acceptodds
Under review as a conference paper at ICLR 2027

Causal Integrated Gradients: Scalable Feature Attribution Along Causal Paths

Abstract

Existing feature attribution methods that account for feature dependencies do not scale to high dimensions. We propose causal integrated gradients (CIG) and causal expected gradients (CEG), which extend integrated gradients to incorporate prior causal structure among input features while retaining its computational efficiency and, under conditions on the causal graph, its axiomatic guarantees. In many settings, some causal structure between features is already known—for example, that a transcription factor causally regulates a target gene's expression. Yet existing gradient-based methods cannot exploit this structure, instead crediting a feature's downstream effects over its true upstream cause, such as ranking the target gene above the transcription factor that regulates it. Our key idea is to embed this causal structure by routing the integration path along the causal graph. The method thereby decomposes each feature's attribution into a direct effect and effects mediated through its causal descendants. We evaluate CIG in three settings of increasing realism. First, on a synthetic benchmark with a known causal model, CIG's attributions best match the ground-truth causal effect from Pearl's do-calculus, matching the accuracy of causal Shapley values at a fraction of the computational cost. Second, on simulated single-cell gene expression data with a known gene regulatory network (Dyngen), CIG most accurately recovers the transcription factors driving the simulation. Third, on real interferon- stimulation data, where only an approximate causal graph is available, CIG better prioritizes the transcription factors driving the measured response than DeepSHAP and integrated gradients. Together, these results show that incorporating causal structure into gradient-based attributions recovers upstream drivers that existing methods miss, without sacrificing scalability.

Primary area:
interpretability and explainable AI
Keywords:
feature attribution, explainability, integrated gradients, causal structure, interpretability, gene regulatory networks, single-cell RNA-seq

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

This paper proposes Causal Integrated Gradients (CIG) and Causal Expected Gradients (CEG), extending traditional gradient-based feature attribution to account for causal dependencies between input features. Given a pretrained predictive model and a causal DAG, the method learns surrogate causal mechanisms from observational data and uses their derivatives to decompose feature attributions into direct and indirect effects, redistributing importance toward upstream causes. Its main contribution is incorporating causal structure into gradient-based attribution while maintaining computational efficiency compared with causal Shapley values. The authors validate their approach on synthetic data with known causal effects, simulated gene regulatory networks (Dyngen), and real single-cell gene expression data (IFN-β), demonstrating improved recovery of causal regulators and substantially lower computational costs. Additional experiments assess robustness to graph misspecification and surrogate estimation errors.

Critical strengths
  1. The proposed method is computationally more efficient than existing dependency-aware approaches, particularly causal Shapley values. This makes it suitable for high-dimensional applications such as RNA-seq gene expression data, where computational costs can become prohibitive.

  2. The idea of incorporating causal dependencies into integrated gradients through differentiable surrogate models is interesting and relatively straightforward. Using surrogate derivatives instead of sampling from interventional distributions provides a practical alternative while building on an established attribution method.

  3. The authors evaluate the method under different settings, ranging from a known SCM and DAG, to a known DAG with learned mechanisms, and finally to an approximate DAG in a real biological dataset. This progression provides additional credibility to the proposed methodology and its potential applicability.

Critical weaknesses
  1. While the authors investigate graph misspecification and surrogate approximation errors, the evaluation does not explicitly examine the impact of latent confounders on the learned causal mechanisms. Since CIG relies on surrogate derivatives estimated from observational data, it would be useful to assess how sensitive its attributions are to unobserved common causes, particularly in biological settings.

  2. The main synthetic comparison uses the true SCM for both CIG and causal Shapley values, whereas practical applications rely on estimated causal mechanisms. Comparing both approaches under the same learned causal model would help clarify their relative robustness to estimation errors, beyond the demonstrated computational advantages.

Questions for authors
  1. Have the authors considered evaluating CIG under different levels of latent confounding? Since the surrogate models are trained on observational data, it would be useful to understand how errors in the estimated causal mechanisms affect attribution quality, beyond the graph misspecification experiments already presented.

  2. Have the authors considered introducing sparsity-inducing regularization (e.g., L1 or group Lasso) in the surrogate models to reduce the influence of potentially spurious causal edges? Given that the method relies on an approximate DAG in real-world settings, it would be interesting to assess whether such constraints improve attribution robustness or reduce computational costs.

3 comments
AuthorsOpenReview ↗

Response to Reviewer fvrG (1/2)

(Part 1/2)

We thank the reviewer for the positive assessment and the concrete suggestions.

W1 / Q1 (latent confounding).

We agree this was missing. To quantify it, we added hidden confounders to the synthetic benchmark (100 features, MLP, 5 seeds; new Appendix L). The confounders are neither features nor nodes of the supplied graph. We test two types: per-edge confounders, each acting on both ends of one of 25% of the edges, and 3 global factors, each acting on 30% of the nodes. We vary their strength ρ\rho, the share of a confounded node's exogenous variance that comes from its confounders, from 0 (the unconfounded benchmark) to 0.75, keeping total variances fixed. The surrogate is learned from the confounded data on the true graph. To separate the cost of its bias from everything else, we also run CIG and CEG with the true mechanisms. The ground-truth CATE treats each sample's confounder values as fixed background: they are held at their realized values, and only the independent noise is resampled.

As expected, confounding strongly biases the learned mechanisms. Attribution quality, however, degrades gradually rather than collapsing. On root nodes (Fig. L.1), CIG with the learned surrogate goes from 0.79 at ρ=0\rho=0 to about 0.70 at ρ=0.75\rho=0.75, with CEG close to it, still about twice as high as IG (0.31–0.35). About half of the loss comes from the surrogate's bias. The other half also occurs with the true mechanisms, because the CATE holds each sample's confounder contributions fixed as background, whereas CIG credits each node for its full residual change, including the part driven by the confounders. All nodes show the same pattern with smaller margins; with global confounders at ρ=0.75\rho=0.75, the learned-surrogate curves approach IG. More accurate mechanisms would reduce the part due to the surrogate's bias, which we now note in the Limitations. On real data, we added the analogous check for the most likely confounder in the IFN-β\beta data, the stimulation itself: fitting the surrogate on control cells only, so that the stimulation cannot bias the learned mechanisms, leaves the IFN-β\beta conclusions unchanged (Appendix I; see our answer to Reviewer UQnW, Q3). We also tested a latent confounder in a biological simulation (new Appendix L.2): a hidden batch variable in Dyngen, coupled to time in the training data and not part of the graph, which the model exploits as a shortcut. With the surrogate fitted to these confounded cells, CIG and CEG still rank the TFs near the top (AUROC 0.98), whereas DeepSHAP, EG, IG and DeepLIFT rank them at or below chance. CIG and CEG also give the batch genes, which the model uses as a shortcut, about 40% less importance than IG and EG, less than a uniform share. They still rank these genes above chance (AUROC 0.56–0.63), as expected, since the model does use them: CIG moves part of their credit to upstream regulators rather than removing it. Unlike those of the graph-agnostic methods, their attributions do not cluster by batch.

AuthorsOpenReview ↗

Response to Reviewer fvrG (2/2)

(Part 2/2)

W2 (compare CIG and causal Shapley under the same learned causal model).

We agree this is the comparison that matters in practice and have added it at 100 features (Appendix F, "Matched model evaluations", Fig. F.2). In the learned-surrogate setting, causal Shapley values sample from the learned surrogate as an additive-noise SCM, with Gaussian noise whose per-node standard deviation is that of the surrogate's training residuals. CIG and CEG use the same surrogate. We vary each method's budget, from 1 to 512 cycles for causal Shapley values. At the one budget both reach (about 10410^4 model evaluations per sample), CEG scores 0.76 against 0.43 for causal Shapley values on root nodes (Fig. F.2, right), and even at 512 cycles causal Shapley values stay below CIG and CEG. Estimation error lowers all methods by a similar amount (0.05–0.07, since the surrogate is trained for only 10 epochs in this experiment), so the ranking is the same as with the true SCM, and CIG is not more sensitive to estimation error than causal Shapley values. With the 50 epochs used below, the learned surrogate matches the true mechanisms almost exactly.

Across 100–1000 features, we reran the comparison of CIG with true and learned mechanisms on the revised benchmark, for both the MLP and the Lasso (Appendix F, Figs. F.5–F.6). With the MLP, the two give nearly identical curves, and with the Lasso the learned surrogate is at most about 0.02 lower; both stay far above IG. The learned surrogate is also faster, because its Jacobian is computed analytically for the graph's edges.

Q2 (sparsity-inducing regularization on the surrogate).

The surrogate is already sparse in structure, since it only parameterizes edges in the supplied graph, and our corruption study (Appendix K) shows that spurious edges barely affect root-cause attribution, unlike missing or reversed edges.

The reviewer's suggestion fits our surrogate naturally: each edge has its own coefficient vector wijw_{ij} over the basis expansion, so a group Lasso λ∑(i,j)∥wij∥2\lambda\sum_{(i,j)}\lVert w_{ij}\rVert_2 with one group per edge can switch an edge off entirely, removing it from the surrogate Jacobian. A plain ℓ1\ell_1 penalty would zero individual basis coefficients, so an edge would leave the surrogate Jacobian only once all of its coefficients vanish; the group Lasso is its edge-level version. We implemented it (proximal block soft-thresholding after each optimizer step, so edges reach exact zeros) and tested it with 50% spurious edges added (new Table K.2). Root-cause attribution stays at 0.93–0.97 up to λ=2\lambda=2, because spurious edges never hurt it: the unregularized surrogate already shrinks them enough. On all nodes, spurious edges do have a small cost (0.08 at d=1000d=1000), and the penalty removes it: at λ=2\lambda=2, which prunes all spurious edges and about a fifth of the true ones, the graph with spurious edges scores as well as the true graph with the same penalty (Table K.2). Too large a penalty (λ=3\lambda=3) deletes the graph, bringing root-cause attribution down to the level of IG. In practice we would recommend choosing λ\lambda by held-out reconstruction error of the surrogate. Sparsity cannot fix missing or reversed edges, the corruptions that do degrade CIG, since it can only delete edges. On cost, pruning does not make training cheaper, since every candidate edge is still fitted, and makes attribution cheaper only by a constant factor (Appendix D). We added the table and this discussion to Appendix K and a sentence to the Limitations.

I appreciate the author's detailed responses and the performed experiments. I decide to maintain my score.

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

The paper proposes Causal Integrated Gradients (CIG) and Causal Expected Gradients (CEG), which extend integrated gradients to incorporate a known causal graph over input features. CIG models each feature as a surrogate-predicted function of its parents plus a residual, and uses the chain rule to split each feature's attribution into a direct effect and an indirect effect passed back from its children, tracked for one or two hops (with a general H-hop formula in Appendix A). The surrogate is a sparse, additive, basis-expanded model whose Jacobian is analytic, which keeps the method scalable to thousands of features. The authors claim completeness and sensitivity(a) for the direct term, plus implementation invariance, linearity, and symmetry under stated preconditions on the graph and surrogate, while deliberately giving up the dummy axiom. On a synthetic benchmark, CIG tracks ground-truth CATE slightly better than causal Shapley values at far lower runtime and clearly better than conditional SHAP and IG, and on simulated Dyngen data it best recovers the driving transcription factors. On real IFN-β stimulation data, CIG ranks upstream and stimulation-responsive transcription factors higher than DeepSHAP and IG, but recovers differentially expressed genes worse than both.

