MeSH Hierarchy Ranking: GBM Features and Poincaré Coordinates

Explicit hierarchy features and hyperbolic coordinates provide separate structural ranking evidence.

RI.MV.003·Maintained reference·Revised 2026-08-06

MeSH terms sit in a tree of broader and narrower concepts. Ranking can use that structure as extra evidence. A gradient-boosted model, which combines many small decision rules, can read plain tree facts. Poincaré coordinates, a way of placing a tree in curved space so deep branches have room, capture the same parent-child structure as measured positions.

tree depthancestor overlapGBMPoincaré disk

Geometry: Hierarchies expand faster than Euclidean space

SpaceRepresentationBoundary
EuclideanEqual physical distance represents equal vector distance.Deep branches compete for finite area.
PoincaréSpace grows toward the disk boundary.Distance represents structure, not word meaning.

Explicit hierarchy features: Auditable ranker inputs

FeatureQuestion answered
DepthHow specific is the candidate?
Ancestor overlapDo candidates share a branch?
Path distanceHow far apart are the concepts?
Sibling densityDoes a crowded branch require stronger separation?

Model sequence: Add structural evidence in measured stages

Inference: The deterministic index defines the candidate pool

ordinary query → normalized words, phrases, and identifiers → associated MeSH candidates → hierarchy features → GBM scores → current descriptors

Controlled comparison: One evidence layer changes at a time

ModelEvidenceExcluded evidence
Frequency formulaPair frequency, lift, phrase weight, cross-source bonusLearned ranking and hierarchy
GBM plus hierarchyFrequency, depth, ancestors, branch, relatives, positionsLearned graph coordinates
GBM plus PoincaréAll explicit features, coordinates, centroid distance, radial depth, separationNo additional evidence layer

Poincaré signals: Coordinates supplement explicit tree facts

  • Candidate coordinates represent location in the learned graph.
  • Token-associated centroid distance compares a candidate with query-linked terms.
  • Nearest-candidate distance represents local structural density.
  • Radial position approximates depth.
  • Cross-branch separation distinguishes distant subtrees.