Applicability to binary representations?

[rajarshi]

Hi, I've been playing with the code and it's speed is very impressive. Most of the example employ real valued representations of the objects. Is the method applicable to a binary representation? In particular, I guess the similarity calculations should be appropriate for the representation (e.g., jaccard for binary data, euclidean for real valued data etc).

So I guess the question is, is it appropriate to run this on binary data?

[dkobak]

No distance metric apart from the Euclidean is currently supported. So you can run it on binary data, but only using Euclidean distances.

@linqiaozhi Annoy supports various metrics out of the box. Should we maybe have a metric parameter that simply gets passed on to Annoy? It should be trivial to implement. (One would want to switch off the normalization of the data for custom metrics.)

@rajarshi In case you are using Python, you can try a Python re-implementation of FIt-SNE https://github.com/pavlin-policar/openTSNE which does support various metrics.

[linqiaozhi]

@linqiaozhi Annoy supports various metrics out of the box. Should we maybe have a metric parameter that simply gets passed on to Annoy? It should be trivial to implement. (One would want to switch off the normalization of the data for custom metrics.)

This is a great idea. And we can just give an error if the user tries to use VP trees with a non-Euclidean metric.