- Rating
- 4
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
Event logs in complex systems frequently exhibit non-stationarity in both background activity and causal dependencies, causing traditional segmentation-based discovery methods to misattribute background drift to true influence and generate spurious edges. The authors introduce CausalField, a continuous point-process framework that models causal influence as a state-conditional field parameterized by a coordinate-based neural network, explicitly separating time-varying excitation from background drift while learning directly from the point-process likelihood.
Critical strengths
- Overall, I consider this a very strong paper. It is clearly written and easy to follow. The authors explicitly state the research questions in lines 71–74 and enumerate the contributions in lines 85–95. Figure 1 provides a clear overview of the proposed approach and its motivation. I particularly appreciate the effort put into the presentation, which I believe makes the paper accessible to readers beyond the causal inference community.
- The results are presented clearly, and the limitations are explicitly discussed in the appendix.
Critical weaknesses
- In the experimental section, I would like to see a justification for the choice of evaluation metrics and benchmarks.
- I would also like to see a comparison of runtimes across the evaluated methods.
Questions for authors
- How does the runtime of the proposed method compare with those of the baselines?
- Why were these particular baselines selected?
3 comments
I am happy with the authors' answer and I maintain my score.
We thank the reviewer for the positive feedback and thoughtful evaluation of our work. If you have any further questions or concerns during the discussion period, please feel free to raise them. We will respond promptly.
Evaluation design, baselines, and runtime (W1–W2, Q1–Q2)
Thank you for your thoughtful review and constructive questions. Unless stated otherwise, table and algorithm line numbers refer to the submission; additions to the revised manuscript are marked in blue.
W1 Choice of evaluation metrics and benchmarks.
We designed the evaluation to assess graph recovery, coefficient accuracy, and prediction from complementary perspectives. Precision, recall, F1, and MCC assess the recovered edge set at a threshold; ROC AUC measures ranking across thresholds, and AUPRC summarizes precision–recall trade-offs in sparse graphs (Section 4.1 and Appendix B.2). Separate background and influence RMSE assess component recovery on held-out synthetic histories (Section 5.1 and Appendix C.2). Held-out alarm NLL per event assesses predictive fit (Appendix C, Table 18).
We chose benchmarks according to the available ground truth. The main synthetic benchmark varies background and coupling separately and measures their trajectory errors (Sections 4.1 and 5.1). Alternative generators test departures in delay, memory, and observation link (Section 5.2 and Appendix B.1). Recorded alarms test aggregate recovery in an operational setting (Section 4.2). The new semi-synthetic dataset adds known temporal and history-dependent switches for regime-specific evaluation (revised Sections 4.1–4.2, Appendix B.1 and C.4).
Q2 Choice of baselines.
We included StateHPs as the direct comparison for state-dependent modeling. The other nine semi-synthetic baselines cover Hawkes structure estimators (L0Hawkes, NPHC, SHP), neural approaches (CAUSE, NGC, NHP), event-sequence discovery (CASCADE), and topology-aware point-process methods (THP, TTPM). Section 4.1 and Appendix B.3 describe the baselines and tuning. The separate synthetic drift experiment includes seg-ADM4 as a segment-wise comparator. Time-only, blocked-time, and quantized-state variants test representation choices within CausalField (Section 5.1 and Appendix C).
W2 & Q1 Runtime comparison.
At 100 event types, CausalField took 164 seconds per seed on the synthetic scalability benchmark. NPHC took 29 seconds, NGC 52, L0Hawkes 117, THP 572, and CAUSE 5,834. SHP reached the time limit, as did TTPM at δ=0.1. CausalField completed all five tested sizes, from 24 seconds at 20 types to 491 seconds at 200 (Table 23, Appendix C.3). The sweep contains no matched StateHPs, CASCADE, or NHP timings.
We appreciate your feedback and will clarify these points in the revised manuscript.