AI 용어집
인공지능 완전 사전
DAG (Directed Acyclic Graph)
Graphical representation of causal relationships where nodes are variables and directed edges indicate cause-to-effect relationships without closed cycles.
Block randomization
Experimental method where subjects are grouped into homogeneous blocks before random assignment of treatment, reducing variability and increasing statistical power.
Instrumental variable
Variable correlated with treatment but not directly with the outcome, used to identify causal effects in the presence of unobserved confounding variables.
Back-door criterion
Sufficient condition for identifying a causal effect by adjusting on a set of variables that block all non-causal paths between treatment and outcome.
Front-door criterion
Alternative to the back-door criterion allowing causal identification when an observable mediator blocks all direct paths between treatment and outcome.
Markov equivalence
Principle establishing that conditional independences observed in data correspond to separations in the causal graph representing the data-generating process.
Average Treatment Effect (ATE)
Expected mean difference between potential outcomes with and without treatment in the entire population, fundamental measure of causal effect.
Average Treatment Effect on the Treated (ATT)
Average causal effect calculated specifically on the subpopulation that actually received the treatment, relevant for policy evaluation.
Difference-in-Differences Method
Quasi-experimental identification strategy comparing changes in outcomes before and after treatment between treatment and control groups.
Regression Discontinuity
Causal identification method exploiting treatment thresholds where treatment assignment changes abruptly while other factors vary continuously.
Structural Causal Inference
Causal analysis paradigm based on explicit structural models of relationships between variables, enabling distinction between correlation and causation.
Structural Causal Model
Mathematical formalization of underlying causal mechanisms describing how variables generate other variables, foundation of modern causal analysis.
Potential Outcome
Conceptual framework defining for each unit the potential outcomes under each treatment level, even though only one is observable in practice.