acceptodds
Under review as a conference paper at ICLR 2027

Cyclic Structural Hawkes Processes for Learning Causal Structure from Discrete-Time Event Sequences

Abstract

Learning causal structure from discrete-time event sequences is particularly challenging because the exact temporal order of events within the same time interval is unobserved. As a result, causal relations that occur within one interval cannot be captured by methods that rely only on strictly lagged dependence. To tackle this issue, existing approaches such as Structural Hawkes Processes (SHPs) explicitly model causal interactions within the same interval, but require the corresponding causal structure to be acyclic to ensure a well-defined recursive generative process. Such an assumption can be restrictive, since events within one interval may repeatedly excite each other and naturally form directed cycles. In this work, we propose Cyclic Structural Hawkes Processes (C-SHPs) for learning cyclic causal structures from discrete-time event sequences. By connecting the observed event counts to an underlying continuous-time Hawkes process through the probability generating functional (PGFL), we develop theoretical tools for characterizing self-loops, adjacency, and edge directions from temporally aggregated observations. We further establish the identifiability of the directed causal structure in the presence of self-loops, reciprocal edges, and arbitrary directed cycles. Based on these results, we develop a practical causal discovery algorithm that operates directly on discrete-time event counts. Experiments on cyclic graphs demonstrate robust performance across varying time resolutions, model parameters, sample sizes, graph dimensions, and graph densities.

Primary area:
causal reasoning
Keywords:
Causal Discovery, Hawkes Processes

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_L8oTOpenReview ↗
Rating
3
Confidence
1
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper addresses causal structure learning from event counts aggregated into time intervals. For example, given only the number of alarms of each type in each interval, the goal is to determine whether one alarm type triggers another, whether their influence is reciprocal, and whether an alarm type triggers further events of the same type. Building on a continuous-time Hawkes process, the authors establish that the complete count distribution identifies the underlying connection structure, including self-excitation and directed cycles, under the stated assumptions. They also propose C-SHP, a practical algorithm combining approximate computations with comparisons among candidate graph structures. Experiments on synthetic data and a selected real-world alarm dataset report higher directed edge recovery performance than the compared methods.

Critical strengths
  1. The paper addresses a fundamental question of whether causal connections can still be distinguished after exact event times are lost. Under its assumptions, it establishes that the complete count distribution identifies connections involving self-excitation and directed cycles. The analysis explicitly addresses effects carried over from earlier events and the ambiguity that self-excitation can introduce when determining connection directions.
  2. The practical algorithm combines several checks when selecting connections. For each candidate graph, it adjusts the baseline event rates and connection strengths, then compares how well the graph explains the data while penalizing additional connections. It also uses several patterns in the event counts to assess direction and retains existing two-way connections when the evidence is inconclusive. This design accounts for uncertainty when applying theory-inspired criteria to a limited amount of observed data.
Critical weaknesses

Self-excitation is discussed extensively in the theoretical analysis, and the proposed algorithm explicitly includes a step for detecting self-loops. However, self-loop recovery is not directly evaluated in the reported experiments. The synthetic experiments report edge recovery scores excluding self-loops, while the real-world reference graph contains no self-loops. The experimental results therefore do not show how accurately the proposed algorithm identifies self-loops.

Questions for authors
  1. Why was self-loop recovery excluded from the reported evaluation? Since the synthetic graphs contain self-loops and the algorithm includes a step for detecting them, could you report self-loop recovery performance from the existing experiments, if available?
  2. While Assumption 1 defines the theoretical setting, could you discuss the main challenges in applying C-SHP to more complex real-world event datasets than the one evaluated here? In particular, what should practitioners consider when the data may not fully satisfy these assumptions, and which aspects of the method might need to be adapted?
4 comments

W1 & Q1: Why was self-loop recovery excluded from the reported evaluation? Since the synthetic graphs contain self-loops and the algorithm includes a step for detecting them, could you report self-loop recovery performance from the existing experiments, if available?

Thank you for pointing out the need to explicitly report the recovery of self-loops.

The original Figure 2 reported off-diagonal directed F1, which excluded self-loops from that metric. We now report F1 including self-loops for all six original sensitivity experiments in Appendix D.2 (Figure 5). For the original default 20-seed datasets, C-SHP achieved overall F1 of 0.9144 ± 0.0465 and self-loop F1 of 1.0000 ± 0.0000 (Appendix D.3). Self-loop recovery was therefore accurate in the existing default experiments as well.

To more comprehensively evaluate the ability to recover self-loops, we conducted a supplementary synthetic experiment in which the proportion of self-loops was increased from 20% to 40%. Each graph contains 20 event types and 8 self-loops, in addition to bidirectional edges and longer directed cycles. The decay rate used during the data generation process was set to 0.35, while the decay rate used by C-SHP during model fitting was 0.175. The results are reported as the mean ± standard deviation across 20 random seeds.

MethodOverall F1 (including self-loops)Self-loop F1
C-SHP (Ours)0.9739 ± 0.02351.0000 ± 0.0000
SHP0.6646 ± 0.03480.0000 ± 0.0000
MLE-SGL0.6581 ± 0.09310.9797 ± 0.0336
NPHC0.3149 ± 0.12420.6372 ± 0.1148
ADM40.5814 ± 0.02980.9477 ± 0.0312

In every graph evaluated, C-SHP accurately recovered all 8 self-loops without producing any false-positive self-loops, while maintaining an overall F1 score of 0.9739. These results demonstrate that, under the tested settings—even when the decay rate used for model fitting differs from the one used to generate the data—the method can still accurately recover self-loops.

