250520697,
  author       = {Pith},
  title        = {Pith review of: Generating Hypotheses of Dynamic Causal Graphs in Neuroscience: Leveraging Generative Factor Models of Observed Time Series},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZALQ7YU2}},
  note         = {Machine review of arXiv:2505.20697}
}
read the original abstract

The field of hypothesis generation promises to reduce costs in neuroscience by narrowing the range of interventional studies needed to study various phenomena. Existing machine learning methods can generate scientific hypotheses from complex datasets, but many approaches assume causal relationships are static over time, limiting their applicability to systems with dynamic, state-dependent behavior, such as the brain. While some techniques attempt dynamic causal discovery through factor models, they often restrict relationships to linear patterns or impose other simplifying assumptions. We propose a novel method that models dynamic graphs as a conditionally weighted superposition of static graphs, where each static graph can capture nonlinear relationships. This approach enables the detection of complex, time-varying interactions between variables beyond linear limitations. Our method improves f1-scores of predicted dynamic causal patterns by roughly 22-28% on average over baselines in some of our experiments, with some improvements reaching well over 60%. A case study on real brain data demonstrates our method's ability to uncover relationships linked to specific behavioral states, offering valuable insights into neural dynamics.

Figures

Figures reproduced from arXiv: 2505.20697 by the authors.

