causal feature attribution
Latest papers
9 papers3 papers in the last four weeks, with none the four weeks before. 0.0% of all new papers.
In regression and classification tasks with tabular datasets, most of the literature on explanation in ML (e.g. SHAP, permutation importance) has focused on predictive attribution: does feature help predict outcome ? At its heart, predictive attribution captures correlation, not causation. We instead ask a causal question: would manipulating feature change outcome ? We define a natural quantity to answer this question, the feature intervention effect: . In general, estimating this quantity from observational data alone would entail strong causal assumptions. We propose FADO (Feature-wise Amortized Do-Operators), an amortized inference method using prior-fitted networks trained on synthetic tabular datasets generated by structural causal models. During training, the synthetic data-generating process is known, so we can compute ground-truth intervention effects , and train a model to predict them at test time from a purely observational tabular dataset. As we demonstrate through extensive experiments, our new causal explainability method is easy to train and provides better explanations under a range of settings for which standard predictive attribution fails, including proxy variables, confounders, colliders, and placebos. We evaluate it on real, synthetic and semi-synthetic datasets, with and without hidden confounders, and compare it to existing predictive attribution methods and state-of-the-art causal inference methods.
Dropping Tokens Without Dropping Them: Causal Attribution for Diabetic Retinopathy
In this work, we investigate whether attention-based attribution maps in a vision transformer provide causal explanations for diabetic retinopathy (DR) classification. We introduce a principled decomposition of attention into two attribution maps: the contribution (), which satisfies the completeness axiom in explainable artificial intelligence (xAI), and the pull (), which quantifies causal explanations. We define an identity relating and that characterizes when fails to capture causality. We test whether each map is causally faithful on the MAPLES-DR dataset by dropping/inserting tokens, and report Necessity (), Sufficiency (), and Faithfulness Correlation (). The pull achieves the highest causal alignment (, , ), outperforming Attention Rollout, Grad-CAM, and AttnLRP. Occlusion retains better sufficiency () but is about 800 times slower. Furthermore, on MAPLES-DR, achieves of the average precision (AP) of HRDecoder, a state-of-the-art supervised lesion segmentation model, despite the absence of pixel-level supervision. Together, these findings show that the pull and the contribution form computationally efficient attribution maps for reliable causal interpretation in safety-critical medical diagnostics. The code and demo are available at https://github.com/pramath-haritz/dr-xai.
Kinematic Causal Attribution: Black-Box Explanations of Motion versus Geometry for 4D Point-Cloud Video Models
4D perception models, which classify 3D point-cloud videos, are being built into robots, assistive monitoring and AR, yet before deploying one a practitioner cannot check which cue it relies on: how a body part moves, or what shape it has. 3D attribution is single-frame, video attribution is confined to RGB pixels, gradient maps need white-box access, attention is absent from state-space backbones, and point-level perturbations assume a point correspondence across frames that real depth and LiDAR data lack. We present Kinematic Causal Attribution (KCA), a black-box, interventional attribution method that queries only the model's outputs. Its freeze operator removes a region's motion while keeping its shape, and a shape-free proxy removes the shape while keeping the motion. Together they span a motiongeometry cube for every tracked region, whose order-symmetric Shapley decomposition is complete and whose interaction is measured rather than assumed away. The operators rewrite per-frame point sets, so they need no correspondences, and a context-averaged estimator handles structures that span several regions. We validate KCA on the official PSTNet, P4Transformer and Mamba4D trained on MSR-Action3D. The decisive test is an audit: two binary tasks are built from real clips whose classes differ only in the motion or only in the shape of the same arm, and the motion share of the arm's importance follows the defining cue on all three backbones (0.83 vs. 0.60 on P4Transformer, 0.86 vs. 0.30 on PSTNet, 0.92 vs. 0.61 on Mamba4D), a distinction no single attribution map can make. The split costs no faithfulness: the necessity maps have the lowest deletion AUC on P4Transformer and PSTNet, tie with the best baseline on Mamba4D, and lose their correlation with the trained model's once the weights are randomised. Skeleton ground truth, re-sampling stability and cost complete the validation.
HullProbe: Measuring Interventional Part Reliance in Maritime Vessel Classifiers
Recognizing a vessel's type from an image supports maritime surveillance and sanctions monitoring, but a classifier can reach high accuracy by leaning on the generic hull instead of the more distinctive superstructure, and is then fooled when a vessel is re-hulled or occluded. De-biasing methods claim to remove such reliance, yet the field certifies them with worst-group accuracy, a proxy that needs group labels and only infers reliance from accuracy. We measure the reliance directly. HullProbe swaps a registered anatomical part of a vessel for another instance's and reads how much that part drives the decision, using a same-class swap to control for compositing artifacts. Because a surface vessel is a hull and a superstructure arranged along one axis, its parts form a 2x3 grid whose 2^6 = 64 swaps we evaluate exhaustively, and a single Walsh transform returns each part's interventional influence and its interaction spectrum. On planted shortcuts, where the true reliance is known, the measure tracks worst-group accuracy (r = -0.94 over 45 configurations), and replacing its learned segmenter with human annotations changes the hull-share by less than the audit's resolution. On the 26 vessel types of MARVEL, the classifier places about 46% of its corrected sensitivity on the generic hull yet follows the superstructure when the two conflict, and the group-label-free de-biasers available there do not detectably reduce the hull's share. The findings hold across three architectures and a second maritime dataset, and the method extends to fine-grained vehicle recognition on Stanford Cars.
