unit-level causal attribution

In causal effect estimation, the absence of standardized hyperparameter tuning evaluation criteria impedes reliable model selection and creates a substantial gap between commonly used metrics and true performance. This paper systematically investigates the interplay between hyperparameter tuning and evaluation, jointly analyzing estimators (T-/X-/R-Learner), base learners (random forests, gradient boosting, neural networks), and evaluation metrics (IPW, DR, PEHE) across four benchmark datasets. Key findings are: (1) thorough hyperparameter tuning eliminates performance differences among mainstream causal estimators; (2) the choice of evaluation strategy exerts greater influence on final performance than either the estimator type or base learner architecture; and (3) existing evaluation metrics underestimate the performance gain from optimal model selection by over 35% on average. These results demonstrate that hyperparameter tuning is the primary determinant of causal estimation accuracy, underscoring an urgent need for more robust, theoretically grounded evaluation paradigms in causal machine learning.

This work addresses the challenge of efficiently extracting sparse causal abstractions from pretrained neural networks that maintain high interventional fidelity, without resorting to brute-force interventions or retraining. By framing structured pruning as a search for approximate causal abstractions, the authors model the network as a deterministic structural causal model and introduce a closed-form pruning criterion derived from a second-order Taylor expansion of interventional risk. Under a unified curvature assumption, this criterion reduces to the activation variance method, while also clarifying the conditions under which that heuristic fails. Empirical validation via interventional swapping demonstrates that the extracted abstractions exhibit strong causal faithfulness, enabling efficient and interpretable discovery of causal structure.

The absence of genuine interventions in observational data impedes reliable distinction between causal and spurious features. To address this, we propose Feature-Matching Intervention (FMI), a novel framework that constructs a causal latent graph in the representation space and simulates perfect intervention via mechanism matching—enabling intervention-free identification of causal features. Our key contributions are: (1) the first intervention paradigm grounded in feature matching; (2) formal definition of a causal latent graph to uniformly encode structural causal relationships in latent space; and (3) theoretical guarantees for strong out-of-distribution (OOD) generalization. FMI synergistically integrates causal representation learning, matching estimation, and latent-variable structural causal models (SCMs), operating solely on observational data. Empirically, it achieves significant improvements in causal feature identification accuracy across multiple OOD benchmarks, consistently outperforming state-of-the-art methods.

Identifying intervention targets in single-cell biology—i.e., inferring the set of perturbed variables from combined observational and interventional data—remains challenging due to small sample sizes, high dimensionality, and violations of ideal causal assumptions. Method: We propose Causal Differential Networks (CDN), a novel end-to-end joint training framework that unifies noisy causal graph inference, graph-structural difference modeling, and multi-source feature supervision to jointly optimize for causal interpretability and prediction robustness. Contribution/Results: Evaluated on seven real single-cell transcriptomic datasets and diverse synthetic intervention scenarios, CDN consistently outperforms state-of-the-art baselines. It achieves substantial improvements in both soft and hard target prediction accuracy, offering an interpretable, high-precision computational paradigm for drug target discovery and cellular engineering.

This paper addresses causal effect estimation under network interference, where treated units are connected via a known network, and the interference radius and strength are unknown, varying heterogeneously across local topologies and treatment assignments—rendering the classical no-interference assumption invalid. We propose a synthetic-control-based feature-decoupling method featuring a novel neighborhood-adaptive enumeration mechanism that dynamically identifies the optimal interference radius for each local treatment pattern. Integrating network topology modeling with asymptotic distribution theory, we establish convergence rates and asymptotic normality of the estimator. Simulation and empirical studies demonstrate that our method significantly outperforms fixed-radius benchmarks in estimating the average direct treatment effect (ADTE) on the treated. Our work delivers the first scalable, theoretically grounded, and data-driven framework for modeling network interference.

This work addresses the problem of automatically discovering high-level causal abstraction models that accurately capture interventional behaviors directly from low-level observational data, without relying on expert-specified priors. Building upon a low-rank causal discovery assumption, the authors propose an end-to-end learnable, unsupervised method that jointly integrates structural causal modeling with representation learning to identify high-level latent variables and infer their causal structure from data. The theoretical analysis establishes, for the first time, a formal connection between low-rank graph-generating mechanisms and causal abstractions, proving the identifiability and interventional validity of the learned high-level variables. Experimental results demonstrate that the proposed approach reliably recovers high-level structural causal models, offering a novel paradigm for automated causal abstraction.

This study addresses the challenge of evaluating causal effects on multivariate, latent, and interdependent outcomes—such as brain effective connectivity—by proposing a two-stage causal inference framework. First, effective connectivity is estimated from high-dimensional neuroimaging time series; then, inverse probability weighting combined with multiple testing correction is employed to assess the causal impact of external interventions. The method innovatively incorporates sample splitting to mitigate bias arising from internal dependencies among outcomes and integrates causal modeling with rigorous control of multiple hypothesis testing. Theoretical analysis and simulations demonstrate that the approach achieves asymptotic validity while effectively controlling both Type I error and family-wise error rates. Application to ADNI data successfully uncovers the causal influence of amyloid burden on brain effective connectivity.

This study addresses the challenge of identifying and estimating causal effects under network interference, where an individual’s treatment may spill over and affect others’ outcomes. The authors propose a solution based on a linear outcome model that yields unbiased and consistent estimates of both binary and continuous treatment effects when the interference structure is known or partially known. The approach accommodates both fixed and random interference network specifications and innovatively eliminates interference-induced bias while remaining compatible with standard linear regression software. It also conveniently allows for the incorporation of random effects and heteroskedasticity- and autocorrelation-consistent (HAC) standard errors. Numerical simulations and empirical analyses demonstrate the method’s effectiveness in bias correction and practical applicability.

Standard diffusion models lack the capacity to model causal structures, rendering them unsuitable for interventional sampling and causal inference. This work proposes a causal graph–guided conditional diffusion mechanism that embeds a known directed acyclic graph into the diffusion process for the first time. By appropriately propagating interventional signals during reverse-time sampling and employing resampling to construct null distributions for edge testing, the method enables valid causal discovery. Theoretical analysis guarantees convergence of distribution estimation and control of Type I error in edge testing. Experiments demonstrate that the approach more accurately recovers interventional distributions in simulations, achieves nominal significance levels with high statistical power in edge tests, and successfully validates contested signaling pathways in flow cytometry data.

This study addresses the challenge of distinguishing causal features from spurious ones in pre-trained models without access to training dynamics. The authors propose Normalized Sensitivity Ratio (NSR), a post-hoc, model-agnostic diagnostic that identifies causal features by evaluating the stability of feature sensitivities across environments under structured distribution shifts. Grounded in linear structural causal models, NSR constructs a theoretically justified metric based on the coefficient of variation of sensitivities, enabling precise identification of causal features without any training information and providing formal characterization of its failure conditions. Empirical results demonstrate near-perfect performance on synthetic data satisfying model assumptions (AUROC = 1.000), strong consistency in feature rankings across five diverse model architectures (Kendall’s τ ≥ 0.529), and high precision in real-world settings (Precision@7 = 0.75 on bike-sharing data).