Critical strengths
  1. The paper scales dependency-aware attribution far beyond prior work, which was validated on at most 28 features (Heskes et al. used 7, and Aas et al. used 3–10 with a largest real-data example of 28). Here the method reaches 6,941 genes with transparent cost accounting, taking 3.8 h end-to-end on IFN-β split roughly evenly between surrogate training (1.8 h) and attribution (2.0 h).

  2. The evaluation is candid and spans three settings of increasing realism, including a per-sample CATE ground truth that is rarely available for evaluating attributions. The paper also reports results that do not favor CIG, such as the lowest DEG NDCG in every cell type and the lowest F1 and recovery AUC in six of seven. Its corruption study shows the expected asymmetry, tolerating spurious edges while degrading under removal and reversal, and Appendix L discusses surrogate bias openly.

Critical weaknesses
  1. CIG appears equivalent to IG along the same straight path, with credit reallocated by the surrogate Jacobian, and the paper does not discuss or ablate this. On IG's line, ∂γi/∂α=(xi−bi)−∑k∈pa(i)∂si∂xk(xk−bk)\partial\gamma_i/\partial\alpha = (x_i - b_i) - \sum_{k \in \mathrm{pa}(i)} \frac{\partial s_i}{\partial x_k}(x_k - b_k) in closed form, so each child's parent-explained share of IG credit is simply handed to its parents. Since RiR_i then contains ∂F/∂xj\partial F/\partial x_j terms for j≠ij \neq i, CIG is not a path method in the usual sense, and the axiom inheritance claimed in Sec. 3.5 needs restating. The closed form also makes the finite difference in Eq. 6 unnecessary, and that finite difference appears to cause both the ~10−310^{-3} completeness residual and CEG's high cost. The key baselines, DeepSHAP redistributed through the same Jacobian and naive graph propagation of IG, are missing, and Appendix J cannot substitute for them because graph-propagated IG would also be instance-specific. Without these baselines, the gains cannot yet be attributed to CIG's formulation rather than to graph access.

  2. The evaluations favor CIG's inductive bias, starting with a synthetic benchmark that uses a linear FF, depth ≤2 graphs, and a reduced causal Shapley budget, which confounds both the accuracy and runtime comparisons. On Dyngen and IFN-β, the ground truth rewards ranking upstream TFs, and CIG receives the graph that defines it (Dyngen) or closely resembles it (OmniPath input vs. SCENIC evaluation edges). The only graph-independent readout, DEG recovery, goes against CIG, so the experiments cannot yet separate correct use of the graph from generic upstream credit-shifting.

Questions for authors
  • Q1. Please replace the finite difference with the closed form above and report the completeness delta, along with CEG's accuracy and runtime using one path point per baseline.

  • Q2. Please add DeepSHAP redistributed through the same surrogate Jacobian and naive graph propagation of IG credit as baselines on Dyngen and IFN-β.

  • Q3. Since stimulation is both the prediction target and a likely confounder, please refit the surrogate on control cells only, or add stimulation as a root in the surrogate (not as an input to FF), and check whether Fig. 5 still holds.

  • Q4. If xj=bjx_j = b_j but a parent moved, ∂γj/∂α≠0\partial\gamma_j/\partial\alpha \neq 0, so the unchanged child receives a nonzero direct attribution, and the Appendix F result R2=0R_2 = 0 likewise holds only when the baseline lies on an affine mechanism. Could you clarify whether this is intended and how large the effect is in practice?

  • Q5. Why truncate at H = 1 or 2 rather than use the full (I−J)−1(I - J)^{-1}, which needs only a cheap sparse triangular solve, and do the results depend on H?

  • Q6. Please compare against causal Shapley with a nonlinear FF, deeper graphs, and a matched number of model evaluations at up to 359 features.

  • Q7. Please align the abstract's "integration path along the causal graph" with the straight-line path used in Sec. 3.2, and define precisely what "α as residual information" means.

3 comments
AuthorsOpenReview ↗

Response to Reviewer UQnW (1/3)

(Part 1/3)

We thank the reviewer for an exceptionally careful review. The observation about the closed-form path derivative is correct, and following it through has substantially improved the method (general response, points 1 and 2).

W1, Q1 (closed form; path-method status).

We agree with the analysis. We have replaced the finite difference by the closed form (general response, point 1). CIG's completeness delta now equals that of IG, 10−710^{-7}–10−610^{-6} of the output difference instead of ∼10−3\sim10^{-3} on the synthetic benchmark, and CEG (one path point per baseline) matches CIG in accuracy at a comparable cost. Substituting it into Eq. 3, the integrand of the one-hop attribution is

