Why is `lastfm-64-dot` called `-dot`?

[thomasahle]

From the path http://ann-benchmarks.com/lastfm-64-dot_10_angular.html it seems that this dataset is actually angular.
But the name indicates dot-product, which many of the algorithms don't natively support.

[erikbern]

yeah this is very confusing – I think it's a mistake. https://github.com/erikbern/ann-benchmarks/blob/main/ann_benchmarks/datasets.py#L427 indicates it's angular (cosine) distance too.

Maybe let's remove this dataset from the benchmarks for now.

[maumueller]

@benfred should be able to shed some light on this.

[benfred]

The original intent was to test out inner-product distance (dot), not angular distance: #91 .

IIRC, the rationale was that certain algorithms either didn't support IP distance - or didn't have good performance when applying transforms like https://www.microsoft.com/en-us/research/wp-content/uploads/2016/02/XboxInnerProduct.pdf to convert IP distance to a cosine space

[erikbern]

I think it's nice to have a dataset for dot products. But I'll fix that after I'm done with this run.

[yuzhenmao]

Hi, just a follow-up to the old question. May I ask why lastfm-64-dot has 65 dimensions while there is a '64' in its name? Is there any transformation applied on this dataset? @maumueller @erikbern

May I assume the dataset is transformed based on https://www.microsoft.com/en-us/research/wp-content/uploads/2016/02/XboxInnerProduct.pdf? (So it actually does Maximum-Inner-Product Search during Euclidean-based ANN Search)