Figure 1
Figure 1. Illustration of the REDCLIFF(-S) algorithm. A) The two primary subtasks performed by a REDCLIFF(-S) model. B) An illustration of a nk = 3 factor REDCLIFF-S model’s forward pass. We seek to estimate causal patterns linked to specific behav￾iors in signals. Thus, we follow prior literature and assign behaviors to the first B elements of α (Mague et al., 2022). To enforce better separation between the factors fϕk and s… view at source ↗
Figure 2
Figure 2. Synthetic Systems results from select systems. A) Av￾erage optimal f1-score and standard error of the mean (SEM) between true and estimated inter-variable causal relationships. B) Average pairwise improvement and SEM between optimal f1- scores obtained by REDCLIFF-S and baselines. Appendix D). Here we set the number of lags for each VAR to τ = 2 time steps for visualization purposes, but any in￾teger greater than 0 … view at source ↗
Figure 3
Figure 3. Visualizing true and REDCLIFF-S Top-10 estimated inter-variable causal relationships of a factor from Synthetic Sys￾tem “12-11-5”. REDCLIFF-S captures 7 of 11 true relationships, while incorrect/missed pathways share functional similarities; the fact that REDCLIFF-S extracted these relationships with four other system states/factors adding noise in the dataset marks a sea change in modeling capability over baselines… view at source ↗
Figures from the paper (19 more)
Figure 4
Figure 4. Figure 4: Average optimal f1-score between true and estimated inter-variable causal relationships across the D4IC HSNR dataset. a single dynamic-causal system. Non-stochastic noise was simulated by adding down-weighted recordings from all but one DREAM4 folds to a recording from…
Figure 5
Figure 5. Figure 5: Difference in REDCLIFF-S’ mean, normalized inter￾variable causal estimates in the Open Field vs Home Cage paradigms of our TST Case Study. Accumbens Core (Acb Core) to predict behavior in the Pre￾limbic and Infralimbic cortices, and from the Anterior-Left Lateral Haben…
Figure 6
Figure 6. Figure 6: Pairwise optimal f1-Score Improvement of REDCLIFF-S on the Synthetic Systems 6-2-2 experiment. 60%”; [PITH_FULL_IMAGE:figures/full_fig_p017_6.png]
Figure 7
Figure 7. Figure 7: Optimal f1-score on D4IC-LSNR experiments [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]
Figure 9
Figure 9. Figure 9: Summary plot of pairwise optimal f1-score improvement of REDCLIFF-S across Synthetic Systems experiments. In [PITH_FULL_IMAGE:figures/full_fig_p018_9.png]
Figure 10
Figure 10. Figure 10: Visualizing true (left column) and REDCLIFF-S Top-10 estimated (right column) inter-variable causal relationships of two factors (‘A’ - also depicted in main paper [PITH_FULL_IMAGE:figures/full_fig_p021_10.png]
Figure 12
Figure 12. Figure 12: Average improvement in optimal f1-scores ± SEM by the REDCLIFF-S algorithm over baselines from our Low￾complexity Synthetic Systems experiments [PITH_FULL_IMAGE:figures/full_fig_p022_12.png]
Figure 14
Figure 14. Figure 14: Average improvement in optimal f1-scores ± SEM by the REDCLIFF-S algorithm over baselines from our Low￾complexity Synthetic Systems experiments. 22 [PITH_FULL_IMAGE:figures/full_fig_p022_14.png]
Figure 16
Figure 16. Figure 16: Average improvement in optimal f1-scores ± SEM by the REDCLIFF-S algorithm over baselines from our Low￾complexity Synthetic Systems experiments [PITH_FULL_IMAGE:figures/full_fig_p023_16.png]
Figure 18
Figure 18. Figure 18: Average improvement in optimal f1-scores ± SEM by the REDCLIFF-S algorithm over baselines from our Moderate￾complexity Synthetic Systems experiments. 23 [PITH_FULL_IMAGE:figures/full_fig_p023_18.png]
Figure 20
Figure 20. Figure 20: Average improvement in optimal f1-scores ± SEM by the REDCLIFF-S algorithm over baselines from our Moderate￾complexity Synthetic Systems experiments [PITH_FULL_IMAGE:figures/full_fig_p024_20.png]
Figure 22
Figure 22. Figure 22: Average improvement in optimal f1-scores ± SEM by the REDCLIFF-S algorithm over baselines from our Moderate￾complexity Synthetic Systems experiments. 24 [PITH_FULL_IMAGE:figures/full_fig_p024_22.png]
Figure 24
Figure 24. Figure 24: Average improvement in optimal f1-scores ± SEM by the REDCLIFF-S algorithm over baselines from our High￾complexity Synthetic Systems experiments [PITH_FULL_IMAGE:figures/full_fig_p025_24.png]
Figure 25
Figure 25. Figure 25: Visualizing model selection in the Region-Averaged TST 100 Hz case study [PITH_FULL_IMAGE:figures/full_fig_p025_25.png]
Figure 28
Figure 28. Figure 28: Average estimated strength of causal relationships be￾tween channels recorded in the Tail Suspended state from the Region-Averaged TST 100 Hz experiment [PITH_FULL_IMAGE:figures/full_fig_p026_28.png]
Figure 29
Figure 29. Figure 29: Difference between Open Field vs Tail Suspended aver￾age estimated strength of causal relationships between channels from the Region-Averaged TST 100 Hz experiment [PITH_FULL_IMAGE:figures/full_fig_p026_29.png]
Figure 31
Figure 31. Figure 31: Visualizing model selection in the Region-Averaged SP 100 Hz case study [PITH_FULL_IMAGE:figures/full_fig_p027_31.png]
Figure 32
Figure 32. Figure 32: Average estimated strength of causal relationships be￾tween channels recorded in the Social Preference state from the Region-Averaged SP 100 Hz experiment [PITH_FULL_IMAGE:figures/full_fig_p027_32.png]
Figure 34
Figure 34. Figure 34: Difference in average estimated strength of causal relationships between channels predicted in the Social Preference vs Object Preference behavioral paradigms of the Region-Averaged SP 100 Hz experiment. 28 [PITH_FULL_IMAGE:figures/full_fig_p028_34.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 41 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    A., Liu, J., Kiehl, K

    Allen, E. A., Liu, J., Kiehl, K. A., Gelernter, J., Pearlson, G. D., Perrone-Bizzozero, N. I., and Calhoun, V. D. Components of cross-frequency modulation in health and disease. Frontiers in systems neuroscience, 5: 0 59, 2011

    work page 2011

  3. [3]

    K., Devijver, E., and Gaussier, E

    Assaad, C. K., Devijver, E., and Gaussier, E. Survey and evaluation of causal discovery methods for time series. Journal of Artificial Intelligence Research, 73: 0 767--819, 2022

    2022

  4. [4]

