이미지: METAL LAB 생성
Summary
- Apple ML Research has presented a method for directly analyzing the k-nearest neighbor (kNN) graph that UMAP builds internally
- By applying network science algorithms such as PageRank, k-core decomposition, and clustering coefficient, it identified representative data points, dense regions, and similar clusters
- In experiments on MNIST and Fashion MNIST, the approach produced results comparable to dedicated techniques like k-medoids and HDBSCAN
- 발행
- Apple ML Research, 2026년 7월
- 논문 제목
- Dimensionality Reduction Meets Network Science: Sensemaking on UMAP's kNN Graph
- 저자
- Duen Horng (Polo) Chau, Donghao Ren, Fred Hohman, Dominik Moritz
- 적용 알고리즘
- PageRank, k-core 분해, 클러스터링 계수
- 평가 데이터셋
- MNIST, Fashion MNIST
- 비교 대상 기법
- k-medoids(대표점 선정), HDBSCAN(밀도 기반 군집화)
What was announced
Apple ML Research has released a paper that shines new light on a hidden byproduct of the data visualization tool UMAP. Most people who use UMAP only look at the resulting 2D scatter plot. But this paper focuses on something UMAP builds internally before it draws that plot: a k-nearest neighbor (kNN) graph—a network in which each data point is connected to its closest neighbors. The researchers applied three techniques to this graph: PageRank (an algorithm once used to rank web pages, here used to pick out representative data points), k-core decomposition (which separates dense core regions from sparse peripheral areas), and clustering coefficient (which identifies tightly interconnected groups of similar neighbors). Tested on the MNIST and Fashion MNIST datasets, the method produced results comparable to—and sometimes complementary to—k-medoids, a technique for selecting representative points, and HDBSCAN, a density-based clustering method.
What this means
UMAP (Uniform Manifold Approximation and Projection) is a technique that compresses data with hundreds or thousands of dimensions—things like image features or sentence embeddings—into a 2D map that humans can visually inspect. It's one of the most widely used tools among data scientists for finding clusters and patterns in large datasets. The problem is that the compression process introduces distortion. Points that were close together in the original high-dimensional space can end up looking farther apart once squeezed into 2D, and conversely, unrelated points can end up placed near each other by chance. But before UMAP performs this compression, it first builds a graph representing which neighbors each data point is actually close to. Until now, this intermediate output was treated as a mere computational byproduct and largely discarded. This research recognized that the graph itself preserves the original relational information without distortion, and demonstrated that applying already-established network science algorithms directly to it can yield useful insights.
Why it matters
What makes this approach compelling is that it requires no additional computation. Until now, understanding data has meant running separate algorithms for picking representative points, separate algorithms for splitting clusters, and separate algorithms for measuring density. This research showed experimentally that a single graph—one that's already generated the moment UMAP runs—can substitute for much of that work. This is still an early validation stage, tested on relatively simple datasets like handwritten digits (MNIST) and clothing images (Fashion MNIST), and further validation is needed to see how well it generalizes to more complex real-world data. Even so, for researchers and developers working with data visualization tools, this effectively adds a new layer of analysis available for free within a tool they're likely already using.
