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-β 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.

Forest-weighted causal correlation: A scalable framework for causal inference in ecological time series
S. Kishi
bioRxiv1 Oct 2026PreprintCC BYFull textEcologybioRxiv
A dependency-free, streamable format and cross-language toolkit for scalable LC-MS feature detection: reading only what you need
J. Osorio Mosquera, N. G. Lawler, M. Cox +8
bioRxiv15 Sep 2026PreprintCC BYFull textBioinformatics & computational biologybioRxiv
Scalable phylodynamic inference of mutational and epistatic fitness effects with multi-feature birth-death models
S. Cen, L. M. Kepler, C. Lanzas +2
bioRxiv19 Oct 2026PreprintCC BY-NCFull textEvolutionary biologybioRxiv
EEG-based Alzheimer’s disease detection using two-branch multi-scale deep learning with integrated gradients-guided channel selection
Jungpil Shin, Najmul Hassan, Abu Saleh Musa Miah +1
Scientific Reports20 Oct 2026Peer-reviewedCC BY 4.0Full textNeuroscienceCrossref
An integrated instance segmentation and deep feature fusion model for robust leaf species classification in natural environments
K. Arunkumar, S. Leninisha
Frontiers in Plant Science15 Oct 2026Peer-reviewedCC BY 4.0Full textPlant biologyCrossrefEurope PMCEurope PMCPMC

Cite

BibTeX RIS

BibTeX

@misc{michiels2026causal,
  title = {{Causal Integrated Gradients: Scalable Feature Attribution Along Causal Paths}},
  author = {Joran Michiels and Anis Ismail and Maarten De Vos and Alejandro Sifrim},
  howpublished = {arXiv},
  year = {2026},
  month = oct,
  doi = {10.48550/arxiv.2610.22126},
  eprint = {2610.22126},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.22126},
  note = {CC BY 4.0},
}

RIS

TY  - UNPB
TI  - Causal Integrated Gradients: Scalable Feature Attribution Along Causal Paths
AU  - Joran Michiels
AU  - Anis Ismail
AU  - Maarten De Vos
AU  - Alejandro Sifrim
PY  - 2026
DA  - 2026/10/15
PB  - arXiv
DO  - 10.48550/arxiv.2610.22126
UR  - https://arxiv.org/abs/2610.22126
AB  - 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-β 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.
ER  -

Abstract

Related

Cite