∂F∂γi(xi−bi)⏟IG−∂F∂γi∑k<i∂si∂xk(xk−bk)⏟passed to the parents+∑j>i∂F∂γj∂sj∂xi(xi−bi)⏟received from the children,\underbrace{\frac{\partial F}{\partial \gamma_i}(x_i - b_i)}_{\text{IG}} - \underbrace{\frac{\partial F}{\partial \gamma_i}\sum_{k<i} \frac{\partial s_i}{\partial x_k}(x_k - b_k)}_{\text{passed to the parents}} + \underbrace{\sum_{j>i} \frac{\partial F}{\partial \gamma_j}\frac{\partial s_j}{\partial x_i}(x_i - b_i)}_{\text{received from the children}},

i.e. IG's integrand plus an exact transfer of credit from each child to its parents, which we now state explicitly in Sec. 3.2. We agree that the one-hop version is therefore not a path method in the usual sense, since RiR_i contains ∂F/∂γj\partial F/\partial\gamma_j terms for j≠ij \neq i. It satisfies the axioms of Sec. 3.5, which Appendix C proves directly, and the revised Sec. 3.5 no longer presents them as inherited from the path-method framework. The all-hop version, which uses the full (I−J)−1(I-J)^{-1} as suggested, is exactly a path method, on G(h):=F(Φ(h))G(h) := F(\Phi(h)) over the exogenous residuals, and inherits the axioms with respect to GG (general response, point 2; Appendices A and N). We keep the one-hop derivation in the main text because it is the simplest and makes this transfer of credit explicit; the HH-hop and all-hop versions extend it further upstream, and all experiments use the all-hop version unless stated otherwise.

W1, W2, Q2 (graph access vs. formulation; missing baselines).

We added both as ablations of CIG, each adding one of its ingredients to the model: naive propagation along the graph's edges (IG+MP), and propagation through the mechanisms' Jacobian (DeepLIFT$+J$, DeepSHAP$+J$). We tested them on the synthetic benchmark, where the CATE gives an exact ground truth (new Appendix J.1), and added IG+MP and DeepSHAP$+J$ to both biological experiments (Appendices J.2–J.4).

Naive graph propagation of IG. IG+MP starts from integrated gradients; each non-root feature then passes its credit to its parents, split equally, until all of it sits on the root nodes. It uses only the graph's edges and stays complete, so it is also the control for the generic upstream credit-shifting raised in W2. It does not recover the CATE. On root nodes it stays near IG (above it with the Lasso, below it with the MLP), and on all nodes far below it, since it strips all credit from non-root features (Fig. J.1). On Dyngen, all CIG and CEG variants outperform it in backbone-TF recovery (Fig. J.2), and it is significant in none of the six transitions (Fig. J.3). On IFN-β\beta, it ranks the responsive TFs below the other genes (Fig. J.5). Credit has to be split according to how much of each feature's change its parents explain, which needs the mechanisms' strength and sign, and whether the parent changed at all. We agree that the submission's Appendix J, which only tested instance-level specificity, could not rule this out.

Propagation through the mechanisms. DeepLIFT$+J$ splits each feature's DeepLIFT credit between its own residual change and its parents, in proportion to their contributions under the path-averaged Jacobian Jˉ\bar J (of the true mechanisms on the synthetic benchmark), and continues up to the roots. DeepSHAP$+J$ averages this over CEG's 50 background rows. On the synthetic benchmark, both reach CIG's level, within about 0.01 (Fig. J.1). On Dyngen, DeepSHAP$+J$ scores in a similar range to CIG and CEG and is significant in all six transitions. On IFN-β\beta, it ranks the responsive TFs above the graph-agnostic methods but below CIG and CEG. We read this as an ablation, not a tie between competing methods. The +J+J methods borrow CIG's central component: they replace the gradient along the path by DeepLIFT's multipliers and J(α)J(\alpha) by its path average, and they coincide with CIG when FF and the mechanisms are linear along the path. Together with the failure of naive propagation, this shows that the gain comes from propagating credit through the causal mechanisms, not from graph access. CIG remains the principled form of this idea: a path method over the residuals (general response, point 2) that inherits completeness and implementation invariance, the latter of which DeepLIFT's multipliers violate (Sundararajan et al., 2017). The +J+J methods also offer no runtime advantage, needing the same 100 Jacobian evaluations per sample as CIG.

AuthorsOpenReview ↗

Response to Reviewer UQnW (2/3)

(Part 2/3)

W2, Q6 (evaluation design).

  • Synthetic benchmark. We reran the benchmark (Sec. 4.1, Fig. 3) on deeper graphs (depth 4 instead of ≤2\le 2), with an MLP on a nonlinear target as the main model and the Lasso as a second one (Fig. F.4), and added expected gradients (EG), the non-causal counterpart of CEG. Causal Shapley values now use the same budget at every size up to 1000 features: 4 permutation cycles ×\times 50 rows (the "10 permutations" of the submission, in shap's units), so the previous reduction at $\ge$464 features is gone. That is 400–4,000$\times$ more model evaluations than CIG. On root nodes, CIG and CEG outperform all other methods at every size (0.79–0.87 with the MLP), while causal Shapley values and conditional SHAP degrade as the number of features grows. A new experiment at d=100d=100 (Appendix F, Figs. F.1–F.2) varies each method's budget: at about 10410^4 evaluations per sample, CEG scores 0.82 against 0.48 for causal Shapley values, and even 128 times the standard budget leaves causal Shapley values below CIG. CIG's advantage therefore holds with a nonlinear model and deeper graphs, and does not depend on the budget given to causal Shapley values.
  • Dyngen. The ground truth is indeed defined by the graph that CIG receives; the experiment asks whether, given a correct graph, a method recovers the drivers, which graph-agnostic methods cannot do. The corruption study (Appendix K) tests the dependence on the graph: removing or reversing backbone edges degrades recovery strongly, while spurious edges barely do.
  • IFN-β\beta. To test whether the SCENIC-based results only reflect overlap between the OmniPath input graph and the SCENIC regulons, we added a control, IG+MP, which moves credit along the same OmniPath graph but without the mechanisms. It ranks the OmniPath roots highest by construction (Fig. I.1), but ranks the SCENIC roots below the graph-agnostic methods (Fig. J.4) and the stimulation-responsive TFs below the other genes (Fig. J.5). Following the input graph alone therefore does not reproduce the SCENIC-based results.
  • DEG recovery. We agree that it is the only graph-independent readout and that CIG scores lower than the other methods on it. DEG recovery is a simple check: a DEG is a gene whose own expression changes, tested gene by gene. CIG moves credit from downstream targets, which are typically the strongest DEGs, to the regulators that drive them. These regulators are rarely DEGs themselves, but they are the causal drivers, which are biologically more informative. Consistent with this, the genes that only CIG ranks highly are enriched for pathways of transcription factors that regulate the interferon-β\beta response (Appendix I.4). Lower DEG recovery is therefore expected, but it means DEG recovery alone cannot confirm CIG's rankings, and the IG+MP control provided above addresses this directly.

