New CosMAP method improves dimensionality reduction for complex data

Researchers have developed CosMAP, a new unsupervised dimensionality-reduction method designed to create faithful and interpretable embeddings for complex, high-dimensional datasets. CosMAP extends the UMAP framework by incorporating cosine-similarity neighborhoods and contrastive affinities, optimized through an attractive-repulsive objective. The method includes a two-phase refinement process, first learning an intermediate high-dimensional representation to reconstruct the neighborhood graph and initialize the final low-dimensional embedding. Evaluations on datasets including handwritten digits, single-cell RNA sequencing data, and genealogical information demonstrate that CosMAP produces more coherent visual representations, better neighborhood preservation, and clearer global organization compared to existing state-of-the-art methods. AI

IMPACT Offers a more robust framework for exploratory analysis of complex, sparse, high-dimensional data.

RANK_REASON The cluster contains a research paper detailing a new method for dimensionality reduction. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv
• BALSAC-CARTaGENE
• MNIST database
• Uniform Manifold Approximation and Projection

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →