Exploratory Causal Inference in SAEnce

Core task: unsupervised causal effect discovery from high-dimensional experimental data. The field addresses the challenge of identifying causal relationships and treatment effects when labeled outcome data or explicit intervention annotations are scarce or absent, particularly in settings where the dimensionality of covariates is large. The taxonomy reflects a broad landscape organized around six main branches. Causal Structure Discovery from Observational Data focuses on learning directed acyclic graphs and structural equation models from passive observations, often leveraging independence constraints or functional form assumptions. Treatment Effect Estimation and Causal Inference encompasses methods for quantifying intervention impacts, including both supervised approaches that rely on known treatment assignments and unsupervised variants that must infer effects without direct outcome labels. Semi-Supervised and Unlabeled Causal Inference explores hybrid settings where partial supervision or positive-unlabeled data guides discovery. Causal Inference Under Distribution Shift tackles robustness when training and test distributions differ, while Causal Feature Selection and Discovery aims to identify causally relevant variables in high-dimensional spaces. Specialized Causal Inference Methods includes domain-specific techniques and novel algorithmic frameworks, such as those integrating large language models or leveraging representation learning. Several active lines of work highlight key trade-offs and open questions. One prominent theme is the tension between model flexibility and identifiability: representation learning approaches like Identifiable Causal Representation[11] and Latent Invariant Mechanism[29] seek to uncover latent causal factors from complex observations, yet must impose structural constraints to ensure uniqueness. Another contrast appears between methods that assume known interventions versus those that operate in fully unsupervised regimes. SAEnce Causal Inference[0] sits within the Unsupervised Causal Effect Discovery cluster, emphasizing the extraction of causal signals from experimental data without explicit outcome labels. It shares this unsupervised orientation with Correlation to Causation[4], which also aims to move beyond associational patterns, and contrasts with semi-supervised strategies like Semi-supervised Misspecification[5] that blend labeled and unlabeled information. The original work's focus on high-dimensional experimental settings positions it at the intersection of structure discovery and effect estimation, addressing scenarios where traditional supervised methods are infeasible yet experimental perturbations provide crucial leverage for causal identification.