Fig. 8.

ChEMBL ($n=1.7M$) k-nearest neighbor searches performance of 2048-D MHFP6 indexed using LSH Forest and 2048-D ECFP4 indexed using Annoy Recovery rates for both implementations depend on parameters $k$, $k_c$, and $l$. a While LSH Forest performs better for $k=5$ and $k=10$ nearest neighbors, Annoy surpasses LSH Forest for $k=50$ and $k=100$. b By increasing the number of nearest neighbors by a factor of $k_c$, the performance of both ANN neighbor methods can be greatly improved. While LSH Forest (orange) shows worse performance compared to Annoy (green) for $k_c<20$, it surpasses Annoy for higher values. c Increasing the number of trees $l$ increases the recovery rate for both methods at the expense of main memory. Annoy performs slightly better for $l=8,\dots ,128$, performance of LSH Forest increases at a greater rate, overtaking Annoy at $l=256$. d, e Increasing values of parameters $k_c$ and $k$ affects query times of Annoy negatively. While the average query time for LSH Forest remains below 100 ms for $k=50$ and $k=100$, Annoys average query time increases to above 100 and 200 ms respectively. f As the number of prefix trees, and thus the recovery rate, in LSH Forest increases, the query time decreases. On the other hand, an increase in Annoy trees, with a beneficial effect on recovery rate, also increases the query time. For subplots a, d; b, e; and c, f; the data has been aggregated over all measured values for $k_c$, $l$; $k$, $l$; and $k_c$, $k$; respectively