We have added F1 results including self-loops for all six original sensitivity experiments in Appendix D.2 (Figure 5). Appendix D.3 separately reports self-loop F1 for the original default 20-seed setting. The supplementary enriched-cycle results appear in Section 5 (Figure 3(c)), with their generation protocol in Appendix D.3 and the evaluation definitions in Appendix D.1.

Q2: While Assumption 1 defines the theoretical setting, could you discuss the main challenges in applying C-SHP to more complex real-world event datasets than the one evaluated here? In particular, what should practitioners consider when the data may not fully satisfy these assumptions, and which aspects of the method might need to be adapted?

We thank the reviewer for this valuable suggestion. We agree that applying C-SHP to more complex real-world event sequences may present additional challenges beyond the theoretical setting considered in our work. We have clarified these practical considerations in Appendix D.8.

First, latent confounding and incomplete event observations. Assumption 1 requires causal sufficiency, whereas real-world event logs may omit relevant event types or common external drivers. Such unobserved factors may induce dependencies that are difficult to distinguish from direct excitation effects. Practitioners should therefore assess whether relevant event types or common external drivers are omitted.

Second, model misspecification and temporal heterogeneity. C-SHP assumes nonnegative excitation with shared-decay exponential kernels and time-invariant parameters. In practice, event dynamics may exhibit heterogeneous decay rates, longer-range temporal dependencies, or time-varying background intensities. Such departures may influence the accuracy of the discrete-time fitting procedure and the recovered structure. Extensions to more flexible excitation kernels or time-varying intensities are possible directions for future work, although their corresponding PGFL representations and identifiability properties would require further theoretical investigation. Kernel shape and changes in background intensities should therefore be assessed when applying the method. The reported robustness tests assess specific departures from these assumptions.

Third, observation quality and scalability. Real-world event logs may contain missing events, irregular observation periods, or substantial differences in event frequencies across types. These characteristics may influence the estimation of historical excitation states and PGFL-based statistics. Careful preprocessing, distinguishing missing observations from genuine zero-count intervals, and assessing the stability of the inferred structure are therefore important in practice. For higher-dimensional event systems, candidate-edge screening and sparsity-aware search strategies may improve computational efficiency, while preserving sufficient coverage of potentially relevant interactions. These considerations call for careful preprocessing and checking stability across decay and threshold choices.

We have added the practical considerations in Appendix D.8 ("Practical scope"). Appendix C.1 explains the observation-window and exposure approximations, Appendix C.6 discusses search cost and computational limits, and Appendix D.4 reports the kernel and hidden-parent tests relevant to these considerations.

Thank you for the detailed response and revisions. The added self-loop recovery results address my main concern, and the discussion of practical limitations answers my second question. I view these additions positively. However, as noted in my review, my expertise in this area is limited, and I am not confident enough in my independent assessment of the theoretical contributions to support a stronger recommendation. I therefore maintain my rating.

We sincerely thank you for your thoughtful feedback and for recognizing the improvements made in our revision. We greatly appreciate your positive assessment of the additional results and clarifications, as well as your careful and constructive review. We fully respect your assessment and are grateful for the time and effort you have devoted to evaluating our work. Your feedback has been invaluable in helping us strengthen our paper.

Reviewer_ye9gOpenReview ↗
Rating
2
Confidence
4(+1)
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Cyclic Structural Hawkes Processes (C-SHP) for causal structure discovery from temporally aggregated event sequences. By leveraging the probability generating functional (PGFL), the authors establish theoretical identifiability results for directed causal structures containing self-loops, reciprocal edges, and arbitrary directed cycles.

The proposed algorithm combines discrete-time graph fitting, PGFL-based structural scores, and factorial cumulants. Experiments on synthetic datasets and a real-world alarm dataset demonstrate improvements over existing Hawkes-based causal discovery methods.

Critical strengths
  1. Interesting problem formulation. The paper addresses causal discovery from low-resolution event sequences with cyclic dependencies, which is an important and challenging problem.

  2. Meaningful theoretical contribution. The PGFL-based analysis provides identifiability results for causal structures containing self-loops and directed cycles.

  3. Promising experimental performance. The proposed method consistently outperforms several existing baselines across different synthetic settings and achieves strong results on the real-world alarm dataset.

Critical weaknesses
  1. Gap between theory and algorithm. The identifiability results rely on the exact population PGFL, while the practical algorithm uses approximate estimators and heuristic graph search. The theoretical guarantees do not directly extend to the proposed estimator.

  2. Restrictive model assumptions. The theoretical analysis assumes correctly specified Hawkes processes with shared-decay exponential kernels, positive baselines, and causal sufficiency, limiting its general applicability.

  3. Insufficient evaluation of cyclic structures. Although cyclic graph recovery is the main contribution, the experiments primarily report overall directed F1 without separately analyzing the recovery of different cycle types and reciprocal edges.

