KI-Glossar
Das vollständige Wörterbuch der Künstlichen Intelligenz
Causal node
Variable represented in a causal graph that can influence or be influenced by other variables through directed arcs. Each node corresponds to an entity or concept in the studied system.
Causal arc
Directed arrow connecting two nodes in a DAG representing a direct cause-and-effect relationship between the corresponding variables. The orientation of the arc indicates the direction of causal influence.
Causal structure
Global configuration of nodes and arcs in a causal graph that captures the complete set of causal relationships in the system. The structure determines the conditional independence constraints observable in the data.
d-separation criterion
Graphical test allowing to determine whether two sets of variables are conditionally independent given a third set in a DAG. d-separation is crucial for identifying causal effects that can be identified from observational data.
Structural equations
System of mathematical equations describing how each variable is generated from its direct causes and independent error terms. Structural equations provide a quantitative model of causal mechanisms.
SCM (Structural Causal Model)
Formal framework combining a causal graph, structural equations, and a distribution over exogenous variables to completely model a causal system. SCMs enable the simulation of interventions and counterfactuals.
Collider
Structure in a DAG where two arrows converge on the same node, indicating that this node is causally influenced by two distinct variables. Colliders create conditional dependencies when controlled for.
Causal chain
Sequence of variables where each directly influences the next, forming a continuous causal path in the DAG. Causal chains are essential for understanding mediation mechanisms.
Fork
Causal structure where a common variable influences two other variables, creating a statistical association between them. Forks represent classic sources of confounding in causal analysis.
Causal ancestor
Variable that can reach another variable by following directed arcs in a DAG, representing a direct or indirect cause. Ancestors are crucial for identifying potential confounding paths.
Causal descendant
Variable accessible from another variable by following directed arcs in a DAG, representing a direct or indirect effect. Descendants must be avoided as adjustment variables to prevent introducing bias.
Moral graph
Transformation of a DAG where unconnected parents of the same node are linked and all arrows are replaced by undirected edges. The moral graph is used for inference in Bayesian networks.
Pearl's causality
Formal theory of causality based on DAGs and do-calculus, developed by Judea Pearl to unify causal inference. This approach hierarchizes knowledge from association to intervention and then to counterfactual.