Q3 (stimulation as a confounder).

We refit the surrogate on control cells only, so that the stimulation cannot bias the learned mechanisms (new paragraph in Appendix I). With this surrogate, CIG still ranks the stimulation-responsive TFs highest of all methods in the three cell types shown (CD14+ and FCGR3A+ monocytes, dendritic cells), together with CIG fitted on all cells (Fig. J.5). It also ranks the SCENIC roots higher than the graph-agnostic methods do (Fig. J.4), and matches CIG fitted on all cells within the seed-to-seed spread in DEG recovery over the whole panel and the full DEG set (Appendix I.3). Fig. 5's conclusions therefore do not depend on the surrogate having seen stimulated cells.

Q4 (an unchanged child with a moved parent receives direct attribution; Appendix F).

This is intended. If xi=bix_i=b_i while a parent moved, xix_i's residual must have changed to cancel the parent's effect. CIG credits the parent for its effect and xix_i for this counteracting change: if a TF went up while its target stayed unchanged, something outside the model repressed the target. We now state this explicitly in the revision. In practice, the effect is small. On IFN-β\beta, inputs are z-scored, so the baseline is the average expression profile and no gene sits exactly at it. Even with a loose tolerance of 0.1 standard deviations, the 3.4% of genes per cell that are near their baseline receive only 1.4% of the total absolute attribution and make up 2.2% of each cell's top-ranked genes (new Appendix I.5, Table I.13).

We agree that the Appendix F result needs conditions. The revised appendix states the idealized setting it assumes (a linear FF, and input and baseline both satisfying the mechanism exactly), and that otherwise R2R_2 need not be exactly zero, while the part of x2x_2's change explained by x1x_1 is still attributed to x1x_1.

AuthorsOpenReview ↗

Response to Reviewer UQnW (3/3)

(Part 3/3)

Q5 (truncation at HH; dependence on HH).

We agree that there is no reason to truncate. The all-hop attribution costs one sparse triangular solve with (I−J⊤)(I-J^\top) per path point, and our implementation now supports any HH. The all-hop solve costs the same order, O(E)O(E), as computing the surrogate Jacobian, which both versions need, whereas truncating at HH hops costs O(H⋅E)O(H\cdot E) (Appendix D, Table D.1). In practice, on Dyngen, the attribution step takes about as long with all hops as with one hop, and less time than with two hops (Appendix E, Table E.1). To test whether results depend on HH, we repeated the synthetic benchmark on deeper graphs (new Fig. A.1, summarized in Sec. 3). They depend on HH only once graphs are deep. With H=1H=1, the Spearman correlation with the CATE on root nodes decreases as the graph gets deeper, for both models. H=2H=2 recovers most of this gap, results do not change from H=3H=3 on, and all-hop CIG is best or tied at every depth.

Q7 (abstract wording; "α\alpha as residual information").

We agree that this abstract sentence was imprecise. CIG integrates along the same straight line as IG. We revised the abstract to: "Our key idea is to split each feature's movement along the integration path into the part explained by its causal parents and its own residual." α\alpha remains the parameter of the straight-line path; Sec. 3.2 now defines the residual along it, hi(α)=γi(α)−si(γ1(α),…,γi−1(α))h_i(\alpha)=\gamma_i(\alpha)-s_i(\gamma_1(\alpha),\dots,\gamma_{i-1}(\alpha)), with ∂γi/∂α=dhi/dα\partial\gamma_i/\partial\alpha = dh_i/d\alpha being xix_i's own movement.

Reviewer_VE9AOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies feature attribution, when the input features have causal relationships described by a DAG. Building on integrated / expected gradient methods, the authors propose causal integrated gradients (CIG) (and causal expected gradients (CEG)), which account for this causal structure by choosing a particular integration path. In particular, their method uses a surrogate model ss to account for feature dependencies, which treats each variable xjx_j as a linear combination of features of its parents; the path is then defined recursively through this surrogate model.

The experiments include a synthetic benchmark (studying correlation b/t attribution strength and the magnitude of ground-truth causal effects), a simulated biological dataset (where the DAG is a gene regulatory network with transcription factors as source nodes), and a real RNA-sequencing dataset (where the DAG is learned using curated prior knowledge and an existing GRN inference method). Compared to other attribution method (including non-causal methods and causal methods based on Shapley values), the method achieves better performance on the synthetic benchmark correlations, on the simulated dataset AUC and DCG, and on recovery of stimulation-responsive regulators on the real-world dataset.