  4. Incomplete evaluation of self-loops. The synthetic experiments report off-diagonal directed F1, excluding self-loops despite their importance in the theoretical contribution.

  5. Limited real-world validation. The real-world evaluation considers only five selected alarm types from a single dataset, making it difficult to assess generalizability to larger and more complex event systems.

  6. Lack of component ablation studies. The paper does not sufficiently isolate the contributions of PGFL-based structural scores, factorial cumulants, and the graph-search procedure.

  7. Unclear computational efficiency and parameter sensitivity. The algorithm involves repeated graph fitting, bootstrap estimation, and multiple calibration parameters, but computational costs and sensitivity to these choices are not adequately evaluated.

Questions for authors
  1. Theory-practice gap: How do the theoretical identifiability guarantees relate to the finite-sample approximations used in the practical algorithm? Can any consistency guarantees be established?

  2. Cycle recovery: Could the authors provide separate evaluations for self-loops, reciprocal edges, and longer directed cycles rather than reporting only overall directed F1?

  3. Model assumptions: How robust is C-SHP when the underlying event process violates the shared-decay exponential kernel assumption or contains unobserved confounders?

  4. Component effectiveness: What are the individual contributions of the PGFL-based scores, factorial cumulants, and graph-search procedure? Could the authors provide additional ablation experiments?

  5. Efficiency and scalability: How does the computational cost of C-SHP compare with existing baselines as the number of event types increases? How sensitive are the results to the decay parameter and calibration thresholds?

7 comments

W1 & Q1: Theory-practice gap: How do the theoretical identifiability guarantees relate to the finite-sample approximations used in the practical algorithm? Can any consistency guarantees be established?

Thanks for asking us to clarify how the theoretical results relate to the practical algorithm and its consistency.

Theoretical identifiability. Under the stated model assumptions, Theorem 2 provides the exact count-PGFL and ODE representation for shared-decay exponential kernels. Theorem 3 characterizes local structural signatures, and Theorem 4 establishes the identifiability of the directed causal structure, including self-loops, reciprocal edges, and longer directed cycles, for any fixed observation width Δ>0\Delta>0, without requiring Δ→0\Delta\to0. To make these results and their connections easier to follow, we have revised the wording and presentation of Theorems 2 and 4 to improve clarity and readability.

Practical algorithm. C-SHP uses these results through PGFL-based scores, factorial cumulants, an approximation of the history state XkX_k computed from window counts, and graph search. Specifically, PGFL-based scores support self-loop and edge decisions, while the empirical factorial cumulants K^ab(2)\widehat{K}^{(2)}_{ab} and K^abb(3)\widehat{K}^{(3)}_{abb} are used for candidate prioritization and direction proposals, respectively. PGFL comparisons are evaluated at finite values of qq, and the initial graph search compares refitted candidate graphs using BIC. Bootstrap resampling of independent sequences assesses the variability of the estimated statistics, while candidate graphs are checked for stability before acceptance. Section 4 and Appendices C.1–C.5 explain how these quantities are computed and how candidate graphs are accepted. C-SHP therefore uses the structural information through quantities computed from the observed window counts.

Consistency. Appendix C.6 gives sufficient conditions for consistent graph selection based on a common scoring criterion. Consider a fixed, finite set of candidate graphs containing the true graph G0G_0. Assume that the population score of G0G_0 exceeds that of every incorrect candidate by at least γ>0\gamma>0. This positive score difference is an additional condition for the consistency argument. For RR independent sequences, let εR\varepsilon_R bound the difference between the estimated and population scores across candidate graphs, including estimation and approximation errors. Let ηR≥0\eta_R\geq0 bound the difference between the best estimated score and the estimated score of the graph returned by the search. Then Pr⁡(G^R≠G0)≤Pr⁡(2εR+ηR≥γ)\Pr(\widehat{G}_R\ne G_0)\leq\Pr(2\varepsilon_R+\eta_R\geq\gamma).

Therefore, if these errors converge to zero in probability, the probability of recovering the true graph approaches one.

We have revised the wording and presentation of Theorems 2 and 4 in Sections 3.1 and 3.3 to improve clarity. The consistency argument and its conditions are stated in Appendix C.6 (Eq. (C.17)).

W3 & W4 & Q2: Cycle recovery: Could the authors provide separate evaluations for self-loops, reciprocal edges, and longer directed cycles rather than reporting only overall directed F1?

Thanks for suggesting separate evaluations of different cycle structures.

For the original default 20-seed datasets, C-SHP achieves self-loop F1 of 1.0000 ± 0.0000 and overall F1 of 0.9144 ± 0.0465 (Appendix D.3). Appendix D.2 (Figure 5) also reports F1 including self-loops for all six original sensitivity experiments. These results directly address the self-loops omitted from the original reporting metric.

To ensure that each cycle type is sufficiently represented, we added a synthetic experiment with a higher proportion of self-loops and explicit coverage of reciprocal edges and longer directed cycles. The self-loop rate was increased from 20% to 40%. Each graph contained 20 event types, 30 non-self-loop edges, eight self-loops, two reciprocal pairs, and one directed cycle of each length from three to five. In this supplementary setting, the generating decay rate was 0.35, while C-SHP, SHP, and ADM4 used a fixed fitting decay rate of 0.175. All methods used the same observations, with observation width Δ=5\Delta=5 and 40 independent sequences of 500 events each. We report overall F1 including self-loops, diagonal F1 for self-loops, and F1 for the sets of edges belonging to cycles of each specified length in the predicted and true graphs. Additional predicted cycle edges count as false positives. The following results are mean ± standard deviation across 20 seeds.

