This paper introduces an innovative approach to dramatically accelerate UMAP using spectral data compression.The proposed method significantly reduces the size of the dataset, preserving its essential manifold structure through an advanced spectral compression technique. This allows UMAP to perform much faster while maintaining the quality of its embeddings. Experiments on real-world datasets, such as USPS, demonstrate the method's ability to achieve substantial data reduction without compromising embedding fidelity.

翻译：本文提出一种利用谱数据压缩显著加速UMAP的创新方法。所提方法通过先进的谱压缩技术，在保持数据集本质流形结构的同时，显著缩减数据规模。这使得UMAP能够在保持嵌入质量的前提下实现更快的计算速度。在USPS等真实数据集上的实验表明，该方法能在不损害嵌入保真度的前提下实现显著的数据压缩。