KI-Glossar
Das vollständige Wörterbuch der Künstlichen Intelligenz
Directed Acyclic Graph (DAG/Causal Graph)
Graphical representation of causal relationships where nodes are variables and directed edges indicate direct causal influences. DAGs cannot contain cycles, reflecting the temporal asymmetry of causal relationships.
d-separation
Graphical criterion for determining whether two sets of variables are conditionally independent in a directed causal graph. Forms the foundation of do-calculus for identifying blocked or open causal paths.
Causal identifiability
Property determining whether a causal effect can be uniquely computed from observational data and the causal graph. Identifiability depends on the graph structure and the set of observable variables.
Confounding
Phenomenon where a third variable simultaneously influences both the treatment and the outcome, creating a non-causal association between them. Confounding is the main obstacle to causal inference from observational data.
Causal mediator
Intermediate variable in the causal chain between treatment and outcome, transmitting the treatment effect. Mediation analysis allows decomposing the total effect into direct and indirect effects.
Causal collider
Structure where a variable is influenced by two other variables, creating a conditional dependence between the latter. Colliders can block or open causal paths depending on which variables are conditioned upon.
Causal fork
Structure where one variable influences two other variables, creating an association between the latter when the common cause is conditioned upon. Forking can introduce selection bias in causal analysis.
Do-operator
Mathematical operator representing an intervention that forces a variable X to take value x, independently of its natural causes. The do-operator is central in Pearl's formalism for distinguishing correlation from causation.
Potential Outcome Framework
Approche alternative à l'inférence causale basée sur les résultats potentiels (ou contrefactuels) pour chaque unité sous différents traitements. Fournit un cadre théorique pour définir les effets causaux même lorsque seul un traitement est observé par unité.