MethodOverall F1Self-loop F1Reciprocal-edge F13-cycle edge F14-cycle edge F15-cycle edge F1
C-SHP (Ours)0.9739 ± 0.02351.0000 ± 0.00000.6833 ± 0.31450.8667 ± 0.25131.0000 ± 0.00000.9833 ± 0.0745
SHP0.6646 ± 0.03480.0000 ± 0.00000.0000 ± 0.00000.0000 ± 0.00000.0000 ± 0.00000.0000 ± 0.0000
MLE-SGL0.6581 ± 0.09310.9797 ± 0.03360.4302 ± 0.16130.2425 ± 0.13090.3200 ± 0.19130.5574 ± 0.4132
NPHC0.3149 ± 0.12420.6372 ± 0.11480.0372 ± 0.06790.0658 ± 0.14640.0588 ± 0.08850.0409 ± 0.0326
ADM40.5814 ± 0.02980.9477 ± 0.03120.3379 ± 0.05170.2038 ± 0.01610.3693 ± 0.05920.6179 ± 0.0922

C-SHP achieved the highest mean F1 in every reported category in this controlled setting. Its overall F1 was 0.9739, with accurate self-loop recovery and high scores for three-, four-, and five-node cycles. Reciprocal-edge recovery was more variable across graphs, but its mean F1 remained higher than those of the evaluated baselines. These results show that C-SHP can recover self-loops, reciprocal edges (two-node cycles), and longer directed cycles under the tested settings.

We have added these separate recovery results in Section 5 ("Cycle recovery"; Figure 3(c)). Appendix D.1 (Eq. (D.1)) defines cycle-edge F1, Appendix D.2 (Figure 5) reports the original sensitivity results including self-loops, and Appendix D.3 gives the original default self-loop results and specifies the enriched-cycle graph generation and decay rates.

W2 & Q3: Model assumptions: How robust is C-SHP when the underlying event process violates the shared-decay exponential kernel assumption or contains unobserved confounders?

Thanks for asking about C-SHP’s robustness to violations of the kernel assumption and the presence of unobserved confounders.

We conducted additional synthetic experiments covering both cases, using 20 random seeds per condition. C-SHP retained its shared single-exponential kernel with a fixed fitting decay rate of 1 throughout these experiments.

Role of the assumptions. Shared exponential decay gives the exact finite-dimensional ODE representation in Section 3.1. Positive baseline intensities allow every type to generate immigrant events. With empty initial history, a type with zero baseline and no incoming excitation would never appear, preventing recovery of its outgoing effects from observations. Causal sufficiency supports interpreting the recovered excitation graph among the observed types. The following experiments assess specified departures from the kernel assumption and causal sufficiency, while the theoretical guarantees retain the assumptions in Section 2.

Kernel misspecification. We evaluated four alternative kernel shapes, each normalized to integrate to one:

  • Heterogeneous exponential kernels: guv(t)=βuve−βuvtg_{uv}(t)=\beta_{uv}e^{-\beta_{uv}t}, with βuv∼U(0.5,2)\beta_{uv}\sim\mathcal U(0.5,2).
  • Two-exponential mixture: g(t)=0.5(0.5e−0.5t)+0.5(2e−2t)g(t)=0.5(0.5e^{-0.5t})+0.5(2e^{-2t}).
  • Power-law kernel: g(t)=2/(1+t)3g(t)=2/(1+t)^3.
  • Gamma kernel: g(t)=4te−2tg(t)=4te^{-2t}, with shape 2 and rate 2. C-SHP continued to fit a shared single-exponential kernel with βfit=1\beta_{\mathrm{fit}}=1 throughout these experiments.

Unobserved confounders. We added one hidden common parent hh that directly excited two of the 20 observed event types. Its outgoing kernels were ϕhv(t)=cBhve−t\phi_{hv}(t)=cB_{hv}e^{-t}, where the base weights BhvB_{hv} were sampled from U(0.1,0.2)\mathcal U(0.1,0.2) and c∈{0.5,1,2}c\in\{0.5,1,2\} controlled the confounding strength. The hidden parent's events were excluded from the observations supplied to every method. Recovery was evaluated against the true direct graph among the observed event types. The following table reports overall F1, including self-loops, as mean ± standard deviation across 20 random seeds per condition.

Generating conditionC-SHP F1
Shared exponential kernel0.9141 ± 0.0346
Heterogeneous exponential kernels0.8952 ± 0.0432
Two-exponential mixture0.8974 ± 0.0499
Power-law kernel0.8889 ± 0.0570
Gamma kernel0.9144 ± 0.0259
Hidden common parent, strength multiplier 0.50.9067 ± 0.0294
Hidden common parent, strength multiplier 10.9147 ± 0.0401
Hidden common parent, strength multiplier 20.9142 ± 0.0304

For the kernel experiments, C-SHP maintained mean F1 between 0.8889 and 0.9144. Gamma-kernel performance was close to the exponential control, while the power-law kernel produced a decrease of approximately 2.5 percentage points. Under the three independently generated hidden-parent settings, mean F1 remained between 0.9067 and 0.9147, above 0.90 in every setting. These results show that C-SHP maintains good structure recovery under the tested conditions where the underlying event process violates the shared-decay exponential kernel assumption or contains an unobserved confounder.

