Constraint and Score-Based Discovery

Routing Summary

Recovering causal structure (DAG / CPDAG / PAG) from observational data, from Glymour, Zhang & Spirtes (2019), “Review of Causal Discovery Methods Based on Graphical Models.”

Concept Map

NoteTypeRole
Causal Discovery - OverviewoverviewGoal of structure learning; three method families; Table 1 comparison; practical issues
Markov and Faithfulness Assumptionsdefinitiond-separation, local Markov, causal Markov + faithfulness, Markov equivalence class / CPDAG
PC Algorithm and Constraint-Based DiscoveryconceptCI tests, skeleton, collider orientation, Meek rules; FCI, PAGs/MAGs for latent confounders
GES and Score-Based DiscoveryconceptGreedy two-phase search over equivalence classes, BIC; GFCI hybrid
Functional Causal Models (LiNGAM, ANM)conceptLiNGAM, ANM, PNL; identify direction via noise independence asymmetry

Notes

  • Causal Discovery - Overview — CONTAINS: definition of causal discovery / DGCMs; goal (recover DAG/CPDAG/PAG from observational data); the three families (constraint / score / FCM); Table 1 method comparison; practical challenges (time series, measurement error, selection bias, missing data).
  • Markov and Faithfulness Assumptions — CONTAINS: paths/colliders/d-separation; local Markov condition; Causal Markov assumption; Causal Faithfulness assumption (no extra independencies); Markov Equivalence Class and CPDAG; equivalence invariants (skeleton + v-structures).
  • PC Algorithm and Constraint-Based Discovery — CONTAINS: adjacency criterion; PC skeleton discovery (growing conditioning sets, separating sets); v-structure orientation rule; Meek orientation propagation; CPDAG output; FCI for latent confounders (PAG/MAG, bidirected edges, “o” marks); RFCI/GFCI.
  • GES and Score-Based Discovery — CONTAINS: score functions (BIC); GES forward (FES) and backward (BES) phases; search over equivalence classes; convergence to same MEC as PC; weaker-than-faithfulness condition; GFCI hybrid for confounders.
  • Functional Causal Models (LiNGAM, ANM) — CONTAINS: general FCM , noise-independence asymmetry test; LiNGAM (linear non-Gaussian, ICA/DirectLiNGAM, Darmois–Skitovich); ANM (nonlinear additive noise); PNL (post-nonlinear, most general); identifiability beyond the equivalence class; limitations.

Sources