Home / Curated / Algorithms / HAC-ML network clustering

HAC-ML network clustering

Algorithms AIMech:HACML DRAFT

Aliases: HAC-ML

A greedy network-clustering algorithm that builds a hierarchical stochastic block structure by merging groups with the least loss of profiled network likelihood, using separate Bayesian criteria to collapse terminal groups and stop top-level merging.

Identity Notes & Concept Scope Compared all 442 pinned AIO and 666 native IDs, names, aliases and scoped synonyms for the three candidates and their publication identifiers; no normalized identity match. Ignored- and hidden-inclusive fixed-string searches of the recovery, older discovery and original Documents checkouts returned 26, 0, 0 lines. All matches were read and are prior source/discovery context, with no existing maintained identity. Semantic assessment covered 44 AIO and 260 native definitions: 41 AIO and 228 native rows were unchanged from earlier reading; three additional AIO and 32 additional native definitions were read. Fourteen complete annotated AIO node/edge/generated contexts and the candidate-specific native records were inspected. Retain the 2011 profiled network-likelihood merge guide with separate Bayesian top/terminal selection. ROCK counts shared neighbors with a size normalization, graph degree linkage multiplies directed degree averages, and CliXO selects persistent informative cliques. Generic HAC, a fitted interactome and the HAC-E/ES/Q score variants are distinct scopes.

Claims & Grounded Evidence

6 Claims

Mechanism #profile-likelihood-guide

For undirected, unweighted graphs without self edges, the method models edges and holes between vertex groups with Bernoulli probabilities. It begins with singleton groups and a full block model at the top level. At each step it selects a pair with greatest profiled-likelihood merge ratio, which depends on their connectivity to other groups. This greedy guide is neither a globally optimal tree search nor posterior averaging over trees.

supports Source: Resolving the structure of interactomes with hierarchical agglomerative clustering · Locator: Methods, PDF pp. 2–3, Equations 1–5 and greedy pseudocode; Conclusions, p. 9

Mechanism #bayesian-boundaries

Uniform-prior integration of a block probability gives a Beta factor. A separate integrated-probability merge ratio supplies the proposed top-level stopping rule. When both children are already collapsed, another integrated ratio compares a single terminal group with their separate within- and between-group structure. The paper explicitly omits multiplicity factors for alternative merges and splits; these are stated model-selection conventions rather than a posterior calculation over all partitions.

supports Source: Resolving the structure of interactomes with hierarchical agglomerative clustering · Locator: Methods, PDF pp. 2–4, Equations 2, 7–8 and model-selection qualifications

Mechanism #multiple-edge-types

For multiple interaction types, each type retains its own likelihood factors under a shared group structure. The type-specific likelihoods multiply, so their log scores add. This conditional independence and shared-structure construction differs from taking the union of all edges and does not guarantee that adding another network improves predictions.

supports Source: Resolving the structure of interactomes with hierarchical agglomerative clustering · Locator: Extension to multiple edge types, PDF p. 4; Synergy in mixed networks, pp. 8–9 and Table 3

Limitation #printed-score-boundary

The singleton mismatch expression printed in the prose is not numerically equal to the ratio obtained from Equations 2 and 5: one differing external adjacency gives a profiled ratio of one quarter, whereas the prose expression gives one half. The record uses the general profiled-likelihood construction and preserves this printed discrepancy. Its effect on author code or empirical results was not checked.

supports Source: Resolving the structure of interactomes with hierarchical agglomerative clustering · Locator: Methods, PDF pp. 2–3, Equations 2 and 5 and singleton-mismatch paragraph; curator calculation in research/discovery/hacml-boundaries.py

Scope #complexity-and-search

The paper contrasts a naive quartic implementation with a faster implementation using a priority queue, lazy evaluation and a restriction to candidate pairs sharing a neighbor. The stated faster complexity depends on those choices and network sparsity. Merge ties, excluded candidates and irreversible greedy decisions require implementation conventions; no universal global-optimum or posterior-coverage guarantee is claimed.

supports Source: Resolving the structure of interactomes with hierarchical agglomerative clustering · Locator: Greedy agglomerative algorithm, PDF pp. 3–4; Conclusions, p. 9

Scope #evaluation-scope

The link-prediction study hides observed edges and evaluates them against an equal-sized random set of holes, reporting rankings and maximum F scores. The authors acknowledge that this balanced test overstates practical performance on sparse networks. They prune isolated and degree-one vertices and restrict joint-network comparisons to shared vertices. Reported advantages are empirical across the inspected networks; Table 3 does not show a consistent benefit from combining interaction types.

supports Source: Resolving the structure of interactomes with hierarchical agglomerative clustering · Locator: Performance Evaluation and Results, PDF pp. 4–8, Tables 1–3 and Figures 1–4

⚠️ Curated Limitations & Negative Results

This is Park and Bader’s 2011 HAC-ML network procedure, not the generic family of hierarchical agglomerative clustering methods or a fitted interactome. The separate HAC-E, HAC-ES and HAC-Q scores are not aliases.
The input-network graph records observed interactions; it is not by itself a conditional-dependence graph, a causal graph or a Bayesian network. The grouping model assumes Bernoulli edge trials conditional on its structure.
The implementation complexity claim uses restricted candidate search. The printed top-level product notation and singleton-score shorthand do not provide a completely unambiguous executable specification; no author-code conformance is asserted.
Missing edges are used as holes for fitting and sampled negatives for evaluation, although biological data can contain unobserved true links. The balanced held-out test is not a calibrated estimate of deployment precision.
No author implementation, experimental dataset or supplementary archive was audited or executed. Finite curator calculations assess printed formulas and bounded examples, not benchmark replication or global optimality. The record remains DRAFT.

Artificial Intelligence Ontology (AIO) Grounding

Mapped AIO Classes

AIO Target Class	Mapping Relation	Mapping Rationale
`AIO:HierarchicalClustering` Hierarchical Clustering	related_match	The algorithm constructs a hierarchy of object groups. The target is a related task, so this mapping does not assert algorithm/task equivalence or subclass membership.
`AIO:UnsupervisedClustering` Unsupervised Clustering	related_match	The grouping procedure uses network relationships rather than training a classifier on object labels. Domain-informed input selection or GO-dependent evaluation does not establish identity with a supervised classifier.

Evidence Sources & Primary Literature

1 Source

Source ID	Publication / Document	Type / Access	Reference & Checksum
`park2011`	Resolving the structure of interactomes with hierarchical agglomerative clustering 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: 2026-10-14T01:04:43Z

Updated: 2026-10-14T01:04:43Z

Source File: data/curated/HACML.yaml