PaCMAP
Algorithms AIMech:PaCMAP DRAFT
Aliases: Pairwise Controlled Manifold Approximation Projection
A nonlinear dimensionality-reduction algorithm that embeds neighbor, mid-near and farther point pairs using separate attractive and repulsive losses whose relative weights change through three optimization phases.
Identity Notes & Concept Scope
PaCMAP uses three classes of pair losses and a changing phase schedule, rather than TriMap’s coupled triplet loss or UMAP’s fuzzy graph construction. The graph-sampling rules and phase weights are defining components, not a proof of universal local/global preservation. All 442 pinned AIO labels/scoped synonyms and 114 preceding native identities were compared, together with hidden/ignored-inclusive searches of both checkouts; source citations and pending frontier mentions were distinguished from existing concept identities.
Claims & Grounded Evidence
1 Claim
Mechanism
#method-scope
Choose neighbors from a Euclidean-neighbor candidate pool using locally scaled distances, choose mid-near partners as the second closest among six sampled observations, and sample non-neighbors for farther pairs. With d̃=||y_a−y_b||²+1, use d̃/(10+d̃) for neighbors, d̃/(10000+d̃) for mid-near pairs and 1/(1+d̃) for farther pairs. Optimize with Adam, ordinarily from PCA initialization, first emphasizing mid-near attraction, then reducing it and finally setting its weight to zero while refining local structure.
⚠️ Curated Limitations & Negative Results
- Random sampling and the initial coordinates affect the result. The authors explicitly state that robustness to initialization is incomplete and that initialization scale matters.
- A mid-near partner is selected by a six-point sample rule, not an exact global distance quantile; the method does not guarantee that every such partner is outside the nearest-neighbor set.
- Locally scaled distances select neighbor pairs, while optimization uses embedded distances and pair-class weights; the original distances are not fitted as an all-pairs metric.
- The three-phase procedure seeks a useful nonconvex layout. No global optimum, exact topology preservation, density preservation or universal performance advantage is asserted.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:DimensionalityReduction
Dimensionality Reduction
|
related_match | A specific algorithm for constructing lower-dimensional representations, associated with the pinned dimensionality-reduction task. The target inherits UnsupervisedLearning under MachineLearningTask; related_match avoids treating the algorithm as identical to or a subclass of that task. Shared embedding components do not make this concept identical to PCA, MDS or t-SNE. |
Evidence Sources & Primary Literature
1 Source| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
original-paper
|
Yingfan Wang, Haiyang Huang, Cynthia Rudin, Yaron Shaposhnik
|
primary paper
full_text
|
ID: https://jmlr.org/papers/v22/20-1061.html
sha256:570ab26e1584…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-11T05:21:17Z
Updated: 2026-10-11T05:21:17Z
Source File:
data/curated/PaCMAP.yaml