We have added the kernel-misspecification results in Section 5 ("Non-exponential kernels"; Figure 3(b)). Appendix D.4 (Eq. (D.2) and Figure 6) specifies the generating kernels and hidden-parent construction and reports the additional robustness results. The identifiability assumptions are stated in Section 2 (Assumption 1), and Appendix D.4 distinguishes these assumptions from the empirical robustness tests.

Reviewer_BW5bOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a novel algorithm for learning causal structures from discrete-time event sequences. Unlike existing Structural Hawkes Processes, which require acyclic causal graphs, the proposed framework models the underlying event generation using a continuous-time Hawkes process that allows directed cycles. The paper derives a probability generating functional representation of the observed event counts and establishes identifiability of the causal graph, including self-loops, reciprocal edges, and arbitrary directed cycles. Based on the theoretical results, the authors propose a causal discovery algorithm. Experiments on synthetic and real-world datasets show that C-SHP outperforms the baselines.

Critical strengths
  1. The paper studies an important and practical problem: learning causal structures from discrete-time event sequences when the exact temporal ordering of events within each interval is unavailable.
  2. The proposed framework removes the acyclicity assumption of the original SHP by modeling the underlying event generation using a continuous-time Hawkes process.
  3. The paper establishes a connection between the continuous-time Hawkes process and the distribution of discrete-time event counts through PGFL. It also proves identifiability of the directed causal graph in the presence of self-loops, reciprocal edges, and arbitrary directed cycles.
  4. The experiments consider both synthetic and real-world datasets. The synthetic experiments vary multiple factors. C-SHP outperforms other baselines across the considered settings.
Critical weaknesses
  1. The main weakness is the gap between the theoretical results and the proposed algorithm. Theorem 4 establishes identifiability based on the population count distribution and limiting behavior as q→−∞q \to -\infty. However, C-SHP relies on finite-sample approximations and decision rules that are not directly covered by the theorems. As acknowledged in Appendix C.4, these approximations are not guaranteed to preserve the population identities underlying the theoretical results. No consistency guarantee is provided. Figure 2(d) does not show a clear trend toward perfect recovery as the sample size increases, even though the data are generated from the assumed model. Can the authors clarify whether C-SHP is consistent?
  2. The proposed algorithm involves many tuning parameters, but their values are not reported in the experiments. The decay parameter used in the synthetic experiments is also unclear.
  3. The experimental evaluation is not comprehensive. The synthetic data are generated from the same Hawkes model assumed by C-SHP, which may favor the proposed method.
  4. Although the paper emphasizes the identifiability of self-loops, reciprocal edges, and directed cycles, the experiments only report off-diagonal directed F1. The recovery of self-loops and reciprocal edges is not separately evaluated.
Questions for authors
  1. What decay parameter and tuning parameter values were used in the synthetic experiments? Was the decay parameter set to its true generating value? How sensitive is C-SHP to an incorrectly specified decay parameter or non-exponential kernels?
  2. Since self-loops and reciprocal edges are central to the theoretical contributions, can the authors report their recovery accuracy separately? It would also be helpful to report the computational runtime of C-SHP and the baselines.
  3. How were the continuous-time baselines, such as NPHC, MLE-SGL, and ADM4, applied to discrete-time event counts? Can the authors explain why NPHC and MLE-SGL show relatively low and nearly constant F1 scores across several experimental settings?
  4. The real-world evaluation only considers five selected alarm types with nine reference edges and no self-loops. How representative is this setting? Can the authors explain why SHP obtains an F1 score of zero and evaluate the sensitivity of C-SHP to the choice of the decay parameter?
7 comments

W1: The gap between population identifiability and the practical algorithm, consistency.

Thanks for asking us to clarify the relationship between identifiability, the practical algorithm, and the sample-size results.

Under the stated model assumptions, Theorem 4 establishes the identifiability of the directed causal structure for any fixed observation width Δ>0\Delta>0. Theorem 2 provides the exact count-PGFL and ODE representation for shared-decay exponential kernels, and Theorem 3 characterizes local structural signatures. C-SHP uses these results through PGFL-based scores, factorial cumulants, an approximation of the history state XkX_k computed from window counts, PGFL comparisons at finite values of qq, and graph search. Section 4 and Appendices C.1–C.5 explain how these quantities are computed and how candidate graphs are accepted.

Appendix C.6 gives sufficient conditions for consistent graph selection based on a common scoring criterion. Consider a fixed, finite set of candidate graphs containing the true graph G0G_0. Assume that the population score of G0G_0 exceeds that of every incorrect candidate by at least γ>0\gamma>0. For RR independent sequences, let εR\varepsilon_R bound the difference between the estimated and population scores across candidate graphs, including estimation and approximation errors. Let ηR\eta_R bound the difference between the best estimated score and the estimated score of the graph returned by the search. Then Pr⁡(G^R≠G0)≤Pr⁡(2εR+ηR≥γ)\Pr(\widehat{G}_R\ne G_0)\leq\Pr(2\varepsilon_R+\eta_R\geq\gamma).

