Using Proximity Graph Cut for Fast and Robust Instance-Based Classification in Large Datasets

Table 3

Algorithm speedup comparison with respect to different graph densities (tuned by NSW hyperparameter) and percentile. Speedup values compared to baseline HNSW-based 1-NN classification are provided for percentile (p99), percentile (p95), and average classification time (Avg).

M	NSW-PATH			HNSW-PATH
	p99	p95	Avg	p99	p95	Avg
4	2.44	2.17	2.85	5.59	4.55	4.16
8	1.56	1.71	2.12	4.07	4.36	4.34
16	1.47	1.42	1.84	4.31	4.18	4.45
32	1.55	1.55	2.10	4.51	4.14	4.62
64	1.32	1.26	1.61	6.44	6.50	7.10
128	1.45	1.42	1.82	4.11	3.18	5.30