Detail: Vertex Checkerο
The Vertex Checker, with the first vertex selected.ο
The Vertex Checker takes the Markov check one step further. Where the Markov Checker tells you whether a graph is Markov to your data, the Vertex Checker localizes the trouble to particular vertices and then offers to repair the graph by searching for edge edits that reduce the violations.
It is the tool for the situation where a search has returned a graph that mostly looks right but fails a Markov check, and you want to know which part of it is responsible.
Find it in a Compare box, under Independence Facts Comparison, as Vertex Checker.
The two tabsο
Check β the per-vertex view of the Markov check. For each vertex, the independence facts the graph implies at that vertex are tested against the data, so you can see where the violations are concentrated rather than only how many there are in total.
Repair β the local search described below.
Repairο
Repair is a greedy local search over edge edits, scored by the same Markov-check diagnostics the Check tab reports. For the selected vertex it enumerates single-step edits β additions, removals and reorientations β that are legal for the graph type you have chosen, applies the one that most reduces the violations, and repeats.
This is a search, not a derivation. It finds a graph that fits the data better in the Markov-check sense; it does not prove that graph is right, and a repair that removes violations by removing information is always possible. Treat the result as a hypothesis to examine, and look at what was changed rather than only at the improved score.
Controlsο
Graph type β which class the repaired graph must stay in: DAG, CPDAG, PDAG, MAG or PAG. This is not a display choice; it determines which edits are legal and how the candidate is normalized afterwards. DAG and CPDAG assume causal sufficiency. MAG and PAG admit latent confounders. CPDAG and PAG are canonicalized back to their class after every edit, so non-compelled orientations may be rewritten; PDAG is the one that preserves orientations exactly as edited, which is what you want if you are deliberately holding an orientation fixed.
Node β the vertex to work on. There are two buttons, each naming the selected node: one searches for candidate adjustments at that vertex, and one repairs that vertex only. On a large graph, repairing a single vertex of interest is much faster than a whole-graph sweep and is often all you want.
Attempt Repair runs the full repair.
Strategy β Local sweep goes node by node, repeating the sweep until nothing improves. Global queue keeps a priority queue across all vertices and always applies the best candidate available anywhere next. Global queue generally reaches a better graph; local sweep is easier to follow, since it works through the vertices in order.
Prune alpha β the significance level for the independence tests that score candidates. As everywhere, a smaller alpha means fewer rejections, so fewer violations to repair and a more conservative repair.
Seed β the search makes arbitrary choices among ties, so it is seeded for reproducibility. Change the seed and repeat if you want to know whether a particular repair is stable or an artifact of tie-breaking.
Undo steps back through applied edits, and Graph shows the current working graph. A long run can be stopped with Cancel, which may take a moment, since it stops between tests.
Two things worth knowing about the searchο
Candidate scoring for PAGs is done in two passes, and this does not affect what gets applied. For PAG repair under the ordered-local-Markov conditioning types, candidates are first screened cheaply β projected only as far as the implied MAG, with violations counted there β and the expensive full canonicalization is skipped for candidates that are merely being ranked. Because all MAGs in one equivalence class imply the same independence facts, screening is sound for ranking; but the screening graph is generally not the canonical member of its class, so screening counts can differ slightly from fully canonical ones. The consequence is confined to the order in which candidates are tried. Every edit that is actually applied is re-scored on its fully canonicalized, knowledge-refined graph first, and must beat a fresh no-op on those exact scores, so the approximation cannot reach the working graph.
Candidate enumeration is capped. For the DAG, CPDAG and PDAG types the search also enumerates multi-edge orientation patterns around a vertex, and that enumeration is bounded. A vertex with more free edges than the cap allows falls back to a smaller set of pattern moves, so the search is less thorough there. As with the determinism searches in the Data Audit, the absence of a proposed repair at a high-degree vertex does not mean none exists.