TriMap
Algorithms AIMech:TriMap DRAFT
A nonlinear dimensionality-reduction algorithm that selects weighted triplet comparisons from high-dimensional data and optimizes embedding coordinates to favor the corresponding ordering of pairwise similarities.
Identity Notes & Concept Scope
TriMap denotes the sampled triplet embedding algorithm, not a three-valued segmentation trimap, generic ordinal embedding, its PCA-based global score, or PaCMAP’s pair losses. Version updates to sampling and weights are retained as method variations, not silently mixed with the older experiments. 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
2 Claims
Mechanism
#method-scope
Sample neighbor-versus-farther triplets and additional randomly selected triplets ordered by input proximity. Weight triplets using input-distance differences and the version-specific smoothing rule. Minimize the sum of weighted s(i,k)/(s(i,j)+s(i,k)) losses with s(i,j)=1/(1+||y_i−y_j||²), commonly starting from scaled PCA coordinates and using momentum-based full-batch optimization. The inspected revision describes shifted tempered-log weights and updated triplet counts separately from older experimental settings.
Limitation
#initialization-and-version-limit
Wang et al. challenge attributing global layout solely to the triplet loss: their analysis emphasizes initialization and the limited range of attractive and repulsive forces. Their inspected TriMap definition uses 55 triplets per point and the older exponential/log weight rule, so this qualification is version-scoped rather than a claim about all later code.
⚠️ Curated Limitations & Negative Results
- Triplet sampling, weighting, initialization and input metric influence the layout; minimizing the sampled objective does not guarantee satisfaction of every possible triplet or recovery of a unique manifold.
- The v2 PDF says its displayed experiments use older code. It updates counts from 10×5+5 to 12×4+3 and replaces the earlier weight transformation; these are not results for a single identical software version.
- Wang et al.’s initialization critique studies the older 55-triplet construction and should not be treated as a direct benchmark of every later revision.
- The separately proposed global score is based on linear reconstruction relative to PCA; it is not a universal geometric-fidelity certificate, and its displayed normalization requires nonzero PCA residual error.
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
2 Sources| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
original-paper
|
Ehsan Amid, Manfred K. Warmuth
|
primary paper
full_text
|
ID: arXiv:1910.00204
sha256:e2c2c2d33b7e…
|
initialization-analysis
|
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/TriMap.yaml