    Causal discovery from conditionally stationary time series

    Balsells-Rodas, C., Tu, R., Li, Y., and Kjellstrom, H. Causal discovery from conditionally stationary time series. In UAI 2022 Workshop on Causal Representation Learning, 2022

    work page 2022

  5. [5]

    Neural additive vector autoregression models for causal discovery in time series

    Bussmann, B., Nys, J., and Latr \'e , S. Neural additive vector autoregression models for causal discovery in time series. In Discovery Science: 24th International Conference, DS 2021, Halifax, NS, Canada, October 11--13, 2021, Proceedings 24, pp.\ 446--460. Springer, 2021

    work page 2021

  6. [6]

    D., Miller, R., Pearlson, G., and Adal , T

    Calhoun, V. D., Miller, R., Pearlson, G., and Adal , T. The chronnectome: time-varying connectivity networks as the next frontier in fmri data discovery. Neuron, 84 0 (2): 0 262--274, 2014

    work page 2014

  7. [7]

    Multi-region local field potential recordings during a tail-suspension test

    Carlson, D., Kumar, S., and Dzirasa, K. Multi-region local field potential recordings during a tail-suspension test. Duke Research Data Repository, 2023

    work page 2023

  8. [8]

    Dycast: learning dynamic causal structure from time series

    Cheng, Y., Lyu, B., Xing, W., and Zhu, Z. Dycast: learning dynamic causal structure from time series. 2025

    work page 2025

Show all 46 references
  1. [9]

    Multiscale causal structure learning

    D'Acunto, G., Di Lorenzo, P., and Barbarossa, S. Multiscale causal structure learning. arXiv preprint arXiv:2207.07908, 2022

    2022 arXiv

  2. [10]

    M., Kumar, S., Dancy, E

    Dzirasa, K., Coque, L., Sidor, M. M., Kumar, S., Dancy, E. A., Takahashi, J. S., McClung, C. A., and Nicolelis, M. A. Lithium ameliorates nucleus accumbens phase-signaling dysfunction in a genetic mouse model of mania. Journal of Neuroscience, 30 0 (48): 0 16314--16323, 2010

    2010

  3. [11]

    B., Jordan, M

    Fox, E., Sudderth, E. B., Jordan, M. I., and Willsky, A. S. Bayesian nonparametric inference of switching dynamic linear models. IEEE Transactions on signal processing, 59 0 (4): 0 1569--1585, 2011

    2011

  4. [12]

    Friston, K., Moran, R., and Seth, A. K. Analysing connectivity with granger causality and dynamic causal modelling. Current opinion in neurobiology, 23 0 (2): 0 175, 2013

    2013

  5. [13]

    J., Harrison, L., and Penny, W

    Friston, K. J., Harrison, L., and Penny, W. Dynamic causal modelling. Neuroimage, 19 0 (4): 0 1273--1302, 2003

    2003

  6. [14]

    Causal discovery for non-stationary non-linear time series data using just-in-time modeling

    Fujiwara, D., Koyama, K., Kiritoshi, K., Okawachi, T., Izumitani, T., and Shimizu, S. Causal discovery for non-stationary non-linear time series data using just-in-time modeling. In Conference on Causal Learning and Reasoning, pp.\ 880--894. PMLR, 2023

    2023

  7. [15]

    Directed spectrum measures improve latent network models of neural populations

    Gallagher, N., Dzirasa, K., and Carlson, D. Directed spectrum measures improve latent network models of neural populations. Advances in neural information processing systems, 34: 0 7421--7435, 2021

    2021

  8. [16]

    and Runge, J

    Gerhardus, A. and Runge, J. High-recall causal discovery for autocorrelated time series with latent confounders. Advances in Neural Information Processing Systems, 33: 0 12615--12625, 2020

    2020

  9. [17]

    Switching neural network systems for nonlinear tracking

    Ghimire, M. Switching neural network systems for nonlinear tracking. Master's thesis, Wright State University, 2018

    2018

  10. [18]

    Adjustment identification distance: A gadjid for causal structure learning

    Henckel, L., W \"u rtzen, T., and Weichwald, S. Adjustment identification distance: A gadjid for causal structure learning. In The 40th Conference on Uncertainty in Artificial Intelligence

  11. [19]

    Causal discovery and forecasting in nonstationary environments with state-space models

    Huang, B., Zhang, K., Gong, M., and Glymour, C. Causal discovery and forecasting in nonstationary environments with state-space models. In Chaudhuri, K. and Salakhutdinov, R. (eds.), Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings...