Spatial Attention Noise Masking for Causally Sufficient Interpretability
We present a novel causal approach to interpretability for computer vision models that dynamically masks the input image prior to classification. The interpretability of deep learning predictions is critical in high-stakes fields such as medical imaging, security, and autonomous driving. Most interpretability methods are applied passively to already trained models, which typically result in correlational rather than causal explanations. Existing causal interpretability methods are limited to post hoc analysis, weakening the causal claims. Additionally, existing active methods generally lack explanations that explicitly assign responsibility to input features. This work proposes a spatial attention noise masking framework that provides causal explanations about the features sufficient for the prediction. The proposed framework consists of: 1) a UNet-style mask generator, and 2) a Resnet18 encoder and linear classifier that classifies both masked and unmasked versions of an input image. The generated masks are regularized to be sparse and spatially smooth, while masked image embeddings are constrained to remain consistent with embeddings from the corresponding unmasked images. The resulting masks can be interpreted as feature attribution maps that are competitive with related interpretability methods while additionally providing strong causal explanations of model predictions. Quantitative evaluations demonstrate mask faithfulness, near-baseline classification performance across five classification tasks despite substantial masking of image information, and robustness to distribution shifts such as background swapping and natural adversarial examples. Qualitative comparisons further demonstrate mask behavior and competitive interpretability relative to state-of-the-art feature attribution methods.
From Training to Deployment: Post-Hoc Causal Feature Identification via Sensitivity Ratios
Given a model that is already trained, which features does it rely on causally versus spuriously? Existing methods require access to the training procedure and cannot answer this post-hoc. We introduce the Normalised Sensitivity Ratio~(NSR), a post-hoc, model-agnostic diagnostic for this question under a structured-shift regime: environments differ primarily in the mean of spurious features while the causal mechanism and causal marginals remain stable, as in multi-site clinical data or multi-batch genomics. Within this regime, causal features induce constant model sensitivity across environments while spurious features track shift. NSR formalises this as the squared coefficient of variation of per-environment sensitivity. Under a linear structural causal model (SCM) with non-degenerate environments, NSR achieves exact identification (Theorem~1). We fully characterise failure: weak shifts ( collapse), degenerate geometry, and proxy attenuation (), giving practitioners quantitative criteria for assessing whether the regime holds. Finite-sample rates are under the null and under the alternative. Experiments confirm all theoretical predictions on synthetic data (area under the ROC curve [AUROC] under conditions satisfying the regime), show consistent rankings across five model families (Kendall ), and recover six of eight causal features on bike-sharing data (Precision@7 ) without modifying any trained model.
Faithfulness to Refusal: A Causal Audit of Neuron Selectors
Attribution scores increasingly identify which neuron rows of a language model matter for applications such as pruning, interpretability, and editing for safety, yet whether they identify causally important rows is rarely tested directly. We address this with two paired audits built on one-shot neuron-row zeroing. We first audit selectors at the language-modeling level: attribution methods substantially outperform activation and magnitude-based baselines at identifying dispensable rows across five LLMs. We then adapt the same intervention into a behavior test by driving it with a contrastive harmful-versus-benign signal; the attributed rows are sufficient to install refusal on hate and crime while keeping benign over-refusal low and preserving language model fluency, and specific in that layer-matched random controls at the same depths fail. Highly rank-stable selectors can be among the least causally valid. Refusal moreover lives in a redundant subspace, where different attribution methods install it through largely disjoint row sets, so the recovered edit is one realization of a sufficient set rather than a unique mechanism. Together, these findings show that rank-stability proxies miss the kinds of selector failures a direct causal audit can surface.%
Beyond Shapley: Efficient Computation of Asymmetric Shapley Values
We address the problem of explainability in machine learning models through feature attribution methods. In particular, we consider a variant of Shapley values known as Asymmetric Shapley Values (ASV), which enables the incorporation of causal knowledge into model-agnostic explanations through the use of a causal graph. We show that in certain contexts in which the computation of SHAP is -hard, the exact computation of ASV can be done in polynomial time. To extend this algorithmic result, we introduce a notion of equivalence classes over the topological orderings of the underlying causal graph, which is useful to reduce the time to compute ASV. In particular, we present a polynomial-time algorithm (in the number of equivalence classes) to compute it whenever the causal graph is a rooted directed tree. Finally, we develop an algorithm for approximating ASV in arbitrary causal DAGs which relies on a procedure to sample topological orderings uniformly at random. To implement this sampling mechanism we leverage known algorithms as well as simpler alternatives. Our experimental results demonstrate the practical viability of the proposed approach in realistic causal structures.
AG-REPA: Causal Layer Selection for Representation Alignment in Audio Flow Matching
REPresentation Alignment (REPA) improves the training of generative flow models by aligning intermediate hidden states with pretrained teacher features, but its effectiveness in token-conditioned audio Flow Matching critically depends on the choice of supervised layers, which is typically made heuristically based on the depth. In this work, we introduce Attribution-Guided REPresentation Alignment (AG-REPA), a novel causal layer selection strategy for representation alignment in audio Flow Matching. Firstly, we find that layers that best store semantic/acoustic information (high teacher-space similarity) are not necessarily the layers that contribute most to the velocity field that drives generation, and we call it Store-Contribute Dissociation (SCD). To turn this insight into an actionable training guidance, we propose a forward-only gate ablation (FoG-A) that quantifies each layer's causal contribution via the induced change in the predicted velocity field, enabling sparse layer selection and adaptive weighting for alignment. Across unified speech and general-audio training (LibriSpeech + AudioSet) under different token-conditioning topologies, AG-REPA consistently outperforms REPA baselines. Overall, our results show that alignment is most effective when applied to the causally dominant layers that drive the velocity field, rather than to layers that are representationally rich but functionally passive.