Therefore, if these errors converge to zero in probability, the probability of recovering the true graph approaches one. Figure 2(d) varies the total number of events while keeping the other settings fixed, including the generating decay rate of 5 and fitting decay rate of 0.35. The number of independent sequences remains fixed. F1 can fluctuate across the tested sample sizes because of sampling variability and changes in the selected graph. This experiment evaluates finite-sample structure recovery under the reported settings.

We have detailed the finite-sample approximations in Section 4 and Appendices C.1–C.5, and stated the consistency argument and its conditions in Appendix C.6 (Eq. (C.17)). The sample-size results are reported in Figure 2(d), with F1 including self-loops in Appendix D.2 (Figure 5(d)).

W2 & W3 & Q1: What decay parameter and tuning parameter values were used in the synthetic experiments? Was the decay parameter set to its true generating value? How sensitive is C-SHP to an incorrectly specified decay parameter or non-exponential kernels?

Thanks for raising these questions about the parameter settings and C-SHP’s sensitivity to decay misspecification and non-exponential kernels.

Parameter Settings and Sensitivity Analysis. In the synthetic experiments, the generating decay rate was βtrue=5\beta_{\mathrm{true}}=5, while C-SHP used a fixed fitting decay rate of βfit=0.35\beta_{\mathrm{fit}}=0.35. Thus, the fitting decay rate was not set to the true generating value. The main tuning parameters were the BIC weight ωBIC=1\omega_{\mathrm{BIC}}=1, self-loop thresholds (τweak,τstrong)=(0.08,0.10)(\tau_{\mathrm{weak}},\tau_{\mathrm{strong}})=(0.08,0.10), negative-axis coordinate qH=−3000q_H=-3000, and regularization weight λH=0.005\lambda_H=0.005. We retained this default configuration because our sensitivity analyses showed stable structure recovery across a range of fitting decay rates and calibration thresholds. For example, we evaluated sensitivity to decay misspecification using 20 random seeds, with βtrue=5\beta_{\mathrm{true}}=5, Δ=5\Delta=5, and 40 independent sequences of 500 events per dataset. We varied only the fitting decay rate, using βfit/βtrue∈{0.07,0.25,0.5,1,2,4}\beta_{\mathrm{fit}}/\beta_{\mathrm{true}}\in\{0.07,0.25,0.5,1,2,4\}, where the ratio 0.070.07 corresponds to the default fitting value 0.350.35. All other model settings and the independently calibrated baseline support thresholds were held fixed in this sweep. The following table reports F1, including self-loops, as mean ± standard deviation across the 20 seeds.

Method0.070.250.5124
C-SHP (Ours)0.9144 ± 0.04650.9081 ± 0.05130.9056 ± 0.05410.9056 ± 0.05440.9041 ± 0.05090.9041 ± 0.0509
SHP0.5567 ± 0.08510.4988 ± 0.09020.4373 ± 0.09250.4229 ± 0.08740.4048 ± 0.08150.3866 ± 0.0787
ADM40.5099 ± 0.07360.0057 ± 0.01760.0000 ± 0.00000.0000 ± 0.00000.0000 ± 0.00000.0000 ± 0.0000

C-SHP’s mean F1 remained between 0.9041 and 0.9144 across the six tested decay ratios, compared with 0.9056 under the correctly specified rate. Fourfold underestimation and overestimation yielded F1 of 0.9081 and 0.9041, respectively. C-SHP retained higher mean F1 than SHP and ADM4 at every ratio, indicating low sensitivity under the tested settings.

Experiments with Non-Exponential Kernels. We also evaluated Gamma and power-law generating kernels in a separate paired synthetic experiment, while continuing to fit a shared exponential kernel with βfit=1\beta_{\mathrm{fit}}=1. The normalized generating kernels were gExp(t)=e−tg_{\mathrm{Exp}}(t)=e^{-t}, gGamma(t)=4te−2tg_{\mathrm{Gamma}}(t)=4te^{-2t}, and gPL(t)=2/(1+t)3g_{\mathrm{PL}}(t)=2/(1+t)^3, for t≥0t\geq0. All three have unit mass and mean delay 1. Within each of 20 paired model replicates, the graph, integrated kernel weights, and baseline intensities were held fixed across generating kernels. This kernel experiment used a separate set of datasets and its own exponential control. The absolute-weight thresholds selected on independent default-setting validation data were transferred unchanged to these kernel datasets, without condition-specific recalibration (Appendices D.1 and D.4).

