Causal integrated gradients find true feature causes efficiently
Causal Integrated Gradients: Scalable Feature Attribution Along Causal Paths
Machine LearningArtificial Intelligence
Summary
Finding which parts of complex data really cause effects is hard, especially when features depend on each other. The authors propose a new method called causal integrated gradients that uses known cause-and-effect information among features to better identify true drivers. This method follows the cause-effect chains in data to separate direct and indirect influences of each feature. They tested it on synthetic, simulated biological, and real biological data and showed it points to real upstream causes better than existing methods, while still being efficient to compute.
What this means in practice
- •For bioinformatics teams: Identify key upstream genes driving biological responses from complex gene expression data using known causal gene networks.
- •For machine learning engineers: Improve explainability of models with interdependent input features by incorporating causal structure into feature attribution.
Authors
Joran Michiels, Anis Ismail, Maarten De Vos, Alejandro Sifrim
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.