  12. [20]

    R., Sedler, A

    Keshtkaran, M. R., Sedler, A. R., Chowdhury, R. H., Tandon, R., Basrai, D., Nguyen, S. L., Sohn, H., Jazayeri, M., Miller, L. E., and Pandarinath, C. A large-scale neural network training framework for generalized estimation of single-trial population dynamics. Nature Methods,...

    2022

  13. [21]

    Kolda, T. G. and Bader, B. W. Tensor decompositions and applications. SIAM review, 51 0 (3): 0 455--500, 2009

    2009

  14. [22]

    Exploring and making sense of large graphs

    Koutra, D. Exploring and making sense of large graphs. Computer Science Department, Carnegie Mellon University, 2015

    2015

  15. [23]

    State space reconstruction parameters in the analysis of chaotic time series—the role of the time window length

    Kugiumtzis, D. State space reconstruction parameters in the analysis of chaotic time series—the role of the time window length. Physica D: Nonlinear Phenomena, 95 0 (1): 0 13--28, 1996

    1996

  16. [24]

    Bayesian learning and inference in recurrent switching linear dynamical systems

    Linderman, S., Johnson, M., Miller, A., Adams, R., Blei, D., and Paninski, L. Bayesian learning and inference in recurrent switching linear dynamical systems. In Artificial intelligence and statistics, pp.\ 914--922. PMLR, 2017

    2017

  17. [25]

    Amortized causal discovery: Learning to infer causal graphs from time-series data

    L \"o we, S., Madras, D., Zemel, R., and Welling, M. Amortized causal discovery: Learning to infer causal graphs from time-series data. In Conference on Causal Learning and Reasoning, pp.\ 509--525. PMLR, 2022

    2022

  18. [26]

    Maclaren, O. J. and Nicholson, R. What can be estimated? identifiability, estimability, causal inference and ill-posed inverse problems. arXiv preprint arXiv:1904.02826, 2019

    1904 arXiv

  19. [27]

    D., Talbot, A., Blount, C., Walder-Christensen, K

    Mague, S. D., Talbot, A., Blount, C., Walder-Christensen, K. K., Duffney, L. J., Adamson, E., Bey, A. L., Ndubuizu, N., Thomas, G. E., Hughes, D. N., et al. Brain-wide electrical dynamics encode individual appetitive social behavior. Neuron, 110 0 (10): 0 1728--1741, 2022

    2022

  20. [28]

    Generating realistic in silico gene networks for performance assessment of reverse engineering methods

    Marbach, D., Schaffter, T., Mattiussi, C., and Floreano, D. Generating realistic in silico gene networks for performance assessment of reverse engineering methods. Journal of computational biology, 16 0 (2): 0 229--239, 2009

    2009

  21. [29]

    Dynotears: Structure learning from time-series data

    Pamfil, R., Sriwattanaworachai, N., Desai, S., Pilgerstorfer, P., Georgatzis, K., Beaumont, P., and Aragam, B. Dynotears: Structure learning from time-series data. In International Conference on Artificial Intelligence and Statistics, pp.\ 1595--1605. PMLR, 2020

    2020

  22. [30]

    Causality

    Pearl, J. Causality. Cambridge university press, 2009

    2009

  23. [31]

    Discovering contemporaneous and lagged causal relations in autocorrelated nonlinear time series datasets

    Runge, J. Discovering contemporaneous and lagged causal relations in autocorrelated nonlinear time series datasets. In Conference on Uncertainty in Artificial Intelligence, pp.\ 1388--1397. Pmlr, 2020

    2020

  24. [32]

    Causal discovery in financial markets: A framework for nonstationary time-series data, 2024

    Sadeghi, A., Gopal, A., and Fesanghary, M. Causal discovery in financial markets: A framework for nonstationary time-series data, 2024

    2024

  25. [33]

