Benchmark Methodology

Core-SG performance should be interpreted as a build-once, extract-many workflow. The initial build can be more expensive than a single HDBSCAN run, but repeated extraction for multiple k values can amortize that cost.

The exact algorithm="core-sg" path and the approximate algorithm="score-sg" path should be interpreted separately. Exact CoreSG still has a dense pairwise construction step, which creates a practical n_samples limit. ScoreSG is evaluated as the scalable approximate path that avoids that dense all-pairs construction and is therefore the relevant method when sample size becomes the limiting factor.

The benchmark material in benchmarking/ evaluates repeated multi-k workloads over synthetic datasets. The central comparison is cumulative time for all requested k values, not only the first result.

The report assets are copied into the documentation by docs/scripts/generate_benchmark_figures.py.