The HNSW construction time may be wrong in the paper

[ciuji]

Hi,

Thanks for your great work! It is really a comprehensive conclusion and clearly demonstration.
But I found some inconsistent results about the HNSW construction from other papers and my own experiment.
In the paper, Figure 5 shows that HNSW has a very long construction time, which may be 10-100 times slower than the best method. But in other papers' experiments (e.g. NSG), HNSW usually just would be 2-3 times slower than the KGraph. My own experiment also proved that HNSW is not that slow.
Besides, the parameter selection for HNSW seems not very reliable. I try 400 ef_construction and 800 ef_construction, the 800 one increases the indexing time while not significantly improving the query quality.
So maybe the index experiment for HNSW may have some issues.

Thanks,
Chaoji