Critical strengths

The paper addresses a real problem in the attribution literature: how to scale causal attribution methods.

Critical weaknesses

The main weakness of the paper is the incremental nature of the method: given the conceptual basis offered by Causal Shapley values and the computational framework of integrated gradients, the method is quite straightforward to come up with. The applications are interesting, but without more interesting theory or more task-specific methodological innovations, the novelty/originality is too low, especially when accounting for the overall paper abundance at ML conferences.

Questions for authors
  1. Clearer notation and causal interpretation: As a reader coming from the field of causality, the motivation to account for causal dependencies is very natural, and the use of the additive noise surrogate model (or, more generally, a differentiable surrogate model) is well-motivated from a computational perspective. However, I had the feeling that more formalism could make the paper more clear. For example, since RiR_i depends on γ\gamma, it could be useful to build this into the notation (e.g. Ri(γ)R_i(\gamma)). Then, it could be helpful to connect Ri(γ)R_i(\gamma) for the particular "causal" choice of γ\gamma to known quantities in causality / causal mediation analysis, e.g., causal derivatives and causal direct/indirect effects (see e.g., Kawakami and Tian 2026 and references therein). In the response, could you discuss how the method can be viewed from a more formal statistical/causal perspective? In particular, I view RiR_i (and its components Ri,directR_{i,\text{direct}} and Ri,indirectR_{i,\text{indirect}}) as estimands that likely have a rigorous causal interpretation, with the method returning estimates of these quantities (with estimation error from fitting the surrogate model and approximation error from the integration).

  2. Fixed discretization: The choice to use the Riemann sum is unmotivated: since the estimands are 1D integrals, I'd expect better numerical behavior from more sophisticated ODE solvers, as used in the diffusion modeling / flow-modeling literature (see e.g., Karras et al. (2022). I understand that some other feature attribution methods may use the Riemann sum, but (as mentioned by the authors) some feature attribution methods do consider these numerical issues. To reflect these advances, the paper should also use these advances: this could help make the paper less incremental. If time permits, during the rebuttal, could the authors provide results using better integration methods, along with any description of how the use of such methods changes for the causally-informed path γ\gamma described in their paper?

References

Kawakami and Tian (2026), Cumulative natural direct and indirect effects for causal mediation analysis / Karras et al. (2022), Elucidating the design space of diffusion-based generative models

3 comments
AuthorsOpenReview ↗

Response to Reviewer VE9A (1/2)

(Part 1/2)

We thank the reviewer for the careful reading and the pointers to the causal-inference and numerical-integration literature, which helped us sharpen the method's interpretation.

W1 (incremental given causal Shapley values and integrated gradients).

We agree that the motivation is shared with causal Shapley values and the computational framework with integrated gradients, and the simplicity of the resulting method is deliberate: a method that is easy to implement, inspect and verify is more likely to be trusted and adopted, particularly in application domains such as biology. We would like to clarify, however, that the contribution goes beyond combining the two.

Shared motivation, different mechanism. CIG shares the goal of causal Shapley values, explanations that respect the causal structure among the features, but not how it is achieved. Causal Shapley values require sampling from interventional distributions over coalitions of features, which is what limits them to small numbers of features. CIG requires neither: it routes IG's path through a structural causal model, so that each feature is credited for the change in its own residual, using only gradients of the model and the surrogate. In the attribution literature, gradient-based methods have so far assumed independent features, and to our knowledge CIG is the first to incorporate causal dependencies into them.

Dependency-aware explanations at a new scale. As the reviewer notes, scaling causal attribution methods is a real problem in the attribution literature. Existing dependency-aware attribution methods were validated on at most 28 features (Heskes et al. 2020: 7; Aas et al. 2021: 3–10, with one real-data example at 28). At the scale of our biological experiments, they are computationally infeasible and had to be excluded, so dependency-aware explanation was previously out of reach in exactly the settings where causal structure is best known, such as gene regulation. CIG runs on 6,941 genes, and on both biological datasets it ranks the upstream regulators higher than graph-agnostic methods do. Making these explanations available where they could not be applied before is, we believe, a substantive contribution.

Theory. Our primary aim is a theoretically grounded method that is usable in high dimensions. The submission proves the axioms that CIG satisfies (Appendix C), and, following the reviewer's Q1, the revision adds an exact causal characterization (general response, point 2; Q1 below): the all-hop CIG is a path method over the exogenous residuals of a structural causal model, whose gradient is a causal derivative.

AuthorsOpenReview ↗

Response to Reviewer VE9A (2/2)

(Part 2/2)

Q1 (notation and formal causal interpretation).

We thank the reviewer for this perspective, which we found very valuable. We agree and now write attributions as Ri(γ)R_i(\gamma) (Sec. 2). With the exact path derivative (general response, point 1), the all-hop attribution (general response, point 2) can be written as

