acceptodds
Under review as a conference paper at ICLR 2027

Beyond Predictive Invariance: Causal Mechanism Transport for Generalizable Brain Network Analysis

Abstract

Brain network analysis is challenged by substantial distribution shifts arising from population and acquisition heterogeneity across imaging sites. Existing domain generalization methods improve robustness by seeking predictive regularities shared across source domains, while site heterogeneity can arise through distinct biological and measurement pathways. To characterize transportability under these structured variations, we introduce Causal Mechanism Transport (CMT), which models the two pathways in a multi-site selection diagram and derives a conditional criterion for transportable disease-related representations. To operationalize this criterion from observed multi-site data, we develop Causal Information Contrast (CIC), which contrasts representation distributions before and after conditioning on site within comparable disease and biological contexts. Instantiated with conditional Cauchy-Schwarz divergence, CIC provides a differentiable training objective. Experiments spanning controlled biological and measurement shifts, multiple real-world generalization settings, and independent representation audits across five neuroimaging cohorts demonstrate consistent gains across architectures and reduced conditional site dependence.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Brain Network, Domain Generalization, Representation Learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_osZUOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies generalization across neuroimaging sites by distinguishing changes in biological context from changes in measurement mechanisms. It defines Causal Mechanism Transport (CMT) through a stable latent-state-to-representation channel and derives conditional site independence given disease status and biological context as an observable consequence on common support. Causal Information Contrast (CIC) operationalizes this consequence using a conditional Cauchy–Schwarz objective alongside classification. Validation includes controlled shifts, leave-one-site-out and cross-dataset prediction, comparisons across architectures, and diagnostics of residual conditional site dependence.

Critical strengths
  1. The conditional comparison follows a meaningful scientific distinction. Comparing sites within disease and biological contexts aims to reduce measurement-related differences while retaining relevant population variation. Theorem 2.1 gives a useful one-way connection between the proposed transport property and this observable criterion, with a common-support restriction. This provides a coherent reason for the choice of alignment objective.

  2. The evaluation examines the modeling argument as well as predictive performance. Controlled biological–measurement interactions, conditional site probes, and the jointly conditioned MMD comparison in Figure 8 address whether conditioning matters. Together with results across cohorts and architectures, these analyses provide substantive support for conditional site alignment as a practical learning strategy.

Critical weaknesses
  1. The relationship between the population divergence and the implemented surrogate needs a more precise account. Appendix A.5 defines the population criterion on a shared conditioning space, whereas Equations 13–14 use sample-based normalization and kernel smoothing. In the idealized discrete, indicator-kernel case with full common support across kk sites, the empirical expression can approach −log⁡k-\log k under conditional independence. This constant offset does not change the optimizer. It does, however, make it important to distinguish numerical equality from equivalence of minimizers. The relationship under the actual fixed-bandwidth age kernel, adaptive representation kernel, and feature bank remains less explicit. Figure 9 and Table 11 usefully examine neighborhood support and bandwidth sensitivity, but do not by themselves establish that relationship or specify how inadequate cross-site overlap is handled. This limits the precision of the theoretical interpretation assigned to the trained regularizer.

  2. The mechanism-level interpretation is less directly supported than the conditional alignment contribution. The paper explicitly acknowledges that the theoretical implication is one-way. The remaining issue is how far the empirical success supports the latent-channel interpretation in the definition of CMT. Reduced conditional site dependence on observed sites need not persist under a new measurement mechanism. The controlled experiments provide relevant evidence, but the paper should connect the measurement shifts they cover more explicitly to the scope of its transport claims. This distinction affects the claimed conceptual advance over predictive invariance, even though the conditional alignment method remains useful.

Questions for authors
  1. What is the precise relationship between the population criterion in Equation 65 and Equations 13–14 under the implemented kernels and feature bank: equality up to a constant, a limiting relationship, or a surrogate with shared minimizers? How are contexts with inadequate cross-site overlap treated, beyond the reported effective-sample-size diagnostics?

  2. Which aspects of the existing controlled experiments support the mechanism-level interpretation beyond successful conditional alignment, and what class of measurement shifts do those results cover? Please distinguish the formal implication, the empirically supported robustness, and the broader motivating interpretation.

Reviewer_tYaiOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper derives conditional site independence from causal mechanism transport and implements it through a conditional Cauchy–Schwarz penalty. Experiments cover synthetic shifts, unseen sites, and external cohorts across three backbones.