MethodgExp(t)=e−tg_{\mathrm{Exp}}(t)=e^{-t}gPL(t)=2/(1+t)3g_{\mathrm{PL}}(t)=2/(1+t)^3gGamma(t)=4te−2tg_{\mathrm{Gamma}}(t)=4te^{-2t}
C-SHP (Ours)0.9141 ± 0.03460.8889 ± 0.05700.9144 ± 0.0259
SHP0.6752 ± 0.05700.6778 ± 0.09340.6676 ± 0.0441
MLE-SGL0.1596 ± 0.00120.1590 ± 0.00080.1598 ± 0.0008
NPHC0.1587 ± 0.01790.1691 ± 0.01540.1679 ± 0.0149
ADM40.0000 ± 0.00000.0000 ± 0.00000.0000 ± 0.0000

C-SHP achieved F1 = 0.9144 under the Gamma kernel, close to its matched exponential control (0.9141). Under the power-law kernel, F1 was 0.8889, a decrease of 2.52 percentage points from that control. C-SHP retained the highest mean F1 among the evaluated methods for all three generating kernels. These results support robustness under the tested kernel misspecifications.

We have revised Section 5 ("Synthetic experiments") and Appendix C.6 to report the generating and fitting decay rates and the main tuning parameters. Decay sensitivity is reported in Figure 3(a), with diagnostics in Appendix D.3. Non-exponential-kernel results are reported in Figure 3(b), with kernel definitions, the exponential control, and additional robustness results in Appendix D.4 (Eq. (D.2) and Figure 6).

W4 & Q2: Since self-loops and reciprocal edges are central to the theoretical contributions, can the authors report their recovery accuracy separately? It would also be helpful to report the computational runtime of C-SHP and the baselines.

Thanks for requesting separate recovery results for self-loops and reciprocal edges, together with runtime comparisons.

To ensure sufficient representation of different cycle structures, we conducted a supplementary synthetic experiment with the self-loop rate increased from 20% to 40%. Each graph contained 20 event types, 30 non-self-loop edges, eight self-loops, two reciprocal pairs, and one directed cycle of each length from three to five. Each dataset contained 40 independent sequences of 500 events. The generating decay rate was 0.35, while C-SHP, SHP, and ADM4 used a fitting decay rate of 0.175.We report the overall F1 score—including self-loops—as well as the F1 score for self-loops specifically. We also provide F1 scores for bidirectional edges and cycles of length three to five, along with the average runtime. All results are presented as the mean ± standard deviation across 20 random seeds.

MethodOverall F1Self-loop F1Reciprocal-edge F13-cycle edge F14-cycle edge F15-cycle edge F1Runtime (seconds)
C-SHP (Ours)0.9739 ± 0.02351.0000 ± 0.00000.6833 ± 0.31450.8667 ± 0.25131.0000 ± 0.00000.9833 ± 0.074535.64 ± 4.63
SHP0.6646 ± 0.03480.0000 ± 0.00000.0000 ± 0.00000.0000 ± 0.00000.0000 ± 0.00000.0000 ± 0.000010.19 ± 1.09
MLE-SGL0.6581 ± 0.09310.9797 ± 0.03360.4302 ± 0.16130.2425 ± 0.13090.3200 ± 0.19130.5574 ± 0.41320.60 ± 0.06
NPHC0.3149 ± 0.12420.6372 ± 0.11480.0372 ± 0.06790.0658 ± 0.14640.0588 ± 0.08850.0409 ± 0.03264.61 ± 0.20
ADM40.5814 ± 0.02980.9477 ± 0.03120.3379 ± 0.05170.2038 ± 0.01610.3693 ± 0.05920.6179 ± 0.09220.17 ± 0.02

C-SHP achieved the highest mean F1 in every reported structural category. It recovered all eight self-loops in every evaluated graph and obtained high recovery scores for longer directed cycles. Reciprocal-edge recovery showed greater variation across seeds while retaining the highest mean F1 among the evaluated methods.These results show that C-SHP can recover self-loops, reciprocal edges (two-node cycles), and longer directed cycles under the tested settings.

In this synthetic setting with more directed cycles, SHP restricts its search to DAGs and rejects cyclic candidate graphs before parameter fitting, which can reduce candidate-evaluation costs. C-SHP allows self-loops and directed cycles and additionally computes PGFL-based scores, factorial cumulants, and bootstrap-based structural criteria. These differences may contribute to SHP's shorter runtime in this experiment, while C-SHP achieves a higher mean F1 of 0.9739 compared with 0.6646 for SHP.

We have added separate self-loop, reciprocal-edge, and longer-cycle results in Section 5 ("Cycle recovery"; Figure 3(c)). The cycle-edge F1 definition is given in Appendix D.1 (Eq. (D.1)), and the graph-generation protocol and same-data runtime comparison are reported in Appendix D.3 (Table 3).

Comments on the paper

3 comments
Area_Chair_qHf2OpenReview ↗

Please join the author-reviewer discussion

Dear the authors and reviewers,

We noticed the authors did not provide responses. Our author and reviewer discussion will end soon in 18 November (AOE). If you still have some issues to discuss, this may still be a good time to do it. We highly appreciate your time on discussing and exchanging thoughts to keep our ICLR standards. Many thanks!

Bests,
Your AC

Thank you for the careful reviews and constructive suggestions. We have addressed the comments and revised the manuscript as summarized below.

Theorem presentation and consistency We revised the presentation of Theorems 2 and 4, introduced the required notation before their statements, and clarified the progression from the exact count-PGFL and ODE representation in Theorem 2, through the local structural signatures in Theorem 3, to global identifiability in Theorem 4. The mathematical conclusions are unchanged.