Ri=∫01∂G∂hi(h(α))dhidαdα,G(h):=F(Φ(h)),R_i = \int_0^1 \frac{\partial G}{\partial h_i}\big(h(\alpha)\big) \frac{dh_i}{d\alpha} d\alpha, \qquad G(h) := F(\Phi(h)),

where h(α)h(\alpha) are the residuals along IG's straight line (Sec. 3.2). As the reviewer suggests, this is an estimand with a causal interpretation: ∂G/∂hi\partial G/\partial h_i is the derivative of FF with respect to an intervention on xix_i, propagated through xix_i's descendants (a unit-level causal derivative, taken at the residuals h(α)h(\alpha) of that path point, with the other residuals held fixed), and dhi/dαdh_i/d\alpha is the change in xix_i's exogenous residual. For linear mechanisms and a linear model, RiR_i is the total causal effect of xix_i on FF times Δhi\Delta h_i, and Ri,directR_{i,\mathrm{direct}} and Ri,indirectR_{i,\mathrm{indirect}} are the natural direct and indirect effects of a shift of xix_i by Δhi\Delta h_i. In the nonlinear case, the direct and indirect terms at each path point are unit-level analogues of the local natural direct and indirect effects of Kawakami and Tian (2026), which add up to the local total effect, and RiR_i integrates them along the path, as in their cumulative effects. Unlike there, all residuals move jointly along the path, so interactions between features are allocated by the path, as in IG; this is why the dependence on γ\gamma matters. The one-hop version derived in Sec. 3.2 shares the direct term. Its indirect term differs in two ways: it follows xix_i's effect only to its children, and it uses xix_i's total movement, so xix_i is also credited for the part of its movement caused by its parents. Both differences vanish on graphs of depth one. We added these derivations and their proofs, and the relation to causal mediation analysis, in a new appendix (Appendix N), with a pointer from Sec. 3.2. CIG returns estimates of these quantities: all estimation error enters through the surrogate, which we study for the graph (Appendix K) and the mechanisms (Appendices F and M), and with the closed-form derivative, the integration error is pure quadrature error (Q2). More broadly, we believe the connection between feature attribution and causal inference is still underexplored, and we see further correspondences, for instance between path methods and cumulative mediation effects, as a promising direction for future work.

Q2 (Riemann sum vs. better integrators).

We used the Riemann sum because it is the standard choice in the integrated gradients literature, but we agree that the choice should be discussed. Unlike in diffusion sampling, the integrand is an explicit function of α\alpha, so each attribution is a 1D quadrature problem rather than an ODE, and ODE solvers reduce to standard quadrature rules (e.g. Heun's method, used by Karras et al. (2022), becomes the trapezoidal rule). The submitted version approximated the causal path derivative by a finite difference between consecutive nodes, which limited the benefit of a higher-order rule. With the exact closed-form derivative of the revision (general response, point 1), any rule applies exactly as for integrated gradients: the causal path changes only what is evaluated at each node, not how the nodes and weights are chosen. On the synthetic benchmark (Appendix C, Fig. C.1), Gauss–Legendre reaches machine precision at K=16K=16 with the true mechanisms, whereas with the learned surrogate, which is only finitely smooth, all rules converge as K−2K^{-2}. At the setting used in the paper (midpoint rule, K=100K=100), the relative quadrature error is about 3×10−63\times10^{-6} and 10−510^{-5}, respectively. Since the benefit of a higher-order rule depends on the smoothness of the model and the surrogate, the new version of our software, developed alongside this work, will let the user choose the rule, as Captum does.

I thank the authors for the thoughtful responses, and I commend the significant improvements.

Responses

W1: I agree with the authors' point that "Making these explanations available where they could not be applied before is... a substantive contribution", and that the simplicity of the method has its merits (especially in terms of trust and adoption). While I somewhat stand by the original point (I think ML conferences increasingly need a high bar on theoretical/methodological novelty), I also see the authors' perspective. Ultimately, based on the additional connections/improvements (better efficiency by using the closed-form derivative, better causal interpretation, new experimental results that add scientific value), I am satisfied with the novelty of the paper.

Q1: Thank you for investigating this point and sharing the results. The "CIG = IG over exogenous residuals" is, in retrospect, fairly obvious, and I'm glad to see it makes the derivations simpler. I find the causal interpretation/connection more interesting: the linear case is satisfyingly intuitive, and I appreciate how the nonlinear case shows why the dependence on γ\gamma is important. I agree that further correspondences are a promising direction for future work; by introducing this correspondence, the potential impact of the paper has also been increased.

Q2: Thanks for acknowledging and explaining. First, I agree that the quadrature language is a better fit for the 1D case. Further, I appreciate that the software will now let the user choose the rule, and I appreciate that the revision reports (low) integration error.

Summary

My questions and the weaknesses were addressed to my satisfaction; I also enjoyed seeing the additional experiments and improvements in response to the other reviewers. In light of this, I've updated my score to 3 (Weak Accept).

Comments on the paper

2 comments
AuthorsOpenReview ↗

General response to all reviewers (1/2)

(Part 1/2)

We thank all reviewers for their detailed and constructive reviews. We first summarize the main changes, which address points raised by several reviewers, and then answer each reviewer individually. We have uploaded a revised version of the paper in which all text changes are marked in blue.

1. Exact path derivative (Reviewer UQnW).

Reviewer UQnW correctly observed that on the straight-line path we use, the path derivative that we estimated by a finite difference (Eq. 6 of the submission; Eq. 7 in the revision) has a closed form:

∂γi∂α=(xi−bi)−∑k<i∂si∂xk(xk−bk),i.e.dhdα=(I−J)(x−b),\frac{\partial\gamma_i}{\partial\alpha} = (x_i-b_i) - \sum_{k<i} \frac{\partial s_i}{\partial x_k} (x_k-b_k), \qquad\text{i.e.}\qquad \frac{dh}{d\alpha} = (I-J)(x-b),

where hi=xi−si(x1,…,xi−1)h_i = x_i - s_i(x_1,\dots,x_{i-1}) is the residual and Jji=∂sj/∂xiJ_{ji}=\partial s_j/\partial x_i the surrogate Jacobian, which we already compute for the indirect term. We have replaced the finite difference by this expression. As a result, (a) at every path point, CIG's integrands summed over features equal the integrand of integrated gradients, so CIG's completeness error equals that of integrated gradients for any quadrature rule and number of steps (Appendix C, Table C.1), and (b) causal expected gradients no longer needs several path points per baseline: like expected gradients, one point per baseline gives an unbiased estimate. On the synthetic benchmark, CEG with 100 single-point draws matches CIG with K=100K{=}100 in accuracy at a comparable cost (Sec. 4.1), and on Dyngen its attribution step now takes about 200 s instead of 18.4 h (Appendix E, Table E.1).

