- Rating
- 1
- Confidence
- 4
- Informativeness
- 1
- Presentation
- 1
- Soundness
- 2(+1)
Read review
Paper summary
I focus on paper strengths, weaknesses, and contributions in later review fields, as appropriate. My assessment of the paper is it currently stands, overall, is as follows: the paper is a 9-page guide to a 36-page appendix. While technically, appendices are unlimited, the paper should stand on its own; the current version of this paper does not. I am completely unable to read the paper without referring to appendices constantly. Some concepts are not even introduced in the paper. For example, S0-S2, the three spline reference configurations,or M0-M2 labels used in the structure tables, appear in the appendix only. I don't think it's the idea problem; I think it's the presentation problem. So, from the presentation standpoint, the paper could benefit from a serious re-writing.
From the idea perspective, the paper has some merit: the authors are attempting to model a dynamic system of multiple mutually-influencing nodes. While the paper is appropriately using concepts such as recurrent latent regimes, shared candidate score matrix, state-specific sigmoid gates, and top-k masks, the reality of the modeling is much simpler: each node can fire at any time and influence certain other nodes. The system is also multimodal, so some modes can appear at certain frequencies (e.g., seasonality), while others do not. This is a very complex problem, because both the nodes and the modes have to be modeled simultaneously. So the authors are proposing one shared candidate "wiring" diagram (PCDM), where different nodes are allowed to switch individual wires on and off with their own sets of gates. The idea is that doing so would allow to find rare, mode-specific wires that otherwise can be missed. The second piece that the authors propose, INFO, is a simple statistical model that ranks the nodes that matter most in each mode.
What the authors found was that PCDM did not work. INFO provided some useful node-specific wiring, but the idea did not hold, either, because it turned out that most of that added accuracy was already present when the simple wiring rearranged the gates. In other words, neither of the two components, introduced as novel by the paper, provided any improvement to modeling complexity of the dynamic causal system. Given the additional issues with presentation, this paper needs serious rework before it is ready for a conference.
Critical strengths
I think the most important strength of the paper is the effort to find the structure of influence in the multimodal dynamic causal system with nodes having the ability to mutually influence each other. Any work that gets us closer to discovering "true" causality outside the experimental design is a worthwhile attempt.
Specifically in this paper, I liked that the authors used the controlled evaluation objects appropriately. For the question of whether an edge is accurately evaluated to belong to the regime that uses it, PRIVATE AP, owner accuracy, signed margin, and finite-response NMSE are all good choices. Likewise, for evaluating the contribution of INFO, the authors used the matched restart controls such as KEEP, RANDOM with overlap matching, and screening-weight controls (POOLED, PERMUTED). So from that perspective the design was solid.
I also liked that the authors were careful in reporting the results, noting the intervals that include zero, and reporting nearly-null results upfront, in the abstract, stating that the benefit of sharing “remains to be established.” I also have to commend on the effort that produced the 45 pages of material for this study: even though the proposed ideas did not seem to work, at least the authors are no aware of what does not work and can look for what does.
Critical weaknesses
The most critical weakness is that the paper's title and abstract do not match the paper's contribution - or at least it does not seem that they do. The paper claims shared structure with local modulation as its contribution, but the results do not support the conclusion that sharing helps. These results are clearly stated in the paper itself, and given the null results, I wonder if the paper can be reframed in its current state. What the experiments seem to support is a different, much narrower contribution: that a posterior-weighted additive spline screen can assign private edges to their respective regimes.
But that result belongs to the screen rather than to the shared-plus-gated architecture the title indicates, and most importantly, such screens already exist. For example, ranking candidate inputs per target with an additive spline fit: the method proposed by Fan, Feng, and Song (2011) is nonparametric independence screening, which the paper cites but does not compare against. I think the comparison will show that Fan et al., 2011, will perform just as well as INFO, if not better. Another method, CAM, in the causal discovery literature specifically (Bühlmann, Peters and Ernest, 2014) uses preliminary neighborhood selection by additive regression, then order search, then pruning by significance tests in the fitted GAMs. The paper's proposed INFO’s “fit splines, rank, keep top-k per target” is CAM’s first and third steps without the order search.
I also noticed that the authors did not evaluate their proposed method against any of the competitors that they name in related work. I would have expected to see a comparison against at least one of the methods that evaluate dynamic causal discovery, some of which the authors mention: Regime-PCMCI (Saggioro et al. 2020), CD-NOD (Huang et al. 2020), the regime-structure learner of Rahmani and Frossard (2025), non-homogeneous DBNs (Grzegorczyk and Husmeier). On regime-switching data, the authors provide comparisons to PCDM's own ablations, but I don't think it's sufficient for comparison. In external baseline on DREAM3 (stationary, no regimes), PCDM performs worse than DYNOTEARS and NTS-NOTEARS on both macro metrics.