Critical strengths
  1. Source-only model selection and paired backbone comparisons show consistent gains across multiple cohorts.
  2. Synthetic interventions, held-out-source probes, conditioning ablations, and kernel/support analyses go beyond aggregate prediction scores.
Critical weaknesses
  1. The learned constraint does not establish mechanism transport. Theorem 2.1 proves only CMT ⇒ conditional independence, as Appendix A.3 acknowledges. The authors’ XOR example also permits zero source CIC and prediction error but complete target failure. The training objective therefore retains the motivating ambiguity.
  2. The causal interpretation requires unvalidated biological adjustment. Equation 2 assumes site-invariant disease prevalence and manifestation conditional on C, instantiated only by age and sex. Unmeasured clinical or recruitment differences could violate this assumption; reduced site predictability cannot identify which pathway was removed.
  3. The closest conditional-alignment comparison is missing on real data. Figure 8 includes (Y,C)-conditional MMD with performance close to CIC, but Tables 2–4 omit it. Thus, real-data gains do not isolate CIC from existing alignment objectives with identical conditioning.
Questions for authors
  1. What source-diversity assumptions make low source CIC informative about unseen transport, including the XOR example?
  2. How robust is CIC to omitted biological covariates or recruitment shifts?
  3. Does CIC outperform (Y,C)-conditional MMD on real cohorts with matched backbones and tuning budgets?
Reviewer_88UMOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper addresses cross-site generalization in brain-network disorder classification. It argues that predictive invariance across source sites does not guarantee transport to unseen sites and proposes Causal Mechanism Transport (CMT): a selection diagram in which the site S affects biological covariates C (age, sex) and measurement conditions M, while the mechanisms p(y|c), p(z|y,c) and p(x|z,m) are shared across sites. CMT requires the latent-state-to-representation channel to be site-invariant, which implies the observable criterion R ⊥ S | Y, C. The authors operationalize this with Causal Information Contrast (CIC), a kernel estimator of the conditional Cauchy–Schwarz divergence between p(r|y,c) and p(r|y,c,s), added as a regularizer to cross-entropy with a feature bank for conditional support. Validation comprises: i) a synthetic study with separately controlled measurement (ΔM) and biological-context (ΔC) shifts ii) leave-one-site-out evaluation on ABIDE-I, ABIDE-II, REST-meta-MDD and SRPBS-SCZ, where CIC added to GAT, BrainNetCNN and BrainNetTF improves all 24 backbone–metric comparisons (on average +3.05 ACC, +2.77 AUC) iii) three cross-dataset ASD transfers iv) a probe showing reduced conditional site information; and an ablation over the conditioning set.

Critical strengths
  1. Deriving R ⊥ S | Y, C from an explicit causal model gives a clear account of why marginal alignment (CORAL, MMD) fails when sites differ in diagnostic prevalence or demographics, and of what should be removed: residual site dependence within comparable disease–biological contexts. The synthetic ΔC experiment, the κ ablation (Fig. 8) and the real-data conditioning ablation (Table 4) consistently favor (Y, C)-conditioning over unconditional and Y-only variants. This is consequential because it is the paper's central conceptual contribution, and it is corroborated rather than only asserted.
  2. CIC improves three structurally different backbones across four LOSO cohorts and three cross-dataset transfers without changing inference cost. The site-probe analysis, which never touches the held-out site, directly checks that the learned representations carry less conditional site information. This supports the claim that the regularizer does what it is designed to do, not only that it raises accuracy.
Critical weaknesses
  1. On real data, backbone+CIC is compared with the plain backbone and with other architectures, but never with other invariance regularizers applied to the same backbone (e.g., CORAL, marginal MMD, Y- or (Y,C)-conditional MMD, conditional adversarial training). The backbone used for the CORAL and IRM rows in Table 2 is also not stated. In the authors' own synthetic study (Fig. 8), (Y,C)-conditional MMD performs comparably to CIC, suggesting that the improvement comes from the choice of conditioning set rather than from the conditional CS-divergence estimator. The positioning relative to established conditional-invariance domain generalization and to standard site harmonization (ComBat), which is absent as a baseline, is therefore unclear.
  2. Average gains (about 2–3 points) are small compared with the across-site standard deviations (often 5–14 points), and 5 of 18 testable comparisons are not significant. Theoretically, CMT implies the observable criterion but not conversely, so CIC enforces a necessary rather than sufficient condition. With C restricted to age and sex, any other site-dependent biological variation (severity, medication, diagnostic practice) is treated as measurement and removed. Finally, the synthetic generator exposes M directly in X, which makes measurement trivially identifiable and limits how far the controlled results speak to real scanner effects entangled with neural signal.
  3. How does the method perform compared to losses that involve concepts of instrument variables, backdoor and frontdoor adjustment of causality?
  4. Fig. 2 leaves p_s(c, m) unrestricted, so dependence between C and M is accommodated, but the analysis assumes there are no site-dependent factors outside C and M. How does CMT behave when this fails? In particular: (i) unobserved site-varying biology beyond age and sex, such as illness severity or medication, which CIC would treat as measurement variation and remove (ii) a direct S to Y path from site-specific recruitment or diagnostic criteria, which violates the shared p(y | c) (iii) a Y to M path such as disease-related head motion, which places diagnostic information on the measurement pathway, where conditioning on Y does not remove it. A discussion of which of these cases the criterion is robust to, and an empirical check (for example, controlling for mean framewise displacement, which the authors already report), would clarify the scope of the transportability claim.
  5. Theorem 2.1 guarantees p(r | y, c) is site-invariant, but under the paper's own admissible shifts p_t(y, c) can differ from the sources, so the target posterior p_t(y | r) ∝ p(r | y, c) p_t(y, c) can still change even when CIC is exactly zero. The classifier sees neither C nor the target prior at test time, and no bound relates CIC to target-site risk. In addition, the "CIC = 0 if and only if R ⊥ S | Y, C" result holds only in the population.
Questions for authors

Please see the Critical Weaknesses section.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.