While making this change, we noticed that CIG and integrated gradients were not integrated with the same rule (a Riemann sum for CIG, Captum's default Gauss–Legendre rule for IG). Both now use the midpoint rule with the same number of steps, so their completeness errors can be compared directly: they agree to float32 rounding on the synthetic benchmark (to within 10−1310^{-13} in float64) and agree on the biological data for every CIG variant except one (Appendix C, Table C.2), where the difference is consistent with float32 rounding: along long chains of TFs, the all-hop indirect attributions can become large and cancel almost entirely in their sum.

AuthorsOpenReview ↗

General response to all reviewers (2/2)

(Part 2/2)

2. An exact characterization: CIG is integrated gradients over exogenous residuals (Reviewers VE9A, UQnW).

With the closed form, the attribution that follows dependencies to any depth (rather than truncating at H=1H=1 or 22 hops) is

Ri=∫01[(I−J)−⊤∇F]idhidαdα.R_i = \int_0^1 \big[(I-J)^{-\top}\nabla F\big]_i \frac{dh_i}{d\alpha} d\alpha .

It costs one sparse triangular solve per path point, i.e. one backward pass through the surrogate, so the same order as the one-hop version. We show that this is exactly a path method in the sense of Sundararajan et al. (2017), applied to the composed model G(h):=F(Φ(h))G(h) := F(\Phi(h)), where Φ\Phi is the surrogate structural causal model that maps the exogenous residuals hh to the features xx (by recursively computing xi=si(xpa(i))+hix_i = s_i(x_{\mathrm{pa}(i)}) + h_i). Indeed, ∇hG=(I−J)−⊤∇F\nabla_h G = (I-J)^{-\top}\nabla F by implicit differentiation, and dh/dαdh/d\alpha is the velocity of the residuals along the path. This gives CIG a precise meaning: it attributes the change in the model output to the changes in each feature's exogenous residual, weighted by that feature's total causal effect on the output. Consequences:

  • Completeness, implementation invariance, linearity and dummy hold with respect to GG and the residuals, inherited from the path-method structure rather than proven case by case. In particular, the "dummy" axiom that we previously reported as violated holds in residual space: a residual receives zero attribution if neither its feature nor any descendant affects the output.
  • For linear mechanisms and a linear model, RiR_i is the total causal effect of xix_i on FF times the change in xix_i's exogenous residual, and the direct and indirect terms are the natural direct and indirect effects (Appendix N; Reviewer VE9A, Q1).
  • The one-hop version used in the submission is a truncation of this quantity; it is complete and its axioms hold (Appendix C), but, as Reviewer UQnW noted, it is not itself a path method, and we have revised Sec. 3.5 accordingly.

The main text keeps the one-hop derivation and introduces the HH-hop and all-hop versions briefly; their derivation and this characterization are given in Appendices A and N. On deeper synthetic graphs, two hops recover most of the gap between one hop and all hops, and results do not change from three hops on (Sec. 3.1, Appendix A); on Dyngen, two hops likewise recover most of this gap (Fig. 4A). All experiments, including the graph-corruption study (Appendix K), were rerun with the closed-form derivative, the midpoint rule and, unless stated otherwise, the all-hop attribution, and the synthetic benchmark now uses deeper graphs, so most numbers differ from the submission. The main conclusions are unchanged: on the synthetic benchmark, CIG and CEG match the causal effect best at a fraction of the cost of causal Shapley values, and on Dyngen and IFN-β\beta they best recover the driving transcription factors.

3. New experiments.

  • Graph-aware ablations (Reviewer UQnW). To test whether CIG's advantage comes from graph access, we built two ablations of CIG, neither of which is an existing method: integrated gradients with its credit propagated naively to parents along the graph, without the mechanisms (IG+MP), and DeepSHAP with its credit propagated through the mechanisms with CIG's own propagation step (DeepSHAP$+J$). IG+MP fails on all three benchmarks, whereas DeepSHAP$+J$ matches CIG on the synthetic benchmark and recovers most of its advantage on the biological data (Appendix J). The gain therefore comes from propagating credit through the mechanisms, CIG's central component, not from graph access. CIG is the principled form of this propagation: a path method that inherits completeness and implementation invariance.
  • Harder synthetic benchmark (Reviewer UQnW): deeper graphs, a nonlinear MLP model, and the same causal Shapley budget at every size up to 1000 features. CIG's advantage on root nodes holds at every size, and causal Shapley values stay below CIG even with 128 times the standard budget, also when both use the same learned causal model (Reviewer fvrG).
  • Confounding (Reviewer fvrG): a synthetic sweep over latent-confounder strength and a hidden batch variable in Dyngen that the model uses as a shortcut (Appendix L), and on IFN-β\beta a surrogate refit on control cells only (Reviewer UQnW). Attribution quality degrades gradually with confounding strength, and the biological conclusions are unchanged.
  • Quadrature rules (Reviewer VE9A): at the setting used in the paper, the integration error of the attributions is about 10−510^{-5} or less (Appendix C).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.