A new paper examines how covariate balance diagnostics, a tool borrowed from causal inference, can reveal hidden confounding or model misspecification when offline reinforcement learning is used to recommend treatments. The author argues these checks help assess whether learned treatment policies rest on valid assumptions. The work targets researchers applying RL to clinical or policy decision data.
Researchers present a multi-head feed-forward neural network that jointly estimates conditional average treatment effects and the full distribution of those effects. The approach is framed as a generative learner for distributional causal effects, aimed at capturing heterogeneity beyond single-point estimates. It is a preprint posted to arXiv.
A new arXiv paper relaxes the standard no-interference assumption used in conformal counterfactual prediction, where one unit's treatment is assumed not to affect another's outcome. The authors develop an approach that produces prediction sets with finite-sample coverage guarantees for counterfactual outcomes and individual treatment effects when units interact within a network. This matters for settings such as social networks, marketplaces, and trials where spillover effects are common.
A new arXiv preprint introduces Generative Marketing Mix Modeling, a causal inference approach for estimating how exposure to a brand's name inside AI-generated answers affects business outcomes. The method links generative engine optimization and generative engine marketing metrics to sales impact, since conventional marketing datasets do not capture how often users see or notice a company's name in generated responses.
A new arXiv paper formalizes a comparison between causal structures that share the same observed variables. One structure is said to observationally dominate another when the distributions it can generate over the visible variables include all those the other can produce. The work studies the resulting partial order and its implications for因果 inference when hidden (latent) factors are present.
Researchers describe a technique for inferring the moral graph of a causal directed acyclic graph using only the probability distribution over a set of discrete variables. The method represents the distribution with fully connected tensor networks and applies nuclear-norm regularization to the bond correction matrices to guide recovery. The paper adds a tensor-based approach to the toolkit for causal structure learning.
A researcher proposes a technique for estimating average treatment effects in observational studies while preserving differential privacy. The approach uses propensity score blocking to group similar subjects, limiting how much any individual's data influences the result. The preprint is posted on arXiv and is categorized under machine learning.