Decay settings and sensitivity The generating decay was 5, while C-SHP used a fitting decay of 0.35. Across the tested ratios of fitting decay to generating decay from 0.07 to 4, mean F1 remains between 0.9041 and 0.9144 on the original 20 default datasets. These results indicate low sensitivity under the tested settings. We have reported the decay rates and main tuning parameters in Section 5 and Appendix C.6, and added the sensitivity results in Figure 3(a) and Appendix D.3.

Kernel misspecification and hidden confounding We added heterogeneous exponential, two-exponential mixture, power-law, Gamma, and hidden common parent experiments. C-SHP retains mean F1 between 0.8889 and 0.9144 across the tested kernel conditions and between 0.9067 and 0.9147 under the tested hidden-parent conditions. These experiments support empirical robustness to the specified departures. We have added the results and generating-kernel definitions in Section 5 and Appendix D.4, while clarifying that the identifiability guarantees retain their stated assumptions.

Self-loop and cycle recovery We now report F1 including self-loops for all six original sensitivity experiments in Appendix D.2. We also generated 20 supplementary datasets with more self-loops and explicit coverage of reciprocal edges and three- to five-node cycles. C-SHP achieves overall F1 of 0.9739, self-loop F1 of 1.0000, and cycle-edge F1 of 0.6833, 0.8667, 1.0000, and 0.9833 for lengths two through five, respectively. It has the highest mean F1 in every reported structural category. We have added these results in Figure 3(c), with metric definitions and generation details in Appendices D.1 and D.3.

Continuous-time baselines and support thresholds The baselines receive count-preserving quantized timestamps without access to the original within-window event times. The nearly constant NPHC and MLE-SGL scores are explained by complete off-diagonal predictions under the original zero-threshold rule. We added thresholds selected on 10 independent validation graphs and fixed before testing. Calibration improves the baselines, while C-SHP retains a higher mean F1 on the same 20 test datasets. We have clarified the input representation and calibration procedure in Appendix D.1 and added the comparison in Figure 4.

Component effectiveness Paired ablations on the original 20 default datasets show that removing PGFL-based scores, factorial cumulants, or the initial graph-search procedure reduces mean F1 by 1.57, 3.43, and 46.91 percentage points, respectively. We have added the results in Section 5, Table 1, and specified the operations removed and retained in Appendix D.5.

Runtime and calibration sensitivity We added runtime comparisons on the same datasets, scaling experiments from 10 to 50 event types, and sensitivity analyses for self-loop thresholds and reliability thresholds for empirical H. Recovery is stable around the reported default thresholds. These protocols and results are reported in Appendices D.3, D.6, and D.7.

Real-world evaluation and practical scope We clarified that the five alarm types were selected because their reference graph contains directed cycles. Across the tested fitting decay rates, C-SHP recovers seven or eight of the nine reference edges and reaches F1 of 0.8889 at the default. We also explained SHP's DAG restriction and its actual fitted graph. Section 5 and Appendix D.8 now describe the selection rationale, real-data sensitivity, and practical considerations for more complex event systems.

Thank you again for the valuable feedback, which has helped us improve the clarity and experimental evaluation of the paper.

Dear Area Chair,

Thank you for handling our submission. For your convenience, we summarize the outcome of the discussion and the main revisions below. New and revised text is marked in red in the revised PDF.

Current ratings: 3/3/2 (initially 3/2/2).

  • Reviewer L8oT (3): "I am not confident enough in my independent assessment of the theoretical contributions to support a stronger recommendation. I therefore maintain my rating."
  • Reviewer BW5b (2 → 3): "which address most of my concerns, so I have raised my score."
  • Reviewer ye9g (2): Has not responded.

Main concerns and what we added

  1. Identifiability and its link to the algorithm (BW5b, ye9g). We revised the statements of Theorems 2 and 4 and the surrounding text to clarify the path from the exact count-PGFL representation to local structural signatures and global identifiability (Sections 3.1 and 3.3). Appendix C.6 now states sufficient conditions for consistent graph selection under a common scoring criterion.
  2. Self-loops and directed cycles (L8oT, BW5b, ye9g). We added separate results for self-loops, reciprocal edges, and cycles of length three to five. C-SHP achieved the highest mean F1 in every reported structural category and recovered all eight self-loops in each of the 20 new synthetic datasets (Figure 3 and Appendix D.3). We also report F1 including self-loops for the six original sensitivity experiments (Appendix D.2).
  3. Parameter sensitivity and model assumptions (BW5b, ye9g). We reported the main tuning parameters and added experiments on decay misspecification, alternative kernels, and a hidden common parent. C-SHP’s mean F1 ranged from 0.9041 to 0.9144 across the six tested decay ratios and remained above 0.88 in the kernel and hidden-parent experiments (Figure 3 and Appendix D.4).
  4. Ablations, runtime, and real-world evaluation (L8oT, BW5b, ye9g). We added component ablations, runtime comparisons on the same datasets, and separate scalability and calibration experiments (Table 1 and Appendices D.3 and D.5–D.7). We also explained why the real-world evaluation focuses on five alarm types with a cyclic reference graph. C-SHP recovered seven or eight of the nine reference edges at every tested decay rate (Appendix D.8).

We thank the reviewers for their constructive feedback, which helped us clarify the theoretical results and strengthen the experimental evaluation.

Sincerely,

Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.