AI 용어집
인공지능 완전 사전
Causal Bayesian Networks
Directed acyclic graphs explicitly modeling causal relationships between variables, where edges represent causal mechanisms and nodes represent variables, enabling interventions and counterfactual calculations.
Propensity Score Matching with ML
Use of machine learning algorithms (gradient boosting, random forests) to more accurately estimate propensity scores in high-dimensional contexts, thus improving covariate balance between treated and control groups.
Instrumental Variables with Neural Networks
Approach integrating neural networks to model complex non-linear relationships between instrumental variables, treatment, and outcome, overcoming limitations of traditional linear IV methods.
Causal Inference Trees
Modified decision trees specifically designed to identify subpopulations with heterogeneous treatment effects, using splitting criteria based on treatment-covariate interaction tests rather than outcome purity.
Neural Causal Models
Deep neural network architectures designed to learn invariant representations and causal mechanisms, incorporating structural constraints to ensure causal interpretability and robustness to distribution shifts.
Counterfactual Regression
Representation learning method using neural networks to learn balanced embeddings where treated and untreated covariate distributions are indistinguishable, facilitating estimation of individual causal effects.
Meta-learners for CATE
Algorithmic framework including S-learner, T-learner, X-learner and R-learner that use basic ML models as building blocks to flexibly and non-parametrically estimate conditional average treatment effects.
Orthogonal Random Forest
Extension of random forests incorporating orthogonalization principles to reduce bias in causal effect estimation, particularly effective in the presence of high-dimensional confounders with non-linear effects.
Structural Causal Models with Deep Learning
Integration of deep neural networks into the structural equations of SCMs to capture complex non-linear causal relationships while preserving interpretability and counterfactual reasoning capability.
Causal Discovery Algorithms
Algorithms (PC, FCI, GES, NOTEARS) using machine learning to automatically infer causal structure from observational data, often with sparsity constraints and non-parametric independence tests.
Doubly Robust Neural Estimators
Causal estimators combining neural networks to model outcomes and propensity scores, ensuring consistency if either model is correctly specified, particularly useful in high dimensions.
Invariant Risk Minimization
Learning principle seeking representations where outcome predictors remain invariant across different environments, serving as a proxy for learning causal mechanisms robust to distribution shifts.
G-computation with ML
Parametric causal estimation method using advanced ML models (gradient boosting, neural nets) to approximate the conditional distribution of outcomes, allowing simulation of interventions and estimation of marginal causal effects.
Heterogeneous Treatment Effect Discovery
Process using ML algorithms to automatically identify subgroups with significantly different treatment responses, combining clustering and effect estimation to personalize interventions.