dire-evoc Clustering Method Update

leland_mcinnes · x · 2026-07-11

A clustering method dire-evoc has been fully tested in RAPIDS 26.06, with code, examples, and full benchmarks provided. The method builds a layout first, making it heavier, but this layout helps when clustering in the original space easily misses structure.

The author provides benchmarks on NVIDIA A10: under matched data and base controls, dire-evoc improves clustering scores on OptDigits, PenDigits, SatImage, Image Segmentation; matches on several synthetic datasets; and slightly underperforms pure EVoC on MNIST. Conclusion: performance depends on dataset, but it can capture more important structure in some scenarios.

Original post →

More from Research

Research channel →