HAC-ES network clustering
Algorithms AIMech:HACES DRAFT
Aliases: HAC-ES
A hierarchical network-clustering variant that guides agglomeration by the sum of cross-group edge density and Jaccard overlap of the groups’ external neighboring vertices.
Identity Notes & Concept Scope
Compared all 442 pinned AIO and 669 native IDs, names, aliases and scoped synonyms for the three candidates and publication identifiers; no normalized identity match. Ignored- and hidden-inclusive fixed-string searches of the recovery, older discovery and original Documents checkouts returned 158, 0 and 5 lines. All 158 public matches were inspected and belong to source/discovery/review context and rendered existing records, not these maintained identities. The five original-checkout matches are unrelated Spanish hace text in ignored dependencies. Semantic assessment covered 46 AIO and 264 native definitions: 44 AIO and 260 native rows were unchanged from prior reading; two additional AIO and four additional native definitions were read. Sixteen complete annotated AIO node/edge/generated contexts and the candidate-specific native records were assessed. Retain the source-named HAC-ES variant: edge density plus Jaccard overlap of external neighboring vertices. ROCK uses cross-cluster shared-neighbor link counts and a different size normalization. HAC-ML uses profiled network likelihood, and neither HAC-E nor HAC-Q has this combined score.
Claims & Grounded Evidence
5 Claims
Mechanism
#combined-merge-guide
The guide adds the cross-group edge density to a shared-neighbor term. That term counts outside vertices connected to both groups and divides by the number connected to either group. Vertices in the two candidate groups are excluded, and each outside vertex contributes a Boolean presence indicator rather than its number of incident edges.
Scope
#ranking-and-probability-boundary
The unweighted sum can change candidate-pair rankings relative to edge density alone and can exceed one. A constructed pair with edge density one half and complete external-neighbor overlap scores three halves, exceeding a pair with density one and no overlap. The combined score is therefore a merge heuristic, not a Bernoulli edge probability.
Limitation
#empty-neighborhood-boundary
If neither candidate group has an external neighbor, the printed shared-neighbor expression has zero numerator and denominator. The paper does not state a numerical convention for this case. Candidate restrictions or code might handle it, but no such behavior was inspected or assumed here.
Scope
#hierarchical-link-estimation
The method belongs to the paper’s hierarchical link-prediction comparison, where densities at tree joins estimate missing links. The shared-neighbor merge heuristic and those subsequent density estimates play different roles. Improved agreement of local scores with common neighbors does not itself establish biological identity or universal predictive improvement.
Scope
#evaluation-scope
The study compares this named variant with HAC-ML and other methods on biological interaction networks using held-out-edge prediction. Positive test pairs are balanced by an equal number of sampled holes, and maximum F scores summarize ranked predictions. This selection changes the negative prevalence relative to deployment. The reported comparisons support a setting-specific empirical assessment, not universal superiority or calibrated real-world link probabilities.
⚠️ Curated Limitations & Negative Results
- The external-neighbor term is a Jaccard ratio over vertices, not over the current cluster labels or a sum of all shared-neighbor paths. It is distinct from ROCK’s link-count normalization.
- The printed sum supplies no fitted mixing weight. Its two components have different information content despite each lying between zero and one when defined.
- This record covers a named 2011 score variant in Park and Bader’s network-clustering study, not a fitted interactome or the entire family of agglomerative clustering methods.
- The paper defines each alternative by replacing the merge guide. It does not separately specify every candidate-restriction, tie, stopping and terminal-collapse convention for each variant. The detailed HAC-ML implementation and its complexity are not automatically asserted for this variant.
- The experiments use undirected, unweighted interaction graphs with iterative removal of isolated and degree-one vertices. Missing observed edges can include unobserved biological interactions; sampled holes are not established biological negatives.
- No author implementation, data or supplementary archive was audited or executed. Finite constructed checks assess local scores and algebraic distinctions, not complete clustering runs, benchmark replication or global optimality. DRAFT remains.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:HierarchicalClustering
Hierarchical Clustering
|
related_match | This algorithm constructs a hierarchy of vertex groups. The target is a related task, not an equivalent algorithm or asserted superclass. |
AIO:UnsupervisedClustering
Unsupervised Clustering
|
related_match | The procedure groups graph vertices using observed relationships rather than supervised label prediction. The task relation does not establish algorithm identity. |
Evidence Sources & Primary Literature
1 Source| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
park2011
|
Yongjin Park, Joel S Bader
|
primary paper
full_text
|
ID: doi:10.1186/1471-2105-12-S1-S44
sha256:e930abd762f0…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-14T01:20:34Z
Updated: 2026-10-14T01:20:34Z
Source File:
data/curated/HACES.yaml