Questions for authors
Most of my questions are suggestions for future work. However, there are a couple of questions that could potentially affect my assessment.
-
The run against the Regime-PCMCI, CD-NOD and the Rahmani-Frossard regime learner. If they were already run during the development and omitted in the paper, it's an answer in its own. But if they haven't been run, maybe they should be. If PCDM performs better than any of them, both the direction and the magnitude would matter.
-
What distinguishes INFO’s screening step from the literature in the posterior-weighted additive spline neighborhood selection that I've mentioned in the previous section?
2 comments
Thank you for running all the experiments I've asked for. I will raise my soundness score to 2 because your work is clearly accurate and competent. However, the results, as expected, confirmed my earlier concerns about lack of substantive contribution. I also looked at the revised paper, and it did not improve much. While undefined labels appear to be gone from the main paper, the appendix grew by another 10 pages or so, so the ratio of the main text to appendix got worse. I know appendices are unlimited, and now there are some cross-references (with the exception of sections 4.4 and 4.5, where I didn't find any), but the paper should be understandable as a stand-alone - and it still is not.
I will retain the score of clear rejection because this work did not fundamentally improve the paper's contribution, and in its present form, it is not ready for publication.
Response to 4KvQ
Response to 4KvQ
Thank you for reading the paper closely. On rereading the experiments, we also found that understanding the comparisons meant going back and forth to the appendix. Section 4 now introduces the controls where they are used. Table 5(a) puts the complete spline pipeline alongside the two neural continuations.
1. Comparisons with external methods
We checked the archives and found no verifiable pre-submission records of completed runs for the requested methods. We have now compared linear Regime-PCMCI on five matched graphs per family using common training timestamps. On additive data, canonical INFO improved state ARI by 0.050 and boundary F1 by 0.437. Both gains passed correction. PRIVATE AP differences remain unresolved in both families. CASTOR used a single state in each run.
We also adapted the released SDCI code to the same observations and lag-one targets, with three states supplied to both models. Conditioned INFO gave response NMSE of 0.113/0.157 compared with SDCI's 1.028/0.992. Time-weighted ALL AP was 0.991/0.989 and 0.489/0.499, respectively. Table 5(a) gives both endpoints. Appendix F.6.1 describes the different state definitions, priors, capacities and budgets.
The nonlinear Regime-PCMCI and CD-NOD attempts used three paired seeds per family and method. Our rough estimate for completing these runs on a Ryzen 9 9950X3D workstation with an RTX 6000D was more than a month. We stopped the runs before completion and have no comparable results from those runs. On DREAM3, DYNOTEARS and NTS-NOTEARS still score higher than INFO on both macro metrics (Appendix B).
On OPPORTUNITY we hold out each of four participants and infer states offline from their sensor windows. Mean locomotion ARI is 0.1192 for conditioned INFO, 0.0808 for the complete simple pipeline and 0.0351 for the spline-HMM. INFO has higher ARI than both routes for all four participants. It also has lower reconstruction error than the spline-HMM for all four and lower error than the simple route for two (Appendix I).
2. The relationship to NIS, INIS-SpAM and CAM
We agree about the link to existing additive screens. INFO weights lagged observations by the inferred state posterior, ranks candidate parents and reassigns the existing gates at restart. We compared posterior-weighted INIS-SpAM and lagged CAM with the same observations and state weights. The joint screen's PRIVATE AP was higher than INIS-SpAM's by 0.00152 on additive data and 0.00080 on weak interactions. The advantage remains unresolved after correction against both comparators. Section 4.3 now explains how we use the screen to initialize the neural restart.
Sharing and subsequent learning
A fixed-posterior comparison showed a small response improvement from shared scores. Broader localization gains remain unresolved. Appendix A.1 gives an example in which a common score matrix produces disjoint hard supports and explains why the theory assumes common support.
Against KEEP and RANDOM, the clearest gain from INFO is the initial parent assignment. Decoder conditioning improved neural fitting on the original and new graphs. On the original graphs, fitting a spline on the same final parents and posterior still reduced NMSE by 0.0806/0.0429. None of the 12 condition–family combinations met the prespecified response-benefit criterion against both complete spline routes. The main text now reports these findings separately.