    Reconstructing regime-dependent causal relationships from observational time series

    Saggioro, E., de Wiljes, J., Kretschmer, M., and Runge, J. Reconstructing regime-dependent causal relationships from observational time series. Chaos: An Interdisciplinary Journal of Nonlinear Science, 30 0 (11), 2020

    2020

  26. [34]

    Granger causality

    Seth, A. Granger causality. Scholarpedia, 2 0 (7): 0 1667, 2007

    2007

  27. [35]

    and Fox, E

    Shojaie, A. and Fox, E. B. Granger causality: A review and recent advances. Annual Review of Statistics and Its Application, 9 0 (1): 0 289--319, 2022

    2022

  28. [36]

    Eeg emotion recognition using dynamical graph convolutional neural networks

    Song, T., Zheng, W., Song, P., and Cui, Z. Eeg emotion recognition using dynamical graph convolutional neural networks. IEEE Transactions on Affective Computing, 11 0 (3): 0 532--541, 2018

    2018

  29. [37]

    E., Penny, W

    Stephan, K. E., Penny, W. D., Moran, R. J., den Ouden, H. E., Daunizeau, J., and Friston, K. J. Ten simple rules for dynamic causal modeling. Neuroimage, 49 0 (4): 0 3101--3102, 2010

    2010

  30. [38]

    Estimating a brain network predictive of stress and genotype with supervised autoencoders

    Talbot, A., Dunson, D., Dzirasa, K., and Carlson, D. Estimating a brain network predictive of stress and genotype with supervised autoencoders. Journal of the Royal Statistical Society Series C: Applied Statistics, 72 0 (4): 0 912--936, 05 2023. ISSN 0035-9254. doi:10.1093/jrs...

    2023 doi

  31. [39]

    Anxiety control by astrocytes in the lateral habenula

    Tan, W., Ikoma, Y., Takahashi, Y., Konno, A., Hirai, H., Hirase, H., and Matsui, K. Anxiety control by astrocytes in the lateral habenula. Neuroscience Research, 2024

    2024

  32. [40]

    Tank, A., Covert, I., Foti, N., Shojaie, A., and Fox, E. B. Neural granger causality. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44 0 (8): 0 4267--4279, 2021

    2021

  33. [41]

    E., Mogensen, P

    Weichwald, S., Jakobsen, M. E., Mogensen, P. B., Petersen, L., Thams, N., and Varando, G. Causal structure learning from time series: Large regression coefficients may predict causal links better in practice than small p-values . volume 123 of Proceedings of the NeurIPS 2019 C...

    2019

  34. [42]

    Dag-gnn: Dag structure learning with graph neural networks

    Yu, Y., Chen, J., Gao, T., and Yu, M. Dag-gnn: Dag structure learning with graph neural networks. In International conference on machine learning, pp.\ 7154--7163. PMLR, 2019

    2019

  35. [43]

    Causal discovery from nonstationary/heterogeneous data: Skeleton estimation and orientation determination

    Zhang, K., Huang, B., Zhang, J., Glymour, C., and Sch \"o lkopf, B. Causal discovery from nonstationary/heterogeneous data: Skeleton estimation and orientation determination. In IJCAI: Proceedings of the Conference, volume 2017, pp.\ 1347, 2017 a

    2017

  36. [44]

    Causal discovery from nonstationary/heterogeneous data: Skeleton estimation and orientation determination

    Zhang, K., Huang, B., Zhang, J., Glymour, C., and Sch \"o lkopf, B. Causal discovery from nonstationary/heterogeneous data: Skeleton estimation and orientation determination. In IJCAI: Proceedings of the Conference, volume 2017, pp.\ 1347. NIH Public Access, 2017 b

    2017

  37. [45]

    and Wang, J

    Zivot, E. and Wang, J. Vector autoregressive models for multivariate time series. Modeling financial time series with S-PLUS , pp.\ 385--429, 2006

    2006

  38. [46]

    and Feng, J

    Zou, C. and Feng, J. Granger causality vs. dynamic bayesian network inference: a comparative study. BMC bioinformatics, 10: 0 1--17, 2009

    2009

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.